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>
24 #include <linux/sched.h>
27 * Targeted preemption latency for CPU-bound tasks:
28 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 unsigned int sysctl_sched_latency
= 6000000ULL;
39 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
42 * The initial- and re-scaling of tunables is configurable
43 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 enum sched_tunable_scaling sysctl_sched_tunable_scaling
51 = SCHED_TUNABLESCALING_LOG
;
54 * Minimal preemption granularity for CPU-bound tasks:
55 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
57 unsigned int sysctl_sched_min_granularity
= 750000ULL;
58 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
61 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
63 static unsigned int sched_nr_latency
= 8;
66 * After fork, child runs first. If set to 0 (default) then
67 * parent will (try to) run first.
69 unsigned int sysctl_sched_child_runs_first __read_mostly
;
72 * sys_sched_yield() compat mode
74 * This option switches the agressive yield implementation of the
75 * old scheduler back on.
77 unsigned int __read_mostly sysctl_sched_compat_yield
;
80 * SCHED_OTHER wake-up granularity.
81 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 * This option delays the preemption effects of decoupled workloads
84 * and reduces their over-scheduling. Synchronous workloads will still
85 * have immediate wakeup/sleep latencies.
87 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
88 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
90 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
92 static const struct sched_class fair_sched_class
;
94 /**************************************************************
95 * CFS operations on generic schedulable entities:
98 #ifdef CONFIG_FAIR_GROUP_SCHED
100 /* cpu runqueue to which this cfs_rq is attached */
101 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
106 /* An entity is a task if it doesn't "own" a runqueue */
107 #define entity_is_task(se) (!se->my_q)
109 static inline struct task_struct
*task_of(struct sched_entity
*se
)
111 #ifdef CONFIG_SCHED_DEBUG
112 WARN_ON_ONCE(!entity_is_task(se
));
114 return container_of(se
, struct task_struct
, se
);
117 /* Walk up scheduling entities hierarchy */
118 #define for_each_sched_entity(se) \
119 for (; se; se = se->parent)
121 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
126 /* runqueue on which this entity is (to be) queued */
127 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
132 /* runqueue "owned" by this group */
133 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
138 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
139 * another cpu ('this_cpu')
141 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
143 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
146 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
148 if (!cfs_rq
->on_list
) {
149 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
150 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
156 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
158 if (cfs_rq
->on_list
) {
159 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
164 /* Iterate thr' all leaf cfs_rq's on a runqueue */
165 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
166 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
168 /* Do the two (enqueued) entities belong to the same group ? */
170 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
172 if (se
->cfs_rq
== pse
->cfs_rq
)
178 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
183 /* return depth at which a sched entity is present in the hierarchy */
184 static inline int depth_se(struct sched_entity
*se
)
188 for_each_sched_entity(se
)
195 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
197 int se_depth
, pse_depth
;
200 * preemption test can be made between sibling entities who are in the
201 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
202 * both tasks until we find their ancestors who are siblings of common
206 /* First walk up until both entities are at same depth */
207 se_depth
= depth_se(*se
);
208 pse_depth
= depth_se(*pse
);
210 while (se_depth
> pse_depth
) {
212 *se
= parent_entity(*se
);
215 while (pse_depth
> se_depth
) {
217 *pse
= parent_entity(*pse
);
220 while (!is_same_group(*se
, *pse
)) {
221 *se
= parent_entity(*se
);
222 *pse
= parent_entity(*pse
);
226 #else /* !CONFIG_FAIR_GROUP_SCHED */
228 static inline struct task_struct
*task_of(struct sched_entity
*se
)
230 return container_of(se
, struct task_struct
, se
);
233 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
235 return container_of(cfs_rq
, struct rq
, cfs
);
238 #define entity_is_task(se) 1
240 #define for_each_sched_entity(se) \
241 for (; se; se = NULL)
243 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
245 return &task_rq(p
)->cfs
;
248 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
250 struct task_struct
*p
= task_of(se
);
251 struct rq
*rq
= task_rq(p
);
256 /* runqueue "owned" by this group */
257 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
262 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
264 return &cpu_rq(this_cpu
)->cfs
;
267 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
271 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
275 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
276 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
279 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
284 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
290 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
294 #endif /* CONFIG_FAIR_GROUP_SCHED */
297 /**************************************************************
298 * Scheduling class tree data structure manipulation methods:
301 static inline u64
max_vruntime(u64 min_vruntime
, u64 vruntime
)
303 s64 delta
= (s64
)(vruntime
- min_vruntime
);
305 min_vruntime
= vruntime
;
310 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
312 s64 delta
= (s64
)(vruntime
- min_vruntime
);
314 min_vruntime
= vruntime
;
319 static inline int entity_before(struct sched_entity
*a
,
320 struct sched_entity
*b
)
322 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
325 static inline s64
entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
327 return se
->vruntime
- cfs_rq
->min_vruntime
;
330 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
332 u64 vruntime
= cfs_rq
->min_vruntime
;
335 vruntime
= cfs_rq
->curr
->vruntime
;
337 if (cfs_rq
->rb_leftmost
) {
338 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
343 vruntime
= se
->vruntime
;
345 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
348 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
352 * Enqueue an entity into the rb-tree:
354 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
356 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
357 struct rb_node
*parent
= NULL
;
358 struct sched_entity
*entry
;
359 s64 key
= entity_key(cfs_rq
, se
);
363 * Find the right place in the rbtree:
367 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
369 * We dont care about collisions. Nodes with
370 * the same key stay together.
372 if (key
< entity_key(cfs_rq
, entry
)) {
373 link
= &parent
->rb_left
;
375 link
= &parent
->rb_right
;
381 * Maintain a cache of leftmost tree entries (it is frequently
385 cfs_rq
->rb_leftmost
= &se
->run_node
;
387 rb_link_node(&se
->run_node
, parent
, link
);
388 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
391 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
393 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
394 struct rb_node
*next_node
;
396 next_node
= rb_next(&se
->run_node
);
397 cfs_rq
->rb_leftmost
= next_node
;
400 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
403 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
405 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
410 return rb_entry(left
, struct sched_entity
, run_node
);
413 static struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
415 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
420 return rb_entry(last
, struct sched_entity
, run_node
);
423 /**************************************************************
424 * Scheduling class statistics methods:
427 #ifdef CONFIG_SCHED_DEBUG
428 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
429 void __user
*buffer
, size_t *lenp
,
432 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
433 int factor
= get_update_sysctl_factor();
438 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
439 sysctl_sched_min_granularity
);
441 #define WRT_SYSCTL(name) \
442 (normalized_sysctl_##name = sysctl_##name / (factor))
443 WRT_SYSCTL(sched_min_granularity
);
444 WRT_SYSCTL(sched_latency
);
445 WRT_SYSCTL(sched_wakeup_granularity
);
455 static inline unsigned long
456 calc_delta_fair(unsigned long delta
, struct sched_entity
*se
)
458 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
459 delta
= calc_delta_mine(delta
, NICE_0_LOAD
, &se
->load
);
465 * The idea is to set a period in which each task runs once.
467 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
468 * this period because otherwise the slices get too small.
470 * p = (nr <= nl) ? l : l*nr/nl
472 static u64
__sched_period(unsigned long nr_running
)
474 u64 period
= sysctl_sched_latency
;
475 unsigned long nr_latency
= sched_nr_latency
;
477 if (unlikely(nr_running
> nr_latency
)) {
478 period
= sysctl_sched_min_granularity
;
479 period
*= nr_running
;
486 * We calculate the wall-time slice from the period by taking a part
487 * proportional to the weight.
491 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
493 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
495 for_each_sched_entity(se
) {
496 struct load_weight
*load
;
497 struct load_weight lw
;
499 cfs_rq
= cfs_rq_of(se
);
500 load
= &cfs_rq
->load
;
502 if (unlikely(!se
->on_rq
)) {
505 update_load_add(&lw
, se
->load
.weight
);
508 slice
= calc_delta_mine(slice
, se
->load
.weight
, load
);
514 * We calculate the vruntime slice of a to be inserted task
518 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
520 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
524 * Update the current task's runtime statistics. Skip current tasks that
525 * are not in our scheduling class.
528 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
529 unsigned long delta_exec
)
531 unsigned long delta_exec_weighted
;
533 schedstat_set(curr
->statistics
.exec_max
,
534 max((u64
)delta_exec
, curr
->statistics
.exec_max
));
536 curr
->sum_exec_runtime
+= delta_exec
;
537 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
538 delta_exec_weighted
= calc_delta_fair(delta_exec
, curr
);
540 curr
->vruntime
+= delta_exec_weighted
;
541 update_min_vruntime(cfs_rq
);
544 static void update_curr(struct cfs_rq
*cfs_rq
)
546 struct sched_entity
*curr
= cfs_rq
->curr
;
547 u64 now
= rq_of(cfs_rq
)->clock_task
;
548 unsigned long delta_exec
;
554 * Get the amount of time the current task was running
555 * since the last time we changed load (this cannot
556 * overflow on 32 bits):
558 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
562 __update_curr(cfs_rq
, curr
, delta_exec
);
563 curr
->exec_start
= now
;
565 if (entity_is_task(curr
)) {
566 struct task_struct
*curtask
= task_of(curr
);
568 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
569 cpuacct_charge(curtask
, delta_exec
);
570 account_group_exec_runtime(curtask
, delta_exec
);
575 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
577 schedstat_set(se
->statistics
.wait_start
, rq_of(cfs_rq
)->clock
);
581 * Task is being enqueued - update stats:
583 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
586 * Are we enqueueing a waiting task? (for current tasks
587 * a dequeue/enqueue event is a NOP)
589 if (se
!= cfs_rq
->curr
)
590 update_stats_wait_start(cfs_rq
, se
);
594 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
596 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
597 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
));
598 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
599 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
600 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
);
601 #ifdef CONFIG_SCHEDSTATS
602 if (entity_is_task(se
)) {
603 trace_sched_stat_wait(task_of(se
),
604 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
);
607 schedstat_set(se
->statistics
.wait_start
, 0);
611 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
614 * Mark the end of the wait period if dequeueing a
617 if (se
!= cfs_rq
->curr
)
618 update_stats_wait_end(cfs_rq
, se
);
622 * We are picking a new current task - update its stats:
625 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
628 * We are starting a new run period:
630 se
->exec_start
= rq_of(cfs_rq
)->clock_task
;
633 /**************************************************
634 * Scheduling class queueing methods:
637 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
639 add_cfs_task_weight(struct cfs_rq
*cfs_rq
, unsigned long weight
)
641 cfs_rq
->task_weight
+= weight
;
645 add_cfs_task_weight(struct cfs_rq
*cfs_rq
, unsigned long weight
)
651 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
653 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
654 if (!parent_entity(se
))
655 inc_cpu_load(rq_of(cfs_rq
), se
->load
.weight
);
656 if (entity_is_task(se
)) {
657 add_cfs_task_weight(cfs_rq
, se
->load
.weight
);
658 list_add(&se
->group_node
, &cfs_rq
->tasks
);
660 cfs_rq
->nr_running
++;
664 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
666 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
667 if (!parent_entity(se
))
668 dec_cpu_load(rq_of(cfs_rq
), se
->load
.weight
);
669 if (entity_is_task(se
)) {
670 add_cfs_task_weight(cfs_rq
, -se
->load
.weight
);
671 list_del_init(&se
->group_node
);
673 cfs_rq
->nr_running
--;
676 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
677 static void update_cfs_load(struct cfs_rq
*cfs_rq
, int lb
)
679 u64 period
= sched_avg_period();
685 now
= rq_of(cfs_rq
)->clock
;
686 delta
= now
- cfs_rq
->load_stamp
;
688 cfs_rq
->load_stamp
= now
;
689 cfs_rq
->load_period
+= delta
;
690 cfs_rq
->load_avg
+= delta
* cfs_rq
->load
.weight
;
692 while (cfs_rq
->load_period
> period
) {
694 * Inline assembly required to prevent the compiler
695 * optimising this loop into a divmod call.
696 * See __iter_div_u64_rem() for another example of this.
698 asm("" : "+rm" (cfs_rq
->load_period
));
699 cfs_rq
->load_period
/= 2;
700 cfs_rq
->load_avg
/= 2;
703 if (lb
&& !cfs_rq
->nr_running
) {
704 if (cfs_rq
->load_avg
< (period
/ 8))
705 list_del_leaf_cfs_rq(cfs_rq
);
709 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
710 unsigned long weight
)
713 account_entity_dequeue(cfs_rq
, se
);
715 update_load_set(&se
->load
, weight
);
718 account_entity_enqueue(cfs_rq
, se
);
721 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
723 struct task_group
*tg
;
724 struct sched_entity
*se
;
725 long load_weight
, load
, shares
;
731 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
735 load
= cfs_rq
->load
.weight
;
737 load_weight
= atomic_read(&tg
->load_weight
);
738 load_weight
-= cfs_rq
->load_contribution
;
741 shares
= (tg
->shares
* load
);
743 shares
/= load_weight
;
745 if (shares
< MIN_SHARES
)
747 if (shares
> tg
->shares
)
750 reweight_entity(cfs_rq_of(se
), se
, shares
);
752 #else /* CONFIG_FAIR_GROUP_SCHED */
753 static inline void update_cfs_load(struct cfs_rq
*cfs_rq
, int lb
)
757 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
760 #endif /* CONFIG_FAIR_GROUP_SCHED */
762 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
764 #ifdef CONFIG_SCHEDSTATS
765 struct task_struct
*tsk
= NULL
;
767 if (entity_is_task(se
))
770 if (se
->statistics
.sleep_start
) {
771 u64 delta
= rq_of(cfs_rq
)->clock
- se
->statistics
.sleep_start
;
776 if (unlikely(delta
> se
->statistics
.sleep_max
))
777 se
->statistics
.sleep_max
= delta
;
779 se
->statistics
.sleep_start
= 0;
780 se
->statistics
.sum_sleep_runtime
+= delta
;
783 account_scheduler_latency(tsk
, delta
>> 10, 1);
784 trace_sched_stat_sleep(tsk
, delta
);
787 if (se
->statistics
.block_start
) {
788 u64 delta
= rq_of(cfs_rq
)->clock
- se
->statistics
.block_start
;
793 if (unlikely(delta
> se
->statistics
.block_max
))
794 se
->statistics
.block_max
= delta
;
796 se
->statistics
.block_start
= 0;
797 se
->statistics
.sum_sleep_runtime
+= delta
;
800 if (tsk
->in_iowait
) {
801 se
->statistics
.iowait_sum
+= delta
;
802 se
->statistics
.iowait_count
++;
803 trace_sched_stat_iowait(tsk
, delta
);
807 * Blocking time is in units of nanosecs, so shift by
808 * 20 to get a milliseconds-range estimation of the
809 * amount of time that the task spent sleeping:
811 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
812 profile_hits(SLEEP_PROFILING
,
813 (void *)get_wchan(tsk
),
816 account_scheduler_latency(tsk
, delta
>> 10, 0);
822 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
824 #ifdef CONFIG_SCHED_DEBUG
825 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
830 if (d
> 3*sysctl_sched_latency
)
831 schedstat_inc(cfs_rq
, nr_spread_over
);
836 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
838 u64 vruntime
= cfs_rq
->min_vruntime
;
841 * The 'current' period is already promised to the current tasks,
842 * however the extra weight of the new task will slow them down a
843 * little, place the new task so that it fits in the slot that
844 * stays open at the end.
846 if (initial
&& sched_feat(START_DEBIT
))
847 vruntime
+= sched_vslice(cfs_rq
, se
);
849 /* sleeps up to a single latency don't count. */
851 unsigned long thresh
= sysctl_sched_latency
;
854 * Halve their sleep time's effect, to allow
855 * for a gentler effect of sleepers:
857 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
863 /* ensure we never gain time by being placed backwards. */
864 vruntime
= max_vruntime(se
->vruntime
, vruntime
);
866 se
->vruntime
= vruntime
;
870 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
873 * Update the normalized vruntime before updating min_vruntime
874 * through callig update_curr().
876 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
877 se
->vruntime
+= cfs_rq
->min_vruntime
;
880 * Update run-time statistics of the 'current'.
883 update_cfs_load(cfs_rq
, 0);
884 account_entity_enqueue(cfs_rq
, se
);
885 update_cfs_shares(cfs_rq
);
887 if (flags
& ENQUEUE_WAKEUP
) {
888 place_entity(cfs_rq
, se
, 0);
889 enqueue_sleeper(cfs_rq
, se
);
892 update_stats_enqueue(cfs_rq
, se
);
893 check_spread(cfs_rq
, se
);
894 if (se
!= cfs_rq
->curr
)
895 __enqueue_entity(cfs_rq
, se
);
898 if (cfs_rq
->nr_running
== 1)
899 list_add_leaf_cfs_rq(cfs_rq
);
902 static void __clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
904 if (!se
|| cfs_rq
->last
== se
)
907 if (!se
|| cfs_rq
->next
== se
)
911 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
913 for_each_sched_entity(se
)
914 __clear_buddies(cfs_rq_of(se
), se
);
918 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
921 * Update run-time statistics of the 'current'.
925 update_stats_dequeue(cfs_rq
, se
);
926 if (flags
& DEQUEUE_SLEEP
) {
927 #ifdef CONFIG_SCHEDSTATS
928 if (entity_is_task(se
)) {
929 struct task_struct
*tsk
= task_of(se
);
931 if (tsk
->state
& TASK_INTERRUPTIBLE
)
932 se
->statistics
.sleep_start
= rq_of(cfs_rq
)->clock
;
933 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
934 se
->statistics
.block_start
= rq_of(cfs_rq
)->clock
;
939 clear_buddies(cfs_rq
, se
);
941 if (se
!= cfs_rq
->curr
)
942 __dequeue_entity(cfs_rq
, se
);
944 update_cfs_load(cfs_rq
, 0);
945 account_entity_dequeue(cfs_rq
, se
);
946 update_min_vruntime(cfs_rq
);
947 update_cfs_shares(cfs_rq
);
950 * Normalize the entity after updating the min_vruntime because the
951 * update can refer to the ->curr item and we need to reflect this
952 * movement in our normalized position.
954 if (!(flags
& DEQUEUE_SLEEP
))
955 se
->vruntime
-= cfs_rq
->min_vruntime
;
959 * Preempt the current task with a newly woken task if needed:
962 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
964 unsigned long ideal_runtime
, delta_exec
;
966 ideal_runtime
= sched_slice(cfs_rq
, curr
);
967 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
968 if (delta_exec
> ideal_runtime
) {
969 resched_task(rq_of(cfs_rq
)->curr
);
971 * The current task ran long enough, ensure it doesn't get
972 * re-elected due to buddy favours.
974 clear_buddies(cfs_rq
, curr
);
979 * Ensure that a task that missed wakeup preemption by a
980 * narrow margin doesn't have to wait for a full slice.
981 * This also mitigates buddy induced latencies under load.
983 if (!sched_feat(WAKEUP_PREEMPT
))
986 if (delta_exec
< sysctl_sched_min_granularity
)
989 if (cfs_rq
->nr_running
> 1) {
990 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
991 s64 delta
= curr
->vruntime
- se
->vruntime
;
993 if (delta
> ideal_runtime
)
994 resched_task(rq_of(cfs_rq
)->curr
);
999 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1001 /* 'current' is not kept within the tree. */
1004 * Any task has to be enqueued before it get to execute on
1005 * a CPU. So account for the time it spent waiting on the
1008 update_stats_wait_end(cfs_rq
, se
);
1009 __dequeue_entity(cfs_rq
, se
);
1012 update_stats_curr_start(cfs_rq
, se
);
1014 #ifdef CONFIG_SCHEDSTATS
1016 * Track our maximum slice length, if the CPU's load is at
1017 * least twice that of our own weight (i.e. dont track it
1018 * when there are only lesser-weight tasks around):
1020 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
1021 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
1022 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
1025 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
1029 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
1031 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
1033 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
1034 struct sched_entity
*left
= se
;
1036 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
1040 * Prefer last buddy, try to return the CPU to a preempted task.
1042 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
1045 clear_buddies(cfs_rq
, se
);
1050 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
1053 * If still on the runqueue then deactivate_task()
1054 * was not called and update_curr() has to be done:
1057 update_curr(cfs_rq
);
1059 check_spread(cfs_rq
, prev
);
1061 update_stats_wait_start(cfs_rq
, prev
);
1062 /* Put 'current' back into the tree. */
1063 __enqueue_entity(cfs_rq
, prev
);
1065 cfs_rq
->curr
= NULL
;
1069 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
1072 * Update run-time statistics of the 'current'.
1074 update_curr(cfs_rq
);
1076 #ifdef CONFIG_SCHED_HRTICK
1078 * queued ticks are scheduled to match the slice, so don't bother
1079 * validating it and just reschedule.
1082 resched_task(rq_of(cfs_rq
)->curr
);
1086 * don't let the period tick interfere with the hrtick preemption
1088 if (!sched_feat(DOUBLE_TICK
) &&
1089 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
1093 if (cfs_rq
->nr_running
> 1 || !sched_feat(WAKEUP_PREEMPT
))
1094 check_preempt_tick(cfs_rq
, curr
);
1097 /**************************************************
1098 * CFS operations on tasks:
1101 #ifdef CONFIG_SCHED_HRTICK
1102 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
1104 struct sched_entity
*se
= &p
->se
;
1105 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1107 WARN_ON(task_rq(p
) != rq
);
1109 if (hrtick_enabled(rq
) && cfs_rq
->nr_running
> 1) {
1110 u64 slice
= sched_slice(cfs_rq
, se
);
1111 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
1112 s64 delta
= slice
- ran
;
1121 * Don't schedule slices shorter than 10000ns, that just
1122 * doesn't make sense. Rely on vruntime for fairness.
1125 delta
= max_t(s64
, 10000LL, delta
);
1127 hrtick_start(rq
, delta
);
1132 * called from enqueue/dequeue and updates the hrtick when the
1133 * current task is from our class and nr_running is low enough
1136 static void hrtick_update(struct rq
*rq
)
1138 struct task_struct
*curr
= rq
->curr
;
1140 if (curr
->sched_class
!= &fair_sched_class
)
1143 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
1144 hrtick_start_fair(rq
, curr
);
1146 #else /* !CONFIG_SCHED_HRTICK */
1148 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
1152 static inline void hrtick_update(struct rq
*rq
)
1158 * The enqueue_task method is called before nr_running is
1159 * increased. Here we update the fair scheduling stats and
1160 * then put the task into the rbtree:
1163 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
1165 struct cfs_rq
*cfs_rq
;
1166 struct sched_entity
*se
= &p
->se
;
1168 for_each_sched_entity(se
) {
1171 cfs_rq
= cfs_rq_of(se
);
1172 enqueue_entity(cfs_rq
, se
, flags
);
1173 flags
= ENQUEUE_WAKEUP
;
1176 for_each_sched_entity(se
) {
1177 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1179 update_cfs_load(cfs_rq
, 0);
1180 update_cfs_shares(cfs_rq
);
1187 * The dequeue_task method is called before nr_running is
1188 * decreased. We remove the task from the rbtree and
1189 * update the fair scheduling stats:
1191 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
1193 struct cfs_rq
*cfs_rq
;
1194 struct sched_entity
*se
= &p
->se
;
1196 for_each_sched_entity(se
) {
1197 cfs_rq
= cfs_rq_of(se
);
1198 dequeue_entity(cfs_rq
, se
, flags
);
1200 /* Don't dequeue parent if it has other entities besides us */
1201 if (cfs_rq
->load
.weight
)
1203 flags
|= DEQUEUE_SLEEP
;
1206 for_each_sched_entity(se
) {
1207 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1209 update_cfs_load(cfs_rq
, 0);
1210 update_cfs_shares(cfs_rq
);
1217 * sched_yield() support is very simple - we dequeue and enqueue.
1219 * If compat_yield is turned on then we requeue to the end of the tree.
1221 static void yield_task_fair(struct rq
*rq
)
1223 struct task_struct
*curr
= rq
->curr
;
1224 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
1225 struct sched_entity
*rightmost
, *se
= &curr
->se
;
1228 * Are we the only task in the tree?
1230 if (unlikely(cfs_rq
->nr_running
== 1))
1233 clear_buddies(cfs_rq
, se
);
1235 if (likely(!sysctl_sched_compat_yield
) && curr
->policy
!= SCHED_BATCH
) {
1236 update_rq_clock(rq
);
1238 * Update run-time statistics of the 'current'.
1240 update_curr(cfs_rq
);
1245 * Find the rightmost entry in the rbtree:
1247 rightmost
= __pick_last_entity(cfs_rq
);
1249 * Already in the rightmost position?
1251 if (unlikely(!rightmost
|| entity_before(rightmost
, se
)))
1255 * Minimally necessary key value to be last in the tree:
1256 * Upon rescheduling, sched_class::put_prev_task() will place
1257 * 'current' within the tree based on its new key value.
1259 se
->vruntime
= rightmost
->vruntime
+ 1;
1264 static void task_waking_fair(struct rq
*rq
, struct task_struct
*p
)
1266 struct sched_entity
*se
= &p
->se
;
1267 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1269 se
->vruntime
-= cfs_rq
->min_vruntime
;
1272 #ifdef CONFIG_FAIR_GROUP_SCHED
1274 * effective_load() calculates the load change as seen from the root_task_group
1276 * Adding load to a group doesn't make a group heavier, but can cause movement
1277 * of group shares between cpus. Assuming the shares were perfectly aligned one
1278 * can calculate the shift in shares.
1280 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
1282 struct sched_entity
*se
= tg
->se
[cpu
];
1287 for_each_sched_entity(se
) {
1288 long S
, rw
, s
, a
, b
;
1290 S
= se
->my_q
->tg
->shares
;
1291 s
= se
->load
.weight
;
1292 rw
= se
->my_q
->load
.weight
;
1303 * Assume the group is already running and will
1304 * thus already be accounted for in the weight.
1306 * That is, moving shares between CPUs, does not
1307 * alter the group weight.
1317 static inline unsigned long effective_load(struct task_group
*tg
, int cpu
,
1318 unsigned long wl
, unsigned long wg
)
1325 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
1327 unsigned long this_load
, load
;
1328 int idx
, this_cpu
, prev_cpu
;
1329 unsigned long tl_per_task
;
1330 struct task_group
*tg
;
1331 unsigned long weight
;
1335 this_cpu
= smp_processor_id();
1336 prev_cpu
= task_cpu(p
);
1337 load
= source_load(prev_cpu
, idx
);
1338 this_load
= target_load(this_cpu
, idx
);
1341 * If sync wakeup then subtract the (maximum possible)
1342 * effect of the currently running task from the load
1343 * of the current CPU:
1347 tg
= task_group(current
);
1348 weight
= current
->se
.load
.weight
;
1350 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
1351 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
1355 weight
= p
->se
.load
.weight
;
1358 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1359 * due to the sync cause above having dropped this_load to 0, we'll
1360 * always have an imbalance, but there's really nothing you can do
1361 * about that, so that's good too.
1363 * Otherwise check if either cpus are near enough in load to allow this
1364 * task to be woken on this_cpu.
1367 unsigned long this_eff_load
, prev_eff_load
;
1369 this_eff_load
= 100;
1370 this_eff_load
*= power_of(prev_cpu
);
1371 this_eff_load
*= this_load
+
1372 effective_load(tg
, this_cpu
, weight
, weight
);
1374 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
1375 prev_eff_load
*= power_of(this_cpu
);
1376 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
1378 balanced
= this_eff_load
<= prev_eff_load
;
1384 * If the currently running task will sleep within
1385 * a reasonable amount of time then attract this newly
1388 if (sync
&& balanced
)
1391 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
1392 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1395 (this_load
<= load
&&
1396 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
1398 * This domain has SD_WAKE_AFFINE and
1399 * p is cache cold in this domain, and
1400 * there is no bad imbalance.
1402 schedstat_inc(sd
, ttwu_move_affine
);
1403 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
1411 * find_idlest_group finds and returns the least busy CPU group within the
1414 static struct sched_group
*
1415 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
1416 int this_cpu
, int load_idx
)
1418 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
1419 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1420 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1423 unsigned long load
, avg_load
;
1427 /* Skip over this group if it has no CPUs allowed */
1428 if (!cpumask_intersects(sched_group_cpus(group
),
1432 local_group
= cpumask_test_cpu(this_cpu
,
1433 sched_group_cpus(group
));
1435 /* Tally up the load of all CPUs in the group */
1438 for_each_cpu(i
, sched_group_cpus(group
)) {
1439 /* Bias balancing toward cpus of our domain */
1441 load
= source_load(i
, load_idx
);
1443 load
= target_load(i
, load_idx
);
1448 /* Adjust by relative CPU power of the group */
1449 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1452 this_load
= avg_load
;
1453 } else if (avg_load
< min_load
) {
1454 min_load
= avg_load
;
1457 } while (group
= group
->next
, group
!= sd
->groups
);
1459 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1465 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1468 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1470 unsigned long load
, min_load
= ULONG_MAX
;
1474 /* Traverse only the allowed CPUs */
1475 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
1476 load
= weighted_cpuload(i
);
1478 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1488 * Try and locate an idle CPU in the sched_domain.
1490 static int select_idle_sibling(struct task_struct
*p
, int target
)
1492 int cpu
= smp_processor_id();
1493 int prev_cpu
= task_cpu(p
);
1494 struct sched_domain
*sd
;
1498 * If the task is going to be woken-up on this cpu and if it is
1499 * already idle, then it is the right target.
1501 if (target
== cpu
&& idle_cpu(cpu
))
1505 * If the task is going to be woken-up on the cpu where it previously
1506 * ran and if it is currently idle, then it the right target.
1508 if (target
== prev_cpu
&& idle_cpu(prev_cpu
))
1512 * Otherwise, iterate the domains and find an elegible idle cpu.
1514 for_each_domain(target
, sd
) {
1515 if (!(sd
->flags
& SD_SHARE_PKG_RESOURCES
))
1518 for_each_cpu_and(i
, sched_domain_span(sd
), &p
->cpus_allowed
) {
1526 * Lets stop looking for an idle sibling when we reached
1527 * the domain that spans the current cpu and prev_cpu.
1529 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
)) &&
1530 cpumask_test_cpu(prev_cpu
, sched_domain_span(sd
)))
1538 * sched_balance_self: balance the current task (running on cpu) in domains
1539 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1542 * Balance, ie. select the least loaded group.
1544 * Returns the target CPU number, or the same CPU if no balancing is needed.
1546 * preempt must be disabled.
1549 select_task_rq_fair(struct rq
*rq
, struct task_struct
*p
, int sd_flag
, int wake_flags
)
1551 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
1552 int cpu
= smp_processor_id();
1553 int prev_cpu
= task_cpu(p
);
1555 int want_affine
= 0;
1557 int sync
= wake_flags
& WF_SYNC
;
1559 if (sd_flag
& SD_BALANCE_WAKE
) {
1560 if (cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
1565 for_each_domain(cpu
, tmp
) {
1566 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
1570 * If power savings logic is enabled for a domain, see if we
1571 * are not overloaded, if so, don't balance wider.
1573 if (tmp
->flags
& (SD_POWERSAVINGS_BALANCE
|SD_PREFER_LOCAL
)) {
1574 unsigned long power
= 0;
1575 unsigned long nr_running
= 0;
1576 unsigned long capacity
;
1579 for_each_cpu(i
, sched_domain_span(tmp
)) {
1580 power
+= power_of(i
);
1581 nr_running
+= cpu_rq(i
)->cfs
.nr_running
;
1584 capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
1586 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1589 if (nr_running
< capacity
)
1594 * If both cpu and prev_cpu are part of this domain,
1595 * cpu is a valid SD_WAKE_AFFINE target.
1597 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
1598 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
1603 if (!want_sd
&& !want_affine
)
1606 if (!(tmp
->flags
& sd_flag
))
1614 if (cpu
== prev_cpu
|| wake_affine(affine_sd
, p
, sync
))
1615 return select_idle_sibling(p
, cpu
);
1617 return select_idle_sibling(p
, prev_cpu
);
1621 int load_idx
= sd
->forkexec_idx
;
1622 struct sched_group
*group
;
1625 if (!(sd
->flags
& sd_flag
)) {
1630 if (sd_flag
& SD_BALANCE_WAKE
)
1631 load_idx
= sd
->wake_idx
;
1633 group
= find_idlest_group(sd
, p
, cpu
, load_idx
);
1639 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
1640 if (new_cpu
== -1 || new_cpu
== cpu
) {
1641 /* Now try balancing at a lower domain level of cpu */
1646 /* Now try balancing at a lower domain level of new_cpu */
1648 weight
= sd
->span_weight
;
1650 for_each_domain(cpu
, tmp
) {
1651 if (weight
<= tmp
->span_weight
)
1653 if (tmp
->flags
& sd_flag
)
1656 /* while loop will break here if sd == NULL */
1661 #endif /* CONFIG_SMP */
1663 static unsigned long
1664 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
1666 unsigned long gran
= sysctl_sched_wakeup_granularity
;
1669 * Since its curr running now, convert the gran from real-time
1670 * to virtual-time in his units.
1672 * By using 'se' instead of 'curr' we penalize light tasks, so
1673 * they get preempted easier. That is, if 'se' < 'curr' then
1674 * the resulting gran will be larger, therefore penalizing the
1675 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1676 * be smaller, again penalizing the lighter task.
1678 * This is especially important for buddies when the leftmost
1679 * task is higher priority than the buddy.
1681 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
1682 gran
= calc_delta_fair(gran
, se
);
1688 * Should 'se' preempt 'curr'.
1702 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
1704 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
1709 gran
= wakeup_gran(curr
, se
);
1716 static void set_last_buddy(struct sched_entity
*se
)
1718 if (likely(task_of(se
)->policy
!= SCHED_IDLE
)) {
1719 for_each_sched_entity(se
)
1720 cfs_rq_of(se
)->last
= se
;
1724 static void set_next_buddy(struct sched_entity
*se
)
1726 if (likely(task_of(se
)->policy
!= SCHED_IDLE
)) {
1727 for_each_sched_entity(se
)
1728 cfs_rq_of(se
)->next
= se
;
1733 * Preempt the current task with a newly woken task if needed:
1735 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1737 struct task_struct
*curr
= rq
->curr
;
1738 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
1739 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
1740 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
1742 if (unlikely(rt_prio(p
->prio
)))
1745 if (unlikely(p
->sched_class
!= &fair_sched_class
))
1748 if (unlikely(se
== pse
))
1751 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
))
1752 set_next_buddy(pse
);
1755 * We can come here with TIF_NEED_RESCHED already set from new task
1758 if (test_tsk_need_resched(curr
))
1762 * Batch and idle tasks do not preempt (their preemption is driven by
1765 if (unlikely(p
->policy
!= SCHED_NORMAL
))
1768 /* Idle tasks are by definition preempted by everybody. */
1769 if (unlikely(curr
->policy
== SCHED_IDLE
))
1772 if (!sched_feat(WAKEUP_PREEMPT
))
1775 update_curr(cfs_rq
);
1776 find_matching_se(&se
, &pse
);
1778 if (wakeup_preempt_entity(se
, pse
) == 1)
1786 * Only set the backward buddy when the current task is still
1787 * on the rq. This can happen when a wakeup gets interleaved
1788 * with schedule on the ->pre_schedule() or idle_balance()
1789 * point, either of which can * drop the rq lock.
1791 * Also, during early boot the idle thread is in the fair class,
1792 * for obvious reasons its a bad idea to schedule back to it.
1794 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
1797 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
1801 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
1803 struct task_struct
*p
;
1804 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
1805 struct sched_entity
*se
;
1807 if (!cfs_rq
->nr_running
)
1811 se
= pick_next_entity(cfs_rq
);
1812 set_next_entity(cfs_rq
, se
);
1813 cfs_rq
= group_cfs_rq(se
);
1817 hrtick_start_fair(rq
, p
);
1823 * Account for a descheduled task:
1825 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
1827 struct sched_entity
*se
= &prev
->se
;
1828 struct cfs_rq
*cfs_rq
;
1830 for_each_sched_entity(se
) {
1831 cfs_rq
= cfs_rq_of(se
);
1832 put_prev_entity(cfs_rq
, se
);
1837 /**************************************************
1838 * Fair scheduling class load-balancing methods:
1842 * pull_task - move a task from a remote runqueue to the local runqueue.
1843 * Both runqueues must be locked.
1845 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
1846 struct rq
*this_rq
, int this_cpu
)
1848 deactivate_task(src_rq
, p
, 0);
1849 set_task_cpu(p
, this_cpu
);
1850 activate_task(this_rq
, p
, 0);
1851 check_preempt_curr(this_rq
, p
, 0);
1853 /* re-arm NEWIDLE balancing when moving tasks */
1854 src_rq
->avg_idle
= this_rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
1855 this_rq
->idle_stamp
= 0;
1859 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1862 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
1863 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1866 int tsk_cache_hot
= 0;
1868 * We do not migrate tasks that are:
1869 * 1) running (obviously), or
1870 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1871 * 3) are cache-hot on their current CPU.
1873 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
1874 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
1879 if (task_running(rq
, p
)) {
1880 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
1885 * Aggressive migration if:
1886 * 1) task is cache cold, or
1887 * 2) too many balance attempts have failed.
1890 tsk_cache_hot
= task_hot(p
, rq
->clock_task
, sd
);
1891 if (!tsk_cache_hot
||
1892 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
1893 #ifdef CONFIG_SCHEDSTATS
1894 if (tsk_cache_hot
) {
1895 schedstat_inc(sd
, lb_hot_gained
[idle
]);
1896 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
1902 if (tsk_cache_hot
) {
1903 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
1910 * move_one_task tries to move exactly one task from busiest to this_rq, as
1911 * part of active balancing operations within "domain".
1912 * Returns 1 if successful and 0 otherwise.
1914 * Called with both runqueues locked.
1917 move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1918 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1920 struct task_struct
*p
, *n
;
1921 struct cfs_rq
*cfs_rq
;
1924 for_each_leaf_cfs_rq(busiest
, cfs_rq
) {
1925 list_for_each_entry_safe(p
, n
, &cfs_rq
->tasks
, se
.group_node
) {
1927 if (!can_migrate_task(p
, busiest
, this_cpu
,
1931 pull_task(busiest
, p
, this_rq
, this_cpu
);
1933 * Right now, this is only the second place pull_task()
1934 * is called, so we can safely collect pull_task()
1935 * stats here rather than inside pull_task().
1937 schedstat_inc(sd
, lb_gained
[idle
]);
1945 static unsigned long
1946 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1947 unsigned long max_load_move
, struct sched_domain
*sd
,
1948 enum cpu_idle_type idle
, int *all_pinned
,
1949 int *this_best_prio
, struct cfs_rq
*busiest_cfs_rq
)
1951 int loops
= 0, pulled
= 0, pinned
= 0;
1952 long rem_load_move
= max_load_move
;
1953 struct task_struct
*p
, *n
;
1955 if (max_load_move
== 0)
1960 list_for_each_entry_safe(p
, n
, &busiest_cfs_rq
->tasks
, se
.group_node
) {
1961 if (loops
++ > sysctl_sched_nr_migrate
)
1964 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
1965 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
))
1968 pull_task(busiest
, p
, this_rq
, this_cpu
);
1970 rem_load_move
-= p
->se
.load
.weight
;
1972 #ifdef CONFIG_PREEMPT
1974 * NEWIDLE balancing is a source of latency, so preemptible
1975 * kernels will stop after the first task is pulled to minimize
1976 * the critical section.
1978 if (idle
== CPU_NEWLY_IDLE
)
1983 * We only want to steal up to the prescribed amount of
1986 if (rem_load_move
<= 0)
1989 if (p
->prio
< *this_best_prio
)
1990 *this_best_prio
= p
->prio
;
1994 * Right now, this is one of only two places pull_task() is called,
1995 * so we can safely collect pull_task() stats here rather than
1996 * inside pull_task().
1998 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2001 *all_pinned
= pinned
;
2003 return max_load_move
- rem_load_move
;
2006 #ifdef CONFIG_FAIR_GROUP_SCHED
2008 * update tg->load_weight by folding this cpu's load_avg
2010 static int tg_shares_up(struct task_group
*tg
, int cpu
)
2012 struct cfs_rq
*cfs_rq
;
2013 unsigned long flags
;
2021 cfs_rq
= tg
->cfs_rq
[cpu
];
2023 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2025 update_rq_clock(rq
);
2026 update_cfs_load(cfs_rq
, 1);
2028 load_avg
= div64_u64(cfs_rq
->load_avg
, cfs_rq
->load_period
+1);
2029 load_avg
-= cfs_rq
->load_contribution
;
2030 atomic_add(load_avg
, &tg
->load_weight
);
2031 cfs_rq
->load_contribution
+= load_avg
;
2034 * We need to update shares after updating tg->load_weight in
2035 * order to adjust the weight of groups with long running tasks.
2037 update_cfs_shares(cfs_rq
);
2039 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2044 static void update_shares(int cpu
)
2046 struct cfs_rq
*cfs_rq
;
2047 struct rq
*rq
= cpu_rq(cpu
);
2050 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
2051 struct task_group
*tg
= cfs_rq
->tg
;
2054 tg_shares_up(tg
, cpu
);
2061 static unsigned long
2062 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2063 unsigned long max_load_move
,
2064 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2065 int *all_pinned
, int *this_best_prio
)
2067 long rem_load_move
= max_load_move
;
2068 int busiest_cpu
= cpu_of(busiest
);
2069 struct task_group
*tg
;
2072 update_h_load(busiest_cpu
);
2074 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
2075 struct cfs_rq
*busiest_cfs_rq
= tg
->cfs_rq
[busiest_cpu
];
2076 unsigned long busiest_h_load
= busiest_cfs_rq
->h_load
;
2077 unsigned long busiest_weight
= busiest_cfs_rq
->load
.weight
;
2078 u64 rem_load
, moved_load
;
2083 if (!busiest_cfs_rq
->task_weight
)
2086 rem_load
= (u64
)rem_load_move
* busiest_weight
;
2087 rem_load
= div_u64(rem_load
, busiest_h_load
+ 1);
2089 moved_load
= balance_tasks(this_rq
, this_cpu
, busiest
,
2090 rem_load
, sd
, idle
, all_pinned
, this_best_prio
,
2096 moved_load
*= busiest_h_load
;
2097 moved_load
= div_u64(moved_load
, busiest_weight
+ 1);
2099 rem_load_move
-= moved_load
;
2100 if (rem_load_move
< 0)
2105 return max_load_move
- rem_load_move
;
2108 static inline void update_shares(int cpu
)
2112 static unsigned long
2113 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2114 unsigned long max_load_move
,
2115 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2116 int *all_pinned
, int *this_best_prio
)
2118 return balance_tasks(this_rq
, this_cpu
, busiest
,
2119 max_load_move
, sd
, idle
, all_pinned
,
2120 this_best_prio
, &busiest
->cfs
);
2125 * move_tasks tries to move up to max_load_move weighted load from busiest to
2126 * this_rq, as part of a balancing operation within domain "sd".
2127 * Returns 1 if successful and 0 otherwise.
2129 * Called with both runqueues locked.
2131 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2132 unsigned long max_load_move
,
2133 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2136 unsigned long total_load_moved
= 0, load_moved
;
2137 int this_best_prio
= this_rq
->curr
->prio
;
2140 load_moved
= load_balance_fair(this_rq
, this_cpu
, busiest
,
2141 max_load_move
- total_load_moved
,
2142 sd
, idle
, all_pinned
, &this_best_prio
);
2144 total_load_moved
+= load_moved
;
2146 #ifdef CONFIG_PREEMPT
2148 * NEWIDLE balancing is a source of latency, so preemptible
2149 * kernels will stop after the first task is pulled to minimize
2150 * the critical section.
2152 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
2155 if (raw_spin_is_contended(&this_rq
->lock
) ||
2156 raw_spin_is_contended(&busiest
->lock
))
2159 } while (load_moved
&& max_load_move
> total_load_moved
);
2161 return total_load_moved
> 0;
2164 /********** Helpers for find_busiest_group ************************/
2166 * sd_lb_stats - Structure to store the statistics of a sched_domain
2167 * during load balancing.
2169 struct sd_lb_stats
{
2170 struct sched_group
*busiest
; /* Busiest group in this sd */
2171 struct sched_group
*this; /* Local group in this sd */
2172 unsigned long total_load
; /* Total load of all groups in sd */
2173 unsigned long total_pwr
; /* Total power of all groups in sd */
2174 unsigned long avg_load
; /* Average load across all groups in sd */
2176 /** Statistics of this group */
2177 unsigned long this_load
;
2178 unsigned long this_load_per_task
;
2179 unsigned long this_nr_running
;
2180 unsigned long this_has_capacity
;
2182 /* Statistics of the busiest group */
2183 unsigned long max_load
;
2184 unsigned long busiest_load_per_task
;
2185 unsigned long busiest_nr_running
;
2186 unsigned long busiest_group_capacity
;
2187 unsigned long busiest_has_capacity
;
2189 int group_imb
; /* Is there imbalance in this sd */
2190 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2191 int power_savings_balance
; /* Is powersave balance needed for this sd */
2192 struct sched_group
*group_min
; /* Least loaded group in sd */
2193 struct sched_group
*group_leader
; /* Group which relieves group_min */
2194 unsigned long min_load_per_task
; /* load_per_task in group_min */
2195 unsigned long leader_nr_running
; /* Nr running of group_leader */
2196 unsigned long min_nr_running
; /* Nr running of group_min */
2201 * sg_lb_stats - stats of a sched_group required for load_balancing
2203 struct sg_lb_stats
{
2204 unsigned long avg_load
; /*Avg load across the CPUs of the group */
2205 unsigned long group_load
; /* Total load over the CPUs of the group */
2206 unsigned long sum_nr_running
; /* Nr tasks running in the group */
2207 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
2208 unsigned long group_capacity
;
2209 int group_imb
; /* Is there an imbalance in the group ? */
2210 int group_has_capacity
; /* Is there extra capacity in the group? */
2214 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2215 * @group: The group whose first cpu is to be returned.
2217 static inline unsigned int group_first_cpu(struct sched_group
*group
)
2219 return cpumask_first(sched_group_cpus(group
));
2223 * get_sd_load_idx - Obtain the load index for a given sched domain.
2224 * @sd: The sched_domain whose load_idx is to be obtained.
2225 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2227 static inline int get_sd_load_idx(struct sched_domain
*sd
,
2228 enum cpu_idle_type idle
)
2234 load_idx
= sd
->busy_idx
;
2237 case CPU_NEWLY_IDLE
:
2238 load_idx
= sd
->newidle_idx
;
2241 load_idx
= sd
->idle_idx
;
2249 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2251 * init_sd_power_savings_stats - Initialize power savings statistics for
2252 * the given sched_domain, during load balancing.
2254 * @sd: Sched domain whose power-savings statistics are to be initialized.
2255 * @sds: Variable containing the statistics for sd.
2256 * @idle: Idle status of the CPU at which we're performing load-balancing.
2258 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
2259 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
2262 * Busy processors will not participate in power savings
2265 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2266 sds
->power_savings_balance
= 0;
2268 sds
->power_savings_balance
= 1;
2269 sds
->min_nr_running
= ULONG_MAX
;
2270 sds
->leader_nr_running
= 0;
2275 * update_sd_power_savings_stats - Update the power saving stats for a
2276 * sched_domain while performing load balancing.
2278 * @group: sched_group belonging to the sched_domain under consideration.
2279 * @sds: Variable containing the statistics of the sched_domain
2280 * @local_group: Does group contain the CPU for which we're performing
2282 * @sgs: Variable containing the statistics of the group.
2284 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
2285 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
2288 if (!sds
->power_savings_balance
)
2292 * If the local group is idle or completely loaded
2293 * no need to do power savings balance at this domain
2295 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
2296 !sds
->this_nr_running
))
2297 sds
->power_savings_balance
= 0;
2300 * If a group is already running at full capacity or idle,
2301 * don't include that group in power savings calculations
2303 if (!sds
->power_savings_balance
||
2304 sgs
->sum_nr_running
>= sgs
->group_capacity
||
2305 !sgs
->sum_nr_running
)
2309 * Calculate the group which has the least non-idle load.
2310 * This is the group from where we need to pick up the load
2313 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
2314 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
2315 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
2316 sds
->group_min
= group
;
2317 sds
->min_nr_running
= sgs
->sum_nr_running
;
2318 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
2319 sgs
->sum_nr_running
;
2323 * Calculate the group which is almost near its
2324 * capacity but still has some space to pick up some load
2325 * from other group and save more power
2327 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
2330 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
2331 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
2332 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
2333 sds
->group_leader
= group
;
2334 sds
->leader_nr_running
= sgs
->sum_nr_running
;
2339 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2340 * @sds: Variable containing the statistics of the sched_domain
2341 * under consideration.
2342 * @this_cpu: Cpu at which we're currently performing load-balancing.
2343 * @imbalance: Variable to store the imbalance.
2346 * Check if we have potential to perform some power-savings balance.
2347 * If yes, set the busiest group to be the least loaded group in the
2348 * sched_domain, so that it's CPUs can be put to idle.
2350 * Returns 1 if there is potential to perform power-savings balance.
2353 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
2354 int this_cpu
, unsigned long *imbalance
)
2356 if (!sds
->power_savings_balance
)
2359 if (sds
->this != sds
->group_leader
||
2360 sds
->group_leader
== sds
->group_min
)
2363 *imbalance
= sds
->min_load_per_task
;
2364 sds
->busiest
= sds
->group_min
;
2369 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2370 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
2371 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
2376 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
2377 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
2382 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
2383 int this_cpu
, unsigned long *imbalance
)
2387 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2390 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
2392 return SCHED_LOAD_SCALE
;
2395 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
2397 return default_scale_freq_power(sd
, cpu
);
2400 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
2402 unsigned long weight
= sd
->span_weight
;
2403 unsigned long smt_gain
= sd
->smt_gain
;
2410 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
2412 return default_scale_smt_power(sd
, cpu
);
2415 unsigned long scale_rt_power(int cpu
)
2417 struct rq
*rq
= cpu_rq(cpu
);
2418 u64 total
, available
;
2420 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
2422 if (unlikely(total
< rq
->rt_avg
)) {
2423 /* Ensures that power won't end up being negative */
2426 available
= total
- rq
->rt_avg
;
2429 if (unlikely((s64
)total
< SCHED_LOAD_SCALE
))
2430 total
= SCHED_LOAD_SCALE
;
2432 total
>>= SCHED_LOAD_SHIFT
;
2434 return div_u64(available
, total
);
2437 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
2439 unsigned long weight
= sd
->span_weight
;
2440 unsigned long power
= SCHED_LOAD_SCALE
;
2441 struct sched_group
*sdg
= sd
->groups
;
2443 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
2444 if (sched_feat(ARCH_POWER
))
2445 power
*= arch_scale_smt_power(sd
, cpu
);
2447 power
*= default_scale_smt_power(sd
, cpu
);
2449 power
>>= SCHED_LOAD_SHIFT
;
2452 sdg
->cpu_power_orig
= power
;
2454 if (sched_feat(ARCH_POWER
))
2455 power
*= arch_scale_freq_power(sd
, cpu
);
2457 power
*= default_scale_freq_power(sd
, cpu
);
2459 power
>>= SCHED_LOAD_SHIFT
;
2461 power
*= scale_rt_power(cpu
);
2462 power
>>= SCHED_LOAD_SHIFT
;
2467 cpu_rq(cpu
)->cpu_power
= power
;
2468 sdg
->cpu_power
= power
;
2471 static void update_group_power(struct sched_domain
*sd
, int cpu
)
2473 struct sched_domain
*child
= sd
->child
;
2474 struct sched_group
*group
, *sdg
= sd
->groups
;
2475 unsigned long power
;
2478 update_cpu_power(sd
, cpu
);
2484 group
= child
->groups
;
2486 power
+= group
->cpu_power
;
2487 group
= group
->next
;
2488 } while (group
!= child
->groups
);
2490 sdg
->cpu_power
= power
;
2494 * Try and fix up capacity for tiny siblings, this is needed when
2495 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2496 * which on its own isn't powerful enough.
2498 * See update_sd_pick_busiest() and check_asym_packing().
2501 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
2504 * Only siblings can have significantly less than SCHED_LOAD_SCALE
2506 if (sd
->level
!= SD_LV_SIBLING
)
2510 * If ~90% of the cpu_power is still there, we're good.
2512 if (group
->cpu_power
* 32 > group
->cpu_power_orig
* 29)
2519 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2520 * @sd: The sched_domain whose statistics are to be updated.
2521 * @group: sched_group whose statistics are to be updated.
2522 * @this_cpu: Cpu for which load balance is currently performed.
2523 * @idle: Idle status of this_cpu
2524 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2525 * @sd_idle: Idle status of the sched_domain containing group.
2526 * @local_group: Does group contain this_cpu.
2527 * @cpus: Set of cpus considered for load balancing.
2528 * @balance: Should we balance.
2529 * @sgs: variable to hold the statistics for this group.
2531 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
2532 struct sched_group
*group
, int this_cpu
,
2533 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
2534 int local_group
, const struct cpumask
*cpus
,
2535 int *balance
, struct sg_lb_stats
*sgs
)
2537 unsigned long load
, max_cpu_load
, min_cpu_load
, max_nr_running
;
2539 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2540 unsigned long avg_load_per_task
= 0;
2543 balance_cpu
= group_first_cpu(group
);
2545 /* Tally up the load of all CPUs in the group */
2547 min_cpu_load
= ~0UL;
2550 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
2551 struct rq
*rq
= cpu_rq(i
);
2553 if (*sd_idle
&& rq
->nr_running
)
2556 /* Bias balancing toward cpus of our domain */
2558 if (idle_cpu(i
) && !first_idle_cpu
) {
2563 load
= target_load(i
, load_idx
);
2565 load
= source_load(i
, load_idx
);
2566 if (load
> max_cpu_load
) {
2567 max_cpu_load
= load
;
2568 max_nr_running
= rq
->nr_running
;
2570 if (min_cpu_load
> load
)
2571 min_cpu_load
= load
;
2574 sgs
->group_load
+= load
;
2575 sgs
->sum_nr_running
+= rq
->nr_running
;
2576 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
2581 * First idle cpu or the first cpu(busiest) in this sched group
2582 * is eligible for doing load balancing at this and above
2583 * domains. In the newly idle case, we will allow all the cpu's
2584 * to do the newly idle load balance.
2586 if (idle
!= CPU_NEWLY_IDLE
&& local_group
) {
2587 if (balance_cpu
!= this_cpu
) {
2591 update_group_power(sd
, this_cpu
);
2594 /* Adjust by relative CPU power of the group */
2595 sgs
->avg_load
= (sgs
->group_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2598 * Consider the group unbalanced when the imbalance is larger
2599 * than the average weight of two tasks.
2601 * APZ: with cgroup the avg task weight can vary wildly and
2602 * might not be a suitable number - should we keep a
2603 * normalized nr_running number somewhere that negates
2606 if (sgs
->sum_nr_running
)
2607 avg_load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
2609 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
&& max_nr_running
> 1)
2612 sgs
->group_capacity
= DIV_ROUND_CLOSEST(group
->cpu_power
, SCHED_LOAD_SCALE
);
2613 if (!sgs
->group_capacity
)
2614 sgs
->group_capacity
= fix_small_capacity(sd
, group
);
2616 if (sgs
->group_capacity
> sgs
->sum_nr_running
)
2617 sgs
->group_has_capacity
= 1;
2621 * update_sd_pick_busiest - return 1 on busiest group
2622 * @sd: sched_domain whose statistics are to be checked
2623 * @sds: sched_domain statistics
2624 * @sg: sched_group candidate to be checked for being the busiest
2625 * @sgs: sched_group statistics
2626 * @this_cpu: the current cpu
2628 * Determine if @sg is a busier group than the previously selected
2631 static bool update_sd_pick_busiest(struct sched_domain
*sd
,
2632 struct sd_lb_stats
*sds
,
2633 struct sched_group
*sg
,
2634 struct sg_lb_stats
*sgs
,
2637 if (sgs
->avg_load
<= sds
->max_load
)
2640 if (sgs
->sum_nr_running
> sgs
->group_capacity
)
2647 * ASYM_PACKING needs to move all the work to the lowest
2648 * numbered CPUs in the group, therefore mark all groups
2649 * higher than ourself as busy.
2651 if ((sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
2652 this_cpu
< group_first_cpu(sg
)) {
2656 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
2664 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2665 * @sd: sched_domain whose statistics are to be updated.
2666 * @this_cpu: Cpu for which load balance is currently performed.
2667 * @idle: Idle status of this_cpu
2668 * @sd_idle: Idle status of the sched_domain containing sg.
2669 * @cpus: Set of cpus considered for load balancing.
2670 * @balance: Should we balance.
2671 * @sds: variable to hold the statistics for this sched_domain.
2673 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
2674 enum cpu_idle_type idle
, int *sd_idle
,
2675 const struct cpumask
*cpus
, int *balance
,
2676 struct sd_lb_stats
*sds
)
2678 struct sched_domain
*child
= sd
->child
;
2679 struct sched_group
*sg
= sd
->groups
;
2680 struct sg_lb_stats sgs
;
2681 int load_idx
, prefer_sibling
= 0;
2683 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
2686 init_sd_power_savings_stats(sd
, sds
, idle
);
2687 load_idx
= get_sd_load_idx(sd
, idle
);
2692 local_group
= cpumask_test_cpu(this_cpu
, sched_group_cpus(sg
));
2693 memset(&sgs
, 0, sizeof(sgs
));
2694 update_sg_lb_stats(sd
, sg
, this_cpu
, idle
, load_idx
, sd_idle
,
2695 local_group
, cpus
, balance
, &sgs
);
2697 if (local_group
&& !(*balance
))
2700 sds
->total_load
+= sgs
.group_load
;
2701 sds
->total_pwr
+= sg
->cpu_power
;
2704 * In case the child domain prefers tasks go to siblings
2705 * first, lower the sg capacity to one so that we'll try
2706 * and move all the excess tasks away. We lower the capacity
2707 * of a group only if the local group has the capacity to fit
2708 * these excess tasks, i.e. nr_running < group_capacity. The
2709 * extra check prevents the case where you always pull from the
2710 * heaviest group when it is already under-utilized (possible
2711 * with a large weight task outweighs the tasks on the system).
2713 if (prefer_sibling
&& !local_group
&& sds
->this_has_capacity
)
2714 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
2717 sds
->this_load
= sgs
.avg_load
;
2719 sds
->this_nr_running
= sgs
.sum_nr_running
;
2720 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
2721 sds
->this_has_capacity
= sgs
.group_has_capacity
;
2722 } else if (update_sd_pick_busiest(sd
, sds
, sg
, &sgs
, this_cpu
)) {
2723 sds
->max_load
= sgs
.avg_load
;
2725 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
2726 sds
->busiest_group_capacity
= sgs
.group_capacity
;
2727 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
2728 sds
->busiest_has_capacity
= sgs
.group_has_capacity
;
2729 sds
->group_imb
= sgs
.group_imb
;
2732 update_sd_power_savings_stats(sg
, sds
, local_group
, &sgs
);
2734 } while (sg
!= sd
->groups
);
2737 int __weak
arch_sd_sibling_asym_packing(void)
2739 return 0*SD_ASYM_PACKING
;
2743 * check_asym_packing - Check to see if the group is packed into the
2746 * This is primarily intended to used at the sibling level. Some
2747 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2748 * case of POWER7, it can move to lower SMT modes only when higher
2749 * threads are idle. When in lower SMT modes, the threads will
2750 * perform better since they share less core resources. Hence when we
2751 * have idle threads, we want them to be the higher ones.
2753 * This packing function is run on idle threads. It checks to see if
2754 * the busiest CPU in this domain (core in the P7 case) has a higher
2755 * CPU number than the packing function is being run on. Here we are
2756 * assuming lower CPU number will be equivalent to lower a SMT thread
2759 * Returns 1 when packing is required and a task should be moved to
2760 * this CPU. The amount of the imbalance is returned in *imbalance.
2762 * @sd: The sched_domain whose packing is to be checked.
2763 * @sds: Statistics of the sched_domain which is to be packed
2764 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2765 * @imbalance: returns amount of imbalanced due to packing.
2767 static int check_asym_packing(struct sched_domain
*sd
,
2768 struct sd_lb_stats
*sds
,
2769 int this_cpu
, unsigned long *imbalance
)
2773 if (!(sd
->flags
& SD_ASYM_PACKING
))
2779 busiest_cpu
= group_first_cpu(sds
->busiest
);
2780 if (this_cpu
> busiest_cpu
)
2783 *imbalance
= DIV_ROUND_CLOSEST(sds
->max_load
* sds
->busiest
->cpu_power
,
2789 * fix_small_imbalance - Calculate the minor imbalance that exists
2790 * amongst the groups of a sched_domain, during
2792 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2793 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2794 * @imbalance: Variable to store the imbalance.
2796 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
2797 int this_cpu
, unsigned long *imbalance
)
2799 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
2800 unsigned int imbn
= 2;
2801 unsigned long scaled_busy_load_per_task
;
2803 if (sds
->this_nr_running
) {
2804 sds
->this_load_per_task
/= sds
->this_nr_running
;
2805 if (sds
->busiest_load_per_task
>
2806 sds
->this_load_per_task
)
2809 sds
->this_load_per_task
=
2810 cpu_avg_load_per_task(this_cpu
);
2812 scaled_busy_load_per_task
= sds
->busiest_load_per_task
2814 scaled_busy_load_per_task
/= sds
->busiest
->cpu_power
;
2816 if (sds
->max_load
- sds
->this_load
+ scaled_busy_load_per_task
>=
2817 (scaled_busy_load_per_task
* imbn
)) {
2818 *imbalance
= sds
->busiest_load_per_task
;
2823 * OK, we don't have enough imbalance to justify moving tasks,
2824 * however we may be able to increase total CPU power used by
2828 pwr_now
+= sds
->busiest
->cpu_power
*
2829 min(sds
->busiest_load_per_task
, sds
->max_load
);
2830 pwr_now
+= sds
->this->cpu_power
*
2831 min(sds
->this_load_per_task
, sds
->this_load
);
2832 pwr_now
/= SCHED_LOAD_SCALE
;
2834 /* Amount of load we'd subtract */
2835 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
2836 sds
->busiest
->cpu_power
;
2837 if (sds
->max_load
> tmp
)
2838 pwr_move
+= sds
->busiest
->cpu_power
*
2839 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
2841 /* Amount of load we'd add */
2842 if (sds
->max_load
* sds
->busiest
->cpu_power
<
2843 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
2844 tmp
= (sds
->max_load
* sds
->busiest
->cpu_power
) /
2845 sds
->this->cpu_power
;
2847 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
2848 sds
->this->cpu_power
;
2849 pwr_move
+= sds
->this->cpu_power
*
2850 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
2851 pwr_move
/= SCHED_LOAD_SCALE
;
2853 /* Move if we gain throughput */
2854 if (pwr_move
> pwr_now
)
2855 *imbalance
= sds
->busiest_load_per_task
;
2859 * calculate_imbalance - Calculate the amount of imbalance present within the
2860 * groups of a given sched_domain during load balance.
2861 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
2862 * @this_cpu: Cpu for which currently load balance is being performed.
2863 * @imbalance: The variable to store the imbalance.
2865 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
2866 unsigned long *imbalance
)
2868 unsigned long max_pull
, load_above_capacity
= ~0UL;
2870 sds
->busiest_load_per_task
/= sds
->busiest_nr_running
;
2871 if (sds
->group_imb
) {
2872 sds
->busiest_load_per_task
=
2873 min(sds
->busiest_load_per_task
, sds
->avg_load
);
2877 * In the presence of smp nice balancing, certain scenarios can have
2878 * max load less than avg load(as we skip the groups at or below
2879 * its cpu_power, while calculating max_load..)
2881 if (sds
->max_load
< sds
->avg_load
) {
2883 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
2886 if (!sds
->group_imb
) {
2888 * Don't want to pull so many tasks that a group would go idle.
2890 load_above_capacity
= (sds
->busiest_nr_running
-
2891 sds
->busiest_group_capacity
);
2893 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_LOAD_SCALE
);
2895 load_above_capacity
/= sds
->busiest
->cpu_power
;
2899 * We're trying to get all the cpus to the average_load, so we don't
2900 * want to push ourselves above the average load, nor do we wish to
2901 * reduce the max loaded cpu below the average load. At the same time,
2902 * we also don't want to reduce the group load below the group capacity
2903 * (so that we can implement power-savings policies etc). Thus we look
2904 * for the minimum possible imbalance.
2905 * Be careful of negative numbers as they'll appear as very large values
2906 * with unsigned longs.
2908 max_pull
= min(sds
->max_load
- sds
->avg_load
, load_above_capacity
);
2910 /* How much load to actually move to equalise the imbalance */
2911 *imbalance
= min(max_pull
* sds
->busiest
->cpu_power
,
2912 (sds
->avg_load
- sds
->this_load
) * sds
->this->cpu_power
)
2916 * if *imbalance is less than the average load per runnable task
2917 * there is no gaurantee that any tasks will be moved so we'll have
2918 * a think about bumping its value to force at least one task to be
2921 if (*imbalance
< sds
->busiest_load_per_task
)
2922 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
2926 /******* find_busiest_group() helpers end here *********************/
2929 * find_busiest_group - Returns the busiest group within the sched_domain
2930 * if there is an imbalance. If there isn't an imbalance, and
2931 * the user has opted for power-savings, it returns a group whose
2932 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
2933 * such a group exists.
2935 * Also calculates the amount of weighted load which should be moved
2936 * to restore balance.
2938 * @sd: The sched_domain whose busiest group is to be returned.
2939 * @this_cpu: The cpu for which load balancing is currently being performed.
2940 * @imbalance: Variable which stores amount of weighted load which should
2941 * be moved to restore balance/put a group to idle.
2942 * @idle: The idle status of this_cpu.
2943 * @sd_idle: The idleness of sd
2944 * @cpus: The set of CPUs under consideration for load-balancing.
2945 * @balance: Pointer to a variable indicating if this_cpu
2946 * is the appropriate cpu to perform load balancing at this_level.
2948 * Returns: - the busiest group if imbalance exists.
2949 * - If no imbalance and user has opted for power-savings balance,
2950 * return the least loaded group whose CPUs can be
2951 * put to idle by rebalancing its tasks onto our group.
2953 static struct sched_group
*
2954 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2955 unsigned long *imbalance
, enum cpu_idle_type idle
,
2956 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
2958 struct sd_lb_stats sds
;
2960 memset(&sds
, 0, sizeof(sds
));
2963 * Compute the various statistics relavent for load balancing at
2966 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
2969 /* Cases where imbalance does not exist from POV of this_cpu */
2970 /* 1) this_cpu is not the appropriate cpu to perform load balancing
2972 * 2) There is no busy sibling group to pull from.
2973 * 3) This group is the busiest group.
2974 * 4) This group is more busy than the avg busieness at this
2976 * 5) The imbalance is within the specified limit.
2978 * Note: when doing newidle balance, if the local group has excess
2979 * capacity (i.e. nr_running < group_capacity) and the busiest group
2980 * does not have any capacity, we force a load balance to pull tasks
2981 * to the local group. In this case, we skip past checks 3, 4 and 5.
2986 if ((idle
== CPU_IDLE
|| idle
== CPU_NEWLY_IDLE
) &&
2987 check_asym_packing(sd
, &sds
, this_cpu
, imbalance
))
2990 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
2993 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
2994 if (idle
== CPU_NEWLY_IDLE
&& sds
.this_has_capacity
&&
2995 !sds
.busiest_has_capacity
)
2998 if (sds
.this_load
>= sds
.max_load
)
3001 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
3003 if (sds
.this_load
>= sds
.avg_load
)
3006 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
3010 /* Looks like there is an imbalance. Compute it */
3011 calculate_imbalance(&sds
, this_cpu
, imbalance
);
3016 * There is no obvious imbalance. But check if we can do some balancing
3019 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
3027 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3030 find_busiest_queue(struct sched_domain
*sd
, struct sched_group
*group
,
3031 enum cpu_idle_type idle
, unsigned long imbalance
,
3032 const struct cpumask
*cpus
)
3034 struct rq
*busiest
= NULL
, *rq
;
3035 unsigned long max_load
= 0;
3038 for_each_cpu(i
, sched_group_cpus(group
)) {
3039 unsigned long power
= power_of(i
);
3040 unsigned long capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
3044 capacity
= fix_small_capacity(sd
, group
);
3046 if (!cpumask_test_cpu(i
, cpus
))
3050 wl
= weighted_cpuload(i
);
3053 * When comparing with imbalance, use weighted_cpuload()
3054 * which is not scaled with the cpu power.
3056 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
3060 * For the load comparisons with the other cpu's, consider
3061 * the weighted_cpuload() scaled with the cpu power, so that
3062 * the load can be moved away from the cpu that is potentially
3063 * running at a lower capacity.
3065 wl
= (wl
* SCHED_LOAD_SCALE
) / power
;
3067 if (wl
> max_load
) {
3077 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3078 * so long as it is large enough.
3080 #define MAX_PINNED_INTERVAL 512
3082 /* Working cpumask for load_balance and load_balance_newidle. */
3083 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
3085 static int need_active_balance(struct sched_domain
*sd
, int sd_idle
, int idle
,
3086 int busiest_cpu
, int this_cpu
)
3088 if (idle
== CPU_NEWLY_IDLE
) {
3091 * ASYM_PACKING needs to force migrate tasks from busy but
3092 * higher numbered CPUs in order to pack all tasks in the
3093 * lowest numbered CPUs.
3095 if ((sd
->flags
& SD_ASYM_PACKING
) && busiest_cpu
> this_cpu
)
3099 * The only task running in a non-idle cpu can be moved to this
3100 * cpu in an attempt to completely freeup the other CPU
3103 * The package power saving logic comes from
3104 * find_busiest_group(). If there are no imbalance, then
3105 * f_b_g() will return NULL. However when sched_mc={1,2} then
3106 * f_b_g() will select a group from which a running task may be
3107 * pulled to this cpu in order to make the other package idle.
3108 * If there is no opportunity to make a package idle and if
3109 * there are no imbalance, then f_b_g() will return NULL and no
3110 * action will be taken in load_balance_newidle().
3112 * Under normal task pull operation due to imbalance, there
3113 * will be more than one task in the source run queue and
3114 * move_tasks() will succeed. ld_moved will be true and this
3115 * active balance code will not be triggered.
3117 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3118 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3121 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
3125 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
3128 static int active_load_balance_cpu_stop(void *data
);
3131 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3132 * tasks if there is an imbalance.
3134 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3135 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3138 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3139 struct sched_group
*group
;
3140 unsigned long imbalance
;
3142 unsigned long flags
;
3143 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
3145 cpumask_copy(cpus
, cpu_active_mask
);
3148 * When power savings policy is enabled for the parent domain, idle
3149 * sibling can pick up load irrespective of busy siblings. In this case,
3150 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3151 * portraying it as CPU_NOT_IDLE.
3153 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3154 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3157 schedstat_inc(sd
, lb_count
[idle
]);
3160 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3167 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3171 busiest
= find_busiest_queue(sd
, group
, idle
, imbalance
, cpus
);
3173 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3177 BUG_ON(busiest
== this_rq
);
3179 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3182 if (busiest
->nr_running
> 1) {
3184 * Attempt to move tasks. If find_busiest_group has found
3185 * an imbalance but busiest->nr_running <= 1, the group is
3186 * still unbalanced. ld_moved simply stays zero, so it is
3187 * correctly treated as an imbalance.
3189 local_irq_save(flags
);
3190 double_rq_lock(this_rq
, busiest
);
3191 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3192 imbalance
, sd
, idle
, &all_pinned
);
3193 double_rq_unlock(this_rq
, busiest
);
3194 local_irq_restore(flags
);
3197 * some other cpu did the load balance for us.
3199 if (ld_moved
&& this_cpu
!= smp_processor_id())
3200 resched_cpu(this_cpu
);
3202 /* All tasks on this runqueue were pinned by CPU affinity */
3203 if (unlikely(all_pinned
)) {
3204 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3205 if (!cpumask_empty(cpus
))
3212 schedstat_inc(sd
, lb_failed
[idle
]);
3214 * Increment the failure counter only on periodic balance.
3215 * We do not want newidle balance, which can be very
3216 * frequent, pollute the failure counter causing
3217 * excessive cache_hot migrations and active balances.
3219 if (idle
!= CPU_NEWLY_IDLE
)
3220 sd
->nr_balance_failed
++;
3222 if (need_active_balance(sd
, sd_idle
, idle
, cpu_of(busiest
),
3224 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
3226 /* don't kick the active_load_balance_cpu_stop,
3227 * if the curr task on busiest cpu can't be
3230 if (!cpumask_test_cpu(this_cpu
,
3231 &busiest
->curr
->cpus_allowed
)) {
3232 raw_spin_unlock_irqrestore(&busiest
->lock
,
3235 goto out_one_pinned
;
3239 * ->active_balance synchronizes accesses to
3240 * ->active_balance_work. Once set, it's cleared
3241 * only after active load balance is finished.
3243 if (!busiest
->active_balance
) {
3244 busiest
->active_balance
= 1;
3245 busiest
->push_cpu
= this_cpu
;
3248 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
3251 stop_one_cpu_nowait(cpu_of(busiest
),
3252 active_load_balance_cpu_stop
, busiest
,
3253 &busiest
->active_balance_work
);
3256 * We've kicked active balancing, reset the failure
3259 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3262 sd
->nr_balance_failed
= 0;
3264 if (likely(!active_balance
)) {
3265 /* We were unbalanced, so reset the balancing interval */
3266 sd
->balance_interval
= sd
->min_interval
;
3269 * If we've begun active balancing, start to back off. This
3270 * case may not be covered by the all_pinned logic if there
3271 * is only 1 task on the busy runqueue (because we don't call
3274 if (sd
->balance_interval
< sd
->max_interval
)
3275 sd
->balance_interval
*= 2;
3278 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3279 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3285 schedstat_inc(sd
, lb_balanced
[idle
]);
3287 sd
->nr_balance_failed
= 0;
3290 /* tune up the balancing interval */
3291 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3292 (sd
->balance_interval
< sd
->max_interval
))
3293 sd
->balance_interval
*= 2;
3295 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3296 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3305 * idle_balance is called by schedule() if this_cpu is about to become
3306 * idle. Attempts to pull tasks from other CPUs.
3308 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3310 struct sched_domain
*sd
;
3311 int pulled_task
= 0;
3312 unsigned long next_balance
= jiffies
+ HZ
;
3314 this_rq
->idle_stamp
= this_rq
->clock
;
3316 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
3320 * Drop the rq->lock, but keep IRQ/preempt disabled.
3322 raw_spin_unlock(&this_rq
->lock
);
3324 for_each_domain(this_cpu
, sd
) {
3325 unsigned long interval
;
3328 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3331 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
3332 /* If we've pulled tasks over stop searching: */
3333 pulled_task
= load_balance(this_cpu
, this_rq
,
3334 sd
, CPU_NEWLY_IDLE
, &balance
);
3337 interval
= msecs_to_jiffies(sd
->balance_interval
);
3338 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3339 next_balance
= sd
->last_balance
+ interval
;
3344 raw_spin_lock(&this_rq
->lock
);
3346 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3348 * We are going idle. next_balance may be set based on
3349 * a busy processor. So reset next_balance.
3351 this_rq
->next_balance
= next_balance
;
3356 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3357 * running tasks off the busiest CPU onto idle CPUs. It requires at
3358 * least 1 task to be running on each physical CPU where possible, and
3359 * avoids physical / logical imbalances.
3361 static int active_load_balance_cpu_stop(void *data
)
3363 struct rq
*busiest_rq
= data
;
3364 int busiest_cpu
= cpu_of(busiest_rq
);
3365 int target_cpu
= busiest_rq
->push_cpu
;
3366 struct rq
*target_rq
= cpu_rq(target_cpu
);
3367 struct sched_domain
*sd
;
3369 raw_spin_lock_irq(&busiest_rq
->lock
);
3371 /* make sure the requested cpu hasn't gone down in the meantime */
3372 if (unlikely(busiest_cpu
!= smp_processor_id() ||
3373 !busiest_rq
->active_balance
))
3376 /* Is there any task to move? */
3377 if (busiest_rq
->nr_running
<= 1)
3381 * This condition is "impossible", if it occurs
3382 * we need to fix it. Originally reported by
3383 * Bjorn Helgaas on a 128-cpu setup.
3385 BUG_ON(busiest_rq
== target_rq
);
3387 /* move a task from busiest_rq to target_rq */
3388 double_lock_balance(busiest_rq
, target_rq
);
3390 /* Search for an sd spanning us and the target CPU. */
3391 for_each_domain(target_cpu
, sd
) {
3392 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3393 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
3398 schedstat_inc(sd
, alb_count
);
3400 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3402 schedstat_inc(sd
, alb_pushed
);
3404 schedstat_inc(sd
, alb_failed
);
3406 double_unlock_balance(busiest_rq
, target_rq
);
3408 busiest_rq
->active_balance
= 0;
3409 raw_spin_unlock_irq(&busiest_rq
->lock
);
3415 static DEFINE_PER_CPU(struct call_single_data
, remote_sched_softirq_cb
);
3417 static void trigger_sched_softirq(void *data
)
3419 raise_softirq_irqoff(SCHED_SOFTIRQ
);
3422 static inline void init_sched_softirq_csd(struct call_single_data
*csd
)
3424 csd
->func
= trigger_sched_softirq
;
3431 * idle load balancing details
3432 * - One of the idle CPUs nominates itself as idle load_balancer, while
3434 * - This idle load balancer CPU will also go into tickless mode when
3435 * it is idle, just like all other idle CPUs
3436 * - When one of the busy CPUs notice that there may be an idle rebalancing
3437 * needed, they will kick the idle load balancer, which then does idle
3438 * load balancing for all the idle CPUs.
3441 atomic_t load_balancer
;
3442 atomic_t first_pick_cpu
;
3443 atomic_t second_pick_cpu
;
3444 cpumask_var_t idle_cpus_mask
;
3445 cpumask_var_t grp_idle_mask
;
3446 unsigned long next_balance
; /* in jiffy units */
3447 } nohz ____cacheline_aligned
;
3449 int get_nohz_load_balancer(void)
3451 return atomic_read(&nohz
.load_balancer
);
3454 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3456 * lowest_flag_domain - Return lowest sched_domain containing flag.
3457 * @cpu: The cpu whose lowest level of sched domain is to
3459 * @flag: The flag to check for the lowest sched_domain
3460 * for the given cpu.
3462 * Returns the lowest sched_domain of a cpu which contains the given flag.
3464 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
3466 struct sched_domain
*sd
;
3468 for_each_domain(cpu
, sd
)
3469 if (sd
&& (sd
->flags
& flag
))
3476 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3477 * @cpu: The cpu whose domains we're iterating over.
3478 * @sd: variable holding the value of the power_savings_sd
3480 * @flag: The flag to filter the sched_domains to be iterated.
3482 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3483 * set, starting from the lowest sched_domain to the highest.
3485 #define for_each_flag_domain(cpu, sd, flag) \
3486 for (sd = lowest_flag_domain(cpu, flag); \
3487 (sd && (sd->flags & flag)); sd = sd->parent)
3490 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3491 * @ilb_group: group to be checked for semi-idleness
3493 * Returns: 1 if the group is semi-idle. 0 otherwise.
3495 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3496 * and atleast one non-idle CPU. This helper function checks if the given
3497 * sched_group is semi-idle or not.
3499 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
3501 cpumask_and(nohz
.grp_idle_mask
, nohz
.idle_cpus_mask
,
3502 sched_group_cpus(ilb_group
));
3505 * A sched_group is semi-idle when it has atleast one busy cpu
3506 * and atleast one idle cpu.
3508 if (cpumask_empty(nohz
.grp_idle_mask
))
3511 if (cpumask_equal(nohz
.grp_idle_mask
, sched_group_cpus(ilb_group
)))
3517 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3518 * @cpu: The cpu which is nominating a new idle_load_balancer.
3520 * Returns: Returns the id of the idle load balancer if it exists,
3521 * Else, returns >= nr_cpu_ids.
3523 * This algorithm picks the idle load balancer such that it belongs to a
3524 * semi-idle powersavings sched_domain. The idea is to try and avoid
3525 * completely idle packages/cores just for the purpose of idle load balancing
3526 * when there are other idle cpu's which are better suited for that job.
3528 static int find_new_ilb(int cpu
)
3530 struct sched_domain
*sd
;
3531 struct sched_group
*ilb_group
;
3534 * Have idle load balancer selection from semi-idle packages only
3535 * when power-aware load balancing is enabled
3537 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
3541 * Optimize for the case when we have no idle CPUs or only one
3542 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3544 if (cpumask_weight(nohz
.idle_cpus_mask
) < 2)
3547 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
3548 ilb_group
= sd
->groups
;
3551 if (is_semi_idle_group(ilb_group
))
3552 return cpumask_first(nohz
.grp_idle_mask
);
3554 ilb_group
= ilb_group
->next
;
3556 } while (ilb_group
!= sd
->groups
);
3562 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3563 static inline int find_new_ilb(int call_cpu
)
3570 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3571 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3572 * CPU (if there is one).
3574 static void nohz_balancer_kick(int cpu
)
3578 nohz
.next_balance
++;
3580 ilb_cpu
= get_nohz_load_balancer();
3582 if (ilb_cpu
>= nr_cpu_ids
) {
3583 ilb_cpu
= cpumask_first(nohz
.idle_cpus_mask
);
3584 if (ilb_cpu
>= nr_cpu_ids
)
3588 if (!cpu_rq(ilb_cpu
)->nohz_balance_kick
) {
3589 struct call_single_data
*cp
;
3591 cpu_rq(ilb_cpu
)->nohz_balance_kick
= 1;
3592 cp
= &per_cpu(remote_sched_softirq_cb
, cpu
);
3593 __smp_call_function_single(ilb_cpu
, cp
, 0);
3599 * This routine will try to nominate the ilb (idle load balancing)
3600 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3601 * load balancing on behalf of all those cpus.
3603 * When the ilb owner becomes busy, we will not have new ilb owner until some
3604 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3605 * idle load balancing by kicking one of the idle CPUs.
3607 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3608 * ilb owner CPU in future (when there is a need for idle load balancing on
3609 * behalf of all idle CPUs).
3611 void select_nohz_load_balancer(int stop_tick
)
3613 int cpu
= smp_processor_id();
3616 if (!cpu_active(cpu
)) {
3617 if (atomic_read(&nohz
.load_balancer
) != cpu
)
3621 * If we are going offline and still the leader,
3624 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
,
3631 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
3633 if (atomic_read(&nohz
.first_pick_cpu
) == cpu
)
3634 atomic_cmpxchg(&nohz
.first_pick_cpu
, cpu
, nr_cpu_ids
);
3635 if (atomic_read(&nohz
.second_pick_cpu
) == cpu
)
3636 atomic_cmpxchg(&nohz
.second_pick_cpu
, cpu
, nr_cpu_ids
);
3638 if (atomic_read(&nohz
.load_balancer
) >= nr_cpu_ids
) {
3641 /* make me the ilb owner */
3642 if (atomic_cmpxchg(&nohz
.load_balancer
, nr_cpu_ids
,
3647 * Check to see if there is a more power-efficient
3650 new_ilb
= find_new_ilb(cpu
);
3651 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
3652 atomic_set(&nohz
.load_balancer
, nr_cpu_ids
);
3653 resched_cpu(new_ilb
);
3659 if (!cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))
3662 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
3664 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3665 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
,
3673 static DEFINE_SPINLOCK(balancing
);
3676 * It checks each scheduling domain to see if it is due to be balanced,
3677 * and initiates a balancing operation if so.
3679 * Balancing parameters are set up in arch_init_sched_domains.
3681 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3684 struct rq
*rq
= cpu_rq(cpu
);
3685 unsigned long interval
;
3686 struct sched_domain
*sd
;
3687 /* Earliest time when we have to do rebalance again */
3688 unsigned long next_balance
= jiffies
+ 60*HZ
;
3689 int update_next_balance
= 0;
3694 for_each_domain(cpu
, sd
) {
3695 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3698 interval
= sd
->balance_interval
;
3699 if (idle
!= CPU_IDLE
)
3700 interval
*= sd
->busy_factor
;
3702 /* scale ms to jiffies */
3703 interval
= msecs_to_jiffies(interval
);
3704 if (unlikely(!interval
))
3706 if (interval
> HZ
*NR_CPUS
/10)
3707 interval
= HZ
*NR_CPUS
/10;
3709 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3711 if (need_serialize
) {
3712 if (!spin_trylock(&balancing
))
3716 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3717 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3719 * We've pulled tasks over so either we're no
3720 * longer idle, or one of our SMT siblings is
3723 idle
= CPU_NOT_IDLE
;
3725 sd
->last_balance
= jiffies
;
3728 spin_unlock(&balancing
);
3730 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3731 next_balance
= sd
->last_balance
+ interval
;
3732 update_next_balance
= 1;
3736 * Stop the load balance at this level. There is another
3737 * CPU in our sched group which is doing load balancing more
3745 * next_balance will be updated only when there is a need.
3746 * When the cpu is attached to null domain for ex, it will not be
3749 if (likely(update_next_balance
))
3750 rq
->next_balance
= next_balance
;
3755 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3756 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3758 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
)
3760 struct rq
*this_rq
= cpu_rq(this_cpu
);
3764 if (idle
!= CPU_IDLE
|| !this_rq
->nohz_balance_kick
)
3767 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
3768 if (balance_cpu
== this_cpu
)
3772 * If this cpu gets work to do, stop the load balancing
3773 * work being done for other cpus. Next load
3774 * balancing owner will pick it up.
3776 if (need_resched()) {
3777 this_rq
->nohz_balance_kick
= 0;
3781 raw_spin_lock_irq(&this_rq
->lock
);
3782 update_rq_clock(this_rq
);
3783 update_cpu_load(this_rq
);
3784 raw_spin_unlock_irq(&this_rq
->lock
);
3786 rebalance_domains(balance_cpu
, CPU_IDLE
);
3788 rq
= cpu_rq(balance_cpu
);
3789 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3790 this_rq
->next_balance
= rq
->next_balance
;
3792 nohz
.next_balance
= this_rq
->next_balance
;
3793 this_rq
->nohz_balance_kick
= 0;
3797 * Current heuristic for kicking the idle load balancer
3798 * - first_pick_cpu is the one of the busy CPUs. It will kick
3799 * idle load balancer when it has more than one process active. This
3800 * eliminates the need for idle load balancing altogether when we have
3801 * only one running process in the system (common case).
3802 * - If there are more than one busy CPU, idle load balancer may have
3803 * to run for active_load_balance to happen (i.e., two busy CPUs are
3804 * SMT or core siblings and can run better if they move to different
3805 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
3806 * which will kick idle load balancer as soon as it has any load.
3808 static inline int nohz_kick_needed(struct rq
*rq
, int cpu
)
3810 unsigned long now
= jiffies
;
3812 int first_pick_cpu
, second_pick_cpu
;
3814 if (time_before(now
, nohz
.next_balance
))
3817 if (rq
->idle_at_tick
)
3820 first_pick_cpu
= atomic_read(&nohz
.first_pick_cpu
);
3821 second_pick_cpu
= atomic_read(&nohz
.second_pick_cpu
);
3823 if (first_pick_cpu
< nr_cpu_ids
&& first_pick_cpu
!= cpu
&&
3824 second_pick_cpu
< nr_cpu_ids
&& second_pick_cpu
!= cpu
)
3827 ret
= atomic_cmpxchg(&nohz
.first_pick_cpu
, nr_cpu_ids
, cpu
);
3828 if (ret
== nr_cpu_ids
|| ret
== cpu
) {
3829 atomic_cmpxchg(&nohz
.second_pick_cpu
, cpu
, nr_cpu_ids
);
3830 if (rq
->nr_running
> 1)
3833 ret
= atomic_cmpxchg(&nohz
.second_pick_cpu
, nr_cpu_ids
, cpu
);
3834 if (ret
== nr_cpu_ids
|| ret
== cpu
) {
3842 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
) { }
3846 * run_rebalance_domains is triggered when needed from the scheduler tick.
3847 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
3849 static void run_rebalance_domains(struct softirq_action
*h
)
3851 int this_cpu
= smp_processor_id();
3852 struct rq
*this_rq
= cpu_rq(this_cpu
);
3853 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3854 CPU_IDLE
: CPU_NOT_IDLE
;
3856 rebalance_domains(this_cpu
, idle
);
3859 * If this cpu has a pending nohz_balance_kick, then do the
3860 * balancing on behalf of the other idle cpus whose ticks are
3863 nohz_idle_balance(this_cpu
, idle
);
3866 static inline int on_null_domain(int cpu
)
3868 return !rcu_dereference_sched(cpu_rq(cpu
)->sd
);
3872 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3874 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3876 /* Don't need to rebalance while attached to NULL domain */
3877 if (time_after_eq(jiffies
, rq
->next_balance
) &&
3878 likely(!on_null_domain(cpu
)))
3879 raise_softirq(SCHED_SOFTIRQ
);
3881 else if (nohz_kick_needed(rq
, cpu
) && likely(!on_null_domain(cpu
)))
3882 nohz_balancer_kick(cpu
);
3886 static void rq_online_fair(struct rq
*rq
)
3891 static void rq_offline_fair(struct rq
*rq
)
3896 #else /* CONFIG_SMP */
3899 * on UP we do not need to balance between CPUs:
3901 static inline void idle_balance(int cpu
, struct rq
*rq
)
3905 #endif /* CONFIG_SMP */
3908 * scheduler tick hitting a task of our scheduling class:
3910 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
3912 struct cfs_rq
*cfs_rq
;
3913 struct sched_entity
*se
= &curr
->se
;
3915 for_each_sched_entity(se
) {
3916 cfs_rq
= cfs_rq_of(se
);
3917 entity_tick(cfs_rq
, se
, queued
);
3922 * called on fork with the child task as argument from the parent's context
3923 * - child not yet on the tasklist
3924 * - preemption disabled
3926 static void task_fork_fair(struct task_struct
*p
)
3928 struct cfs_rq
*cfs_rq
= task_cfs_rq(current
);
3929 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
3930 int this_cpu
= smp_processor_id();
3931 struct rq
*rq
= this_rq();
3932 unsigned long flags
;
3934 raw_spin_lock_irqsave(&rq
->lock
, flags
);
3936 update_rq_clock(rq
);
3938 if (unlikely(task_cpu(p
) != this_cpu
)) {
3940 __set_task_cpu(p
, this_cpu
);
3944 update_curr(cfs_rq
);
3947 se
->vruntime
= curr
->vruntime
;
3948 place_entity(cfs_rq
, se
, 1);
3950 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
3952 * Upon rescheduling, sched_class::put_prev_task() will place
3953 * 'current' within the tree based on its new key value.
3955 swap(curr
->vruntime
, se
->vruntime
);
3956 resched_task(rq
->curr
);
3959 se
->vruntime
-= cfs_rq
->min_vruntime
;
3961 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
3965 * Priority of the task has changed. Check to see if we preempt
3968 static void prio_changed_fair(struct rq
*rq
, struct task_struct
*p
,
3969 int oldprio
, int running
)
3972 * Reschedule if we are currently running on this runqueue and
3973 * our priority decreased, or if we are not currently running on
3974 * this runqueue and our priority is higher than the current's
3977 if (p
->prio
> oldprio
)
3978 resched_task(rq
->curr
);
3980 check_preempt_curr(rq
, p
, 0);
3984 * We switched to the sched_fair class.
3986 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
,
3990 * We were most likely switched from sched_rt, so
3991 * kick off the schedule if running, otherwise just see
3992 * if we can still preempt the current task.
3995 resched_task(rq
->curr
);
3997 check_preempt_curr(rq
, p
, 0);
4000 /* Account for a task changing its policy or group.
4002 * This routine is mostly called to set cfs_rq->curr field when a task
4003 * migrates between groups/classes.
4005 static void set_curr_task_fair(struct rq
*rq
)
4007 struct sched_entity
*se
= &rq
->curr
->se
;
4009 for_each_sched_entity(se
)
4010 set_next_entity(cfs_rq_of(se
), se
);
4013 #ifdef CONFIG_FAIR_GROUP_SCHED
4014 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
4017 * If the task was not on the rq at the time of this cgroup movement
4018 * it must have been asleep, sleeping tasks keep their ->vruntime
4019 * absolute on their old rq until wakeup (needed for the fair sleeper
4020 * bonus in place_entity()).
4022 * If it was on the rq, we've just 'preempted' it, which does convert
4023 * ->vruntime to a relative base.
4025 * Make sure both cases convert their relative position when migrating
4026 * to another cgroup's rq. This does somewhat interfere with the
4027 * fair sleeper stuff for the first placement, but who cares.
4030 p
->se
.vruntime
-= cfs_rq_of(&p
->se
)->min_vruntime
;
4031 set_task_rq(p
, task_cpu(p
));
4033 p
->se
.vruntime
+= cfs_rq_of(&p
->se
)->min_vruntime
;
4037 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
4039 struct sched_entity
*se
= &task
->se
;
4040 unsigned int rr_interval
= 0;
4043 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4046 if (rq
->cfs
.load
.weight
)
4047 rr_interval
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
4053 * All the scheduling class methods:
4055 static const struct sched_class fair_sched_class
= {
4056 .next
= &idle_sched_class
,
4057 .enqueue_task
= enqueue_task_fair
,
4058 .dequeue_task
= dequeue_task_fair
,
4059 .yield_task
= yield_task_fair
,
4061 .check_preempt_curr
= check_preempt_wakeup
,
4063 .pick_next_task
= pick_next_task_fair
,
4064 .put_prev_task
= put_prev_task_fair
,
4067 .select_task_rq
= select_task_rq_fair
,
4069 .rq_online
= rq_online_fair
,
4070 .rq_offline
= rq_offline_fair
,
4072 .task_waking
= task_waking_fair
,
4075 .set_curr_task
= set_curr_task_fair
,
4076 .task_tick
= task_tick_fair
,
4077 .task_fork
= task_fork_fair
,
4079 .prio_changed
= prio_changed_fair
,
4080 .switched_to
= switched_to_fair
,
4082 .get_rr_interval
= get_rr_interval_fair
,
4084 #ifdef CONFIG_FAIR_GROUP_SCHED
4085 .task_move_group
= task_move_group_fair
,
4089 #ifdef CONFIG_SCHED_DEBUG
4090 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
4092 struct cfs_rq
*cfs_rq
;
4095 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
4096 print_cfs_rq(m
, cpu
, cfs_rq
);