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: 5ms * (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: 2 msec * (1 + ilog(ncpus)), units: nanoseconds)
57 unsigned int sysctl_sched_min_granularity
= 2000000ULL;
58 unsigned int normalized_sysctl_sched_min_granularity
= 2000000ULL;
61 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
63 static unsigned int sched_nr_latency
= 3;
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 /* Iterate thr' all leaf cfs_rq's on a runqueue */
147 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
148 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
150 /* Do the two (enqueued) entities belong to the same group ? */
152 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
154 if (se
->cfs_rq
== pse
->cfs_rq
)
160 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
165 /* return depth at which a sched entity is present in the hierarchy */
166 static inline int depth_se(struct sched_entity
*se
)
170 for_each_sched_entity(se
)
177 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
179 int se_depth
, pse_depth
;
182 * preemption test can be made between sibling entities who are in the
183 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
184 * both tasks until we find their ancestors who are siblings of common
188 /* First walk up until both entities are at same depth */
189 se_depth
= depth_se(*se
);
190 pse_depth
= depth_se(*pse
);
192 while (se_depth
> pse_depth
) {
194 *se
= parent_entity(*se
);
197 while (pse_depth
> se_depth
) {
199 *pse
= parent_entity(*pse
);
202 while (!is_same_group(*se
, *pse
)) {
203 *se
= parent_entity(*se
);
204 *pse
= parent_entity(*pse
);
208 #else /* !CONFIG_FAIR_GROUP_SCHED */
210 static inline struct task_struct
*task_of(struct sched_entity
*se
)
212 return container_of(se
, struct task_struct
, se
);
215 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
217 return container_of(cfs_rq
, struct rq
, cfs
);
220 #define entity_is_task(se) 1
222 #define for_each_sched_entity(se) \
223 for (; se; se = NULL)
225 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
227 return &task_rq(p
)->cfs
;
230 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
232 struct task_struct
*p
= task_of(se
);
233 struct rq
*rq
= task_rq(p
);
238 /* runqueue "owned" by this group */
239 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
244 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
246 return &cpu_rq(this_cpu
)->cfs
;
249 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
250 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
253 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
258 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
264 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
268 #endif /* CONFIG_FAIR_GROUP_SCHED */
271 /**************************************************************
272 * Scheduling class tree data structure manipulation methods:
275 static inline u64
max_vruntime(u64 min_vruntime
, u64 vruntime
)
277 s64 delta
= (s64
)(vruntime
- min_vruntime
);
279 min_vruntime
= vruntime
;
284 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
286 s64 delta
= (s64
)(vruntime
- min_vruntime
);
288 min_vruntime
= vruntime
;
293 static inline int entity_before(struct sched_entity
*a
,
294 struct sched_entity
*b
)
296 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
299 static inline s64
entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
301 return se
->vruntime
- cfs_rq
->min_vruntime
;
304 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
306 u64 vruntime
= cfs_rq
->min_vruntime
;
309 vruntime
= cfs_rq
->curr
->vruntime
;
311 if (cfs_rq
->rb_leftmost
) {
312 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
317 vruntime
= se
->vruntime
;
319 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
322 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
326 * Enqueue an entity into the rb-tree:
328 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
330 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
331 struct rb_node
*parent
= NULL
;
332 struct sched_entity
*entry
;
333 s64 key
= entity_key(cfs_rq
, se
);
337 * Find the right place in the rbtree:
341 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
343 * We dont care about collisions. Nodes with
344 * the same key stay together.
346 if (key
< entity_key(cfs_rq
, entry
)) {
347 link
= &parent
->rb_left
;
349 link
= &parent
->rb_right
;
355 * Maintain a cache of leftmost tree entries (it is frequently
359 cfs_rq
->rb_leftmost
= &se
->run_node
;
361 rb_link_node(&se
->run_node
, parent
, link
);
362 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
365 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
367 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
368 struct rb_node
*next_node
;
370 next_node
= rb_next(&se
->run_node
);
371 cfs_rq
->rb_leftmost
= next_node
;
374 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
377 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
379 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
384 return rb_entry(left
, struct sched_entity
, run_node
);
387 static struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
389 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
394 return rb_entry(last
, struct sched_entity
, run_node
);
397 /**************************************************************
398 * Scheduling class statistics methods:
401 #ifdef CONFIG_SCHED_DEBUG
402 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
403 void __user
*buffer
, size_t *lenp
,
406 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
407 int factor
= get_update_sysctl_factor();
412 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
413 sysctl_sched_min_granularity
);
415 #define WRT_SYSCTL(name) \
416 (normalized_sysctl_##name = sysctl_##name / (factor))
417 WRT_SYSCTL(sched_min_granularity
);
418 WRT_SYSCTL(sched_latency
);
419 WRT_SYSCTL(sched_wakeup_granularity
);
420 WRT_SYSCTL(sched_shares_ratelimit
);
430 static inline unsigned long
431 calc_delta_fair(unsigned long delta
, struct sched_entity
*se
)
433 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
434 delta
= calc_delta_mine(delta
, NICE_0_LOAD
, &se
->load
);
440 * The idea is to set a period in which each task runs once.
442 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
443 * this period because otherwise the slices get too small.
445 * p = (nr <= nl) ? l : l*nr/nl
447 static u64
__sched_period(unsigned long nr_running
)
449 u64 period
= sysctl_sched_latency
;
450 unsigned long nr_latency
= sched_nr_latency
;
452 if (unlikely(nr_running
> nr_latency
)) {
453 period
= sysctl_sched_min_granularity
;
454 period
*= nr_running
;
461 * We calculate the wall-time slice from the period by taking a part
462 * proportional to the weight.
466 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
468 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
470 for_each_sched_entity(se
) {
471 struct load_weight
*load
;
472 struct load_weight lw
;
474 cfs_rq
= cfs_rq_of(se
);
475 load
= &cfs_rq
->load
;
477 if (unlikely(!se
->on_rq
)) {
480 update_load_add(&lw
, se
->load
.weight
);
483 slice
= calc_delta_mine(slice
, se
->load
.weight
, load
);
489 * We calculate the vruntime slice of a to be inserted task
493 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
495 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
499 * Update the current task's runtime statistics. Skip current tasks that
500 * are not in our scheduling class.
503 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
504 unsigned long delta_exec
)
506 unsigned long delta_exec_weighted
;
508 schedstat_set(curr
->statistics
.exec_max
,
509 max((u64
)delta_exec
, curr
->statistics
.exec_max
));
511 curr
->sum_exec_runtime
+= delta_exec
;
512 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
513 delta_exec_weighted
= calc_delta_fair(delta_exec
, curr
);
515 curr
->vruntime
+= delta_exec_weighted
;
516 update_min_vruntime(cfs_rq
);
519 static void update_curr(struct cfs_rq
*cfs_rq
)
521 struct sched_entity
*curr
= cfs_rq
->curr
;
522 u64 now
= rq_of(cfs_rq
)->clock
;
523 unsigned long delta_exec
;
529 * Get the amount of time the current task was running
530 * since the last time we changed load (this cannot
531 * overflow on 32 bits):
533 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
537 __update_curr(cfs_rq
, curr
, delta_exec
);
538 curr
->exec_start
= now
;
540 if (entity_is_task(curr
)) {
541 struct task_struct
*curtask
= task_of(curr
);
543 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
544 cpuacct_charge(curtask
, delta_exec
);
545 account_group_exec_runtime(curtask
, delta_exec
);
550 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
552 schedstat_set(se
->statistics
.wait_start
, rq_of(cfs_rq
)->clock
);
556 * Task is being enqueued - update stats:
558 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
561 * Are we enqueueing a waiting task? (for current tasks
562 * a dequeue/enqueue event is a NOP)
564 if (se
!= cfs_rq
->curr
)
565 update_stats_wait_start(cfs_rq
, se
);
569 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
571 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
572 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
));
573 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
574 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
575 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
);
576 #ifdef CONFIG_SCHEDSTATS
577 if (entity_is_task(se
)) {
578 trace_sched_stat_wait(task_of(se
),
579 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
);
582 schedstat_set(se
->statistics
.wait_start
, 0);
586 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
589 * Mark the end of the wait period if dequeueing a
592 if (se
!= cfs_rq
->curr
)
593 update_stats_wait_end(cfs_rq
, se
);
597 * We are picking a new current task - update its stats:
600 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
603 * We are starting a new run period:
605 se
->exec_start
= rq_of(cfs_rq
)->clock
;
608 /**************************************************
609 * Scheduling class queueing methods:
612 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
614 add_cfs_task_weight(struct cfs_rq
*cfs_rq
, unsigned long weight
)
616 cfs_rq
->task_weight
+= weight
;
620 add_cfs_task_weight(struct cfs_rq
*cfs_rq
, unsigned long weight
)
626 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
628 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
629 if (!parent_entity(se
))
630 inc_cpu_load(rq_of(cfs_rq
), se
->load
.weight
);
631 if (entity_is_task(se
)) {
632 add_cfs_task_weight(cfs_rq
, se
->load
.weight
);
633 list_add(&se
->group_node
, &cfs_rq
->tasks
);
635 cfs_rq
->nr_running
++;
640 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
642 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
643 if (!parent_entity(se
))
644 dec_cpu_load(rq_of(cfs_rq
), se
->load
.weight
);
645 if (entity_is_task(se
)) {
646 add_cfs_task_weight(cfs_rq
, -se
->load
.weight
);
647 list_del_init(&se
->group_node
);
649 cfs_rq
->nr_running
--;
653 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
655 #ifdef CONFIG_SCHEDSTATS
656 struct task_struct
*tsk
= NULL
;
658 if (entity_is_task(se
))
661 if (se
->statistics
.sleep_start
) {
662 u64 delta
= rq_of(cfs_rq
)->clock
- se
->statistics
.sleep_start
;
667 if (unlikely(delta
> se
->statistics
.sleep_max
))
668 se
->statistics
.sleep_max
= delta
;
670 se
->statistics
.sleep_start
= 0;
671 se
->statistics
.sum_sleep_runtime
+= delta
;
674 account_scheduler_latency(tsk
, delta
>> 10, 1);
675 trace_sched_stat_sleep(tsk
, delta
);
678 if (se
->statistics
.block_start
) {
679 u64 delta
= rq_of(cfs_rq
)->clock
- se
->statistics
.block_start
;
684 if (unlikely(delta
> se
->statistics
.block_max
))
685 se
->statistics
.block_max
= delta
;
687 se
->statistics
.block_start
= 0;
688 se
->statistics
.sum_sleep_runtime
+= delta
;
691 if (tsk
->in_iowait
) {
692 se
->statistics
.iowait_sum
+= delta
;
693 se
->statistics
.iowait_count
++;
694 trace_sched_stat_iowait(tsk
, delta
);
698 * Blocking time is in units of nanosecs, so shift by
699 * 20 to get a milliseconds-range estimation of the
700 * amount of time that the task spent sleeping:
702 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
703 profile_hits(SLEEP_PROFILING
,
704 (void *)get_wchan(tsk
),
707 account_scheduler_latency(tsk
, delta
>> 10, 0);
713 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
715 #ifdef CONFIG_SCHED_DEBUG
716 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
721 if (d
> 3*sysctl_sched_latency
)
722 schedstat_inc(cfs_rq
, nr_spread_over
);
727 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
729 u64 vruntime
= cfs_rq
->min_vruntime
;
732 * The 'current' period is already promised to the current tasks,
733 * however the extra weight of the new task will slow them down a
734 * little, place the new task so that it fits in the slot that
735 * stays open at the end.
737 if (initial
&& sched_feat(START_DEBIT
))
738 vruntime
+= sched_vslice(cfs_rq
, se
);
740 /* sleeps up to a single latency don't count. */
742 unsigned long thresh
= sysctl_sched_latency
;
745 * Halve their sleep time's effect, to allow
746 * for a gentler effect of sleepers:
748 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
754 /* ensure we never gain time by being placed backwards. */
755 vruntime
= max_vruntime(se
->vruntime
, vruntime
);
757 se
->vruntime
= vruntime
;
760 #define ENQUEUE_WAKEUP 1
761 #define ENQUEUE_MIGRATE 2
764 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
767 * Update the normalized vruntime before updating min_vruntime
768 * through callig update_curr().
770 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_MIGRATE
))
771 se
->vruntime
+= cfs_rq
->min_vruntime
;
774 * Update run-time statistics of the 'current'.
777 account_entity_enqueue(cfs_rq
, se
);
779 if (flags
& ENQUEUE_WAKEUP
) {
780 place_entity(cfs_rq
, se
, 0);
781 enqueue_sleeper(cfs_rq
, se
);
784 update_stats_enqueue(cfs_rq
, se
);
785 check_spread(cfs_rq
, se
);
786 if (se
!= cfs_rq
->curr
)
787 __enqueue_entity(cfs_rq
, se
);
790 static void __clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
792 if (!se
|| cfs_rq
->last
== se
)
795 if (!se
|| cfs_rq
->next
== se
)
799 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
801 for_each_sched_entity(se
)
802 __clear_buddies(cfs_rq_of(se
), se
);
806 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
809 * Update run-time statistics of the 'current'.
813 update_stats_dequeue(cfs_rq
, se
);
815 #ifdef CONFIG_SCHEDSTATS
816 if (entity_is_task(se
)) {
817 struct task_struct
*tsk
= task_of(se
);
819 if (tsk
->state
& TASK_INTERRUPTIBLE
)
820 se
->statistics
.sleep_start
= rq_of(cfs_rq
)->clock
;
821 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
822 se
->statistics
.block_start
= rq_of(cfs_rq
)->clock
;
827 clear_buddies(cfs_rq
, se
);
829 if (se
!= cfs_rq
->curr
)
830 __dequeue_entity(cfs_rq
, se
);
831 account_entity_dequeue(cfs_rq
, se
);
832 update_min_vruntime(cfs_rq
);
835 * Normalize the entity after updating the min_vruntime because the
836 * update can refer to the ->curr item and we need to reflect this
837 * movement in our normalized position.
840 se
->vruntime
-= cfs_rq
->min_vruntime
;
844 * Preempt the current task with a newly woken task if needed:
847 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
849 unsigned long ideal_runtime
, delta_exec
;
851 ideal_runtime
= sched_slice(cfs_rq
, curr
);
852 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
853 if (delta_exec
> ideal_runtime
) {
854 resched_task(rq_of(cfs_rq
)->curr
);
856 * The current task ran long enough, ensure it doesn't get
857 * re-elected due to buddy favours.
859 clear_buddies(cfs_rq
, curr
);
864 * Ensure that a task that missed wakeup preemption by a
865 * narrow margin doesn't have to wait for a full slice.
866 * This also mitigates buddy induced latencies under load.
868 if (!sched_feat(WAKEUP_PREEMPT
))
871 if (delta_exec
< sysctl_sched_min_granularity
)
874 if (cfs_rq
->nr_running
> 1) {
875 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
876 s64 delta
= curr
->vruntime
- se
->vruntime
;
878 if (delta
> ideal_runtime
)
879 resched_task(rq_of(cfs_rq
)->curr
);
884 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
886 /* 'current' is not kept within the tree. */
889 * Any task has to be enqueued before it get to execute on
890 * a CPU. So account for the time it spent waiting on the
893 update_stats_wait_end(cfs_rq
, se
);
894 __dequeue_entity(cfs_rq
, se
);
897 update_stats_curr_start(cfs_rq
, se
);
899 #ifdef CONFIG_SCHEDSTATS
901 * Track our maximum slice length, if the CPU's load is at
902 * least twice that of our own weight (i.e. dont track it
903 * when there are only lesser-weight tasks around):
905 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
906 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
907 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
910 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
914 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
916 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
918 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
919 struct sched_entity
*left
= se
;
921 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
925 * Prefer last buddy, try to return the CPU to a preempted task.
927 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
930 clear_buddies(cfs_rq
, se
);
935 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
938 * If still on the runqueue then deactivate_task()
939 * was not called and update_curr() has to be done:
944 check_spread(cfs_rq
, prev
);
946 update_stats_wait_start(cfs_rq
, prev
);
947 /* Put 'current' back into the tree. */
948 __enqueue_entity(cfs_rq
, prev
);
954 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
957 * Update run-time statistics of the 'current'.
961 #ifdef CONFIG_SCHED_HRTICK
963 * queued ticks are scheduled to match the slice, so don't bother
964 * validating it and just reschedule.
967 resched_task(rq_of(cfs_rq
)->curr
);
971 * don't let the period tick interfere with the hrtick preemption
973 if (!sched_feat(DOUBLE_TICK
) &&
974 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
978 if (cfs_rq
->nr_running
> 1 || !sched_feat(WAKEUP_PREEMPT
))
979 check_preempt_tick(cfs_rq
, curr
);
982 /**************************************************
983 * CFS operations on tasks:
986 #ifdef CONFIG_SCHED_HRTICK
987 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
989 struct sched_entity
*se
= &p
->se
;
990 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
992 WARN_ON(task_rq(p
) != rq
);
994 if (hrtick_enabled(rq
) && cfs_rq
->nr_running
> 1) {
995 u64 slice
= sched_slice(cfs_rq
, se
);
996 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
997 s64 delta
= slice
- ran
;
1006 * Don't schedule slices shorter than 10000ns, that just
1007 * doesn't make sense. Rely on vruntime for fairness.
1010 delta
= max_t(s64
, 10000LL, delta
);
1012 hrtick_start(rq
, delta
);
1017 * called from enqueue/dequeue and updates the hrtick when the
1018 * current task is from our class and nr_running is low enough
1021 static void hrtick_update(struct rq
*rq
)
1023 struct task_struct
*curr
= rq
->curr
;
1025 if (curr
->sched_class
!= &fair_sched_class
)
1028 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
1029 hrtick_start_fair(rq
, curr
);
1031 #else /* !CONFIG_SCHED_HRTICK */
1033 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
1037 static inline void hrtick_update(struct rq
*rq
)
1043 * The enqueue_task method is called before nr_running is
1044 * increased. Here we update the fair scheduling stats and
1045 * then put the task into the rbtree:
1048 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
, bool head
)
1050 struct cfs_rq
*cfs_rq
;
1051 struct sched_entity
*se
= &p
->se
;
1055 flags
|= ENQUEUE_WAKEUP
;
1056 if (p
->state
== TASK_WAKING
)
1057 flags
|= ENQUEUE_MIGRATE
;
1059 for_each_sched_entity(se
) {
1062 cfs_rq
= cfs_rq_of(se
);
1063 enqueue_entity(cfs_rq
, se
, flags
);
1064 flags
= ENQUEUE_WAKEUP
;
1071 * The dequeue_task method is called before nr_running is
1072 * decreased. We remove the task from the rbtree and
1073 * update the fair scheduling stats:
1075 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1077 struct cfs_rq
*cfs_rq
;
1078 struct sched_entity
*se
= &p
->se
;
1080 for_each_sched_entity(se
) {
1081 cfs_rq
= cfs_rq_of(se
);
1082 dequeue_entity(cfs_rq
, se
, sleep
);
1083 /* Don't dequeue parent if it has other entities besides us */
1084 if (cfs_rq
->load
.weight
)
1093 * sched_yield() support is very simple - we dequeue and enqueue.
1095 * If compat_yield is turned on then we requeue to the end of the tree.
1097 static void yield_task_fair(struct rq
*rq
)
1099 struct task_struct
*curr
= rq
->curr
;
1100 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
1101 struct sched_entity
*rightmost
, *se
= &curr
->se
;
1104 * Are we the only task in the tree?
1106 if (unlikely(cfs_rq
->nr_running
== 1))
1109 clear_buddies(cfs_rq
, se
);
1111 if (likely(!sysctl_sched_compat_yield
) && curr
->policy
!= SCHED_BATCH
) {
1112 update_rq_clock(rq
);
1114 * Update run-time statistics of the 'current'.
1116 update_curr(cfs_rq
);
1121 * Find the rightmost entry in the rbtree:
1123 rightmost
= __pick_last_entity(cfs_rq
);
1125 * Already in the rightmost position?
1127 if (unlikely(!rightmost
|| entity_before(rightmost
, se
)))
1131 * Minimally necessary key value to be last in the tree:
1132 * Upon rescheduling, sched_class::put_prev_task() will place
1133 * 'current' within the tree based on its new key value.
1135 se
->vruntime
= rightmost
->vruntime
+ 1;
1140 static void task_waking_fair(struct rq
*rq
, struct task_struct
*p
)
1142 struct sched_entity
*se
= &p
->se
;
1143 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1145 se
->vruntime
-= cfs_rq
->min_vruntime
;
1148 #ifdef CONFIG_FAIR_GROUP_SCHED
1150 * effective_load() calculates the load change as seen from the root_task_group
1152 * Adding load to a group doesn't make a group heavier, but can cause movement
1153 * of group shares between cpus. Assuming the shares were perfectly aligned one
1154 * can calculate the shift in shares.
1156 * The problem is that perfectly aligning the shares is rather expensive, hence
1157 * we try to avoid doing that too often - see update_shares(), which ratelimits
1160 * We compensate this by not only taking the current delta into account, but
1161 * also considering the delta between when the shares were last adjusted and
1164 * We still saw a performance dip, some tracing learned us that between
1165 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1166 * significantly. Therefore try to bias the error in direction of failing
1167 * the affine wakeup.
1170 static long effective_load(struct task_group
*tg
, int cpu
,
1173 struct sched_entity
*se
= tg
->se
[cpu
];
1179 * By not taking the decrease of shares on the other cpu into
1180 * account our error leans towards reducing the affine wakeups.
1182 if (!wl
&& sched_feat(ASYM_EFF_LOAD
))
1185 for_each_sched_entity(se
) {
1186 long S
, rw
, s
, a
, b
;
1190 * Instead of using this increment, also add the difference
1191 * between when the shares were last updated and now.
1193 more_w
= se
->my_q
->load
.weight
- se
->my_q
->rq_weight
;
1197 S
= se
->my_q
->tg
->shares
;
1198 s
= se
->my_q
->shares
;
1199 rw
= se
->my_q
->rq_weight
;
1210 * Assume the group is already running and will
1211 * thus already be accounted for in the weight.
1213 * That is, moving shares between CPUs, does not
1214 * alter the group weight.
1224 static inline unsigned long effective_load(struct task_group
*tg
, int cpu
,
1225 unsigned long wl
, unsigned long wg
)
1232 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
1234 unsigned long this_load
, load
;
1235 int idx
, this_cpu
, prev_cpu
;
1236 unsigned long tl_per_task
;
1237 unsigned int imbalance
;
1238 struct task_group
*tg
;
1239 unsigned long weight
;
1243 this_cpu
= smp_processor_id();
1244 prev_cpu
= task_cpu(p
);
1245 load
= source_load(prev_cpu
, idx
);
1246 this_load
= target_load(this_cpu
, idx
);
1249 * If sync wakeup then subtract the (maximum possible)
1250 * effect of the currently running task from the load
1251 * of the current CPU:
1254 tg
= task_group(current
);
1255 weight
= current
->se
.load
.weight
;
1257 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
1258 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
1262 weight
= p
->se
.load
.weight
;
1264 imbalance
= 100 + (sd
->imbalance_pct
- 100) / 2;
1267 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1268 * due to the sync cause above having dropped this_load to 0, we'll
1269 * always have an imbalance, but there's really nothing you can do
1270 * about that, so that's good too.
1272 * Otherwise check if either cpus are near enough in load to allow this
1273 * task to be woken on this_cpu.
1275 balanced
= !this_load
||
1276 100*(this_load
+ effective_load(tg
, this_cpu
, weight
, weight
)) <=
1277 imbalance
*(load
+ effective_load(tg
, prev_cpu
, 0, weight
));
1280 * If the currently running task will sleep within
1281 * a reasonable amount of time then attract this newly
1284 if (sync
&& balanced
)
1287 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
1288 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1291 (this_load
<= load
&&
1292 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
1294 * This domain has SD_WAKE_AFFINE and
1295 * p is cache cold in this domain, and
1296 * there is no bad imbalance.
1298 schedstat_inc(sd
, ttwu_move_affine
);
1299 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
1307 * find_idlest_group finds and returns the least busy CPU group within the
1310 static struct sched_group
*
1311 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
1312 int this_cpu
, int load_idx
)
1314 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1315 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1316 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1319 unsigned long load
, avg_load
;
1323 /* Skip over this group if it has no CPUs allowed */
1324 if (!cpumask_intersects(sched_group_cpus(group
),
1328 local_group
= cpumask_test_cpu(this_cpu
,
1329 sched_group_cpus(group
));
1331 /* Tally up the load of all CPUs in the group */
1334 for_each_cpu(i
, sched_group_cpus(group
)) {
1335 /* Bias balancing toward cpus of our domain */
1337 load
= source_load(i
, load_idx
);
1339 load
= target_load(i
, load_idx
);
1344 /* Adjust by relative CPU power of the group */
1345 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1348 this_load
= avg_load
;
1350 } else if (avg_load
< min_load
) {
1351 min_load
= avg_load
;
1354 } while (group
= group
->next
, group
!= sd
->groups
);
1356 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1362 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1365 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1367 unsigned long load
, min_load
= ULONG_MAX
;
1371 /* Traverse only the allowed CPUs */
1372 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
1373 load
= weighted_cpuload(i
);
1375 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1385 * Try and locate an idle CPU in the sched_domain.
1388 select_idle_sibling(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
1390 int cpu
= smp_processor_id();
1391 int prev_cpu
= task_cpu(p
);
1395 * If this domain spans both cpu and prev_cpu (see the SD_WAKE_AFFINE
1396 * test in select_task_rq_fair) and the prev_cpu is idle then that's
1397 * always a better target than the current cpu.
1399 if (target
== cpu
&& !cpu_rq(prev_cpu
)->cfs
.nr_running
)
1403 * Otherwise, iterate the domain and find an elegible idle cpu.
1405 for_each_cpu_and(i
, sched_domain_span(sd
), &p
->cpus_allowed
) {
1406 if (!cpu_rq(i
)->cfs
.nr_running
) {
1416 * sched_balance_self: balance the current task (running on cpu) in domains
1417 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1420 * Balance, ie. select the least loaded group.
1422 * Returns the target CPU number, or the same CPU if no balancing is needed.
1424 * preempt must be disabled.
1426 static int select_task_rq_fair(struct task_struct
*p
, int sd_flag
, int wake_flags
)
1428 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
1429 int cpu
= smp_processor_id();
1430 int prev_cpu
= task_cpu(p
);
1432 int want_affine
= 0, cpu_idle
= !current
->pid
;
1434 int sync
= wake_flags
& WF_SYNC
;
1436 if (sd_flag
& SD_BALANCE_WAKE
) {
1437 if (sched_feat(AFFINE_WAKEUPS
) &&
1438 cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
1443 for_each_domain(cpu
, tmp
) {
1444 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
1448 * If power savings logic is enabled for a domain, see if we
1449 * are not overloaded, if so, don't balance wider.
1451 if (tmp
->flags
& (SD_POWERSAVINGS_BALANCE
|SD_PREFER_LOCAL
)) {
1452 unsigned long power
= 0;
1453 unsigned long nr_running
= 0;
1454 unsigned long capacity
;
1457 for_each_cpu(i
, sched_domain_span(tmp
)) {
1458 power
+= power_of(i
);
1459 nr_running
+= cpu_rq(i
)->cfs
.nr_running
;
1462 capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
1464 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1467 if (nr_running
< capacity
)
1472 * While iterating the domains looking for a spanning
1473 * WAKE_AFFINE domain, adjust the affine target to any idle cpu
1474 * in cache sharing domains along the way.
1480 * If both cpu and prev_cpu are part of this domain,
1481 * cpu is a valid SD_WAKE_AFFINE target.
1483 if (cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
)))
1487 * If there's an idle sibling in this domain, make that
1488 * the wake_affine target instead of the current cpu.
1490 if (!cpu_idle
&& tmp
->flags
& SD_SHARE_PKG_RESOURCES
)
1491 target
= select_idle_sibling(p
, tmp
, target
);
1494 if (tmp
->flags
& SD_WAKE_AFFINE
) {
1504 if (!want_sd
&& !want_affine
)
1507 if (!(tmp
->flags
& sd_flag
))
1514 #ifdef CONFIG_FAIR_GROUP_SCHED
1515 if (sched_feat(LB_SHARES_UPDATE
)) {
1517 * Pick the largest domain to update shares over
1520 if (affine_sd
&& (!tmp
||
1521 cpumask_weight(sched_domain_span(affine_sd
)) >
1522 cpumask_weight(sched_domain_span(sd
))))
1531 if (cpu_idle
|| cpu
== prev_cpu
|| wake_affine(affine_sd
, p
, sync
))
1536 int load_idx
= sd
->forkexec_idx
;
1537 struct sched_group
*group
;
1540 if (!(sd
->flags
& sd_flag
)) {
1545 if (sd_flag
& SD_BALANCE_WAKE
)
1546 load_idx
= sd
->wake_idx
;
1548 group
= find_idlest_group(sd
, p
, cpu
, load_idx
);
1554 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
1555 if (new_cpu
== -1 || new_cpu
== cpu
) {
1556 /* Now try balancing at a lower domain level of cpu */
1561 /* Now try balancing at a lower domain level of new_cpu */
1563 weight
= cpumask_weight(sched_domain_span(sd
));
1565 for_each_domain(cpu
, tmp
) {
1566 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
1568 if (tmp
->flags
& sd_flag
)
1571 /* while loop will break here if sd == NULL */
1576 #endif /* CONFIG_SMP */
1578 static unsigned long
1579 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
1581 unsigned long gran
= sysctl_sched_wakeup_granularity
;
1584 * Since its curr running now, convert the gran from real-time
1585 * to virtual-time in his units.
1587 if (sched_feat(ASYM_GRAN
)) {
1589 * By using 'se' instead of 'curr' we penalize light tasks, so
1590 * they get preempted easier. That is, if 'se' < 'curr' then
1591 * the resulting gran will be larger, therefore penalizing the
1592 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1593 * be smaller, again penalizing the lighter task.
1595 * This is especially important for buddies when the leftmost
1596 * task is higher priority than the buddy.
1598 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
1599 gran
= calc_delta_fair(gran
, se
);
1601 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
))
1602 gran
= calc_delta_fair(gran
, curr
);
1609 * Should 'se' preempt 'curr'.
1623 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
1625 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
1630 gran
= wakeup_gran(curr
, se
);
1637 static void set_last_buddy(struct sched_entity
*se
)
1639 if (likely(task_of(se
)->policy
!= SCHED_IDLE
)) {
1640 for_each_sched_entity(se
)
1641 cfs_rq_of(se
)->last
= se
;
1645 static void set_next_buddy(struct sched_entity
*se
)
1647 if (likely(task_of(se
)->policy
!= SCHED_IDLE
)) {
1648 for_each_sched_entity(se
)
1649 cfs_rq_of(se
)->next
= se
;
1654 * Preempt the current task with a newly woken task if needed:
1656 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1658 struct task_struct
*curr
= rq
->curr
;
1659 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
1660 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
1661 int sync
= wake_flags
& WF_SYNC
;
1662 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
1664 if (unlikely(rt_prio(p
->prio
)))
1667 if (unlikely(p
->sched_class
!= &fair_sched_class
))
1670 if (unlikely(se
== pse
))
1673 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
))
1674 set_next_buddy(pse
);
1677 * We can come here with TIF_NEED_RESCHED already set from new task
1680 if (test_tsk_need_resched(curr
))
1684 * Batch and idle tasks do not preempt (their preemption is driven by
1687 if (unlikely(p
->policy
!= SCHED_NORMAL
))
1690 /* Idle tasks are by definition preempted by everybody. */
1691 if (unlikely(curr
->policy
== SCHED_IDLE
))
1694 if (sched_feat(WAKEUP_SYNC
) && sync
)
1697 if (!sched_feat(WAKEUP_PREEMPT
))
1700 update_curr(cfs_rq
);
1701 find_matching_se(&se
, &pse
);
1703 if (wakeup_preempt_entity(se
, pse
) == 1)
1711 * Only set the backward buddy when the current task is still
1712 * on the rq. This can happen when a wakeup gets interleaved
1713 * with schedule on the ->pre_schedule() or idle_balance()
1714 * point, either of which can * drop the rq lock.
1716 * Also, during early boot the idle thread is in the fair class,
1717 * for obvious reasons its a bad idea to schedule back to it.
1719 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
1722 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
1726 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
1728 struct task_struct
*p
;
1729 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
1730 struct sched_entity
*se
;
1732 if (!cfs_rq
->nr_running
)
1736 se
= pick_next_entity(cfs_rq
);
1737 set_next_entity(cfs_rq
, se
);
1738 cfs_rq
= group_cfs_rq(se
);
1742 hrtick_start_fair(rq
, p
);
1748 * Account for a descheduled task:
1750 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
1752 struct sched_entity
*se
= &prev
->se
;
1753 struct cfs_rq
*cfs_rq
;
1755 for_each_sched_entity(se
) {
1756 cfs_rq
= cfs_rq_of(se
);
1757 put_prev_entity(cfs_rq
, se
);
1762 /**************************************************
1763 * Fair scheduling class load-balancing methods:
1767 * pull_task - move a task from a remote runqueue to the local runqueue.
1768 * Both runqueues must be locked.
1770 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
1771 struct rq
*this_rq
, int this_cpu
)
1773 deactivate_task(src_rq
, p
, 0);
1774 set_task_cpu(p
, this_cpu
);
1775 activate_task(this_rq
, p
, 0);
1776 check_preempt_curr(this_rq
, p
, 0);
1780 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1783 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
1784 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1787 int tsk_cache_hot
= 0;
1789 * We do not migrate tasks that are:
1790 * 1) running (obviously), or
1791 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1792 * 3) are cache-hot on their current CPU.
1794 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
1795 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
1800 if (task_running(rq
, p
)) {
1801 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
1806 * Aggressive migration if:
1807 * 1) task is cache cold, or
1808 * 2) too many balance attempts have failed.
1811 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
1812 if (!tsk_cache_hot
||
1813 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
1814 #ifdef CONFIG_SCHEDSTATS
1815 if (tsk_cache_hot
) {
1816 schedstat_inc(sd
, lb_hot_gained
[idle
]);
1817 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
1823 if (tsk_cache_hot
) {
1824 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
1831 * move_one_task tries to move exactly one task from busiest to this_rq, as
1832 * part of active balancing operations within "domain".
1833 * Returns 1 if successful and 0 otherwise.
1835 * Called with both runqueues locked.
1838 move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1839 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1841 struct task_struct
*p
, *n
;
1842 struct cfs_rq
*cfs_rq
;
1845 for_each_leaf_cfs_rq(busiest
, cfs_rq
) {
1846 list_for_each_entry_safe(p
, n
, &cfs_rq
->tasks
, se
.group_node
) {
1848 if (!can_migrate_task(p
, busiest
, this_cpu
,
1852 pull_task(busiest
, p
, this_rq
, this_cpu
);
1854 * Right now, this is only the second place pull_task()
1855 * is called, so we can safely collect pull_task()
1856 * stats here rather than inside pull_task().
1858 schedstat_inc(sd
, lb_gained
[idle
]);
1866 static unsigned long
1867 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1868 unsigned long max_load_move
, struct sched_domain
*sd
,
1869 enum cpu_idle_type idle
, int *all_pinned
,
1870 int *this_best_prio
, struct cfs_rq
*busiest_cfs_rq
)
1872 int loops
= 0, pulled
= 0, pinned
= 0;
1873 long rem_load_move
= max_load_move
;
1874 struct task_struct
*p
, *n
;
1876 if (max_load_move
== 0)
1881 list_for_each_entry_safe(p
, n
, &busiest_cfs_rq
->tasks
, se
.group_node
) {
1882 if (loops
++ > sysctl_sched_nr_migrate
)
1885 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
1886 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
))
1889 pull_task(busiest
, p
, this_rq
, this_cpu
);
1891 rem_load_move
-= p
->se
.load
.weight
;
1893 #ifdef CONFIG_PREEMPT
1895 * NEWIDLE balancing is a source of latency, so preemptible
1896 * kernels will stop after the first task is pulled to minimize
1897 * the critical section.
1899 if (idle
== CPU_NEWLY_IDLE
)
1904 * We only want to steal up to the prescribed amount of
1907 if (rem_load_move
<= 0)
1910 if (p
->prio
< *this_best_prio
)
1911 *this_best_prio
= p
->prio
;
1915 * Right now, this is one of only two places pull_task() is called,
1916 * so we can safely collect pull_task() stats here rather than
1917 * inside pull_task().
1919 schedstat_add(sd
, lb_gained
[idle
], pulled
);
1922 *all_pinned
= pinned
;
1924 return max_load_move
- rem_load_move
;
1927 #ifdef CONFIG_FAIR_GROUP_SCHED
1928 static unsigned long
1929 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1930 unsigned long max_load_move
,
1931 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1932 int *all_pinned
, int *this_best_prio
)
1934 long rem_load_move
= max_load_move
;
1935 int busiest_cpu
= cpu_of(busiest
);
1936 struct task_group
*tg
;
1939 update_h_load(busiest_cpu
);
1941 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
1942 struct cfs_rq
*busiest_cfs_rq
= tg
->cfs_rq
[busiest_cpu
];
1943 unsigned long busiest_h_load
= busiest_cfs_rq
->h_load
;
1944 unsigned long busiest_weight
= busiest_cfs_rq
->load
.weight
;
1945 u64 rem_load
, moved_load
;
1950 if (!busiest_cfs_rq
->task_weight
)
1953 rem_load
= (u64
)rem_load_move
* busiest_weight
;
1954 rem_load
= div_u64(rem_load
, busiest_h_load
+ 1);
1956 moved_load
= balance_tasks(this_rq
, this_cpu
, busiest
,
1957 rem_load
, sd
, idle
, all_pinned
, this_best_prio
,
1963 moved_load
*= busiest_h_load
;
1964 moved_load
= div_u64(moved_load
, busiest_weight
+ 1);
1966 rem_load_move
-= moved_load
;
1967 if (rem_load_move
< 0)
1972 return max_load_move
- rem_load_move
;
1975 static unsigned long
1976 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1977 unsigned long max_load_move
,
1978 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1979 int *all_pinned
, int *this_best_prio
)
1981 return balance_tasks(this_rq
, this_cpu
, busiest
,
1982 max_load_move
, sd
, idle
, all_pinned
,
1983 this_best_prio
, &busiest
->cfs
);
1988 * move_tasks tries to move up to max_load_move weighted load from busiest to
1989 * this_rq, as part of a balancing operation within domain "sd".
1990 * Returns 1 if successful and 0 otherwise.
1992 * Called with both runqueues locked.
1994 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1995 unsigned long max_load_move
,
1996 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1999 unsigned long total_load_moved
= 0, load_moved
;
2000 int this_best_prio
= this_rq
->curr
->prio
;
2003 load_moved
= load_balance_fair(this_rq
, this_cpu
, busiest
,
2004 max_load_move
- total_load_moved
,
2005 sd
, idle
, all_pinned
, &this_best_prio
);
2007 total_load_moved
+= load_moved
;
2009 #ifdef CONFIG_PREEMPT
2011 * NEWIDLE balancing is a source of latency, so preemptible
2012 * kernels will stop after the first task is pulled to minimize
2013 * the critical section.
2015 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
2018 if (raw_spin_is_contended(&this_rq
->lock
) ||
2019 raw_spin_is_contended(&busiest
->lock
))
2022 } while (load_moved
&& max_load_move
> total_load_moved
);
2024 return total_load_moved
> 0;
2027 /********** Helpers for find_busiest_group ************************/
2029 * sd_lb_stats - Structure to store the statistics of a sched_domain
2030 * during load balancing.
2032 struct sd_lb_stats
{
2033 struct sched_group
*busiest
; /* Busiest group in this sd */
2034 struct sched_group
*this; /* Local group in this sd */
2035 unsigned long total_load
; /* Total load of all groups in sd */
2036 unsigned long total_pwr
; /* Total power of all groups in sd */
2037 unsigned long avg_load
; /* Average load across all groups in sd */
2039 /** Statistics of this group */
2040 unsigned long this_load
;
2041 unsigned long this_load_per_task
;
2042 unsigned long this_nr_running
;
2044 /* Statistics of the busiest group */
2045 unsigned long max_load
;
2046 unsigned long busiest_load_per_task
;
2047 unsigned long busiest_nr_running
;
2048 unsigned long busiest_group_capacity
;
2050 int group_imb
; /* Is there imbalance in this sd */
2051 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2052 int power_savings_balance
; /* Is powersave balance needed for this sd */
2053 struct sched_group
*group_min
; /* Least loaded group in sd */
2054 struct sched_group
*group_leader
; /* Group which relieves group_min */
2055 unsigned long min_load_per_task
; /* load_per_task in group_min */
2056 unsigned long leader_nr_running
; /* Nr running of group_leader */
2057 unsigned long min_nr_running
; /* Nr running of group_min */
2062 * sg_lb_stats - stats of a sched_group required for load_balancing
2064 struct sg_lb_stats
{
2065 unsigned long avg_load
; /*Avg load across the CPUs of the group */
2066 unsigned long group_load
; /* Total load over the CPUs of the group */
2067 unsigned long sum_nr_running
; /* Nr tasks running in the group */
2068 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
2069 unsigned long group_capacity
;
2070 int group_imb
; /* Is there an imbalance in the group ? */
2074 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2075 * @group: The group whose first cpu is to be returned.
2077 static inline unsigned int group_first_cpu(struct sched_group
*group
)
2079 return cpumask_first(sched_group_cpus(group
));
2083 * get_sd_load_idx - Obtain the load index for a given sched domain.
2084 * @sd: The sched_domain whose load_idx is to be obtained.
2085 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2087 static inline int get_sd_load_idx(struct sched_domain
*sd
,
2088 enum cpu_idle_type idle
)
2094 load_idx
= sd
->busy_idx
;
2097 case CPU_NEWLY_IDLE
:
2098 load_idx
= sd
->newidle_idx
;
2101 load_idx
= sd
->idle_idx
;
2109 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2111 * init_sd_power_savings_stats - Initialize power savings statistics for
2112 * the given sched_domain, during load balancing.
2114 * @sd: Sched domain whose power-savings statistics are to be initialized.
2115 * @sds: Variable containing the statistics for sd.
2116 * @idle: Idle status of the CPU at which we're performing load-balancing.
2118 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
2119 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
2122 * Busy processors will not participate in power savings
2125 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2126 sds
->power_savings_balance
= 0;
2128 sds
->power_savings_balance
= 1;
2129 sds
->min_nr_running
= ULONG_MAX
;
2130 sds
->leader_nr_running
= 0;
2135 * update_sd_power_savings_stats - Update the power saving stats for a
2136 * sched_domain while performing load balancing.
2138 * @group: sched_group belonging to the sched_domain under consideration.
2139 * @sds: Variable containing the statistics of the sched_domain
2140 * @local_group: Does group contain the CPU for which we're performing
2142 * @sgs: Variable containing the statistics of the group.
2144 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
2145 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
2148 if (!sds
->power_savings_balance
)
2152 * If the local group is idle or completely loaded
2153 * no need to do power savings balance at this domain
2155 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
2156 !sds
->this_nr_running
))
2157 sds
->power_savings_balance
= 0;
2160 * If a group is already running at full capacity or idle,
2161 * don't include that group in power savings calculations
2163 if (!sds
->power_savings_balance
||
2164 sgs
->sum_nr_running
>= sgs
->group_capacity
||
2165 !sgs
->sum_nr_running
)
2169 * Calculate the group which has the least non-idle load.
2170 * This is the group from where we need to pick up the load
2173 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
2174 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
2175 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
2176 sds
->group_min
= group
;
2177 sds
->min_nr_running
= sgs
->sum_nr_running
;
2178 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
2179 sgs
->sum_nr_running
;
2183 * Calculate the group which is almost near its
2184 * capacity but still has some space to pick up some load
2185 * from other group and save more power
2187 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
2190 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
2191 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
2192 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
2193 sds
->group_leader
= group
;
2194 sds
->leader_nr_running
= sgs
->sum_nr_running
;
2199 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2200 * @sds: Variable containing the statistics of the sched_domain
2201 * under consideration.
2202 * @this_cpu: Cpu at which we're currently performing load-balancing.
2203 * @imbalance: Variable to store the imbalance.
2206 * Check if we have potential to perform some power-savings balance.
2207 * If yes, set the busiest group to be the least loaded group in the
2208 * sched_domain, so that it's CPUs can be put to idle.
2210 * Returns 1 if there is potential to perform power-savings balance.
2213 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
2214 int this_cpu
, unsigned long *imbalance
)
2216 if (!sds
->power_savings_balance
)
2219 if (sds
->this != sds
->group_leader
||
2220 sds
->group_leader
== sds
->group_min
)
2223 *imbalance
= sds
->min_load_per_task
;
2224 sds
->busiest
= sds
->group_min
;
2229 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2230 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
2231 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
2236 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
2237 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
2242 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
2243 int this_cpu
, unsigned long *imbalance
)
2247 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2250 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
2252 return SCHED_LOAD_SCALE
;
2255 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
2257 return default_scale_freq_power(sd
, cpu
);
2260 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
2262 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
2263 unsigned long smt_gain
= sd
->smt_gain
;
2270 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
2272 return default_scale_smt_power(sd
, cpu
);
2275 unsigned long scale_rt_power(int cpu
)
2277 struct rq
*rq
= cpu_rq(cpu
);
2278 u64 total
, available
;
2280 sched_avg_update(rq
);
2282 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
2283 available
= total
- rq
->rt_avg
;
2285 if (unlikely((s64
)total
< SCHED_LOAD_SCALE
))
2286 total
= SCHED_LOAD_SCALE
;
2288 total
>>= SCHED_LOAD_SHIFT
;
2290 return div_u64(available
, total
);
2293 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
2295 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
2296 unsigned long power
= SCHED_LOAD_SCALE
;
2297 struct sched_group
*sdg
= sd
->groups
;
2299 if (sched_feat(ARCH_POWER
))
2300 power
*= arch_scale_freq_power(sd
, cpu
);
2302 power
*= default_scale_freq_power(sd
, cpu
);
2304 power
>>= SCHED_LOAD_SHIFT
;
2306 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
2307 if (sched_feat(ARCH_POWER
))
2308 power
*= arch_scale_smt_power(sd
, cpu
);
2310 power
*= default_scale_smt_power(sd
, cpu
);
2312 power
>>= SCHED_LOAD_SHIFT
;
2315 power
*= scale_rt_power(cpu
);
2316 power
>>= SCHED_LOAD_SHIFT
;
2321 sdg
->cpu_power
= power
;
2324 static void update_group_power(struct sched_domain
*sd
, int cpu
)
2326 struct sched_domain
*child
= sd
->child
;
2327 struct sched_group
*group
, *sdg
= sd
->groups
;
2328 unsigned long power
;
2331 update_cpu_power(sd
, cpu
);
2337 group
= child
->groups
;
2339 power
+= group
->cpu_power
;
2340 group
= group
->next
;
2341 } while (group
!= child
->groups
);
2343 sdg
->cpu_power
= power
;
2347 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2348 * @sd: The sched_domain whose statistics are to be updated.
2349 * @group: sched_group whose statistics are to be updated.
2350 * @this_cpu: Cpu for which load balance is currently performed.
2351 * @idle: Idle status of this_cpu
2352 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2353 * @sd_idle: Idle status of the sched_domain containing group.
2354 * @local_group: Does group contain this_cpu.
2355 * @cpus: Set of cpus considered for load balancing.
2356 * @balance: Should we balance.
2357 * @sgs: variable to hold the statistics for this group.
2359 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
2360 struct sched_group
*group
, int this_cpu
,
2361 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
2362 int local_group
, const struct cpumask
*cpus
,
2363 int *balance
, struct sg_lb_stats
*sgs
)
2365 unsigned long load
, max_cpu_load
, min_cpu_load
;
2367 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2368 unsigned long avg_load_per_task
= 0;
2371 balance_cpu
= group_first_cpu(group
);
2373 /* Tally up the load of all CPUs in the group */
2375 min_cpu_load
= ~0UL;
2377 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
2378 struct rq
*rq
= cpu_rq(i
);
2380 if (*sd_idle
&& rq
->nr_running
)
2383 /* Bias balancing toward cpus of our domain */
2385 if (idle_cpu(i
) && !first_idle_cpu
) {
2390 load
= target_load(i
, load_idx
);
2392 load
= source_load(i
, load_idx
);
2393 if (load
> max_cpu_load
)
2394 max_cpu_load
= load
;
2395 if (min_cpu_load
> load
)
2396 min_cpu_load
= load
;
2399 sgs
->group_load
+= load
;
2400 sgs
->sum_nr_running
+= rq
->nr_running
;
2401 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
2406 * First idle cpu or the first cpu(busiest) in this sched group
2407 * is eligible for doing load balancing at this and above
2408 * domains. In the newly idle case, we will allow all the cpu's
2409 * to do the newly idle load balance.
2411 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2412 balance_cpu
!= this_cpu
) {
2417 update_group_power(sd
, this_cpu
);
2419 /* Adjust by relative CPU power of the group */
2420 sgs
->avg_load
= (sgs
->group_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2423 * Consider the group unbalanced when the imbalance is larger
2424 * than the average weight of two tasks.
2426 * APZ: with cgroup the avg task weight can vary wildly and
2427 * might not be a suitable number - should we keep a
2428 * normalized nr_running number somewhere that negates
2431 if (sgs
->sum_nr_running
)
2432 avg_load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
2434 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
2437 sgs
->group_capacity
=
2438 DIV_ROUND_CLOSEST(group
->cpu_power
, SCHED_LOAD_SCALE
);
2442 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2443 * @sd: sched_domain whose statistics are to be updated.
2444 * @this_cpu: Cpu for which load balance is currently performed.
2445 * @idle: Idle status of this_cpu
2446 * @sd_idle: Idle status of the sched_domain containing group.
2447 * @cpus: Set of cpus considered for load balancing.
2448 * @balance: Should we balance.
2449 * @sds: variable to hold the statistics for this sched_domain.
2451 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
2452 enum cpu_idle_type idle
, int *sd_idle
,
2453 const struct cpumask
*cpus
, int *balance
,
2454 struct sd_lb_stats
*sds
)
2456 struct sched_domain
*child
= sd
->child
;
2457 struct sched_group
*group
= sd
->groups
;
2458 struct sg_lb_stats sgs
;
2459 int load_idx
, prefer_sibling
= 0;
2461 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
2464 init_sd_power_savings_stats(sd
, sds
, idle
);
2465 load_idx
= get_sd_load_idx(sd
, idle
);
2470 local_group
= cpumask_test_cpu(this_cpu
,
2471 sched_group_cpus(group
));
2472 memset(&sgs
, 0, sizeof(sgs
));
2473 update_sg_lb_stats(sd
, group
, this_cpu
, idle
, load_idx
, sd_idle
,
2474 local_group
, cpus
, balance
, &sgs
);
2476 if (local_group
&& !(*balance
))
2479 sds
->total_load
+= sgs
.group_load
;
2480 sds
->total_pwr
+= group
->cpu_power
;
2483 * In case the child domain prefers tasks go to siblings
2484 * first, lower the group capacity to one so that we'll try
2485 * and move all the excess tasks away.
2488 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
2491 sds
->this_load
= sgs
.avg_load
;
2493 sds
->this_nr_running
= sgs
.sum_nr_running
;
2494 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
2495 } else if (sgs
.avg_load
> sds
->max_load
&&
2496 (sgs
.sum_nr_running
> sgs
.group_capacity
||
2498 sds
->max_load
= sgs
.avg_load
;
2499 sds
->busiest
= group
;
2500 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
2501 sds
->busiest_group_capacity
= sgs
.group_capacity
;
2502 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
2503 sds
->group_imb
= sgs
.group_imb
;
2506 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
2507 group
= group
->next
;
2508 } while (group
!= sd
->groups
);
2512 * fix_small_imbalance - Calculate the minor imbalance that exists
2513 * amongst the groups of a sched_domain, during
2515 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2516 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2517 * @imbalance: Variable to store the imbalance.
2519 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
2520 int this_cpu
, unsigned long *imbalance
)
2522 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
2523 unsigned int imbn
= 2;
2524 unsigned long scaled_busy_load_per_task
;
2526 if (sds
->this_nr_running
) {
2527 sds
->this_load_per_task
/= sds
->this_nr_running
;
2528 if (sds
->busiest_load_per_task
>
2529 sds
->this_load_per_task
)
2532 sds
->this_load_per_task
=
2533 cpu_avg_load_per_task(this_cpu
);
2535 scaled_busy_load_per_task
= sds
->busiest_load_per_task
2537 scaled_busy_load_per_task
/= sds
->busiest
->cpu_power
;
2539 if (sds
->max_load
- sds
->this_load
+ scaled_busy_load_per_task
>=
2540 (scaled_busy_load_per_task
* imbn
)) {
2541 *imbalance
= sds
->busiest_load_per_task
;
2546 * OK, we don't have enough imbalance to justify moving tasks,
2547 * however we may be able to increase total CPU power used by
2551 pwr_now
+= sds
->busiest
->cpu_power
*
2552 min(sds
->busiest_load_per_task
, sds
->max_load
);
2553 pwr_now
+= sds
->this->cpu_power
*
2554 min(sds
->this_load_per_task
, sds
->this_load
);
2555 pwr_now
/= SCHED_LOAD_SCALE
;
2557 /* Amount of load we'd subtract */
2558 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
2559 sds
->busiest
->cpu_power
;
2560 if (sds
->max_load
> tmp
)
2561 pwr_move
+= sds
->busiest
->cpu_power
*
2562 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
2564 /* Amount of load we'd add */
2565 if (sds
->max_load
* sds
->busiest
->cpu_power
<
2566 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
2567 tmp
= (sds
->max_load
* sds
->busiest
->cpu_power
) /
2568 sds
->this->cpu_power
;
2570 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
2571 sds
->this->cpu_power
;
2572 pwr_move
+= sds
->this->cpu_power
*
2573 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
2574 pwr_move
/= SCHED_LOAD_SCALE
;
2576 /* Move if we gain throughput */
2577 if (pwr_move
> pwr_now
)
2578 *imbalance
= sds
->busiest_load_per_task
;
2582 * calculate_imbalance - Calculate the amount of imbalance present within the
2583 * groups of a given sched_domain during load balance.
2584 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
2585 * @this_cpu: Cpu for which currently load balance is being performed.
2586 * @imbalance: The variable to store the imbalance.
2588 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
2589 unsigned long *imbalance
)
2591 unsigned long max_pull
, load_above_capacity
= ~0UL;
2593 sds
->busiest_load_per_task
/= sds
->busiest_nr_running
;
2594 if (sds
->group_imb
) {
2595 sds
->busiest_load_per_task
=
2596 min(sds
->busiest_load_per_task
, sds
->avg_load
);
2600 * In the presence of smp nice balancing, certain scenarios can have
2601 * max load less than avg load(as we skip the groups at or below
2602 * its cpu_power, while calculating max_load..)
2604 if (sds
->max_load
< sds
->avg_load
) {
2606 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
2609 if (!sds
->group_imb
) {
2611 * Don't want to pull so many tasks that a group would go idle.
2613 load_above_capacity
= (sds
->busiest_nr_running
-
2614 sds
->busiest_group_capacity
);
2616 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_LOAD_SCALE
);
2618 load_above_capacity
/= sds
->busiest
->cpu_power
;
2622 * We're trying to get all the cpus to the average_load, so we don't
2623 * want to push ourselves above the average load, nor do we wish to
2624 * reduce the max loaded cpu below the average load. At the same time,
2625 * we also don't want to reduce the group load below the group capacity
2626 * (so that we can implement power-savings policies etc). Thus we look
2627 * for the minimum possible imbalance.
2628 * Be careful of negative numbers as they'll appear as very large values
2629 * with unsigned longs.
2631 max_pull
= min(sds
->max_load
- sds
->avg_load
, load_above_capacity
);
2633 /* How much load to actually move to equalise the imbalance */
2634 *imbalance
= min(max_pull
* sds
->busiest
->cpu_power
,
2635 (sds
->avg_load
- sds
->this_load
) * sds
->this->cpu_power
)
2639 * if *imbalance is less than the average load per runnable task
2640 * there is no gaurantee that any tasks will be moved so we'll have
2641 * a think about bumping its value to force at least one task to be
2644 if (*imbalance
< sds
->busiest_load_per_task
)
2645 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
2648 /******* find_busiest_group() helpers end here *********************/
2651 * find_busiest_group - Returns the busiest group within the sched_domain
2652 * if there is an imbalance. If there isn't an imbalance, and
2653 * the user has opted for power-savings, it returns a group whose
2654 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
2655 * such a group exists.
2657 * Also calculates the amount of weighted load which should be moved
2658 * to restore balance.
2660 * @sd: The sched_domain whose busiest group is to be returned.
2661 * @this_cpu: The cpu for which load balancing is currently being performed.
2662 * @imbalance: Variable which stores amount of weighted load which should
2663 * be moved to restore balance/put a group to idle.
2664 * @idle: The idle status of this_cpu.
2665 * @sd_idle: The idleness of sd
2666 * @cpus: The set of CPUs under consideration for load-balancing.
2667 * @balance: Pointer to a variable indicating if this_cpu
2668 * is the appropriate cpu to perform load balancing at this_level.
2670 * Returns: - the busiest group if imbalance exists.
2671 * - If no imbalance and user has opted for power-savings balance,
2672 * return the least loaded group whose CPUs can be
2673 * put to idle by rebalancing its tasks onto our group.
2675 static struct sched_group
*
2676 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2677 unsigned long *imbalance
, enum cpu_idle_type idle
,
2678 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
2680 struct sd_lb_stats sds
;
2682 memset(&sds
, 0, sizeof(sds
));
2685 * Compute the various statistics relavent for load balancing at
2688 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
2691 /* Cases where imbalance does not exist from POV of this_cpu */
2692 /* 1) this_cpu is not the appropriate cpu to perform load balancing
2694 * 2) There is no busy sibling group to pull from.
2695 * 3) This group is the busiest group.
2696 * 4) This group is more busy than the avg busieness at this
2698 * 5) The imbalance is within the specified limit.
2703 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
2706 if (sds
.this_load
>= sds
.max_load
)
2709 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
2711 if (sds
.this_load
>= sds
.avg_load
)
2714 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
2717 /* Looks like there is an imbalance. Compute it */
2718 calculate_imbalance(&sds
, this_cpu
, imbalance
);
2723 * There is no obvious imbalance. But check if we can do some balancing
2726 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
2734 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2737 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2738 unsigned long imbalance
, const struct cpumask
*cpus
)
2740 struct rq
*busiest
= NULL
, *rq
;
2741 unsigned long max_load
= 0;
2744 for_each_cpu(i
, sched_group_cpus(group
)) {
2745 unsigned long power
= power_of(i
);
2746 unsigned long capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
2749 if (!cpumask_test_cpu(i
, cpus
))
2753 wl
= weighted_cpuload(i
);
2756 * When comparing with imbalance, use weighted_cpuload()
2757 * which is not scaled with the cpu power.
2759 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
2763 * For the load comparisons with the other cpu's, consider
2764 * the weighted_cpuload() scaled with the cpu power, so that
2765 * the load can be moved away from the cpu that is potentially
2766 * running at a lower capacity.
2768 wl
= (wl
* SCHED_LOAD_SCALE
) / power
;
2770 if (wl
> max_load
) {
2780 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2781 * so long as it is large enough.
2783 #define MAX_PINNED_INTERVAL 512
2785 /* Working cpumask for load_balance and load_balance_newidle. */
2786 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
2788 static int need_active_balance(struct sched_domain
*sd
, int sd_idle
, int idle
)
2790 if (idle
== CPU_NEWLY_IDLE
) {
2792 * The only task running in a non-idle cpu can be moved to this
2793 * cpu in an attempt to completely freeup the other CPU
2796 * The package power saving logic comes from
2797 * find_busiest_group(). If there are no imbalance, then
2798 * f_b_g() will return NULL. However when sched_mc={1,2} then
2799 * f_b_g() will select a group from which a running task may be
2800 * pulled to this cpu in order to make the other package idle.
2801 * If there is no opportunity to make a package idle and if
2802 * there are no imbalance, then f_b_g() will return NULL and no
2803 * action will be taken in load_balance_newidle().
2805 * Under normal task pull operation due to imbalance, there
2806 * will be more than one task in the source run queue and
2807 * move_tasks() will succeed. ld_moved will be true and this
2808 * active balance code will not be triggered.
2810 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2811 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2814 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
2818 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
2822 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2823 * tasks if there is an imbalance.
2825 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2826 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2829 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2830 struct sched_group
*group
;
2831 unsigned long imbalance
;
2833 unsigned long flags
;
2834 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
2836 cpumask_copy(cpus
, cpu_active_mask
);
2839 * When power savings policy is enabled for the parent domain, idle
2840 * sibling can pick up load irrespective of busy siblings. In this case,
2841 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2842 * portraying it as CPU_NOT_IDLE.
2844 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2845 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2848 schedstat_inc(sd
, lb_count
[idle
]);
2852 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2859 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2863 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
2865 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2869 BUG_ON(busiest
== this_rq
);
2871 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2874 if (busiest
->nr_running
> 1) {
2876 * Attempt to move tasks. If find_busiest_group has found
2877 * an imbalance but busiest->nr_running <= 1, the group is
2878 * still unbalanced. ld_moved simply stays zero, so it is
2879 * correctly treated as an imbalance.
2881 local_irq_save(flags
);
2882 double_rq_lock(this_rq
, busiest
);
2883 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2884 imbalance
, sd
, idle
, &all_pinned
);
2885 double_rq_unlock(this_rq
, busiest
);
2886 local_irq_restore(flags
);
2889 * some other cpu did the load balance for us.
2891 if (ld_moved
&& this_cpu
!= smp_processor_id())
2892 resched_cpu(this_cpu
);
2894 /* All tasks on this runqueue were pinned by CPU affinity */
2895 if (unlikely(all_pinned
)) {
2896 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
2897 if (!cpumask_empty(cpus
))
2904 schedstat_inc(sd
, lb_failed
[idle
]);
2905 sd
->nr_balance_failed
++;
2907 if (need_active_balance(sd
, sd_idle
, idle
)) {
2908 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
2910 /* don't kick the migration_thread, if the curr
2911 * task on busiest cpu can't be moved to this_cpu
2913 if (!cpumask_test_cpu(this_cpu
,
2914 &busiest
->curr
->cpus_allowed
)) {
2915 raw_spin_unlock_irqrestore(&busiest
->lock
,
2918 goto out_one_pinned
;
2921 if (!busiest
->active_balance
) {
2922 busiest
->active_balance
= 1;
2923 busiest
->push_cpu
= this_cpu
;
2926 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
2928 wake_up_process(busiest
->migration_thread
);
2931 * We've kicked active balancing, reset the failure
2934 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2937 sd
->nr_balance_failed
= 0;
2939 if (likely(!active_balance
)) {
2940 /* We were unbalanced, so reset the balancing interval */
2941 sd
->balance_interval
= sd
->min_interval
;
2944 * If we've begun active balancing, start to back off. This
2945 * case may not be covered by the all_pinned logic if there
2946 * is only 1 task on the busy runqueue (because we don't call
2949 if (sd
->balance_interval
< sd
->max_interval
)
2950 sd
->balance_interval
*= 2;
2953 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2954 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2960 schedstat_inc(sd
, lb_balanced
[idle
]);
2962 sd
->nr_balance_failed
= 0;
2965 /* tune up the balancing interval */
2966 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2967 (sd
->balance_interval
< sd
->max_interval
))
2968 sd
->balance_interval
*= 2;
2970 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2971 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2982 * idle_balance is called by schedule() if this_cpu is about to become
2983 * idle. Attempts to pull tasks from other CPUs.
2985 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2987 struct sched_domain
*sd
;
2988 int pulled_task
= 0;
2989 unsigned long next_balance
= jiffies
+ HZ
;
2991 this_rq
->idle_stamp
= this_rq
->clock
;
2993 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
2997 * Drop the rq->lock, but keep IRQ/preempt disabled.
2999 raw_spin_unlock(&this_rq
->lock
);
3001 for_each_domain(this_cpu
, sd
) {
3002 unsigned long interval
;
3005 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3008 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
3009 /* If we've pulled tasks over stop searching: */
3010 pulled_task
= load_balance(this_cpu
, this_rq
,
3011 sd
, CPU_NEWLY_IDLE
, &balance
);
3014 interval
= msecs_to_jiffies(sd
->balance_interval
);
3015 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3016 next_balance
= sd
->last_balance
+ interval
;
3018 this_rq
->idle_stamp
= 0;
3023 raw_spin_lock(&this_rq
->lock
);
3025 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3027 * We are going idle. next_balance may be set based on
3028 * a busy processor. So reset next_balance.
3030 this_rq
->next_balance
= next_balance
;
3035 * active_load_balance is run by migration threads. It pushes running tasks
3036 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3037 * running on each physical CPU where possible, and avoids physical /
3038 * logical imbalances.
3040 * Called with busiest_rq locked.
3042 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3044 int target_cpu
= busiest_rq
->push_cpu
;
3045 struct sched_domain
*sd
;
3046 struct rq
*target_rq
;
3048 /* Is there any task to move? */
3049 if (busiest_rq
->nr_running
<= 1)
3052 target_rq
= cpu_rq(target_cpu
);
3055 * This condition is "impossible", if it occurs
3056 * we need to fix it. Originally reported by
3057 * Bjorn Helgaas on a 128-cpu setup.
3059 BUG_ON(busiest_rq
== target_rq
);
3061 /* move a task from busiest_rq to target_rq */
3062 double_lock_balance(busiest_rq
, target_rq
);
3064 /* Search for an sd spanning us and the target CPU. */
3065 for_each_domain(target_cpu
, sd
) {
3066 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3067 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
3072 schedstat_inc(sd
, alb_count
);
3074 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3076 schedstat_inc(sd
, alb_pushed
);
3078 schedstat_inc(sd
, alb_failed
);
3080 double_unlock_balance(busiest_rq
, target_rq
);
3085 atomic_t load_balancer
;
3086 cpumask_var_t cpu_mask
;
3087 cpumask_var_t ilb_grp_nohz_mask
;
3088 } nohz ____cacheline_aligned
= {
3089 .load_balancer
= ATOMIC_INIT(-1),
3092 int get_nohz_load_balancer(void)
3094 return atomic_read(&nohz
.load_balancer
);
3097 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3099 * lowest_flag_domain - Return lowest sched_domain containing flag.
3100 * @cpu: The cpu whose lowest level of sched domain is to
3102 * @flag: The flag to check for the lowest sched_domain
3103 * for the given cpu.
3105 * Returns the lowest sched_domain of a cpu which contains the given flag.
3107 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
3109 struct sched_domain
*sd
;
3111 for_each_domain(cpu
, sd
)
3112 if (sd
&& (sd
->flags
& flag
))
3119 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3120 * @cpu: The cpu whose domains we're iterating over.
3121 * @sd: variable holding the value of the power_savings_sd
3123 * @flag: The flag to filter the sched_domains to be iterated.
3125 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3126 * set, starting from the lowest sched_domain to the highest.
3128 #define for_each_flag_domain(cpu, sd, flag) \
3129 for (sd = lowest_flag_domain(cpu, flag); \
3130 (sd && (sd->flags & flag)); sd = sd->parent)
3133 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3134 * @ilb_group: group to be checked for semi-idleness
3136 * Returns: 1 if the group is semi-idle. 0 otherwise.
3138 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3139 * and atleast one non-idle CPU. This helper function checks if the given
3140 * sched_group is semi-idle or not.
3142 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
3144 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
3145 sched_group_cpus(ilb_group
));
3148 * A sched_group is semi-idle when it has atleast one busy cpu
3149 * and atleast one idle cpu.
3151 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
3154 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
3160 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3161 * @cpu: The cpu which is nominating a new idle_load_balancer.
3163 * Returns: Returns the id of the idle load balancer if it exists,
3164 * Else, returns >= nr_cpu_ids.
3166 * This algorithm picks the idle load balancer such that it belongs to a
3167 * semi-idle powersavings sched_domain. The idea is to try and avoid
3168 * completely idle packages/cores just for the purpose of idle load balancing
3169 * when there are other idle cpu's which are better suited for that job.
3171 static int find_new_ilb(int cpu
)
3173 struct sched_domain
*sd
;
3174 struct sched_group
*ilb_group
;
3177 * Have idle load balancer selection from semi-idle packages only
3178 * when power-aware load balancing is enabled
3180 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
3184 * Optimize for the case when we have no idle CPUs or only one
3185 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3187 if (cpumask_weight(nohz
.cpu_mask
) < 2)
3190 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
3191 ilb_group
= sd
->groups
;
3194 if (is_semi_idle_group(ilb_group
))
3195 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
3197 ilb_group
= ilb_group
->next
;
3199 } while (ilb_group
!= sd
->groups
);
3203 return cpumask_first(nohz
.cpu_mask
);
3205 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3206 static inline int find_new_ilb(int call_cpu
)
3208 return cpumask_first(nohz
.cpu_mask
);
3213 * This routine will try to nominate the ilb (idle load balancing)
3214 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3215 * load balancing on behalf of all those cpus. If all the cpus in the system
3216 * go into this tickless mode, then there will be no ilb owner (as there is
3217 * no need for one) and all the cpus will sleep till the next wakeup event
3220 * For the ilb owner, tick is not stopped. And this tick will be used
3221 * for idle load balancing. ilb owner will still be part of
3224 * While stopping the tick, this cpu will become the ilb owner if there
3225 * is no other owner. And will be the owner till that cpu becomes busy
3226 * or if all cpus in the system stop their ticks at which point
3227 * there is no need for ilb owner.
3229 * When the ilb owner becomes busy, it nominates another owner, during the
3230 * next busy scheduler_tick()
3232 int select_nohz_load_balancer(int stop_tick
)
3234 int cpu
= smp_processor_id();
3237 cpu_rq(cpu
)->in_nohz_recently
= 1;
3239 if (!cpu_active(cpu
)) {
3240 if (atomic_read(&nohz
.load_balancer
) != cpu
)
3244 * If we are going offline and still the leader,
3247 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3253 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
3255 /* time for ilb owner also to sleep */
3256 if (cpumask_weight(nohz
.cpu_mask
) == num_active_cpus()) {
3257 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3258 atomic_set(&nohz
.load_balancer
, -1);
3262 if (atomic_read(&nohz
.load_balancer
) == -1) {
3263 /* make me the ilb owner */
3264 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3266 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3269 if (!(sched_smt_power_savings
||
3270 sched_mc_power_savings
))
3273 * Check to see if there is a more power-efficient
3276 new_ilb
= find_new_ilb(cpu
);
3277 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
3278 atomic_set(&nohz
.load_balancer
, -1);
3279 resched_cpu(new_ilb
);
3285 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
3288 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
3290 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3291 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3298 static DEFINE_SPINLOCK(balancing
);
3301 * It checks each scheduling domain to see if it is due to be balanced,
3302 * and initiates a balancing operation if so.
3304 * Balancing parameters are set up in arch_init_sched_domains.
3306 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3309 struct rq
*rq
= cpu_rq(cpu
);
3310 unsigned long interval
;
3311 struct sched_domain
*sd
;
3312 /* Earliest time when we have to do rebalance again */
3313 unsigned long next_balance
= jiffies
+ 60*HZ
;
3314 int update_next_balance
= 0;
3317 for_each_domain(cpu
, sd
) {
3318 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3321 interval
= sd
->balance_interval
;
3322 if (idle
!= CPU_IDLE
)
3323 interval
*= sd
->busy_factor
;
3325 /* scale ms to jiffies */
3326 interval
= msecs_to_jiffies(interval
);
3327 if (unlikely(!interval
))
3329 if (interval
> HZ
*NR_CPUS
/10)
3330 interval
= HZ
*NR_CPUS
/10;
3332 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3334 if (need_serialize
) {
3335 if (!spin_trylock(&balancing
))
3339 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3340 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3342 * We've pulled tasks over so either we're no
3343 * longer idle, or one of our SMT siblings is
3346 idle
= CPU_NOT_IDLE
;
3348 sd
->last_balance
= jiffies
;
3351 spin_unlock(&balancing
);
3353 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3354 next_balance
= sd
->last_balance
+ interval
;
3355 update_next_balance
= 1;
3359 * Stop the load balance at this level. There is another
3360 * CPU in our sched group which is doing load balancing more
3368 * next_balance will be updated only when there is a need.
3369 * When the cpu is attached to null domain for ex, it will not be
3372 if (likely(update_next_balance
))
3373 rq
->next_balance
= next_balance
;
3377 * run_rebalance_domains is triggered when needed from the scheduler tick.
3378 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3379 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3381 static void run_rebalance_domains(struct softirq_action
*h
)
3383 int this_cpu
= smp_processor_id();
3384 struct rq
*this_rq
= cpu_rq(this_cpu
);
3385 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3386 CPU_IDLE
: CPU_NOT_IDLE
;
3388 rebalance_domains(this_cpu
, idle
);
3392 * If this cpu is the owner for idle load balancing, then do the
3393 * balancing on behalf of the other idle cpus whose ticks are
3396 if (this_rq
->idle_at_tick
&&
3397 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3401 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
3402 if (balance_cpu
== this_cpu
)
3406 * If this cpu gets work to do, stop the load balancing
3407 * work being done for other cpus. Next load
3408 * balancing owner will pick it up.
3413 rebalance_domains(balance_cpu
, CPU_IDLE
);
3415 rq
= cpu_rq(balance_cpu
);
3416 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3417 this_rq
->next_balance
= rq
->next_balance
;
3423 static inline int on_null_domain(int cpu
)
3425 return !rcu_dereference(cpu_rq(cpu
)->sd
);
3429 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3431 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3432 * idle load balancing owner or decide to stop the periodic load balancing,
3433 * if the whole system is idle.
3435 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3439 * If we were in the nohz mode recently and busy at the current
3440 * scheduler tick, then check if we need to nominate new idle
3443 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3444 rq
->in_nohz_recently
= 0;
3446 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3447 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
3448 atomic_set(&nohz
.load_balancer
, -1);
3451 if (atomic_read(&nohz
.load_balancer
) == -1) {
3452 int ilb
= find_new_ilb(cpu
);
3454 if (ilb
< nr_cpu_ids
)
3460 * If this cpu is idle and doing idle load balancing for all the
3461 * cpus with ticks stopped, is it time for that to stop?
3463 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3464 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3470 * If this cpu is idle and the idle load balancing is done by
3471 * someone else, then no need raise the SCHED_SOFTIRQ
3473 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3474 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
3477 /* Don't need to rebalance while attached to NULL domain */
3478 if (time_after_eq(jiffies
, rq
->next_balance
) &&
3479 likely(!on_null_domain(cpu
)))
3480 raise_softirq(SCHED_SOFTIRQ
);
3483 static void rq_online_fair(struct rq
*rq
)
3488 static void rq_offline_fair(struct rq
*rq
)
3493 #else /* CONFIG_SMP */
3496 * on UP we do not need to balance between CPUs:
3498 static inline void idle_balance(int cpu
, struct rq
*rq
)
3502 #endif /* CONFIG_SMP */
3505 * scheduler tick hitting a task of our scheduling class:
3507 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
3509 struct cfs_rq
*cfs_rq
;
3510 struct sched_entity
*se
= &curr
->se
;
3512 for_each_sched_entity(se
) {
3513 cfs_rq
= cfs_rq_of(se
);
3514 entity_tick(cfs_rq
, se
, queued
);
3519 * called on fork with the child task as argument from the parent's context
3520 * - child not yet on the tasklist
3521 * - preemption disabled
3523 static void task_fork_fair(struct task_struct
*p
)
3525 struct cfs_rq
*cfs_rq
= task_cfs_rq(current
);
3526 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
3527 int this_cpu
= smp_processor_id();
3528 struct rq
*rq
= this_rq();
3529 unsigned long flags
;
3531 raw_spin_lock_irqsave(&rq
->lock
, flags
);
3533 if (unlikely(task_cpu(p
) != this_cpu
))
3534 __set_task_cpu(p
, this_cpu
);
3536 update_curr(cfs_rq
);
3539 se
->vruntime
= curr
->vruntime
;
3540 place_entity(cfs_rq
, se
, 1);
3542 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
3544 * Upon rescheduling, sched_class::put_prev_task() will place
3545 * 'current' within the tree based on its new key value.
3547 swap(curr
->vruntime
, se
->vruntime
);
3548 resched_task(rq
->curr
);
3551 se
->vruntime
-= cfs_rq
->min_vruntime
;
3553 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
3557 * Priority of the task has changed. Check to see if we preempt
3560 static void prio_changed_fair(struct rq
*rq
, struct task_struct
*p
,
3561 int oldprio
, int running
)
3564 * Reschedule if we are currently running on this runqueue and
3565 * our priority decreased, or if we are not currently running on
3566 * this runqueue and our priority is higher than the current's
3569 if (p
->prio
> oldprio
)
3570 resched_task(rq
->curr
);
3572 check_preempt_curr(rq
, p
, 0);
3576 * We switched to the sched_fair class.
3578 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
,
3582 * We were most likely switched from sched_rt, so
3583 * kick off the schedule if running, otherwise just see
3584 * if we can still preempt the current task.
3587 resched_task(rq
->curr
);
3589 check_preempt_curr(rq
, p
, 0);
3592 /* Account for a task changing its policy or group.
3594 * This routine is mostly called to set cfs_rq->curr field when a task
3595 * migrates between groups/classes.
3597 static void set_curr_task_fair(struct rq
*rq
)
3599 struct sched_entity
*se
= &rq
->curr
->se
;
3601 for_each_sched_entity(se
)
3602 set_next_entity(cfs_rq_of(se
), se
);
3605 #ifdef CONFIG_FAIR_GROUP_SCHED
3606 static void moved_group_fair(struct task_struct
*p
, int on_rq
)
3608 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
3610 update_curr(cfs_rq
);
3612 place_entity(cfs_rq
, &p
->se
, 1);
3616 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
3618 struct sched_entity
*se
= &task
->se
;
3619 unsigned int rr_interval
= 0;
3622 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
3625 if (rq
->cfs
.load
.weight
)
3626 rr_interval
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
3632 * All the scheduling class methods:
3634 static const struct sched_class fair_sched_class
= {
3635 .next
= &idle_sched_class
,
3636 .enqueue_task
= enqueue_task_fair
,
3637 .dequeue_task
= dequeue_task_fair
,
3638 .yield_task
= yield_task_fair
,
3640 .check_preempt_curr
= check_preempt_wakeup
,
3642 .pick_next_task
= pick_next_task_fair
,
3643 .put_prev_task
= put_prev_task_fair
,
3646 .select_task_rq
= select_task_rq_fair
,
3648 .rq_online
= rq_online_fair
,
3649 .rq_offline
= rq_offline_fair
,
3651 .task_waking
= task_waking_fair
,
3654 .set_curr_task
= set_curr_task_fair
,
3655 .task_tick
= task_tick_fair
,
3656 .task_fork
= task_fork_fair
,
3658 .prio_changed
= prio_changed_fair
,
3659 .switched_to
= switched_to_fair
,
3661 .get_rr_interval
= get_rr_interval_fair
,
3663 #ifdef CONFIG_FAIR_GROUP_SCHED
3664 .moved_group
= moved_group_fair
,
3668 #ifdef CONFIG_SCHED_DEBUG
3669 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
3671 struct cfs_rq
*cfs_rq
;
3674 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
3675 print_cfs_rq(m
, cpu
, cfs_rq
);