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 /* 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
;
761 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
764 * Update the normalized vruntime before updating min_vruntime
765 * through callig update_curr().
767 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
768 se
->vruntime
+= cfs_rq
->min_vruntime
;
771 * Update run-time statistics of the 'current'.
774 account_entity_enqueue(cfs_rq
, se
);
776 if (flags
& ENQUEUE_WAKEUP
) {
777 place_entity(cfs_rq
, se
, 0);
778 enqueue_sleeper(cfs_rq
, se
);
781 update_stats_enqueue(cfs_rq
, se
);
782 check_spread(cfs_rq
, se
);
783 if (se
!= cfs_rq
->curr
)
784 __enqueue_entity(cfs_rq
, se
);
787 static void __clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
789 if (!se
|| cfs_rq
->last
== se
)
792 if (!se
|| cfs_rq
->next
== se
)
796 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
798 for_each_sched_entity(se
)
799 __clear_buddies(cfs_rq_of(se
), se
);
803 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
806 * Update run-time statistics of the 'current'.
810 update_stats_dequeue(cfs_rq
, se
);
811 if (flags
& DEQUEUE_SLEEP
) {
812 #ifdef CONFIG_SCHEDSTATS
813 if (entity_is_task(se
)) {
814 struct task_struct
*tsk
= task_of(se
);
816 if (tsk
->state
& TASK_INTERRUPTIBLE
)
817 se
->statistics
.sleep_start
= rq_of(cfs_rq
)->clock
;
818 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
819 se
->statistics
.block_start
= rq_of(cfs_rq
)->clock
;
824 clear_buddies(cfs_rq
, se
);
826 if (se
!= cfs_rq
->curr
)
827 __dequeue_entity(cfs_rq
, se
);
828 account_entity_dequeue(cfs_rq
, se
);
829 update_min_vruntime(cfs_rq
);
832 * Normalize the entity after updating the min_vruntime because the
833 * update can refer to the ->curr item and we need to reflect this
834 * movement in our normalized position.
836 if (!(flags
& DEQUEUE_SLEEP
))
837 se
->vruntime
-= cfs_rq
->min_vruntime
;
841 * Preempt the current task with a newly woken task if needed:
844 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
846 unsigned long ideal_runtime
, delta_exec
;
848 ideal_runtime
= sched_slice(cfs_rq
, curr
);
849 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
850 if (delta_exec
> ideal_runtime
) {
851 resched_task(rq_of(cfs_rq
)->curr
);
853 * The current task ran long enough, ensure it doesn't get
854 * re-elected due to buddy favours.
856 clear_buddies(cfs_rq
, curr
);
861 * Ensure that a task that missed wakeup preemption by a
862 * narrow margin doesn't have to wait for a full slice.
863 * This also mitigates buddy induced latencies under load.
865 if (!sched_feat(WAKEUP_PREEMPT
))
868 if (delta_exec
< sysctl_sched_min_granularity
)
871 if (cfs_rq
->nr_running
> 1) {
872 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
873 s64 delta
= curr
->vruntime
- se
->vruntime
;
875 if (delta
> ideal_runtime
)
876 resched_task(rq_of(cfs_rq
)->curr
);
881 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
883 /* 'current' is not kept within the tree. */
886 * Any task has to be enqueued before it get to execute on
887 * a CPU. So account for the time it spent waiting on the
890 update_stats_wait_end(cfs_rq
, se
);
891 __dequeue_entity(cfs_rq
, se
);
894 update_stats_curr_start(cfs_rq
, se
);
896 #ifdef CONFIG_SCHEDSTATS
898 * Track our maximum slice length, if the CPU's load is at
899 * least twice that of our own weight (i.e. dont track it
900 * when there are only lesser-weight tasks around):
902 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
903 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
904 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
907 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
911 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
913 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
915 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
916 struct sched_entity
*left
= se
;
918 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
922 * Prefer last buddy, try to return the CPU to a preempted task.
924 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
927 clear_buddies(cfs_rq
, se
);
932 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
935 * If still on the runqueue then deactivate_task()
936 * was not called and update_curr() has to be done:
941 check_spread(cfs_rq
, prev
);
943 update_stats_wait_start(cfs_rq
, prev
);
944 /* Put 'current' back into the tree. */
945 __enqueue_entity(cfs_rq
, prev
);
951 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
954 * Update run-time statistics of the 'current'.
958 #ifdef CONFIG_SCHED_HRTICK
960 * queued ticks are scheduled to match the slice, so don't bother
961 * validating it and just reschedule.
964 resched_task(rq_of(cfs_rq
)->curr
);
968 * don't let the period tick interfere with the hrtick preemption
970 if (!sched_feat(DOUBLE_TICK
) &&
971 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
975 if (cfs_rq
->nr_running
> 1 || !sched_feat(WAKEUP_PREEMPT
))
976 check_preempt_tick(cfs_rq
, curr
);
979 /**************************************************
980 * CFS operations on tasks:
983 #ifdef CONFIG_SCHED_HRTICK
984 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
986 struct sched_entity
*se
= &p
->se
;
987 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
989 WARN_ON(task_rq(p
) != rq
);
991 if (hrtick_enabled(rq
) && cfs_rq
->nr_running
> 1) {
992 u64 slice
= sched_slice(cfs_rq
, se
);
993 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
994 s64 delta
= slice
- ran
;
1003 * Don't schedule slices shorter than 10000ns, that just
1004 * doesn't make sense. Rely on vruntime for fairness.
1007 delta
= max_t(s64
, 10000LL, delta
);
1009 hrtick_start(rq
, delta
);
1014 * called from enqueue/dequeue and updates the hrtick when the
1015 * current task is from our class and nr_running is low enough
1018 static void hrtick_update(struct rq
*rq
)
1020 struct task_struct
*curr
= rq
->curr
;
1022 if (curr
->sched_class
!= &fair_sched_class
)
1025 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
1026 hrtick_start_fair(rq
, curr
);
1028 #else /* !CONFIG_SCHED_HRTICK */
1030 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
1034 static inline void hrtick_update(struct rq
*rq
)
1040 * The enqueue_task method is called before nr_running is
1041 * increased. Here we update the fair scheduling stats and
1042 * then put the task into the rbtree:
1045 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
1047 struct cfs_rq
*cfs_rq
;
1048 struct sched_entity
*se
= &p
->se
;
1050 for_each_sched_entity(se
) {
1053 cfs_rq
= cfs_rq_of(se
);
1054 enqueue_entity(cfs_rq
, se
, flags
);
1055 flags
= ENQUEUE_WAKEUP
;
1062 * The dequeue_task method is called before nr_running is
1063 * decreased. We remove the task from the rbtree and
1064 * update the fair scheduling stats:
1066 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
1068 struct cfs_rq
*cfs_rq
;
1069 struct sched_entity
*se
= &p
->se
;
1071 for_each_sched_entity(se
) {
1072 cfs_rq
= cfs_rq_of(se
);
1073 dequeue_entity(cfs_rq
, se
, flags
);
1074 /* Don't dequeue parent if it has other entities besides us */
1075 if (cfs_rq
->load
.weight
)
1077 flags
|= DEQUEUE_SLEEP
;
1084 * sched_yield() support is very simple - we dequeue and enqueue.
1086 * If compat_yield is turned on then we requeue to the end of the tree.
1088 static void yield_task_fair(struct rq
*rq
)
1090 struct task_struct
*curr
= rq
->curr
;
1091 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
1092 struct sched_entity
*rightmost
, *se
= &curr
->se
;
1095 * Are we the only task in the tree?
1097 if (unlikely(cfs_rq
->nr_running
== 1))
1100 clear_buddies(cfs_rq
, se
);
1102 if (likely(!sysctl_sched_compat_yield
) && curr
->policy
!= SCHED_BATCH
) {
1103 update_rq_clock(rq
);
1105 * Update run-time statistics of the 'current'.
1107 update_curr(cfs_rq
);
1112 * Find the rightmost entry in the rbtree:
1114 rightmost
= __pick_last_entity(cfs_rq
);
1116 * Already in the rightmost position?
1118 if (unlikely(!rightmost
|| entity_before(rightmost
, se
)))
1122 * Minimally necessary key value to be last in the tree:
1123 * Upon rescheduling, sched_class::put_prev_task() will place
1124 * 'current' within the tree based on its new key value.
1126 se
->vruntime
= rightmost
->vruntime
+ 1;
1131 static void task_waking_fair(struct rq
*rq
, struct task_struct
*p
)
1133 struct sched_entity
*se
= &p
->se
;
1134 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1136 se
->vruntime
-= cfs_rq
->min_vruntime
;
1139 #ifdef CONFIG_FAIR_GROUP_SCHED
1141 * effective_load() calculates the load change as seen from the root_task_group
1143 * Adding load to a group doesn't make a group heavier, but can cause movement
1144 * of group shares between cpus. Assuming the shares were perfectly aligned one
1145 * can calculate the shift in shares.
1147 * The problem is that perfectly aligning the shares is rather expensive, hence
1148 * we try to avoid doing that too often - see update_shares(), which ratelimits
1151 * We compensate this by not only taking the current delta into account, but
1152 * also considering the delta between when the shares were last adjusted and
1155 * We still saw a performance dip, some tracing learned us that between
1156 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1157 * significantly. Therefore try to bias the error in direction of failing
1158 * the affine wakeup.
1161 static long effective_load(struct task_group
*tg
, int cpu
,
1164 struct sched_entity
*se
= tg
->se
[cpu
];
1170 * By not taking the decrease of shares on the other cpu into
1171 * account our error leans towards reducing the affine wakeups.
1173 if (!wl
&& sched_feat(ASYM_EFF_LOAD
))
1176 for_each_sched_entity(se
) {
1177 long S
, rw
, s
, a
, b
;
1181 * Instead of using this increment, also add the difference
1182 * between when the shares were last updated and now.
1184 more_w
= se
->my_q
->load
.weight
- se
->my_q
->rq_weight
;
1188 S
= se
->my_q
->tg
->shares
;
1189 s
= se
->my_q
->shares
;
1190 rw
= se
->my_q
->rq_weight
;
1201 * Assume the group is already running and will
1202 * thus already be accounted for in the weight.
1204 * That is, moving shares between CPUs, does not
1205 * alter the group weight.
1215 static inline unsigned long effective_load(struct task_group
*tg
, int cpu
,
1216 unsigned long wl
, unsigned long wg
)
1223 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
1225 unsigned long this_load
, load
;
1226 int idx
, this_cpu
, prev_cpu
;
1227 unsigned long tl_per_task
;
1228 struct task_group
*tg
;
1229 unsigned long weight
;
1233 this_cpu
= smp_processor_id();
1234 prev_cpu
= task_cpu(p
);
1235 load
= source_load(prev_cpu
, idx
);
1236 this_load
= target_load(this_cpu
, idx
);
1239 * If sync wakeup then subtract the (maximum possible)
1240 * effect of the currently running task from the load
1241 * of the current CPU:
1245 tg
= task_group(current
);
1246 weight
= current
->se
.load
.weight
;
1248 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
1249 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
1253 weight
= p
->se
.load
.weight
;
1256 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1257 * due to the sync cause above having dropped this_load to 0, we'll
1258 * always have an imbalance, but there's really nothing you can do
1259 * about that, so that's good too.
1261 * Otherwise check if either cpus are near enough in load to allow this
1262 * task to be woken on this_cpu.
1265 unsigned long this_eff_load
, prev_eff_load
;
1267 this_eff_load
= 100;
1268 this_eff_load
*= power_of(prev_cpu
);
1269 this_eff_load
*= this_load
+
1270 effective_load(tg
, this_cpu
, weight
, weight
);
1272 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
1273 prev_eff_load
*= power_of(this_cpu
);
1274 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
1276 balanced
= this_eff_load
<= prev_eff_load
;
1282 * If the currently running task will sleep within
1283 * a reasonable amount of time then attract this newly
1286 if (sync
&& balanced
)
1289 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
1290 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1293 (this_load
<= load
&&
1294 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
1296 * This domain has SD_WAKE_AFFINE and
1297 * p is cache cold in this domain, and
1298 * there is no bad imbalance.
1300 schedstat_inc(sd
, ttwu_move_affine
);
1301 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
1309 * find_idlest_group finds and returns the least busy CPU group within the
1312 static struct sched_group
*
1313 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
1314 int this_cpu
, int load_idx
)
1316 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
1317 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1318 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1321 unsigned long load
, avg_load
;
1325 /* Skip over this group if it has no CPUs allowed */
1326 if (!cpumask_intersects(sched_group_cpus(group
),
1330 local_group
= cpumask_test_cpu(this_cpu
,
1331 sched_group_cpus(group
));
1333 /* Tally up the load of all CPUs in the group */
1336 for_each_cpu(i
, sched_group_cpus(group
)) {
1337 /* Bias balancing toward cpus of our domain */
1339 load
= source_load(i
, load_idx
);
1341 load
= target_load(i
, load_idx
);
1346 /* Adjust by relative CPU power of the group */
1347 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1350 this_load
= avg_load
;
1351 } else if (avg_load
< min_load
) {
1352 min_load
= avg_load
;
1355 } while (group
= group
->next
, group
!= sd
->groups
);
1357 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1363 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1366 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1368 unsigned long load
, min_load
= ULONG_MAX
;
1372 /* Traverse only the allowed CPUs */
1373 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
1374 load
= weighted_cpuload(i
);
1376 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1386 * Try and locate an idle CPU in the sched_domain.
1388 static int select_idle_sibling(struct task_struct
*p
, int target
)
1390 int cpu
= smp_processor_id();
1391 int prev_cpu
= task_cpu(p
);
1392 struct sched_domain
*sd
;
1396 * If the task is going to be woken-up on this cpu and if it is
1397 * already idle, then it is the right target.
1399 if (target
== cpu
&& idle_cpu(cpu
))
1403 * If the task is going to be woken-up on the cpu where it previously
1404 * ran and if it is currently idle, then it the right target.
1406 if (target
== prev_cpu
&& idle_cpu(prev_cpu
))
1410 * Otherwise, iterate the domains and find an elegible idle cpu.
1412 for_each_domain(target
, sd
) {
1413 if (!(sd
->flags
& SD_SHARE_PKG_RESOURCES
))
1416 for_each_cpu_and(i
, sched_domain_span(sd
), &p
->cpus_allowed
) {
1424 * Lets stop looking for an idle sibling when we reached
1425 * the domain that spans the current cpu and prev_cpu.
1427 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
)) &&
1428 cpumask_test_cpu(prev_cpu
, sched_domain_span(sd
)))
1436 * sched_balance_self: balance the current task (running on cpu) in domains
1437 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1440 * Balance, ie. select the least loaded group.
1442 * Returns the target CPU number, or the same CPU if no balancing is needed.
1444 * preempt must be disabled.
1447 select_task_rq_fair(struct rq
*rq
, struct task_struct
*p
, int sd_flag
, int wake_flags
)
1449 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
1450 int cpu
= smp_processor_id();
1451 int prev_cpu
= task_cpu(p
);
1453 int want_affine
= 0;
1455 int sync
= wake_flags
& WF_SYNC
;
1457 if (sd_flag
& SD_BALANCE_WAKE
) {
1458 if (cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
1463 for_each_domain(cpu
, tmp
) {
1464 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
1468 * If power savings logic is enabled for a domain, see if we
1469 * are not overloaded, if so, don't balance wider.
1471 if (tmp
->flags
& (SD_POWERSAVINGS_BALANCE
|SD_PREFER_LOCAL
)) {
1472 unsigned long power
= 0;
1473 unsigned long nr_running
= 0;
1474 unsigned long capacity
;
1477 for_each_cpu(i
, sched_domain_span(tmp
)) {
1478 power
+= power_of(i
);
1479 nr_running
+= cpu_rq(i
)->cfs
.nr_running
;
1482 capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
1484 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1487 if (nr_running
< capacity
)
1492 * If both cpu and prev_cpu are part of this domain,
1493 * cpu is a valid SD_WAKE_AFFINE target.
1495 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
1496 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
1501 if (!want_sd
&& !want_affine
)
1504 if (!(tmp
->flags
& sd_flag
))
1511 #ifdef CONFIG_FAIR_GROUP_SCHED
1512 if (sched_feat(LB_SHARES_UPDATE
)) {
1514 * Pick the largest domain to update shares over
1517 if (affine_sd
&& (!tmp
|| affine_sd
->span_weight
> sd
->span_weight
))
1521 raw_spin_unlock(&rq
->lock
);
1523 raw_spin_lock(&rq
->lock
);
1529 if (cpu
== prev_cpu
|| wake_affine(affine_sd
, p
, sync
))
1530 return select_idle_sibling(p
, cpu
);
1532 return select_idle_sibling(p
, prev_cpu
);
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
= sd
->span_weight
;
1565 for_each_domain(cpu
, tmp
) {
1566 if (weight
<= tmp
->span_weight
)
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 * By using 'se' instead of 'curr' we penalize light tasks, so
1588 * they get preempted easier. That is, if 'se' < 'curr' then
1589 * the resulting gran will be larger, therefore penalizing the
1590 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1591 * be smaller, again penalizing the lighter task.
1593 * This is especially important for buddies when the leftmost
1594 * task is higher priority than the buddy.
1596 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
1597 gran
= calc_delta_fair(gran
, se
);
1603 * Should 'se' preempt 'curr'.
1617 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
1619 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
1624 gran
= wakeup_gran(curr
, se
);
1631 static void set_last_buddy(struct sched_entity
*se
)
1633 if (likely(task_of(se
)->policy
!= SCHED_IDLE
)) {
1634 for_each_sched_entity(se
)
1635 cfs_rq_of(se
)->last
= se
;
1639 static void set_next_buddy(struct sched_entity
*se
)
1641 if (likely(task_of(se
)->policy
!= SCHED_IDLE
)) {
1642 for_each_sched_entity(se
)
1643 cfs_rq_of(se
)->next
= se
;
1648 * Preempt the current task with a newly woken task if needed:
1650 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1652 struct task_struct
*curr
= rq
->curr
;
1653 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
1654 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
1655 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
1657 if (unlikely(rt_prio(p
->prio
)))
1660 if (unlikely(p
->sched_class
!= &fair_sched_class
))
1663 if (unlikely(se
== pse
))
1666 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
))
1667 set_next_buddy(pse
);
1670 * We can come here with TIF_NEED_RESCHED already set from new task
1673 if (test_tsk_need_resched(curr
))
1677 * Batch and idle tasks do not preempt (their preemption is driven by
1680 if (unlikely(p
->policy
!= SCHED_NORMAL
))
1683 /* Idle tasks are by definition preempted by everybody. */
1684 if (unlikely(curr
->policy
== SCHED_IDLE
))
1687 if (!sched_feat(WAKEUP_PREEMPT
))
1690 update_curr(cfs_rq
);
1691 find_matching_se(&se
, &pse
);
1693 if (wakeup_preempt_entity(se
, pse
) == 1)
1701 * Only set the backward buddy when the current task is still
1702 * on the rq. This can happen when a wakeup gets interleaved
1703 * with schedule on the ->pre_schedule() or idle_balance()
1704 * point, either of which can * drop the rq lock.
1706 * Also, during early boot the idle thread is in the fair class,
1707 * for obvious reasons its a bad idea to schedule back to it.
1709 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
1712 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
1716 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
1718 struct task_struct
*p
;
1719 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
1720 struct sched_entity
*se
;
1722 if (!cfs_rq
->nr_running
)
1726 se
= pick_next_entity(cfs_rq
);
1727 set_next_entity(cfs_rq
, se
);
1728 cfs_rq
= group_cfs_rq(se
);
1732 hrtick_start_fair(rq
, p
);
1738 * Account for a descheduled task:
1740 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
1742 struct sched_entity
*se
= &prev
->se
;
1743 struct cfs_rq
*cfs_rq
;
1745 for_each_sched_entity(se
) {
1746 cfs_rq
= cfs_rq_of(se
);
1747 put_prev_entity(cfs_rq
, se
);
1752 /**************************************************
1753 * Fair scheduling class load-balancing methods:
1757 * pull_task - move a task from a remote runqueue to the local runqueue.
1758 * Both runqueues must be locked.
1760 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
1761 struct rq
*this_rq
, int this_cpu
)
1763 deactivate_task(src_rq
, p
, 0);
1764 set_task_cpu(p
, this_cpu
);
1765 activate_task(this_rq
, p
, 0);
1766 check_preempt_curr(this_rq
, p
, 0);
1768 /* re-arm NEWIDLE balancing when moving tasks */
1769 src_rq
->avg_idle
= this_rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
1770 this_rq
->idle_stamp
= 0;
1774 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1777 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
1778 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1781 int tsk_cache_hot
= 0;
1783 * We do not migrate tasks that are:
1784 * 1) running (obviously), or
1785 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1786 * 3) are cache-hot on their current CPU.
1788 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
1789 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
1794 if (task_running(rq
, p
)) {
1795 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
1800 * Aggressive migration if:
1801 * 1) task is cache cold, or
1802 * 2) too many balance attempts have failed.
1805 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
1806 if (!tsk_cache_hot
||
1807 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
1808 #ifdef CONFIG_SCHEDSTATS
1809 if (tsk_cache_hot
) {
1810 schedstat_inc(sd
, lb_hot_gained
[idle
]);
1811 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
1817 if (tsk_cache_hot
) {
1818 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
1825 * move_one_task tries to move exactly one task from busiest to this_rq, as
1826 * part of active balancing operations within "domain".
1827 * Returns 1 if successful and 0 otherwise.
1829 * Called with both runqueues locked.
1832 move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1833 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1835 struct task_struct
*p
, *n
;
1836 struct cfs_rq
*cfs_rq
;
1839 for_each_leaf_cfs_rq(busiest
, cfs_rq
) {
1840 list_for_each_entry_safe(p
, n
, &cfs_rq
->tasks
, se
.group_node
) {
1842 if (!can_migrate_task(p
, busiest
, this_cpu
,
1846 pull_task(busiest
, p
, this_rq
, this_cpu
);
1848 * Right now, this is only the second place pull_task()
1849 * is called, so we can safely collect pull_task()
1850 * stats here rather than inside pull_task().
1852 schedstat_inc(sd
, lb_gained
[idle
]);
1860 static unsigned long
1861 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1862 unsigned long max_load_move
, struct sched_domain
*sd
,
1863 enum cpu_idle_type idle
, int *all_pinned
,
1864 int *this_best_prio
, struct cfs_rq
*busiest_cfs_rq
)
1866 int loops
= 0, pulled
= 0, pinned
= 0;
1867 long rem_load_move
= max_load_move
;
1868 struct task_struct
*p
, *n
;
1870 if (max_load_move
== 0)
1875 list_for_each_entry_safe(p
, n
, &busiest_cfs_rq
->tasks
, se
.group_node
) {
1876 if (loops
++ > sysctl_sched_nr_migrate
)
1879 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
1880 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
))
1883 pull_task(busiest
, p
, this_rq
, this_cpu
);
1885 rem_load_move
-= p
->se
.load
.weight
;
1887 #ifdef CONFIG_PREEMPT
1889 * NEWIDLE balancing is a source of latency, so preemptible
1890 * kernels will stop after the first task is pulled to minimize
1891 * the critical section.
1893 if (idle
== CPU_NEWLY_IDLE
)
1898 * We only want to steal up to the prescribed amount of
1901 if (rem_load_move
<= 0)
1904 if (p
->prio
< *this_best_prio
)
1905 *this_best_prio
= p
->prio
;
1909 * Right now, this is one of only two places pull_task() is called,
1910 * so we can safely collect pull_task() stats here rather than
1911 * inside pull_task().
1913 schedstat_add(sd
, lb_gained
[idle
], pulled
);
1916 *all_pinned
= pinned
;
1918 return max_load_move
- rem_load_move
;
1921 #ifdef CONFIG_FAIR_GROUP_SCHED
1922 static unsigned long
1923 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1924 unsigned long max_load_move
,
1925 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1926 int *all_pinned
, int *this_best_prio
)
1928 long rem_load_move
= max_load_move
;
1929 int busiest_cpu
= cpu_of(busiest
);
1930 struct task_group
*tg
;
1933 update_h_load(busiest_cpu
);
1935 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
1936 struct cfs_rq
*busiest_cfs_rq
= tg
->cfs_rq
[busiest_cpu
];
1937 unsigned long busiest_h_load
= busiest_cfs_rq
->h_load
;
1938 unsigned long busiest_weight
= busiest_cfs_rq
->load
.weight
;
1939 u64 rem_load
, moved_load
;
1944 if (!busiest_cfs_rq
->task_weight
)
1947 rem_load
= (u64
)rem_load_move
* busiest_weight
;
1948 rem_load
= div_u64(rem_load
, busiest_h_load
+ 1);
1950 moved_load
= balance_tasks(this_rq
, this_cpu
, busiest
,
1951 rem_load
, sd
, idle
, all_pinned
, this_best_prio
,
1957 moved_load
*= busiest_h_load
;
1958 moved_load
= div_u64(moved_load
, busiest_weight
+ 1);
1960 rem_load_move
-= moved_load
;
1961 if (rem_load_move
< 0)
1966 return max_load_move
- rem_load_move
;
1969 static unsigned long
1970 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1971 unsigned long max_load_move
,
1972 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1973 int *all_pinned
, int *this_best_prio
)
1975 return balance_tasks(this_rq
, this_cpu
, busiest
,
1976 max_load_move
, sd
, idle
, all_pinned
,
1977 this_best_prio
, &busiest
->cfs
);
1982 * move_tasks tries to move up to max_load_move weighted load from busiest to
1983 * this_rq, as part of a balancing operation within domain "sd".
1984 * Returns 1 if successful and 0 otherwise.
1986 * Called with both runqueues locked.
1988 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1989 unsigned long max_load_move
,
1990 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1993 unsigned long total_load_moved
= 0, load_moved
;
1994 int this_best_prio
= this_rq
->curr
->prio
;
1997 load_moved
= load_balance_fair(this_rq
, this_cpu
, busiest
,
1998 max_load_move
- total_load_moved
,
1999 sd
, idle
, all_pinned
, &this_best_prio
);
2001 total_load_moved
+= load_moved
;
2003 #ifdef CONFIG_PREEMPT
2005 * NEWIDLE balancing is a source of latency, so preemptible
2006 * kernels will stop after the first task is pulled to minimize
2007 * the critical section.
2009 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
2012 if (raw_spin_is_contended(&this_rq
->lock
) ||
2013 raw_spin_is_contended(&busiest
->lock
))
2016 } while (load_moved
&& max_load_move
> total_load_moved
);
2018 return total_load_moved
> 0;
2021 /********** Helpers for find_busiest_group ************************/
2023 * sd_lb_stats - Structure to store the statistics of a sched_domain
2024 * during load balancing.
2026 struct sd_lb_stats
{
2027 struct sched_group
*busiest
; /* Busiest group in this sd */
2028 struct sched_group
*this; /* Local group in this sd */
2029 unsigned long total_load
; /* Total load of all groups in sd */
2030 unsigned long total_pwr
; /* Total power of all groups in sd */
2031 unsigned long avg_load
; /* Average load across all groups in sd */
2033 /** Statistics of this group */
2034 unsigned long this_load
;
2035 unsigned long this_load_per_task
;
2036 unsigned long this_nr_running
;
2037 unsigned long this_has_capacity
;
2039 /* Statistics of the busiest group */
2040 unsigned long max_load
;
2041 unsigned long busiest_load_per_task
;
2042 unsigned long busiest_nr_running
;
2043 unsigned long busiest_group_capacity
;
2044 unsigned long busiest_has_capacity
;
2046 int group_imb
; /* Is there imbalance in this sd */
2047 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2048 int power_savings_balance
; /* Is powersave balance needed for this sd */
2049 struct sched_group
*group_min
; /* Least loaded group in sd */
2050 struct sched_group
*group_leader
; /* Group which relieves group_min */
2051 unsigned long min_load_per_task
; /* load_per_task in group_min */
2052 unsigned long leader_nr_running
; /* Nr running of group_leader */
2053 unsigned long min_nr_running
; /* Nr running of group_min */
2058 * sg_lb_stats - stats of a sched_group required for load_balancing
2060 struct sg_lb_stats
{
2061 unsigned long avg_load
; /*Avg load across the CPUs of the group */
2062 unsigned long group_load
; /* Total load over the CPUs of the group */
2063 unsigned long sum_nr_running
; /* Nr tasks running in the group */
2064 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
2065 unsigned long group_capacity
;
2066 int group_imb
; /* Is there an imbalance in the group ? */
2067 int group_has_capacity
; /* Is there extra capacity in the group? */
2071 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2072 * @group: The group whose first cpu is to be returned.
2074 static inline unsigned int group_first_cpu(struct sched_group
*group
)
2076 return cpumask_first(sched_group_cpus(group
));
2080 * get_sd_load_idx - Obtain the load index for a given sched domain.
2081 * @sd: The sched_domain whose load_idx is to be obtained.
2082 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2084 static inline int get_sd_load_idx(struct sched_domain
*sd
,
2085 enum cpu_idle_type idle
)
2091 load_idx
= sd
->busy_idx
;
2094 case CPU_NEWLY_IDLE
:
2095 load_idx
= sd
->newidle_idx
;
2098 load_idx
= sd
->idle_idx
;
2106 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2108 * init_sd_power_savings_stats - Initialize power savings statistics for
2109 * the given sched_domain, during load balancing.
2111 * @sd: Sched domain whose power-savings statistics are to be initialized.
2112 * @sds: Variable containing the statistics for sd.
2113 * @idle: Idle status of the CPU at which we're performing load-balancing.
2115 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
2116 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
2119 * Busy processors will not participate in power savings
2122 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2123 sds
->power_savings_balance
= 0;
2125 sds
->power_savings_balance
= 1;
2126 sds
->min_nr_running
= ULONG_MAX
;
2127 sds
->leader_nr_running
= 0;
2132 * update_sd_power_savings_stats - Update the power saving stats for a
2133 * sched_domain while performing load balancing.
2135 * @group: sched_group belonging to the sched_domain under consideration.
2136 * @sds: Variable containing the statistics of the sched_domain
2137 * @local_group: Does group contain the CPU for which we're performing
2139 * @sgs: Variable containing the statistics of the group.
2141 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
2142 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
2145 if (!sds
->power_savings_balance
)
2149 * If the local group is idle or completely loaded
2150 * no need to do power savings balance at this domain
2152 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
2153 !sds
->this_nr_running
))
2154 sds
->power_savings_balance
= 0;
2157 * If a group is already running at full capacity or idle,
2158 * don't include that group in power savings calculations
2160 if (!sds
->power_savings_balance
||
2161 sgs
->sum_nr_running
>= sgs
->group_capacity
||
2162 !sgs
->sum_nr_running
)
2166 * Calculate the group which has the least non-idle load.
2167 * This is the group from where we need to pick up the load
2170 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
2171 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
2172 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
2173 sds
->group_min
= group
;
2174 sds
->min_nr_running
= sgs
->sum_nr_running
;
2175 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
2176 sgs
->sum_nr_running
;
2180 * Calculate the group which is almost near its
2181 * capacity but still has some space to pick up some load
2182 * from other group and save more power
2184 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
2187 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
2188 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
2189 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
2190 sds
->group_leader
= group
;
2191 sds
->leader_nr_running
= sgs
->sum_nr_running
;
2196 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2197 * @sds: Variable containing the statistics of the sched_domain
2198 * under consideration.
2199 * @this_cpu: Cpu at which we're currently performing load-balancing.
2200 * @imbalance: Variable to store the imbalance.
2203 * Check if we have potential to perform some power-savings balance.
2204 * If yes, set the busiest group to be the least loaded group in the
2205 * sched_domain, so that it's CPUs can be put to idle.
2207 * Returns 1 if there is potential to perform power-savings balance.
2210 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
2211 int this_cpu
, unsigned long *imbalance
)
2213 if (!sds
->power_savings_balance
)
2216 if (sds
->this != sds
->group_leader
||
2217 sds
->group_leader
== sds
->group_min
)
2220 *imbalance
= sds
->min_load_per_task
;
2221 sds
->busiest
= sds
->group_min
;
2226 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2227 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
2228 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
2233 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
2234 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
2239 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
2240 int this_cpu
, unsigned long *imbalance
)
2244 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2247 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
2249 return SCHED_LOAD_SCALE
;
2252 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
2254 return default_scale_freq_power(sd
, cpu
);
2257 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
2259 unsigned long weight
= sd
->span_weight
;
2260 unsigned long smt_gain
= sd
->smt_gain
;
2267 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
2269 return default_scale_smt_power(sd
, cpu
);
2272 unsigned long scale_rt_power(int cpu
)
2274 struct rq
*rq
= cpu_rq(cpu
);
2275 u64 total
, available
;
2277 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
2278 available
= total
- rq
->rt_avg
;
2280 if (unlikely((s64
)total
< SCHED_LOAD_SCALE
))
2281 total
= SCHED_LOAD_SCALE
;
2283 total
>>= SCHED_LOAD_SHIFT
;
2285 return div_u64(available
, total
);
2288 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
2290 unsigned long weight
= sd
->span_weight
;
2291 unsigned long power
= SCHED_LOAD_SCALE
;
2292 struct sched_group
*sdg
= sd
->groups
;
2294 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
2295 if (sched_feat(ARCH_POWER
))
2296 power
*= arch_scale_smt_power(sd
, cpu
);
2298 power
*= default_scale_smt_power(sd
, cpu
);
2300 power
>>= SCHED_LOAD_SHIFT
;
2303 sdg
->cpu_power_orig
= power
;
2305 if (sched_feat(ARCH_POWER
))
2306 power
*= arch_scale_freq_power(sd
, cpu
);
2308 power
*= default_scale_freq_power(sd
, cpu
);
2310 power
>>= SCHED_LOAD_SHIFT
;
2312 power
*= scale_rt_power(cpu
);
2313 power
>>= SCHED_LOAD_SHIFT
;
2318 cpu_rq(cpu
)->cpu_power
= power
;
2319 sdg
->cpu_power
= power
;
2322 static void update_group_power(struct sched_domain
*sd
, int cpu
)
2324 struct sched_domain
*child
= sd
->child
;
2325 struct sched_group
*group
, *sdg
= sd
->groups
;
2326 unsigned long power
;
2329 update_cpu_power(sd
, cpu
);
2335 group
= child
->groups
;
2337 power
+= group
->cpu_power
;
2338 group
= group
->next
;
2339 } while (group
!= child
->groups
);
2341 sdg
->cpu_power
= power
;
2345 * Try and fix up capacity for tiny siblings, this is needed when
2346 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2347 * which on its own isn't powerful enough.
2349 * See update_sd_pick_busiest() and check_asym_packing().
2352 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
2355 * Only siblings can have significantly less than SCHED_LOAD_SCALE
2357 if (sd
->level
!= SD_LV_SIBLING
)
2361 * If ~90% of the cpu_power is still there, we're good.
2363 if (group
->cpu_power
* 32 > group
->cpu_power_orig
* 29)
2370 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2371 * @sd: The sched_domain whose statistics are to be updated.
2372 * @group: sched_group whose statistics are to be updated.
2373 * @this_cpu: Cpu for which load balance is currently performed.
2374 * @idle: Idle status of this_cpu
2375 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2376 * @sd_idle: Idle status of the sched_domain containing group.
2377 * @local_group: Does group contain this_cpu.
2378 * @cpus: Set of cpus considered for load balancing.
2379 * @balance: Should we balance.
2380 * @sgs: variable to hold the statistics for this group.
2382 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
2383 struct sched_group
*group
, int this_cpu
,
2384 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
2385 int local_group
, const struct cpumask
*cpus
,
2386 int *balance
, struct sg_lb_stats
*sgs
)
2388 unsigned long load
, max_cpu_load
, min_cpu_load
, max_nr_running
;
2390 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2391 unsigned long avg_load_per_task
= 0;
2394 balance_cpu
= group_first_cpu(group
);
2396 /* Tally up the load of all CPUs in the group */
2398 min_cpu_load
= ~0UL;
2401 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
2402 struct rq
*rq
= cpu_rq(i
);
2404 if (*sd_idle
&& rq
->nr_running
)
2407 /* Bias balancing toward cpus of our domain */
2409 if (idle_cpu(i
) && !first_idle_cpu
) {
2414 load
= target_load(i
, load_idx
);
2416 load
= source_load(i
, load_idx
);
2417 if (load
> max_cpu_load
) {
2418 max_cpu_load
= load
;
2419 max_nr_running
= rq
->nr_running
;
2421 if (min_cpu_load
> load
)
2422 min_cpu_load
= load
;
2425 sgs
->group_load
+= load
;
2426 sgs
->sum_nr_running
+= rq
->nr_running
;
2427 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
2432 * First idle cpu or the first cpu(busiest) in this sched group
2433 * is eligible for doing load balancing at this and above
2434 * domains. In the newly idle case, we will allow all the cpu's
2435 * to do the newly idle load balance.
2437 if (idle
!= CPU_NEWLY_IDLE
&& local_group
) {
2438 if (balance_cpu
!= this_cpu
) {
2442 update_group_power(sd
, this_cpu
);
2445 /* Adjust by relative CPU power of the group */
2446 sgs
->avg_load
= (sgs
->group_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2449 * Consider the group unbalanced when the imbalance is larger
2450 * than the average weight of two tasks.
2452 * APZ: with cgroup the avg task weight can vary wildly and
2453 * might not be a suitable number - should we keep a
2454 * normalized nr_running number somewhere that negates
2457 if (sgs
->sum_nr_running
)
2458 avg_load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
2460 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
&& max_nr_running
> 1)
2463 sgs
->group_capacity
= DIV_ROUND_CLOSEST(group
->cpu_power
, SCHED_LOAD_SCALE
);
2464 if (!sgs
->group_capacity
)
2465 sgs
->group_capacity
= fix_small_capacity(sd
, group
);
2467 if (sgs
->group_capacity
> sgs
->sum_nr_running
)
2468 sgs
->group_has_capacity
= 1;
2472 * update_sd_pick_busiest - return 1 on busiest group
2473 * @sd: sched_domain whose statistics are to be checked
2474 * @sds: sched_domain statistics
2475 * @sg: sched_group candidate to be checked for being the busiest
2476 * @sgs: sched_group statistics
2477 * @this_cpu: the current cpu
2479 * Determine if @sg is a busier group than the previously selected
2482 static bool update_sd_pick_busiest(struct sched_domain
*sd
,
2483 struct sd_lb_stats
*sds
,
2484 struct sched_group
*sg
,
2485 struct sg_lb_stats
*sgs
,
2488 if (sgs
->avg_load
<= sds
->max_load
)
2491 if (sgs
->sum_nr_running
> sgs
->group_capacity
)
2498 * ASYM_PACKING needs to move all the work to the lowest
2499 * numbered CPUs in the group, therefore mark all groups
2500 * higher than ourself as busy.
2502 if ((sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
2503 this_cpu
< group_first_cpu(sg
)) {
2507 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
2515 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2516 * @sd: sched_domain whose statistics are to be updated.
2517 * @this_cpu: Cpu for which load balance is currently performed.
2518 * @idle: Idle status of this_cpu
2519 * @sd_idle: Idle status of the sched_domain containing sg.
2520 * @cpus: Set of cpus considered for load balancing.
2521 * @balance: Should we balance.
2522 * @sds: variable to hold the statistics for this sched_domain.
2524 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
2525 enum cpu_idle_type idle
, int *sd_idle
,
2526 const struct cpumask
*cpus
, int *balance
,
2527 struct sd_lb_stats
*sds
)
2529 struct sched_domain
*child
= sd
->child
;
2530 struct sched_group
*sg
= sd
->groups
;
2531 struct sg_lb_stats sgs
;
2532 int load_idx
, prefer_sibling
= 0;
2534 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
2537 init_sd_power_savings_stats(sd
, sds
, idle
);
2538 load_idx
= get_sd_load_idx(sd
, idle
);
2543 local_group
= cpumask_test_cpu(this_cpu
, sched_group_cpus(sg
));
2544 memset(&sgs
, 0, sizeof(sgs
));
2545 update_sg_lb_stats(sd
, sg
, this_cpu
, idle
, load_idx
, sd_idle
,
2546 local_group
, cpus
, balance
, &sgs
);
2548 if (local_group
&& !(*balance
))
2551 sds
->total_load
+= sgs
.group_load
;
2552 sds
->total_pwr
+= sg
->cpu_power
;
2555 * In case the child domain prefers tasks go to siblings
2556 * first, lower the sg capacity to one so that we'll try
2557 * and move all the excess tasks away. We lower the capacity
2558 * of a group only if the local group has the capacity to fit
2559 * these excess tasks, i.e. nr_running < group_capacity. The
2560 * extra check prevents the case where you always pull from the
2561 * heaviest group when it is already under-utilized (possible
2562 * with a large weight task outweighs the tasks on the system).
2564 if (prefer_sibling
&& !local_group
&& sds
->this_has_capacity
)
2565 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
2568 sds
->this_load
= sgs
.avg_load
;
2570 sds
->this_nr_running
= sgs
.sum_nr_running
;
2571 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
2572 sds
->this_has_capacity
= sgs
.group_has_capacity
;
2573 } else if (update_sd_pick_busiest(sd
, sds
, sg
, &sgs
, this_cpu
)) {
2574 sds
->max_load
= sgs
.avg_load
;
2576 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
2577 sds
->busiest_group_capacity
= sgs
.group_capacity
;
2578 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
2579 sds
->busiest_has_capacity
= sgs
.group_has_capacity
;
2580 sds
->group_imb
= sgs
.group_imb
;
2583 update_sd_power_savings_stats(sg
, sds
, local_group
, &sgs
);
2585 } while (sg
!= sd
->groups
);
2588 int __weak
arch_sd_sibling_asym_packing(void)
2590 return 0*SD_ASYM_PACKING
;
2594 * check_asym_packing - Check to see if the group is packed into the
2597 * This is primarily intended to used at the sibling level. Some
2598 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2599 * case of POWER7, it can move to lower SMT modes only when higher
2600 * threads are idle. When in lower SMT modes, the threads will
2601 * perform better since they share less core resources. Hence when we
2602 * have idle threads, we want them to be the higher ones.
2604 * This packing function is run on idle threads. It checks to see if
2605 * the busiest CPU in this domain (core in the P7 case) has a higher
2606 * CPU number than the packing function is being run on. Here we are
2607 * assuming lower CPU number will be equivalent to lower a SMT thread
2610 * Returns 1 when packing is required and a task should be moved to
2611 * this CPU. The amount of the imbalance is returned in *imbalance.
2613 * @sd: The sched_domain whose packing is to be checked.
2614 * @sds: Statistics of the sched_domain which is to be packed
2615 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2616 * @imbalance: returns amount of imbalanced due to packing.
2618 static int check_asym_packing(struct sched_domain
*sd
,
2619 struct sd_lb_stats
*sds
,
2620 int this_cpu
, unsigned long *imbalance
)
2624 if (!(sd
->flags
& SD_ASYM_PACKING
))
2630 busiest_cpu
= group_first_cpu(sds
->busiest
);
2631 if (this_cpu
> busiest_cpu
)
2634 *imbalance
= DIV_ROUND_CLOSEST(sds
->max_load
* sds
->busiest
->cpu_power
,
2640 * fix_small_imbalance - Calculate the minor imbalance that exists
2641 * amongst the groups of a sched_domain, during
2643 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2644 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2645 * @imbalance: Variable to store the imbalance.
2647 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
2648 int this_cpu
, unsigned long *imbalance
)
2650 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
2651 unsigned int imbn
= 2;
2652 unsigned long scaled_busy_load_per_task
;
2654 if (sds
->this_nr_running
) {
2655 sds
->this_load_per_task
/= sds
->this_nr_running
;
2656 if (sds
->busiest_load_per_task
>
2657 sds
->this_load_per_task
)
2660 sds
->this_load_per_task
=
2661 cpu_avg_load_per_task(this_cpu
);
2663 scaled_busy_load_per_task
= sds
->busiest_load_per_task
2665 scaled_busy_load_per_task
/= sds
->busiest
->cpu_power
;
2667 if (sds
->max_load
- sds
->this_load
+ scaled_busy_load_per_task
>=
2668 (scaled_busy_load_per_task
* imbn
)) {
2669 *imbalance
= sds
->busiest_load_per_task
;
2674 * OK, we don't have enough imbalance to justify moving tasks,
2675 * however we may be able to increase total CPU power used by
2679 pwr_now
+= sds
->busiest
->cpu_power
*
2680 min(sds
->busiest_load_per_task
, sds
->max_load
);
2681 pwr_now
+= sds
->this->cpu_power
*
2682 min(sds
->this_load_per_task
, sds
->this_load
);
2683 pwr_now
/= SCHED_LOAD_SCALE
;
2685 /* Amount of load we'd subtract */
2686 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
2687 sds
->busiest
->cpu_power
;
2688 if (sds
->max_load
> tmp
)
2689 pwr_move
+= sds
->busiest
->cpu_power
*
2690 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
2692 /* Amount of load we'd add */
2693 if (sds
->max_load
* sds
->busiest
->cpu_power
<
2694 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
2695 tmp
= (sds
->max_load
* sds
->busiest
->cpu_power
) /
2696 sds
->this->cpu_power
;
2698 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
2699 sds
->this->cpu_power
;
2700 pwr_move
+= sds
->this->cpu_power
*
2701 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
2702 pwr_move
/= SCHED_LOAD_SCALE
;
2704 /* Move if we gain throughput */
2705 if (pwr_move
> pwr_now
)
2706 *imbalance
= sds
->busiest_load_per_task
;
2710 * calculate_imbalance - Calculate the amount of imbalance present within the
2711 * groups of a given sched_domain during load balance.
2712 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
2713 * @this_cpu: Cpu for which currently load balance is being performed.
2714 * @imbalance: The variable to store the imbalance.
2716 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
2717 unsigned long *imbalance
)
2719 unsigned long max_pull
, load_above_capacity
= ~0UL;
2721 sds
->busiest_load_per_task
/= sds
->busiest_nr_running
;
2722 if (sds
->group_imb
) {
2723 sds
->busiest_load_per_task
=
2724 min(sds
->busiest_load_per_task
, sds
->avg_load
);
2728 * In the presence of smp nice balancing, certain scenarios can have
2729 * max load less than avg load(as we skip the groups at or below
2730 * its cpu_power, while calculating max_load..)
2732 if (sds
->max_load
< sds
->avg_load
) {
2734 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
2737 if (!sds
->group_imb
) {
2739 * Don't want to pull so many tasks that a group would go idle.
2741 load_above_capacity
= (sds
->busiest_nr_running
-
2742 sds
->busiest_group_capacity
);
2744 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_LOAD_SCALE
);
2746 load_above_capacity
/= sds
->busiest
->cpu_power
;
2750 * We're trying to get all the cpus to the average_load, so we don't
2751 * want to push ourselves above the average load, nor do we wish to
2752 * reduce the max loaded cpu below the average load. At the same time,
2753 * we also don't want to reduce the group load below the group capacity
2754 * (so that we can implement power-savings policies etc). Thus we look
2755 * for the minimum possible imbalance.
2756 * Be careful of negative numbers as they'll appear as very large values
2757 * with unsigned longs.
2759 max_pull
= min(sds
->max_load
- sds
->avg_load
, load_above_capacity
);
2761 /* How much load to actually move to equalise the imbalance */
2762 *imbalance
= min(max_pull
* sds
->busiest
->cpu_power
,
2763 (sds
->avg_load
- sds
->this_load
) * sds
->this->cpu_power
)
2767 * if *imbalance is less than the average load per runnable task
2768 * there is no gaurantee that any tasks will be moved so we'll have
2769 * a think about bumping its value to force at least one task to be
2772 if (*imbalance
< sds
->busiest_load_per_task
)
2773 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
2777 /******* find_busiest_group() helpers end here *********************/
2780 * find_busiest_group - Returns the busiest group within the sched_domain
2781 * if there is an imbalance. If there isn't an imbalance, and
2782 * the user has opted for power-savings, it returns a group whose
2783 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
2784 * such a group exists.
2786 * Also calculates the amount of weighted load which should be moved
2787 * to restore balance.
2789 * @sd: The sched_domain whose busiest group is to be returned.
2790 * @this_cpu: The cpu for which load balancing is currently being performed.
2791 * @imbalance: Variable which stores amount of weighted load which should
2792 * be moved to restore balance/put a group to idle.
2793 * @idle: The idle status of this_cpu.
2794 * @sd_idle: The idleness of sd
2795 * @cpus: The set of CPUs under consideration for load-balancing.
2796 * @balance: Pointer to a variable indicating if this_cpu
2797 * is the appropriate cpu to perform load balancing at this_level.
2799 * Returns: - the busiest group if imbalance exists.
2800 * - If no imbalance and user has opted for power-savings balance,
2801 * return the least loaded group whose CPUs can be
2802 * put to idle by rebalancing its tasks onto our group.
2804 static struct sched_group
*
2805 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2806 unsigned long *imbalance
, enum cpu_idle_type idle
,
2807 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
2809 struct sd_lb_stats sds
;
2811 memset(&sds
, 0, sizeof(sds
));
2814 * Compute the various statistics relavent for load balancing at
2817 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
2820 /* Cases where imbalance does not exist from POV of this_cpu */
2821 /* 1) this_cpu is not the appropriate cpu to perform load balancing
2823 * 2) There is no busy sibling group to pull from.
2824 * 3) This group is the busiest group.
2825 * 4) This group is more busy than the avg busieness at this
2827 * 5) The imbalance is within the specified limit.
2829 * Note: when doing newidle balance, if the local group has excess
2830 * capacity (i.e. nr_running < group_capacity) and the busiest group
2831 * does not have any capacity, we force a load balance to pull tasks
2832 * to the local group. In this case, we skip past checks 3, 4 and 5.
2837 if ((idle
== CPU_IDLE
|| idle
== CPU_NEWLY_IDLE
) &&
2838 check_asym_packing(sd
, &sds
, this_cpu
, imbalance
))
2841 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
2844 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
2845 if (idle
== CPU_NEWLY_IDLE
&& sds
.this_has_capacity
&&
2846 !sds
.busiest_has_capacity
)
2849 if (sds
.this_load
>= sds
.max_load
)
2852 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
2854 if (sds
.this_load
>= sds
.avg_load
)
2857 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
2861 /* Looks like there is an imbalance. Compute it */
2862 calculate_imbalance(&sds
, this_cpu
, imbalance
);
2867 * There is no obvious imbalance. But check if we can do some balancing
2870 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
2878 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2881 find_busiest_queue(struct sched_domain
*sd
, struct sched_group
*group
,
2882 enum cpu_idle_type idle
, unsigned long imbalance
,
2883 const struct cpumask
*cpus
)
2885 struct rq
*busiest
= NULL
, *rq
;
2886 unsigned long max_load
= 0;
2889 for_each_cpu(i
, sched_group_cpus(group
)) {
2890 unsigned long power
= power_of(i
);
2891 unsigned long capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
2895 capacity
= fix_small_capacity(sd
, group
);
2897 if (!cpumask_test_cpu(i
, cpus
))
2901 wl
= weighted_cpuload(i
);
2904 * When comparing with imbalance, use weighted_cpuload()
2905 * which is not scaled with the cpu power.
2907 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
2911 * For the load comparisons with the other cpu's, consider
2912 * the weighted_cpuload() scaled with the cpu power, so that
2913 * the load can be moved away from the cpu that is potentially
2914 * running at a lower capacity.
2916 wl
= (wl
* SCHED_LOAD_SCALE
) / power
;
2918 if (wl
> max_load
) {
2928 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2929 * so long as it is large enough.
2931 #define MAX_PINNED_INTERVAL 512
2933 /* Working cpumask for load_balance and load_balance_newidle. */
2934 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
2936 static int need_active_balance(struct sched_domain
*sd
, int sd_idle
, int idle
,
2937 int busiest_cpu
, int this_cpu
)
2939 if (idle
== CPU_NEWLY_IDLE
) {
2942 * ASYM_PACKING needs to force migrate tasks from busy but
2943 * higher numbered CPUs in order to pack all tasks in the
2944 * lowest numbered CPUs.
2946 if ((sd
->flags
& SD_ASYM_PACKING
) && busiest_cpu
> this_cpu
)
2950 * The only task running in a non-idle cpu can be moved to this
2951 * cpu in an attempt to completely freeup the other CPU
2954 * The package power saving logic comes from
2955 * find_busiest_group(). If there are no imbalance, then
2956 * f_b_g() will return NULL. However when sched_mc={1,2} then
2957 * f_b_g() will select a group from which a running task may be
2958 * pulled to this cpu in order to make the other package idle.
2959 * If there is no opportunity to make a package idle and if
2960 * there are no imbalance, then f_b_g() will return NULL and no
2961 * action will be taken in load_balance_newidle().
2963 * Under normal task pull operation due to imbalance, there
2964 * will be more than one task in the source run queue and
2965 * move_tasks() will succeed. ld_moved will be true and this
2966 * active balance code will not be triggered.
2968 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2969 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2972 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
2976 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
2979 static int active_load_balance_cpu_stop(void *data
);
2982 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2983 * tasks if there is an imbalance.
2985 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2986 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2989 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2990 struct sched_group
*group
;
2991 unsigned long imbalance
;
2993 unsigned long flags
;
2994 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
2996 cpumask_copy(cpus
, cpu_active_mask
);
2999 * When power savings policy is enabled for the parent domain, idle
3000 * sibling can pick up load irrespective of busy siblings. In this case,
3001 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3002 * portraying it as CPU_NOT_IDLE.
3004 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3005 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3008 schedstat_inc(sd
, lb_count
[idle
]);
3012 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3019 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3023 busiest
= find_busiest_queue(sd
, group
, idle
, imbalance
, cpus
);
3025 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3029 BUG_ON(busiest
== this_rq
);
3031 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3034 if (busiest
->nr_running
> 1) {
3036 * Attempt to move tasks. If find_busiest_group has found
3037 * an imbalance but busiest->nr_running <= 1, the group is
3038 * still unbalanced. ld_moved simply stays zero, so it is
3039 * correctly treated as an imbalance.
3041 local_irq_save(flags
);
3042 double_rq_lock(this_rq
, busiest
);
3043 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3044 imbalance
, sd
, idle
, &all_pinned
);
3045 double_rq_unlock(this_rq
, busiest
);
3046 local_irq_restore(flags
);
3049 * some other cpu did the load balance for us.
3051 if (ld_moved
&& this_cpu
!= smp_processor_id())
3052 resched_cpu(this_cpu
);
3054 /* All tasks on this runqueue were pinned by CPU affinity */
3055 if (unlikely(all_pinned
)) {
3056 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3057 if (!cpumask_empty(cpus
))
3064 schedstat_inc(sd
, lb_failed
[idle
]);
3066 * Increment the failure counter only on periodic balance.
3067 * We do not want newidle balance, which can be very
3068 * frequent, pollute the failure counter causing
3069 * excessive cache_hot migrations and active balances.
3071 if (idle
!= CPU_NEWLY_IDLE
)
3072 sd
->nr_balance_failed
++;
3074 if (need_active_balance(sd
, sd_idle
, idle
, cpu_of(busiest
),
3076 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
3078 /* don't kick the active_load_balance_cpu_stop,
3079 * if the curr task on busiest cpu can't be
3082 if (!cpumask_test_cpu(this_cpu
,
3083 &busiest
->curr
->cpus_allowed
)) {
3084 raw_spin_unlock_irqrestore(&busiest
->lock
,
3087 goto out_one_pinned
;
3091 * ->active_balance synchronizes accesses to
3092 * ->active_balance_work. Once set, it's cleared
3093 * only after active load balance is finished.
3095 if (!busiest
->active_balance
) {
3096 busiest
->active_balance
= 1;
3097 busiest
->push_cpu
= this_cpu
;
3100 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
3103 stop_one_cpu_nowait(cpu_of(busiest
),
3104 active_load_balance_cpu_stop
, busiest
,
3105 &busiest
->active_balance_work
);
3108 * We've kicked active balancing, reset the failure
3111 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3114 sd
->nr_balance_failed
= 0;
3116 if (likely(!active_balance
)) {
3117 /* We were unbalanced, so reset the balancing interval */
3118 sd
->balance_interval
= sd
->min_interval
;
3121 * If we've begun active balancing, start to back off. This
3122 * case may not be covered by the all_pinned logic if there
3123 * is only 1 task on the busy runqueue (because we don't call
3126 if (sd
->balance_interval
< sd
->max_interval
)
3127 sd
->balance_interval
*= 2;
3130 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3131 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3137 schedstat_inc(sd
, lb_balanced
[idle
]);
3139 sd
->nr_balance_failed
= 0;
3142 /* tune up the balancing interval */
3143 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3144 (sd
->balance_interval
< sd
->max_interval
))
3145 sd
->balance_interval
*= 2;
3147 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3148 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3159 * idle_balance is called by schedule() if this_cpu is about to become
3160 * idle. Attempts to pull tasks from other CPUs.
3162 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3164 struct sched_domain
*sd
;
3165 int pulled_task
= 0;
3166 unsigned long next_balance
= jiffies
+ HZ
;
3168 this_rq
->idle_stamp
= this_rq
->clock
;
3170 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
3174 * Drop the rq->lock, but keep IRQ/preempt disabled.
3176 raw_spin_unlock(&this_rq
->lock
);
3178 for_each_domain(this_cpu
, sd
) {
3179 unsigned long interval
;
3182 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3185 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
3186 /* If we've pulled tasks over stop searching: */
3187 pulled_task
= load_balance(this_cpu
, this_rq
,
3188 sd
, CPU_NEWLY_IDLE
, &balance
);
3191 interval
= msecs_to_jiffies(sd
->balance_interval
);
3192 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3193 next_balance
= sd
->last_balance
+ interval
;
3198 raw_spin_lock(&this_rq
->lock
);
3200 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3202 * We are going idle. next_balance may be set based on
3203 * a busy processor. So reset next_balance.
3205 this_rq
->next_balance
= next_balance
;
3210 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3211 * running tasks off the busiest CPU onto idle CPUs. It requires at
3212 * least 1 task to be running on each physical CPU where possible, and
3213 * avoids physical / logical imbalances.
3215 static int active_load_balance_cpu_stop(void *data
)
3217 struct rq
*busiest_rq
= data
;
3218 int busiest_cpu
= cpu_of(busiest_rq
);
3219 int target_cpu
= busiest_rq
->push_cpu
;
3220 struct rq
*target_rq
= cpu_rq(target_cpu
);
3221 struct sched_domain
*sd
;
3223 raw_spin_lock_irq(&busiest_rq
->lock
);
3225 /* make sure the requested cpu hasn't gone down in the meantime */
3226 if (unlikely(busiest_cpu
!= smp_processor_id() ||
3227 !busiest_rq
->active_balance
))
3230 /* Is there any task to move? */
3231 if (busiest_rq
->nr_running
<= 1)
3235 * This condition is "impossible", if it occurs
3236 * we need to fix it. Originally reported by
3237 * Bjorn Helgaas on a 128-cpu setup.
3239 BUG_ON(busiest_rq
== target_rq
);
3241 /* move a task from busiest_rq to target_rq */
3242 double_lock_balance(busiest_rq
, target_rq
);
3244 /* Search for an sd spanning us and the target CPU. */
3245 for_each_domain(target_cpu
, sd
) {
3246 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3247 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
3252 schedstat_inc(sd
, alb_count
);
3254 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3256 schedstat_inc(sd
, alb_pushed
);
3258 schedstat_inc(sd
, alb_failed
);
3260 double_unlock_balance(busiest_rq
, target_rq
);
3262 busiest_rq
->active_balance
= 0;
3263 raw_spin_unlock_irq(&busiest_rq
->lock
);
3269 static DEFINE_PER_CPU(struct call_single_data
, remote_sched_softirq_cb
);
3271 static void trigger_sched_softirq(void *data
)
3273 raise_softirq_irqoff(SCHED_SOFTIRQ
);
3276 static inline void init_sched_softirq_csd(struct call_single_data
*csd
)
3278 csd
->func
= trigger_sched_softirq
;
3285 * idle load balancing details
3286 * - One of the idle CPUs nominates itself as idle load_balancer, while
3288 * - This idle load balancer CPU will also go into tickless mode when
3289 * it is idle, just like all other idle CPUs
3290 * - When one of the busy CPUs notice that there may be an idle rebalancing
3291 * needed, they will kick the idle load balancer, which then does idle
3292 * load balancing for all the idle CPUs.
3295 atomic_t load_balancer
;
3296 atomic_t first_pick_cpu
;
3297 atomic_t second_pick_cpu
;
3298 cpumask_var_t idle_cpus_mask
;
3299 cpumask_var_t grp_idle_mask
;
3300 unsigned long next_balance
; /* in jiffy units */
3301 } nohz ____cacheline_aligned
;
3303 int get_nohz_load_balancer(void)
3305 return atomic_read(&nohz
.load_balancer
);
3308 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3310 * lowest_flag_domain - Return lowest sched_domain containing flag.
3311 * @cpu: The cpu whose lowest level of sched domain is to
3313 * @flag: The flag to check for the lowest sched_domain
3314 * for the given cpu.
3316 * Returns the lowest sched_domain of a cpu which contains the given flag.
3318 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
3320 struct sched_domain
*sd
;
3322 for_each_domain(cpu
, sd
)
3323 if (sd
&& (sd
->flags
& flag
))
3330 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3331 * @cpu: The cpu whose domains we're iterating over.
3332 * @sd: variable holding the value of the power_savings_sd
3334 * @flag: The flag to filter the sched_domains to be iterated.
3336 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3337 * set, starting from the lowest sched_domain to the highest.
3339 #define for_each_flag_domain(cpu, sd, flag) \
3340 for (sd = lowest_flag_domain(cpu, flag); \
3341 (sd && (sd->flags & flag)); sd = sd->parent)
3344 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3345 * @ilb_group: group to be checked for semi-idleness
3347 * Returns: 1 if the group is semi-idle. 0 otherwise.
3349 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3350 * and atleast one non-idle CPU. This helper function checks if the given
3351 * sched_group is semi-idle or not.
3353 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
3355 cpumask_and(nohz
.grp_idle_mask
, nohz
.idle_cpus_mask
,
3356 sched_group_cpus(ilb_group
));
3359 * A sched_group is semi-idle when it has atleast one busy cpu
3360 * and atleast one idle cpu.
3362 if (cpumask_empty(nohz
.grp_idle_mask
))
3365 if (cpumask_equal(nohz
.grp_idle_mask
, sched_group_cpus(ilb_group
)))
3371 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3372 * @cpu: The cpu which is nominating a new idle_load_balancer.
3374 * Returns: Returns the id of the idle load balancer if it exists,
3375 * Else, returns >= nr_cpu_ids.
3377 * This algorithm picks the idle load balancer such that it belongs to a
3378 * semi-idle powersavings sched_domain. The idea is to try and avoid
3379 * completely idle packages/cores just for the purpose of idle load balancing
3380 * when there are other idle cpu's which are better suited for that job.
3382 static int find_new_ilb(int cpu
)
3384 struct sched_domain
*sd
;
3385 struct sched_group
*ilb_group
;
3388 * Have idle load balancer selection from semi-idle packages only
3389 * when power-aware load balancing is enabled
3391 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
3395 * Optimize for the case when we have no idle CPUs or only one
3396 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3398 if (cpumask_weight(nohz
.idle_cpus_mask
) < 2)
3401 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
3402 ilb_group
= sd
->groups
;
3405 if (is_semi_idle_group(ilb_group
))
3406 return cpumask_first(nohz
.grp_idle_mask
);
3408 ilb_group
= ilb_group
->next
;
3410 } while (ilb_group
!= sd
->groups
);
3416 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3417 static inline int find_new_ilb(int call_cpu
)
3424 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3425 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3426 * CPU (if there is one).
3428 static void nohz_balancer_kick(int cpu
)
3432 nohz
.next_balance
++;
3434 ilb_cpu
= get_nohz_load_balancer();
3436 if (ilb_cpu
>= nr_cpu_ids
) {
3437 ilb_cpu
= cpumask_first(nohz
.idle_cpus_mask
);
3438 if (ilb_cpu
>= nr_cpu_ids
)
3442 if (!cpu_rq(ilb_cpu
)->nohz_balance_kick
) {
3443 struct call_single_data
*cp
;
3445 cpu_rq(ilb_cpu
)->nohz_balance_kick
= 1;
3446 cp
= &per_cpu(remote_sched_softirq_cb
, cpu
);
3447 __smp_call_function_single(ilb_cpu
, cp
, 0);
3453 * This routine will try to nominate the ilb (idle load balancing)
3454 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3455 * load balancing on behalf of all those cpus.
3457 * When the ilb owner becomes busy, we will not have new ilb owner until some
3458 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3459 * idle load balancing by kicking one of the idle CPUs.
3461 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3462 * ilb owner CPU in future (when there is a need for idle load balancing on
3463 * behalf of all idle CPUs).
3465 void select_nohz_load_balancer(int stop_tick
)
3467 int cpu
= smp_processor_id();
3470 if (!cpu_active(cpu
)) {
3471 if (atomic_read(&nohz
.load_balancer
) != cpu
)
3475 * If we are going offline and still the leader,
3478 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
,
3485 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
3487 if (atomic_read(&nohz
.first_pick_cpu
) == cpu
)
3488 atomic_cmpxchg(&nohz
.first_pick_cpu
, cpu
, nr_cpu_ids
);
3489 if (atomic_read(&nohz
.second_pick_cpu
) == cpu
)
3490 atomic_cmpxchg(&nohz
.second_pick_cpu
, cpu
, nr_cpu_ids
);
3492 if (atomic_read(&nohz
.load_balancer
) >= nr_cpu_ids
) {
3495 /* make me the ilb owner */
3496 if (atomic_cmpxchg(&nohz
.load_balancer
, nr_cpu_ids
,
3501 * Check to see if there is a more power-efficient
3504 new_ilb
= find_new_ilb(cpu
);
3505 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
3506 atomic_set(&nohz
.load_balancer
, nr_cpu_ids
);
3507 resched_cpu(new_ilb
);
3513 if (!cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))
3516 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
3518 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3519 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
,
3527 static DEFINE_SPINLOCK(balancing
);
3530 * It checks each scheduling domain to see if it is due to be balanced,
3531 * and initiates a balancing operation if so.
3533 * Balancing parameters are set up in arch_init_sched_domains.
3535 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3538 struct rq
*rq
= cpu_rq(cpu
);
3539 unsigned long interval
;
3540 struct sched_domain
*sd
;
3541 /* Earliest time when we have to do rebalance again */
3542 unsigned long next_balance
= jiffies
+ 60*HZ
;
3543 int update_next_balance
= 0;
3546 for_each_domain(cpu
, sd
) {
3547 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3550 interval
= sd
->balance_interval
;
3551 if (idle
!= CPU_IDLE
)
3552 interval
*= sd
->busy_factor
;
3554 /* scale ms to jiffies */
3555 interval
= msecs_to_jiffies(interval
);
3556 if (unlikely(!interval
))
3558 if (interval
> HZ
*NR_CPUS
/10)
3559 interval
= HZ
*NR_CPUS
/10;
3561 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3563 if (need_serialize
) {
3564 if (!spin_trylock(&balancing
))
3568 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3569 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3571 * We've pulled tasks over so either we're no
3572 * longer idle, or one of our SMT siblings is
3575 idle
= CPU_NOT_IDLE
;
3577 sd
->last_balance
= jiffies
;
3580 spin_unlock(&balancing
);
3582 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3583 next_balance
= sd
->last_balance
+ interval
;
3584 update_next_balance
= 1;
3588 * Stop the load balance at this level. There is another
3589 * CPU in our sched group which is doing load balancing more
3597 * next_balance will be updated only when there is a need.
3598 * When the cpu is attached to null domain for ex, it will not be
3601 if (likely(update_next_balance
))
3602 rq
->next_balance
= next_balance
;
3607 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3608 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3610 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
)
3612 struct rq
*this_rq
= cpu_rq(this_cpu
);
3616 if (idle
!= CPU_IDLE
|| !this_rq
->nohz_balance_kick
)
3619 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
3620 if (balance_cpu
== this_cpu
)
3624 * If this cpu gets work to do, stop the load balancing
3625 * work being done for other cpus. Next load
3626 * balancing owner will pick it up.
3628 if (need_resched()) {
3629 this_rq
->nohz_balance_kick
= 0;
3633 raw_spin_lock_irq(&this_rq
->lock
);
3634 update_rq_clock(this_rq
);
3635 update_cpu_load(this_rq
);
3636 raw_spin_unlock_irq(&this_rq
->lock
);
3638 rebalance_domains(balance_cpu
, CPU_IDLE
);
3640 rq
= cpu_rq(balance_cpu
);
3641 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3642 this_rq
->next_balance
= rq
->next_balance
;
3644 nohz
.next_balance
= this_rq
->next_balance
;
3645 this_rq
->nohz_balance_kick
= 0;
3649 * Current heuristic for kicking the idle load balancer
3650 * - first_pick_cpu is the one of the busy CPUs. It will kick
3651 * idle load balancer when it has more than one process active. This
3652 * eliminates the need for idle load balancing altogether when we have
3653 * only one running process in the system (common case).
3654 * - If there are more than one busy CPU, idle load balancer may have
3655 * to run for active_load_balance to happen (i.e., two busy CPUs are
3656 * SMT or core siblings and can run better if they move to different
3657 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
3658 * which will kick idle load balancer as soon as it has any load.
3660 static inline int nohz_kick_needed(struct rq
*rq
, int cpu
)
3662 unsigned long now
= jiffies
;
3664 int first_pick_cpu
, second_pick_cpu
;
3666 if (time_before(now
, nohz
.next_balance
))
3669 if (rq
->idle_at_tick
)
3672 first_pick_cpu
= atomic_read(&nohz
.first_pick_cpu
);
3673 second_pick_cpu
= atomic_read(&nohz
.second_pick_cpu
);
3675 if (first_pick_cpu
< nr_cpu_ids
&& first_pick_cpu
!= cpu
&&
3676 second_pick_cpu
< nr_cpu_ids
&& second_pick_cpu
!= cpu
)
3679 ret
= atomic_cmpxchg(&nohz
.first_pick_cpu
, nr_cpu_ids
, cpu
);
3680 if (ret
== nr_cpu_ids
|| ret
== cpu
) {
3681 atomic_cmpxchg(&nohz
.second_pick_cpu
, cpu
, nr_cpu_ids
);
3682 if (rq
->nr_running
> 1)
3685 ret
= atomic_cmpxchg(&nohz
.second_pick_cpu
, nr_cpu_ids
, cpu
);
3686 if (ret
== nr_cpu_ids
|| ret
== cpu
) {
3694 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
) { }
3698 * run_rebalance_domains is triggered when needed from the scheduler tick.
3699 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
3701 static void run_rebalance_domains(struct softirq_action
*h
)
3703 int this_cpu
= smp_processor_id();
3704 struct rq
*this_rq
= cpu_rq(this_cpu
);
3705 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3706 CPU_IDLE
: CPU_NOT_IDLE
;
3708 rebalance_domains(this_cpu
, idle
);
3711 * If this cpu has a pending nohz_balance_kick, then do the
3712 * balancing on behalf of the other idle cpus whose ticks are
3715 nohz_idle_balance(this_cpu
, idle
);
3718 static inline int on_null_domain(int cpu
)
3720 return !rcu_dereference_sched(cpu_rq(cpu
)->sd
);
3724 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3726 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3728 /* Don't need to rebalance while attached to NULL domain */
3729 if (time_after_eq(jiffies
, rq
->next_balance
) &&
3730 likely(!on_null_domain(cpu
)))
3731 raise_softirq(SCHED_SOFTIRQ
);
3733 else if (nohz_kick_needed(rq
, cpu
) && likely(!on_null_domain(cpu
)))
3734 nohz_balancer_kick(cpu
);
3738 static void rq_online_fair(struct rq
*rq
)
3743 static void rq_offline_fair(struct rq
*rq
)
3748 #else /* CONFIG_SMP */
3751 * on UP we do not need to balance between CPUs:
3753 static inline void idle_balance(int cpu
, struct rq
*rq
)
3757 #endif /* CONFIG_SMP */
3760 * scheduler tick hitting a task of our scheduling class:
3762 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
3764 struct cfs_rq
*cfs_rq
;
3765 struct sched_entity
*se
= &curr
->se
;
3767 for_each_sched_entity(se
) {
3768 cfs_rq
= cfs_rq_of(se
);
3769 entity_tick(cfs_rq
, se
, queued
);
3774 * called on fork with the child task as argument from the parent's context
3775 * - child not yet on the tasklist
3776 * - preemption disabled
3778 static void task_fork_fair(struct task_struct
*p
)
3780 struct cfs_rq
*cfs_rq
= task_cfs_rq(current
);
3781 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
3782 int this_cpu
= smp_processor_id();
3783 struct rq
*rq
= this_rq();
3784 unsigned long flags
;
3786 raw_spin_lock_irqsave(&rq
->lock
, flags
);
3788 update_rq_clock(rq
);
3790 if (unlikely(task_cpu(p
) != this_cpu
))
3791 __set_task_cpu(p
, this_cpu
);
3793 update_curr(cfs_rq
);
3796 se
->vruntime
= curr
->vruntime
;
3797 place_entity(cfs_rq
, se
, 1);
3799 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
3801 * Upon rescheduling, sched_class::put_prev_task() will place
3802 * 'current' within the tree based on its new key value.
3804 swap(curr
->vruntime
, se
->vruntime
);
3805 resched_task(rq
->curr
);
3808 se
->vruntime
-= cfs_rq
->min_vruntime
;
3810 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
3814 * Priority of the task has changed. Check to see if we preempt
3817 static void prio_changed_fair(struct rq
*rq
, struct task_struct
*p
,
3818 int oldprio
, int running
)
3821 * Reschedule if we are currently running on this runqueue and
3822 * our priority decreased, or if we are not currently running on
3823 * this runqueue and our priority is higher than the current's
3826 if (p
->prio
> oldprio
)
3827 resched_task(rq
->curr
);
3829 check_preempt_curr(rq
, p
, 0);
3833 * We switched to the sched_fair class.
3835 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
,
3839 * We were most likely switched from sched_rt, so
3840 * kick off the schedule if running, otherwise just see
3841 * if we can still preempt the current task.
3844 resched_task(rq
->curr
);
3846 check_preempt_curr(rq
, p
, 0);
3849 /* Account for a task changing its policy or group.
3851 * This routine is mostly called to set cfs_rq->curr field when a task
3852 * migrates between groups/classes.
3854 static void set_curr_task_fair(struct rq
*rq
)
3856 struct sched_entity
*se
= &rq
->curr
->se
;
3858 for_each_sched_entity(se
)
3859 set_next_entity(cfs_rq_of(se
), se
);
3862 #ifdef CONFIG_FAIR_GROUP_SCHED
3863 static void moved_group_fair(struct task_struct
*p
, int on_rq
)
3865 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
3867 update_curr(cfs_rq
);
3869 place_entity(cfs_rq
, &p
->se
, 1);
3873 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
3875 struct sched_entity
*se
= &task
->se
;
3876 unsigned int rr_interval
= 0;
3879 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
3882 if (rq
->cfs
.load
.weight
)
3883 rr_interval
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
3889 * All the scheduling class methods:
3891 static const struct sched_class fair_sched_class
= {
3892 .next
= &idle_sched_class
,
3893 .enqueue_task
= enqueue_task_fair
,
3894 .dequeue_task
= dequeue_task_fair
,
3895 .yield_task
= yield_task_fair
,
3897 .check_preempt_curr
= check_preempt_wakeup
,
3899 .pick_next_task
= pick_next_task_fair
,
3900 .put_prev_task
= put_prev_task_fair
,
3903 .select_task_rq
= select_task_rq_fair
,
3905 .rq_online
= rq_online_fair
,
3906 .rq_offline
= rq_offline_fair
,
3908 .task_waking
= task_waking_fair
,
3911 .set_curr_task
= set_curr_task_fair
,
3912 .task_tick
= task_tick_fair
,
3913 .task_fork
= task_fork_fair
,
3915 .prio_changed
= prio_changed_fair
,
3916 .switched_to
= switched_to_fair
,
3918 .get_rr_interval
= get_rr_interval_fair
,
3920 #ifdef CONFIG_FAIR_GROUP_SCHED
3921 .moved_group
= moved_group_fair
,
3925 #ifdef CONFIG_SCHED_DEBUG
3926 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
3928 struct cfs_rq
*cfs_rq
;
3931 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
3932 print_cfs_rq(m
, cpu
, cfs_rq
);