1 // SPDX-License-Identifier: GPL-2.0
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
26 * Targeted preemption latency for CPU-bound tasks:
28 * NOTE: this latency value is not the same as the concept of
29 * 'timeslice length' - timeslices in CFS are of variable length
30 * and have no persistent notion like in traditional, time-slice
31 * based scheduling concepts.
33 * (to see the precise effective timeslice length of your workload,
34 * run vmstat and monitor the context-switches (cs) field)
36 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
38 unsigned int sysctl_sched_latency
= 6000000ULL;
39 static unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
42 * The initial- and re-scaling of tunables is configurable
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
52 unsigned int sysctl_sched_tunable_scaling
= SCHED_TUNABLESCALING_LOG
;
55 * Minimal preemption granularity for CPU-bound tasks:
57 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
59 unsigned int sysctl_sched_min_granularity
= 750000ULL;
60 static unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
63 * Minimal preemption granularity for CPU-bound SCHED_IDLE tasks.
64 * Applies only when SCHED_IDLE tasks compete with normal tasks.
66 * (default: 0.75 msec)
68 unsigned int sysctl_sched_idle_min_granularity
= 750000ULL;
71 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
73 static unsigned int sched_nr_latency
= 8;
76 * After fork, child runs first. If set to 0 (default) then
77 * parent will (try to) run first.
79 unsigned int sysctl_sched_child_runs_first __read_mostly
;
82 * SCHED_OTHER wake-up granularity.
84 * This option delays the preemption effects of decoupled workloads
85 * and reduces their over-scheduling. Synchronous workloads will still
86 * have immediate wakeup/sleep latencies.
88 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
90 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
91 static unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
93 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
95 int sched_thermal_decay_shift
;
96 static int __init
setup_sched_thermal_decay_shift(char *str
)
100 if (kstrtoint(str
, 0, &_shift
))
101 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
103 sched_thermal_decay_shift
= clamp(_shift
, 0, 10);
106 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift
);
110 * For asym packing, by default the lower numbered CPU has higher priority.
112 int __weak
arch_asym_cpu_priority(int cpu
)
118 * The margin used when comparing utilization with CPU capacity.
122 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
125 * The margin used when comparing CPU capacities.
126 * is 'cap1' noticeably greater than 'cap2'
130 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
133 #ifdef CONFIG_CFS_BANDWIDTH
135 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
136 * each time a cfs_rq requests quota.
138 * Note: in the case that the slice exceeds the runtime remaining (either due
139 * to consumption or the quota being specified to be smaller than the slice)
140 * we will always only issue the remaining available time.
142 * (default: 5 msec, units: microseconds)
144 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
147 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
153 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
159 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
166 * Increase the granularity value when there are more CPUs,
167 * because with more CPUs the 'effective latency' as visible
168 * to users decreases. But the relationship is not linear,
169 * so pick a second-best guess by going with the log2 of the
172 * This idea comes from the SD scheduler of Con Kolivas:
174 static unsigned int get_update_sysctl_factor(void)
176 unsigned int cpus
= min_t(unsigned int, num_online_cpus(), 8);
179 switch (sysctl_sched_tunable_scaling
) {
180 case SCHED_TUNABLESCALING_NONE
:
183 case SCHED_TUNABLESCALING_LINEAR
:
186 case SCHED_TUNABLESCALING_LOG
:
188 factor
= 1 + ilog2(cpus
);
195 static void update_sysctl(void)
197 unsigned int factor
= get_update_sysctl_factor();
199 #define SET_SYSCTL(name) \
200 (sysctl_##name = (factor) * normalized_sysctl_##name)
201 SET_SYSCTL(sched_min_granularity
);
202 SET_SYSCTL(sched_latency
);
203 SET_SYSCTL(sched_wakeup_granularity
);
207 void __init
sched_init_granularity(void)
212 #define WMULT_CONST (~0U)
213 #define WMULT_SHIFT 32
215 static void __update_inv_weight(struct load_weight
*lw
)
219 if (likely(lw
->inv_weight
))
222 w
= scale_load_down(lw
->weight
);
224 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
226 else if (unlikely(!w
))
227 lw
->inv_weight
= WMULT_CONST
;
229 lw
->inv_weight
= WMULT_CONST
/ w
;
233 * delta_exec * weight / lw.weight
235 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
237 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
238 * we're guaranteed shift stays positive because inv_weight is guaranteed to
239 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
241 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
242 * weight/lw.weight <= 1, and therefore our shift will also be positive.
244 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
246 u64 fact
= scale_load_down(weight
);
247 u32 fact_hi
= (u32
)(fact
>> 32);
248 int shift
= WMULT_SHIFT
;
251 __update_inv_weight(lw
);
253 if (unlikely(fact_hi
)) {
259 fact
= mul_u32_u32(fact
, lw
->inv_weight
);
261 fact_hi
= (u32
)(fact
>> 32);
268 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
272 const struct sched_class fair_sched_class
;
274 /**************************************************************
275 * CFS operations on generic schedulable entities:
278 #ifdef CONFIG_FAIR_GROUP_SCHED
280 /* Walk up scheduling entities hierarchy */
281 #define for_each_sched_entity(se) \
282 for (; se; se = se->parent)
284 static inline void cfs_rq_tg_path(struct cfs_rq
*cfs_rq
, char *path
, int len
)
289 if (cfs_rq
&& task_group_is_autogroup(cfs_rq
->tg
))
290 autogroup_path(cfs_rq
->tg
, path
, len
);
291 else if (cfs_rq
&& cfs_rq
->tg
->css
.cgroup
)
292 cgroup_path(cfs_rq
->tg
->css
.cgroup
, path
, len
);
294 strlcpy(path
, "(null)", len
);
297 static inline bool list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
299 struct rq
*rq
= rq_of(cfs_rq
);
300 int cpu
= cpu_of(rq
);
303 return rq
->tmp_alone_branch
== &rq
->leaf_cfs_rq_list
;
308 * Ensure we either appear before our parent (if already
309 * enqueued) or force our parent to appear after us when it is
310 * enqueued. The fact that we always enqueue bottom-up
311 * reduces this to two cases and a special case for the root
312 * cfs_rq. Furthermore, it also means that we will always reset
313 * tmp_alone_branch either when the branch is connected
314 * to a tree or when we reach the top of the tree
316 if (cfs_rq
->tg
->parent
&&
317 cfs_rq
->tg
->parent
->cfs_rq
[cpu
]->on_list
) {
319 * If parent is already on the list, we add the child
320 * just before. Thanks to circular linked property of
321 * the list, this means to put the child at the tail
322 * of the list that starts by parent.
324 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
325 &(cfs_rq
->tg
->parent
->cfs_rq
[cpu
]->leaf_cfs_rq_list
));
327 * The branch is now connected to its tree so we can
328 * reset tmp_alone_branch to the beginning of the
331 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
335 if (!cfs_rq
->tg
->parent
) {
337 * cfs rq without parent should be put
338 * at the tail of the list.
340 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
341 &rq
->leaf_cfs_rq_list
);
343 * We have reach the top of a tree so we can reset
344 * tmp_alone_branch to the beginning of the list.
346 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
351 * The parent has not already been added so we want to
352 * make sure that it will be put after us.
353 * tmp_alone_branch points to the begin of the branch
354 * where we will add parent.
356 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, rq
->tmp_alone_branch
);
358 * update tmp_alone_branch to points to the new begin
361 rq
->tmp_alone_branch
= &cfs_rq
->leaf_cfs_rq_list
;
365 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
367 if (cfs_rq
->on_list
) {
368 struct rq
*rq
= rq_of(cfs_rq
);
371 * With cfs_rq being unthrottled/throttled during an enqueue,
372 * it can happen the tmp_alone_branch points the a leaf that
373 * we finally want to del. In this case, tmp_alone_branch moves
374 * to the prev element but it will point to rq->leaf_cfs_rq_list
375 * at the end of the enqueue.
377 if (rq
->tmp_alone_branch
== &cfs_rq
->leaf_cfs_rq_list
)
378 rq
->tmp_alone_branch
= cfs_rq
->leaf_cfs_rq_list
.prev
;
380 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
385 static inline void assert_list_leaf_cfs_rq(struct rq
*rq
)
387 SCHED_WARN_ON(rq
->tmp_alone_branch
!= &rq
->leaf_cfs_rq_list
);
390 /* Iterate thr' all leaf cfs_rq's on a runqueue */
391 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
392 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
395 /* Do the two (enqueued) entities belong to the same group ? */
396 static inline struct cfs_rq
*
397 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
399 if (se
->cfs_rq
== pse
->cfs_rq
)
405 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
411 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
413 int se_depth
, pse_depth
;
416 * preemption test can be made between sibling entities who are in the
417 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
418 * both tasks until we find their ancestors who are siblings of common
422 /* First walk up until both entities are at same depth */
423 se_depth
= (*se
)->depth
;
424 pse_depth
= (*pse
)->depth
;
426 while (se_depth
> pse_depth
) {
428 *se
= parent_entity(*se
);
431 while (pse_depth
> se_depth
) {
433 *pse
= parent_entity(*pse
);
436 while (!is_same_group(*se
, *pse
)) {
437 *se
= parent_entity(*se
);
438 *pse
= parent_entity(*pse
);
442 static int tg_is_idle(struct task_group
*tg
)
447 static int cfs_rq_is_idle(struct cfs_rq
*cfs_rq
)
449 return cfs_rq
->idle
> 0;
452 static int se_is_idle(struct sched_entity
*se
)
454 if (entity_is_task(se
))
455 return task_has_idle_policy(task_of(se
));
456 return cfs_rq_is_idle(group_cfs_rq(se
));
459 #else /* !CONFIG_FAIR_GROUP_SCHED */
461 #define for_each_sched_entity(se) \
462 for (; se; se = NULL)
464 static inline void cfs_rq_tg_path(struct cfs_rq
*cfs_rq
, char *path
, int len
)
467 strlcpy(path
, "(null)", len
);
470 static inline bool list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
475 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
479 static inline void assert_list_leaf_cfs_rq(struct rq
*rq
)
483 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
484 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
486 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
492 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
496 static inline int tg_is_idle(struct task_group
*tg
)
501 static int cfs_rq_is_idle(struct cfs_rq
*cfs_rq
)
506 static int se_is_idle(struct sched_entity
*se
)
511 #endif /* CONFIG_FAIR_GROUP_SCHED */
513 static __always_inline
514 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
516 /**************************************************************
517 * Scheduling class tree data structure manipulation methods:
520 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
522 s64 delta
= (s64
)(vruntime
- max_vruntime
);
524 max_vruntime
= vruntime
;
529 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
531 s64 delta
= (s64
)(vruntime
- min_vruntime
);
533 min_vruntime
= vruntime
;
538 static inline bool entity_before(struct sched_entity
*a
,
539 struct sched_entity
*b
)
541 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
544 #define __node_2_se(node) \
545 rb_entry((node), struct sched_entity, run_node)
547 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
549 struct sched_entity
*curr
= cfs_rq
->curr
;
550 struct rb_node
*leftmost
= rb_first_cached(&cfs_rq
->tasks_timeline
);
552 u64 vruntime
= cfs_rq
->min_vruntime
;
556 vruntime
= curr
->vruntime
;
561 if (leftmost
) { /* non-empty tree */
562 struct sched_entity
*se
= __node_2_se(leftmost
);
565 vruntime
= se
->vruntime
;
567 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
570 /* ensure we never gain time by being placed backwards. */
571 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
574 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
578 static inline bool __entity_less(struct rb_node
*a
, const struct rb_node
*b
)
580 return entity_before(__node_2_se(a
), __node_2_se(b
));
584 * Enqueue an entity into the rb-tree:
586 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
588 rb_add_cached(&se
->run_node
, &cfs_rq
->tasks_timeline
, __entity_less
);
591 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
593 rb_erase_cached(&se
->run_node
, &cfs_rq
->tasks_timeline
);
596 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
598 struct rb_node
*left
= rb_first_cached(&cfs_rq
->tasks_timeline
);
603 return __node_2_se(left
);
606 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
608 struct rb_node
*next
= rb_next(&se
->run_node
);
613 return __node_2_se(next
);
616 #ifdef CONFIG_SCHED_DEBUG
617 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
619 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
.rb_root
);
624 return __node_2_se(last
);
627 /**************************************************************
628 * Scheduling class statistics methods:
631 int sched_update_scaling(void)
633 unsigned int factor
= get_update_sysctl_factor();
635 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
636 sysctl_sched_min_granularity
);
638 #define WRT_SYSCTL(name) \
639 (normalized_sysctl_##name = sysctl_##name / (factor))
640 WRT_SYSCTL(sched_min_granularity
);
641 WRT_SYSCTL(sched_latency
);
642 WRT_SYSCTL(sched_wakeup_granularity
);
652 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
654 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
655 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
661 * The idea is to set a period in which each task runs once.
663 * When there are too many tasks (sched_nr_latency) we have to stretch
664 * this period because otherwise the slices get too small.
666 * p = (nr <= nl) ? l : l*nr/nl
668 static u64
__sched_period(unsigned long nr_running
)
670 if (unlikely(nr_running
> sched_nr_latency
))
671 return nr_running
* sysctl_sched_min_granularity
;
673 return sysctl_sched_latency
;
676 static bool sched_idle_cfs_rq(struct cfs_rq
*cfs_rq
);
679 * We calculate the wall-time slice from the period by taking a part
680 * proportional to the weight.
684 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
686 unsigned int nr_running
= cfs_rq
->nr_running
;
687 struct sched_entity
*init_se
= se
;
688 unsigned int min_gran
;
691 if (sched_feat(ALT_PERIOD
))
692 nr_running
= rq_of(cfs_rq
)->cfs
.h_nr_running
;
694 slice
= __sched_period(nr_running
+ !se
->on_rq
);
696 for_each_sched_entity(se
) {
697 struct load_weight
*load
;
698 struct load_weight lw
;
699 struct cfs_rq
*qcfs_rq
;
701 qcfs_rq
= cfs_rq_of(se
);
702 load
= &qcfs_rq
->load
;
704 if (unlikely(!se
->on_rq
)) {
707 update_load_add(&lw
, se
->load
.weight
);
710 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
713 if (sched_feat(BASE_SLICE
)) {
714 if (se_is_idle(init_se
) && !sched_idle_cfs_rq(cfs_rq
))
715 min_gran
= sysctl_sched_idle_min_granularity
;
717 min_gran
= sysctl_sched_min_granularity
;
719 slice
= max_t(u64
, slice
, min_gran
);
726 * We calculate the vruntime slice of a to-be-inserted task.
730 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
732 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
738 static int select_idle_sibling(struct task_struct
*p
, int prev_cpu
, int cpu
);
739 static unsigned long task_h_load(struct task_struct
*p
);
740 static unsigned long capacity_of(int cpu
);
742 /* Give new sched_entity start runnable values to heavy its load in infant time */
743 void init_entity_runnable_average(struct sched_entity
*se
)
745 struct sched_avg
*sa
= &se
->avg
;
747 memset(sa
, 0, sizeof(*sa
));
750 * Tasks are initialized with full load to be seen as heavy tasks until
751 * they get a chance to stabilize to their real load level.
752 * Group entities are initialized with zero load to reflect the fact that
753 * nothing has been attached to the task group yet.
755 if (entity_is_task(se
))
756 sa
->load_avg
= scale_load_down(se
->load
.weight
);
758 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
761 static void attach_entity_cfs_rq(struct sched_entity
*se
);
764 * With new tasks being created, their initial util_avgs are extrapolated
765 * based on the cfs_rq's current util_avg:
767 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
769 * However, in many cases, the above util_avg does not give a desired
770 * value. Moreover, the sum of the util_avgs may be divergent, such
771 * as when the series is a harmonic series.
773 * To solve this problem, we also cap the util_avg of successive tasks to
774 * only 1/2 of the left utilization budget:
776 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
778 * where n denotes the nth task and cpu_scale the CPU capacity.
780 * For example, for a CPU with 1024 of capacity, a simplest series from
781 * the beginning would be like:
783 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
784 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
786 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
787 * if util_avg > util_avg_cap.
789 void post_init_entity_util_avg(struct task_struct
*p
)
791 struct sched_entity
*se
= &p
->se
;
792 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
793 struct sched_avg
*sa
= &se
->avg
;
794 long cpu_scale
= arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq
)));
795 long cap
= (long)(cpu_scale
- cfs_rq
->avg
.util_avg
) / 2;
798 if (cfs_rq
->avg
.util_avg
!= 0) {
799 sa
->util_avg
= cfs_rq
->avg
.util_avg
* se
->load
.weight
;
800 sa
->util_avg
/= (cfs_rq
->avg
.load_avg
+ 1);
802 if (sa
->util_avg
> cap
)
809 sa
->runnable_avg
= sa
->util_avg
;
811 if (p
->sched_class
!= &fair_sched_class
) {
813 * For !fair tasks do:
815 update_cfs_rq_load_avg(now, cfs_rq);
816 attach_entity_load_avg(cfs_rq, se);
817 switched_from_fair(rq, p);
819 * such that the next switched_to_fair() has the
822 se
->avg
.last_update_time
= cfs_rq_clock_pelt(cfs_rq
);
826 attach_entity_cfs_rq(se
);
829 #else /* !CONFIG_SMP */
830 void init_entity_runnable_average(struct sched_entity
*se
)
833 void post_init_entity_util_avg(struct task_struct
*p
)
836 static void update_tg_load_avg(struct cfs_rq
*cfs_rq
)
839 #endif /* CONFIG_SMP */
842 * Update the current task's runtime statistics.
844 static void update_curr(struct cfs_rq
*cfs_rq
)
846 struct sched_entity
*curr
= cfs_rq
->curr
;
847 u64 now
= rq_clock_task(rq_of(cfs_rq
));
853 delta_exec
= now
- curr
->exec_start
;
854 if (unlikely((s64
)delta_exec
<= 0))
857 curr
->exec_start
= now
;
859 if (schedstat_enabled()) {
860 struct sched_statistics
*stats
;
862 stats
= __schedstats_from_se(curr
);
863 __schedstat_set(stats
->exec_max
,
864 max(delta_exec
, stats
->exec_max
));
867 curr
->sum_exec_runtime
+= delta_exec
;
868 schedstat_add(cfs_rq
->exec_clock
, delta_exec
);
870 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
871 update_min_vruntime(cfs_rq
);
873 if (entity_is_task(curr
)) {
874 struct task_struct
*curtask
= task_of(curr
);
876 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
877 cgroup_account_cputime(curtask
, delta_exec
);
878 account_group_exec_runtime(curtask
, delta_exec
);
881 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
884 static void update_curr_fair(struct rq
*rq
)
886 update_curr(cfs_rq_of(&rq
->curr
->se
));
890 update_stats_wait_start_fair(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
892 struct sched_statistics
*stats
;
893 struct task_struct
*p
= NULL
;
895 if (!schedstat_enabled())
898 stats
= __schedstats_from_se(se
);
900 if (entity_is_task(se
))
903 __update_stats_wait_start(rq_of(cfs_rq
), p
, stats
);
907 update_stats_wait_end_fair(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
909 struct sched_statistics
*stats
;
910 struct task_struct
*p
= NULL
;
912 if (!schedstat_enabled())
915 stats
= __schedstats_from_se(se
);
918 * When the sched_schedstat changes from 0 to 1, some sched se
919 * maybe already in the runqueue, the se->statistics.wait_start
920 * will be 0.So it will let the delta wrong. We need to avoid this
923 if (unlikely(!schedstat_val(stats
->wait_start
)))
926 if (entity_is_task(se
))
929 __update_stats_wait_end(rq_of(cfs_rq
), p
, stats
);
933 update_stats_enqueue_sleeper_fair(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
935 struct sched_statistics
*stats
;
936 struct task_struct
*tsk
= NULL
;
938 if (!schedstat_enabled())
941 stats
= __schedstats_from_se(se
);
943 if (entity_is_task(se
))
946 __update_stats_enqueue_sleeper(rq_of(cfs_rq
), tsk
, stats
);
950 * Task is being enqueued - update stats:
953 update_stats_enqueue_fair(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
955 if (!schedstat_enabled())
959 * Are we enqueueing a waiting task? (for current tasks
960 * a dequeue/enqueue event is a NOP)
962 if (se
!= cfs_rq
->curr
)
963 update_stats_wait_start_fair(cfs_rq
, se
);
965 if (flags
& ENQUEUE_WAKEUP
)
966 update_stats_enqueue_sleeper_fair(cfs_rq
, se
);
970 update_stats_dequeue_fair(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
973 if (!schedstat_enabled())
977 * Mark the end of the wait period if dequeueing a
980 if (se
!= cfs_rq
->curr
)
981 update_stats_wait_end_fair(cfs_rq
, se
);
983 if ((flags
& DEQUEUE_SLEEP
) && entity_is_task(se
)) {
984 struct task_struct
*tsk
= task_of(se
);
987 /* XXX racy against TTWU */
988 state
= READ_ONCE(tsk
->__state
);
989 if (state
& TASK_INTERRUPTIBLE
)
990 __schedstat_set(tsk
->stats
.sleep_start
,
991 rq_clock(rq_of(cfs_rq
)));
992 if (state
& TASK_UNINTERRUPTIBLE
)
993 __schedstat_set(tsk
->stats
.block_start
,
994 rq_clock(rq_of(cfs_rq
)));
999 * We are picking a new current task - update its stats:
1002 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1005 * We are starting a new run period:
1007 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
1010 /**************************************************
1011 * Scheduling class queueing methods:
1014 #ifdef CONFIG_NUMA_BALANCING
1016 * Approximate time to scan a full NUMA task in ms. The task scan period is
1017 * calculated based on the tasks virtual memory size and
1018 * numa_balancing_scan_size.
1020 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
1021 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
1023 /* Portion of address space to scan in MB */
1024 unsigned int sysctl_numa_balancing_scan_size
= 256;
1026 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1027 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
1030 refcount_t refcount
;
1032 spinlock_t lock
; /* nr_tasks, tasks */
1037 struct rcu_head rcu
;
1038 unsigned long total_faults
;
1039 unsigned long max_faults_cpu
;
1041 * faults[] array is split into two regions: faults_mem and faults_cpu.
1043 * Faults_cpu is used to decide whether memory should move
1044 * towards the CPU. As a consequence, these stats are weighted
1045 * more by CPU use than by memory faults.
1047 unsigned long faults
[];
1051 * For functions that can be called in multiple contexts that permit reading
1052 * ->numa_group (see struct task_struct for locking rules).
1054 static struct numa_group
*deref_task_numa_group(struct task_struct
*p
)
1056 return rcu_dereference_check(p
->numa_group
, p
== current
||
1057 (lockdep_is_held(__rq_lockp(task_rq(p
))) && !READ_ONCE(p
->on_cpu
)));
1060 static struct numa_group
*deref_curr_numa_group(struct task_struct
*p
)
1062 return rcu_dereference_protected(p
->numa_group
, p
== current
);
1065 static inline unsigned long group_faults_priv(struct numa_group
*ng
);
1066 static inline unsigned long group_faults_shared(struct numa_group
*ng
);
1068 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
1070 unsigned long rss
= 0;
1071 unsigned long nr_scan_pages
;
1074 * Calculations based on RSS as non-present and empty pages are skipped
1075 * by the PTE scanner and NUMA hinting faults should be trapped based
1078 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
1079 rss
= get_mm_rss(p
->mm
);
1081 rss
= nr_scan_pages
;
1083 rss
= round_up(rss
, nr_scan_pages
);
1084 return rss
/ nr_scan_pages
;
1087 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1088 #define MAX_SCAN_WINDOW 2560
1090 static unsigned int task_scan_min(struct task_struct
*p
)
1092 unsigned int scan_size
= READ_ONCE(sysctl_numa_balancing_scan_size
);
1093 unsigned int scan
, floor
;
1094 unsigned int windows
= 1;
1096 if (scan_size
< MAX_SCAN_WINDOW
)
1097 windows
= MAX_SCAN_WINDOW
/ scan_size
;
1098 floor
= 1000 / windows
;
1100 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
1101 return max_t(unsigned int, floor
, scan
);
1104 static unsigned int task_scan_start(struct task_struct
*p
)
1106 unsigned long smin
= task_scan_min(p
);
1107 unsigned long period
= smin
;
1108 struct numa_group
*ng
;
1110 /* Scale the maximum scan period with the amount of shared memory. */
1112 ng
= rcu_dereference(p
->numa_group
);
1114 unsigned long shared
= group_faults_shared(ng
);
1115 unsigned long private = group_faults_priv(ng
);
1117 period
*= refcount_read(&ng
->refcount
);
1118 period
*= shared
+ 1;
1119 period
/= private + shared
+ 1;
1123 return max(smin
, period
);
1126 static unsigned int task_scan_max(struct task_struct
*p
)
1128 unsigned long smin
= task_scan_min(p
);
1130 struct numa_group
*ng
;
1132 /* Watch for min being lower than max due to floor calculations */
1133 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
1135 /* Scale the maximum scan period with the amount of shared memory. */
1136 ng
= deref_curr_numa_group(p
);
1138 unsigned long shared
= group_faults_shared(ng
);
1139 unsigned long private = group_faults_priv(ng
);
1140 unsigned long period
= smax
;
1142 period
*= refcount_read(&ng
->refcount
);
1143 period
*= shared
+ 1;
1144 period
/= private + shared
+ 1;
1146 smax
= max(smax
, period
);
1149 return max(smin
, smax
);
1152 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
1154 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= NUMA_NO_NODE
);
1155 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
1158 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
1160 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= NUMA_NO_NODE
);
1161 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
1164 /* Shared or private faults. */
1165 #define NR_NUMA_HINT_FAULT_TYPES 2
1167 /* Memory and CPU locality */
1168 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1170 /* Averaged statistics, and temporary buffers. */
1171 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1173 pid_t
task_numa_group_id(struct task_struct
*p
)
1175 struct numa_group
*ng
;
1179 ng
= rcu_dereference(p
->numa_group
);
1188 * The averaged statistics, shared & private, memory & CPU,
1189 * occupy the first half of the array. The second half of the
1190 * array is for current counters, which are averaged into the
1191 * first set by task_numa_placement.
1193 static inline int task_faults_idx(enum numa_faults_stats s
, int nid
, int priv
)
1195 return NR_NUMA_HINT_FAULT_TYPES
* (s
* nr_node_ids
+ nid
) + priv
;
1198 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
1200 if (!p
->numa_faults
)
1203 return p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1204 p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1207 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
1209 struct numa_group
*ng
= deref_task_numa_group(p
);
1214 return ng
->faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1215 ng
->faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1218 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
1220 return group
->faults
[task_faults_idx(NUMA_CPU
, nid
, 0)] +
1221 group
->faults
[task_faults_idx(NUMA_CPU
, nid
, 1)];
1224 static inline unsigned long group_faults_priv(struct numa_group
*ng
)
1226 unsigned long faults
= 0;
1229 for_each_online_node(node
) {
1230 faults
+= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
1236 static inline unsigned long group_faults_shared(struct numa_group
*ng
)
1238 unsigned long faults
= 0;
1241 for_each_online_node(node
) {
1242 faults
+= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
1249 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1250 * considered part of a numa group's pseudo-interleaving set. Migrations
1251 * between these nodes are slowed down, to allow things to settle down.
1253 #define ACTIVE_NODE_FRACTION 3
1255 static bool numa_is_active_node(int nid
, struct numa_group
*ng
)
1257 return group_faults_cpu(ng
, nid
) * ACTIVE_NODE_FRACTION
> ng
->max_faults_cpu
;
1260 /* Handle placement on systems where not all nodes are directly connected. */
1261 static unsigned long score_nearby_nodes(struct task_struct
*p
, int nid
,
1262 int maxdist
, bool task
)
1264 unsigned long score
= 0;
1268 * All nodes are directly connected, and the same distance
1269 * from each other. No need for fancy placement algorithms.
1271 if (sched_numa_topology_type
== NUMA_DIRECT
)
1275 * This code is called for each node, introducing N^2 complexity,
1276 * which should be ok given the number of nodes rarely exceeds 8.
1278 for_each_online_node(node
) {
1279 unsigned long faults
;
1280 int dist
= node_distance(nid
, node
);
1283 * The furthest away nodes in the system are not interesting
1284 * for placement; nid was already counted.
1286 if (dist
== sched_max_numa_distance
|| node
== nid
)
1290 * On systems with a backplane NUMA topology, compare groups
1291 * of nodes, and move tasks towards the group with the most
1292 * memory accesses. When comparing two nodes at distance
1293 * "hoplimit", only nodes closer by than "hoplimit" are part
1294 * of each group. Skip other nodes.
1296 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1300 /* Add up the faults from nearby nodes. */
1302 faults
= task_faults(p
, node
);
1304 faults
= group_faults(p
, node
);
1307 * On systems with a glueless mesh NUMA topology, there are
1308 * no fixed "groups of nodes". Instead, nodes that are not
1309 * directly connected bounce traffic through intermediate
1310 * nodes; a numa_group can occupy any set of nodes.
1311 * The further away a node is, the less the faults count.
1312 * This seems to result in good task placement.
1314 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1315 faults
*= (sched_max_numa_distance
- dist
);
1316 faults
/= (sched_max_numa_distance
- LOCAL_DISTANCE
);
1326 * These return the fraction of accesses done by a particular task, or
1327 * task group, on a particular numa node. The group weight is given a
1328 * larger multiplier, in order to group tasks together that are almost
1329 * evenly spread out between numa nodes.
1331 static inline unsigned long task_weight(struct task_struct
*p
, int nid
,
1334 unsigned long faults
, total_faults
;
1336 if (!p
->numa_faults
)
1339 total_faults
= p
->total_numa_faults
;
1344 faults
= task_faults(p
, nid
);
1345 faults
+= score_nearby_nodes(p
, nid
, dist
, true);
1347 return 1000 * faults
/ total_faults
;
1350 static inline unsigned long group_weight(struct task_struct
*p
, int nid
,
1353 struct numa_group
*ng
= deref_task_numa_group(p
);
1354 unsigned long faults
, total_faults
;
1359 total_faults
= ng
->total_faults
;
1364 faults
= group_faults(p
, nid
);
1365 faults
+= score_nearby_nodes(p
, nid
, dist
, false);
1367 return 1000 * faults
/ total_faults
;
1370 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
1371 int src_nid
, int dst_cpu
)
1373 struct numa_group
*ng
= deref_curr_numa_group(p
);
1374 int dst_nid
= cpu_to_node(dst_cpu
);
1375 int last_cpupid
, this_cpupid
;
1377 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
1378 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
1381 * Allow first faults or private faults to migrate immediately early in
1382 * the lifetime of a task. The magic number 4 is based on waiting for
1383 * two full passes of the "multi-stage node selection" test that is
1386 if ((p
->numa_preferred_nid
== NUMA_NO_NODE
|| p
->numa_scan_seq
<= 4) &&
1387 (cpupid_pid_unset(last_cpupid
) || cpupid_match_pid(p
, last_cpupid
)))
1391 * Multi-stage node selection is used in conjunction with a periodic
1392 * migration fault to build a temporal task<->page relation. By using
1393 * a two-stage filter we remove short/unlikely relations.
1395 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1396 * a task's usage of a particular page (n_p) per total usage of this
1397 * page (n_t) (in a given time-span) to a probability.
1399 * Our periodic faults will sample this probability and getting the
1400 * same result twice in a row, given these samples are fully
1401 * independent, is then given by P(n)^2, provided our sample period
1402 * is sufficiently short compared to the usage pattern.
1404 * This quadric squishes small probabilities, making it less likely we
1405 * act on an unlikely task<->page relation.
1407 if (!cpupid_pid_unset(last_cpupid
) &&
1408 cpupid_to_nid(last_cpupid
) != dst_nid
)
1411 /* Always allow migrate on private faults */
1412 if (cpupid_match_pid(p
, last_cpupid
))
1415 /* A shared fault, but p->numa_group has not been set up yet. */
1420 * Destination node is much more heavily used than the source
1421 * node? Allow migration.
1423 if (group_faults_cpu(ng
, dst_nid
) > group_faults_cpu(ng
, src_nid
) *
1424 ACTIVE_NODE_FRACTION
)
1428 * Distribute memory according to CPU & memory use on each node,
1429 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1431 * faults_cpu(dst) 3 faults_cpu(src)
1432 * --------------- * - > ---------------
1433 * faults_mem(dst) 4 faults_mem(src)
1435 return group_faults_cpu(ng
, dst_nid
) * group_faults(p
, src_nid
) * 3 >
1436 group_faults_cpu(ng
, src_nid
) * group_faults(p
, dst_nid
) * 4;
1440 * 'numa_type' describes the node at the moment of load balancing.
1443 /* The node has spare capacity that can be used to run more tasks. */
1446 * The node is fully used and the tasks don't compete for more CPU
1447 * cycles. Nevertheless, some tasks might wait before running.
1451 * The node is overloaded and can't provide expected CPU cycles to all
1457 /* Cached statistics for all CPUs within a node */
1460 unsigned long runnable
;
1462 /* Total compute capacity of CPUs on a node */
1463 unsigned long compute_capacity
;
1464 unsigned int nr_running
;
1465 unsigned int weight
;
1466 enum numa_type node_type
;
1470 static inline bool is_core_idle(int cpu
)
1472 #ifdef CONFIG_SCHED_SMT
1475 for_each_cpu(sibling
, cpu_smt_mask(cpu
)) {
1479 if (!idle_cpu(sibling
))
1487 struct task_numa_env
{
1488 struct task_struct
*p
;
1490 int src_cpu
, src_nid
;
1491 int dst_cpu
, dst_nid
;
1493 struct numa_stats src_stats
, dst_stats
;
1498 struct task_struct
*best_task
;
1503 static unsigned long cpu_load(struct rq
*rq
);
1504 static unsigned long cpu_runnable(struct rq
*rq
);
1505 static inline long adjust_numa_imbalance(int imbalance
,
1506 int dst_running
, int dst_weight
);
1509 numa_type
numa_classify(unsigned int imbalance_pct
,
1510 struct numa_stats
*ns
)
1512 if ((ns
->nr_running
> ns
->weight
) &&
1513 (((ns
->compute_capacity
* 100) < (ns
->util
* imbalance_pct
)) ||
1514 ((ns
->compute_capacity
* imbalance_pct
) < (ns
->runnable
* 100))))
1515 return node_overloaded
;
1517 if ((ns
->nr_running
< ns
->weight
) ||
1518 (((ns
->compute_capacity
* 100) > (ns
->util
* imbalance_pct
)) &&
1519 ((ns
->compute_capacity
* imbalance_pct
) > (ns
->runnable
* 100))))
1520 return node_has_spare
;
1522 return node_fully_busy
;
1525 #ifdef CONFIG_SCHED_SMT
1526 /* Forward declarations of select_idle_sibling helpers */
1527 static inline bool test_idle_cores(int cpu
, bool def
);
1528 static inline int numa_idle_core(int idle_core
, int cpu
)
1530 if (!static_branch_likely(&sched_smt_present
) ||
1531 idle_core
>= 0 || !test_idle_cores(cpu
, false))
1535 * Prefer cores instead of packing HT siblings
1536 * and triggering future load balancing.
1538 if (is_core_idle(cpu
))
1544 static inline int numa_idle_core(int idle_core
, int cpu
)
1551 * Gather all necessary information to make NUMA balancing placement
1552 * decisions that are compatible with standard load balancer. This
1553 * borrows code and logic from update_sg_lb_stats but sharing a
1554 * common implementation is impractical.
1556 static void update_numa_stats(struct task_numa_env
*env
,
1557 struct numa_stats
*ns
, int nid
,
1560 int cpu
, idle_core
= -1;
1562 memset(ns
, 0, sizeof(*ns
));
1566 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1567 struct rq
*rq
= cpu_rq(cpu
);
1569 ns
->load
+= cpu_load(rq
);
1570 ns
->runnable
+= cpu_runnable(rq
);
1571 ns
->util
+= cpu_util_cfs(cpu
);
1572 ns
->nr_running
+= rq
->cfs
.h_nr_running
;
1573 ns
->compute_capacity
+= capacity_of(cpu
);
1575 if (find_idle
&& !rq
->nr_running
&& idle_cpu(cpu
)) {
1576 if (READ_ONCE(rq
->numa_migrate_on
) ||
1577 !cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
))
1580 if (ns
->idle_cpu
== -1)
1583 idle_core
= numa_idle_core(idle_core
, cpu
);
1588 ns
->weight
= cpumask_weight(cpumask_of_node(nid
));
1590 ns
->node_type
= numa_classify(env
->imbalance_pct
, ns
);
1593 ns
->idle_cpu
= idle_core
;
1596 static void task_numa_assign(struct task_numa_env
*env
,
1597 struct task_struct
*p
, long imp
)
1599 struct rq
*rq
= cpu_rq(env
->dst_cpu
);
1601 /* Check if run-queue part of active NUMA balance. */
1602 if (env
->best_cpu
!= env
->dst_cpu
&& xchg(&rq
->numa_migrate_on
, 1)) {
1604 int start
= env
->dst_cpu
;
1606 /* Find alternative idle CPU. */
1607 for_each_cpu_wrap(cpu
, cpumask_of_node(env
->dst_nid
), start
) {
1608 if (cpu
== env
->best_cpu
|| !idle_cpu(cpu
) ||
1609 !cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
)) {
1614 rq
= cpu_rq(env
->dst_cpu
);
1615 if (!xchg(&rq
->numa_migrate_on
, 1))
1619 /* Failed to find an alternative idle CPU */
1625 * Clear previous best_cpu/rq numa-migrate flag, since task now
1626 * found a better CPU to move/swap.
1628 if (env
->best_cpu
!= -1 && env
->best_cpu
!= env
->dst_cpu
) {
1629 rq
= cpu_rq(env
->best_cpu
);
1630 WRITE_ONCE(rq
->numa_migrate_on
, 0);
1634 put_task_struct(env
->best_task
);
1639 env
->best_imp
= imp
;
1640 env
->best_cpu
= env
->dst_cpu
;
1643 static bool load_too_imbalanced(long src_load
, long dst_load
,
1644 struct task_numa_env
*env
)
1647 long orig_src_load
, orig_dst_load
;
1648 long src_capacity
, dst_capacity
;
1651 * The load is corrected for the CPU capacity available on each node.
1654 * ------------ vs ---------
1655 * src_capacity dst_capacity
1657 src_capacity
= env
->src_stats
.compute_capacity
;
1658 dst_capacity
= env
->dst_stats
.compute_capacity
;
1660 imb
= abs(dst_load
* src_capacity
- src_load
* dst_capacity
);
1662 orig_src_load
= env
->src_stats
.load
;
1663 orig_dst_load
= env
->dst_stats
.load
;
1665 old_imb
= abs(orig_dst_load
* src_capacity
- orig_src_load
* dst_capacity
);
1667 /* Would this change make things worse? */
1668 return (imb
> old_imb
);
1672 * Maximum NUMA importance can be 1998 (2*999);
1673 * SMALLIMP @ 30 would be close to 1998/64.
1674 * Used to deter task migration.
1679 * This checks if the overall compute and NUMA accesses of the system would
1680 * be improved if the source tasks was migrated to the target dst_cpu taking
1681 * into account that it might be best if task running on the dst_cpu should
1682 * be exchanged with the source task
1684 static bool task_numa_compare(struct task_numa_env
*env
,
1685 long taskimp
, long groupimp
, bool maymove
)
1687 struct numa_group
*cur_ng
, *p_ng
= deref_curr_numa_group(env
->p
);
1688 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1689 long imp
= p_ng
? groupimp
: taskimp
;
1690 struct task_struct
*cur
;
1691 long src_load
, dst_load
;
1692 int dist
= env
->dist
;
1695 bool stopsearch
= false;
1697 if (READ_ONCE(dst_rq
->numa_migrate_on
))
1701 cur
= rcu_dereference(dst_rq
->curr
);
1702 if (cur
&& ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
)))
1706 * Because we have preemption enabled we can get migrated around and
1707 * end try selecting ourselves (current == env->p) as a swap candidate.
1709 if (cur
== env
->p
) {
1715 if (maymove
&& moveimp
>= env
->best_imp
)
1721 /* Skip this swap candidate if cannot move to the source cpu. */
1722 if (!cpumask_test_cpu(env
->src_cpu
, cur
->cpus_ptr
))
1726 * Skip this swap candidate if it is not moving to its preferred
1727 * node and the best task is.
1729 if (env
->best_task
&&
1730 env
->best_task
->numa_preferred_nid
== env
->src_nid
&&
1731 cur
->numa_preferred_nid
!= env
->src_nid
) {
1736 * "imp" is the fault differential for the source task between the
1737 * source and destination node. Calculate the total differential for
1738 * the source task and potential destination task. The more negative
1739 * the value is, the more remote accesses that would be expected to
1740 * be incurred if the tasks were swapped.
1742 * If dst and source tasks are in the same NUMA group, or not
1743 * in any group then look only at task weights.
1745 cur_ng
= rcu_dereference(cur
->numa_group
);
1746 if (cur_ng
== p_ng
) {
1747 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1748 task_weight(cur
, env
->dst_nid
, dist
);
1750 * Add some hysteresis to prevent swapping the
1751 * tasks within a group over tiny differences.
1757 * Compare the group weights. If a task is all by itself
1758 * (not part of a group), use the task weight instead.
1761 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1762 group_weight(cur
, env
->dst_nid
, dist
);
1764 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1765 task_weight(cur
, env
->dst_nid
, dist
);
1768 /* Discourage picking a task already on its preferred node */
1769 if (cur
->numa_preferred_nid
== env
->dst_nid
)
1773 * Encourage picking a task that moves to its preferred node.
1774 * This potentially makes imp larger than it's maximum of
1775 * 1998 (see SMALLIMP and task_weight for why) but in this
1776 * case, it does not matter.
1778 if (cur
->numa_preferred_nid
== env
->src_nid
)
1781 if (maymove
&& moveimp
> imp
&& moveimp
> env
->best_imp
) {
1788 * Prefer swapping with a task moving to its preferred node over a
1791 if (env
->best_task
&& cur
->numa_preferred_nid
== env
->src_nid
&&
1792 env
->best_task
->numa_preferred_nid
!= env
->src_nid
) {
1797 * If the NUMA importance is less than SMALLIMP,
1798 * task migration might only result in ping pong
1799 * of tasks and also hurt performance due to cache
1802 if (imp
< SMALLIMP
|| imp
<= env
->best_imp
+ SMALLIMP
/ 2)
1806 * In the overloaded case, try and keep the load balanced.
1808 load
= task_h_load(env
->p
) - task_h_load(cur
);
1812 dst_load
= env
->dst_stats
.load
+ load
;
1813 src_load
= env
->src_stats
.load
- load
;
1815 if (load_too_imbalanced(src_load
, dst_load
, env
))
1819 /* Evaluate an idle CPU for a task numa move. */
1821 int cpu
= env
->dst_stats
.idle_cpu
;
1823 /* Nothing cached so current CPU went idle since the search. */
1828 * If the CPU is no longer truly idle and the previous best CPU
1829 * is, keep using it.
1831 if (!idle_cpu(cpu
) && env
->best_cpu
>= 0 &&
1832 idle_cpu(env
->best_cpu
)) {
1833 cpu
= env
->best_cpu
;
1839 task_numa_assign(env
, cur
, imp
);
1842 * If a move to idle is allowed because there is capacity or load
1843 * balance improves then stop the search. While a better swap
1844 * candidate may exist, a search is not free.
1846 if (maymove
&& !cur
&& env
->best_cpu
>= 0 && idle_cpu(env
->best_cpu
))
1850 * If a swap candidate must be identified and the current best task
1851 * moves its preferred node then stop the search.
1853 if (!maymove
&& env
->best_task
&&
1854 env
->best_task
->numa_preferred_nid
== env
->src_nid
) {
1863 static void task_numa_find_cpu(struct task_numa_env
*env
,
1864 long taskimp
, long groupimp
)
1866 bool maymove
= false;
1870 * If dst node has spare capacity, then check if there is an
1871 * imbalance that would be overruled by the load balancer.
1873 if (env
->dst_stats
.node_type
== node_has_spare
) {
1874 unsigned int imbalance
;
1875 int src_running
, dst_running
;
1878 * Would movement cause an imbalance? Note that if src has
1879 * more running tasks that the imbalance is ignored as the
1880 * move improves the imbalance from the perspective of the
1881 * CPU load balancer.
1883 src_running
= env
->src_stats
.nr_running
- 1;
1884 dst_running
= env
->dst_stats
.nr_running
+ 1;
1885 imbalance
= max(0, dst_running
- src_running
);
1886 imbalance
= adjust_numa_imbalance(imbalance
, dst_running
,
1887 env
->dst_stats
.weight
);
1889 /* Use idle CPU if there is no imbalance */
1892 if (env
->dst_stats
.idle_cpu
>= 0) {
1893 env
->dst_cpu
= env
->dst_stats
.idle_cpu
;
1894 task_numa_assign(env
, NULL
, 0);
1899 long src_load
, dst_load
, load
;
1901 * If the improvement from just moving env->p direction is better
1902 * than swapping tasks around, check if a move is possible.
1904 load
= task_h_load(env
->p
);
1905 dst_load
= env
->dst_stats
.load
+ load
;
1906 src_load
= env
->src_stats
.load
- load
;
1907 maymove
= !load_too_imbalanced(src_load
, dst_load
, env
);
1910 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1911 /* Skip this CPU if the source task cannot migrate */
1912 if (!cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
))
1916 if (task_numa_compare(env
, taskimp
, groupimp
, maymove
))
1921 static int task_numa_migrate(struct task_struct
*p
)
1923 struct task_numa_env env
= {
1926 .src_cpu
= task_cpu(p
),
1927 .src_nid
= task_node(p
),
1929 .imbalance_pct
= 112,
1935 unsigned long taskweight
, groupweight
;
1936 struct sched_domain
*sd
;
1937 long taskimp
, groupimp
;
1938 struct numa_group
*ng
;
1943 * Pick the lowest SD_NUMA domain, as that would have the smallest
1944 * imbalance and would be the first to start moving tasks about.
1946 * And we want to avoid any moving of tasks about, as that would create
1947 * random movement of tasks -- counter the numa conditions we're trying
1951 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1953 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1957 * Cpusets can break the scheduler domain tree into smaller
1958 * balance domains, some of which do not cross NUMA boundaries.
1959 * Tasks that are "trapped" in such domains cannot be migrated
1960 * elsewhere, so there is no point in (re)trying.
1962 if (unlikely(!sd
)) {
1963 sched_setnuma(p
, task_node(p
));
1967 env
.dst_nid
= p
->numa_preferred_nid
;
1968 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1969 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1970 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1971 update_numa_stats(&env
, &env
.src_stats
, env
.src_nid
, false);
1972 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
1973 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
1974 update_numa_stats(&env
, &env
.dst_stats
, env
.dst_nid
, true);
1976 /* Try to find a spot on the preferred nid. */
1977 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1980 * Look at other nodes in these cases:
1981 * - there is no space available on the preferred_nid
1982 * - the task is part of a numa_group that is interleaved across
1983 * multiple NUMA nodes; in order to better consolidate the group,
1984 * we need to check other locations.
1986 ng
= deref_curr_numa_group(p
);
1987 if (env
.best_cpu
== -1 || (ng
&& ng
->active_nodes
> 1)) {
1988 for_each_online_node(nid
) {
1989 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1992 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1993 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1995 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1996 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1999 /* Only consider nodes where both task and groups benefit */
2000 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
2001 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
2002 if (taskimp
< 0 && groupimp
< 0)
2007 update_numa_stats(&env
, &env
.dst_stats
, env
.dst_nid
, true);
2008 task_numa_find_cpu(&env
, taskimp
, groupimp
);
2013 * If the task is part of a workload that spans multiple NUMA nodes,
2014 * and is migrating into one of the workload's active nodes, remember
2015 * this node as the task's preferred numa node, so the workload can
2017 * A task that migrated to a second choice node will be better off
2018 * trying for a better one later. Do not set the preferred node here.
2021 if (env
.best_cpu
== -1)
2024 nid
= cpu_to_node(env
.best_cpu
);
2026 if (nid
!= p
->numa_preferred_nid
)
2027 sched_setnuma(p
, nid
);
2030 /* No better CPU than the current one was found. */
2031 if (env
.best_cpu
== -1) {
2032 trace_sched_stick_numa(p
, env
.src_cpu
, NULL
, -1);
2036 best_rq
= cpu_rq(env
.best_cpu
);
2037 if (env
.best_task
== NULL
) {
2038 ret
= migrate_task_to(p
, env
.best_cpu
);
2039 WRITE_ONCE(best_rq
->numa_migrate_on
, 0);
2041 trace_sched_stick_numa(p
, env
.src_cpu
, NULL
, env
.best_cpu
);
2045 ret
= migrate_swap(p
, env
.best_task
, env
.best_cpu
, env
.src_cpu
);
2046 WRITE_ONCE(best_rq
->numa_migrate_on
, 0);
2049 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_task
, env
.best_cpu
);
2050 put_task_struct(env
.best_task
);
2054 /* Attempt to migrate a task to a CPU on the preferred node. */
2055 static void numa_migrate_preferred(struct task_struct
*p
)
2057 unsigned long interval
= HZ
;
2059 /* This task has no NUMA fault statistics yet */
2060 if (unlikely(p
->numa_preferred_nid
== NUMA_NO_NODE
|| !p
->numa_faults
))
2063 /* Periodically retry migrating the task to the preferred node */
2064 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
2065 p
->numa_migrate_retry
= jiffies
+ interval
;
2067 /* Success if task is already running on preferred CPU */
2068 if (task_node(p
) == p
->numa_preferred_nid
)
2071 /* Otherwise, try migrate to a CPU on the preferred node */
2072 task_numa_migrate(p
);
2076 * Find out how many nodes the workload is actively running on. Do this by
2077 * tracking the nodes from which NUMA hinting faults are triggered. This can
2078 * be different from the set of nodes where the workload's memory is currently
2081 static void numa_group_count_active_nodes(struct numa_group
*numa_group
)
2083 unsigned long faults
, max_faults
= 0;
2084 int nid
, active_nodes
= 0;
2086 for_each_online_node(nid
) {
2087 faults
= group_faults_cpu(numa_group
, nid
);
2088 if (faults
> max_faults
)
2089 max_faults
= faults
;
2092 for_each_online_node(nid
) {
2093 faults
= group_faults_cpu(numa_group
, nid
);
2094 if (faults
* ACTIVE_NODE_FRACTION
> max_faults
)
2098 numa_group
->max_faults_cpu
= max_faults
;
2099 numa_group
->active_nodes
= active_nodes
;
2103 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2104 * increments. The more local the fault statistics are, the higher the scan
2105 * period will be for the next scan window. If local/(local+remote) ratio is
2106 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2107 * the scan period will decrease. Aim for 70% local accesses.
2109 #define NUMA_PERIOD_SLOTS 10
2110 #define NUMA_PERIOD_THRESHOLD 7
2113 * Increase the scan period (slow down scanning) if the majority of
2114 * our memory is already on our local node, or if the majority of
2115 * the page accesses are shared with other processes.
2116 * Otherwise, decrease the scan period.
2118 static void update_task_scan_period(struct task_struct
*p
,
2119 unsigned long shared
, unsigned long private)
2121 unsigned int period_slot
;
2122 int lr_ratio
, ps_ratio
;
2125 unsigned long remote
= p
->numa_faults_locality
[0];
2126 unsigned long local
= p
->numa_faults_locality
[1];
2129 * If there were no record hinting faults then either the task is
2130 * completely idle or all activity is in areas that are not of interest
2131 * to automatic numa balancing. Related to that, if there were failed
2132 * migration then it implies we are migrating too quickly or the local
2133 * node is overloaded. In either case, scan slower
2135 if (local
+ shared
== 0 || p
->numa_faults_locality
[2]) {
2136 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
2137 p
->numa_scan_period
<< 1);
2139 p
->mm
->numa_next_scan
= jiffies
+
2140 msecs_to_jiffies(p
->numa_scan_period
);
2146 * Prepare to scale scan period relative to the current period.
2147 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2148 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2149 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2151 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
2152 lr_ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
2153 ps_ratio
= (private * NUMA_PERIOD_SLOTS
) / (private + shared
);
2155 if (ps_ratio
>= NUMA_PERIOD_THRESHOLD
) {
2157 * Most memory accesses are local. There is no need to
2158 * do fast NUMA scanning, since memory is already local.
2160 int slot
= ps_ratio
- NUMA_PERIOD_THRESHOLD
;
2163 diff
= slot
* period_slot
;
2164 } else if (lr_ratio
>= NUMA_PERIOD_THRESHOLD
) {
2166 * Most memory accesses are shared with other tasks.
2167 * There is no point in continuing fast NUMA scanning,
2168 * since other tasks may just move the memory elsewhere.
2170 int slot
= lr_ratio
- NUMA_PERIOD_THRESHOLD
;
2173 diff
= slot
* period_slot
;
2176 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2177 * yet they are not on the local NUMA node. Speed up
2178 * NUMA scanning to get the memory moved over.
2180 int ratio
= max(lr_ratio
, ps_ratio
);
2181 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
2184 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
2185 task_scan_min(p
), task_scan_max(p
));
2186 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2190 * Get the fraction of time the task has been running since the last
2191 * NUMA placement cycle. The scheduler keeps similar statistics, but
2192 * decays those on a 32ms period, which is orders of magnitude off
2193 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2194 * stats only if the task is so new there are no NUMA statistics yet.
2196 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
2198 u64 runtime
, delta
, now
;
2199 /* Use the start of this time slice to avoid calculations. */
2200 now
= p
->se
.exec_start
;
2201 runtime
= p
->se
.sum_exec_runtime
;
2203 if (p
->last_task_numa_placement
) {
2204 delta
= runtime
- p
->last_sum_exec_runtime
;
2205 *period
= now
- p
->last_task_numa_placement
;
2207 /* Avoid time going backwards, prevent potential divide error: */
2208 if (unlikely((s64
)*period
< 0))
2211 delta
= p
->se
.avg
.load_sum
;
2212 *period
= LOAD_AVG_MAX
;
2215 p
->last_sum_exec_runtime
= runtime
;
2216 p
->last_task_numa_placement
= now
;
2222 * Determine the preferred nid for a task in a numa_group. This needs to
2223 * be done in a way that produces consistent results with group_weight,
2224 * otherwise workloads might not converge.
2226 static int preferred_group_nid(struct task_struct
*p
, int nid
)
2231 /* Direct connections between all NUMA nodes. */
2232 if (sched_numa_topology_type
== NUMA_DIRECT
)
2236 * On a system with glueless mesh NUMA topology, group_weight
2237 * scores nodes according to the number of NUMA hinting faults on
2238 * both the node itself, and on nearby nodes.
2240 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
2241 unsigned long score
, max_score
= 0;
2242 int node
, max_node
= nid
;
2244 dist
= sched_max_numa_distance
;
2246 for_each_online_node(node
) {
2247 score
= group_weight(p
, node
, dist
);
2248 if (score
> max_score
) {
2257 * Finding the preferred nid in a system with NUMA backplane
2258 * interconnect topology is more involved. The goal is to locate
2259 * tasks from numa_groups near each other in the system, and
2260 * untangle workloads from different sides of the system. This requires
2261 * searching down the hierarchy of node groups, recursively searching
2262 * inside the highest scoring group of nodes. The nodemask tricks
2263 * keep the complexity of the search down.
2265 nodes
= node_online_map
;
2266 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
2267 unsigned long max_faults
= 0;
2268 nodemask_t max_group
= NODE_MASK_NONE
;
2271 /* Are there nodes at this distance from each other? */
2272 if (!find_numa_distance(dist
))
2275 for_each_node_mask(a
, nodes
) {
2276 unsigned long faults
= 0;
2277 nodemask_t this_group
;
2278 nodes_clear(this_group
);
2280 /* Sum group's NUMA faults; includes a==b case. */
2281 for_each_node_mask(b
, nodes
) {
2282 if (node_distance(a
, b
) < dist
) {
2283 faults
+= group_faults(p
, b
);
2284 node_set(b
, this_group
);
2285 node_clear(b
, nodes
);
2289 /* Remember the top group. */
2290 if (faults
> max_faults
) {
2291 max_faults
= faults
;
2292 max_group
= this_group
;
2294 * subtle: at the smallest distance there is
2295 * just one node left in each "group", the
2296 * winner is the preferred nid.
2301 /* Next round, evaluate the nodes within max_group. */
2309 static void task_numa_placement(struct task_struct
*p
)
2311 int seq
, nid
, max_nid
= NUMA_NO_NODE
;
2312 unsigned long max_faults
= 0;
2313 unsigned long fault_types
[2] = { 0, 0 };
2314 unsigned long total_faults
;
2315 u64 runtime
, period
;
2316 spinlock_t
*group_lock
= NULL
;
2317 struct numa_group
*ng
;
2320 * The p->mm->numa_scan_seq field gets updated without
2321 * exclusive access. Use READ_ONCE() here to ensure
2322 * that the field is read in a single access:
2324 seq
= READ_ONCE(p
->mm
->numa_scan_seq
);
2325 if (p
->numa_scan_seq
== seq
)
2327 p
->numa_scan_seq
= seq
;
2328 p
->numa_scan_period_max
= task_scan_max(p
);
2330 total_faults
= p
->numa_faults_locality
[0] +
2331 p
->numa_faults_locality
[1];
2332 runtime
= numa_get_avg_runtime(p
, &period
);
2334 /* If the task is part of a group prevent parallel updates to group stats */
2335 ng
= deref_curr_numa_group(p
);
2337 group_lock
= &ng
->lock
;
2338 spin_lock_irq(group_lock
);
2341 /* Find the node with the highest number of faults */
2342 for_each_online_node(nid
) {
2343 /* Keep track of the offsets in numa_faults array */
2344 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
2345 unsigned long faults
= 0, group_faults
= 0;
2348 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
2349 long diff
, f_diff
, f_weight
;
2351 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
2352 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
2353 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
2354 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
2356 /* Decay existing window, copy faults since last scan */
2357 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
2358 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
2359 p
->numa_faults
[membuf_idx
] = 0;
2362 * Normalize the faults_from, so all tasks in a group
2363 * count according to CPU use, instead of by the raw
2364 * number of faults. Tasks with little runtime have
2365 * little over-all impact on throughput, and thus their
2366 * faults are less important.
2368 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
2369 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
2371 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
2372 p
->numa_faults
[cpubuf_idx
] = 0;
2374 p
->numa_faults
[mem_idx
] += diff
;
2375 p
->numa_faults
[cpu_idx
] += f_diff
;
2376 faults
+= p
->numa_faults
[mem_idx
];
2377 p
->total_numa_faults
+= diff
;
2380 * safe because we can only change our own group
2382 * mem_idx represents the offset for a given
2383 * nid and priv in a specific region because it
2384 * is at the beginning of the numa_faults array.
2386 ng
->faults
[mem_idx
] += diff
;
2387 ng
->faults
[cpu_idx
] += f_diff
;
2388 ng
->total_faults
+= diff
;
2389 group_faults
+= ng
->faults
[mem_idx
];
2394 if (faults
> max_faults
) {
2395 max_faults
= faults
;
2398 } else if (group_faults
> max_faults
) {
2399 max_faults
= group_faults
;
2405 numa_group_count_active_nodes(ng
);
2406 spin_unlock_irq(group_lock
);
2407 max_nid
= preferred_group_nid(p
, max_nid
);
2411 /* Set the new preferred node */
2412 if (max_nid
!= p
->numa_preferred_nid
)
2413 sched_setnuma(p
, max_nid
);
2416 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
2419 static inline int get_numa_group(struct numa_group
*grp
)
2421 return refcount_inc_not_zero(&grp
->refcount
);
2424 static inline void put_numa_group(struct numa_group
*grp
)
2426 if (refcount_dec_and_test(&grp
->refcount
))
2427 kfree_rcu(grp
, rcu
);
2430 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
2433 struct numa_group
*grp
, *my_grp
;
2434 struct task_struct
*tsk
;
2436 int cpu
= cpupid_to_cpu(cpupid
);
2439 if (unlikely(!deref_curr_numa_group(p
))) {
2440 unsigned int size
= sizeof(struct numa_group
) +
2441 NR_NUMA_HINT_FAULT_STATS
*
2442 nr_node_ids
* sizeof(unsigned long);
2444 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
2448 refcount_set(&grp
->refcount
, 1);
2449 grp
->active_nodes
= 1;
2450 grp
->max_faults_cpu
= 0;
2451 spin_lock_init(&grp
->lock
);
2454 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2455 grp
->faults
[i
] = p
->numa_faults
[i
];
2457 grp
->total_faults
= p
->total_numa_faults
;
2460 rcu_assign_pointer(p
->numa_group
, grp
);
2464 tsk
= READ_ONCE(cpu_rq(cpu
)->curr
);
2466 if (!cpupid_match_pid(tsk
, cpupid
))
2469 grp
= rcu_dereference(tsk
->numa_group
);
2473 my_grp
= deref_curr_numa_group(p
);
2478 * Only join the other group if its bigger; if we're the bigger group,
2479 * the other task will join us.
2481 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
2485 * Tie-break on the grp address.
2487 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
2490 /* Always join threads in the same process. */
2491 if (tsk
->mm
== current
->mm
)
2494 /* Simple filter to avoid false positives due to PID collisions */
2495 if (flags
& TNF_SHARED
)
2498 /* Update priv based on whether false sharing was detected */
2501 if (join
&& !get_numa_group(grp
))
2509 BUG_ON(irqs_disabled());
2510 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
2512 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
2513 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
2514 grp
->faults
[i
] += p
->numa_faults
[i
];
2516 my_grp
->total_faults
-= p
->total_numa_faults
;
2517 grp
->total_faults
+= p
->total_numa_faults
;
2522 spin_unlock(&my_grp
->lock
);
2523 spin_unlock_irq(&grp
->lock
);
2525 rcu_assign_pointer(p
->numa_group
, grp
);
2527 put_numa_group(my_grp
);
2536 * Get rid of NUMA statistics associated with a task (either current or dead).
2537 * If @final is set, the task is dead and has reached refcount zero, so we can
2538 * safely free all relevant data structures. Otherwise, there might be
2539 * concurrent reads from places like load balancing and procfs, and we should
2540 * reset the data back to default state without freeing ->numa_faults.
2542 void task_numa_free(struct task_struct
*p
, bool final
)
2544 /* safe: p either is current or is being freed by current */
2545 struct numa_group
*grp
= rcu_dereference_raw(p
->numa_group
);
2546 unsigned long *numa_faults
= p
->numa_faults
;
2547 unsigned long flags
;
2554 spin_lock_irqsave(&grp
->lock
, flags
);
2555 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2556 grp
->faults
[i
] -= p
->numa_faults
[i
];
2557 grp
->total_faults
-= p
->total_numa_faults
;
2560 spin_unlock_irqrestore(&grp
->lock
, flags
);
2561 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2562 put_numa_group(grp
);
2566 p
->numa_faults
= NULL
;
2569 p
->total_numa_faults
= 0;
2570 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2576 * Got a PROT_NONE fault for a page on @node.
2578 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2580 struct task_struct
*p
= current
;
2581 bool migrated
= flags
& TNF_MIGRATED
;
2582 int cpu_node
= task_node(current
);
2583 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2584 struct numa_group
*ng
;
2587 if (!static_branch_likely(&sched_numa_balancing
))
2590 /* for example, ksmd faulting in a user's mm */
2594 /* Allocate buffer to track faults on a per-node basis */
2595 if (unlikely(!p
->numa_faults
)) {
2596 int size
= sizeof(*p
->numa_faults
) *
2597 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2599 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2600 if (!p
->numa_faults
)
2603 p
->total_numa_faults
= 0;
2604 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2608 * First accesses are treated as private, otherwise consider accesses
2609 * to be private if the accessing pid has not changed
2611 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2614 priv
= cpupid_match_pid(p
, last_cpupid
);
2615 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2616 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2620 * If a workload spans multiple NUMA nodes, a shared fault that
2621 * occurs wholly within the set of nodes that the workload is
2622 * actively using should be counted as local. This allows the
2623 * scan rate to slow down when a workload has settled down.
2625 ng
= deref_curr_numa_group(p
);
2626 if (!priv
&& !local
&& ng
&& ng
->active_nodes
> 1 &&
2627 numa_is_active_node(cpu_node
, ng
) &&
2628 numa_is_active_node(mem_node
, ng
))
2632 * Retry to migrate task to preferred node periodically, in case it
2633 * previously failed, or the scheduler moved us.
2635 if (time_after(jiffies
, p
->numa_migrate_retry
)) {
2636 task_numa_placement(p
);
2637 numa_migrate_preferred(p
);
2641 p
->numa_pages_migrated
+= pages
;
2642 if (flags
& TNF_MIGRATE_FAIL
)
2643 p
->numa_faults_locality
[2] += pages
;
2645 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2646 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2647 p
->numa_faults_locality
[local
] += pages
;
2650 static void reset_ptenuma_scan(struct task_struct
*p
)
2653 * We only did a read acquisition of the mmap sem, so
2654 * p->mm->numa_scan_seq is written to without exclusive access
2655 * and the update is not guaranteed to be atomic. That's not
2656 * much of an issue though, since this is just used for
2657 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2658 * expensive, to avoid any form of compiler optimizations:
2660 WRITE_ONCE(p
->mm
->numa_scan_seq
, READ_ONCE(p
->mm
->numa_scan_seq
) + 1);
2661 p
->mm
->numa_scan_offset
= 0;
2665 * The expensive part of numa migration is done from task_work context.
2666 * Triggered from task_tick_numa().
2668 static void task_numa_work(struct callback_head
*work
)
2670 unsigned long migrate
, next_scan
, now
= jiffies
;
2671 struct task_struct
*p
= current
;
2672 struct mm_struct
*mm
= p
->mm
;
2673 u64 runtime
= p
->se
.sum_exec_runtime
;
2674 struct vm_area_struct
*vma
;
2675 unsigned long start
, end
;
2676 unsigned long nr_pte_updates
= 0;
2677 long pages
, virtpages
;
2679 SCHED_WARN_ON(p
!= container_of(work
, struct task_struct
, numa_work
));
2683 * Who cares about NUMA placement when they're dying.
2685 * NOTE: make sure not to dereference p->mm before this check,
2686 * exit_task_work() happens _after_ exit_mm() so we could be called
2687 * without p->mm even though we still had it when we enqueued this
2690 if (p
->flags
& PF_EXITING
)
2693 if (!mm
->numa_next_scan
) {
2694 mm
->numa_next_scan
= now
+
2695 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2699 * Enforce maximal scan/migration frequency..
2701 migrate
= mm
->numa_next_scan
;
2702 if (time_before(now
, migrate
))
2705 if (p
->numa_scan_period
== 0) {
2706 p
->numa_scan_period_max
= task_scan_max(p
);
2707 p
->numa_scan_period
= task_scan_start(p
);
2710 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2711 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2715 * Delay this task enough that another task of this mm will likely win
2716 * the next time around.
2718 p
->node_stamp
+= 2 * TICK_NSEC
;
2720 start
= mm
->numa_scan_offset
;
2721 pages
= sysctl_numa_balancing_scan_size
;
2722 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2723 virtpages
= pages
* 8; /* Scan up to this much virtual space */
2728 if (!mmap_read_trylock(mm
))
2730 vma
= find_vma(mm
, start
);
2732 reset_ptenuma_scan(p
);
2736 for (; vma
; vma
= vma
->vm_next
) {
2737 if (!vma_migratable(vma
) || !vma_policy_mof(vma
) ||
2738 is_vm_hugetlb_page(vma
) || (vma
->vm_flags
& VM_MIXEDMAP
)) {
2743 * Shared library pages mapped by multiple processes are not
2744 * migrated as it is expected they are cache replicated. Avoid
2745 * hinting faults in read-only file-backed mappings or the vdso
2746 * as migrating the pages will be of marginal benefit.
2749 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2753 * Skip inaccessible VMAs to avoid any confusion between
2754 * PROT_NONE and NUMA hinting ptes
2756 if (!vma_is_accessible(vma
))
2760 start
= max(start
, vma
->vm_start
);
2761 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2762 end
= min(end
, vma
->vm_end
);
2763 nr_pte_updates
= change_prot_numa(vma
, start
, end
);
2766 * Try to scan sysctl_numa_balancing_size worth of
2767 * hpages that have at least one present PTE that
2768 * is not already pte-numa. If the VMA contains
2769 * areas that are unused or already full of prot_numa
2770 * PTEs, scan up to virtpages, to skip through those
2774 pages
-= (end
- start
) >> PAGE_SHIFT
;
2775 virtpages
-= (end
- start
) >> PAGE_SHIFT
;
2778 if (pages
<= 0 || virtpages
<= 0)
2782 } while (end
!= vma
->vm_end
);
2787 * It is possible to reach the end of the VMA list but the last few
2788 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2789 * would find the !migratable VMA on the next scan but not reset the
2790 * scanner to the start so check it now.
2793 mm
->numa_scan_offset
= start
;
2795 reset_ptenuma_scan(p
);
2796 mmap_read_unlock(mm
);
2799 * Make sure tasks use at least 32x as much time to run other code
2800 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2801 * Usually update_task_scan_period slows down scanning enough; on an
2802 * overloaded system we need to limit overhead on a per task basis.
2804 if (unlikely(p
->se
.sum_exec_runtime
!= runtime
)) {
2805 u64 diff
= p
->se
.sum_exec_runtime
- runtime
;
2806 p
->node_stamp
+= 32 * diff
;
2810 void init_numa_balancing(unsigned long clone_flags
, struct task_struct
*p
)
2813 struct mm_struct
*mm
= p
->mm
;
2816 mm_users
= atomic_read(&mm
->mm_users
);
2817 if (mm_users
== 1) {
2818 mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2819 mm
->numa_scan_seq
= 0;
2823 p
->numa_scan_seq
= mm
? mm
->numa_scan_seq
: 0;
2824 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2825 /* Protect against double add, see task_tick_numa and task_numa_work */
2826 p
->numa_work
.next
= &p
->numa_work
;
2827 p
->numa_faults
= NULL
;
2828 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2829 p
->last_task_numa_placement
= 0;
2830 p
->last_sum_exec_runtime
= 0;
2832 init_task_work(&p
->numa_work
, task_numa_work
);
2834 /* New address space, reset the preferred nid */
2835 if (!(clone_flags
& CLONE_VM
)) {
2836 p
->numa_preferred_nid
= NUMA_NO_NODE
;
2841 * New thread, keep existing numa_preferred_nid which should be copied
2842 * already by arch_dup_task_struct but stagger when scans start.
2847 delay
= min_t(unsigned int, task_scan_max(current
),
2848 current
->numa_scan_period
* mm_users
* NSEC_PER_MSEC
);
2849 delay
+= 2 * TICK_NSEC
;
2850 p
->node_stamp
= delay
;
2855 * Drive the periodic memory faults..
2857 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2859 struct callback_head
*work
= &curr
->numa_work
;
2863 * We don't care about NUMA placement if we don't have memory.
2865 if ((curr
->flags
& (PF_EXITING
| PF_KTHREAD
)) || work
->next
!= work
)
2869 * Using runtime rather than walltime has the dual advantage that
2870 * we (mostly) drive the selection from busy threads and that the
2871 * task needs to have done some actual work before we bother with
2874 now
= curr
->se
.sum_exec_runtime
;
2875 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2877 if (now
> curr
->node_stamp
+ period
) {
2878 if (!curr
->node_stamp
)
2879 curr
->numa_scan_period
= task_scan_start(curr
);
2880 curr
->node_stamp
+= period
;
2882 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
))
2883 task_work_add(curr
, work
, TWA_RESUME
);
2887 static void update_scan_period(struct task_struct
*p
, int new_cpu
)
2889 int src_nid
= cpu_to_node(task_cpu(p
));
2890 int dst_nid
= cpu_to_node(new_cpu
);
2892 if (!static_branch_likely(&sched_numa_balancing
))
2895 if (!p
->mm
|| !p
->numa_faults
|| (p
->flags
& PF_EXITING
))
2898 if (src_nid
== dst_nid
)
2902 * Allow resets if faults have been trapped before one scan
2903 * has completed. This is most likely due to a new task that
2904 * is pulled cross-node due to wakeups or load balancing.
2906 if (p
->numa_scan_seq
) {
2908 * Avoid scan adjustments if moving to the preferred
2909 * node or if the task was not previously running on
2910 * the preferred node.
2912 if (dst_nid
== p
->numa_preferred_nid
||
2913 (p
->numa_preferred_nid
!= NUMA_NO_NODE
&&
2914 src_nid
!= p
->numa_preferred_nid
))
2918 p
->numa_scan_period
= task_scan_start(p
);
2922 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2926 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2930 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2934 static inline void update_scan_period(struct task_struct
*p
, int new_cpu
)
2938 #endif /* CONFIG_NUMA_BALANCING */
2941 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2943 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2945 if (entity_is_task(se
)) {
2946 struct rq
*rq
= rq_of(cfs_rq
);
2948 account_numa_enqueue(rq
, task_of(se
));
2949 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2952 cfs_rq
->nr_running
++;
2954 cfs_rq
->idle_nr_running
++;
2958 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2960 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2962 if (entity_is_task(se
)) {
2963 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2964 list_del_init(&se
->group_node
);
2967 cfs_rq
->nr_running
--;
2969 cfs_rq
->idle_nr_running
--;
2973 * Signed add and clamp on underflow.
2975 * Explicitly do a load-store to ensure the intermediate value never hits
2976 * memory. This allows lockless observations without ever seeing the negative
2979 #define add_positive(_ptr, _val) do { \
2980 typeof(_ptr) ptr = (_ptr); \
2981 typeof(_val) val = (_val); \
2982 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2986 if (val < 0 && res > var) \
2989 WRITE_ONCE(*ptr, res); \
2993 * Unsigned subtract and clamp on underflow.
2995 * Explicitly do a load-store to ensure the intermediate value never hits
2996 * memory. This allows lockless observations without ever seeing the negative
2999 #define sub_positive(_ptr, _val) do { \
3000 typeof(_ptr) ptr = (_ptr); \
3001 typeof(*ptr) val = (_val); \
3002 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3006 WRITE_ONCE(*ptr, res); \
3010 * Remove and clamp on negative, from a local variable.
3012 * A variant of sub_positive(), which does not use explicit load-store
3013 * and is thus optimized for local variable updates.
3015 #define lsub_positive(_ptr, _val) do { \
3016 typeof(_ptr) ptr = (_ptr); \
3017 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3022 enqueue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3024 cfs_rq
->avg
.load_avg
+= se
->avg
.load_avg
;
3025 cfs_rq
->avg
.load_sum
+= se_weight(se
) * se
->avg
.load_sum
;
3029 dequeue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3031 u32 divider
= get_pelt_divider(&se
->avg
);
3032 sub_positive(&cfs_rq
->avg
.load_avg
, se
->avg
.load_avg
);
3033 cfs_rq
->avg
.load_sum
= cfs_rq
->avg
.load_avg
* divider
;
3037 enqueue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
3039 dequeue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
3042 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
3043 unsigned long weight
)
3046 /* commit outstanding execution time */
3047 if (cfs_rq
->curr
== se
)
3048 update_curr(cfs_rq
);
3049 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
3051 dequeue_load_avg(cfs_rq
, se
);
3053 update_load_set(&se
->load
, weight
);
3057 u32 divider
= get_pelt_divider(&se
->avg
);
3059 se
->avg
.load_avg
= div_u64(se_weight(se
) * se
->avg
.load_sum
, divider
);
3063 enqueue_load_avg(cfs_rq
, se
);
3065 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
3069 void reweight_task(struct task_struct
*p
, int prio
)
3071 struct sched_entity
*se
= &p
->se
;
3072 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3073 struct load_weight
*load
= &se
->load
;
3074 unsigned long weight
= scale_load(sched_prio_to_weight
[prio
]);
3076 reweight_entity(cfs_rq
, se
, weight
);
3077 load
->inv_weight
= sched_prio_to_wmult
[prio
];
3080 #ifdef CONFIG_FAIR_GROUP_SCHED
3083 * All this does is approximate the hierarchical proportion which includes that
3084 * global sum we all love to hate.
3086 * That is, the weight of a group entity, is the proportional share of the
3087 * group weight based on the group runqueue weights. That is:
3089 * tg->weight * grq->load.weight
3090 * ge->load.weight = ----------------------------- (1)
3091 * \Sum grq->load.weight
3093 * Now, because computing that sum is prohibitively expensive to compute (been
3094 * there, done that) we approximate it with this average stuff. The average
3095 * moves slower and therefore the approximation is cheaper and more stable.
3097 * So instead of the above, we substitute:
3099 * grq->load.weight -> grq->avg.load_avg (2)
3101 * which yields the following:
3103 * tg->weight * grq->avg.load_avg
3104 * ge->load.weight = ------------------------------ (3)
3107 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3109 * That is shares_avg, and it is right (given the approximation (2)).
3111 * The problem with it is that because the average is slow -- it was designed
3112 * to be exactly that of course -- this leads to transients in boundary
3113 * conditions. In specific, the case where the group was idle and we start the
3114 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3115 * yielding bad latency etc..
3117 * Now, in that special case (1) reduces to:
3119 * tg->weight * grq->load.weight
3120 * ge->load.weight = ----------------------------- = tg->weight (4)
3123 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3125 * So what we do is modify our approximation (3) to approach (4) in the (near)
3130 * tg->weight * grq->load.weight
3131 * --------------------------------------------------- (5)
3132 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3134 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3135 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3138 * tg->weight * grq->load.weight
3139 * ge->load.weight = ----------------------------- (6)
3144 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3145 * max(grq->load.weight, grq->avg.load_avg)
3147 * And that is shares_weight and is icky. In the (near) UP case it approaches
3148 * (4) while in the normal case it approaches (3). It consistently
3149 * overestimates the ge->load.weight and therefore:
3151 * \Sum ge->load.weight >= tg->weight
3155 static long calc_group_shares(struct cfs_rq
*cfs_rq
)
3157 long tg_weight
, tg_shares
, load
, shares
;
3158 struct task_group
*tg
= cfs_rq
->tg
;
3160 tg_shares
= READ_ONCE(tg
->shares
);
3162 load
= max(scale_load_down(cfs_rq
->load
.weight
), cfs_rq
->avg
.load_avg
);
3164 tg_weight
= atomic_long_read(&tg
->load_avg
);
3166 /* Ensure tg_weight >= load */
3167 tg_weight
-= cfs_rq
->tg_load_avg_contrib
;
3170 shares
= (tg_shares
* load
);
3172 shares
/= tg_weight
;
3175 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3176 * of a group with small tg->shares value. It is a floor value which is
3177 * assigned as a minimum load.weight to the sched_entity representing
3178 * the group on a CPU.
3180 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3181 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3182 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3183 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3186 return clamp_t(long, shares
, MIN_SHARES
, tg_shares
);
3188 #endif /* CONFIG_SMP */
3190 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
3193 * Recomputes the group entity based on the current state of its group
3196 static void update_cfs_group(struct sched_entity
*se
)
3198 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3204 if (throttled_hierarchy(gcfs_rq
))
3208 shares
= READ_ONCE(gcfs_rq
->tg
->shares
);
3210 if (likely(se
->load
.weight
== shares
))
3213 shares
= calc_group_shares(gcfs_rq
);
3216 reweight_entity(cfs_rq_of(se
), se
, shares
);
3219 #else /* CONFIG_FAIR_GROUP_SCHED */
3220 static inline void update_cfs_group(struct sched_entity
*se
)
3223 #endif /* CONFIG_FAIR_GROUP_SCHED */
3225 static inline void cfs_rq_util_change(struct cfs_rq
*cfs_rq
, int flags
)
3227 struct rq
*rq
= rq_of(cfs_rq
);
3229 if (&rq
->cfs
== cfs_rq
) {
3231 * There are a few boundary cases this might miss but it should
3232 * get called often enough that that should (hopefully) not be
3235 * It will not get called when we go idle, because the idle
3236 * thread is a different class (!fair), nor will the utilization
3237 * number include things like RT tasks.
3239 * As is, the util number is not freq-invariant (we'd have to
3240 * implement arch_scale_freq_capacity() for that).
3242 * See cpu_util_cfs().
3244 cpufreq_update_util(rq
, flags
);
3249 #ifdef CONFIG_FAIR_GROUP_SCHED
3251 * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
3252 * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
3253 * bottom-up, we only have to test whether the cfs_rq before us on the list
3255 * If cfs_rq is not on the list, test whether a child needs its to be added to
3256 * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
3258 static inline bool child_cfs_rq_on_list(struct cfs_rq
*cfs_rq
)
3260 struct cfs_rq
*prev_cfs_rq
;
3261 struct list_head
*prev
;
3263 if (cfs_rq
->on_list
) {
3264 prev
= cfs_rq
->leaf_cfs_rq_list
.prev
;
3266 struct rq
*rq
= rq_of(cfs_rq
);
3268 prev
= rq
->tmp_alone_branch
;
3271 prev_cfs_rq
= container_of(prev
, struct cfs_rq
, leaf_cfs_rq_list
);
3273 return (prev_cfs_rq
->tg
->parent
== cfs_rq
->tg
);
3276 static inline bool cfs_rq_is_decayed(struct cfs_rq
*cfs_rq
)
3278 if (cfs_rq
->load
.weight
)
3281 if (cfs_rq
->avg
.load_sum
)
3284 if (cfs_rq
->avg
.util_sum
)
3287 if (cfs_rq
->avg
.runnable_sum
)
3290 if (child_cfs_rq_on_list(cfs_rq
))
3294 * _avg must be null when _sum are null because _avg = _sum / divider
3295 * Make sure that rounding and/or propagation of PELT values never
3298 SCHED_WARN_ON(cfs_rq
->avg
.load_avg
||
3299 cfs_rq
->avg
.util_avg
||
3300 cfs_rq
->avg
.runnable_avg
);
3306 * update_tg_load_avg - update the tg's load avg
3307 * @cfs_rq: the cfs_rq whose avg changed
3309 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3310 * However, because tg->load_avg is a global value there are performance
3313 * In order to avoid having to look at the other cfs_rq's, we use a
3314 * differential update where we store the last value we propagated. This in
3315 * turn allows skipping updates if the differential is 'small'.
3317 * Updating tg's load_avg is necessary before update_cfs_share().
3319 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
)
3321 long delta
= cfs_rq
->avg
.load_avg
- cfs_rq
->tg_load_avg_contrib
;
3324 * No need to update load_avg for root_task_group as it is not used.
3326 if (cfs_rq
->tg
== &root_task_group
)
3329 if (abs(delta
) > cfs_rq
->tg_load_avg_contrib
/ 64) {
3330 atomic_long_add(delta
, &cfs_rq
->tg
->load_avg
);
3331 cfs_rq
->tg_load_avg_contrib
= cfs_rq
->avg
.load_avg
;
3336 * Called within set_task_rq() right before setting a task's CPU. The
3337 * caller only guarantees p->pi_lock is held; no other assumptions,
3338 * including the state of rq->lock, should be made.
3340 void set_task_rq_fair(struct sched_entity
*se
,
3341 struct cfs_rq
*prev
, struct cfs_rq
*next
)
3343 u64 p_last_update_time
;
3344 u64 n_last_update_time
;
3346 if (!sched_feat(ATTACH_AGE_LOAD
))
3350 * We are supposed to update the task to "current" time, then its up to
3351 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3352 * getting what current time is, so simply throw away the out-of-date
3353 * time. This will result in the wakee task is less decayed, but giving
3354 * the wakee more load sounds not bad.
3356 if (!(se
->avg
.last_update_time
&& prev
))
3359 #ifndef CONFIG_64BIT
3361 u64 p_last_update_time_copy
;
3362 u64 n_last_update_time_copy
;
3365 p_last_update_time_copy
= prev
->load_last_update_time_copy
;
3366 n_last_update_time_copy
= next
->load_last_update_time_copy
;
3370 p_last_update_time
= prev
->avg
.last_update_time
;
3371 n_last_update_time
= next
->avg
.last_update_time
;
3373 } while (p_last_update_time
!= p_last_update_time_copy
||
3374 n_last_update_time
!= n_last_update_time_copy
);
3377 p_last_update_time
= prev
->avg
.last_update_time
;
3378 n_last_update_time
= next
->avg
.last_update_time
;
3380 __update_load_avg_blocked_se(p_last_update_time
, se
);
3381 se
->avg
.last_update_time
= n_last_update_time
;
3386 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3387 * propagate its contribution. The key to this propagation is the invariant
3388 * that for each group:
3390 * ge->avg == grq->avg (1)
3392 * _IFF_ we look at the pure running and runnable sums. Because they
3393 * represent the very same entity, just at different points in the hierarchy.
3395 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3396 * and simply copies the running/runnable sum over (but still wrong, because
3397 * the group entity and group rq do not have their PELT windows aligned).
3399 * However, update_tg_cfs_load() is more complex. So we have:
3401 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3403 * And since, like util, the runnable part should be directly transferable,
3404 * the following would _appear_ to be the straight forward approach:
3406 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3408 * And per (1) we have:
3410 * ge->avg.runnable_avg == grq->avg.runnable_avg
3414 * ge->load.weight * grq->avg.load_avg
3415 * ge->avg.load_avg = ----------------------------------- (4)
3418 * Except that is wrong!
3420 * Because while for entities historical weight is not important and we
3421 * really only care about our future and therefore can consider a pure
3422 * runnable sum, runqueues can NOT do this.
3424 * We specifically want runqueues to have a load_avg that includes
3425 * historical weights. Those represent the blocked load, the load we expect
3426 * to (shortly) return to us. This only works by keeping the weights as
3427 * integral part of the sum. We therefore cannot decompose as per (3).
3429 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3430 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3431 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3432 * runnable section of these tasks overlap (or not). If they were to perfectly
3433 * align the rq as a whole would be runnable 2/3 of the time. If however we
3434 * always have at least 1 runnable task, the rq as a whole is always runnable.
3436 * So we'll have to approximate.. :/
3438 * Given the constraint:
3440 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3442 * We can construct a rule that adds runnable to a rq by assuming minimal
3445 * On removal, we'll assume each task is equally runnable; which yields:
3447 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3449 * XXX: only do this for the part of runnable > running ?
3454 update_tg_cfs_util(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3456 long delta
= gcfs_rq
->avg
.util_avg
- se
->avg
.util_avg
;
3459 /* Nothing to update */
3464 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3465 * See ___update_load_avg() for details.
3467 divider
= get_pelt_divider(&cfs_rq
->avg
);
3469 /* Set new sched_entity's utilization */
3470 se
->avg
.util_avg
= gcfs_rq
->avg
.util_avg
;
3471 se
->avg
.util_sum
= se
->avg
.util_avg
* divider
;
3473 /* Update parent cfs_rq utilization */
3474 add_positive(&cfs_rq
->avg
.util_avg
, delta
);
3475 cfs_rq
->avg
.util_sum
= cfs_rq
->avg
.util_avg
* divider
;
3479 update_tg_cfs_runnable(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3481 long delta
= gcfs_rq
->avg
.runnable_avg
- se
->avg
.runnable_avg
;
3484 /* Nothing to update */
3489 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3490 * See ___update_load_avg() for details.
3492 divider
= get_pelt_divider(&cfs_rq
->avg
);
3494 /* Set new sched_entity's runnable */
3495 se
->avg
.runnable_avg
= gcfs_rq
->avg
.runnable_avg
;
3496 se
->avg
.runnable_sum
= se
->avg
.runnable_avg
* divider
;
3498 /* Update parent cfs_rq runnable */
3499 add_positive(&cfs_rq
->avg
.runnable_avg
, delta
);
3500 cfs_rq
->avg
.runnable_sum
= cfs_rq
->avg
.runnable_avg
* divider
;
3504 update_tg_cfs_load(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3506 long delta
, running_sum
, runnable_sum
= gcfs_rq
->prop_runnable_sum
;
3507 unsigned long load_avg
;
3514 gcfs_rq
->prop_runnable_sum
= 0;
3517 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3518 * See ___update_load_avg() for details.
3520 divider
= get_pelt_divider(&cfs_rq
->avg
);
3522 if (runnable_sum
>= 0) {
3524 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3525 * the CPU is saturated running == runnable.
3527 runnable_sum
+= se
->avg
.load_sum
;
3528 runnable_sum
= min_t(long, runnable_sum
, divider
);
3531 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3532 * assuming all tasks are equally runnable.
3534 if (scale_load_down(gcfs_rq
->load
.weight
)) {
3535 load_sum
= div_s64(gcfs_rq
->avg
.load_sum
,
3536 scale_load_down(gcfs_rq
->load
.weight
));
3539 /* But make sure to not inflate se's runnable */
3540 runnable_sum
= min(se
->avg
.load_sum
, load_sum
);
3544 * runnable_sum can't be lower than running_sum
3545 * Rescale running sum to be in the same range as runnable sum
3546 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3547 * runnable_sum is in [0 : LOAD_AVG_MAX]
3549 running_sum
= se
->avg
.util_sum
>> SCHED_CAPACITY_SHIFT
;
3550 runnable_sum
= max(runnable_sum
, running_sum
);
3552 load_sum
= (s64
)se_weight(se
) * runnable_sum
;
3553 load_avg
= div_s64(load_sum
, divider
);
3555 se
->avg
.load_sum
= runnable_sum
;
3557 delta
= load_avg
- se
->avg
.load_avg
;
3561 se
->avg
.load_avg
= load_avg
;
3563 add_positive(&cfs_rq
->avg
.load_avg
, delta
);
3564 cfs_rq
->avg
.load_sum
= cfs_rq
->avg
.load_avg
* divider
;
3567 static inline void add_tg_cfs_propagate(struct cfs_rq
*cfs_rq
, long runnable_sum
)
3569 cfs_rq
->propagate
= 1;
3570 cfs_rq
->prop_runnable_sum
+= runnable_sum
;
3573 /* Update task and its cfs_rq load average */
3574 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3576 struct cfs_rq
*cfs_rq
, *gcfs_rq
;
3578 if (entity_is_task(se
))
3581 gcfs_rq
= group_cfs_rq(se
);
3582 if (!gcfs_rq
->propagate
)
3585 gcfs_rq
->propagate
= 0;
3587 cfs_rq
= cfs_rq_of(se
);
3589 add_tg_cfs_propagate(cfs_rq
, gcfs_rq
->prop_runnable_sum
);
3591 update_tg_cfs_util(cfs_rq
, se
, gcfs_rq
);
3592 update_tg_cfs_runnable(cfs_rq
, se
, gcfs_rq
);
3593 update_tg_cfs_load(cfs_rq
, se
, gcfs_rq
);
3595 trace_pelt_cfs_tp(cfs_rq
);
3596 trace_pelt_se_tp(se
);
3602 * Check if we need to update the load and the utilization of a blocked
3605 static inline bool skip_blocked_update(struct sched_entity
*se
)
3607 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3610 * If sched_entity still have not zero load or utilization, we have to
3613 if (se
->avg
.load_avg
|| se
->avg
.util_avg
)
3617 * If there is a pending propagation, we have to update the load and
3618 * the utilization of the sched_entity:
3620 if (gcfs_rq
->propagate
)
3624 * Otherwise, the load and the utilization of the sched_entity is
3625 * already zero and there is no pending propagation, so it will be a
3626 * waste of time to try to decay it:
3631 #else /* CONFIG_FAIR_GROUP_SCHED */
3633 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
) {}
3635 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3640 static inline void add_tg_cfs_propagate(struct cfs_rq
*cfs_rq
, long runnable_sum
) {}
3642 #endif /* CONFIG_FAIR_GROUP_SCHED */
3645 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3646 * @now: current time, as per cfs_rq_clock_pelt()
3647 * @cfs_rq: cfs_rq to update
3649 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3650 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3651 * post_init_entity_util_avg().
3653 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3655 * Returns true if the load decayed or we removed load.
3657 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3658 * call update_tg_load_avg() when this function returns true.
3661 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
)
3663 unsigned long removed_load
= 0, removed_util
= 0, removed_runnable
= 0;
3664 struct sched_avg
*sa
= &cfs_rq
->avg
;
3667 if (cfs_rq
->removed
.nr
) {
3669 u32 divider
= get_pelt_divider(&cfs_rq
->avg
);
3671 raw_spin_lock(&cfs_rq
->removed
.lock
);
3672 swap(cfs_rq
->removed
.util_avg
, removed_util
);
3673 swap(cfs_rq
->removed
.load_avg
, removed_load
);
3674 swap(cfs_rq
->removed
.runnable_avg
, removed_runnable
);
3675 cfs_rq
->removed
.nr
= 0;
3676 raw_spin_unlock(&cfs_rq
->removed
.lock
);
3679 sub_positive(&sa
->load_avg
, r
);
3680 sa
->load_sum
= sa
->load_avg
* divider
;
3683 sub_positive(&sa
->util_avg
, r
);
3684 sa
->util_sum
= sa
->util_avg
* divider
;
3686 r
= removed_runnable
;
3687 sub_positive(&sa
->runnable_avg
, r
);
3688 sa
->runnable_sum
= sa
->runnable_avg
* divider
;
3691 * removed_runnable is the unweighted version of removed_load so we
3692 * can use it to estimate removed_load_sum.
3694 add_tg_cfs_propagate(cfs_rq
,
3695 -(long)(removed_runnable
* divider
) >> SCHED_CAPACITY_SHIFT
);
3700 decayed
|= __update_load_avg_cfs_rq(now
, cfs_rq
);
3702 #ifndef CONFIG_64BIT
3704 cfs_rq
->load_last_update_time_copy
= sa
->last_update_time
;
3711 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3712 * @cfs_rq: cfs_rq to attach to
3713 * @se: sched_entity to attach
3715 * Must call update_cfs_rq_load_avg() before this, since we rely on
3716 * cfs_rq->avg.last_update_time being current.
3718 static void attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3721 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3722 * See ___update_load_avg() for details.
3724 u32 divider
= get_pelt_divider(&cfs_rq
->avg
);
3727 * When we attach the @se to the @cfs_rq, we must align the decay
3728 * window because without that, really weird and wonderful things can
3733 se
->avg
.last_update_time
= cfs_rq
->avg
.last_update_time
;
3734 se
->avg
.period_contrib
= cfs_rq
->avg
.period_contrib
;
3737 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3738 * period_contrib. This isn't strictly correct, but since we're
3739 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3742 se
->avg
.util_sum
= se
->avg
.util_avg
* divider
;
3744 se
->avg
.runnable_sum
= se
->avg
.runnable_avg
* divider
;
3746 se
->avg
.load_sum
= divider
;
3747 if (se_weight(se
)) {
3749 div_u64(se
->avg
.load_avg
* se
->avg
.load_sum
, se_weight(se
));
3752 enqueue_load_avg(cfs_rq
, se
);
3753 cfs_rq
->avg
.util_avg
+= se
->avg
.util_avg
;
3754 cfs_rq
->avg
.util_sum
+= se
->avg
.util_sum
;
3755 cfs_rq
->avg
.runnable_avg
+= se
->avg
.runnable_avg
;
3756 cfs_rq
->avg
.runnable_sum
+= se
->avg
.runnable_sum
;
3758 add_tg_cfs_propagate(cfs_rq
, se
->avg
.load_sum
);
3760 cfs_rq_util_change(cfs_rq
, 0);
3762 trace_pelt_cfs_tp(cfs_rq
);
3766 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3767 * @cfs_rq: cfs_rq to detach from
3768 * @se: sched_entity to detach
3770 * Must call update_cfs_rq_load_avg() before this, since we rely on
3771 * cfs_rq->avg.last_update_time being current.
3773 static void detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3776 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3777 * See ___update_load_avg() for details.
3779 u32 divider
= get_pelt_divider(&cfs_rq
->avg
);
3781 dequeue_load_avg(cfs_rq
, se
);
3782 sub_positive(&cfs_rq
->avg
.util_avg
, se
->avg
.util_avg
);
3783 cfs_rq
->avg
.util_sum
= cfs_rq
->avg
.util_avg
* divider
;
3784 sub_positive(&cfs_rq
->avg
.runnable_avg
, se
->avg
.runnable_avg
);
3785 cfs_rq
->avg
.runnable_sum
= cfs_rq
->avg
.runnable_avg
* divider
;
3787 add_tg_cfs_propagate(cfs_rq
, -se
->avg
.load_sum
);
3789 cfs_rq_util_change(cfs_rq
, 0);
3791 trace_pelt_cfs_tp(cfs_rq
);
3795 * Optional action to be done while updating the load average
3797 #define UPDATE_TG 0x1
3798 #define SKIP_AGE_LOAD 0x2
3799 #define DO_ATTACH 0x4
3801 /* Update task and its cfs_rq load average */
3802 static inline void update_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3804 u64 now
= cfs_rq_clock_pelt(cfs_rq
);
3808 * Track task load average for carrying it to new CPU after migrated, and
3809 * track group sched_entity load average for task_h_load calc in migration
3811 if (se
->avg
.last_update_time
&& !(flags
& SKIP_AGE_LOAD
))
3812 __update_load_avg_se(now
, cfs_rq
, se
);
3814 decayed
= update_cfs_rq_load_avg(now
, cfs_rq
);
3815 decayed
|= propagate_entity_load_avg(se
);
3817 if (!se
->avg
.last_update_time
&& (flags
& DO_ATTACH
)) {
3820 * DO_ATTACH means we're here from enqueue_entity().
3821 * !last_update_time means we've passed through
3822 * migrate_task_rq_fair() indicating we migrated.
3824 * IOW we're enqueueing a task on a new CPU.
3826 attach_entity_load_avg(cfs_rq
, se
);
3827 update_tg_load_avg(cfs_rq
);
3829 } else if (decayed
) {
3830 cfs_rq_util_change(cfs_rq
, 0);
3832 if (flags
& UPDATE_TG
)
3833 update_tg_load_avg(cfs_rq
);
3837 #ifndef CONFIG_64BIT
3838 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3840 u64 last_update_time_copy
;
3841 u64 last_update_time
;
3844 last_update_time_copy
= cfs_rq
->load_last_update_time_copy
;
3846 last_update_time
= cfs_rq
->avg
.last_update_time
;
3847 } while (last_update_time
!= last_update_time_copy
);
3849 return last_update_time
;
3852 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3854 return cfs_rq
->avg
.last_update_time
;
3859 * Synchronize entity load avg of dequeued entity without locking
3862 static void sync_entity_load_avg(struct sched_entity
*se
)
3864 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3865 u64 last_update_time
;
3867 last_update_time
= cfs_rq_last_update_time(cfs_rq
);
3868 __update_load_avg_blocked_se(last_update_time
, se
);
3872 * Task first catches up with cfs_rq, and then subtract
3873 * itself from the cfs_rq (task must be off the queue now).
3875 static void remove_entity_load_avg(struct sched_entity
*se
)
3877 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3878 unsigned long flags
;
3881 * tasks cannot exit without having gone through wake_up_new_task() ->
3882 * post_init_entity_util_avg() which will have added things to the
3883 * cfs_rq, so we can remove unconditionally.
3886 sync_entity_load_avg(se
);
3888 raw_spin_lock_irqsave(&cfs_rq
->removed
.lock
, flags
);
3889 ++cfs_rq
->removed
.nr
;
3890 cfs_rq
->removed
.util_avg
+= se
->avg
.util_avg
;
3891 cfs_rq
->removed
.load_avg
+= se
->avg
.load_avg
;
3892 cfs_rq
->removed
.runnable_avg
+= se
->avg
.runnable_avg
;
3893 raw_spin_unlock_irqrestore(&cfs_rq
->removed
.lock
, flags
);
3896 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq
*cfs_rq
)
3898 return cfs_rq
->avg
.runnable_avg
;
3901 static inline unsigned long cfs_rq_load_avg(struct cfs_rq
*cfs_rq
)
3903 return cfs_rq
->avg
.load_avg
;
3906 static int newidle_balance(struct rq
*this_rq
, struct rq_flags
*rf
);
3908 static inline unsigned long task_util(struct task_struct
*p
)
3910 return READ_ONCE(p
->se
.avg
.util_avg
);
3913 static inline unsigned long _task_util_est(struct task_struct
*p
)
3915 struct util_est ue
= READ_ONCE(p
->se
.avg
.util_est
);
3917 return max(ue
.ewma
, (ue
.enqueued
& ~UTIL_AVG_UNCHANGED
));
3920 static inline unsigned long task_util_est(struct task_struct
*p
)
3922 return max(task_util(p
), _task_util_est(p
));
3925 #ifdef CONFIG_UCLAMP_TASK
3926 static inline unsigned long uclamp_task_util(struct task_struct
*p
)
3928 return clamp(task_util_est(p
),
3929 uclamp_eff_value(p
, UCLAMP_MIN
),
3930 uclamp_eff_value(p
, UCLAMP_MAX
));
3933 static inline unsigned long uclamp_task_util(struct task_struct
*p
)
3935 return task_util_est(p
);
3939 static inline void util_est_enqueue(struct cfs_rq
*cfs_rq
,
3940 struct task_struct
*p
)
3942 unsigned int enqueued
;
3944 if (!sched_feat(UTIL_EST
))
3947 /* Update root cfs_rq's estimated utilization */
3948 enqueued
= cfs_rq
->avg
.util_est
.enqueued
;
3949 enqueued
+= _task_util_est(p
);
3950 WRITE_ONCE(cfs_rq
->avg
.util_est
.enqueued
, enqueued
);
3952 trace_sched_util_est_cfs_tp(cfs_rq
);
3955 static inline void util_est_dequeue(struct cfs_rq
*cfs_rq
,
3956 struct task_struct
*p
)
3958 unsigned int enqueued
;
3960 if (!sched_feat(UTIL_EST
))
3963 /* Update root cfs_rq's estimated utilization */
3964 enqueued
= cfs_rq
->avg
.util_est
.enqueued
;
3965 enqueued
-= min_t(unsigned int, enqueued
, _task_util_est(p
));
3966 WRITE_ONCE(cfs_rq
->avg
.util_est
.enqueued
, enqueued
);
3968 trace_sched_util_est_cfs_tp(cfs_rq
);
3971 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
3974 * Check if a (signed) value is within a specified (unsigned) margin,
3975 * based on the observation that:
3977 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3979 * NOTE: this only works when value + margin < INT_MAX.
3981 static inline bool within_margin(int value
, int margin
)
3983 return ((unsigned int)(value
+ margin
- 1) < (2 * margin
- 1));
3986 static inline void util_est_update(struct cfs_rq
*cfs_rq
,
3987 struct task_struct
*p
,
3990 long last_ewma_diff
, last_enqueued_diff
;
3993 if (!sched_feat(UTIL_EST
))
3997 * Skip update of task's estimated utilization when the task has not
3998 * yet completed an activation, e.g. being migrated.
4004 * If the PELT values haven't changed since enqueue time,
4005 * skip the util_est update.
4007 ue
= p
->se
.avg
.util_est
;
4008 if (ue
.enqueued
& UTIL_AVG_UNCHANGED
)
4011 last_enqueued_diff
= ue
.enqueued
;
4014 * Reset EWMA on utilization increases, the moving average is used only
4015 * to smooth utilization decreases.
4017 ue
.enqueued
= task_util(p
);
4018 if (sched_feat(UTIL_EST_FASTUP
)) {
4019 if (ue
.ewma
< ue
.enqueued
) {
4020 ue
.ewma
= ue
.enqueued
;
4026 * Skip update of task's estimated utilization when its members are
4027 * already ~1% close to its last activation value.
4029 last_ewma_diff
= ue
.enqueued
- ue
.ewma
;
4030 last_enqueued_diff
-= ue
.enqueued
;
4031 if (within_margin(last_ewma_diff
, UTIL_EST_MARGIN
)) {
4032 if (!within_margin(last_enqueued_diff
, UTIL_EST_MARGIN
))
4039 * To avoid overestimation of actual task utilization, skip updates if
4040 * we cannot grant there is idle time in this CPU.
4042 if (task_util(p
) > capacity_orig_of(cpu_of(rq_of(cfs_rq
))))
4046 * Update Task's estimated utilization
4048 * When *p completes an activation we can consolidate another sample
4049 * of the task size. This is done by storing the current PELT value
4050 * as ue.enqueued and by using this value to update the Exponential
4051 * Weighted Moving Average (EWMA):
4053 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4054 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4055 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4056 * = w * ( last_ewma_diff ) + ewma(t-1)
4057 * = w * (last_ewma_diff + ewma(t-1) / w)
4059 * Where 'w' is the weight of new samples, which is configured to be
4060 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4062 ue
.ewma
<<= UTIL_EST_WEIGHT_SHIFT
;
4063 ue
.ewma
+= last_ewma_diff
;
4064 ue
.ewma
>>= UTIL_EST_WEIGHT_SHIFT
;
4066 ue
.enqueued
|= UTIL_AVG_UNCHANGED
;
4067 WRITE_ONCE(p
->se
.avg
.util_est
, ue
);
4069 trace_sched_util_est_se_tp(&p
->se
);
4072 static inline int task_fits_capacity(struct task_struct
*p
,
4073 unsigned long capacity
)
4075 return fits_capacity(uclamp_task_util(p
), capacity
);
4078 static inline void update_misfit_status(struct task_struct
*p
, struct rq
*rq
)
4080 if (!static_branch_unlikely(&sched_asym_cpucapacity
))
4083 if (!p
|| p
->nr_cpus_allowed
== 1) {
4084 rq
->misfit_task_load
= 0;
4088 if (task_fits_capacity(p
, capacity_of(cpu_of(rq
)))) {
4089 rq
->misfit_task_load
= 0;
4094 * Make sure that misfit_task_load will not be null even if
4095 * task_h_load() returns 0.
4097 rq
->misfit_task_load
= max_t(unsigned long, task_h_load(p
), 1);
4100 #else /* CONFIG_SMP */
4102 static inline bool cfs_rq_is_decayed(struct cfs_rq
*cfs_rq
)
4107 #define UPDATE_TG 0x0
4108 #define SKIP_AGE_LOAD 0x0
4109 #define DO_ATTACH 0x0
4111 static inline void update_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int not_used1
)
4113 cfs_rq_util_change(cfs_rq
, 0);
4116 static inline void remove_entity_load_avg(struct sched_entity
*se
) {}
4119 attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
4121 detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
4123 static inline int newidle_balance(struct rq
*rq
, struct rq_flags
*rf
)
4129 util_est_enqueue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
) {}
4132 util_est_dequeue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
) {}
4135 util_est_update(struct cfs_rq
*cfs_rq
, struct task_struct
*p
,
4137 static inline void update_misfit_status(struct task_struct
*p
, struct rq
*rq
) {}
4139 #endif /* CONFIG_SMP */
4141 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4143 #ifdef CONFIG_SCHED_DEBUG
4144 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
4149 if (d
> 3*sysctl_sched_latency
)
4150 schedstat_inc(cfs_rq
->nr_spread_over
);
4155 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
4157 u64 vruntime
= cfs_rq
->min_vruntime
;
4160 * The 'current' period is already promised to the current tasks,
4161 * however the extra weight of the new task will slow them down a
4162 * little, place the new task so that it fits in the slot that
4163 * stays open at the end.
4165 if (initial
&& sched_feat(START_DEBIT
))
4166 vruntime
+= sched_vslice(cfs_rq
, se
);
4168 /* sleeps up to a single latency don't count. */
4170 unsigned long thresh
;
4173 thresh
= sysctl_sched_min_granularity
;
4175 thresh
= sysctl_sched_latency
;
4178 * Halve their sleep time's effect, to allow
4179 * for a gentler effect of sleepers:
4181 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
4187 /* ensure we never gain time by being placed backwards. */
4188 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
4191 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
4193 static inline bool cfs_bandwidth_used(void);
4200 * update_min_vruntime()
4201 * vruntime -= min_vruntime
4205 * update_min_vruntime()
4206 * vruntime += min_vruntime
4208 * this way the vruntime transition between RQs is done when both
4209 * min_vruntime are up-to-date.
4213 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4214 * vruntime -= min_vruntime
4218 * update_min_vruntime()
4219 * vruntime += min_vruntime
4221 * this way we don't have the most up-to-date min_vruntime on the originating
4222 * CPU and an up-to-date min_vruntime on the destination CPU.
4226 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
4228 bool renorm
= !(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_MIGRATED
);
4229 bool curr
= cfs_rq
->curr
== se
;
4232 * If we're the current task, we must renormalise before calling
4236 se
->vruntime
+= cfs_rq
->min_vruntime
;
4238 update_curr(cfs_rq
);
4241 * Otherwise, renormalise after, such that we're placed at the current
4242 * moment in time, instead of some random moment in the past. Being
4243 * placed in the past could significantly boost this task to the
4244 * fairness detriment of existing tasks.
4246 if (renorm
&& !curr
)
4247 se
->vruntime
+= cfs_rq
->min_vruntime
;
4250 * When enqueuing a sched_entity, we must:
4251 * - Update loads to have both entity and cfs_rq synced with now.
4252 * - Add its load to cfs_rq->runnable_avg
4253 * - For group_entity, update its weight to reflect the new share of
4255 * - Add its new weight to cfs_rq->load.weight
4257 update_load_avg(cfs_rq
, se
, UPDATE_TG
| DO_ATTACH
);
4258 se_update_runnable(se
);
4259 update_cfs_group(se
);
4260 account_entity_enqueue(cfs_rq
, se
);
4262 if (flags
& ENQUEUE_WAKEUP
)
4263 place_entity(cfs_rq
, se
, 0);
4265 check_schedstat_required();
4266 update_stats_enqueue_fair(cfs_rq
, se
, flags
);
4267 check_spread(cfs_rq
, se
);
4269 __enqueue_entity(cfs_rq
, se
);
4273 * When bandwidth control is enabled, cfs might have been removed
4274 * because of a parent been throttled but cfs->nr_running > 1. Try to
4275 * add it unconditionally.
4277 if (cfs_rq
->nr_running
== 1 || cfs_bandwidth_used())
4278 list_add_leaf_cfs_rq(cfs_rq
);
4280 if (cfs_rq
->nr_running
== 1)
4281 check_enqueue_throttle(cfs_rq
);
4284 static void __clear_buddies_last(struct sched_entity
*se
)
4286 for_each_sched_entity(se
) {
4287 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4288 if (cfs_rq
->last
!= se
)
4291 cfs_rq
->last
= NULL
;
4295 static void __clear_buddies_next(struct sched_entity
*se
)
4297 for_each_sched_entity(se
) {
4298 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4299 if (cfs_rq
->next
!= se
)
4302 cfs_rq
->next
= NULL
;
4306 static void __clear_buddies_skip(struct sched_entity
*se
)
4308 for_each_sched_entity(se
) {
4309 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4310 if (cfs_rq
->skip
!= se
)
4313 cfs_rq
->skip
= NULL
;
4317 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4319 if (cfs_rq
->last
== se
)
4320 __clear_buddies_last(se
);
4322 if (cfs_rq
->next
== se
)
4323 __clear_buddies_next(se
);
4325 if (cfs_rq
->skip
== se
)
4326 __clear_buddies_skip(se
);
4329 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
4332 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
4335 * Update run-time statistics of the 'current'.
4337 update_curr(cfs_rq
);
4340 * When dequeuing a sched_entity, we must:
4341 * - Update loads to have both entity and cfs_rq synced with now.
4342 * - Subtract its load from the cfs_rq->runnable_avg.
4343 * - Subtract its previous weight from cfs_rq->load.weight.
4344 * - For group entity, update its weight to reflect the new share
4345 * of its group cfs_rq.
4347 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4348 se_update_runnable(se
);
4350 update_stats_dequeue_fair(cfs_rq
, se
, flags
);
4352 clear_buddies(cfs_rq
, se
);
4354 if (se
!= cfs_rq
->curr
)
4355 __dequeue_entity(cfs_rq
, se
);
4357 account_entity_dequeue(cfs_rq
, se
);
4360 * Normalize after update_curr(); which will also have moved
4361 * min_vruntime if @se is the one holding it back. But before doing
4362 * update_min_vruntime() again, which will discount @se's position and
4363 * can move min_vruntime forward still more.
4365 if (!(flags
& DEQUEUE_SLEEP
))
4366 se
->vruntime
-= cfs_rq
->min_vruntime
;
4368 /* return excess runtime on last dequeue */
4369 return_cfs_rq_runtime(cfs_rq
);
4371 update_cfs_group(se
);
4374 * Now advance min_vruntime if @se was the entity holding it back,
4375 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4376 * put back on, and if we advance min_vruntime, we'll be placed back
4377 * further than we started -- ie. we'll be penalized.
4379 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) != DEQUEUE_SAVE
)
4380 update_min_vruntime(cfs_rq
);
4384 * Preempt the current task with a newly woken task if needed:
4387 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
4389 unsigned long ideal_runtime
, delta_exec
;
4390 struct sched_entity
*se
;
4393 ideal_runtime
= sched_slice(cfs_rq
, curr
);
4394 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
4395 if (delta_exec
> ideal_runtime
) {
4396 resched_curr(rq_of(cfs_rq
));
4398 * The current task ran long enough, ensure it doesn't get
4399 * re-elected due to buddy favours.
4401 clear_buddies(cfs_rq
, curr
);
4406 * Ensure that a task that missed wakeup preemption by a
4407 * narrow margin doesn't have to wait for a full slice.
4408 * This also mitigates buddy induced latencies under load.
4410 if (delta_exec
< sysctl_sched_min_granularity
)
4413 se
= __pick_first_entity(cfs_rq
);
4414 delta
= curr
->vruntime
- se
->vruntime
;
4419 if (delta
> ideal_runtime
)
4420 resched_curr(rq_of(cfs_rq
));
4424 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4426 clear_buddies(cfs_rq
, se
);
4428 /* 'current' is not kept within the tree. */
4431 * Any task has to be enqueued before it get to execute on
4432 * a CPU. So account for the time it spent waiting on the
4435 update_stats_wait_end_fair(cfs_rq
, se
);
4436 __dequeue_entity(cfs_rq
, se
);
4437 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4440 update_stats_curr_start(cfs_rq
, se
);
4444 * Track our maximum slice length, if the CPU's load is at
4445 * least twice that of our own weight (i.e. dont track it
4446 * when there are only lesser-weight tasks around):
4448 if (schedstat_enabled() &&
4449 rq_of(cfs_rq
)->cfs
.load
.weight
>= 2*se
->load
.weight
) {
4450 struct sched_statistics
*stats
;
4452 stats
= __schedstats_from_se(se
);
4453 __schedstat_set(stats
->slice_max
,
4454 max((u64
)stats
->slice_max
,
4455 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
));
4458 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
4462 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
4465 * Pick the next process, keeping these things in mind, in this order:
4466 * 1) keep things fair between processes/task groups
4467 * 2) pick the "next" process, since someone really wants that to run
4468 * 3) pick the "last" process, for cache locality
4469 * 4) do not run the "skip" process, if something else is available
4471 static struct sched_entity
*
4472 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
4474 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
4475 struct sched_entity
*se
;
4478 * If curr is set we have to see if its left of the leftmost entity
4479 * still in the tree, provided there was anything in the tree at all.
4481 if (!left
|| (curr
&& entity_before(curr
, left
)))
4484 se
= left
; /* ideally we run the leftmost entity */
4487 * Avoid running the skip buddy, if running something else can
4488 * be done without getting too unfair.
4490 if (cfs_rq
->skip
&& cfs_rq
->skip
== se
) {
4491 struct sched_entity
*second
;
4494 second
= __pick_first_entity(cfs_rq
);
4496 second
= __pick_next_entity(se
);
4497 if (!second
|| (curr
&& entity_before(curr
, second
)))
4501 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
4505 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1) {
4507 * Someone really wants this to run. If it's not unfair, run it.
4510 } else if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1) {
4512 * Prefer last buddy, try to return the CPU to a preempted task.
4520 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
4522 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
4525 * If still on the runqueue then deactivate_task()
4526 * was not called and update_curr() has to be done:
4529 update_curr(cfs_rq
);
4531 /* throttle cfs_rqs exceeding runtime */
4532 check_cfs_rq_runtime(cfs_rq
);
4534 check_spread(cfs_rq
, prev
);
4537 update_stats_wait_start_fair(cfs_rq
, prev
);
4538 /* Put 'current' back into the tree. */
4539 __enqueue_entity(cfs_rq
, prev
);
4540 /* in !on_rq case, update occurred at dequeue */
4541 update_load_avg(cfs_rq
, prev
, 0);
4543 cfs_rq
->curr
= NULL
;
4547 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
4550 * Update run-time statistics of the 'current'.
4552 update_curr(cfs_rq
);
4555 * Ensure that runnable average is periodically updated.
4557 update_load_avg(cfs_rq
, curr
, UPDATE_TG
);
4558 update_cfs_group(curr
);
4560 #ifdef CONFIG_SCHED_HRTICK
4562 * queued ticks are scheduled to match the slice, so don't bother
4563 * validating it and just reschedule.
4566 resched_curr(rq_of(cfs_rq
));
4570 * don't let the period tick interfere with the hrtick preemption
4572 if (!sched_feat(DOUBLE_TICK
) &&
4573 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
4577 if (cfs_rq
->nr_running
> 1)
4578 check_preempt_tick(cfs_rq
, curr
);
4582 /**************************************************
4583 * CFS bandwidth control machinery
4586 #ifdef CONFIG_CFS_BANDWIDTH
4588 #ifdef CONFIG_JUMP_LABEL
4589 static struct static_key __cfs_bandwidth_used
;
4591 static inline bool cfs_bandwidth_used(void)
4593 return static_key_false(&__cfs_bandwidth_used
);
4596 void cfs_bandwidth_usage_inc(void)
4598 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used
);
4601 void cfs_bandwidth_usage_dec(void)
4603 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used
);
4605 #else /* CONFIG_JUMP_LABEL */
4606 static bool cfs_bandwidth_used(void)
4611 void cfs_bandwidth_usage_inc(void) {}
4612 void cfs_bandwidth_usage_dec(void) {}
4613 #endif /* CONFIG_JUMP_LABEL */
4616 * default period for cfs group bandwidth.
4617 * default: 0.1s, units: nanoseconds
4619 static inline u64
default_cfs_period(void)
4621 return 100000000ULL;
4624 static inline u64
sched_cfs_bandwidth_slice(void)
4626 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
4630 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4631 * directly instead of rq->clock to avoid adding additional synchronization
4634 * requires cfs_b->lock
4636 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
4640 if (unlikely(cfs_b
->quota
== RUNTIME_INF
))
4643 cfs_b
->runtime
+= cfs_b
->quota
;
4644 runtime
= cfs_b
->runtime_snap
- cfs_b
->runtime
;
4646 cfs_b
->burst_time
+= runtime
;
4650 cfs_b
->runtime
= min(cfs_b
->runtime
, cfs_b
->quota
+ cfs_b
->burst
);
4651 cfs_b
->runtime_snap
= cfs_b
->runtime
;
4654 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4656 return &tg
->cfs_bandwidth
;
4659 /* returns 0 on failure to allocate runtime */
4660 static int __assign_cfs_rq_runtime(struct cfs_bandwidth
*cfs_b
,
4661 struct cfs_rq
*cfs_rq
, u64 target_runtime
)
4663 u64 min_amount
, amount
= 0;
4665 lockdep_assert_held(&cfs_b
->lock
);
4667 /* note: this is a positive sum as runtime_remaining <= 0 */
4668 min_amount
= target_runtime
- cfs_rq
->runtime_remaining
;
4670 if (cfs_b
->quota
== RUNTIME_INF
)
4671 amount
= min_amount
;
4673 start_cfs_bandwidth(cfs_b
);
4675 if (cfs_b
->runtime
> 0) {
4676 amount
= min(cfs_b
->runtime
, min_amount
);
4677 cfs_b
->runtime
-= amount
;
4682 cfs_rq
->runtime_remaining
+= amount
;
4684 return cfs_rq
->runtime_remaining
> 0;
4687 /* returns 0 on failure to allocate runtime */
4688 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4690 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4693 raw_spin_lock(&cfs_b
->lock
);
4694 ret
= __assign_cfs_rq_runtime(cfs_b
, cfs_rq
, sched_cfs_bandwidth_slice());
4695 raw_spin_unlock(&cfs_b
->lock
);
4700 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4702 /* dock delta_exec before expiring quota (as it could span periods) */
4703 cfs_rq
->runtime_remaining
-= delta_exec
;
4705 if (likely(cfs_rq
->runtime_remaining
> 0))
4708 if (cfs_rq
->throttled
)
4711 * if we're unable to extend our runtime we resched so that the active
4712 * hierarchy can be throttled
4714 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
4715 resched_curr(rq_of(cfs_rq
));
4718 static __always_inline
4719 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4721 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
4724 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
4727 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4729 return cfs_bandwidth_used() && cfs_rq
->throttled
;
4732 /* check whether cfs_rq, or any parent, is throttled */
4733 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4735 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
4739 * Ensure that neither of the group entities corresponding to src_cpu or
4740 * dest_cpu are members of a throttled hierarchy when performing group
4741 * load-balance operations.
4743 static inline int throttled_lb_pair(struct task_group
*tg
,
4744 int src_cpu
, int dest_cpu
)
4746 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
4748 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
4749 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
4751 return throttled_hierarchy(src_cfs_rq
) ||
4752 throttled_hierarchy(dest_cfs_rq
);
4755 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
4757 struct rq
*rq
= data
;
4758 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4760 cfs_rq
->throttle_count
--;
4761 if (!cfs_rq
->throttle_count
) {
4762 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
4763 cfs_rq
->throttled_clock_task
;
4765 /* Add cfs_rq with load or one or more already running entities to the list */
4766 if (!cfs_rq_is_decayed(cfs_rq
) || cfs_rq
->nr_running
)
4767 list_add_leaf_cfs_rq(cfs_rq
);
4773 static int tg_throttle_down(struct task_group
*tg
, void *data
)
4775 struct rq
*rq
= data
;
4776 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4778 /* group is entering throttled state, stop time */
4779 if (!cfs_rq
->throttle_count
) {
4780 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
4781 list_del_leaf_cfs_rq(cfs_rq
);
4783 cfs_rq
->throttle_count
++;
4788 static bool throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
4790 struct rq
*rq
= rq_of(cfs_rq
);
4791 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4792 struct sched_entity
*se
;
4793 long task_delta
, idle_task_delta
, dequeue
= 1;
4795 raw_spin_lock(&cfs_b
->lock
);
4796 /* This will start the period timer if necessary */
4797 if (__assign_cfs_rq_runtime(cfs_b
, cfs_rq
, 1)) {
4799 * We have raced with bandwidth becoming available, and if we
4800 * actually throttled the timer might not unthrottle us for an
4801 * entire period. We additionally needed to make sure that any
4802 * subsequent check_cfs_rq_runtime calls agree not to throttle
4803 * us, as we may commit to do cfs put_prev+pick_next, so we ask
4804 * for 1ns of runtime rather than just check cfs_b.
4808 list_add_tail_rcu(&cfs_rq
->throttled_list
,
4809 &cfs_b
->throttled_cfs_rq
);
4811 raw_spin_unlock(&cfs_b
->lock
);
4814 return false; /* Throttle no longer required. */
4816 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
4818 /* freeze hierarchy runnable averages while throttled */
4820 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
4823 task_delta
= cfs_rq
->h_nr_running
;
4824 idle_task_delta
= cfs_rq
->idle_h_nr_running
;
4825 for_each_sched_entity(se
) {
4826 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4827 /* throttled entity or throttle-on-deactivate */
4831 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
4833 if (cfs_rq_is_idle(group_cfs_rq(se
)))
4834 idle_task_delta
= cfs_rq
->h_nr_running
;
4836 qcfs_rq
->h_nr_running
-= task_delta
;
4837 qcfs_rq
->idle_h_nr_running
-= idle_task_delta
;
4839 if (qcfs_rq
->load
.weight
) {
4840 /* Avoid re-evaluating load for this entity: */
4841 se
= parent_entity(se
);
4846 for_each_sched_entity(se
) {
4847 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4848 /* throttled entity or throttle-on-deactivate */
4852 update_load_avg(qcfs_rq
, se
, 0);
4853 se_update_runnable(se
);
4855 if (cfs_rq_is_idle(group_cfs_rq(se
)))
4856 idle_task_delta
= cfs_rq
->h_nr_running
;
4858 qcfs_rq
->h_nr_running
-= task_delta
;
4859 qcfs_rq
->idle_h_nr_running
-= idle_task_delta
;
4862 /* At this point se is NULL and we are at root level*/
4863 sub_nr_running(rq
, task_delta
);
4867 * Note: distribution will already see us throttled via the
4868 * throttled-list. rq->lock protects completion.
4870 cfs_rq
->throttled
= 1;
4871 cfs_rq
->throttled_clock
= rq_clock(rq
);
4875 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
4877 struct rq
*rq
= rq_of(cfs_rq
);
4878 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4879 struct sched_entity
*se
;
4880 long task_delta
, idle_task_delta
;
4882 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
4884 cfs_rq
->throttled
= 0;
4886 update_rq_clock(rq
);
4888 raw_spin_lock(&cfs_b
->lock
);
4889 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
4890 list_del_rcu(&cfs_rq
->throttled_list
);
4891 raw_spin_unlock(&cfs_b
->lock
);
4893 /* update hierarchical throttle state */
4894 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
4896 /* Nothing to run but something to decay (on_list)? Complete the branch */
4897 if (!cfs_rq
->load
.weight
) {
4898 if (cfs_rq
->on_list
)
4899 goto unthrottle_throttle
;
4903 task_delta
= cfs_rq
->h_nr_running
;
4904 idle_task_delta
= cfs_rq
->idle_h_nr_running
;
4905 for_each_sched_entity(se
) {
4906 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4910 enqueue_entity(qcfs_rq
, se
, ENQUEUE_WAKEUP
);
4912 if (cfs_rq_is_idle(group_cfs_rq(se
)))
4913 idle_task_delta
= cfs_rq
->h_nr_running
;
4915 qcfs_rq
->h_nr_running
+= task_delta
;
4916 qcfs_rq
->idle_h_nr_running
+= idle_task_delta
;
4918 /* end evaluation on encountering a throttled cfs_rq */
4919 if (cfs_rq_throttled(qcfs_rq
))
4920 goto unthrottle_throttle
;
4923 for_each_sched_entity(se
) {
4924 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4926 update_load_avg(qcfs_rq
, se
, UPDATE_TG
);
4927 se_update_runnable(se
);
4929 if (cfs_rq_is_idle(group_cfs_rq(se
)))
4930 idle_task_delta
= cfs_rq
->h_nr_running
;
4932 qcfs_rq
->h_nr_running
+= task_delta
;
4933 qcfs_rq
->idle_h_nr_running
+= idle_task_delta
;
4935 /* end evaluation on encountering a throttled cfs_rq */
4936 if (cfs_rq_throttled(qcfs_rq
))
4937 goto unthrottle_throttle
;
4940 * One parent has been throttled and cfs_rq removed from the
4941 * list. Add it back to not break the leaf list.
4943 if (throttled_hierarchy(qcfs_rq
))
4944 list_add_leaf_cfs_rq(qcfs_rq
);
4947 /* At this point se is NULL and we are at root level*/
4948 add_nr_running(rq
, task_delta
);
4950 unthrottle_throttle
:
4952 * The cfs_rq_throttled() breaks in the above iteration can result in
4953 * incomplete leaf list maintenance, resulting in triggering the
4956 for_each_sched_entity(se
) {
4957 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4959 if (list_add_leaf_cfs_rq(qcfs_rq
))
4963 assert_list_leaf_cfs_rq(rq
);
4965 /* Determine whether we need to wake up potentially idle CPU: */
4966 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
4970 static void distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
)
4972 struct cfs_rq
*cfs_rq
;
4973 u64 runtime
, remaining
= 1;
4976 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
4978 struct rq
*rq
= rq_of(cfs_rq
);
4981 rq_lock_irqsave(rq
, &rf
);
4982 if (!cfs_rq_throttled(cfs_rq
))
4985 /* By the above check, this should never be true */
4986 SCHED_WARN_ON(cfs_rq
->runtime_remaining
> 0);
4988 raw_spin_lock(&cfs_b
->lock
);
4989 runtime
= -cfs_rq
->runtime_remaining
+ 1;
4990 if (runtime
> cfs_b
->runtime
)
4991 runtime
= cfs_b
->runtime
;
4992 cfs_b
->runtime
-= runtime
;
4993 remaining
= cfs_b
->runtime
;
4994 raw_spin_unlock(&cfs_b
->lock
);
4996 cfs_rq
->runtime_remaining
+= runtime
;
4998 /* we check whether we're throttled above */
4999 if (cfs_rq
->runtime_remaining
> 0)
5000 unthrottle_cfs_rq(cfs_rq
);
5003 rq_unlock_irqrestore(rq
, &rf
);
5012 * Responsible for refilling a task_group's bandwidth and unthrottling its
5013 * cfs_rqs as appropriate. If there has been no activity within the last
5014 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5015 * used to track this state.
5017 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
, unsigned long flags
)
5021 /* no need to continue the timer with no bandwidth constraint */
5022 if (cfs_b
->quota
== RUNTIME_INF
)
5023 goto out_deactivate
;
5025 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
5026 cfs_b
->nr_periods
+= overrun
;
5028 /* Refill extra burst quota even if cfs_b->idle */
5029 __refill_cfs_bandwidth_runtime(cfs_b
);
5032 * idle depends on !throttled (for the case of a large deficit), and if
5033 * we're going inactive then everything else can be deferred
5035 if (cfs_b
->idle
&& !throttled
)
5036 goto out_deactivate
;
5039 /* mark as potentially idle for the upcoming period */
5044 /* account preceding periods in which throttling occurred */
5045 cfs_b
->nr_throttled
+= overrun
;
5048 * This check is repeated as we release cfs_b->lock while we unthrottle.
5050 while (throttled
&& cfs_b
->runtime
> 0) {
5051 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5052 /* we can't nest cfs_b->lock while distributing bandwidth */
5053 distribute_cfs_runtime(cfs_b
);
5054 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5056 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
5060 * While we are ensured activity in the period following an
5061 * unthrottle, this also covers the case in which the new bandwidth is
5062 * insufficient to cover the existing bandwidth deficit. (Forcing the
5063 * timer to remain active while there are any throttled entities.)
5073 /* a cfs_rq won't donate quota below this amount */
5074 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
5075 /* minimum remaining period time to redistribute slack quota */
5076 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
5077 /* how long we wait to gather additional slack before distributing */
5078 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
5081 * Are we near the end of the current quota period?
5083 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5084 * hrtimer base being cleared by hrtimer_start. In the case of
5085 * migrate_hrtimers, base is never cleared, so we are fine.
5087 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
5089 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
5092 /* if the call-back is running a quota refresh is already occurring */
5093 if (hrtimer_callback_running(refresh_timer
))
5096 /* is a quota refresh about to occur? */
5097 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
5098 if (remaining
< (s64
)min_expire
)
5104 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
5106 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
5108 /* if there's a quota refresh soon don't bother with slack */
5109 if (runtime_refresh_within(cfs_b
, min_left
))
5112 /* don't push forwards an existing deferred unthrottle */
5113 if (cfs_b
->slack_started
)
5115 cfs_b
->slack_started
= true;
5117 hrtimer_start(&cfs_b
->slack_timer
,
5118 ns_to_ktime(cfs_bandwidth_slack_period
),
5122 /* we know any runtime found here is valid as update_curr() precedes return */
5123 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5125 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
5126 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
5128 if (slack_runtime
<= 0)
5131 raw_spin_lock(&cfs_b
->lock
);
5132 if (cfs_b
->quota
!= RUNTIME_INF
) {
5133 cfs_b
->runtime
+= slack_runtime
;
5135 /* we are under rq->lock, defer unthrottling using a timer */
5136 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
5137 !list_empty(&cfs_b
->throttled_cfs_rq
))
5138 start_cfs_slack_bandwidth(cfs_b
);
5140 raw_spin_unlock(&cfs_b
->lock
);
5142 /* even if it's not valid for return we don't want to try again */
5143 cfs_rq
->runtime_remaining
-= slack_runtime
;
5146 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5148 if (!cfs_bandwidth_used())
5151 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
5154 __return_cfs_rq_runtime(cfs_rq
);
5158 * This is done with a timer (instead of inline with bandwidth return) since
5159 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5161 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
5163 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
5164 unsigned long flags
;
5166 /* confirm we're still not at a refresh boundary */
5167 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5168 cfs_b
->slack_started
= false;
5170 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
5171 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5175 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
5176 runtime
= cfs_b
->runtime
;
5178 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5183 distribute_cfs_runtime(cfs_b
);
5187 * When a group wakes up we want to make sure that its quota is not already
5188 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5189 * runtime as update_curr() throttling can not trigger until it's on-rq.
5191 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
5193 if (!cfs_bandwidth_used())
5196 /* an active group must be handled by the update_curr()->put() path */
5197 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
5200 /* ensure the group is not already throttled */
5201 if (cfs_rq_throttled(cfs_rq
))
5204 /* update runtime allocation */
5205 account_cfs_rq_runtime(cfs_rq
, 0);
5206 if (cfs_rq
->runtime_remaining
<= 0)
5207 throttle_cfs_rq(cfs_rq
);
5210 static void sync_throttle(struct task_group
*tg
, int cpu
)
5212 struct cfs_rq
*pcfs_rq
, *cfs_rq
;
5214 if (!cfs_bandwidth_used())
5220 cfs_rq
= tg
->cfs_rq
[cpu
];
5221 pcfs_rq
= tg
->parent
->cfs_rq
[cpu
];
5223 cfs_rq
->throttle_count
= pcfs_rq
->throttle_count
;
5224 cfs_rq
->throttled_clock_task
= rq_clock_task(cpu_rq(cpu
));
5227 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5228 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5230 if (!cfs_bandwidth_used())
5233 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
5237 * it's possible for a throttled entity to be forced into a running
5238 * state (e.g. set_curr_task), in this case we're finished.
5240 if (cfs_rq_throttled(cfs_rq
))
5243 return throttle_cfs_rq(cfs_rq
);
5246 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
5248 struct cfs_bandwidth
*cfs_b
=
5249 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
5251 do_sched_cfs_slack_timer(cfs_b
);
5253 return HRTIMER_NORESTART
;
5256 extern const u64 max_cfs_quota_period
;
5258 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
5260 struct cfs_bandwidth
*cfs_b
=
5261 container_of(timer
, struct cfs_bandwidth
, period_timer
);
5262 unsigned long flags
;
5267 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5269 overrun
= hrtimer_forward_now(timer
, cfs_b
->period
);
5273 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
, flags
);
5276 u64
new, old
= ktime_to_ns(cfs_b
->period
);
5279 * Grow period by a factor of 2 to avoid losing precision.
5280 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5284 if (new < max_cfs_quota_period
) {
5285 cfs_b
->period
= ns_to_ktime(new);
5289 pr_warn_ratelimited(
5290 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5292 div_u64(new, NSEC_PER_USEC
),
5293 div_u64(cfs_b
->quota
, NSEC_PER_USEC
));
5295 pr_warn_ratelimited(
5296 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5298 div_u64(old
, NSEC_PER_USEC
),
5299 div_u64(cfs_b
->quota
, NSEC_PER_USEC
));
5302 /* reset count so we don't come right back in here */
5307 cfs_b
->period_active
= 0;
5308 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5310 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
5313 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5315 raw_spin_lock_init(&cfs_b
->lock
);
5317 cfs_b
->quota
= RUNTIME_INF
;
5318 cfs_b
->period
= ns_to_ktime(default_cfs_period());
5321 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
5322 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_PINNED
);
5323 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
5324 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5325 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
5326 cfs_b
->slack_started
= false;
5329 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5331 cfs_rq
->runtime_enabled
= 0;
5332 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
5335 void start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5337 lockdep_assert_held(&cfs_b
->lock
);
5339 if (cfs_b
->period_active
)
5342 cfs_b
->period_active
= 1;
5343 hrtimer_forward_now(&cfs_b
->period_timer
, cfs_b
->period
);
5344 hrtimer_start_expires(&cfs_b
->period_timer
, HRTIMER_MODE_ABS_PINNED
);
5347 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5349 /* init_cfs_bandwidth() was not called */
5350 if (!cfs_b
->throttled_cfs_rq
.next
)
5353 hrtimer_cancel(&cfs_b
->period_timer
);
5354 hrtimer_cancel(&cfs_b
->slack_timer
);
5358 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5360 * The race is harmless, since modifying bandwidth settings of unhooked group
5361 * bits doesn't do much.
5364 /* cpu online callback */
5365 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
5367 struct task_group
*tg
;
5369 lockdep_assert_rq_held(rq
);
5372 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
5373 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
5374 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
5376 raw_spin_lock(&cfs_b
->lock
);
5377 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
5378 raw_spin_unlock(&cfs_b
->lock
);
5383 /* cpu offline callback */
5384 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
5386 struct task_group
*tg
;
5388 lockdep_assert_rq_held(rq
);
5391 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
5392 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
5394 if (!cfs_rq
->runtime_enabled
)
5398 * clock_task is not advancing so we just need to make sure
5399 * there's some valid quota amount
5401 cfs_rq
->runtime_remaining
= 1;
5403 * Offline rq is schedulable till CPU is completely disabled
5404 * in take_cpu_down(), so we prevent new cfs throttling here.
5406 cfs_rq
->runtime_enabled
= 0;
5408 if (cfs_rq_throttled(cfs_rq
))
5409 unthrottle_cfs_rq(cfs_rq
);
5414 #else /* CONFIG_CFS_BANDWIDTH */
5416 static inline bool cfs_bandwidth_used(void)
5421 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
5422 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
5423 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
5424 static inline void sync_throttle(struct task_group
*tg
, int cpu
) {}
5425 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
5427 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
5432 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
5437 static inline int throttled_lb_pair(struct task_group
*tg
,
5438 int src_cpu
, int dest_cpu
)
5443 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
5445 #ifdef CONFIG_FAIR_GROUP_SCHED
5446 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
5449 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
5453 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
5454 static inline void update_runtime_enabled(struct rq
*rq
) {}
5455 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
5457 #endif /* CONFIG_CFS_BANDWIDTH */
5459 /**************************************************
5460 * CFS operations on tasks:
5463 #ifdef CONFIG_SCHED_HRTICK
5464 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
5466 struct sched_entity
*se
= &p
->se
;
5467 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5469 SCHED_WARN_ON(task_rq(p
) != rq
);
5471 if (rq
->cfs
.h_nr_running
> 1) {
5472 u64 slice
= sched_slice(cfs_rq
, se
);
5473 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
5474 s64 delta
= slice
- ran
;
5477 if (task_current(rq
, p
))
5481 hrtick_start(rq
, delta
);
5486 * called from enqueue/dequeue and updates the hrtick when the
5487 * current task is from our class and nr_running is low enough
5490 static void hrtick_update(struct rq
*rq
)
5492 struct task_struct
*curr
= rq
->curr
;
5494 if (!hrtick_enabled_fair(rq
) || curr
->sched_class
!= &fair_sched_class
)
5497 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
5498 hrtick_start_fair(rq
, curr
);
5500 #else /* !CONFIG_SCHED_HRTICK */
5502 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
5506 static inline void hrtick_update(struct rq
*rq
)
5512 static inline bool cpu_overutilized(int cpu
)
5514 return !fits_capacity(cpu_util_cfs(cpu
), capacity_of(cpu
));
5517 static inline void update_overutilized_status(struct rq
*rq
)
5519 if (!READ_ONCE(rq
->rd
->overutilized
) && cpu_overutilized(rq
->cpu
)) {
5520 WRITE_ONCE(rq
->rd
->overutilized
, SG_OVERUTILIZED
);
5521 trace_sched_overutilized_tp(rq
->rd
, SG_OVERUTILIZED
);
5525 static inline void update_overutilized_status(struct rq
*rq
) { }
5528 /* Runqueue only has SCHED_IDLE tasks enqueued */
5529 static int sched_idle_rq(struct rq
*rq
)
5531 return unlikely(rq
->nr_running
== rq
->cfs
.idle_h_nr_running
&&
5536 * Returns true if cfs_rq only has SCHED_IDLE entities enqueued. Note the use
5537 * of idle_nr_running, which does not consider idle descendants of normal
5540 static bool sched_idle_cfs_rq(struct cfs_rq
*cfs_rq
)
5542 return cfs_rq
->nr_running
&&
5543 cfs_rq
->nr_running
== cfs_rq
->idle_nr_running
;
5547 static int sched_idle_cpu(int cpu
)
5549 return sched_idle_rq(cpu_rq(cpu
));
5554 * The enqueue_task method is called before nr_running is
5555 * increased. Here we update the fair scheduling stats and
5556 * then put the task into the rbtree:
5559 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
5561 struct cfs_rq
*cfs_rq
;
5562 struct sched_entity
*se
= &p
->se
;
5563 int idle_h_nr_running
= task_has_idle_policy(p
);
5564 int task_new
= !(flags
& ENQUEUE_WAKEUP
);
5567 * The code below (indirectly) updates schedutil which looks at
5568 * the cfs_rq utilization to select a frequency.
5569 * Let's add the task's estimated utilization to the cfs_rq's
5570 * estimated utilization, before we update schedutil.
5572 util_est_enqueue(&rq
->cfs
, p
);
5575 * If in_iowait is set, the code below may not trigger any cpufreq
5576 * utilization updates, so do it here explicitly with the IOWAIT flag
5580 cpufreq_update_util(rq
, SCHED_CPUFREQ_IOWAIT
);
5582 for_each_sched_entity(se
) {
5585 cfs_rq
= cfs_rq_of(se
);
5586 enqueue_entity(cfs_rq
, se
, flags
);
5588 cfs_rq
->h_nr_running
++;
5589 cfs_rq
->idle_h_nr_running
+= idle_h_nr_running
;
5591 if (cfs_rq_is_idle(cfs_rq
))
5592 idle_h_nr_running
= 1;
5594 /* end evaluation on encountering a throttled cfs_rq */
5595 if (cfs_rq_throttled(cfs_rq
))
5596 goto enqueue_throttle
;
5598 flags
= ENQUEUE_WAKEUP
;
5601 for_each_sched_entity(se
) {
5602 cfs_rq
= cfs_rq_of(se
);
5604 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
5605 se_update_runnable(se
);
5606 update_cfs_group(se
);
5608 cfs_rq
->h_nr_running
++;
5609 cfs_rq
->idle_h_nr_running
+= idle_h_nr_running
;
5611 if (cfs_rq_is_idle(cfs_rq
))
5612 idle_h_nr_running
= 1;
5614 /* end evaluation on encountering a throttled cfs_rq */
5615 if (cfs_rq_throttled(cfs_rq
))
5616 goto enqueue_throttle
;
5619 * One parent has been throttled and cfs_rq removed from the
5620 * list. Add it back to not break the leaf list.
5622 if (throttled_hierarchy(cfs_rq
))
5623 list_add_leaf_cfs_rq(cfs_rq
);
5626 /* At this point se is NULL and we are at root level*/
5627 add_nr_running(rq
, 1);
5630 * Since new tasks are assigned an initial util_avg equal to
5631 * half of the spare capacity of their CPU, tiny tasks have the
5632 * ability to cross the overutilized threshold, which will
5633 * result in the load balancer ruining all the task placement
5634 * done by EAS. As a way to mitigate that effect, do not account
5635 * for the first enqueue operation of new tasks during the
5636 * overutilized flag detection.
5638 * A better way of solving this problem would be to wait for
5639 * the PELT signals of tasks to converge before taking them
5640 * into account, but that is not straightforward to implement,
5641 * and the following generally works well enough in practice.
5644 update_overutilized_status(rq
);
5647 if (cfs_bandwidth_used()) {
5649 * When bandwidth control is enabled; the cfs_rq_throttled()
5650 * breaks in the above iteration can result in incomplete
5651 * leaf list maintenance, resulting in triggering the assertion
5654 for_each_sched_entity(se
) {
5655 cfs_rq
= cfs_rq_of(se
);
5657 if (list_add_leaf_cfs_rq(cfs_rq
))
5662 assert_list_leaf_cfs_rq(rq
);
5667 static void set_next_buddy(struct sched_entity
*se
);
5670 * The dequeue_task method is called before nr_running is
5671 * decreased. We remove the task from the rbtree and
5672 * update the fair scheduling stats:
5674 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
5676 struct cfs_rq
*cfs_rq
;
5677 struct sched_entity
*se
= &p
->se
;
5678 int task_sleep
= flags
& DEQUEUE_SLEEP
;
5679 int idle_h_nr_running
= task_has_idle_policy(p
);
5680 bool was_sched_idle
= sched_idle_rq(rq
);
5682 util_est_dequeue(&rq
->cfs
, p
);
5684 for_each_sched_entity(se
) {
5685 cfs_rq
= cfs_rq_of(se
);
5686 dequeue_entity(cfs_rq
, se
, flags
);
5688 cfs_rq
->h_nr_running
--;
5689 cfs_rq
->idle_h_nr_running
-= idle_h_nr_running
;
5691 if (cfs_rq_is_idle(cfs_rq
))
5692 idle_h_nr_running
= 1;
5694 /* end evaluation on encountering a throttled cfs_rq */
5695 if (cfs_rq_throttled(cfs_rq
))
5696 goto dequeue_throttle
;
5698 /* Don't dequeue parent if it has other entities besides us */
5699 if (cfs_rq
->load
.weight
) {
5700 /* Avoid re-evaluating load for this entity: */
5701 se
= parent_entity(se
);
5703 * Bias pick_next to pick a task from this cfs_rq, as
5704 * p is sleeping when it is within its sched_slice.
5706 if (task_sleep
&& se
&& !throttled_hierarchy(cfs_rq
))
5710 flags
|= DEQUEUE_SLEEP
;
5713 for_each_sched_entity(se
) {
5714 cfs_rq
= cfs_rq_of(se
);
5716 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
5717 se_update_runnable(se
);
5718 update_cfs_group(se
);
5720 cfs_rq
->h_nr_running
--;
5721 cfs_rq
->idle_h_nr_running
-= idle_h_nr_running
;
5723 if (cfs_rq_is_idle(cfs_rq
))
5724 idle_h_nr_running
= 1;
5726 /* end evaluation on encountering a throttled cfs_rq */
5727 if (cfs_rq_throttled(cfs_rq
))
5728 goto dequeue_throttle
;
5732 /* At this point se is NULL and we are at root level*/
5733 sub_nr_running(rq
, 1);
5735 /* balance early to pull high priority tasks */
5736 if (unlikely(!was_sched_idle
&& sched_idle_rq(rq
)))
5737 rq
->next_balance
= jiffies
;
5740 util_est_update(&rq
->cfs
, p
, task_sleep
);
5746 /* Working cpumask for: load_balance, load_balance_newidle. */
5747 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5748 DEFINE_PER_CPU(cpumask_var_t
, select_idle_mask
);
5750 #ifdef CONFIG_NO_HZ_COMMON
5753 cpumask_var_t idle_cpus_mask
;
5755 int has_blocked
; /* Idle CPUS has blocked load */
5756 int needs_update
; /* Newly idle CPUs need their next_balance collated */
5757 unsigned long next_balance
; /* in jiffy units */
5758 unsigned long next_blocked
; /* Next update of blocked load in jiffies */
5759 } nohz ____cacheline_aligned
;
5761 #endif /* CONFIG_NO_HZ_COMMON */
5763 static unsigned long cpu_load(struct rq
*rq
)
5765 return cfs_rq_load_avg(&rq
->cfs
);
5769 * cpu_load_without - compute CPU load without any contributions from *p
5770 * @cpu: the CPU which load is requested
5771 * @p: the task which load should be discounted
5773 * The load of a CPU is defined by the load of tasks currently enqueued on that
5774 * CPU as well as tasks which are currently sleeping after an execution on that
5777 * This method returns the load of the specified CPU by discounting the load of
5778 * the specified task, whenever the task is currently contributing to the CPU
5781 static unsigned long cpu_load_without(struct rq
*rq
, struct task_struct
*p
)
5783 struct cfs_rq
*cfs_rq
;
5786 /* Task has no contribution or is new */
5787 if (cpu_of(rq
) != task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
5788 return cpu_load(rq
);
5791 load
= READ_ONCE(cfs_rq
->avg
.load_avg
);
5793 /* Discount task's util from CPU's util */
5794 lsub_positive(&load
, task_h_load(p
));
5799 static unsigned long cpu_runnable(struct rq
*rq
)
5801 return cfs_rq_runnable_avg(&rq
->cfs
);
5804 static unsigned long cpu_runnable_without(struct rq
*rq
, struct task_struct
*p
)
5806 struct cfs_rq
*cfs_rq
;
5807 unsigned int runnable
;
5809 /* Task has no contribution or is new */
5810 if (cpu_of(rq
) != task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
5811 return cpu_runnable(rq
);
5814 runnable
= READ_ONCE(cfs_rq
->avg
.runnable_avg
);
5816 /* Discount task's runnable from CPU's runnable */
5817 lsub_positive(&runnable
, p
->se
.avg
.runnable_avg
);
5822 static unsigned long capacity_of(int cpu
)
5824 return cpu_rq(cpu
)->cpu_capacity
;
5827 static void record_wakee(struct task_struct
*p
)
5830 * Only decay a single time; tasks that have less then 1 wakeup per
5831 * jiffy will not have built up many flips.
5833 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
5834 current
->wakee_flips
>>= 1;
5835 current
->wakee_flip_decay_ts
= jiffies
;
5838 if (current
->last_wakee
!= p
) {
5839 current
->last_wakee
= p
;
5840 current
->wakee_flips
++;
5845 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5847 * A waker of many should wake a different task than the one last awakened
5848 * at a frequency roughly N times higher than one of its wakees.
5850 * In order to determine whether we should let the load spread vs consolidating
5851 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5852 * partner, and a factor of lls_size higher frequency in the other.
5854 * With both conditions met, we can be relatively sure that the relationship is
5855 * non-monogamous, with partner count exceeding socket size.
5857 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5858 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5861 static int wake_wide(struct task_struct
*p
)
5863 unsigned int master
= current
->wakee_flips
;
5864 unsigned int slave
= p
->wakee_flips
;
5865 int factor
= __this_cpu_read(sd_llc_size
);
5868 swap(master
, slave
);
5869 if (slave
< factor
|| master
< slave
* factor
)
5875 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5876 * soonest. For the purpose of speed we only consider the waking and previous
5879 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5880 * cache-affine and is (or will be) idle.
5882 * wake_affine_weight() - considers the weight to reflect the average
5883 * scheduling latency of the CPUs. This seems to work
5884 * for the overloaded case.
5887 wake_affine_idle(int this_cpu
, int prev_cpu
, int sync
)
5890 * If this_cpu is idle, it implies the wakeup is from interrupt
5891 * context. Only allow the move if cache is shared. Otherwise an
5892 * interrupt intensive workload could force all tasks onto one
5893 * node depending on the IO topology or IRQ affinity settings.
5895 * If the prev_cpu is idle and cache affine then avoid a migration.
5896 * There is no guarantee that the cache hot data from an interrupt
5897 * is more important than cache hot data on the prev_cpu and from
5898 * a cpufreq perspective, it's better to have higher utilisation
5901 if (available_idle_cpu(this_cpu
) && cpus_share_cache(this_cpu
, prev_cpu
))
5902 return available_idle_cpu(prev_cpu
) ? prev_cpu
: this_cpu
;
5904 if (sync
&& cpu_rq(this_cpu
)->nr_running
== 1)
5907 if (available_idle_cpu(prev_cpu
))
5910 return nr_cpumask_bits
;
5914 wake_affine_weight(struct sched_domain
*sd
, struct task_struct
*p
,
5915 int this_cpu
, int prev_cpu
, int sync
)
5917 s64 this_eff_load
, prev_eff_load
;
5918 unsigned long task_load
;
5920 this_eff_load
= cpu_load(cpu_rq(this_cpu
));
5923 unsigned long current_load
= task_h_load(current
);
5925 if (current_load
> this_eff_load
)
5928 this_eff_load
-= current_load
;
5931 task_load
= task_h_load(p
);
5933 this_eff_load
+= task_load
;
5934 if (sched_feat(WA_BIAS
))
5935 this_eff_load
*= 100;
5936 this_eff_load
*= capacity_of(prev_cpu
);
5938 prev_eff_load
= cpu_load(cpu_rq(prev_cpu
));
5939 prev_eff_load
-= task_load
;
5940 if (sched_feat(WA_BIAS
))
5941 prev_eff_load
*= 100 + (sd
->imbalance_pct
- 100) / 2;
5942 prev_eff_load
*= capacity_of(this_cpu
);
5945 * If sync, adjust the weight of prev_eff_load such that if
5946 * prev_eff == this_eff that select_idle_sibling() will consider
5947 * stacking the wakee on top of the waker if no other CPU is
5953 return this_eff_load
< prev_eff_load
? this_cpu
: nr_cpumask_bits
;
5956 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
,
5957 int this_cpu
, int prev_cpu
, int sync
)
5959 int target
= nr_cpumask_bits
;
5961 if (sched_feat(WA_IDLE
))
5962 target
= wake_affine_idle(this_cpu
, prev_cpu
, sync
);
5964 if (sched_feat(WA_WEIGHT
) && target
== nr_cpumask_bits
)
5965 target
= wake_affine_weight(sd
, p
, this_cpu
, prev_cpu
, sync
);
5967 schedstat_inc(p
->stats
.nr_wakeups_affine_attempts
);
5968 if (target
== nr_cpumask_bits
)
5971 schedstat_inc(sd
->ttwu_move_affine
);
5972 schedstat_inc(p
->stats
.nr_wakeups_affine
);
5976 static struct sched_group
*
5977 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
);
5980 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5983 find_idlest_group_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
5985 unsigned long load
, min_load
= ULONG_MAX
;
5986 unsigned int min_exit_latency
= UINT_MAX
;
5987 u64 latest_idle_timestamp
= 0;
5988 int least_loaded_cpu
= this_cpu
;
5989 int shallowest_idle_cpu
= -1;
5992 /* Check if we have any choice: */
5993 if (group
->group_weight
== 1)
5994 return cpumask_first(sched_group_span(group
));
5996 /* Traverse only the allowed CPUs */
5997 for_each_cpu_and(i
, sched_group_span(group
), p
->cpus_ptr
) {
5998 struct rq
*rq
= cpu_rq(i
);
6000 if (!sched_core_cookie_match(rq
, p
))
6003 if (sched_idle_cpu(i
))
6006 if (available_idle_cpu(i
)) {
6007 struct cpuidle_state
*idle
= idle_get_state(rq
);
6008 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
6010 * We give priority to a CPU whose idle state
6011 * has the smallest exit latency irrespective
6012 * of any idle timestamp.
6014 min_exit_latency
= idle
->exit_latency
;
6015 latest_idle_timestamp
= rq
->idle_stamp
;
6016 shallowest_idle_cpu
= i
;
6017 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
6018 rq
->idle_stamp
> latest_idle_timestamp
) {
6020 * If equal or no active idle state, then
6021 * the most recently idled CPU might have
6024 latest_idle_timestamp
= rq
->idle_stamp
;
6025 shallowest_idle_cpu
= i
;
6027 } else if (shallowest_idle_cpu
== -1) {
6028 load
= cpu_load(cpu_rq(i
));
6029 if (load
< min_load
) {
6031 least_loaded_cpu
= i
;
6036 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
6039 static inline int find_idlest_cpu(struct sched_domain
*sd
, struct task_struct
*p
,
6040 int cpu
, int prev_cpu
, int sd_flag
)
6044 if (!cpumask_intersects(sched_domain_span(sd
), p
->cpus_ptr
))
6048 * We need task's util for cpu_util_without, sync it up to
6049 * prev_cpu's last_update_time.
6051 if (!(sd_flag
& SD_BALANCE_FORK
))
6052 sync_entity_load_avg(&p
->se
);
6055 struct sched_group
*group
;
6056 struct sched_domain
*tmp
;
6059 if (!(sd
->flags
& sd_flag
)) {
6064 group
= find_idlest_group(sd
, p
, cpu
);
6070 new_cpu
= find_idlest_group_cpu(group
, p
, cpu
);
6071 if (new_cpu
== cpu
) {
6072 /* Now try balancing at a lower domain level of 'cpu': */
6077 /* Now try balancing at a lower domain level of 'new_cpu': */
6079 weight
= sd
->span_weight
;
6081 for_each_domain(cpu
, tmp
) {
6082 if (weight
<= tmp
->span_weight
)
6084 if (tmp
->flags
& sd_flag
)
6092 static inline int __select_idle_cpu(int cpu
, struct task_struct
*p
)
6094 if ((available_idle_cpu(cpu
) || sched_idle_cpu(cpu
)) &&
6095 sched_cpu_cookie_match(cpu_rq(cpu
), p
))
6101 #ifdef CONFIG_SCHED_SMT
6102 DEFINE_STATIC_KEY_FALSE(sched_smt_present
);
6103 EXPORT_SYMBOL_GPL(sched_smt_present
);
6105 static inline void set_idle_cores(int cpu
, int val
)
6107 struct sched_domain_shared
*sds
;
6109 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
6111 WRITE_ONCE(sds
->has_idle_cores
, val
);
6114 static inline bool test_idle_cores(int cpu
, bool def
)
6116 struct sched_domain_shared
*sds
;
6118 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
6120 return READ_ONCE(sds
->has_idle_cores
);
6126 * Scans the local SMT mask to see if the entire core is idle, and records this
6127 * information in sd_llc_shared->has_idle_cores.
6129 * Since SMT siblings share all cache levels, inspecting this limited remote
6130 * state should be fairly cheap.
6132 void __update_idle_core(struct rq
*rq
)
6134 int core
= cpu_of(rq
);
6138 if (test_idle_cores(core
, true))
6141 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
6145 if (!available_idle_cpu(cpu
))
6149 set_idle_cores(core
, 1);
6155 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6156 * there are no idle cores left in the system; tracked through
6157 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6159 static int select_idle_core(struct task_struct
*p
, int core
, struct cpumask
*cpus
, int *idle_cpu
)
6164 if (!static_branch_likely(&sched_smt_present
))
6165 return __select_idle_cpu(core
, p
);
6167 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
6168 if (!available_idle_cpu(cpu
)) {
6170 if (*idle_cpu
== -1) {
6171 if (sched_idle_cpu(cpu
) && cpumask_test_cpu(cpu
, p
->cpus_ptr
)) {
6179 if (*idle_cpu
== -1 && cpumask_test_cpu(cpu
, p
->cpus_ptr
))
6186 cpumask_andnot(cpus
, cpus
, cpu_smt_mask(core
));
6191 * Scan the local SMT mask for idle CPUs.
6193 static int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6197 for_each_cpu(cpu
, cpu_smt_mask(target
)) {
6198 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
) ||
6199 !cpumask_test_cpu(cpu
, sched_domain_span(sd
)))
6201 if (available_idle_cpu(cpu
) || sched_idle_cpu(cpu
))
6208 #else /* CONFIG_SCHED_SMT */
6210 static inline void set_idle_cores(int cpu
, int val
)
6214 static inline bool test_idle_cores(int cpu
, bool def
)
6219 static inline int select_idle_core(struct task_struct
*p
, int core
, struct cpumask
*cpus
, int *idle_cpu
)
6221 return __select_idle_cpu(core
, p
);
6224 static inline int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6229 #endif /* CONFIG_SCHED_SMT */
6232 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6233 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6234 * average idle time for this rq (as found in rq->avg_idle).
6236 static int select_idle_cpu(struct task_struct
*p
, struct sched_domain
*sd
, bool has_idle_core
, int target
)
6238 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6239 int i
, cpu
, idle_cpu
= -1, nr
= INT_MAX
;
6240 struct rq
*this_rq
= this_rq();
6241 int this = smp_processor_id();
6242 struct sched_domain
*this_sd
;
6245 this_sd
= rcu_dereference(*this_cpu_ptr(&sd_llc
));
6249 cpumask_and(cpus
, sched_domain_span(sd
), p
->cpus_ptr
);
6251 if (sched_feat(SIS_PROP
) && !has_idle_core
) {
6252 u64 avg_cost
, avg_idle
, span_avg
;
6253 unsigned long now
= jiffies
;
6256 * If we're busy, the assumption that the last idle period
6257 * predicts the future is flawed; age away the remaining
6258 * predicted idle time.
6260 if (unlikely(this_rq
->wake_stamp
< now
)) {
6261 while (this_rq
->wake_stamp
< now
&& this_rq
->wake_avg_idle
) {
6262 this_rq
->wake_stamp
++;
6263 this_rq
->wake_avg_idle
>>= 1;
6267 avg_idle
= this_rq
->wake_avg_idle
;
6268 avg_cost
= this_sd
->avg_scan_cost
+ 1;
6270 span_avg
= sd
->span_weight
* avg_idle
;
6271 if (span_avg
> 4*avg_cost
)
6272 nr
= div_u64(span_avg
, avg_cost
);
6276 time
= cpu_clock(this);
6279 for_each_cpu_wrap(cpu
, cpus
, target
+ 1) {
6280 if (has_idle_core
) {
6281 i
= select_idle_core(p
, cpu
, cpus
, &idle_cpu
);
6282 if ((unsigned int)i
< nr_cpumask_bits
)
6288 idle_cpu
= __select_idle_cpu(cpu
, p
);
6289 if ((unsigned int)idle_cpu
< nr_cpumask_bits
)
6295 set_idle_cores(target
, false);
6297 if (sched_feat(SIS_PROP
) && !has_idle_core
) {
6298 time
= cpu_clock(this) - time
;
6301 * Account for the scan cost of wakeups against the average
6304 this_rq
->wake_avg_idle
-= min(this_rq
->wake_avg_idle
, time
);
6306 update_avg(&this_sd
->avg_scan_cost
, time
);
6313 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6314 * the task fits. If no CPU is big enough, but there are idle ones, try to
6315 * maximize capacity.
6318 select_idle_capacity(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6320 unsigned long task_util
, best_cap
= 0;
6321 int cpu
, best_cpu
= -1;
6322 struct cpumask
*cpus
;
6324 cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6325 cpumask_and(cpus
, sched_domain_span(sd
), p
->cpus_ptr
);
6327 task_util
= uclamp_task_util(p
);
6329 for_each_cpu_wrap(cpu
, cpus
, target
) {
6330 unsigned long cpu_cap
= capacity_of(cpu
);
6332 if (!available_idle_cpu(cpu
) && !sched_idle_cpu(cpu
))
6334 if (fits_capacity(task_util
, cpu_cap
))
6337 if (cpu_cap
> best_cap
) {
6346 static inline bool asym_fits_capacity(unsigned long task_util
, int cpu
)
6348 if (static_branch_unlikely(&sched_asym_cpucapacity
))
6349 return fits_capacity(task_util
, capacity_of(cpu
));
6355 * Try and locate an idle core/thread in the LLC cache domain.
6357 static int select_idle_sibling(struct task_struct
*p
, int prev
, int target
)
6359 bool has_idle_core
= false;
6360 struct sched_domain
*sd
;
6361 unsigned long task_util
;
6362 int i
, recent_used_cpu
;
6365 * On asymmetric system, update task utilization because we will check
6366 * that the task fits with cpu's capacity.
6368 if (static_branch_unlikely(&sched_asym_cpucapacity
)) {
6369 sync_entity_load_avg(&p
->se
);
6370 task_util
= uclamp_task_util(p
);
6374 * per-cpu select_idle_mask usage
6376 lockdep_assert_irqs_disabled();
6378 if ((available_idle_cpu(target
) || sched_idle_cpu(target
)) &&
6379 asym_fits_capacity(task_util
, target
))
6383 * If the previous CPU is cache affine and idle, don't be stupid:
6385 if (prev
!= target
&& cpus_share_cache(prev
, target
) &&
6386 (available_idle_cpu(prev
) || sched_idle_cpu(prev
)) &&
6387 asym_fits_capacity(task_util
, prev
))
6391 * Allow a per-cpu kthread to stack with the wakee if the
6392 * kworker thread and the tasks previous CPUs are the same.
6393 * The assumption is that the wakee queued work for the
6394 * per-cpu kthread that is now complete and the wakeup is
6395 * essentially a sync wakeup. An obvious example of this
6396 * pattern is IO completions.
6398 if (is_per_cpu_kthread(current
) &&
6400 prev
== smp_processor_id() &&
6401 this_rq()->nr_running
<= 1 &&
6402 asym_fits_capacity(task_util
, prev
)) {
6406 /* Check a recently used CPU as a potential idle candidate: */
6407 recent_used_cpu
= p
->recent_used_cpu
;
6408 p
->recent_used_cpu
= prev
;
6409 if (recent_used_cpu
!= prev
&&
6410 recent_used_cpu
!= target
&&
6411 cpus_share_cache(recent_used_cpu
, target
) &&
6412 (available_idle_cpu(recent_used_cpu
) || sched_idle_cpu(recent_used_cpu
)) &&
6413 cpumask_test_cpu(p
->recent_used_cpu
, p
->cpus_ptr
) &&
6414 asym_fits_capacity(task_util
, recent_used_cpu
)) {
6415 return recent_used_cpu
;
6419 * For asymmetric CPU capacity systems, our domain of interest is
6420 * sd_asym_cpucapacity rather than sd_llc.
6422 if (static_branch_unlikely(&sched_asym_cpucapacity
)) {
6423 sd
= rcu_dereference(per_cpu(sd_asym_cpucapacity
, target
));
6425 * On an asymmetric CPU capacity system where an exclusive
6426 * cpuset defines a symmetric island (i.e. one unique
6427 * capacity_orig value through the cpuset), the key will be set
6428 * but the CPUs within that cpuset will not have a domain with
6429 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6433 i
= select_idle_capacity(p
, sd
, target
);
6434 return ((unsigned)i
< nr_cpumask_bits
) ? i
: target
;
6438 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
6442 if (sched_smt_active()) {
6443 has_idle_core
= test_idle_cores(target
, false);
6445 if (!has_idle_core
&& cpus_share_cache(prev
, target
)) {
6446 i
= select_idle_smt(p
, sd
, prev
);
6447 if ((unsigned int)i
< nr_cpumask_bits
)
6452 i
= select_idle_cpu(p
, sd
, has_idle_core
, target
);
6453 if ((unsigned)i
< nr_cpumask_bits
)
6460 * cpu_util_without: compute cpu utilization without any contributions from *p
6461 * @cpu: the CPU which utilization is requested
6462 * @p: the task which utilization should be discounted
6464 * The utilization of a CPU is defined by the utilization of tasks currently
6465 * enqueued on that CPU as well as tasks which are currently sleeping after an
6466 * execution on that CPU.
6468 * This method returns the utilization of the specified CPU by discounting the
6469 * utilization of the specified task, whenever the task is currently
6470 * contributing to the CPU utilization.
6472 static unsigned long cpu_util_without(int cpu
, struct task_struct
*p
)
6474 struct cfs_rq
*cfs_rq
;
6477 /* Task has no contribution or is new */
6478 if (cpu
!= task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
6479 return cpu_util_cfs(cpu
);
6481 cfs_rq
= &cpu_rq(cpu
)->cfs
;
6482 util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6484 /* Discount task's util from CPU's util */
6485 lsub_positive(&util
, task_util(p
));
6490 * a) if *p is the only task sleeping on this CPU, then:
6491 * cpu_util (== task_util) > util_est (== 0)
6492 * and thus we return:
6493 * cpu_util_without = (cpu_util - task_util) = 0
6495 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6497 * cpu_util >= task_util
6498 * cpu_util > util_est (== 0)
6499 * and thus we discount *p's blocked utilization to return:
6500 * cpu_util_without = (cpu_util - task_util) >= 0
6502 * c) if other tasks are RUNNABLE on that CPU and
6503 * util_est > cpu_util
6504 * then we use util_est since it returns a more restrictive
6505 * estimation of the spare capacity on that CPU, by just
6506 * considering the expected utilization of tasks already
6507 * runnable on that CPU.
6509 * Cases a) and b) are covered by the above code, while case c) is
6510 * covered by the following code when estimated utilization is
6513 if (sched_feat(UTIL_EST
)) {
6514 unsigned int estimated
=
6515 READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
);
6518 * Despite the following checks we still have a small window
6519 * for a possible race, when an execl's select_task_rq_fair()
6520 * races with LB's detach_task():
6523 * p->on_rq = TASK_ON_RQ_MIGRATING;
6524 * ---------------------------------- A
6525 * deactivate_task() \
6526 * dequeue_task() + RaceTime
6527 * util_est_dequeue() /
6528 * ---------------------------------- B
6530 * The additional check on "current == p" it's required to
6531 * properly fix the execl regression and it helps in further
6532 * reducing the chances for the above race.
6534 if (unlikely(task_on_rq_queued(p
) || current
== p
))
6535 lsub_positive(&estimated
, _task_util_est(p
));
6537 util
= max(util
, estimated
);
6541 * Utilization (estimated) can exceed the CPU capacity, thus let's
6542 * clamp to the maximum CPU capacity to ensure consistency with
6545 return min_t(unsigned long, util
, capacity_orig_of(cpu
));
6549 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6552 static unsigned long cpu_util_next(int cpu
, struct task_struct
*p
, int dst_cpu
)
6554 struct cfs_rq
*cfs_rq
= &cpu_rq(cpu
)->cfs
;
6555 unsigned long util_est
, util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6558 * If @p migrates from @cpu to another, remove its contribution. Or,
6559 * if @p migrates from another CPU to @cpu, add its contribution. In
6560 * the other cases, @cpu is not impacted by the migration, so the
6561 * util_avg should already be correct.
6563 if (task_cpu(p
) == cpu
&& dst_cpu
!= cpu
)
6564 lsub_positive(&util
, task_util(p
));
6565 else if (task_cpu(p
) != cpu
&& dst_cpu
== cpu
)
6566 util
+= task_util(p
);
6568 if (sched_feat(UTIL_EST
)) {
6569 util_est
= READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
);
6572 * During wake-up, the task isn't enqueued yet and doesn't
6573 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6574 * so just add it (if needed) to "simulate" what will be
6575 * cpu_util after the task has been enqueued.
6578 util_est
+= _task_util_est(p
);
6580 util
= max(util
, util_est
);
6583 return min(util
, capacity_orig_of(cpu
));
6587 * compute_energy(): Estimates the energy that @pd would consume if @p was
6588 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6589 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6590 * to compute what would be the energy if we decided to actually migrate that
6594 compute_energy(struct task_struct
*p
, int dst_cpu
, struct perf_domain
*pd
)
6596 struct cpumask
*pd_mask
= perf_domain_span(pd
);
6597 unsigned long cpu_cap
= arch_scale_cpu_capacity(cpumask_first(pd_mask
));
6598 unsigned long max_util
= 0, sum_util
= 0;
6599 unsigned long _cpu_cap
= cpu_cap
;
6602 _cpu_cap
-= arch_scale_thermal_pressure(cpumask_first(pd_mask
));
6605 * The capacity state of CPUs of the current rd can be driven by CPUs
6606 * of another rd if they belong to the same pd. So, account for the
6607 * utilization of these CPUs too by masking pd with cpu_online_mask
6608 * instead of the rd span.
6610 * If an entire pd is outside of the current rd, it will not appear in
6611 * its pd list and will not be accounted by compute_energy().
6613 for_each_cpu_and(cpu
, pd_mask
, cpu_online_mask
) {
6614 unsigned long util_freq
= cpu_util_next(cpu
, p
, dst_cpu
);
6615 unsigned long cpu_util
, util_running
= util_freq
;
6616 struct task_struct
*tsk
= NULL
;
6619 * When @p is placed on @cpu:
6621 * util_running = max(cpu_util, cpu_util_est) +
6622 * max(task_util, _task_util_est)
6624 * while cpu_util_next is: max(cpu_util + task_util,
6625 * cpu_util_est + _task_util_est)
6627 if (cpu
== dst_cpu
) {
6630 cpu_util_next(cpu
, p
, -1) + task_util_est(p
);
6634 * Busy time computation: utilization clamping is not
6635 * required since the ratio (sum_util / cpu_capacity)
6636 * is already enough to scale the EM reported power
6637 * consumption at the (eventually clamped) cpu_capacity.
6639 cpu_util
= effective_cpu_util(cpu
, util_running
, cpu_cap
,
6642 sum_util
+= min(cpu_util
, _cpu_cap
);
6645 * Performance domain frequency: utilization clamping
6646 * must be considered since it affects the selection
6647 * of the performance domain frequency.
6648 * NOTE: in case RT tasks are running, by default the
6649 * FREQUENCY_UTIL's utilization can be max OPP.
6651 cpu_util
= effective_cpu_util(cpu
, util_freq
, cpu_cap
,
6652 FREQUENCY_UTIL
, tsk
);
6653 max_util
= max(max_util
, min(cpu_util
, _cpu_cap
));
6656 return em_cpu_energy(pd
->em_pd
, max_util
, sum_util
, _cpu_cap
);
6660 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6661 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6662 * spare capacity in each performance domain and uses it as a potential
6663 * candidate to execute the task. Then, it uses the Energy Model to figure
6664 * out which of the CPU candidates is the most energy-efficient.
6666 * The rationale for this heuristic is as follows. In a performance domain,
6667 * all the most energy efficient CPU candidates (according to the Energy
6668 * Model) are those for which we'll request a low frequency. When there are
6669 * several CPUs for which the frequency request will be the same, we don't
6670 * have enough data to break the tie between them, because the Energy Model
6671 * only includes active power costs. With this model, if we assume that
6672 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6673 * the maximum spare capacity in a performance domain is guaranteed to be among
6674 * the best candidates of the performance domain.
6676 * In practice, it could be preferable from an energy standpoint to pack
6677 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6678 * but that could also hurt our chances to go cluster idle, and we have no
6679 * ways to tell with the current Energy Model if this is actually a good
6680 * idea or not. So, find_energy_efficient_cpu() basically favors
6681 * cluster-packing, and spreading inside a cluster. That should at least be
6682 * a good thing for latency, and this is consistent with the idea that most
6683 * of the energy savings of EAS come from the asymmetry of the system, and
6684 * not so much from breaking the tie between identical CPUs. That's also the
6685 * reason why EAS is enabled in the topology code only for systems where
6686 * SD_ASYM_CPUCAPACITY is set.
6688 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6689 * they don't have any useful utilization data yet and it's not possible to
6690 * forecast their impact on energy consumption. Consequently, they will be
6691 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6692 * to be energy-inefficient in some use-cases. The alternative would be to
6693 * bias new tasks towards specific types of CPUs first, or to try to infer
6694 * their util_avg from the parent task, but those heuristics could hurt
6695 * other use-cases too. So, until someone finds a better way to solve this,
6696 * let's keep things simple by re-using the existing slow path.
6698 static int find_energy_efficient_cpu(struct task_struct
*p
, int prev_cpu
)
6700 unsigned long prev_delta
= ULONG_MAX
, best_delta
= ULONG_MAX
;
6701 struct root_domain
*rd
= cpu_rq(smp_processor_id())->rd
;
6702 int cpu
, best_energy_cpu
= prev_cpu
, target
= -1;
6703 unsigned long cpu_cap
, util
, base_energy
= 0;
6704 struct sched_domain
*sd
;
6705 struct perf_domain
*pd
;
6708 pd
= rcu_dereference(rd
->pd
);
6709 if (!pd
|| READ_ONCE(rd
->overutilized
))
6713 * Energy-aware wake-up happens on the lowest sched_domain starting
6714 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6716 sd
= rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity
));
6717 while (sd
&& !cpumask_test_cpu(prev_cpu
, sched_domain_span(sd
)))
6724 sync_entity_load_avg(&p
->se
);
6725 if (!task_util_est(p
))
6728 for (; pd
; pd
= pd
->next
) {
6729 unsigned long cur_delta
, spare_cap
, max_spare_cap
= 0;
6730 bool compute_prev_delta
= false;
6731 unsigned long base_energy_pd
;
6732 int max_spare_cap_cpu
= -1;
6734 for_each_cpu_and(cpu
, perf_domain_span(pd
), sched_domain_span(sd
)) {
6735 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
6738 util
= cpu_util_next(cpu
, p
, cpu
);
6739 cpu_cap
= capacity_of(cpu
);
6740 spare_cap
= cpu_cap
;
6741 lsub_positive(&spare_cap
, util
);
6744 * Skip CPUs that cannot satisfy the capacity request.
6745 * IOW, placing the task there would make the CPU
6746 * overutilized. Take uclamp into account to see how
6747 * much capacity we can get out of the CPU; this is
6748 * aligned with sched_cpu_util().
6750 util
= uclamp_rq_util_with(cpu_rq(cpu
), util
, p
);
6751 if (!fits_capacity(util
, cpu_cap
))
6754 if (cpu
== prev_cpu
) {
6755 /* Always use prev_cpu as a candidate. */
6756 compute_prev_delta
= true;
6757 } else if (spare_cap
> max_spare_cap
) {
6759 * Find the CPU with the maximum spare capacity
6760 * in the performance domain.
6762 max_spare_cap
= spare_cap
;
6763 max_spare_cap_cpu
= cpu
;
6767 if (max_spare_cap_cpu
< 0 && !compute_prev_delta
)
6770 /* Compute the 'base' energy of the pd, without @p */
6771 base_energy_pd
= compute_energy(p
, -1, pd
);
6772 base_energy
+= base_energy_pd
;
6774 /* Evaluate the energy impact of using prev_cpu. */
6775 if (compute_prev_delta
) {
6776 prev_delta
= compute_energy(p
, prev_cpu
, pd
);
6777 if (prev_delta
< base_energy_pd
)
6779 prev_delta
-= base_energy_pd
;
6780 best_delta
= min(best_delta
, prev_delta
);
6783 /* Evaluate the energy impact of using max_spare_cap_cpu. */
6784 if (max_spare_cap_cpu
>= 0) {
6785 cur_delta
= compute_energy(p
, max_spare_cap_cpu
, pd
);
6786 if (cur_delta
< base_energy_pd
)
6788 cur_delta
-= base_energy_pd
;
6789 if (cur_delta
< best_delta
) {
6790 best_delta
= cur_delta
;
6791 best_energy_cpu
= max_spare_cap_cpu
;
6798 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6799 * least 6% of the energy used by prev_cpu.
6801 if ((prev_delta
== ULONG_MAX
) ||
6802 (prev_delta
- best_delta
) > ((prev_delta
+ base_energy
) >> 4))
6803 target
= best_energy_cpu
;
6814 * select_task_rq_fair: Select target runqueue for the waking task in domains
6815 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
6816 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6818 * Balances load by selecting the idlest CPU in the idlest group, or under
6819 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6821 * Returns the target CPU number.
6824 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int wake_flags
)
6826 int sync
= (wake_flags
& WF_SYNC
) && !(current
->flags
& PF_EXITING
);
6827 struct sched_domain
*tmp
, *sd
= NULL
;
6828 int cpu
= smp_processor_id();
6829 int new_cpu
= prev_cpu
;
6830 int want_affine
= 0;
6831 /* SD_flags and WF_flags share the first nibble */
6832 int sd_flag
= wake_flags
& 0xF;
6835 * required for stable ->cpus_allowed
6837 lockdep_assert_held(&p
->pi_lock
);
6838 if (wake_flags
& WF_TTWU
) {
6841 if (sched_energy_enabled()) {
6842 new_cpu
= find_energy_efficient_cpu(p
, prev_cpu
);
6848 want_affine
= !wake_wide(p
) && cpumask_test_cpu(cpu
, p
->cpus_ptr
);
6852 for_each_domain(cpu
, tmp
) {
6854 * If both 'cpu' and 'prev_cpu' are part of this domain,
6855 * cpu is a valid SD_WAKE_AFFINE target.
6857 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
6858 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
6859 if (cpu
!= prev_cpu
)
6860 new_cpu
= wake_affine(tmp
, p
, cpu
, prev_cpu
, sync
);
6862 sd
= NULL
; /* Prefer wake_affine over balance flags */
6867 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
6868 * usually do not have SD_BALANCE_WAKE set. That means wakeup
6869 * will usually go to the fast path.
6871 if (tmp
->flags
& sd_flag
)
6873 else if (!want_affine
)
6879 new_cpu
= find_idlest_cpu(sd
, p
, cpu
, prev_cpu
, sd_flag
);
6880 } else if (wake_flags
& WF_TTWU
) { /* XXX always ? */
6882 new_cpu
= select_idle_sibling(p
, prev_cpu
, new_cpu
);
6889 static void detach_entity_cfs_rq(struct sched_entity
*se
);
6892 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6893 * cfs_rq_of(p) references at time of call are still valid and identify the
6894 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6896 static void migrate_task_rq_fair(struct task_struct
*p
, int new_cpu
)
6899 * As blocked tasks retain absolute vruntime the migration needs to
6900 * deal with this by subtracting the old and adding the new
6901 * min_vruntime -- the latter is done by enqueue_entity() when placing
6902 * the task on the new runqueue.
6904 if (READ_ONCE(p
->__state
) == TASK_WAKING
) {
6905 struct sched_entity
*se
= &p
->se
;
6906 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6909 #ifndef CONFIG_64BIT
6910 u64 min_vruntime_copy
;
6913 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
6915 min_vruntime
= cfs_rq
->min_vruntime
;
6916 } while (min_vruntime
!= min_vruntime_copy
);
6918 min_vruntime
= cfs_rq
->min_vruntime
;
6921 se
->vruntime
-= min_vruntime
;
6924 if (p
->on_rq
== TASK_ON_RQ_MIGRATING
) {
6926 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6927 * rq->lock and can modify state directly.
6929 lockdep_assert_rq_held(task_rq(p
));
6930 detach_entity_cfs_rq(&p
->se
);
6934 * We are supposed to update the task to "current" time, then
6935 * its up to date and ready to go to new CPU/cfs_rq. But we
6936 * have difficulty in getting what current time is, so simply
6937 * throw away the out-of-date time. This will result in the
6938 * wakee task is less decayed, but giving the wakee more load
6941 remove_entity_load_avg(&p
->se
);
6944 /* Tell new CPU we are migrated */
6945 p
->se
.avg
.last_update_time
= 0;
6947 /* We have migrated, no longer consider this task hot */
6948 p
->se
.exec_start
= 0;
6950 update_scan_period(p
, new_cpu
);
6953 static void task_dead_fair(struct task_struct
*p
)
6955 remove_entity_load_avg(&p
->se
);
6959 balance_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
6964 return newidle_balance(rq
, rf
) != 0;
6966 #endif /* CONFIG_SMP */
6968 static unsigned long wakeup_gran(struct sched_entity
*se
)
6970 unsigned long gran
= sysctl_sched_wakeup_granularity
;
6973 * Since its curr running now, convert the gran from real-time
6974 * to virtual-time in his units.
6976 * By using 'se' instead of 'curr' we penalize light tasks, so
6977 * they get preempted easier. That is, if 'se' < 'curr' then
6978 * the resulting gran will be larger, therefore penalizing the
6979 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6980 * be smaller, again penalizing the lighter task.
6982 * This is especially important for buddies when the leftmost
6983 * task is higher priority than the buddy.
6985 return calc_delta_fair(gran
, se
);
6989 * Should 'se' preempt 'curr'.
7003 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
7005 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
7010 gran
= wakeup_gran(se
);
7017 static void set_last_buddy(struct sched_entity
*se
)
7019 for_each_sched_entity(se
) {
7020 if (SCHED_WARN_ON(!se
->on_rq
))
7024 cfs_rq_of(se
)->last
= se
;
7028 static void set_next_buddy(struct sched_entity
*se
)
7030 for_each_sched_entity(se
) {
7031 if (SCHED_WARN_ON(!se
->on_rq
))
7035 cfs_rq_of(se
)->next
= se
;
7039 static void set_skip_buddy(struct sched_entity
*se
)
7041 for_each_sched_entity(se
)
7042 cfs_rq_of(se
)->skip
= se
;
7046 * Preempt the current task with a newly woken task if needed:
7048 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
7050 struct task_struct
*curr
= rq
->curr
;
7051 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
7052 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
7053 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
7054 int next_buddy_marked
= 0;
7055 int cse_is_idle
, pse_is_idle
;
7057 if (unlikely(se
== pse
))
7061 * This is possible from callers such as attach_tasks(), in which we
7062 * unconditionally check_preempt_curr() after an enqueue (which may have
7063 * lead to a throttle). This both saves work and prevents false
7064 * next-buddy nomination below.
7066 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
7069 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
7070 set_next_buddy(pse
);
7071 next_buddy_marked
= 1;
7075 * We can come here with TIF_NEED_RESCHED already set from new task
7078 * Note: this also catches the edge-case of curr being in a throttled
7079 * group (e.g. via set_curr_task), since update_curr() (in the
7080 * enqueue of curr) will have resulted in resched being set. This
7081 * prevents us from potentially nominating it as a false LAST_BUDDY
7084 if (test_tsk_need_resched(curr
))
7087 /* Idle tasks are by definition preempted by non-idle tasks. */
7088 if (unlikely(task_has_idle_policy(curr
)) &&
7089 likely(!task_has_idle_policy(p
)))
7093 * Batch and idle tasks do not preempt non-idle tasks (their preemption
7094 * is driven by the tick):
7096 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
7099 find_matching_se(&se
, &pse
);
7102 cse_is_idle
= se_is_idle(se
);
7103 pse_is_idle
= se_is_idle(pse
);
7106 * Preempt an idle group in favor of a non-idle group (and don't preempt
7107 * in the inverse case).
7109 if (cse_is_idle
&& !pse_is_idle
)
7111 if (cse_is_idle
!= pse_is_idle
)
7114 update_curr(cfs_rq_of(se
));
7115 if (wakeup_preempt_entity(se
, pse
) == 1) {
7117 * Bias pick_next to pick the sched entity that is
7118 * triggering this preemption.
7120 if (!next_buddy_marked
)
7121 set_next_buddy(pse
);
7130 * Only set the backward buddy when the current task is still
7131 * on the rq. This can happen when a wakeup gets interleaved
7132 * with schedule on the ->pre_schedule() or idle_balance()
7133 * point, either of which can * drop the rq lock.
7135 * Also, during early boot the idle thread is in the fair class,
7136 * for obvious reasons its a bad idea to schedule back to it.
7138 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
7141 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
7146 static struct task_struct
*pick_task_fair(struct rq
*rq
)
7148 struct sched_entity
*se
;
7149 struct cfs_rq
*cfs_rq
;
7153 if (!cfs_rq
->nr_running
)
7157 struct sched_entity
*curr
= cfs_rq
->curr
;
7159 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
7162 update_curr(cfs_rq
);
7166 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
7170 se
= pick_next_entity(cfs_rq
, curr
);
7171 cfs_rq
= group_cfs_rq(se
);
7178 struct task_struct
*
7179 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
7181 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
7182 struct sched_entity
*se
;
7183 struct task_struct
*p
;
7187 if (!sched_fair_runnable(rq
))
7190 #ifdef CONFIG_FAIR_GROUP_SCHED
7191 if (!prev
|| prev
->sched_class
!= &fair_sched_class
)
7195 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7196 * likely that a next task is from the same cgroup as the current.
7198 * Therefore attempt to avoid putting and setting the entire cgroup
7199 * hierarchy, only change the part that actually changes.
7203 struct sched_entity
*curr
= cfs_rq
->curr
;
7206 * Since we got here without doing put_prev_entity() we also
7207 * have to consider cfs_rq->curr. If it is still a runnable
7208 * entity, update_curr() will update its vruntime, otherwise
7209 * forget we've ever seen it.
7213 update_curr(cfs_rq
);
7218 * This call to check_cfs_rq_runtime() will do the
7219 * throttle and dequeue its entity in the parent(s).
7220 * Therefore the nr_running test will indeed
7223 if (unlikely(check_cfs_rq_runtime(cfs_rq
))) {
7226 if (!cfs_rq
->nr_running
)
7233 se
= pick_next_entity(cfs_rq
, curr
);
7234 cfs_rq
= group_cfs_rq(se
);
7240 * Since we haven't yet done put_prev_entity and if the selected task
7241 * is a different task than we started out with, try and touch the
7242 * least amount of cfs_rqs.
7245 struct sched_entity
*pse
= &prev
->se
;
7247 while (!(cfs_rq
= is_same_group(se
, pse
))) {
7248 int se_depth
= se
->depth
;
7249 int pse_depth
= pse
->depth
;
7251 if (se_depth
<= pse_depth
) {
7252 put_prev_entity(cfs_rq_of(pse
), pse
);
7253 pse
= parent_entity(pse
);
7255 if (se_depth
>= pse_depth
) {
7256 set_next_entity(cfs_rq_of(se
), se
);
7257 se
= parent_entity(se
);
7261 put_prev_entity(cfs_rq
, pse
);
7262 set_next_entity(cfs_rq
, se
);
7269 put_prev_task(rq
, prev
);
7272 se
= pick_next_entity(cfs_rq
, NULL
);
7273 set_next_entity(cfs_rq
, se
);
7274 cfs_rq
= group_cfs_rq(se
);
7279 done
: __maybe_unused
;
7282 * Move the next running task to the front of
7283 * the list, so our cfs_tasks list becomes MRU
7286 list_move(&p
->se
.group_node
, &rq
->cfs_tasks
);
7289 if (hrtick_enabled_fair(rq
))
7290 hrtick_start_fair(rq
, p
);
7292 update_misfit_status(p
, rq
);
7300 new_tasks
= newidle_balance(rq
, rf
);
7303 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7304 * possible for any higher priority task to appear. In that case we
7305 * must re-start the pick_next_entity() loop.
7314 * rq is about to be idle, check if we need to update the
7315 * lost_idle_time of clock_pelt
7317 update_idle_rq_clock_pelt(rq
);
7322 static struct task_struct
*__pick_next_task_fair(struct rq
*rq
)
7324 return pick_next_task_fair(rq
, NULL
, NULL
);
7328 * Account for a descheduled task:
7330 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
7332 struct sched_entity
*se
= &prev
->se
;
7333 struct cfs_rq
*cfs_rq
;
7335 for_each_sched_entity(se
) {
7336 cfs_rq
= cfs_rq_of(se
);
7337 put_prev_entity(cfs_rq
, se
);
7342 * sched_yield() is very simple
7344 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7346 static void yield_task_fair(struct rq
*rq
)
7348 struct task_struct
*curr
= rq
->curr
;
7349 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
7350 struct sched_entity
*se
= &curr
->se
;
7353 * Are we the only task in the tree?
7355 if (unlikely(rq
->nr_running
== 1))
7358 clear_buddies(cfs_rq
, se
);
7360 if (curr
->policy
!= SCHED_BATCH
) {
7361 update_rq_clock(rq
);
7363 * Update run-time statistics of the 'current'.
7365 update_curr(cfs_rq
);
7367 * Tell update_rq_clock() that we've just updated,
7368 * so we don't do microscopic update in schedule()
7369 * and double the fastpath cost.
7371 rq_clock_skip_update(rq
);
7377 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
)
7379 struct sched_entity
*se
= &p
->se
;
7381 /* throttled hierarchies are not runnable */
7382 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
7385 /* Tell the scheduler that we'd really like pse to run next. */
7388 yield_task_fair(rq
);
7394 /**************************************************
7395 * Fair scheduling class load-balancing methods.
7399 * The purpose of load-balancing is to achieve the same basic fairness the
7400 * per-CPU scheduler provides, namely provide a proportional amount of compute
7401 * time to each task. This is expressed in the following equation:
7403 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7405 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7406 * W_i,0 is defined as:
7408 * W_i,0 = \Sum_j w_i,j (2)
7410 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7411 * is derived from the nice value as per sched_prio_to_weight[].
7413 * The weight average is an exponential decay average of the instantaneous
7416 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7418 * C_i is the compute capacity of CPU i, typically it is the
7419 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7420 * can also include other factors [XXX].
7422 * To achieve this balance we define a measure of imbalance which follows
7423 * directly from (1):
7425 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7427 * We them move tasks around to minimize the imbalance. In the continuous
7428 * function space it is obvious this converges, in the discrete case we get
7429 * a few fun cases generally called infeasible weight scenarios.
7432 * - infeasible weights;
7433 * - local vs global optima in the discrete case. ]
7438 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7439 * for all i,j solution, we create a tree of CPUs that follows the hardware
7440 * topology where each level pairs two lower groups (or better). This results
7441 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7442 * tree to only the first of the previous level and we decrease the frequency
7443 * of load-balance at each level inv. proportional to the number of CPUs in
7449 * \Sum { --- * --- * 2^i } = O(n) (5)
7451 * `- size of each group
7452 * | | `- number of CPUs doing load-balance
7454 * `- sum over all levels
7456 * Coupled with a limit on how many tasks we can migrate every balance pass,
7457 * this makes (5) the runtime complexity of the balancer.
7459 * An important property here is that each CPU is still (indirectly) connected
7460 * to every other CPU in at most O(log n) steps:
7462 * The adjacency matrix of the resulting graph is given by:
7465 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7468 * And you'll find that:
7470 * A^(log_2 n)_i,j != 0 for all i,j (7)
7472 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7473 * The task movement gives a factor of O(m), giving a convergence complexity
7476 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7481 * In order to avoid CPUs going idle while there's still work to do, new idle
7482 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7483 * tree itself instead of relying on other CPUs to bring it work.
7485 * This adds some complexity to both (5) and (8) but it reduces the total idle
7493 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7496 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7501 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7503 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7505 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7508 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7509 * rewrite all of this once again.]
7512 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
7514 enum fbq_type
{ regular
, remote
, all
};
7517 * 'group_type' describes the group of CPUs at the moment of load balancing.
7519 * The enum is ordered by pulling priority, with the group with lowest priority
7520 * first so the group_type can simply be compared when selecting the busiest
7521 * group. See update_sd_pick_busiest().
7524 /* The group has spare capacity that can be used to run more tasks. */
7525 group_has_spare
= 0,
7527 * The group is fully used and the tasks don't compete for more CPU
7528 * cycles. Nevertheless, some tasks might wait before running.
7532 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
7533 * and must be migrated to a more powerful CPU.
7537 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7538 * and the task should be migrated to it instead of running on the
7543 * The tasks' affinity constraints previously prevented the scheduler
7544 * from balancing the load across the system.
7548 * The CPU is overloaded and can't provide expected CPU cycles to all
7554 enum migration_type
{
7561 #define LBF_ALL_PINNED 0x01
7562 #define LBF_NEED_BREAK 0x02
7563 #define LBF_DST_PINNED 0x04
7564 #define LBF_SOME_PINNED 0x08
7565 #define LBF_ACTIVE_LB 0x10
7568 struct sched_domain
*sd
;
7576 struct cpumask
*dst_grpmask
;
7578 enum cpu_idle_type idle
;
7580 /* The set of CPUs under consideration for load-balancing */
7581 struct cpumask
*cpus
;
7586 unsigned int loop_break
;
7587 unsigned int loop_max
;
7589 enum fbq_type fbq_type
;
7590 enum migration_type migration_type
;
7591 struct list_head tasks
;
7595 * Is this task likely cache-hot:
7597 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
7601 lockdep_assert_rq_held(env
->src_rq
);
7603 if (p
->sched_class
!= &fair_sched_class
)
7606 if (unlikely(task_has_idle_policy(p
)))
7609 /* SMT siblings share cache */
7610 if (env
->sd
->flags
& SD_SHARE_CPUCAPACITY
)
7614 * Buddy candidates are cache hot:
7616 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
7617 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
7618 &p
->se
== cfs_rq_of(&p
->se
)->last
))
7621 if (sysctl_sched_migration_cost
== -1)
7625 * Don't migrate task if the task's cookie does not match
7626 * with the destination CPU's core cookie.
7628 if (!sched_core_cookie_match(cpu_rq(env
->dst_cpu
), p
))
7631 if (sysctl_sched_migration_cost
== 0)
7634 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
7636 return delta
< (s64
)sysctl_sched_migration_cost
;
7639 #ifdef CONFIG_NUMA_BALANCING
7641 * Returns 1, if task migration degrades locality
7642 * Returns 0, if task migration improves locality i.e migration preferred.
7643 * Returns -1, if task migration is not affected by locality.
7645 static int migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
7647 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
7648 unsigned long src_weight
, dst_weight
;
7649 int src_nid
, dst_nid
, dist
;
7651 if (!static_branch_likely(&sched_numa_balancing
))
7654 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
7657 src_nid
= cpu_to_node(env
->src_cpu
);
7658 dst_nid
= cpu_to_node(env
->dst_cpu
);
7660 if (src_nid
== dst_nid
)
7663 /* Migrating away from the preferred node is always bad. */
7664 if (src_nid
== p
->numa_preferred_nid
) {
7665 if (env
->src_rq
->nr_running
> env
->src_rq
->nr_preferred_running
)
7671 /* Encourage migration to the preferred node. */
7672 if (dst_nid
== p
->numa_preferred_nid
)
7675 /* Leaving a core idle is often worse than degrading locality. */
7676 if (env
->idle
== CPU_IDLE
)
7679 dist
= node_distance(src_nid
, dst_nid
);
7681 src_weight
= group_weight(p
, src_nid
, dist
);
7682 dst_weight
= group_weight(p
, dst_nid
, dist
);
7684 src_weight
= task_weight(p
, src_nid
, dist
);
7685 dst_weight
= task_weight(p
, dst_nid
, dist
);
7688 return dst_weight
< src_weight
;
7692 static inline int migrate_degrades_locality(struct task_struct
*p
,
7700 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7703 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
7707 lockdep_assert_rq_held(env
->src_rq
);
7710 * We do not migrate tasks that are:
7711 * 1) throttled_lb_pair, or
7712 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7713 * 3) running (obviously), or
7714 * 4) are cache-hot on their current CPU.
7716 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
7719 /* Disregard pcpu kthreads; they are where they need to be. */
7720 if (kthread_is_per_cpu(p
))
7723 if (!cpumask_test_cpu(env
->dst_cpu
, p
->cpus_ptr
)) {
7726 schedstat_inc(p
->stats
.nr_failed_migrations_affine
);
7728 env
->flags
|= LBF_SOME_PINNED
;
7731 * Remember if this task can be migrated to any other CPU in
7732 * our sched_group. We may want to revisit it if we couldn't
7733 * meet load balance goals by pulling other tasks on src_cpu.
7735 * Avoid computing new_dst_cpu
7737 * - if we have already computed one in current iteration
7738 * - if it's an active balance
7740 if (env
->idle
== CPU_NEWLY_IDLE
||
7741 env
->flags
& (LBF_DST_PINNED
| LBF_ACTIVE_LB
))
7744 /* Prevent to re-select dst_cpu via env's CPUs: */
7745 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
7746 if (cpumask_test_cpu(cpu
, p
->cpus_ptr
)) {
7747 env
->flags
|= LBF_DST_PINNED
;
7748 env
->new_dst_cpu
= cpu
;
7756 /* Record that we found at least one task that could run on dst_cpu */
7757 env
->flags
&= ~LBF_ALL_PINNED
;
7759 if (task_running(env
->src_rq
, p
)) {
7760 schedstat_inc(p
->stats
.nr_failed_migrations_running
);
7765 * Aggressive migration if:
7767 * 2) destination numa is preferred
7768 * 3) task is cache cold, or
7769 * 4) too many balance attempts have failed.
7771 if (env
->flags
& LBF_ACTIVE_LB
)
7774 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
7775 if (tsk_cache_hot
== -1)
7776 tsk_cache_hot
= task_hot(p
, env
);
7778 if (tsk_cache_hot
<= 0 ||
7779 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
7780 if (tsk_cache_hot
== 1) {
7781 schedstat_inc(env
->sd
->lb_hot_gained
[env
->idle
]);
7782 schedstat_inc(p
->stats
.nr_forced_migrations
);
7787 schedstat_inc(p
->stats
.nr_failed_migrations_hot
);
7792 * detach_task() -- detach the task for the migration specified in env
7794 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
7796 lockdep_assert_rq_held(env
->src_rq
);
7798 deactivate_task(env
->src_rq
, p
, DEQUEUE_NOCLOCK
);
7799 set_task_cpu(p
, env
->dst_cpu
);
7803 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7804 * part of active balancing operations within "domain".
7806 * Returns a task if successful and NULL otherwise.
7808 static struct task_struct
*detach_one_task(struct lb_env
*env
)
7810 struct task_struct
*p
;
7812 lockdep_assert_rq_held(env
->src_rq
);
7814 list_for_each_entry_reverse(p
,
7815 &env
->src_rq
->cfs_tasks
, se
.group_node
) {
7816 if (!can_migrate_task(p
, env
))
7819 detach_task(p
, env
);
7822 * Right now, this is only the second place where
7823 * lb_gained[env->idle] is updated (other is detach_tasks)
7824 * so we can safely collect stats here rather than
7825 * inside detach_tasks().
7827 schedstat_inc(env
->sd
->lb_gained
[env
->idle
]);
7833 static const unsigned int sched_nr_migrate_break
= 32;
7836 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7837 * busiest_rq, as part of a balancing operation within domain "sd".
7839 * Returns number of detached tasks if successful and 0 otherwise.
7841 static int detach_tasks(struct lb_env
*env
)
7843 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
7844 unsigned long util
, load
;
7845 struct task_struct
*p
;
7848 lockdep_assert_rq_held(env
->src_rq
);
7851 * Source run queue has been emptied by another CPU, clear
7852 * LBF_ALL_PINNED flag as we will not test any task.
7854 if (env
->src_rq
->nr_running
<= 1) {
7855 env
->flags
&= ~LBF_ALL_PINNED
;
7859 if (env
->imbalance
<= 0)
7862 while (!list_empty(tasks
)) {
7864 * We don't want to steal all, otherwise we may be treated likewise,
7865 * which could at worst lead to a livelock crash.
7867 if (env
->idle
!= CPU_NOT_IDLE
&& env
->src_rq
->nr_running
<= 1)
7870 p
= list_last_entry(tasks
, struct task_struct
, se
.group_node
);
7873 /* We've more or less seen every task there is, call it quits */
7874 if (env
->loop
> env
->loop_max
)
7877 /* take a breather every nr_migrate tasks */
7878 if (env
->loop
> env
->loop_break
) {
7879 env
->loop_break
+= sched_nr_migrate_break
;
7880 env
->flags
|= LBF_NEED_BREAK
;
7884 if (!can_migrate_task(p
, env
))
7887 switch (env
->migration_type
) {
7890 * Depending of the number of CPUs and tasks and the
7891 * cgroup hierarchy, task_h_load() can return a null
7892 * value. Make sure that env->imbalance decreases
7893 * otherwise detach_tasks() will stop only after
7894 * detaching up to loop_max tasks.
7896 load
= max_t(unsigned long, task_h_load(p
), 1);
7898 if (sched_feat(LB_MIN
) &&
7899 load
< 16 && !env
->sd
->nr_balance_failed
)
7903 * Make sure that we don't migrate too much load.
7904 * Nevertheless, let relax the constraint if
7905 * scheduler fails to find a good waiting task to
7908 if (shr_bound(load
, env
->sd
->nr_balance_failed
) > env
->imbalance
)
7911 env
->imbalance
-= load
;
7915 util
= task_util_est(p
);
7917 if (util
> env
->imbalance
)
7920 env
->imbalance
-= util
;
7927 case migrate_misfit
:
7928 /* This is not a misfit task */
7929 if (task_fits_capacity(p
, capacity_of(env
->src_cpu
)))
7936 detach_task(p
, env
);
7937 list_add(&p
->se
.group_node
, &env
->tasks
);
7941 #ifdef CONFIG_PREEMPTION
7943 * NEWIDLE balancing is a source of latency, so preemptible
7944 * kernels will stop after the first task is detached to minimize
7945 * the critical section.
7947 if (env
->idle
== CPU_NEWLY_IDLE
)
7952 * We only want to steal up to the prescribed amount of
7955 if (env
->imbalance
<= 0)
7960 list_move(&p
->se
.group_node
, tasks
);
7964 * Right now, this is one of only two places we collect this stat
7965 * so we can safely collect detach_one_task() stats here rather
7966 * than inside detach_one_task().
7968 schedstat_add(env
->sd
->lb_gained
[env
->idle
], detached
);
7974 * attach_task() -- attach the task detached by detach_task() to its new rq.
7976 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
7978 lockdep_assert_rq_held(rq
);
7980 BUG_ON(task_rq(p
) != rq
);
7981 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
7982 check_preempt_curr(rq
, p
, 0);
7986 * attach_one_task() -- attaches the task returned from detach_one_task() to
7989 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
7994 update_rq_clock(rq
);
8000 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
8003 static void attach_tasks(struct lb_env
*env
)
8005 struct list_head
*tasks
= &env
->tasks
;
8006 struct task_struct
*p
;
8009 rq_lock(env
->dst_rq
, &rf
);
8010 update_rq_clock(env
->dst_rq
);
8012 while (!list_empty(tasks
)) {
8013 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
8014 list_del_init(&p
->se
.group_node
);
8016 attach_task(env
->dst_rq
, p
);
8019 rq_unlock(env
->dst_rq
, &rf
);
8022 #ifdef CONFIG_NO_HZ_COMMON
8023 static inline bool cfs_rq_has_blocked(struct cfs_rq
*cfs_rq
)
8025 if (cfs_rq
->avg
.load_avg
)
8028 if (cfs_rq
->avg
.util_avg
)
8034 static inline bool others_have_blocked(struct rq
*rq
)
8036 if (READ_ONCE(rq
->avg_rt
.util_avg
))
8039 if (READ_ONCE(rq
->avg_dl
.util_avg
))
8042 if (thermal_load_avg(rq
))
8045 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
8046 if (READ_ONCE(rq
->avg_irq
.util_avg
))
8053 static inline void update_blocked_load_tick(struct rq
*rq
)
8055 WRITE_ONCE(rq
->last_blocked_load_update_tick
, jiffies
);
8058 static inline void update_blocked_load_status(struct rq
*rq
, bool has_blocked
)
8061 rq
->has_blocked_load
= 0;
8064 static inline bool cfs_rq_has_blocked(struct cfs_rq
*cfs_rq
) { return false; }
8065 static inline bool others_have_blocked(struct rq
*rq
) { return false; }
8066 static inline void update_blocked_load_tick(struct rq
*rq
) {}
8067 static inline void update_blocked_load_status(struct rq
*rq
, bool has_blocked
) {}
8070 static bool __update_blocked_others(struct rq
*rq
, bool *done
)
8072 const struct sched_class
*curr_class
;
8073 u64 now
= rq_clock_pelt(rq
);
8074 unsigned long thermal_pressure
;
8078 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
8079 * DL and IRQ signals have been updated before updating CFS.
8081 curr_class
= rq
->curr
->sched_class
;
8083 thermal_pressure
= arch_scale_thermal_pressure(cpu_of(rq
));
8085 decayed
= update_rt_rq_load_avg(now
, rq
, curr_class
== &rt_sched_class
) |
8086 update_dl_rq_load_avg(now
, rq
, curr_class
== &dl_sched_class
) |
8087 update_thermal_load_avg(rq_clock_thermal(rq
), rq
, thermal_pressure
) |
8088 update_irq_load_avg(rq
, 0);
8090 if (others_have_blocked(rq
))
8096 #ifdef CONFIG_FAIR_GROUP_SCHED
8098 static bool __update_blocked_fair(struct rq
*rq
, bool *done
)
8100 struct cfs_rq
*cfs_rq
, *pos
;
8101 bool decayed
= false;
8102 int cpu
= cpu_of(rq
);
8105 * Iterates the task_group tree in a bottom up fashion, see
8106 * list_add_leaf_cfs_rq() for details.
8108 for_each_leaf_cfs_rq_safe(rq
, cfs_rq
, pos
) {
8109 struct sched_entity
*se
;
8111 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq
), cfs_rq
)) {
8112 update_tg_load_avg(cfs_rq
);
8114 if (cfs_rq
== &rq
->cfs
)
8118 /* Propagate pending load changes to the parent, if any: */
8119 se
= cfs_rq
->tg
->se
[cpu
];
8120 if (se
&& !skip_blocked_update(se
))
8121 update_load_avg(cfs_rq_of(se
), se
, UPDATE_TG
);
8124 * There can be a lot of idle CPU cgroups. Don't let fully
8125 * decayed cfs_rqs linger on the list.
8127 if (cfs_rq_is_decayed(cfs_rq
))
8128 list_del_leaf_cfs_rq(cfs_rq
);
8130 /* Don't need periodic decay once load/util_avg are null */
8131 if (cfs_rq_has_blocked(cfs_rq
))
8139 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
8140 * This needs to be done in a top-down fashion because the load of a child
8141 * group is a fraction of its parents load.
8143 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
8145 struct rq
*rq
= rq_of(cfs_rq
);
8146 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
8147 unsigned long now
= jiffies
;
8150 if (cfs_rq
->last_h_load_update
== now
)
8153 WRITE_ONCE(cfs_rq
->h_load_next
, NULL
);
8154 for_each_sched_entity(se
) {
8155 cfs_rq
= cfs_rq_of(se
);
8156 WRITE_ONCE(cfs_rq
->h_load_next
, se
);
8157 if (cfs_rq
->last_h_load_update
== now
)
8162 cfs_rq
->h_load
= cfs_rq_load_avg(cfs_rq
);
8163 cfs_rq
->last_h_load_update
= now
;
8166 while ((se
= READ_ONCE(cfs_rq
->h_load_next
)) != NULL
) {
8167 load
= cfs_rq
->h_load
;
8168 load
= div64_ul(load
* se
->avg
.load_avg
,
8169 cfs_rq_load_avg(cfs_rq
) + 1);
8170 cfs_rq
= group_cfs_rq(se
);
8171 cfs_rq
->h_load
= load
;
8172 cfs_rq
->last_h_load_update
= now
;
8176 static unsigned long task_h_load(struct task_struct
*p
)
8178 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
8180 update_cfs_rq_h_load(cfs_rq
);
8181 return div64_ul(p
->se
.avg
.load_avg
* cfs_rq
->h_load
,
8182 cfs_rq_load_avg(cfs_rq
) + 1);
8185 static bool __update_blocked_fair(struct rq
*rq
, bool *done
)
8187 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
8190 decayed
= update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq
), cfs_rq
);
8191 if (cfs_rq_has_blocked(cfs_rq
))
8197 static unsigned long task_h_load(struct task_struct
*p
)
8199 return p
->se
.avg
.load_avg
;
8203 static void update_blocked_averages(int cpu
)
8205 bool decayed
= false, done
= true;
8206 struct rq
*rq
= cpu_rq(cpu
);
8209 rq_lock_irqsave(rq
, &rf
);
8210 update_blocked_load_tick(rq
);
8211 update_rq_clock(rq
);
8213 decayed
|= __update_blocked_others(rq
, &done
);
8214 decayed
|= __update_blocked_fair(rq
, &done
);
8216 update_blocked_load_status(rq
, !done
);
8218 cpufreq_update_util(rq
, 0);
8219 rq_unlock_irqrestore(rq
, &rf
);
8222 /********** Helpers for find_busiest_group ************************/
8225 * sg_lb_stats - stats of a sched_group required for load_balancing
8227 struct sg_lb_stats
{
8228 unsigned long avg_load
; /*Avg load across the CPUs of the group */
8229 unsigned long group_load
; /* Total load over the CPUs of the group */
8230 unsigned long group_capacity
;
8231 unsigned long group_util
; /* Total utilization over the CPUs of the group */
8232 unsigned long group_runnable
; /* Total runnable time over the CPUs of the group */
8233 unsigned int sum_nr_running
; /* Nr of tasks running in the group */
8234 unsigned int sum_h_nr_running
; /* Nr of CFS tasks running in the group */
8235 unsigned int idle_cpus
;
8236 unsigned int group_weight
;
8237 enum group_type group_type
;
8238 unsigned int group_asym_packing
; /* Tasks should be moved to preferred CPU */
8239 unsigned long group_misfit_task_load
; /* A CPU has a task too big for its capacity */
8240 #ifdef CONFIG_NUMA_BALANCING
8241 unsigned int nr_numa_running
;
8242 unsigned int nr_preferred_running
;
8247 * sd_lb_stats - Structure to store the statistics of a sched_domain
8248 * during load balancing.
8250 struct sd_lb_stats
{
8251 struct sched_group
*busiest
; /* Busiest group in this sd */
8252 struct sched_group
*local
; /* Local group in this sd */
8253 unsigned long total_load
; /* Total load of all groups in sd */
8254 unsigned long total_capacity
; /* Total capacity of all groups in sd */
8255 unsigned long avg_load
; /* Average load across all groups in sd */
8256 unsigned int prefer_sibling
; /* tasks should go to sibling first */
8258 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
8259 struct sg_lb_stats local_stat
; /* Statistics of the local group */
8262 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
8265 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8266 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8267 * We must however set busiest_stat::group_type and
8268 * busiest_stat::idle_cpus to the worst busiest group because
8269 * update_sd_pick_busiest() reads these before assignment.
8271 *sds
= (struct sd_lb_stats
){
8275 .total_capacity
= 0UL,
8277 .idle_cpus
= UINT_MAX
,
8278 .group_type
= group_has_spare
,
8283 static unsigned long scale_rt_capacity(int cpu
)
8285 struct rq
*rq
= cpu_rq(cpu
);
8286 unsigned long max
= arch_scale_cpu_capacity(cpu
);
8287 unsigned long used
, free
;
8290 irq
= cpu_util_irq(rq
);
8292 if (unlikely(irq
>= max
))
8296 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8297 * (running and not running) with weights 0 and 1024 respectively.
8298 * avg_thermal.load_avg tracks thermal pressure and the weighted
8299 * average uses the actual delta max capacity(load).
8301 used
= READ_ONCE(rq
->avg_rt
.util_avg
);
8302 used
+= READ_ONCE(rq
->avg_dl
.util_avg
);
8303 used
+= thermal_load_avg(rq
);
8305 if (unlikely(used
>= max
))
8310 return scale_irq_capacity(free
, irq
, max
);
8313 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
8315 unsigned long capacity
= scale_rt_capacity(cpu
);
8316 struct sched_group
*sdg
= sd
->groups
;
8318 cpu_rq(cpu
)->cpu_capacity_orig
= arch_scale_cpu_capacity(cpu
);
8323 cpu_rq(cpu
)->cpu_capacity
= capacity
;
8324 trace_sched_cpu_capacity_tp(cpu_rq(cpu
));
8326 sdg
->sgc
->capacity
= capacity
;
8327 sdg
->sgc
->min_capacity
= capacity
;
8328 sdg
->sgc
->max_capacity
= capacity
;
8331 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
8333 struct sched_domain
*child
= sd
->child
;
8334 struct sched_group
*group
, *sdg
= sd
->groups
;
8335 unsigned long capacity
, min_capacity
, max_capacity
;
8336 unsigned long interval
;
8338 interval
= msecs_to_jiffies(sd
->balance_interval
);
8339 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
8340 sdg
->sgc
->next_update
= jiffies
+ interval
;
8343 update_cpu_capacity(sd
, cpu
);
8348 min_capacity
= ULONG_MAX
;
8351 if (child
->flags
& SD_OVERLAP
) {
8353 * SD_OVERLAP domains cannot assume that child groups
8354 * span the current group.
8357 for_each_cpu(cpu
, sched_group_span(sdg
)) {
8358 unsigned long cpu_cap
= capacity_of(cpu
);
8360 capacity
+= cpu_cap
;
8361 min_capacity
= min(cpu_cap
, min_capacity
);
8362 max_capacity
= max(cpu_cap
, max_capacity
);
8366 * !SD_OVERLAP domains can assume that child groups
8367 * span the current group.
8370 group
= child
->groups
;
8372 struct sched_group_capacity
*sgc
= group
->sgc
;
8374 capacity
+= sgc
->capacity
;
8375 min_capacity
= min(sgc
->min_capacity
, min_capacity
);
8376 max_capacity
= max(sgc
->max_capacity
, max_capacity
);
8377 group
= group
->next
;
8378 } while (group
!= child
->groups
);
8381 sdg
->sgc
->capacity
= capacity
;
8382 sdg
->sgc
->min_capacity
= min_capacity
;
8383 sdg
->sgc
->max_capacity
= max_capacity
;
8387 * Check whether the capacity of the rq has been noticeably reduced by side
8388 * activity. The imbalance_pct is used for the threshold.
8389 * Return true is the capacity is reduced
8392 check_cpu_capacity(struct rq
*rq
, struct sched_domain
*sd
)
8394 return ((rq
->cpu_capacity
* sd
->imbalance_pct
) <
8395 (rq
->cpu_capacity_orig
* 100));
8399 * Check whether a rq has a misfit task and if it looks like we can actually
8400 * help that task: we can migrate the task to a CPU of higher capacity, or
8401 * the task's current CPU is heavily pressured.
8403 static inline int check_misfit_status(struct rq
*rq
, struct sched_domain
*sd
)
8405 return rq
->misfit_task_load
&&
8406 (rq
->cpu_capacity_orig
< rq
->rd
->max_cpu_capacity
||
8407 check_cpu_capacity(rq
, sd
));
8411 * Group imbalance indicates (and tries to solve) the problem where balancing
8412 * groups is inadequate due to ->cpus_ptr constraints.
8414 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8415 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8418 * { 0 1 2 3 } { 4 5 6 7 }
8421 * If we were to balance group-wise we'd place two tasks in the first group and
8422 * two tasks in the second group. Clearly this is undesired as it will overload
8423 * cpu 3 and leave one of the CPUs in the second group unused.
8425 * The current solution to this issue is detecting the skew in the first group
8426 * by noticing the lower domain failed to reach balance and had difficulty
8427 * moving tasks due to affinity constraints.
8429 * When this is so detected; this group becomes a candidate for busiest; see
8430 * update_sd_pick_busiest(). And calculate_imbalance() and
8431 * find_busiest_group() avoid some of the usual balance conditions to allow it
8432 * to create an effective group imbalance.
8434 * This is a somewhat tricky proposition since the next run might not find the
8435 * group imbalance and decide the groups need to be balanced again. A most
8436 * subtle and fragile situation.
8439 static inline int sg_imbalanced(struct sched_group
*group
)
8441 return group
->sgc
->imbalance
;
8445 * group_has_capacity returns true if the group has spare capacity that could
8446 * be used by some tasks.
8447 * We consider that a group has spare capacity if the * number of task is
8448 * smaller than the number of CPUs or if the utilization is lower than the
8449 * available capacity for CFS tasks.
8450 * For the latter, we use a threshold to stabilize the state, to take into
8451 * account the variance of the tasks' load and to return true if the available
8452 * capacity in meaningful for the load balancer.
8453 * As an example, an available capacity of 1% can appear but it doesn't make
8454 * any benefit for the load balance.
8457 group_has_capacity(unsigned int imbalance_pct
, struct sg_lb_stats
*sgs
)
8459 if (sgs
->sum_nr_running
< sgs
->group_weight
)
8462 if ((sgs
->group_capacity
* imbalance_pct
) <
8463 (sgs
->group_runnable
* 100))
8466 if ((sgs
->group_capacity
* 100) >
8467 (sgs
->group_util
* imbalance_pct
))
8474 * group_is_overloaded returns true if the group has more tasks than it can
8476 * group_is_overloaded is not equals to !group_has_capacity because a group
8477 * with the exact right number of tasks, has no more spare capacity but is not
8478 * overloaded so both group_has_capacity and group_is_overloaded return
8482 group_is_overloaded(unsigned int imbalance_pct
, struct sg_lb_stats
*sgs
)
8484 if (sgs
->sum_nr_running
<= sgs
->group_weight
)
8487 if ((sgs
->group_capacity
* 100) <
8488 (sgs
->group_util
* imbalance_pct
))
8491 if ((sgs
->group_capacity
* imbalance_pct
) <
8492 (sgs
->group_runnable
* 100))
8499 group_type
group_classify(unsigned int imbalance_pct
,
8500 struct sched_group
*group
,
8501 struct sg_lb_stats
*sgs
)
8503 if (group_is_overloaded(imbalance_pct
, sgs
))
8504 return group_overloaded
;
8506 if (sg_imbalanced(group
))
8507 return group_imbalanced
;
8509 if (sgs
->group_asym_packing
)
8510 return group_asym_packing
;
8512 if (sgs
->group_misfit_task_load
)
8513 return group_misfit_task
;
8515 if (!group_has_capacity(imbalance_pct
, sgs
))
8516 return group_fully_busy
;
8518 return group_has_spare
;
8522 * asym_smt_can_pull_tasks - Check whether the load balancing CPU can pull tasks
8523 * @dst_cpu: Destination CPU of the load balancing
8524 * @sds: Load-balancing data with statistics of the local group
8525 * @sgs: Load-balancing statistics of the candidate busiest group
8526 * @sg: The candidate busiest group
8528 * Check the state of the SMT siblings of both @sds::local and @sg and decide
8529 * if @dst_cpu can pull tasks.
8531 * If @dst_cpu does not have SMT siblings, it can pull tasks if two or more of
8532 * the SMT siblings of @sg are busy. If only one CPU in @sg is busy, pull tasks
8533 * only if @dst_cpu has higher priority.
8535 * If both @dst_cpu and @sg have SMT siblings, and @sg has exactly one more
8536 * busy CPU than @sds::local, let @dst_cpu pull tasks if it has higher priority.
8537 * Bigger imbalances in the number of busy CPUs will be dealt with in
8538 * update_sd_pick_busiest().
8540 * If @sg does not have SMT siblings, only pull tasks if all of the SMT siblings
8541 * of @dst_cpu are idle and @sg has lower priority.
8543 static bool asym_smt_can_pull_tasks(int dst_cpu
, struct sd_lb_stats
*sds
,
8544 struct sg_lb_stats
*sgs
,
8545 struct sched_group
*sg
)
8547 #ifdef CONFIG_SCHED_SMT
8548 bool local_is_smt
, sg_is_smt
;
8551 local_is_smt
= sds
->local
->flags
& SD_SHARE_CPUCAPACITY
;
8552 sg_is_smt
= sg
->flags
& SD_SHARE_CPUCAPACITY
;
8554 sg_busy_cpus
= sgs
->group_weight
- sgs
->idle_cpus
;
8556 if (!local_is_smt
) {
8558 * If we are here, @dst_cpu is idle and does not have SMT
8559 * siblings. Pull tasks if candidate group has two or more
8562 if (sg_busy_cpus
>= 2) /* implies sg_is_smt */
8566 * @dst_cpu does not have SMT siblings. @sg may have SMT
8567 * siblings and only one is busy. In such case, @dst_cpu
8568 * can help if it has higher priority and is idle (i.e.,
8569 * it has no running tasks).
8571 return sched_asym_prefer(dst_cpu
, sg
->asym_prefer_cpu
);
8574 /* @dst_cpu has SMT siblings. */
8577 int local_busy_cpus
= sds
->local
->group_weight
-
8578 sds
->local_stat
.idle_cpus
;
8579 int busy_cpus_delta
= sg_busy_cpus
- local_busy_cpus
;
8581 if (busy_cpus_delta
== 1)
8582 return sched_asym_prefer(dst_cpu
, sg
->asym_prefer_cpu
);
8588 * @sg does not have SMT siblings. Ensure that @sds::local does not end
8589 * up with more than one busy SMT sibling and only pull tasks if there
8590 * are not busy CPUs (i.e., no CPU has running tasks).
8592 if (!sds
->local_stat
.sum_nr_running
)
8593 return sched_asym_prefer(dst_cpu
, sg
->asym_prefer_cpu
);
8597 /* Always return false so that callers deal with non-SMT cases. */
8603 sched_asym(struct lb_env
*env
, struct sd_lb_stats
*sds
, struct sg_lb_stats
*sgs
,
8604 struct sched_group
*group
)
8606 /* Only do SMT checks if either local or candidate have SMT siblings */
8607 if ((sds
->local
->flags
& SD_SHARE_CPUCAPACITY
) ||
8608 (group
->flags
& SD_SHARE_CPUCAPACITY
))
8609 return asym_smt_can_pull_tasks(env
->dst_cpu
, sds
, sgs
, group
);
8611 return sched_asym_prefer(env
->dst_cpu
, group
->asym_prefer_cpu
);
8615 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8616 * @env: The load balancing environment.
8617 * @group: sched_group whose statistics are to be updated.
8618 * @sgs: variable to hold the statistics for this group.
8619 * @sg_status: Holds flag indicating the status of the sched_group
8621 static inline void update_sg_lb_stats(struct lb_env
*env
,
8622 struct sd_lb_stats
*sds
,
8623 struct sched_group
*group
,
8624 struct sg_lb_stats
*sgs
,
8627 int i
, nr_running
, local_group
;
8629 memset(sgs
, 0, sizeof(*sgs
));
8631 local_group
= group
== sds
->local
;
8633 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
8634 struct rq
*rq
= cpu_rq(i
);
8636 sgs
->group_load
+= cpu_load(rq
);
8637 sgs
->group_util
+= cpu_util_cfs(i
);
8638 sgs
->group_runnable
+= cpu_runnable(rq
);
8639 sgs
->sum_h_nr_running
+= rq
->cfs
.h_nr_running
;
8641 nr_running
= rq
->nr_running
;
8642 sgs
->sum_nr_running
+= nr_running
;
8645 *sg_status
|= SG_OVERLOAD
;
8647 if (cpu_overutilized(i
))
8648 *sg_status
|= SG_OVERUTILIZED
;
8650 #ifdef CONFIG_NUMA_BALANCING
8651 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
8652 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
8655 * No need to call idle_cpu() if nr_running is not 0
8657 if (!nr_running
&& idle_cpu(i
)) {
8659 /* Idle cpu can't have misfit task */
8666 /* Check for a misfit task on the cpu */
8667 if (env
->sd
->flags
& SD_ASYM_CPUCAPACITY
&&
8668 sgs
->group_misfit_task_load
< rq
->misfit_task_load
) {
8669 sgs
->group_misfit_task_load
= rq
->misfit_task_load
;
8670 *sg_status
|= SG_OVERLOAD
;
8674 sgs
->group_capacity
= group
->sgc
->capacity
;
8676 sgs
->group_weight
= group
->group_weight
;
8678 /* Check if dst CPU is idle and preferred to this group */
8679 if (!local_group
&& env
->sd
->flags
& SD_ASYM_PACKING
&&
8680 env
->idle
!= CPU_NOT_IDLE
&& sgs
->sum_h_nr_running
&&
8681 sched_asym(env
, sds
, sgs
, group
)) {
8682 sgs
->group_asym_packing
= 1;
8685 sgs
->group_type
= group_classify(env
->sd
->imbalance_pct
, group
, sgs
);
8687 /* Computing avg_load makes sense only when group is overloaded */
8688 if (sgs
->group_type
== group_overloaded
)
8689 sgs
->avg_load
= (sgs
->group_load
* SCHED_CAPACITY_SCALE
) /
8690 sgs
->group_capacity
;
8694 * update_sd_pick_busiest - return 1 on busiest group
8695 * @env: The load balancing environment.
8696 * @sds: sched_domain statistics
8697 * @sg: sched_group candidate to be checked for being the busiest
8698 * @sgs: sched_group statistics
8700 * Determine if @sg is a busier group than the previously selected
8703 * Return: %true if @sg is a busier group than the previously selected
8704 * busiest group. %false otherwise.
8706 static bool update_sd_pick_busiest(struct lb_env
*env
,
8707 struct sd_lb_stats
*sds
,
8708 struct sched_group
*sg
,
8709 struct sg_lb_stats
*sgs
)
8711 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
8713 /* Make sure that there is at least one task to pull */
8714 if (!sgs
->sum_h_nr_running
)
8718 * Don't try to pull misfit tasks we can't help.
8719 * We can use max_capacity here as reduction in capacity on some
8720 * CPUs in the group should either be possible to resolve
8721 * internally or be covered by avg_load imbalance (eventually).
8723 if (sgs
->group_type
== group_misfit_task
&&
8724 (!capacity_greater(capacity_of(env
->dst_cpu
), sg
->sgc
->max_capacity
) ||
8725 sds
->local_stat
.group_type
!= group_has_spare
))
8728 if (sgs
->group_type
> busiest
->group_type
)
8731 if (sgs
->group_type
< busiest
->group_type
)
8735 * The candidate and the current busiest group are the same type of
8736 * group. Let check which one is the busiest according to the type.
8739 switch (sgs
->group_type
) {
8740 case group_overloaded
:
8741 /* Select the overloaded group with highest avg_load. */
8742 if (sgs
->avg_load
<= busiest
->avg_load
)
8746 case group_imbalanced
:
8748 * Select the 1st imbalanced group as we don't have any way to
8749 * choose one more than another.
8753 case group_asym_packing
:
8754 /* Prefer to move from lowest priority CPU's work */
8755 if (sched_asym_prefer(sg
->asym_prefer_cpu
, sds
->busiest
->asym_prefer_cpu
))
8759 case group_misfit_task
:
8761 * If we have more than one misfit sg go with the biggest
8764 if (sgs
->group_misfit_task_load
< busiest
->group_misfit_task_load
)
8768 case group_fully_busy
:
8770 * Select the fully busy group with highest avg_load. In
8771 * theory, there is no need to pull task from such kind of
8772 * group because tasks have all compute capacity that they need
8773 * but we can still improve the overall throughput by reducing
8774 * contention when accessing shared HW resources.
8776 * XXX for now avg_load is not computed and always 0 so we
8777 * select the 1st one.
8779 if (sgs
->avg_load
<= busiest
->avg_load
)
8783 case group_has_spare
:
8785 * Select not overloaded group with lowest number of idle cpus
8786 * and highest number of running tasks. We could also compare
8787 * the spare capacity which is more stable but it can end up
8788 * that the group has less spare capacity but finally more idle
8789 * CPUs which means less opportunity to pull tasks.
8791 if (sgs
->idle_cpus
> busiest
->idle_cpus
)
8793 else if ((sgs
->idle_cpus
== busiest
->idle_cpus
) &&
8794 (sgs
->sum_nr_running
<= busiest
->sum_nr_running
))
8801 * Candidate sg has no more than one task per CPU and has higher
8802 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8803 * throughput. Maximize throughput, power/energy consequences are not
8806 if ((env
->sd
->flags
& SD_ASYM_CPUCAPACITY
) &&
8807 (sgs
->group_type
<= group_fully_busy
) &&
8808 (capacity_greater(sg
->sgc
->min_capacity
, capacity_of(env
->dst_cpu
))))
8814 #ifdef CONFIG_NUMA_BALANCING
8815 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
8817 if (sgs
->sum_h_nr_running
> sgs
->nr_numa_running
)
8819 if (sgs
->sum_h_nr_running
> sgs
->nr_preferred_running
)
8824 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
8826 if (rq
->nr_running
> rq
->nr_numa_running
)
8828 if (rq
->nr_running
> rq
->nr_preferred_running
)
8833 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
8838 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
8842 #endif /* CONFIG_NUMA_BALANCING */
8848 * task_running_on_cpu - return 1 if @p is running on @cpu.
8851 static unsigned int task_running_on_cpu(int cpu
, struct task_struct
*p
)
8853 /* Task has no contribution or is new */
8854 if (cpu
!= task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
8857 if (task_on_rq_queued(p
))
8864 * idle_cpu_without - would a given CPU be idle without p ?
8865 * @cpu: the processor on which idleness is tested.
8866 * @p: task which should be ignored.
8868 * Return: 1 if the CPU would be idle. 0 otherwise.
8870 static int idle_cpu_without(int cpu
, struct task_struct
*p
)
8872 struct rq
*rq
= cpu_rq(cpu
);
8874 if (rq
->curr
!= rq
->idle
&& rq
->curr
!= p
)
8878 * rq->nr_running can't be used but an updated version without the
8879 * impact of p on cpu must be used instead. The updated nr_running
8880 * be computed and tested before calling idle_cpu_without().
8884 if (rq
->ttwu_pending
)
8892 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8893 * @sd: The sched_domain level to look for idlest group.
8894 * @group: sched_group whose statistics are to be updated.
8895 * @sgs: variable to hold the statistics for this group.
8896 * @p: The task for which we look for the idlest group/CPU.
8898 static inline void update_sg_wakeup_stats(struct sched_domain
*sd
,
8899 struct sched_group
*group
,
8900 struct sg_lb_stats
*sgs
,
8901 struct task_struct
*p
)
8905 memset(sgs
, 0, sizeof(*sgs
));
8907 for_each_cpu(i
, sched_group_span(group
)) {
8908 struct rq
*rq
= cpu_rq(i
);
8911 sgs
->group_load
+= cpu_load_without(rq
, p
);
8912 sgs
->group_util
+= cpu_util_without(i
, p
);
8913 sgs
->group_runnable
+= cpu_runnable_without(rq
, p
);
8914 local
= task_running_on_cpu(i
, p
);
8915 sgs
->sum_h_nr_running
+= rq
->cfs
.h_nr_running
- local
;
8917 nr_running
= rq
->nr_running
- local
;
8918 sgs
->sum_nr_running
+= nr_running
;
8921 * No need to call idle_cpu_without() if nr_running is not 0
8923 if (!nr_running
&& idle_cpu_without(i
, p
))
8928 /* Check if task fits in the group */
8929 if (sd
->flags
& SD_ASYM_CPUCAPACITY
&&
8930 !task_fits_capacity(p
, group
->sgc
->max_capacity
)) {
8931 sgs
->group_misfit_task_load
= 1;
8934 sgs
->group_capacity
= group
->sgc
->capacity
;
8936 sgs
->group_weight
= group
->group_weight
;
8938 sgs
->group_type
= group_classify(sd
->imbalance_pct
, group
, sgs
);
8941 * Computing avg_load makes sense only when group is fully busy or
8944 if (sgs
->group_type
== group_fully_busy
||
8945 sgs
->group_type
== group_overloaded
)
8946 sgs
->avg_load
= (sgs
->group_load
* SCHED_CAPACITY_SCALE
) /
8947 sgs
->group_capacity
;
8950 static bool update_pick_idlest(struct sched_group
*idlest
,
8951 struct sg_lb_stats
*idlest_sgs
,
8952 struct sched_group
*group
,
8953 struct sg_lb_stats
*sgs
)
8955 if (sgs
->group_type
< idlest_sgs
->group_type
)
8958 if (sgs
->group_type
> idlest_sgs
->group_type
)
8962 * The candidate and the current idlest group are the same type of
8963 * group. Let check which one is the idlest according to the type.
8966 switch (sgs
->group_type
) {
8967 case group_overloaded
:
8968 case group_fully_busy
:
8969 /* Select the group with lowest avg_load. */
8970 if (idlest_sgs
->avg_load
<= sgs
->avg_load
)
8974 case group_imbalanced
:
8975 case group_asym_packing
:
8976 /* Those types are not used in the slow wakeup path */
8979 case group_misfit_task
:
8980 /* Select group with the highest max capacity */
8981 if (idlest
->sgc
->max_capacity
>= group
->sgc
->max_capacity
)
8985 case group_has_spare
:
8986 /* Select group with most idle CPUs */
8987 if (idlest_sgs
->idle_cpus
> sgs
->idle_cpus
)
8990 /* Select group with lowest group_util */
8991 if (idlest_sgs
->idle_cpus
== sgs
->idle_cpus
&&
8992 idlest_sgs
->group_util
<= sgs
->group_util
)
9002 * Allow a NUMA imbalance if busy CPUs is less than 25% of the domain.
9003 * This is an approximation as the number of running tasks may not be
9004 * related to the number of busy CPUs due to sched_setaffinity.
9006 static inline bool allow_numa_imbalance(int dst_running
, int dst_weight
)
9008 return (dst_running
< (dst_weight
>> 2));
9012 * find_idlest_group() finds and returns the least busy CPU group within the
9015 * Assumes p is allowed on at least one CPU in sd.
9017 static struct sched_group
*
9018 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
9020 struct sched_group
*idlest
= NULL
, *local
= NULL
, *group
= sd
->groups
;
9021 struct sg_lb_stats local_sgs
, tmp_sgs
;
9022 struct sg_lb_stats
*sgs
;
9023 unsigned long imbalance
;
9024 struct sg_lb_stats idlest_sgs
= {
9025 .avg_load
= UINT_MAX
,
9026 .group_type
= group_overloaded
,
9032 /* Skip over this group if it has no CPUs allowed */
9033 if (!cpumask_intersects(sched_group_span(group
),
9037 /* Skip over this group if no cookie matched */
9038 if (!sched_group_cookie_match(cpu_rq(this_cpu
), p
, group
))
9041 local_group
= cpumask_test_cpu(this_cpu
,
9042 sched_group_span(group
));
9051 update_sg_wakeup_stats(sd
, group
, sgs
, p
);
9053 if (!local_group
&& update_pick_idlest(idlest
, &idlest_sgs
, group
, sgs
)) {
9058 } while (group
= group
->next
, group
!= sd
->groups
);
9061 /* There is no idlest group to push tasks to */
9065 /* The local group has been skipped because of CPU affinity */
9070 * If the local group is idler than the selected idlest group
9071 * don't try and push the task.
9073 if (local_sgs
.group_type
< idlest_sgs
.group_type
)
9077 * If the local group is busier than the selected idlest group
9078 * try and push the task.
9080 if (local_sgs
.group_type
> idlest_sgs
.group_type
)
9083 switch (local_sgs
.group_type
) {
9084 case group_overloaded
:
9085 case group_fully_busy
:
9087 /* Calculate allowed imbalance based on load */
9088 imbalance
= scale_load_down(NICE_0_LOAD
) *
9089 (sd
->imbalance_pct
-100) / 100;
9092 * When comparing groups across NUMA domains, it's possible for
9093 * the local domain to be very lightly loaded relative to the
9094 * remote domains but "imbalance" skews the comparison making
9095 * remote CPUs look much more favourable. When considering
9096 * cross-domain, add imbalance to the load on the remote node
9097 * and consider staying local.
9100 if ((sd
->flags
& SD_NUMA
) &&
9101 ((idlest_sgs
.avg_load
+ imbalance
) >= local_sgs
.avg_load
))
9105 * If the local group is less loaded than the selected
9106 * idlest group don't try and push any tasks.
9108 if (idlest_sgs
.avg_load
>= (local_sgs
.avg_load
+ imbalance
))
9111 if (100 * local_sgs
.avg_load
<= sd
->imbalance_pct
* idlest_sgs
.avg_load
)
9115 case group_imbalanced
:
9116 case group_asym_packing
:
9117 /* Those type are not used in the slow wakeup path */
9120 case group_misfit_task
:
9121 /* Select group with the highest max capacity */
9122 if (local
->sgc
->max_capacity
>= idlest
->sgc
->max_capacity
)
9126 case group_has_spare
:
9127 if (sd
->flags
& SD_NUMA
) {
9128 #ifdef CONFIG_NUMA_BALANCING
9131 * If there is spare capacity at NUMA, try to select
9132 * the preferred node
9134 if (cpu_to_node(this_cpu
) == p
->numa_preferred_nid
)
9137 idlest_cpu
= cpumask_first(sched_group_span(idlest
));
9138 if (cpu_to_node(idlest_cpu
) == p
->numa_preferred_nid
)
9142 * Otherwise, keep the task on this node to stay close
9143 * its wakeup source and improve locality. If there is
9144 * a real need of migration, periodic load balance will
9147 if (allow_numa_imbalance(local_sgs
.sum_nr_running
, sd
->span_weight
))
9152 * Select group with highest number of idle CPUs. We could also
9153 * compare the utilization which is more stable but it can end
9154 * up that the group has less spare capacity but finally more
9155 * idle CPUs which means more opportunity to run task.
9157 if (local_sgs
.idle_cpus
>= idlest_sgs
.idle_cpus
)
9166 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
9167 * @env: The load balancing environment.
9168 * @sds: variable to hold the statistics for this sched_domain.
9171 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
9173 struct sched_domain
*child
= env
->sd
->child
;
9174 struct sched_group
*sg
= env
->sd
->groups
;
9175 struct sg_lb_stats
*local
= &sds
->local_stat
;
9176 struct sg_lb_stats tmp_sgs
;
9180 struct sg_lb_stats
*sgs
= &tmp_sgs
;
9183 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_span(sg
));
9188 if (env
->idle
!= CPU_NEWLY_IDLE
||
9189 time_after_eq(jiffies
, sg
->sgc
->next_update
))
9190 update_group_capacity(env
->sd
, env
->dst_cpu
);
9193 update_sg_lb_stats(env
, sds
, sg
, sgs
, &sg_status
);
9199 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
9201 sds
->busiest_stat
= *sgs
;
9205 /* Now, start updating sd_lb_stats */
9206 sds
->total_load
+= sgs
->group_load
;
9207 sds
->total_capacity
+= sgs
->group_capacity
;
9210 } while (sg
!= env
->sd
->groups
);
9212 /* Tag domain that child domain prefers tasks go to siblings first */
9213 sds
->prefer_sibling
= child
&& child
->flags
& SD_PREFER_SIBLING
;
9216 if (env
->sd
->flags
& SD_NUMA
)
9217 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
9219 if (!env
->sd
->parent
) {
9220 struct root_domain
*rd
= env
->dst_rq
->rd
;
9222 /* update overload indicator if we are at root domain */
9223 WRITE_ONCE(rd
->overload
, sg_status
& SG_OVERLOAD
);
9225 /* Update over-utilization (tipping point, U >= 0) indicator */
9226 WRITE_ONCE(rd
->overutilized
, sg_status
& SG_OVERUTILIZED
);
9227 trace_sched_overutilized_tp(rd
, sg_status
& SG_OVERUTILIZED
);
9228 } else if (sg_status
& SG_OVERUTILIZED
) {
9229 struct root_domain
*rd
= env
->dst_rq
->rd
;
9231 WRITE_ONCE(rd
->overutilized
, SG_OVERUTILIZED
);
9232 trace_sched_overutilized_tp(rd
, SG_OVERUTILIZED
);
9236 #define NUMA_IMBALANCE_MIN 2
9238 static inline long adjust_numa_imbalance(int imbalance
,
9239 int dst_running
, int dst_weight
)
9241 if (!allow_numa_imbalance(dst_running
, dst_weight
))
9245 * Allow a small imbalance based on a simple pair of communicating
9246 * tasks that remain local when the destination is lightly loaded.
9248 if (imbalance
<= NUMA_IMBALANCE_MIN
)
9255 * calculate_imbalance - Calculate the amount of imbalance present within the
9256 * groups of a given sched_domain during load balance.
9257 * @env: load balance environment
9258 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
9260 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
9262 struct sg_lb_stats
*local
, *busiest
;
9264 local
= &sds
->local_stat
;
9265 busiest
= &sds
->busiest_stat
;
9267 if (busiest
->group_type
== group_misfit_task
) {
9268 /* Set imbalance to allow misfit tasks to be balanced. */
9269 env
->migration_type
= migrate_misfit
;
9274 if (busiest
->group_type
== group_asym_packing
) {
9276 * In case of asym capacity, we will try to migrate all load to
9277 * the preferred CPU.
9279 env
->migration_type
= migrate_task
;
9280 env
->imbalance
= busiest
->sum_h_nr_running
;
9284 if (busiest
->group_type
== group_imbalanced
) {
9286 * In the group_imb case we cannot rely on group-wide averages
9287 * to ensure CPU-load equilibrium, try to move any task to fix
9288 * the imbalance. The next load balance will take care of
9289 * balancing back the system.
9291 env
->migration_type
= migrate_task
;
9297 * Try to use spare capacity of local group without overloading it or
9300 if (local
->group_type
== group_has_spare
) {
9301 if ((busiest
->group_type
> group_fully_busy
) &&
9302 !(env
->sd
->flags
& SD_SHARE_PKG_RESOURCES
)) {
9304 * If busiest is overloaded, try to fill spare
9305 * capacity. This might end up creating spare capacity
9306 * in busiest or busiest still being overloaded but
9307 * there is no simple way to directly compute the
9308 * amount of load to migrate in order to balance the
9311 env
->migration_type
= migrate_util
;
9312 env
->imbalance
= max(local
->group_capacity
, local
->group_util
) -
9316 * In some cases, the group's utilization is max or even
9317 * higher than capacity because of migrations but the
9318 * local CPU is (newly) idle. There is at least one
9319 * waiting task in this overloaded busiest group. Let's
9322 if (env
->idle
!= CPU_NOT_IDLE
&& env
->imbalance
== 0) {
9323 env
->migration_type
= migrate_task
;
9330 if (busiest
->group_weight
== 1 || sds
->prefer_sibling
) {
9331 unsigned int nr_diff
= busiest
->sum_nr_running
;
9333 * When prefer sibling, evenly spread running tasks on
9336 env
->migration_type
= migrate_task
;
9337 lsub_positive(&nr_diff
, local
->sum_nr_running
);
9338 env
->imbalance
= nr_diff
>> 1;
9342 * If there is no overload, we just want to even the number of
9345 env
->migration_type
= migrate_task
;
9346 env
->imbalance
= max_t(long, 0, (local
->idle_cpus
-
9347 busiest
->idle_cpus
) >> 1);
9350 /* Consider allowing a small imbalance between NUMA groups */
9351 if (env
->sd
->flags
& SD_NUMA
) {
9352 env
->imbalance
= adjust_numa_imbalance(env
->imbalance
,
9353 busiest
->sum_nr_running
, busiest
->group_weight
);
9360 * Local is fully busy but has to take more load to relieve the
9363 if (local
->group_type
< group_overloaded
) {
9365 * Local will become overloaded so the avg_load metrics are
9369 local
->avg_load
= (local
->group_load
* SCHED_CAPACITY_SCALE
) /
9370 local
->group_capacity
;
9372 sds
->avg_load
= (sds
->total_load
* SCHED_CAPACITY_SCALE
) /
9373 sds
->total_capacity
;
9375 * If the local group is more loaded than the selected
9376 * busiest group don't try to pull any tasks.
9378 if (local
->avg_load
>= busiest
->avg_load
) {
9385 * Both group are or will become overloaded and we're trying to get all
9386 * the CPUs to the average_load, so we don't want to push ourselves
9387 * above the average load, nor do we wish to reduce the max loaded CPU
9388 * below the average load. At the same time, we also don't want to
9389 * reduce the group load below the group capacity. Thus we look for
9390 * the minimum possible imbalance.
9392 env
->migration_type
= migrate_load
;
9393 env
->imbalance
= min(
9394 (busiest
->avg_load
- sds
->avg_load
) * busiest
->group_capacity
,
9395 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
9396 ) / SCHED_CAPACITY_SCALE
;
9399 /******* find_busiest_group() helpers end here *********************/
9402 * Decision matrix according to the local and busiest group type:
9404 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9405 * has_spare nr_idle balanced N/A N/A balanced balanced
9406 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9407 * misfit_task force N/A N/A N/A force force
9408 * asym_packing force force N/A N/A force force
9409 * imbalanced force force N/A N/A force force
9410 * overloaded force force N/A N/A force avg_load
9412 * N/A : Not Applicable because already filtered while updating
9414 * balanced : The system is balanced for these 2 groups.
9415 * force : Calculate the imbalance as load migration is probably needed.
9416 * avg_load : Only if imbalance is significant enough.
9417 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9418 * different in groups.
9422 * find_busiest_group - Returns the busiest group within the sched_domain
9423 * if there is an imbalance.
9425 * Also calculates the amount of runnable load which should be moved
9426 * to restore balance.
9428 * @env: The load balancing environment.
9430 * Return: - The busiest group if imbalance exists.
9432 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
9434 struct sg_lb_stats
*local
, *busiest
;
9435 struct sd_lb_stats sds
;
9437 init_sd_lb_stats(&sds
);
9440 * Compute the various statistics relevant for load balancing at
9443 update_sd_lb_stats(env
, &sds
);
9445 if (sched_energy_enabled()) {
9446 struct root_domain
*rd
= env
->dst_rq
->rd
;
9448 if (rcu_dereference(rd
->pd
) && !READ_ONCE(rd
->overutilized
))
9452 local
= &sds
.local_stat
;
9453 busiest
= &sds
.busiest_stat
;
9455 /* There is no busy sibling group to pull tasks from */
9459 /* Misfit tasks should be dealt with regardless of the avg load */
9460 if (busiest
->group_type
== group_misfit_task
)
9463 /* ASYM feature bypasses nice load balance check */
9464 if (busiest
->group_type
== group_asym_packing
)
9468 * If the busiest group is imbalanced the below checks don't
9469 * work because they assume all things are equal, which typically
9470 * isn't true due to cpus_ptr constraints and the like.
9472 if (busiest
->group_type
== group_imbalanced
)
9476 * If the local group is busier than the selected busiest group
9477 * don't try and pull any tasks.
9479 if (local
->group_type
> busiest
->group_type
)
9483 * When groups are overloaded, use the avg_load to ensure fairness
9486 if (local
->group_type
== group_overloaded
) {
9488 * If the local group is more loaded than the selected
9489 * busiest group don't try to pull any tasks.
9491 if (local
->avg_load
>= busiest
->avg_load
)
9494 /* XXX broken for overlapping NUMA groups */
9495 sds
.avg_load
= (sds
.total_load
* SCHED_CAPACITY_SCALE
) /
9499 * Don't pull any tasks if this group is already above the
9500 * domain average load.
9502 if (local
->avg_load
>= sds
.avg_load
)
9506 * If the busiest group is more loaded, use imbalance_pct to be
9509 if (100 * busiest
->avg_load
<=
9510 env
->sd
->imbalance_pct
* local
->avg_load
)
9514 /* Try to move all excess tasks to child's sibling domain */
9515 if (sds
.prefer_sibling
&& local
->group_type
== group_has_spare
&&
9516 busiest
->sum_nr_running
> local
->sum_nr_running
+ 1)
9519 if (busiest
->group_type
!= group_overloaded
) {
9520 if (env
->idle
== CPU_NOT_IDLE
)
9522 * If the busiest group is not overloaded (and as a
9523 * result the local one too) but this CPU is already
9524 * busy, let another idle CPU try to pull task.
9528 if (busiest
->group_weight
> 1 &&
9529 local
->idle_cpus
<= (busiest
->idle_cpus
+ 1))
9531 * If the busiest group is not overloaded
9532 * and there is no imbalance between this and busiest
9533 * group wrt idle CPUs, it is balanced. The imbalance
9534 * becomes significant if the diff is greater than 1
9535 * otherwise we might end up to just move the imbalance
9536 * on another group. Of course this applies only if
9537 * there is more than 1 CPU per group.
9541 if (busiest
->sum_h_nr_running
== 1)
9543 * busiest doesn't have any tasks waiting to run
9549 /* Looks like there is an imbalance. Compute it */
9550 calculate_imbalance(env
, &sds
);
9551 return env
->imbalance
? sds
.busiest
: NULL
;
9559 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9561 static struct rq
*find_busiest_queue(struct lb_env
*env
,
9562 struct sched_group
*group
)
9564 struct rq
*busiest
= NULL
, *rq
;
9565 unsigned long busiest_util
= 0, busiest_load
= 0, busiest_capacity
= 1;
9566 unsigned int busiest_nr
= 0;
9569 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
9570 unsigned long capacity
, load
, util
;
9571 unsigned int nr_running
;
9575 rt
= fbq_classify_rq(rq
);
9578 * We classify groups/runqueues into three groups:
9579 * - regular: there are !numa tasks
9580 * - remote: there are numa tasks that run on the 'wrong' node
9581 * - all: there is no distinction
9583 * In order to avoid migrating ideally placed numa tasks,
9584 * ignore those when there's better options.
9586 * If we ignore the actual busiest queue to migrate another
9587 * task, the next balance pass can still reduce the busiest
9588 * queue by moving tasks around inside the node.
9590 * If we cannot move enough load due to this classification
9591 * the next pass will adjust the group classification and
9592 * allow migration of more tasks.
9594 * Both cases only affect the total convergence complexity.
9596 if (rt
> env
->fbq_type
)
9599 nr_running
= rq
->cfs
.h_nr_running
;
9603 capacity
= capacity_of(i
);
9606 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9607 * eventually lead to active_balancing high->low capacity.
9608 * Higher per-CPU capacity is considered better than balancing
9611 if (env
->sd
->flags
& SD_ASYM_CPUCAPACITY
&&
9612 !capacity_greater(capacity_of(env
->dst_cpu
), capacity
) &&
9616 /* Make sure we only pull tasks from a CPU of lower priority */
9617 if ((env
->sd
->flags
& SD_ASYM_PACKING
) &&
9618 sched_asym_prefer(i
, env
->dst_cpu
) &&
9622 switch (env
->migration_type
) {
9625 * When comparing with load imbalance, use cpu_load()
9626 * which is not scaled with the CPU capacity.
9628 load
= cpu_load(rq
);
9630 if (nr_running
== 1 && load
> env
->imbalance
&&
9631 !check_cpu_capacity(rq
, env
->sd
))
9635 * For the load comparisons with the other CPUs,
9636 * consider the cpu_load() scaled with the CPU
9637 * capacity, so that the load can be moved away
9638 * from the CPU that is potentially running at a
9641 * Thus we're looking for max(load_i / capacity_i),
9642 * crosswise multiplication to rid ourselves of the
9643 * division works out to:
9644 * load_i * capacity_j > load_j * capacity_i;
9645 * where j is our previous maximum.
9647 if (load
* busiest_capacity
> busiest_load
* capacity
) {
9648 busiest_load
= load
;
9649 busiest_capacity
= capacity
;
9655 util
= cpu_util_cfs(i
);
9658 * Don't try to pull utilization from a CPU with one
9659 * running task. Whatever its utilization, we will fail
9662 if (nr_running
<= 1)
9665 if (busiest_util
< util
) {
9666 busiest_util
= util
;
9672 if (busiest_nr
< nr_running
) {
9673 busiest_nr
= nr_running
;
9678 case migrate_misfit
:
9680 * For ASYM_CPUCAPACITY domains with misfit tasks we
9681 * simply seek the "biggest" misfit task.
9683 if (rq
->misfit_task_load
> busiest_load
) {
9684 busiest_load
= rq
->misfit_task_load
;
9697 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9698 * so long as it is large enough.
9700 #define MAX_PINNED_INTERVAL 512
9703 asym_active_balance(struct lb_env
*env
)
9706 * ASYM_PACKING needs to force migrate tasks from busy but
9707 * lower priority CPUs in order to pack all tasks in the
9708 * highest priority CPUs.
9710 return env
->idle
!= CPU_NOT_IDLE
&& (env
->sd
->flags
& SD_ASYM_PACKING
) &&
9711 sched_asym_prefer(env
->dst_cpu
, env
->src_cpu
);
9715 imbalanced_active_balance(struct lb_env
*env
)
9717 struct sched_domain
*sd
= env
->sd
;
9720 * The imbalanced case includes the case of pinned tasks preventing a fair
9721 * distribution of the load on the system but also the even distribution of the
9722 * threads on a system with spare capacity
9724 if ((env
->migration_type
== migrate_task
) &&
9725 (sd
->nr_balance_failed
> sd
->cache_nice_tries
+2))
9731 static int need_active_balance(struct lb_env
*env
)
9733 struct sched_domain
*sd
= env
->sd
;
9735 if (asym_active_balance(env
))
9738 if (imbalanced_active_balance(env
))
9742 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
9743 * It's worth migrating the task if the src_cpu's capacity is reduced
9744 * because of other sched_class or IRQs if more capacity stays
9745 * available on dst_cpu.
9747 if ((env
->idle
!= CPU_NOT_IDLE
) &&
9748 (env
->src_rq
->cfs
.h_nr_running
== 1)) {
9749 if ((check_cpu_capacity(env
->src_rq
, sd
)) &&
9750 (capacity_of(env
->src_cpu
)*sd
->imbalance_pct
< capacity_of(env
->dst_cpu
)*100))
9754 if (env
->migration_type
== migrate_misfit
)
9760 static int active_load_balance_cpu_stop(void *data
);
9762 static int should_we_balance(struct lb_env
*env
)
9764 struct sched_group
*sg
= env
->sd
->groups
;
9768 * Ensure the balancing environment is consistent; can happen
9769 * when the softirq triggers 'during' hotplug.
9771 if (!cpumask_test_cpu(env
->dst_cpu
, env
->cpus
))
9775 * In the newly idle case, we will allow all the CPUs
9776 * to do the newly idle load balance.
9778 if (env
->idle
== CPU_NEWLY_IDLE
)
9781 /* Try to find first idle CPU */
9782 for_each_cpu_and(cpu
, group_balance_mask(sg
), env
->cpus
) {
9786 /* Are we the first idle CPU? */
9787 return cpu
== env
->dst_cpu
;
9790 /* Are we the first CPU of this group ? */
9791 return group_balance_cpu(sg
) == env
->dst_cpu
;
9795 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9796 * tasks if there is an imbalance.
9798 static int load_balance(int this_cpu
, struct rq
*this_rq
,
9799 struct sched_domain
*sd
, enum cpu_idle_type idle
,
9800 int *continue_balancing
)
9802 int ld_moved
, cur_ld_moved
, active_balance
= 0;
9803 struct sched_domain
*sd_parent
= sd
->parent
;
9804 struct sched_group
*group
;
9807 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
9809 struct lb_env env
= {
9811 .dst_cpu
= this_cpu
,
9813 .dst_grpmask
= sched_group_span(sd
->groups
),
9815 .loop_break
= sched_nr_migrate_break
,
9818 .tasks
= LIST_HEAD_INIT(env
.tasks
),
9821 cpumask_and(cpus
, sched_domain_span(sd
), cpu_active_mask
);
9823 schedstat_inc(sd
->lb_count
[idle
]);
9826 if (!should_we_balance(&env
)) {
9827 *continue_balancing
= 0;
9831 group
= find_busiest_group(&env
);
9833 schedstat_inc(sd
->lb_nobusyg
[idle
]);
9837 busiest
= find_busiest_queue(&env
, group
);
9839 schedstat_inc(sd
->lb_nobusyq
[idle
]);
9843 BUG_ON(busiest
== env
.dst_rq
);
9845 schedstat_add(sd
->lb_imbalance
[idle
], env
.imbalance
);
9847 env
.src_cpu
= busiest
->cpu
;
9848 env
.src_rq
= busiest
;
9851 /* Clear this flag as soon as we find a pullable task */
9852 env
.flags
|= LBF_ALL_PINNED
;
9853 if (busiest
->nr_running
> 1) {
9855 * Attempt to move tasks. If find_busiest_group has found
9856 * an imbalance but busiest->nr_running <= 1, the group is
9857 * still unbalanced. ld_moved simply stays zero, so it is
9858 * correctly treated as an imbalance.
9860 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
9863 rq_lock_irqsave(busiest
, &rf
);
9864 update_rq_clock(busiest
);
9867 * cur_ld_moved - load moved in current iteration
9868 * ld_moved - cumulative load moved across iterations
9870 cur_ld_moved
= detach_tasks(&env
);
9873 * We've detached some tasks from busiest_rq. Every
9874 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9875 * unlock busiest->lock, and we are able to be sure
9876 * that nobody can manipulate the tasks in parallel.
9877 * See task_rq_lock() family for the details.
9880 rq_unlock(busiest
, &rf
);
9884 ld_moved
+= cur_ld_moved
;
9887 local_irq_restore(rf
.flags
);
9889 if (env
.flags
& LBF_NEED_BREAK
) {
9890 env
.flags
&= ~LBF_NEED_BREAK
;
9895 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9896 * us and move them to an alternate dst_cpu in our sched_group
9897 * where they can run. The upper limit on how many times we
9898 * iterate on same src_cpu is dependent on number of CPUs in our
9901 * This changes load balance semantics a bit on who can move
9902 * load to a given_cpu. In addition to the given_cpu itself
9903 * (or a ilb_cpu acting on its behalf where given_cpu is
9904 * nohz-idle), we now have balance_cpu in a position to move
9905 * load to given_cpu. In rare situations, this may cause
9906 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9907 * _independently_ and at _same_ time to move some load to
9908 * given_cpu) causing excess load to be moved to given_cpu.
9909 * This however should not happen so much in practice and
9910 * moreover subsequent load balance cycles should correct the
9911 * excess load moved.
9913 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
9915 /* Prevent to re-select dst_cpu via env's CPUs */
9916 __cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
9918 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
9919 env
.dst_cpu
= env
.new_dst_cpu
;
9920 env
.flags
&= ~LBF_DST_PINNED
;
9922 env
.loop_break
= sched_nr_migrate_break
;
9925 * Go back to "more_balance" rather than "redo" since we
9926 * need to continue with same src_cpu.
9932 * We failed to reach balance because of affinity.
9935 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
9937 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
9938 *group_imbalance
= 1;
9941 /* All tasks on this runqueue were pinned by CPU affinity */
9942 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
9943 __cpumask_clear_cpu(cpu_of(busiest
), cpus
);
9945 * Attempting to continue load balancing at the current
9946 * sched_domain level only makes sense if there are
9947 * active CPUs remaining as possible busiest CPUs to
9948 * pull load from which are not contained within the
9949 * destination group that is receiving any migrated
9952 if (!cpumask_subset(cpus
, env
.dst_grpmask
)) {
9954 env
.loop_break
= sched_nr_migrate_break
;
9957 goto out_all_pinned
;
9962 schedstat_inc(sd
->lb_failed
[idle
]);
9964 * Increment the failure counter only on periodic balance.
9965 * We do not want newidle balance, which can be very
9966 * frequent, pollute the failure counter causing
9967 * excessive cache_hot migrations and active balances.
9969 if (idle
!= CPU_NEWLY_IDLE
)
9970 sd
->nr_balance_failed
++;
9972 if (need_active_balance(&env
)) {
9973 unsigned long flags
;
9975 raw_spin_rq_lock_irqsave(busiest
, flags
);
9978 * Don't kick the active_load_balance_cpu_stop,
9979 * if the curr task on busiest CPU can't be
9980 * moved to this_cpu:
9982 if (!cpumask_test_cpu(this_cpu
, busiest
->curr
->cpus_ptr
)) {
9983 raw_spin_rq_unlock_irqrestore(busiest
, flags
);
9984 goto out_one_pinned
;
9987 /* Record that we found at least one task that could run on this_cpu */
9988 env
.flags
&= ~LBF_ALL_PINNED
;
9991 * ->active_balance synchronizes accesses to
9992 * ->active_balance_work. Once set, it's cleared
9993 * only after active load balance is finished.
9995 if (!busiest
->active_balance
) {
9996 busiest
->active_balance
= 1;
9997 busiest
->push_cpu
= this_cpu
;
10000 raw_spin_rq_unlock_irqrestore(busiest
, flags
);
10002 if (active_balance
) {
10003 stop_one_cpu_nowait(cpu_of(busiest
),
10004 active_load_balance_cpu_stop
, busiest
,
10005 &busiest
->active_balance_work
);
10009 sd
->nr_balance_failed
= 0;
10012 if (likely(!active_balance
) || need_active_balance(&env
)) {
10013 /* We were unbalanced, so reset the balancing interval */
10014 sd
->balance_interval
= sd
->min_interval
;
10021 * We reach balance although we may have faced some affinity
10022 * constraints. Clear the imbalance flag only if other tasks got
10023 * a chance to move and fix the imbalance.
10025 if (sd_parent
&& !(env
.flags
& LBF_ALL_PINNED
)) {
10026 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
10028 if (*group_imbalance
)
10029 *group_imbalance
= 0;
10034 * We reach balance because all tasks are pinned at this level so
10035 * we can't migrate them. Let the imbalance flag set so parent level
10036 * can try to migrate them.
10038 schedstat_inc(sd
->lb_balanced
[idle
]);
10040 sd
->nr_balance_failed
= 0;
10046 * newidle_balance() disregards balance intervals, so we could
10047 * repeatedly reach this code, which would lead to balance_interval
10048 * skyrocketing in a short amount of time. Skip the balance_interval
10049 * increase logic to avoid that.
10051 if (env
.idle
== CPU_NEWLY_IDLE
)
10054 /* tune up the balancing interval */
10055 if ((env
.flags
& LBF_ALL_PINNED
&&
10056 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
10057 sd
->balance_interval
< sd
->max_interval
)
10058 sd
->balance_interval
*= 2;
10063 static inline unsigned long
10064 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
10066 unsigned long interval
= sd
->balance_interval
;
10069 interval
*= sd
->busy_factor
;
10071 /* scale ms to jiffies */
10072 interval
= msecs_to_jiffies(interval
);
10075 * Reduce likelihood of busy balancing at higher domains racing with
10076 * balancing at lower domains by preventing their balancing periods
10077 * from being multiples of each other.
10082 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
10088 update_next_balance(struct sched_domain
*sd
, unsigned long *next_balance
)
10090 unsigned long interval
, next
;
10092 /* used by idle balance, so cpu_busy = 0 */
10093 interval
= get_sd_balance_interval(sd
, 0);
10094 next
= sd
->last_balance
+ interval
;
10096 if (time_after(*next_balance
, next
))
10097 *next_balance
= next
;
10101 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
10102 * running tasks off the busiest CPU onto idle CPUs. It requires at
10103 * least 1 task to be running on each physical CPU where possible, and
10104 * avoids physical / logical imbalances.
10106 static int active_load_balance_cpu_stop(void *data
)
10108 struct rq
*busiest_rq
= data
;
10109 int busiest_cpu
= cpu_of(busiest_rq
);
10110 int target_cpu
= busiest_rq
->push_cpu
;
10111 struct rq
*target_rq
= cpu_rq(target_cpu
);
10112 struct sched_domain
*sd
;
10113 struct task_struct
*p
= NULL
;
10114 struct rq_flags rf
;
10116 rq_lock_irq(busiest_rq
, &rf
);
10118 * Between queueing the stop-work and running it is a hole in which
10119 * CPUs can become inactive. We should not move tasks from or to
10122 if (!cpu_active(busiest_cpu
) || !cpu_active(target_cpu
))
10125 /* Make sure the requested CPU hasn't gone down in the meantime: */
10126 if (unlikely(busiest_cpu
!= smp_processor_id() ||
10127 !busiest_rq
->active_balance
))
10130 /* Is there any task to move? */
10131 if (busiest_rq
->nr_running
<= 1)
10135 * This condition is "impossible", if it occurs
10136 * we need to fix it. Originally reported by
10137 * Bjorn Helgaas on a 128-CPU setup.
10139 BUG_ON(busiest_rq
== target_rq
);
10141 /* Search for an sd spanning us and the target CPU. */
10143 for_each_domain(target_cpu
, sd
) {
10144 if (cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
10149 struct lb_env env
= {
10151 .dst_cpu
= target_cpu
,
10152 .dst_rq
= target_rq
,
10153 .src_cpu
= busiest_rq
->cpu
,
10154 .src_rq
= busiest_rq
,
10156 .flags
= LBF_ACTIVE_LB
,
10159 schedstat_inc(sd
->alb_count
);
10160 update_rq_clock(busiest_rq
);
10162 p
= detach_one_task(&env
);
10164 schedstat_inc(sd
->alb_pushed
);
10165 /* Active balancing done, reset the failure counter. */
10166 sd
->nr_balance_failed
= 0;
10168 schedstat_inc(sd
->alb_failed
);
10173 busiest_rq
->active_balance
= 0;
10174 rq_unlock(busiest_rq
, &rf
);
10177 attach_one_task(target_rq
, p
);
10179 local_irq_enable();
10184 static DEFINE_SPINLOCK(balancing
);
10187 * Scale the max load_balance interval with the number of CPUs in the system.
10188 * This trades load-balance latency on larger machines for less cross talk.
10190 void update_max_interval(void)
10192 max_load_balance_interval
= HZ
*num_online_cpus()/10;
10195 static inline bool update_newidle_cost(struct sched_domain
*sd
, u64 cost
)
10197 if (cost
> sd
->max_newidle_lb_cost
) {
10199 * Track max cost of a domain to make sure to not delay the
10200 * next wakeup on the CPU.
10202 sd
->max_newidle_lb_cost
= cost
;
10203 sd
->last_decay_max_lb_cost
= jiffies
;
10204 } else if (time_after(jiffies
, sd
->last_decay_max_lb_cost
+ HZ
)) {
10206 * Decay the newidle max times by ~1% per second to ensure that
10207 * it is not outdated and the current max cost is actually
10210 sd
->max_newidle_lb_cost
= (sd
->max_newidle_lb_cost
* 253) / 256;
10211 sd
->last_decay_max_lb_cost
= jiffies
;
10220 * It checks each scheduling domain to see if it is due to be balanced,
10221 * and initiates a balancing operation if so.
10223 * Balancing parameters are set up in init_sched_domains.
10225 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
10227 int continue_balancing
= 1;
10229 int busy
= idle
!= CPU_IDLE
&& !sched_idle_cpu(cpu
);
10230 unsigned long interval
;
10231 struct sched_domain
*sd
;
10232 /* Earliest time when we have to do rebalance again */
10233 unsigned long next_balance
= jiffies
+ 60*HZ
;
10234 int update_next_balance
= 0;
10235 int need_serialize
, need_decay
= 0;
10239 for_each_domain(cpu
, sd
) {
10241 * Decay the newidle max times here because this is a regular
10242 * visit to all the domains.
10244 need_decay
= update_newidle_cost(sd
, 0);
10245 max_cost
+= sd
->max_newidle_lb_cost
;
10248 * Stop the load balance at this level. There is another
10249 * CPU in our sched group which is doing load balancing more
10252 if (!continue_balancing
) {
10258 interval
= get_sd_balance_interval(sd
, busy
);
10260 need_serialize
= sd
->flags
& SD_SERIALIZE
;
10261 if (need_serialize
) {
10262 if (!spin_trylock(&balancing
))
10266 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
10267 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
10269 * The LBF_DST_PINNED logic could have changed
10270 * env->dst_cpu, so we can't know our idle
10271 * state even if we migrated tasks. Update it.
10273 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
10274 busy
= idle
!= CPU_IDLE
&& !sched_idle_cpu(cpu
);
10276 sd
->last_balance
= jiffies
;
10277 interval
= get_sd_balance_interval(sd
, busy
);
10279 if (need_serialize
)
10280 spin_unlock(&balancing
);
10282 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
10283 next_balance
= sd
->last_balance
+ interval
;
10284 update_next_balance
= 1;
10289 * Ensure the rq-wide value also decays but keep it at a
10290 * reasonable floor to avoid funnies with rq->avg_idle.
10292 rq
->max_idle_balance_cost
=
10293 max((u64
)sysctl_sched_migration_cost
, max_cost
);
10298 * next_balance will be updated only when there is a need.
10299 * When the cpu is attached to null domain for ex, it will not be
10302 if (likely(update_next_balance
))
10303 rq
->next_balance
= next_balance
;
10307 static inline int on_null_domain(struct rq
*rq
)
10309 return unlikely(!rcu_dereference_sched(rq
->sd
));
10312 #ifdef CONFIG_NO_HZ_COMMON
10314 * idle load balancing details
10315 * - When one of the busy CPUs notice that there may be an idle rebalancing
10316 * needed, they will kick the idle load balancer, which then does idle
10317 * load balancing for all the idle CPUs.
10318 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
10322 static inline int find_new_ilb(void)
10325 const struct cpumask
*hk_mask
;
10327 hk_mask
= housekeeping_cpumask(HK_FLAG_MISC
);
10329 for_each_cpu_and(ilb
, nohz
.idle_cpus_mask
, hk_mask
) {
10331 if (ilb
== smp_processor_id())
10342 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
10343 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
10345 static void kick_ilb(unsigned int flags
)
10350 * Increase nohz.next_balance only when if full ilb is triggered but
10351 * not if we only update stats.
10353 if (flags
& NOHZ_BALANCE_KICK
)
10354 nohz
.next_balance
= jiffies
+1;
10356 ilb_cpu
= find_new_ilb();
10358 if (ilb_cpu
>= nr_cpu_ids
)
10362 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
10363 * the first flag owns it; cleared by nohz_csd_func().
10365 flags
= atomic_fetch_or(flags
, nohz_flags(ilb_cpu
));
10366 if (flags
& NOHZ_KICK_MASK
)
10370 * This way we generate an IPI on the target CPU which
10371 * is idle. And the softirq performing nohz idle load balance
10372 * will be run before returning from the IPI.
10374 smp_call_function_single_async(ilb_cpu
, &cpu_rq(ilb_cpu
)->nohz_csd
);
10378 * Current decision point for kicking the idle load balancer in the presence
10379 * of idle CPUs in the system.
10381 static void nohz_balancer_kick(struct rq
*rq
)
10383 unsigned long now
= jiffies
;
10384 struct sched_domain_shared
*sds
;
10385 struct sched_domain
*sd
;
10386 int nr_busy
, i
, cpu
= rq
->cpu
;
10387 unsigned int flags
= 0;
10389 if (unlikely(rq
->idle_balance
))
10393 * We may be recently in ticked or tickless idle mode. At the first
10394 * busy tick after returning from idle, we will update the busy stats.
10396 nohz_balance_exit_idle(rq
);
10399 * None are in tickless mode and hence no need for NOHZ idle load
10402 if (likely(!atomic_read(&nohz
.nr_cpus
)))
10405 if (READ_ONCE(nohz
.has_blocked
) &&
10406 time_after(now
, READ_ONCE(nohz
.next_blocked
)))
10407 flags
= NOHZ_STATS_KICK
;
10409 if (time_before(now
, nohz
.next_balance
))
10412 if (rq
->nr_running
>= 2) {
10413 flags
= NOHZ_STATS_KICK
| NOHZ_BALANCE_KICK
;
10419 sd
= rcu_dereference(rq
->sd
);
10422 * If there's a CFS task and the current CPU has reduced
10423 * capacity; kick the ILB to see if there's a better CPU to run
10426 if (rq
->cfs
.h_nr_running
>= 1 && check_cpu_capacity(rq
, sd
)) {
10427 flags
= NOHZ_STATS_KICK
| NOHZ_BALANCE_KICK
;
10432 sd
= rcu_dereference(per_cpu(sd_asym_packing
, cpu
));
10435 * When ASYM_PACKING; see if there's a more preferred CPU
10436 * currently idle; in which case, kick the ILB to move tasks
10439 for_each_cpu_and(i
, sched_domain_span(sd
), nohz
.idle_cpus_mask
) {
10440 if (sched_asym_prefer(i
, cpu
)) {
10441 flags
= NOHZ_STATS_KICK
| NOHZ_BALANCE_KICK
;
10447 sd
= rcu_dereference(per_cpu(sd_asym_cpucapacity
, cpu
));
10450 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10451 * to run the misfit task on.
10453 if (check_misfit_status(rq
, sd
)) {
10454 flags
= NOHZ_STATS_KICK
| NOHZ_BALANCE_KICK
;
10459 * For asymmetric systems, we do not want to nicely balance
10460 * cache use, instead we want to embrace asymmetry and only
10461 * ensure tasks have enough CPU capacity.
10463 * Skip the LLC logic because it's not relevant in that case.
10468 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
10471 * If there is an imbalance between LLC domains (IOW we could
10472 * increase the overall cache use), we need some less-loaded LLC
10473 * domain to pull some load. Likewise, we may need to spread
10474 * load within the current LLC domain (e.g. packed SMT cores but
10475 * other CPUs are idle). We can't really know from here how busy
10476 * the others are - so just get a nohz balance going if it looks
10477 * like this LLC domain has tasks we could move.
10479 nr_busy
= atomic_read(&sds
->nr_busy_cpus
);
10481 flags
= NOHZ_STATS_KICK
| NOHZ_BALANCE_KICK
;
10488 if (READ_ONCE(nohz
.needs_update
))
10489 flags
|= NOHZ_NEXT_KICK
;
10495 static void set_cpu_sd_state_busy(int cpu
)
10497 struct sched_domain
*sd
;
10500 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
10502 if (!sd
|| !sd
->nohz_idle
)
10506 atomic_inc(&sd
->shared
->nr_busy_cpus
);
10511 void nohz_balance_exit_idle(struct rq
*rq
)
10513 SCHED_WARN_ON(rq
!= this_rq());
10515 if (likely(!rq
->nohz_tick_stopped
))
10518 rq
->nohz_tick_stopped
= 0;
10519 cpumask_clear_cpu(rq
->cpu
, nohz
.idle_cpus_mask
);
10520 atomic_dec(&nohz
.nr_cpus
);
10522 set_cpu_sd_state_busy(rq
->cpu
);
10525 static void set_cpu_sd_state_idle(int cpu
)
10527 struct sched_domain
*sd
;
10530 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
10532 if (!sd
|| sd
->nohz_idle
)
10536 atomic_dec(&sd
->shared
->nr_busy_cpus
);
10542 * This routine will record that the CPU is going idle with tick stopped.
10543 * This info will be used in performing idle load balancing in the future.
10545 void nohz_balance_enter_idle(int cpu
)
10547 struct rq
*rq
= cpu_rq(cpu
);
10549 SCHED_WARN_ON(cpu
!= smp_processor_id());
10551 /* If this CPU is going down, then nothing needs to be done: */
10552 if (!cpu_active(cpu
))
10555 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10556 if (!housekeeping_cpu(cpu
, HK_FLAG_SCHED
))
10560 * Can be set safely without rq->lock held
10561 * If a clear happens, it will have evaluated last additions because
10562 * rq->lock is held during the check and the clear
10564 rq
->has_blocked_load
= 1;
10567 * The tick is still stopped but load could have been added in the
10568 * meantime. We set the nohz.has_blocked flag to trig a check of the
10569 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10570 * of nohz.has_blocked can only happen after checking the new load
10572 if (rq
->nohz_tick_stopped
)
10575 /* If we're a completely isolated CPU, we don't play: */
10576 if (on_null_domain(rq
))
10579 rq
->nohz_tick_stopped
= 1;
10581 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
10582 atomic_inc(&nohz
.nr_cpus
);
10585 * Ensures that if nohz_idle_balance() fails to observe our
10586 * @idle_cpus_mask store, it must observe the @has_blocked
10587 * and @needs_update stores.
10589 smp_mb__after_atomic();
10591 set_cpu_sd_state_idle(cpu
);
10593 WRITE_ONCE(nohz
.needs_update
, 1);
10596 * Each time a cpu enter idle, we assume that it has blocked load and
10597 * enable the periodic update of the load of idle cpus
10599 WRITE_ONCE(nohz
.has_blocked
, 1);
10602 static bool update_nohz_stats(struct rq
*rq
)
10604 unsigned int cpu
= rq
->cpu
;
10606 if (!rq
->has_blocked_load
)
10609 if (!cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))
10612 if (!time_after(jiffies
, READ_ONCE(rq
->last_blocked_load_update_tick
)))
10615 update_blocked_averages(cpu
);
10617 return rq
->has_blocked_load
;
10621 * Internal function that runs load balance for all idle cpus. The load balance
10622 * can be a simple update of blocked load or a complete load balance with
10623 * tasks movement depending of flags.
10625 static void _nohz_idle_balance(struct rq
*this_rq
, unsigned int flags
,
10626 enum cpu_idle_type idle
)
10628 /* Earliest time when we have to do rebalance again */
10629 unsigned long now
= jiffies
;
10630 unsigned long next_balance
= now
+ 60*HZ
;
10631 bool has_blocked_load
= false;
10632 int update_next_balance
= 0;
10633 int this_cpu
= this_rq
->cpu
;
10637 SCHED_WARN_ON((flags
& NOHZ_KICK_MASK
) == NOHZ_BALANCE_KICK
);
10640 * We assume there will be no idle load after this update and clear
10641 * the has_blocked flag. If a cpu enters idle in the mean time, it will
10642 * set the has_blocked flag and trigger another update of idle load.
10643 * Because a cpu that becomes idle, is added to idle_cpus_mask before
10644 * setting the flag, we are sure to not clear the state and not
10645 * check the load of an idle cpu.
10647 * Same applies to idle_cpus_mask vs needs_update.
10649 if (flags
& NOHZ_STATS_KICK
)
10650 WRITE_ONCE(nohz
.has_blocked
, 0);
10651 if (flags
& NOHZ_NEXT_KICK
)
10652 WRITE_ONCE(nohz
.needs_update
, 0);
10655 * Ensures that if we miss the CPU, we must see the has_blocked
10656 * store from nohz_balance_enter_idle().
10661 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
10662 * chance for other idle cpu to pull load.
10664 for_each_cpu_wrap(balance_cpu
, nohz
.idle_cpus_mask
, this_cpu
+1) {
10665 if (!idle_cpu(balance_cpu
))
10669 * If this CPU gets work to do, stop the load balancing
10670 * work being done for other CPUs. Next load
10671 * balancing owner will pick it up.
10673 if (need_resched()) {
10674 if (flags
& NOHZ_STATS_KICK
)
10675 has_blocked_load
= true;
10676 if (flags
& NOHZ_NEXT_KICK
)
10677 WRITE_ONCE(nohz
.needs_update
, 1);
10681 rq
= cpu_rq(balance_cpu
);
10683 if (flags
& NOHZ_STATS_KICK
)
10684 has_blocked_load
|= update_nohz_stats(rq
);
10687 * If time for next balance is due,
10690 if (time_after_eq(jiffies
, rq
->next_balance
)) {
10691 struct rq_flags rf
;
10693 rq_lock_irqsave(rq
, &rf
);
10694 update_rq_clock(rq
);
10695 rq_unlock_irqrestore(rq
, &rf
);
10697 if (flags
& NOHZ_BALANCE_KICK
)
10698 rebalance_domains(rq
, CPU_IDLE
);
10701 if (time_after(next_balance
, rq
->next_balance
)) {
10702 next_balance
= rq
->next_balance
;
10703 update_next_balance
= 1;
10708 * next_balance will be updated only when there is a need.
10709 * When the CPU is attached to null domain for ex, it will not be
10712 if (likely(update_next_balance
))
10713 nohz
.next_balance
= next_balance
;
10715 if (flags
& NOHZ_STATS_KICK
)
10716 WRITE_ONCE(nohz
.next_blocked
,
10717 now
+ msecs_to_jiffies(LOAD_AVG_PERIOD
));
10720 /* There is still blocked load, enable periodic update */
10721 if (has_blocked_load
)
10722 WRITE_ONCE(nohz
.has_blocked
, 1);
10726 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
10727 * rebalancing for all the cpus for whom scheduler ticks are stopped.
10729 static bool nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
10731 unsigned int flags
= this_rq
->nohz_idle_balance
;
10736 this_rq
->nohz_idle_balance
= 0;
10738 if (idle
!= CPU_IDLE
)
10741 _nohz_idle_balance(this_rq
, flags
, idle
);
10747 * Check if we need to run the ILB for updating blocked load before entering
10750 void nohz_run_idle_balance(int cpu
)
10752 unsigned int flags
;
10754 flags
= atomic_fetch_andnot(NOHZ_NEWILB_KICK
, nohz_flags(cpu
));
10757 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
10758 * (ie NOHZ_STATS_KICK set) and will do the same.
10760 if ((flags
== NOHZ_NEWILB_KICK
) && !need_resched())
10761 _nohz_idle_balance(cpu_rq(cpu
), NOHZ_STATS_KICK
, CPU_IDLE
);
10764 static void nohz_newidle_balance(struct rq
*this_rq
)
10766 int this_cpu
= this_rq
->cpu
;
10769 * This CPU doesn't want to be disturbed by scheduler
10772 if (!housekeeping_cpu(this_cpu
, HK_FLAG_SCHED
))
10775 /* Will wake up very soon. No time for doing anything else*/
10776 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
10779 /* Don't need to update blocked load of idle CPUs*/
10780 if (!READ_ONCE(nohz
.has_blocked
) ||
10781 time_before(jiffies
, READ_ONCE(nohz
.next_blocked
)))
10785 * Set the need to trigger ILB in order to update blocked load
10786 * before entering idle state.
10788 atomic_or(NOHZ_NEWILB_KICK
, nohz_flags(this_cpu
));
10791 #else /* !CONFIG_NO_HZ_COMMON */
10792 static inline void nohz_balancer_kick(struct rq
*rq
) { }
10794 static inline bool nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
10799 static inline void nohz_newidle_balance(struct rq
*this_rq
) { }
10800 #endif /* CONFIG_NO_HZ_COMMON */
10803 * newidle_balance is called by schedule() if this_cpu is about to become
10804 * idle. Attempts to pull tasks from other CPUs.
10807 * < 0 - we released the lock and there are !fair tasks present
10808 * 0 - failed, no new tasks
10809 * > 0 - success, new (fair) tasks present
10811 static int newidle_balance(struct rq
*this_rq
, struct rq_flags
*rf
)
10813 unsigned long next_balance
= jiffies
+ HZ
;
10814 int this_cpu
= this_rq
->cpu
;
10815 u64 t0
, t1
, curr_cost
= 0;
10816 struct sched_domain
*sd
;
10817 int pulled_task
= 0;
10819 update_misfit_status(NULL
, this_rq
);
10822 * There is a task waiting to run. No need to search for one.
10823 * Return 0; the task will be enqueued when switching to idle.
10825 if (this_rq
->ttwu_pending
)
10829 * We must set idle_stamp _before_ calling idle_balance(), such that we
10830 * measure the duration of idle_balance() as idle time.
10832 this_rq
->idle_stamp
= rq_clock(this_rq
);
10835 * Do not pull tasks towards !active CPUs...
10837 if (!cpu_active(this_cpu
))
10841 * This is OK, because current is on_cpu, which avoids it being picked
10842 * for load-balance and preemption/IRQs are still disabled avoiding
10843 * further scheduler activity on it and we're being very careful to
10844 * re-start the picking loop.
10846 rq_unpin_lock(this_rq
, rf
);
10849 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
10851 if (!READ_ONCE(this_rq
->rd
->overload
) ||
10852 (sd
&& this_rq
->avg_idle
< sd
->max_newidle_lb_cost
)) {
10855 update_next_balance(sd
, &next_balance
);
10862 raw_spin_rq_unlock(this_rq
);
10864 t0
= sched_clock_cpu(this_cpu
);
10865 update_blocked_averages(this_cpu
);
10868 for_each_domain(this_cpu
, sd
) {
10869 int continue_balancing
= 1;
10872 update_next_balance(sd
, &next_balance
);
10874 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
)
10877 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
10879 pulled_task
= load_balance(this_cpu
, this_rq
,
10880 sd
, CPU_NEWLY_IDLE
,
10881 &continue_balancing
);
10883 t1
= sched_clock_cpu(this_cpu
);
10884 domain_cost
= t1
- t0
;
10885 update_newidle_cost(sd
, domain_cost
);
10887 curr_cost
+= domain_cost
;
10892 * Stop searching for tasks to pull if there are
10893 * now runnable tasks on this rq.
10895 if (pulled_task
|| this_rq
->nr_running
> 0 ||
10896 this_rq
->ttwu_pending
)
10901 raw_spin_rq_lock(this_rq
);
10903 if (curr_cost
> this_rq
->max_idle_balance_cost
)
10904 this_rq
->max_idle_balance_cost
= curr_cost
;
10907 * While browsing the domains, we released the rq lock, a task could
10908 * have been enqueued in the meantime. Since we're not going idle,
10909 * pretend we pulled a task.
10911 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
10914 /* Is there a task of a high priority class? */
10915 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
10919 /* Move the next balance forward */
10920 if (time_after(this_rq
->next_balance
, next_balance
))
10921 this_rq
->next_balance
= next_balance
;
10924 this_rq
->idle_stamp
= 0;
10926 nohz_newidle_balance(this_rq
);
10928 rq_repin_lock(this_rq
, rf
);
10930 return pulled_task
;
10934 * run_rebalance_domains is triggered when needed from the scheduler tick.
10935 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10937 static __latent_entropy
void run_rebalance_domains(struct softirq_action
*h
)
10939 struct rq
*this_rq
= this_rq();
10940 enum cpu_idle_type idle
= this_rq
->idle_balance
?
10941 CPU_IDLE
: CPU_NOT_IDLE
;
10944 * If this CPU has a pending nohz_balance_kick, then do the
10945 * balancing on behalf of the other idle CPUs whose ticks are
10946 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10947 * give the idle CPUs a chance to load balance. Else we may
10948 * load balance only within the local sched_domain hierarchy
10949 * and abort nohz_idle_balance altogether if we pull some load.
10951 if (nohz_idle_balance(this_rq
, idle
))
10954 /* normal load balance */
10955 update_blocked_averages(this_rq
->cpu
);
10956 rebalance_domains(this_rq
, idle
);
10960 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10962 void trigger_load_balance(struct rq
*rq
)
10965 * Don't need to rebalance while attached to NULL domain or
10966 * runqueue CPU is not active
10968 if (unlikely(on_null_domain(rq
) || !cpu_active(cpu_of(rq
))))
10971 if (time_after_eq(jiffies
, rq
->next_balance
))
10972 raise_softirq(SCHED_SOFTIRQ
);
10974 nohz_balancer_kick(rq
);
10977 static void rq_online_fair(struct rq
*rq
)
10981 update_runtime_enabled(rq
);
10984 static void rq_offline_fair(struct rq
*rq
)
10988 /* Ensure any throttled groups are reachable by pick_next_task */
10989 unthrottle_offline_cfs_rqs(rq
);
10992 #endif /* CONFIG_SMP */
10994 #ifdef CONFIG_SCHED_CORE
10996 __entity_slice_used(struct sched_entity
*se
, int min_nr_tasks
)
10998 u64 slice
= sched_slice(cfs_rq_of(se
), se
);
10999 u64 rtime
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
11001 return (rtime
* min_nr_tasks
> slice
);
11004 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
11005 static inline void task_tick_core(struct rq
*rq
, struct task_struct
*curr
)
11007 if (!sched_core_enabled(rq
))
11011 * If runqueue has only one task which used up its slice and
11012 * if the sibling is forced idle, then trigger schedule to
11013 * give forced idle task a chance.
11015 * sched_slice() considers only this active rq and it gets the
11016 * whole slice. But during force idle, we have siblings acting
11017 * like a single runqueue and hence we need to consider runnable
11018 * tasks on this CPU and the forced idle CPU. Ideally, we should
11019 * go through the forced idle rq, but that would be a perf hit.
11020 * We can assume that the forced idle CPU has at least
11021 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
11022 * if we need to give up the CPU.
11024 if (rq
->core
->core_forceidle_count
&& rq
->cfs
.nr_running
== 1 &&
11025 __entity_slice_used(&curr
->se
, MIN_NR_TASKS_DURING_FORCEIDLE
))
11030 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
11032 static void se_fi_update(struct sched_entity
*se
, unsigned int fi_seq
, bool forceidle
)
11034 for_each_sched_entity(se
) {
11035 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
11038 if (cfs_rq
->forceidle_seq
== fi_seq
)
11040 cfs_rq
->forceidle_seq
= fi_seq
;
11043 cfs_rq
->min_vruntime_fi
= cfs_rq
->min_vruntime
;
11047 void task_vruntime_update(struct rq
*rq
, struct task_struct
*p
, bool in_fi
)
11049 struct sched_entity
*se
= &p
->se
;
11051 if (p
->sched_class
!= &fair_sched_class
)
11054 se_fi_update(se
, rq
->core
->core_forceidle_seq
, in_fi
);
11057 bool cfs_prio_less(struct task_struct
*a
, struct task_struct
*b
, bool in_fi
)
11059 struct rq
*rq
= task_rq(a
);
11060 struct sched_entity
*sea
= &a
->se
;
11061 struct sched_entity
*seb
= &b
->se
;
11062 struct cfs_rq
*cfs_rqa
;
11063 struct cfs_rq
*cfs_rqb
;
11066 SCHED_WARN_ON(task_rq(b
)->core
!= rq
->core
);
11068 #ifdef CONFIG_FAIR_GROUP_SCHED
11070 * Find an se in the hierarchy for tasks a and b, such that the se's
11071 * are immediate siblings.
11073 while (sea
->cfs_rq
->tg
!= seb
->cfs_rq
->tg
) {
11074 int sea_depth
= sea
->depth
;
11075 int seb_depth
= seb
->depth
;
11077 if (sea_depth
>= seb_depth
)
11078 sea
= parent_entity(sea
);
11079 if (sea_depth
<= seb_depth
)
11080 seb
= parent_entity(seb
);
11083 se_fi_update(sea
, rq
->core
->core_forceidle_seq
, in_fi
);
11084 se_fi_update(seb
, rq
->core
->core_forceidle_seq
, in_fi
);
11086 cfs_rqa
= sea
->cfs_rq
;
11087 cfs_rqb
= seb
->cfs_rq
;
11089 cfs_rqa
= &task_rq(a
)->cfs
;
11090 cfs_rqb
= &task_rq(b
)->cfs
;
11094 * Find delta after normalizing se's vruntime with its cfs_rq's
11095 * min_vruntime_fi, which would have been updated in prior calls
11096 * to se_fi_update().
11098 delta
= (s64
)(sea
->vruntime
- seb
->vruntime
) +
11099 (s64
)(cfs_rqb
->min_vruntime_fi
- cfs_rqa
->min_vruntime_fi
);
11104 static inline void task_tick_core(struct rq
*rq
, struct task_struct
*curr
) {}
11108 * scheduler tick hitting a task of our scheduling class.
11110 * NOTE: This function can be called remotely by the tick offload that
11111 * goes along full dynticks. Therefore no local assumption can be made
11112 * and everything must be accessed through the @rq and @curr passed in
11115 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
11117 struct cfs_rq
*cfs_rq
;
11118 struct sched_entity
*se
= &curr
->se
;
11120 for_each_sched_entity(se
) {
11121 cfs_rq
= cfs_rq_of(se
);
11122 entity_tick(cfs_rq
, se
, queued
);
11125 if (static_branch_unlikely(&sched_numa_balancing
))
11126 task_tick_numa(rq
, curr
);
11128 update_misfit_status(curr
, rq
);
11129 update_overutilized_status(task_rq(curr
));
11131 task_tick_core(rq
, curr
);
11135 * called on fork with the child task as argument from the parent's context
11136 * - child not yet on the tasklist
11137 * - preemption disabled
11139 static void task_fork_fair(struct task_struct
*p
)
11141 struct cfs_rq
*cfs_rq
;
11142 struct sched_entity
*se
= &p
->se
, *curr
;
11143 struct rq
*rq
= this_rq();
11144 struct rq_flags rf
;
11147 update_rq_clock(rq
);
11149 cfs_rq
= task_cfs_rq(current
);
11150 curr
= cfs_rq
->curr
;
11152 update_curr(cfs_rq
);
11153 se
->vruntime
= curr
->vruntime
;
11155 place_entity(cfs_rq
, se
, 1);
11157 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
11159 * Upon rescheduling, sched_class::put_prev_task() will place
11160 * 'current' within the tree based on its new key value.
11162 swap(curr
->vruntime
, se
->vruntime
);
11166 se
->vruntime
-= cfs_rq
->min_vruntime
;
11167 rq_unlock(rq
, &rf
);
11171 * Priority of the task has changed. Check to see if we preempt
11172 * the current task.
11175 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
11177 if (!task_on_rq_queued(p
))
11180 if (rq
->cfs
.nr_running
== 1)
11184 * Reschedule if we are currently running on this runqueue and
11185 * our priority decreased, or if we are not currently running on
11186 * this runqueue and our priority is higher than the current's
11188 if (task_current(rq
, p
)) {
11189 if (p
->prio
> oldprio
)
11192 check_preempt_curr(rq
, p
, 0);
11195 static inline bool vruntime_normalized(struct task_struct
*p
)
11197 struct sched_entity
*se
= &p
->se
;
11200 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
11201 * the dequeue_entity(.flags=0) will already have normalized the
11208 * When !on_rq, vruntime of the task has usually NOT been normalized.
11209 * But there are some cases where it has already been normalized:
11211 * - A forked child which is waiting for being woken up by
11212 * wake_up_new_task().
11213 * - A task which has been woken up by try_to_wake_up() and
11214 * waiting for actually being woken up by sched_ttwu_pending().
11216 if (!se
->sum_exec_runtime
||
11217 (READ_ONCE(p
->__state
) == TASK_WAKING
&& p
->sched_remote_wakeup
))
11223 #ifdef CONFIG_FAIR_GROUP_SCHED
11225 * Propagate the changes of the sched_entity across the tg tree to make it
11226 * visible to the root
11228 static void propagate_entity_cfs_rq(struct sched_entity
*se
)
11230 struct cfs_rq
*cfs_rq
;
11232 list_add_leaf_cfs_rq(cfs_rq_of(se
));
11234 /* Start to propagate at parent */
11237 for_each_sched_entity(se
) {
11238 cfs_rq
= cfs_rq_of(se
);
11240 if (!cfs_rq_throttled(cfs_rq
)){
11241 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
11242 list_add_leaf_cfs_rq(cfs_rq
);
11246 if (list_add_leaf_cfs_rq(cfs_rq
))
11251 static void propagate_entity_cfs_rq(struct sched_entity
*se
) { }
11254 static void detach_entity_cfs_rq(struct sched_entity
*se
)
11256 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
11258 /* Catch up with the cfs_rq and remove our load when we leave */
11259 update_load_avg(cfs_rq
, se
, 0);
11260 detach_entity_load_avg(cfs_rq
, se
);
11261 update_tg_load_avg(cfs_rq
);
11262 propagate_entity_cfs_rq(se
);
11265 static void attach_entity_cfs_rq(struct sched_entity
*se
)
11267 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
11269 #ifdef CONFIG_FAIR_GROUP_SCHED
11271 * Since the real-depth could have been changed (only FAIR
11272 * class maintain depth value), reset depth properly.
11274 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
11277 /* Synchronize entity with its cfs_rq */
11278 update_load_avg(cfs_rq
, se
, sched_feat(ATTACH_AGE_LOAD
) ? 0 : SKIP_AGE_LOAD
);
11279 attach_entity_load_avg(cfs_rq
, se
);
11280 update_tg_load_avg(cfs_rq
);
11281 propagate_entity_cfs_rq(se
);
11284 static void detach_task_cfs_rq(struct task_struct
*p
)
11286 struct sched_entity
*se
= &p
->se
;
11287 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
11289 if (!vruntime_normalized(p
)) {
11291 * Fix up our vruntime so that the current sleep doesn't
11292 * cause 'unlimited' sleep bonus.
11294 place_entity(cfs_rq
, se
, 0);
11295 se
->vruntime
-= cfs_rq
->min_vruntime
;
11298 detach_entity_cfs_rq(se
);
11301 static void attach_task_cfs_rq(struct task_struct
*p
)
11303 struct sched_entity
*se
= &p
->se
;
11304 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
11306 attach_entity_cfs_rq(se
);
11308 if (!vruntime_normalized(p
))
11309 se
->vruntime
+= cfs_rq
->min_vruntime
;
11312 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
11314 detach_task_cfs_rq(p
);
11317 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
11319 attach_task_cfs_rq(p
);
11321 if (task_on_rq_queued(p
)) {
11323 * We were most likely switched from sched_rt, so
11324 * kick off the schedule if running, otherwise just see
11325 * if we can still preempt the current task.
11327 if (task_current(rq
, p
))
11330 check_preempt_curr(rq
, p
, 0);
11334 /* Account for a task changing its policy or group.
11336 * This routine is mostly called to set cfs_rq->curr field when a task
11337 * migrates between groups/classes.
11339 static void set_next_task_fair(struct rq
*rq
, struct task_struct
*p
, bool first
)
11341 struct sched_entity
*se
= &p
->se
;
11344 if (task_on_rq_queued(p
)) {
11346 * Move the next running task to the front of the list, so our
11347 * cfs_tasks list becomes MRU one.
11349 list_move(&se
->group_node
, &rq
->cfs_tasks
);
11353 for_each_sched_entity(se
) {
11354 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
11356 set_next_entity(cfs_rq
, se
);
11357 /* ensure bandwidth has been allocated on our new cfs_rq */
11358 account_cfs_rq_runtime(cfs_rq
, 0);
11362 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
11364 cfs_rq
->tasks_timeline
= RB_ROOT_CACHED
;
11365 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
11366 #ifndef CONFIG_64BIT
11367 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
11370 raw_spin_lock_init(&cfs_rq
->removed
.lock
);
11374 #ifdef CONFIG_FAIR_GROUP_SCHED
11375 static void task_set_group_fair(struct task_struct
*p
)
11377 struct sched_entity
*se
= &p
->se
;
11379 set_task_rq(p
, task_cpu(p
));
11380 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
11383 static void task_move_group_fair(struct task_struct
*p
)
11385 detach_task_cfs_rq(p
);
11386 set_task_rq(p
, task_cpu(p
));
11389 /* Tell se's cfs_rq has been changed -- migrated */
11390 p
->se
.avg
.last_update_time
= 0;
11392 attach_task_cfs_rq(p
);
11395 static void task_change_group_fair(struct task_struct
*p
, int type
)
11398 case TASK_SET_GROUP
:
11399 task_set_group_fair(p
);
11402 case TASK_MOVE_GROUP
:
11403 task_move_group_fair(p
);
11408 void free_fair_sched_group(struct task_group
*tg
)
11412 for_each_possible_cpu(i
) {
11414 kfree(tg
->cfs_rq
[i
]);
11423 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
11425 struct sched_entity
*se
;
11426 struct cfs_rq
*cfs_rq
;
11429 tg
->cfs_rq
= kcalloc(nr_cpu_ids
, sizeof(cfs_rq
), GFP_KERNEL
);
11432 tg
->se
= kcalloc(nr_cpu_ids
, sizeof(se
), GFP_KERNEL
);
11436 tg
->shares
= NICE_0_LOAD
;
11438 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
11440 for_each_possible_cpu(i
) {
11441 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
11442 GFP_KERNEL
, cpu_to_node(i
));
11446 se
= kzalloc_node(sizeof(struct sched_entity_stats
),
11447 GFP_KERNEL
, cpu_to_node(i
));
11451 init_cfs_rq(cfs_rq
);
11452 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
11453 init_entity_runnable_average(se
);
11464 void online_fair_sched_group(struct task_group
*tg
)
11466 struct sched_entity
*se
;
11467 struct rq_flags rf
;
11471 for_each_possible_cpu(i
) {
11474 rq_lock_irq(rq
, &rf
);
11475 update_rq_clock(rq
);
11476 attach_entity_cfs_rq(se
);
11477 sync_throttle(tg
, i
);
11478 rq_unlock_irq(rq
, &rf
);
11482 void unregister_fair_sched_group(struct task_group
*tg
)
11484 unsigned long flags
;
11488 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
11490 for_each_possible_cpu(cpu
) {
11492 remove_entity_load_avg(tg
->se
[cpu
]);
11495 * Only empty task groups can be destroyed; so we can speculatively
11496 * check on_list without danger of it being re-added.
11498 if (!tg
->cfs_rq
[cpu
]->on_list
)
11503 raw_spin_rq_lock_irqsave(rq
, flags
);
11504 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
11505 raw_spin_rq_unlock_irqrestore(rq
, flags
);
11509 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
11510 struct sched_entity
*se
, int cpu
,
11511 struct sched_entity
*parent
)
11513 struct rq
*rq
= cpu_rq(cpu
);
11517 init_cfs_rq_runtime(cfs_rq
);
11519 tg
->cfs_rq
[cpu
] = cfs_rq
;
11522 /* se could be NULL for root_task_group */
11527 se
->cfs_rq
= &rq
->cfs
;
11530 se
->cfs_rq
= parent
->my_q
;
11531 se
->depth
= parent
->depth
+ 1;
11535 /* guarantee group entities always have weight */
11536 update_load_set(&se
->load
, NICE_0_LOAD
);
11537 se
->parent
= parent
;
11540 static DEFINE_MUTEX(shares_mutex
);
11542 static int __sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
11546 lockdep_assert_held(&shares_mutex
);
11549 * We can't change the weight of the root cgroup.
11554 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
11556 if (tg
->shares
== shares
)
11559 tg
->shares
= shares
;
11560 for_each_possible_cpu(i
) {
11561 struct rq
*rq
= cpu_rq(i
);
11562 struct sched_entity
*se
= tg
->se
[i
];
11563 struct rq_flags rf
;
11565 /* Propagate contribution to hierarchy */
11566 rq_lock_irqsave(rq
, &rf
);
11567 update_rq_clock(rq
);
11568 for_each_sched_entity(se
) {
11569 update_load_avg(cfs_rq_of(se
), se
, UPDATE_TG
);
11570 update_cfs_group(se
);
11572 rq_unlock_irqrestore(rq
, &rf
);
11578 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
11582 mutex_lock(&shares_mutex
);
11583 if (tg_is_idle(tg
))
11586 ret
= __sched_group_set_shares(tg
, shares
);
11587 mutex_unlock(&shares_mutex
);
11592 int sched_group_set_idle(struct task_group
*tg
, long idle
)
11596 if (tg
== &root_task_group
)
11599 if (idle
< 0 || idle
> 1)
11602 mutex_lock(&shares_mutex
);
11604 if (tg
->idle
== idle
) {
11605 mutex_unlock(&shares_mutex
);
11611 for_each_possible_cpu(i
) {
11612 struct rq
*rq
= cpu_rq(i
);
11613 struct sched_entity
*se
= tg
->se
[i
];
11614 struct cfs_rq
*parent_cfs_rq
, *grp_cfs_rq
= tg
->cfs_rq
[i
];
11615 bool was_idle
= cfs_rq_is_idle(grp_cfs_rq
);
11616 long idle_task_delta
;
11617 struct rq_flags rf
;
11619 rq_lock_irqsave(rq
, &rf
);
11621 grp_cfs_rq
->idle
= idle
;
11622 if (WARN_ON_ONCE(was_idle
== cfs_rq_is_idle(grp_cfs_rq
)))
11626 parent_cfs_rq
= cfs_rq_of(se
);
11627 if (cfs_rq_is_idle(grp_cfs_rq
))
11628 parent_cfs_rq
->idle_nr_running
++;
11630 parent_cfs_rq
->idle_nr_running
--;
11633 idle_task_delta
= grp_cfs_rq
->h_nr_running
-
11634 grp_cfs_rq
->idle_h_nr_running
;
11635 if (!cfs_rq_is_idle(grp_cfs_rq
))
11636 idle_task_delta
*= -1;
11638 for_each_sched_entity(se
) {
11639 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
11644 cfs_rq
->idle_h_nr_running
+= idle_task_delta
;
11646 /* Already accounted at parent level and above. */
11647 if (cfs_rq_is_idle(cfs_rq
))
11652 rq_unlock_irqrestore(rq
, &rf
);
11655 /* Idle groups have minimum weight. */
11656 if (tg_is_idle(tg
))
11657 __sched_group_set_shares(tg
, scale_load(WEIGHT_IDLEPRIO
));
11659 __sched_group_set_shares(tg
, NICE_0_LOAD
);
11661 mutex_unlock(&shares_mutex
);
11665 #else /* CONFIG_FAIR_GROUP_SCHED */
11667 void free_fair_sched_group(struct task_group
*tg
) { }
11669 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
11674 void online_fair_sched_group(struct task_group
*tg
) { }
11676 void unregister_fair_sched_group(struct task_group
*tg
) { }
11678 #endif /* CONFIG_FAIR_GROUP_SCHED */
11681 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
11683 struct sched_entity
*se
= &task
->se
;
11684 unsigned int rr_interval
= 0;
11687 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11690 if (rq
->cfs
.load
.weight
)
11691 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
11693 return rr_interval
;
11697 * All the scheduling class methods:
11699 DEFINE_SCHED_CLASS(fair
) = {
11701 .enqueue_task
= enqueue_task_fair
,
11702 .dequeue_task
= dequeue_task_fair
,
11703 .yield_task
= yield_task_fair
,
11704 .yield_to_task
= yield_to_task_fair
,
11706 .check_preempt_curr
= check_preempt_wakeup
,
11708 .pick_next_task
= __pick_next_task_fair
,
11709 .put_prev_task
= put_prev_task_fair
,
11710 .set_next_task
= set_next_task_fair
,
11713 .balance
= balance_fair
,
11714 .pick_task
= pick_task_fair
,
11715 .select_task_rq
= select_task_rq_fair
,
11716 .migrate_task_rq
= migrate_task_rq_fair
,
11718 .rq_online
= rq_online_fair
,
11719 .rq_offline
= rq_offline_fair
,
11721 .task_dead
= task_dead_fair
,
11722 .set_cpus_allowed
= set_cpus_allowed_common
,
11725 .task_tick
= task_tick_fair
,
11726 .task_fork
= task_fork_fair
,
11728 .prio_changed
= prio_changed_fair
,
11729 .switched_from
= switched_from_fair
,
11730 .switched_to
= switched_to_fair
,
11732 .get_rr_interval
= get_rr_interval_fair
,
11734 .update_curr
= update_curr_fair
,
11736 #ifdef CONFIG_FAIR_GROUP_SCHED
11737 .task_change_group
= task_change_group_fair
,
11740 #ifdef CONFIG_UCLAMP_TASK
11741 .uclamp_enabled
= 1,
11745 #ifdef CONFIG_SCHED_DEBUG
11746 void print_cfs_stats(struct seq_file
*m
, int cpu
)
11748 struct cfs_rq
*cfs_rq
, *pos
;
11751 for_each_leaf_cfs_rq_safe(cpu_rq(cpu
), cfs_rq
, pos
)
11752 print_cfs_rq(m
, cpu
, cfs_rq
);
11756 #ifdef CONFIG_NUMA_BALANCING
11757 void show_numa_stats(struct task_struct
*p
, struct seq_file
*m
)
11760 unsigned long tsf
= 0, tpf
= 0, gsf
= 0, gpf
= 0;
11761 struct numa_group
*ng
;
11764 ng
= rcu_dereference(p
->numa_group
);
11765 for_each_online_node(node
) {
11766 if (p
->numa_faults
) {
11767 tsf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
11768 tpf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
11771 gsf
= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)],
11772 gpf
= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
11774 print_numa_stats(m
, node
, tsf
, tpf
, gsf
, gpf
);
11778 #endif /* CONFIG_NUMA_BALANCING */
11779 #endif /* CONFIG_SCHED_DEBUG */
11781 __init
void init_sched_fair_class(void)
11784 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
11786 #ifdef CONFIG_NO_HZ_COMMON
11787 nohz
.next_balance
= jiffies
;
11788 nohz
.next_blocked
= jiffies
;
11789 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
11796 * Helper functions to facilitate extracting info from tracepoints.
11799 const struct sched_avg
*sched_trace_cfs_rq_avg(struct cfs_rq
*cfs_rq
)
11802 return cfs_rq
? &cfs_rq
->avg
: NULL
;
11807 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg
);
11809 char *sched_trace_cfs_rq_path(struct cfs_rq
*cfs_rq
, char *str
, int len
)
11813 strlcpy(str
, "(null)", len
);
11818 cfs_rq_tg_path(cfs_rq
, str
, len
);
11821 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path
);
11823 int sched_trace_cfs_rq_cpu(struct cfs_rq
*cfs_rq
)
11825 return cfs_rq
? cpu_of(rq_of(cfs_rq
)) : -1;
11827 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu
);
11829 const struct sched_avg
*sched_trace_rq_avg_rt(struct rq
*rq
)
11832 return rq
? &rq
->avg_rt
: NULL
;
11837 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt
);
11839 const struct sched_avg
*sched_trace_rq_avg_dl(struct rq
*rq
)
11842 return rq
? &rq
->avg_dl
: NULL
;
11847 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl
);
11849 const struct sched_avg
*sched_trace_rq_avg_irq(struct rq
*rq
)
11851 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
11852 return rq
? &rq
->avg_irq
: NULL
;
11857 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq
);
11859 int sched_trace_rq_cpu(struct rq
*rq
)
11861 return rq
? cpu_of(rq
) : -1;
11863 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu
);
11865 int sched_trace_rq_cpu_capacity(struct rq
*rq
)
11871 SCHED_CAPACITY_SCALE
11875 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity
);
11877 const struct cpumask
*sched_trace_rd_span(struct root_domain
*rd
)
11880 return rd
? rd
->span
: NULL
;
11885 EXPORT_SYMBOL_GPL(sched_trace_rd_span
);
11887 int sched_trace_rq_nr_running(struct rq
*rq
)
11889 return rq
? rq
->nr_running
: -1;
11891 EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running
);