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 enum sched_tunable_scaling 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 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
65 static unsigned int sched_nr_latency
= 8;
68 * After fork, child runs first. If set to 0 (default) then
69 * parent will (try to) run first.
71 unsigned int sysctl_sched_child_runs_first __read_mostly
;
74 * SCHED_OTHER wake-up granularity.
76 * This option delays the preemption effects of decoupled workloads
77 * and reduces their over-scheduling. Synchronous workloads will still
78 * have immediate wakeup/sleep latencies.
80 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
82 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
83 static unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
85 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
87 int sched_thermal_decay_shift
;
88 static int __init
setup_sched_thermal_decay_shift(char *str
)
92 if (kstrtoint(str
, 0, &_shift
))
93 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
95 sched_thermal_decay_shift
= clamp(_shift
, 0, 10);
98 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift
);
102 * For asym packing, by default the lower numbered CPU has higher priority.
104 int __weak
arch_asym_cpu_priority(int cpu
)
110 * The margin used when comparing utilization with CPU capacity.
114 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
118 #ifdef CONFIG_CFS_BANDWIDTH
120 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
121 * each time a cfs_rq requests quota.
123 * Note: in the case that the slice exceeds the runtime remaining (either due
124 * to consumption or the quota being specified to be smaller than the slice)
125 * we will always only issue the remaining available time.
127 * (default: 5 msec, units: microseconds)
129 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
132 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
138 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
144 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
151 * Increase the granularity value when there are more CPUs,
152 * because with more CPUs the 'effective latency' as visible
153 * to users decreases. But the relationship is not linear,
154 * so pick a second-best guess by going with the log2 of the
157 * This idea comes from the SD scheduler of Con Kolivas:
159 static unsigned int get_update_sysctl_factor(void)
161 unsigned int cpus
= min_t(unsigned int, num_online_cpus(), 8);
164 switch (sysctl_sched_tunable_scaling
) {
165 case SCHED_TUNABLESCALING_NONE
:
168 case SCHED_TUNABLESCALING_LINEAR
:
171 case SCHED_TUNABLESCALING_LOG
:
173 factor
= 1 + ilog2(cpus
);
180 static void update_sysctl(void)
182 unsigned int factor
= get_update_sysctl_factor();
184 #define SET_SYSCTL(name) \
185 (sysctl_##name = (factor) * normalized_sysctl_##name)
186 SET_SYSCTL(sched_min_granularity
);
187 SET_SYSCTL(sched_latency
);
188 SET_SYSCTL(sched_wakeup_granularity
);
192 void __init
sched_init_granularity(void)
197 #define WMULT_CONST (~0U)
198 #define WMULT_SHIFT 32
200 static void __update_inv_weight(struct load_weight
*lw
)
204 if (likely(lw
->inv_weight
))
207 w
= scale_load_down(lw
->weight
);
209 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
211 else if (unlikely(!w
))
212 lw
->inv_weight
= WMULT_CONST
;
214 lw
->inv_weight
= WMULT_CONST
/ w
;
218 * delta_exec * weight / lw.weight
220 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
222 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
223 * we're guaranteed shift stays positive because inv_weight is guaranteed to
224 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
226 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
227 * weight/lw.weight <= 1, and therefore our shift will also be positive.
229 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
231 u64 fact
= scale_load_down(weight
);
232 int shift
= WMULT_SHIFT
;
234 __update_inv_weight(lw
);
236 if (unlikely(fact
>> 32)) {
243 fact
= mul_u32_u32(fact
, lw
->inv_weight
);
250 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
254 const struct sched_class fair_sched_class
;
256 /**************************************************************
257 * CFS operations on generic schedulable entities:
260 #ifdef CONFIG_FAIR_GROUP_SCHED
261 static inline struct task_struct
*task_of(struct sched_entity
*se
)
263 SCHED_WARN_ON(!entity_is_task(se
));
264 return container_of(se
, struct task_struct
, se
);
267 /* Walk up scheduling entities hierarchy */
268 #define for_each_sched_entity(se) \
269 for (; se; se = se->parent)
271 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
276 /* runqueue on which this entity is (to be) queued */
277 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
282 /* runqueue "owned" by this group */
283 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
288 static inline void cfs_rq_tg_path(struct cfs_rq
*cfs_rq
, char *path
, int len
)
293 if (cfs_rq
&& task_group_is_autogroup(cfs_rq
->tg
))
294 autogroup_path(cfs_rq
->tg
, path
, len
);
295 else if (cfs_rq
&& cfs_rq
->tg
->css
.cgroup
)
296 cgroup_path(cfs_rq
->tg
->css
.cgroup
, path
, len
);
298 strlcpy(path
, "(null)", len
);
301 static inline bool list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
303 struct rq
*rq
= rq_of(cfs_rq
);
304 int cpu
= cpu_of(rq
);
307 return rq
->tmp_alone_branch
== &rq
->leaf_cfs_rq_list
;
312 * Ensure we either appear before our parent (if already
313 * enqueued) or force our parent to appear after us when it is
314 * enqueued. The fact that we always enqueue bottom-up
315 * reduces this to two cases and a special case for the root
316 * cfs_rq. Furthermore, it also means that we will always reset
317 * tmp_alone_branch either when the branch is connected
318 * to a tree or when we reach the top of the tree
320 if (cfs_rq
->tg
->parent
&&
321 cfs_rq
->tg
->parent
->cfs_rq
[cpu
]->on_list
) {
323 * If parent is already on the list, we add the child
324 * just before. Thanks to circular linked property of
325 * the list, this means to put the child at the tail
326 * of the list that starts by parent.
328 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
329 &(cfs_rq
->tg
->parent
->cfs_rq
[cpu
]->leaf_cfs_rq_list
));
331 * The branch is now connected to its tree so we can
332 * reset tmp_alone_branch to the beginning of the
335 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
339 if (!cfs_rq
->tg
->parent
) {
341 * cfs rq without parent should be put
342 * at the tail of the list.
344 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
345 &rq
->leaf_cfs_rq_list
);
347 * We have reach the top of a tree so we can reset
348 * tmp_alone_branch to the beginning of the list.
350 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
355 * The parent has not already been added so we want to
356 * make sure that it will be put after us.
357 * tmp_alone_branch points to the begin of the branch
358 * where we will add parent.
360 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, rq
->tmp_alone_branch
);
362 * update tmp_alone_branch to points to the new begin
365 rq
->tmp_alone_branch
= &cfs_rq
->leaf_cfs_rq_list
;
369 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
371 if (cfs_rq
->on_list
) {
372 struct rq
*rq
= rq_of(cfs_rq
);
375 * With cfs_rq being unthrottled/throttled during an enqueue,
376 * it can happen the tmp_alone_branch points the a leaf that
377 * we finally want to del. In this case, tmp_alone_branch moves
378 * to the prev element but it will point to rq->leaf_cfs_rq_list
379 * at the end of the enqueue.
381 if (rq
->tmp_alone_branch
== &cfs_rq
->leaf_cfs_rq_list
)
382 rq
->tmp_alone_branch
= cfs_rq
->leaf_cfs_rq_list
.prev
;
384 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
389 static inline void assert_list_leaf_cfs_rq(struct rq
*rq
)
391 SCHED_WARN_ON(rq
->tmp_alone_branch
!= &rq
->leaf_cfs_rq_list
);
394 /* Iterate thr' all leaf cfs_rq's on a runqueue */
395 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
396 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
399 /* Do the two (enqueued) entities belong to the same group ? */
400 static inline struct cfs_rq
*
401 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
403 if (se
->cfs_rq
== pse
->cfs_rq
)
409 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
415 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
417 int se_depth
, pse_depth
;
420 * preemption test can be made between sibling entities who are in the
421 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
422 * both tasks until we find their ancestors who are siblings of common
426 /* First walk up until both entities are at same depth */
427 se_depth
= (*se
)->depth
;
428 pse_depth
= (*pse
)->depth
;
430 while (se_depth
> pse_depth
) {
432 *se
= parent_entity(*se
);
435 while (pse_depth
> se_depth
) {
437 *pse
= parent_entity(*pse
);
440 while (!is_same_group(*se
, *pse
)) {
441 *se
= parent_entity(*se
);
442 *pse
= parent_entity(*pse
);
446 #else /* !CONFIG_FAIR_GROUP_SCHED */
448 static inline struct task_struct
*task_of(struct sched_entity
*se
)
450 return container_of(se
, struct task_struct
, se
);
453 #define for_each_sched_entity(se) \
454 for (; se; se = NULL)
456 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
458 return &task_rq(p
)->cfs
;
461 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
463 struct task_struct
*p
= task_of(se
);
464 struct rq
*rq
= task_rq(p
);
469 /* runqueue "owned" by this group */
470 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
475 static inline void cfs_rq_tg_path(struct cfs_rq
*cfs_rq
, char *path
, int len
)
478 strlcpy(path
, "(null)", len
);
481 static inline bool list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
486 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
490 static inline void assert_list_leaf_cfs_rq(struct rq
*rq
)
494 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
495 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
497 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
503 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
507 #endif /* CONFIG_FAIR_GROUP_SCHED */
509 static __always_inline
510 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
512 /**************************************************************
513 * Scheduling class tree data structure manipulation methods:
516 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
518 s64 delta
= (s64
)(vruntime
- max_vruntime
);
520 max_vruntime
= vruntime
;
525 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
527 s64 delta
= (s64
)(vruntime
- min_vruntime
);
529 min_vruntime
= vruntime
;
534 static inline int entity_before(struct sched_entity
*a
,
535 struct sched_entity
*b
)
537 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
540 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
542 struct sched_entity
*curr
= cfs_rq
->curr
;
543 struct rb_node
*leftmost
= rb_first_cached(&cfs_rq
->tasks_timeline
);
545 u64 vruntime
= cfs_rq
->min_vruntime
;
549 vruntime
= curr
->vruntime
;
554 if (leftmost
) { /* non-empty tree */
555 struct sched_entity
*se
;
556 se
= rb_entry(leftmost
, struct sched_entity
, run_node
);
559 vruntime
= se
->vruntime
;
561 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
564 /* ensure we never gain time by being placed backwards. */
565 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
568 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
573 * Enqueue an entity into the rb-tree:
575 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
577 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_root
.rb_node
;
578 struct rb_node
*parent
= NULL
;
579 struct sched_entity
*entry
;
580 bool leftmost
= true;
583 * Find the right place in the rbtree:
587 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
589 * We dont care about collisions. Nodes with
590 * the same key stay together.
592 if (entity_before(se
, entry
)) {
593 link
= &parent
->rb_left
;
595 link
= &parent
->rb_right
;
600 rb_link_node(&se
->run_node
, parent
, link
);
601 rb_insert_color_cached(&se
->run_node
,
602 &cfs_rq
->tasks_timeline
, leftmost
);
605 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
607 rb_erase_cached(&se
->run_node
, &cfs_rq
->tasks_timeline
);
610 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
612 struct rb_node
*left
= rb_first_cached(&cfs_rq
->tasks_timeline
);
617 return rb_entry(left
, struct sched_entity
, run_node
);
620 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
622 struct rb_node
*next
= rb_next(&se
->run_node
);
627 return rb_entry(next
, struct sched_entity
, run_node
);
630 #ifdef CONFIG_SCHED_DEBUG
631 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
633 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
.rb_root
);
638 return rb_entry(last
, struct sched_entity
, run_node
);
641 /**************************************************************
642 * Scheduling class statistics methods:
645 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
646 void *buffer
, size_t *lenp
, loff_t
*ppos
)
648 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
649 unsigned int factor
= get_update_sysctl_factor();
654 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
655 sysctl_sched_min_granularity
);
657 #define WRT_SYSCTL(name) \
658 (normalized_sysctl_##name = sysctl_##name / (factor))
659 WRT_SYSCTL(sched_min_granularity
);
660 WRT_SYSCTL(sched_latency
);
661 WRT_SYSCTL(sched_wakeup_granularity
);
671 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
673 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
674 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
680 * The idea is to set a period in which each task runs once.
682 * When there are too many tasks (sched_nr_latency) we have to stretch
683 * this period because otherwise the slices get too small.
685 * p = (nr <= nl) ? l : l*nr/nl
687 static u64
__sched_period(unsigned long nr_running
)
689 if (unlikely(nr_running
> sched_nr_latency
))
690 return nr_running
* sysctl_sched_min_granularity
;
692 return sysctl_sched_latency
;
696 * We calculate the wall-time slice from the period by taking a part
697 * proportional to the weight.
701 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
703 unsigned int nr_running
= cfs_rq
->nr_running
;
706 if (sched_feat(ALT_PERIOD
))
707 nr_running
= rq_of(cfs_rq
)->cfs
.h_nr_running
;
709 slice
= __sched_period(nr_running
+ !se
->on_rq
);
711 for_each_sched_entity(se
) {
712 struct load_weight
*load
;
713 struct load_weight lw
;
715 cfs_rq
= cfs_rq_of(se
);
716 load
= &cfs_rq
->load
;
718 if (unlikely(!se
->on_rq
)) {
721 update_load_add(&lw
, se
->load
.weight
);
724 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
727 if (sched_feat(BASE_SLICE
))
728 slice
= max(slice
, (u64
)sysctl_sched_min_granularity
);
734 * We calculate the vruntime slice of a to-be-inserted task.
738 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
740 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
746 static int select_idle_sibling(struct task_struct
*p
, int prev_cpu
, int cpu
);
747 static unsigned long task_h_load(struct task_struct
*p
);
748 static unsigned long capacity_of(int cpu
);
750 /* Give new sched_entity start runnable values to heavy its load in infant time */
751 void init_entity_runnable_average(struct sched_entity
*se
)
753 struct sched_avg
*sa
= &se
->avg
;
755 memset(sa
, 0, sizeof(*sa
));
758 * Tasks are initialized with full load to be seen as heavy tasks until
759 * they get a chance to stabilize to their real load level.
760 * Group entities are initialized with zero load to reflect the fact that
761 * nothing has been attached to the task group yet.
763 if (entity_is_task(se
))
764 sa
->load_avg
= scale_load_down(se
->load
.weight
);
766 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
769 static void attach_entity_cfs_rq(struct sched_entity
*se
);
772 * With new tasks being created, their initial util_avgs are extrapolated
773 * based on the cfs_rq's current util_avg:
775 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
777 * However, in many cases, the above util_avg does not give a desired
778 * value. Moreover, the sum of the util_avgs may be divergent, such
779 * as when the series is a harmonic series.
781 * To solve this problem, we also cap the util_avg of successive tasks to
782 * only 1/2 of the left utilization budget:
784 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
786 * where n denotes the nth task and cpu_scale the CPU capacity.
788 * For example, for a CPU with 1024 of capacity, a simplest series from
789 * the beginning would be like:
791 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
792 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
794 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
795 * if util_avg > util_avg_cap.
797 void post_init_entity_util_avg(struct task_struct
*p
)
799 struct sched_entity
*se
= &p
->se
;
800 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
801 struct sched_avg
*sa
= &se
->avg
;
802 long cpu_scale
= arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq
)));
803 long cap
= (long)(cpu_scale
- cfs_rq
->avg
.util_avg
) / 2;
806 if (cfs_rq
->avg
.util_avg
!= 0) {
807 sa
->util_avg
= cfs_rq
->avg
.util_avg
* se
->load
.weight
;
808 sa
->util_avg
/= (cfs_rq
->avg
.load_avg
+ 1);
810 if (sa
->util_avg
> cap
)
817 sa
->runnable_avg
= sa
->util_avg
;
819 if (p
->sched_class
!= &fair_sched_class
) {
821 * For !fair tasks do:
823 update_cfs_rq_load_avg(now, cfs_rq);
824 attach_entity_load_avg(cfs_rq, se);
825 switched_from_fair(rq, p);
827 * such that the next switched_to_fair() has the
830 se
->avg
.last_update_time
= cfs_rq_clock_pelt(cfs_rq
);
834 attach_entity_cfs_rq(se
);
837 #else /* !CONFIG_SMP */
838 void init_entity_runnable_average(struct sched_entity
*se
)
841 void post_init_entity_util_avg(struct task_struct
*p
)
844 static void update_tg_load_avg(struct cfs_rq
*cfs_rq
)
847 #endif /* CONFIG_SMP */
850 * Update the current task's runtime statistics.
852 static void update_curr(struct cfs_rq
*cfs_rq
)
854 struct sched_entity
*curr
= cfs_rq
->curr
;
855 u64 now
= rq_clock_task(rq_of(cfs_rq
));
861 delta_exec
= now
- curr
->exec_start
;
862 if (unlikely((s64
)delta_exec
<= 0))
865 curr
->exec_start
= now
;
867 schedstat_set(curr
->statistics
.exec_max
,
868 max(delta_exec
, curr
->statistics
.exec_max
));
870 curr
->sum_exec_runtime
+= delta_exec
;
871 schedstat_add(cfs_rq
->exec_clock
, delta_exec
);
873 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
874 update_min_vruntime(cfs_rq
);
876 if (entity_is_task(curr
)) {
877 struct task_struct
*curtask
= task_of(curr
);
879 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
880 cgroup_account_cputime(curtask
, delta_exec
);
881 account_group_exec_runtime(curtask
, delta_exec
);
884 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
887 static void update_curr_fair(struct rq
*rq
)
889 update_curr(cfs_rq_of(&rq
->curr
->se
));
893 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
895 u64 wait_start
, prev_wait_start
;
897 if (!schedstat_enabled())
900 wait_start
= rq_clock(rq_of(cfs_rq
));
901 prev_wait_start
= schedstat_val(se
->statistics
.wait_start
);
903 if (entity_is_task(se
) && task_on_rq_migrating(task_of(se
)) &&
904 likely(wait_start
> prev_wait_start
))
905 wait_start
-= prev_wait_start
;
907 __schedstat_set(se
->statistics
.wait_start
, wait_start
);
911 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
913 struct task_struct
*p
;
916 if (!schedstat_enabled())
920 * When the sched_schedstat changes from 0 to 1, some sched se
921 * maybe already in the runqueue, the se->statistics.wait_start
922 * will be 0.So it will let the delta wrong. We need to avoid this
925 if (unlikely(!schedstat_val(se
->statistics
.wait_start
)))
928 delta
= rq_clock(rq_of(cfs_rq
)) - schedstat_val(se
->statistics
.wait_start
);
930 if (entity_is_task(se
)) {
932 if (task_on_rq_migrating(p
)) {
934 * Preserve migrating task's wait time so wait_start
935 * time stamp can be adjusted to accumulate wait time
936 * prior to migration.
938 __schedstat_set(se
->statistics
.wait_start
, delta
);
941 trace_sched_stat_wait(p
, delta
);
944 __schedstat_set(se
->statistics
.wait_max
,
945 max(schedstat_val(se
->statistics
.wait_max
), delta
));
946 __schedstat_inc(se
->statistics
.wait_count
);
947 __schedstat_add(se
->statistics
.wait_sum
, delta
);
948 __schedstat_set(se
->statistics
.wait_start
, 0);
952 update_stats_enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
954 struct task_struct
*tsk
= NULL
;
955 u64 sleep_start
, block_start
;
957 if (!schedstat_enabled())
960 sleep_start
= schedstat_val(se
->statistics
.sleep_start
);
961 block_start
= schedstat_val(se
->statistics
.block_start
);
963 if (entity_is_task(se
))
967 u64 delta
= rq_clock(rq_of(cfs_rq
)) - sleep_start
;
972 if (unlikely(delta
> schedstat_val(se
->statistics
.sleep_max
)))
973 __schedstat_set(se
->statistics
.sleep_max
, delta
);
975 __schedstat_set(se
->statistics
.sleep_start
, 0);
976 __schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
979 account_scheduler_latency(tsk
, delta
>> 10, 1);
980 trace_sched_stat_sleep(tsk
, delta
);
984 u64 delta
= rq_clock(rq_of(cfs_rq
)) - block_start
;
989 if (unlikely(delta
> schedstat_val(se
->statistics
.block_max
)))
990 __schedstat_set(se
->statistics
.block_max
, delta
);
992 __schedstat_set(se
->statistics
.block_start
, 0);
993 __schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
996 if (tsk
->in_iowait
) {
997 __schedstat_add(se
->statistics
.iowait_sum
, delta
);
998 __schedstat_inc(se
->statistics
.iowait_count
);
999 trace_sched_stat_iowait(tsk
, delta
);
1002 trace_sched_stat_blocked(tsk
, delta
);
1005 * Blocking time is in units of nanosecs, so shift by
1006 * 20 to get a milliseconds-range estimation of the
1007 * amount of time that the task spent sleeping:
1009 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
1010 profile_hits(SLEEP_PROFILING
,
1011 (void *)get_wchan(tsk
),
1014 account_scheduler_latency(tsk
, delta
>> 10, 0);
1020 * Task is being enqueued - update stats:
1023 update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1025 if (!schedstat_enabled())
1029 * Are we enqueueing a waiting task? (for current tasks
1030 * a dequeue/enqueue event is a NOP)
1032 if (se
!= cfs_rq
->curr
)
1033 update_stats_wait_start(cfs_rq
, se
);
1035 if (flags
& ENQUEUE_WAKEUP
)
1036 update_stats_enqueue_sleeper(cfs_rq
, se
);
1040 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1043 if (!schedstat_enabled())
1047 * Mark the end of the wait period if dequeueing a
1050 if (se
!= cfs_rq
->curr
)
1051 update_stats_wait_end(cfs_rq
, se
);
1053 if ((flags
& DEQUEUE_SLEEP
) && entity_is_task(se
)) {
1054 struct task_struct
*tsk
= task_of(se
);
1056 if (tsk
->state
& TASK_INTERRUPTIBLE
)
1057 __schedstat_set(se
->statistics
.sleep_start
,
1058 rq_clock(rq_of(cfs_rq
)));
1059 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
1060 __schedstat_set(se
->statistics
.block_start
,
1061 rq_clock(rq_of(cfs_rq
)));
1066 * We are picking a new current task - update its stats:
1069 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1072 * We are starting a new run period:
1074 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
1077 /**************************************************
1078 * Scheduling class queueing methods:
1081 #ifdef CONFIG_NUMA_BALANCING
1083 * Approximate time to scan a full NUMA task in ms. The task scan period is
1084 * calculated based on the tasks virtual memory size and
1085 * numa_balancing_scan_size.
1087 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
1088 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
1090 /* Portion of address space to scan in MB */
1091 unsigned int sysctl_numa_balancing_scan_size
= 256;
1093 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1094 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
1097 refcount_t refcount
;
1099 spinlock_t lock
; /* nr_tasks, tasks */
1104 struct rcu_head rcu
;
1105 unsigned long total_faults
;
1106 unsigned long max_faults_cpu
;
1108 * Faults_cpu is used to decide whether memory should move
1109 * towards the CPU. As a consequence, these stats are weighted
1110 * more by CPU use than by memory faults.
1112 unsigned long *faults_cpu
;
1113 unsigned long faults
[];
1117 * For functions that can be called in multiple contexts that permit reading
1118 * ->numa_group (see struct task_struct for locking rules).
1120 static struct numa_group
*deref_task_numa_group(struct task_struct
*p
)
1122 return rcu_dereference_check(p
->numa_group
, p
== current
||
1123 (lockdep_is_held(&task_rq(p
)->lock
) && !READ_ONCE(p
->on_cpu
)));
1126 static struct numa_group
*deref_curr_numa_group(struct task_struct
*p
)
1128 return rcu_dereference_protected(p
->numa_group
, p
== current
);
1131 static inline unsigned long group_faults_priv(struct numa_group
*ng
);
1132 static inline unsigned long group_faults_shared(struct numa_group
*ng
);
1134 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
1136 unsigned long rss
= 0;
1137 unsigned long nr_scan_pages
;
1140 * Calculations based on RSS as non-present and empty pages are skipped
1141 * by the PTE scanner and NUMA hinting faults should be trapped based
1144 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
1145 rss
= get_mm_rss(p
->mm
);
1147 rss
= nr_scan_pages
;
1149 rss
= round_up(rss
, nr_scan_pages
);
1150 return rss
/ nr_scan_pages
;
1153 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1154 #define MAX_SCAN_WINDOW 2560
1156 static unsigned int task_scan_min(struct task_struct
*p
)
1158 unsigned int scan_size
= READ_ONCE(sysctl_numa_balancing_scan_size
);
1159 unsigned int scan
, floor
;
1160 unsigned int windows
= 1;
1162 if (scan_size
< MAX_SCAN_WINDOW
)
1163 windows
= MAX_SCAN_WINDOW
/ scan_size
;
1164 floor
= 1000 / windows
;
1166 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
1167 return max_t(unsigned int, floor
, scan
);
1170 static unsigned int task_scan_start(struct task_struct
*p
)
1172 unsigned long smin
= task_scan_min(p
);
1173 unsigned long period
= smin
;
1174 struct numa_group
*ng
;
1176 /* Scale the maximum scan period with the amount of shared memory. */
1178 ng
= rcu_dereference(p
->numa_group
);
1180 unsigned long shared
= group_faults_shared(ng
);
1181 unsigned long private = group_faults_priv(ng
);
1183 period
*= refcount_read(&ng
->refcount
);
1184 period
*= shared
+ 1;
1185 period
/= private + shared
+ 1;
1189 return max(smin
, period
);
1192 static unsigned int task_scan_max(struct task_struct
*p
)
1194 unsigned long smin
= task_scan_min(p
);
1196 struct numa_group
*ng
;
1198 /* Watch for min being lower than max due to floor calculations */
1199 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
1201 /* Scale the maximum scan period with the amount of shared memory. */
1202 ng
= deref_curr_numa_group(p
);
1204 unsigned long shared
= group_faults_shared(ng
);
1205 unsigned long private = group_faults_priv(ng
);
1206 unsigned long period
= smax
;
1208 period
*= refcount_read(&ng
->refcount
);
1209 period
*= shared
+ 1;
1210 period
/= private + shared
+ 1;
1212 smax
= max(smax
, period
);
1215 return max(smin
, smax
);
1218 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
1220 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= NUMA_NO_NODE
);
1221 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
1224 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
1226 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= NUMA_NO_NODE
);
1227 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
1230 /* Shared or private faults. */
1231 #define NR_NUMA_HINT_FAULT_TYPES 2
1233 /* Memory and CPU locality */
1234 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1236 /* Averaged statistics, and temporary buffers. */
1237 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1239 pid_t
task_numa_group_id(struct task_struct
*p
)
1241 struct numa_group
*ng
;
1245 ng
= rcu_dereference(p
->numa_group
);
1254 * The averaged statistics, shared & private, memory & CPU,
1255 * occupy the first half of the array. The second half of the
1256 * array is for current counters, which are averaged into the
1257 * first set by task_numa_placement.
1259 static inline int task_faults_idx(enum numa_faults_stats s
, int nid
, int priv
)
1261 return NR_NUMA_HINT_FAULT_TYPES
* (s
* nr_node_ids
+ nid
) + priv
;
1264 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
1266 if (!p
->numa_faults
)
1269 return p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1270 p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1273 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
1275 struct numa_group
*ng
= deref_task_numa_group(p
);
1280 return ng
->faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1281 ng
->faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1284 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
1286 return group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1287 group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1290 static inline unsigned long group_faults_priv(struct numa_group
*ng
)
1292 unsigned long faults
= 0;
1295 for_each_online_node(node
) {
1296 faults
+= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
1302 static inline unsigned long group_faults_shared(struct numa_group
*ng
)
1304 unsigned long faults
= 0;
1307 for_each_online_node(node
) {
1308 faults
+= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
1315 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1316 * considered part of a numa group's pseudo-interleaving set. Migrations
1317 * between these nodes are slowed down, to allow things to settle down.
1319 #define ACTIVE_NODE_FRACTION 3
1321 static bool numa_is_active_node(int nid
, struct numa_group
*ng
)
1323 return group_faults_cpu(ng
, nid
) * ACTIVE_NODE_FRACTION
> ng
->max_faults_cpu
;
1326 /* Handle placement on systems where not all nodes are directly connected. */
1327 static unsigned long score_nearby_nodes(struct task_struct
*p
, int nid
,
1328 int maxdist
, bool task
)
1330 unsigned long score
= 0;
1334 * All nodes are directly connected, and the same distance
1335 * from each other. No need for fancy placement algorithms.
1337 if (sched_numa_topology_type
== NUMA_DIRECT
)
1341 * This code is called for each node, introducing N^2 complexity,
1342 * which should be ok given the number of nodes rarely exceeds 8.
1344 for_each_online_node(node
) {
1345 unsigned long faults
;
1346 int dist
= node_distance(nid
, node
);
1349 * The furthest away nodes in the system are not interesting
1350 * for placement; nid was already counted.
1352 if (dist
== sched_max_numa_distance
|| node
== nid
)
1356 * On systems with a backplane NUMA topology, compare groups
1357 * of nodes, and move tasks towards the group with the most
1358 * memory accesses. When comparing two nodes at distance
1359 * "hoplimit", only nodes closer by than "hoplimit" are part
1360 * of each group. Skip other nodes.
1362 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1366 /* Add up the faults from nearby nodes. */
1368 faults
= task_faults(p
, node
);
1370 faults
= group_faults(p
, node
);
1373 * On systems with a glueless mesh NUMA topology, there are
1374 * no fixed "groups of nodes". Instead, nodes that are not
1375 * directly connected bounce traffic through intermediate
1376 * nodes; a numa_group can occupy any set of nodes.
1377 * The further away a node is, the less the faults count.
1378 * This seems to result in good task placement.
1380 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1381 faults
*= (sched_max_numa_distance
- dist
);
1382 faults
/= (sched_max_numa_distance
- LOCAL_DISTANCE
);
1392 * These return the fraction of accesses done by a particular task, or
1393 * task group, on a particular numa node. The group weight is given a
1394 * larger multiplier, in order to group tasks together that are almost
1395 * evenly spread out between numa nodes.
1397 static inline unsigned long task_weight(struct task_struct
*p
, int nid
,
1400 unsigned long faults
, total_faults
;
1402 if (!p
->numa_faults
)
1405 total_faults
= p
->total_numa_faults
;
1410 faults
= task_faults(p
, nid
);
1411 faults
+= score_nearby_nodes(p
, nid
, dist
, true);
1413 return 1000 * faults
/ total_faults
;
1416 static inline unsigned long group_weight(struct task_struct
*p
, int nid
,
1419 struct numa_group
*ng
= deref_task_numa_group(p
);
1420 unsigned long faults
, total_faults
;
1425 total_faults
= ng
->total_faults
;
1430 faults
= group_faults(p
, nid
);
1431 faults
+= score_nearby_nodes(p
, nid
, dist
, false);
1433 return 1000 * faults
/ total_faults
;
1436 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
1437 int src_nid
, int dst_cpu
)
1439 struct numa_group
*ng
= deref_curr_numa_group(p
);
1440 int dst_nid
= cpu_to_node(dst_cpu
);
1441 int last_cpupid
, this_cpupid
;
1443 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
1444 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
1447 * Allow first faults or private faults to migrate immediately early in
1448 * the lifetime of a task. The magic number 4 is based on waiting for
1449 * two full passes of the "multi-stage node selection" test that is
1452 if ((p
->numa_preferred_nid
== NUMA_NO_NODE
|| p
->numa_scan_seq
<= 4) &&
1453 (cpupid_pid_unset(last_cpupid
) || cpupid_match_pid(p
, last_cpupid
)))
1457 * Multi-stage node selection is used in conjunction with a periodic
1458 * migration fault to build a temporal task<->page relation. By using
1459 * a two-stage filter we remove short/unlikely relations.
1461 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1462 * a task's usage of a particular page (n_p) per total usage of this
1463 * page (n_t) (in a given time-span) to a probability.
1465 * Our periodic faults will sample this probability and getting the
1466 * same result twice in a row, given these samples are fully
1467 * independent, is then given by P(n)^2, provided our sample period
1468 * is sufficiently short compared to the usage pattern.
1470 * This quadric squishes small probabilities, making it less likely we
1471 * act on an unlikely task<->page relation.
1473 if (!cpupid_pid_unset(last_cpupid
) &&
1474 cpupid_to_nid(last_cpupid
) != dst_nid
)
1477 /* Always allow migrate on private faults */
1478 if (cpupid_match_pid(p
, last_cpupid
))
1481 /* A shared fault, but p->numa_group has not been set up yet. */
1486 * Destination node is much more heavily used than the source
1487 * node? Allow migration.
1489 if (group_faults_cpu(ng
, dst_nid
) > group_faults_cpu(ng
, src_nid
) *
1490 ACTIVE_NODE_FRACTION
)
1494 * Distribute memory according to CPU & memory use on each node,
1495 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1497 * faults_cpu(dst) 3 faults_cpu(src)
1498 * --------------- * - > ---------------
1499 * faults_mem(dst) 4 faults_mem(src)
1501 return group_faults_cpu(ng
, dst_nid
) * group_faults(p
, src_nid
) * 3 >
1502 group_faults_cpu(ng
, src_nid
) * group_faults(p
, dst_nid
) * 4;
1506 * 'numa_type' describes the node at the moment of load balancing.
1509 /* The node has spare capacity that can be used to run more tasks. */
1512 * The node is fully used and the tasks don't compete for more CPU
1513 * cycles. Nevertheless, some tasks might wait before running.
1517 * The node is overloaded and can't provide expected CPU cycles to all
1523 /* Cached statistics for all CPUs within a node */
1526 unsigned long runnable
;
1528 /* Total compute capacity of CPUs on a node */
1529 unsigned long compute_capacity
;
1530 unsigned int nr_running
;
1531 unsigned int weight
;
1532 enum numa_type node_type
;
1536 static inline bool is_core_idle(int cpu
)
1538 #ifdef CONFIG_SCHED_SMT
1541 for_each_cpu(sibling
, cpu_smt_mask(cpu
)) {
1553 struct task_numa_env
{
1554 struct task_struct
*p
;
1556 int src_cpu
, src_nid
;
1557 int dst_cpu
, dst_nid
;
1559 struct numa_stats src_stats
, dst_stats
;
1564 struct task_struct
*best_task
;
1569 static unsigned long cpu_load(struct rq
*rq
);
1570 static unsigned long cpu_runnable(struct rq
*rq
);
1571 static unsigned long cpu_util(int cpu
);
1572 static inline long adjust_numa_imbalance(int imbalance
,
1573 int dst_running
, int dst_weight
);
1576 numa_type
numa_classify(unsigned int imbalance_pct
,
1577 struct numa_stats
*ns
)
1579 if ((ns
->nr_running
> ns
->weight
) &&
1580 (((ns
->compute_capacity
* 100) < (ns
->util
* imbalance_pct
)) ||
1581 ((ns
->compute_capacity
* imbalance_pct
) < (ns
->runnable
* 100))))
1582 return node_overloaded
;
1584 if ((ns
->nr_running
< ns
->weight
) ||
1585 (((ns
->compute_capacity
* 100) > (ns
->util
* imbalance_pct
)) &&
1586 ((ns
->compute_capacity
* imbalance_pct
) > (ns
->runnable
* 100))))
1587 return node_has_spare
;
1589 return node_fully_busy
;
1592 #ifdef CONFIG_SCHED_SMT
1593 /* Forward declarations of select_idle_sibling helpers */
1594 static inline bool test_idle_cores(int cpu
, bool def
);
1595 static inline int numa_idle_core(int idle_core
, int cpu
)
1597 if (!static_branch_likely(&sched_smt_present
) ||
1598 idle_core
>= 0 || !test_idle_cores(cpu
, false))
1602 * Prefer cores instead of packing HT siblings
1603 * and triggering future load balancing.
1605 if (is_core_idle(cpu
))
1611 static inline int numa_idle_core(int idle_core
, int cpu
)
1618 * Gather all necessary information to make NUMA balancing placement
1619 * decisions that are compatible with standard load balancer. This
1620 * borrows code and logic from update_sg_lb_stats but sharing a
1621 * common implementation is impractical.
1623 static void update_numa_stats(struct task_numa_env
*env
,
1624 struct numa_stats
*ns
, int nid
,
1627 int cpu
, idle_core
= -1;
1629 memset(ns
, 0, sizeof(*ns
));
1633 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1634 struct rq
*rq
= cpu_rq(cpu
);
1636 ns
->load
+= cpu_load(rq
);
1637 ns
->runnable
+= cpu_runnable(rq
);
1638 ns
->util
+= cpu_util(cpu
);
1639 ns
->nr_running
+= rq
->cfs
.h_nr_running
;
1640 ns
->compute_capacity
+= capacity_of(cpu
);
1642 if (find_idle
&& !rq
->nr_running
&& idle_cpu(cpu
)) {
1643 if (READ_ONCE(rq
->numa_migrate_on
) ||
1644 !cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
))
1647 if (ns
->idle_cpu
== -1)
1650 idle_core
= numa_idle_core(idle_core
, cpu
);
1655 ns
->weight
= cpumask_weight(cpumask_of_node(nid
));
1657 ns
->node_type
= numa_classify(env
->imbalance_pct
, ns
);
1660 ns
->idle_cpu
= idle_core
;
1663 static void task_numa_assign(struct task_numa_env
*env
,
1664 struct task_struct
*p
, long imp
)
1666 struct rq
*rq
= cpu_rq(env
->dst_cpu
);
1668 /* Check if run-queue part of active NUMA balance. */
1669 if (env
->best_cpu
!= env
->dst_cpu
&& xchg(&rq
->numa_migrate_on
, 1)) {
1671 int start
= env
->dst_cpu
;
1673 /* Find alternative idle CPU. */
1674 for_each_cpu_wrap(cpu
, cpumask_of_node(env
->dst_nid
), start
) {
1675 if (cpu
== env
->best_cpu
|| !idle_cpu(cpu
) ||
1676 !cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
)) {
1681 rq
= cpu_rq(env
->dst_cpu
);
1682 if (!xchg(&rq
->numa_migrate_on
, 1))
1686 /* Failed to find an alternative idle CPU */
1692 * Clear previous best_cpu/rq numa-migrate flag, since task now
1693 * found a better CPU to move/swap.
1695 if (env
->best_cpu
!= -1 && env
->best_cpu
!= env
->dst_cpu
) {
1696 rq
= cpu_rq(env
->best_cpu
);
1697 WRITE_ONCE(rq
->numa_migrate_on
, 0);
1701 put_task_struct(env
->best_task
);
1706 env
->best_imp
= imp
;
1707 env
->best_cpu
= env
->dst_cpu
;
1710 static bool load_too_imbalanced(long src_load
, long dst_load
,
1711 struct task_numa_env
*env
)
1714 long orig_src_load
, orig_dst_load
;
1715 long src_capacity
, dst_capacity
;
1718 * The load is corrected for the CPU capacity available on each node.
1721 * ------------ vs ---------
1722 * src_capacity dst_capacity
1724 src_capacity
= env
->src_stats
.compute_capacity
;
1725 dst_capacity
= env
->dst_stats
.compute_capacity
;
1727 imb
= abs(dst_load
* src_capacity
- src_load
* dst_capacity
);
1729 orig_src_load
= env
->src_stats
.load
;
1730 orig_dst_load
= env
->dst_stats
.load
;
1732 old_imb
= abs(orig_dst_load
* src_capacity
- orig_src_load
* dst_capacity
);
1734 /* Would this change make things worse? */
1735 return (imb
> old_imb
);
1739 * Maximum NUMA importance can be 1998 (2*999);
1740 * SMALLIMP @ 30 would be close to 1998/64.
1741 * Used to deter task migration.
1746 * This checks if the overall compute and NUMA accesses of the system would
1747 * be improved if the source tasks was migrated to the target dst_cpu taking
1748 * into account that it might be best if task running on the dst_cpu should
1749 * be exchanged with the source task
1751 static bool task_numa_compare(struct task_numa_env
*env
,
1752 long taskimp
, long groupimp
, bool maymove
)
1754 struct numa_group
*cur_ng
, *p_ng
= deref_curr_numa_group(env
->p
);
1755 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1756 long imp
= p_ng
? groupimp
: taskimp
;
1757 struct task_struct
*cur
;
1758 long src_load
, dst_load
;
1759 int dist
= env
->dist
;
1762 bool stopsearch
= false;
1764 if (READ_ONCE(dst_rq
->numa_migrate_on
))
1768 cur
= rcu_dereference(dst_rq
->curr
);
1769 if (cur
&& ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
)))
1773 * Because we have preemption enabled we can get migrated around and
1774 * end try selecting ourselves (current == env->p) as a swap candidate.
1776 if (cur
== env
->p
) {
1782 if (maymove
&& moveimp
>= env
->best_imp
)
1788 /* Skip this swap candidate if cannot move to the source cpu. */
1789 if (!cpumask_test_cpu(env
->src_cpu
, cur
->cpus_ptr
))
1793 * Skip this swap candidate if it is not moving to its preferred
1794 * node and the best task is.
1796 if (env
->best_task
&&
1797 env
->best_task
->numa_preferred_nid
== env
->src_nid
&&
1798 cur
->numa_preferred_nid
!= env
->src_nid
) {
1803 * "imp" is the fault differential for the source task between the
1804 * source and destination node. Calculate the total differential for
1805 * the source task and potential destination task. The more negative
1806 * the value is, the more remote accesses that would be expected to
1807 * be incurred if the tasks were swapped.
1809 * If dst and source tasks are in the same NUMA group, or not
1810 * in any group then look only at task weights.
1812 cur_ng
= rcu_dereference(cur
->numa_group
);
1813 if (cur_ng
== p_ng
) {
1814 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1815 task_weight(cur
, env
->dst_nid
, dist
);
1817 * Add some hysteresis to prevent swapping the
1818 * tasks within a group over tiny differences.
1824 * Compare the group weights. If a task is all by itself
1825 * (not part of a group), use the task weight instead.
1828 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1829 group_weight(cur
, env
->dst_nid
, dist
);
1831 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1832 task_weight(cur
, env
->dst_nid
, dist
);
1835 /* Discourage picking a task already on its preferred node */
1836 if (cur
->numa_preferred_nid
== env
->dst_nid
)
1840 * Encourage picking a task that moves to its preferred node.
1841 * This potentially makes imp larger than it's maximum of
1842 * 1998 (see SMALLIMP and task_weight for why) but in this
1843 * case, it does not matter.
1845 if (cur
->numa_preferred_nid
== env
->src_nid
)
1848 if (maymove
&& moveimp
> imp
&& moveimp
> env
->best_imp
) {
1855 * Prefer swapping with a task moving to its preferred node over a
1858 if (env
->best_task
&& cur
->numa_preferred_nid
== env
->src_nid
&&
1859 env
->best_task
->numa_preferred_nid
!= env
->src_nid
) {
1864 * If the NUMA importance is less than SMALLIMP,
1865 * task migration might only result in ping pong
1866 * of tasks and also hurt performance due to cache
1869 if (imp
< SMALLIMP
|| imp
<= env
->best_imp
+ SMALLIMP
/ 2)
1873 * In the overloaded case, try and keep the load balanced.
1875 load
= task_h_load(env
->p
) - task_h_load(cur
);
1879 dst_load
= env
->dst_stats
.load
+ load
;
1880 src_load
= env
->src_stats
.load
- load
;
1882 if (load_too_imbalanced(src_load
, dst_load
, env
))
1886 /* Evaluate an idle CPU for a task numa move. */
1888 int cpu
= env
->dst_stats
.idle_cpu
;
1890 /* Nothing cached so current CPU went idle since the search. */
1895 * If the CPU is no longer truly idle and the previous best CPU
1896 * is, keep using it.
1898 if (!idle_cpu(cpu
) && env
->best_cpu
>= 0 &&
1899 idle_cpu(env
->best_cpu
)) {
1900 cpu
= env
->best_cpu
;
1906 task_numa_assign(env
, cur
, imp
);
1909 * If a move to idle is allowed because there is capacity or load
1910 * balance improves then stop the search. While a better swap
1911 * candidate may exist, a search is not free.
1913 if (maymove
&& !cur
&& env
->best_cpu
>= 0 && idle_cpu(env
->best_cpu
))
1917 * If a swap candidate must be identified and the current best task
1918 * moves its preferred node then stop the search.
1920 if (!maymove
&& env
->best_task
&&
1921 env
->best_task
->numa_preferred_nid
== env
->src_nid
) {
1930 static void task_numa_find_cpu(struct task_numa_env
*env
,
1931 long taskimp
, long groupimp
)
1933 bool maymove
= false;
1937 * If dst node has spare capacity, then check if there is an
1938 * imbalance that would be overruled by the load balancer.
1940 if (env
->dst_stats
.node_type
== node_has_spare
) {
1941 unsigned int imbalance
;
1942 int src_running
, dst_running
;
1945 * Would movement cause an imbalance? Note that if src has
1946 * more running tasks that the imbalance is ignored as the
1947 * move improves the imbalance from the perspective of the
1948 * CPU load balancer.
1950 src_running
= env
->src_stats
.nr_running
- 1;
1951 dst_running
= env
->dst_stats
.nr_running
+ 1;
1952 imbalance
= max(0, dst_running
- src_running
);
1953 imbalance
= adjust_numa_imbalance(imbalance
, dst_running
,
1954 env
->dst_stats
.weight
);
1956 /* Use idle CPU if there is no imbalance */
1959 if (env
->dst_stats
.idle_cpu
>= 0) {
1960 env
->dst_cpu
= env
->dst_stats
.idle_cpu
;
1961 task_numa_assign(env
, NULL
, 0);
1966 long src_load
, dst_load
, load
;
1968 * If the improvement from just moving env->p direction is better
1969 * than swapping tasks around, check if a move is possible.
1971 load
= task_h_load(env
->p
);
1972 dst_load
= env
->dst_stats
.load
+ load
;
1973 src_load
= env
->src_stats
.load
- load
;
1974 maymove
= !load_too_imbalanced(src_load
, dst_load
, env
);
1977 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1978 /* Skip this CPU if the source task cannot migrate */
1979 if (!cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
))
1983 if (task_numa_compare(env
, taskimp
, groupimp
, maymove
))
1988 static int task_numa_migrate(struct task_struct
*p
)
1990 struct task_numa_env env
= {
1993 .src_cpu
= task_cpu(p
),
1994 .src_nid
= task_node(p
),
1996 .imbalance_pct
= 112,
2002 unsigned long taskweight
, groupweight
;
2003 struct sched_domain
*sd
;
2004 long taskimp
, groupimp
;
2005 struct numa_group
*ng
;
2010 * Pick the lowest SD_NUMA domain, as that would have the smallest
2011 * imbalance and would be the first to start moving tasks about.
2013 * And we want to avoid any moving of tasks about, as that would create
2014 * random movement of tasks -- counter the numa conditions we're trying
2018 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
2020 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
2024 * Cpusets can break the scheduler domain tree into smaller
2025 * balance domains, some of which do not cross NUMA boundaries.
2026 * Tasks that are "trapped" in such domains cannot be migrated
2027 * elsewhere, so there is no point in (re)trying.
2029 if (unlikely(!sd
)) {
2030 sched_setnuma(p
, task_node(p
));
2034 env
.dst_nid
= p
->numa_preferred_nid
;
2035 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
2036 taskweight
= task_weight(p
, env
.src_nid
, dist
);
2037 groupweight
= group_weight(p
, env
.src_nid
, dist
);
2038 update_numa_stats(&env
, &env
.src_stats
, env
.src_nid
, false);
2039 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
2040 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
2041 update_numa_stats(&env
, &env
.dst_stats
, env
.dst_nid
, true);
2043 /* Try to find a spot on the preferred nid. */
2044 task_numa_find_cpu(&env
, taskimp
, groupimp
);
2047 * Look at other nodes in these cases:
2048 * - there is no space available on the preferred_nid
2049 * - the task is part of a numa_group that is interleaved across
2050 * multiple NUMA nodes; in order to better consolidate the group,
2051 * we need to check other locations.
2053 ng
= deref_curr_numa_group(p
);
2054 if (env
.best_cpu
== -1 || (ng
&& ng
->active_nodes
> 1)) {
2055 for_each_online_node(nid
) {
2056 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
2059 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
2060 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
2062 taskweight
= task_weight(p
, env
.src_nid
, dist
);
2063 groupweight
= group_weight(p
, env
.src_nid
, dist
);
2066 /* Only consider nodes where both task and groups benefit */
2067 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
2068 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
2069 if (taskimp
< 0 && groupimp
< 0)
2074 update_numa_stats(&env
, &env
.dst_stats
, env
.dst_nid
, true);
2075 task_numa_find_cpu(&env
, taskimp
, groupimp
);
2080 * If the task is part of a workload that spans multiple NUMA nodes,
2081 * and is migrating into one of the workload's active nodes, remember
2082 * this node as the task's preferred numa node, so the workload can
2084 * A task that migrated to a second choice node will be better off
2085 * trying for a better one later. Do not set the preferred node here.
2088 if (env
.best_cpu
== -1)
2091 nid
= cpu_to_node(env
.best_cpu
);
2093 if (nid
!= p
->numa_preferred_nid
)
2094 sched_setnuma(p
, nid
);
2097 /* No better CPU than the current one was found. */
2098 if (env
.best_cpu
== -1) {
2099 trace_sched_stick_numa(p
, env
.src_cpu
, NULL
, -1);
2103 best_rq
= cpu_rq(env
.best_cpu
);
2104 if (env
.best_task
== NULL
) {
2105 ret
= migrate_task_to(p
, env
.best_cpu
);
2106 WRITE_ONCE(best_rq
->numa_migrate_on
, 0);
2108 trace_sched_stick_numa(p
, env
.src_cpu
, NULL
, env
.best_cpu
);
2112 ret
= migrate_swap(p
, env
.best_task
, env
.best_cpu
, env
.src_cpu
);
2113 WRITE_ONCE(best_rq
->numa_migrate_on
, 0);
2116 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_task
, env
.best_cpu
);
2117 put_task_struct(env
.best_task
);
2121 /* Attempt to migrate a task to a CPU on the preferred node. */
2122 static void numa_migrate_preferred(struct task_struct
*p
)
2124 unsigned long interval
= HZ
;
2126 /* This task has no NUMA fault statistics yet */
2127 if (unlikely(p
->numa_preferred_nid
== NUMA_NO_NODE
|| !p
->numa_faults
))
2130 /* Periodically retry migrating the task to the preferred node */
2131 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
2132 p
->numa_migrate_retry
= jiffies
+ interval
;
2134 /* Success if task is already running on preferred CPU */
2135 if (task_node(p
) == p
->numa_preferred_nid
)
2138 /* Otherwise, try migrate to a CPU on the preferred node */
2139 task_numa_migrate(p
);
2143 * Find out how many nodes on the workload is actively running on. Do this by
2144 * tracking the nodes from which NUMA hinting faults are triggered. This can
2145 * be different from the set of nodes where the workload's memory is currently
2148 static void numa_group_count_active_nodes(struct numa_group
*numa_group
)
2150 unsigned long faults
, max_faults
= 0;
2151 int nid
, active_nodes
= 0;
2153 for_each_online_node(nid
) {
2154 faults
= group_faults_cpu(numa_group
, nid
);
2155 if (faults
> max_faults
)
2156 max_faults
= faults
;
2159 for_each_online_node(nid
) {
2160 faults
= group_faults_cpu(numa_group
, nid
);
2161 if (faults
* ACTIVE_NODE_FRACTION
> max_faults
)
2165 numa_group
->max_faults_cpu
= max_faults
;
2166 numa_group
->active_nodes
= active_nodes
;
2170 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2171 * increments. The more local the fault statistics are, the higher the scan
2172 * period will be for the next scan window. If local/(local+remote) ratio is
2173 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2174 * the scan period will decrease. Aim for 70% local accesses.
2176 #define NUMA_PERIOD_SLOTS 10
2177 #define NUMA_PERIOD_THRESHOLD 7
2180 * Increase the scan period (slow down scanning) if the majority of
2181 * our memory is already on our local node, or if the majority of
2182 * the page accesses are shared with other processes.
2183 * Otherwise, decrease the scan period.
2185 static void update_task_scan_period(struct task_struct
*p
,
2186 unsigned long shared
, unsigned long private)
2188 unsigned int period_slot
;
2189 int lr_ratio
, ps_ratio
;
2192 unsigned long remote
= p
->numa_faults_locality
[0];
2193 unsigned long local
= p
->numa_faults_locality
[1];
2196 * If there were no record hinting faults then either the task is
2197 * completely idle or all activity is areas that are not of interest
2198 * to automatic numa balancing. Related to that, if there were failed
2199 * migration then it implies we are migrating too quickly or the local
2200 * node is overloaded. In either case, scan slower
2202 if (local
+ shared
== 0 || p
->numa_faults_locality
[2]) {
2203 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
2204 p
->numa_scan_period
<< 1);
2206 p
->mm
->numa_next_scan
= jiffies
+
2207 msecs_to_jiffies(p
->numa_scan_period
);
2213 * Prepare to scale scan period relative to the current period.
2214 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2215 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2216 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2218 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
2219 lr_ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
2220 ps_ratio
= (private * NUMA_PERIOD_SLOTS
) / (private + shared
);
2222 if (ps_ratio
>= NUMA_PERIOD_THRESHOLD
) {
2224 * Most memory accesses are local. There is no need to
2225 * do fast NUMA scanning, since memory is already local.
2227 int slot
= ps_ratio
- NUMA_PERIOD_THRESHOLD
;
2230 diff
= slot
* period_slot
;
2231 } else if (lr_ratio
>= NUMA_PERIOD_THRESHOLD
) {
2233 * Most memory accesses are shared with other tasks.
2234 * There is no point in continuing fast NUMA scanning,
2235 * since other tasks may just move the memory elsewhere.
2237 int slot
= lr_ratio
- NUMA_PERIOD_THRESHOLD
;
2240 diff
= slot
* period_slot
;
2243 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2244 * yet they are not on the local NUMA node. Speed up
2245 * NUMA scanning to get the memory moved over.
2247 int ratio
= max(lr_ratio
, ps_ratio
);
2248 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
2251 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
2252 task_scan_min(p
), task_scan_max(p
));
2253 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2257 * Get the fraction of time the task has been running since the last
2258 * NUMA placement cycle. The scheduler keeps similar statistics, but
2259 * decays those on a 32ms period, which is orders of magnitude off
2260 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2261 * stats only if the task is so new there are no NUMA statistics yet.
2263 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
2265 u64 runtime
, delta
, now
;
2266 /* Use the start of this time slice to avoid calculations. */
2267 now
= p
->se
.exec_start
;
2268 runtime
= p
->se
.sum_exec_runtime
;
2270 if (p
->last_task_numa_placement
) {
2271 delta
= runtime
- p
->last_sum_exec_runtime
;
2272 *period
= now
- p
->last_task_numa_placement
;
2274 /* Avoid time going backwards, prevent potential divide error: */
2275 if (unlikely((s64
)*period
< 0))
2278 delta
= p
->se
.avg
.load_sum
;
2279 *period
= LOAD_AVG_MAX
;
2282 p
->last_sum_exec_runtime
= runtime
;
2283 p
->last_task_numa_placement
= now
;
2289 * Determine the preferred nid for a task in a numa_group. This needs to
2290 * be done in a way that produces consistent results with group_weight,
2291 * otherwise workloads might not converge.
2293 static int preferred_group_nid(struct task_struct
*p
, int nid
)
2298 /* Direct connections between all NUMA nodes. */
2299 if (sched_numa_topology_type
== NUMA_DIRECT
)
2303 * On a system with glueless mesh NUMA topology, group_weight
2304 * scores nodes according to the number of NUMA hinting faults on
2305 * both the node itself, and on nearby nodes.
2307 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
2308 unsigned long score
, max_score
= 0;
2309 int node
, max_node
= nid
;
2311 dist
= sched_max_numa_distance
;
2313 for_each_online_node(node
) {
2314 score
= group_weight(p
, node
, dist
);
2315 if (score
> max_score
) {
2324 * Finding the preferred nid in a system with NUMA backplane
2325 * interconnect topology is more involved. The goal is to locate
2326 * tasks from numa_groups near each other in the system, and
2327 * untangle workloads from different sides of the system. This requires
2328 * searching down the hierarchy of node groups, recursively searching
2329 * inside the highest scoring group of nodes. The nodemask tricks
2330 * keep the complexity of the search down.
2332 nodes
= node_online_map
;
2333 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
2334 unsigned long max_faults
= 0;
2335 nodemask_t max_group
= NODE_MASK_NONE
;
2338 /* Are there nodes at this distance from each other? */
2339 if (!find_numa_distance(dist
))
2342 for_each_node_mask(a
, nodes
) {
2343 unsigned long faults
= 0;
2344 nodemask_t this_group
;
2345 nodes_clear(this_group
);
2347 /* Sum group's NUMA faults; includes a==b case. */
2348 for_each_node_mask(b
, nodes
) {
2349 if (node_distance(a
, b
) < dist
) {
2350 faults
+= group_faults(p
, b
);
2351 node_set(b
, this_group
);
2352 node_clear(b
, nodes
);
2356 /* Remember the top group. */
2357 if (faults
> max_faults
) {
2358 max_faults
= faults
;
2359 max_group
= this_group
;
2361 * subtle: at the smallest distance there is
2362 * just one node left in each "group", the
2363 * winner is the preferred nid.
2368 /* Next round, evaluate the nodes within max_group. */
2376 static void task_numa_placement(struct task_struct
*p
)
2378 int seq
, nid
, max_nid
= NUMA_NO_NODE
;
2379 unsigned long max_faults
= 0;
2380 unsigned long fault_types
[2] = { 0, 0 };
2381 unsigned long total_faults
;
2382 u64 runtime
, period
;
2383 spinlock_t
*group_lock
= NULL
;
2384 struct numa_group
*ng
;
2387 * The p->mm->numa_scan_seq field gets updated without
2388 * exclusive access. Use READ_ONCE() here to ensure
2389 * that the field is read in a single access:
2391 seq
= READ_ONCE(p
->mm
->numa_scan_seq
);
2392 if (p
->numa_scan_seq
== seq
)
2394 p
->numa_scan_seq
= seq
;
2395 p
->numa_scan_period_max
= task_scan_max(p
);
2397 total_faults
= p
->numa_faults_locality
[0] +
2398 p
->numa_faults_locality
[1];
2399 runtime
= numa_get_avg_runtime(p
, &period
);
2401 /* If the task is part of a group prevent parallel updates to group stats */
2402 ng
= deref_curr_numa_group(p
);
2404 group_lock
= &ng
->lock
;
2405 spin_lock_irq(group_lock
);
2408 /* Find the node with the highest number of faults */
2409 for_each_online_node(nid
) {
2410 /* Keep track of the offsets in numa_faults array */
2411 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
2412 unsigned long faults
= 0, group_faults
= 0;
2415 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
2416 long diff
, f_diff
, f_weight
;
2418 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
2419 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
2420 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
2421 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
2423 /* Decay existing window, copy faults since last scan */
2424 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
2425 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
2426 p
->numa_faults
[membuf_idx
] = 0;
2429 * Normalize the faults_from, so all tasks in a group
2430 * count according to CPU use, instead of by the raw
2431 * number of faults. Tasks with little runtime have
2432 * little over-all impact on throughput, and thus their
2433 * faults are less important.
2435 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
2436 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
2438 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
2439 p
->numa_faults
[cpubuf_idx
] = 0;
2441 p
->numa_faults
[mem_idx
] += diff
;
2442 p
->numa_faults
[cpu_idx
] += f_diff
;
2443 faults
+= p
->numa_faults
[mem_idx
];
2444 p
->total_numa_faults
+= diff
;
2447 * safe because we can only change our own group
2449 * mem_idx represents the offset for a given
2450 * nid and priv in a specific region because it
2451 * is at the beginning of the numa_faults array.
2453 ng
->faults
[mem_idx
] += diff
;
2454 ng
->faults_cpu
[mem_idx
] += f_diff
;
2455 ng
->total_faults
+= diff
;
2456 group_faults
+= ng
->faults
[mem_idx
];
2461 if (faults
> max_faults
) {
2462 max_faults
= faults
;
2465 } else if (group_faults
> max_faults
) {
2466 max_faults
= group_faults
;
2472 numa_group_count_active_nodes(ng
);
2473 spin_unlock_irq(group_lock
);
2474 max_nid
= preferred_group_nid(p
, max_nid
);
2478 /* Set the new preferred node */
2479 if (max_nid
!= p
->numa_preferred_nid
)
2480 sched_setnuma(p
, max_nid
);
2483 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
2486 static inline int get_numa_group(struct numa_group
*grp
)
2488 return refcount_inc_not_zero(&grp
->refcount
);
2491 static inline void put_numa_group(struct numa_group
*grp
)
2493 if (refcount_dec_and_test(&grp
->refcount
))
2494 kfree_rcu(grp
, rcu
);
2497 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
2500 struct numa_group
*grp
, *my_grp
;
2501 struct task_struct
*tsk
;
2503 int cpu
= cpupid_to_cpu(cpupid
);
2506 if (unlikely(!deref_curr_numa_group(p
))) {
2507 unsigned int size
= sizeof(struct numa_group
) +
2508 4*nr_node_ids
*sizeof(unsigned long);
2510 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
2514 refcount_set(&grp
->refcount
, 1);
2515 grp
->active_nodes
= 1;
2516 grp
->max_faults_cpu
= 0;
2517 spin_lock_init(&grp
->lock
);
2519 /* Second half of the array tracks nids where faults happen */
2520 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
2523 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2524 grp
->faults
[i
] = p
->numa_faults
[i
];
2526 grp
->total_faults
= p
->total_numa_faults
;
2529 rcu_assign_pointer(p
->numa_group
, grp
);
2533 tsk
= READ_ONCE(cpu_rq(cpu
)->curr
);
2535 if (!cpupid_match_pid(tsk
, cpupid
))
2538 grp
= rcu_dereference(tsk
->numa_group
);
2542 my_grp
= deref_curr_numa_group(p
);
2547 * Only join the other group if its bigger; if we're the bigger group,
2548 * the other task will join us.
2550 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
2554 * Tie-break on the grp address.
2556 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
2559 /* Always join threads in the same process. */
2560 if (tsk
->mm
== current
->mm
)
2563 /* Simple filter to avoid false positives due to PID collisions */
2564 if (flags
& TNF_SHARED
)
2567 /* Update priv based on whether false sharing was detected */
2570 if (join
&& !get_numa_group(grp
))
2578 BUG_ON(irqs_disabled());
2579 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
2581 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
2582 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
2583 grp
->faults
[i
] += p
->numa_faults
[i
];
2585 my_grp
->total_faults
-= p
->total_numa_faults
;
2586 grp
->total_faults
+= p
->total_numa_faults
;
2591 spin_unlock(&my_grp
->lock
);
2592 spin_unlock_irq(&grp
->lock
);
2594 rcu_assign_pointer(p
->numa_group
, grp
);
2596 put_numa_group(my_grp
);
2605 * Get rid of NUMA staticstics associated with a task (either current or dead).
2606 * If @final is set, the task is dead and has reached refcount zero, so we can
2607 * safely free all relevant data structures. Otherwise, there might be
2608 * concurrent reads from places like load balancing and procfs, and we should
2609 * reset the data back to default state without freeing ->numa_faults.
2611 void task_numa_free(struct task_struct
*p
, bool final
)
2613 /* safe: p either is current or is being freed by current */
2614 struct numa_group
*grp
= rcu_dereference_raw(p
->numa_group
);
2615 unsigned long *numa_faults
= p
->numa_faults
;
2616 unsigned long flags
;
2623 spin_lock_irqsave(&grp
->lock
, flags
);
2624 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2625 grp
->faults
[i
] -= p
->numa_faults
[i
];
2626 grp
->total_faults
-= p
->total_numa_faults
;
2629 spin_unlock_irqrestore(&grp
->lock
, flags
);
2630 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2631 put_numa_group(grp
);
2635 p
->numa_faults
= NULL
;
2638 p
->total_numa_faults
= 0;
2639 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2645 * Got a PROT_NONE fault for a page on @node.
2647 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2649 struct task_struct
*p
= current
;
2650 bool migrated
= flags
& TNF_MIGRATED
;
2651 int cpu_node
= task_node(current
);
2652 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2653 struct numa_group
*ng
;
2656 if (!static_branch_likely(&sched_numa_balancing
))
2659 /* for example, ksmd faulting in a user's mm */
2663 /* Allocate buffer to track faults on a per-node basis */
2664 if (unlikely(!p
->numa_faults
)) {
2665 int size
= sizeof(*p
->numa_faults
) *
2666 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2668 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2669 if (!p
->numa_faults
)
2672 p
->total_numa_faults
= 0;
2673 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2677 * First accesses are treated as private, otherwise consider accesses
2678 * to be private if the accessing pid has not changed
2680 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2683 priv
= cpupid_match_pid(p
, last_cpupid
);
2684 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2685 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2689 * If a workload spans multiple NUMA nodes, a shared fault that
2690 * occurs wholly within the set of nodes that the workload is
2691 * actively using should be counted as local. This allows the
2692 * scan rate to slow down when a workload has settled down.
2694 ng
= deref_curr_numa_group(p
);
2695 if (!priv
&& !local
&& ng
&& ng
->active_nodes
> 1 &&
2696 numa_is_active_node(cpu_node
, ng
) &&
2697 numa_is_active_node(mem_node
, ng
))
2701 * Retry to migrate task to preferred node periodically, in case it
2702 * previously failed, or the scheduler moved us.
2704 if (time_after(jiffies
, p
->numa_migrate_retry
)) {
2705 task_numa_placement(p
);
2706 numa_migrate_preferred(p
);
2710 p
->numa_pages_migrated
+= pages
;
2711 if (flags
& TNF_MIGRATE_FAIL
)
2712 p
->numa_faults_locality
[2] += pages
;
2714 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2715 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2716 p
->numa_faults_locality
[local
] += pages
;
2719 static void reset_ptenuma_scan(struct task_struct
*p
)
2722 * We only did a read acquisition of the mmap sem, so
2723 * p->mm->numa_scan_seq is written to without exclusive access
2724 * and the update is not guaranteed to be atomic. That's not
2725 * much of an issue though, since this is just used for
2726 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2727 * expensive, to avoid any form of compiler optimizations:
2729 WRITE_ONCE(p
->mm
->numa_scan_seq
, READ_ONCE(p
->mm
->numa_scan_seq
) + 1);
2730 p
->mm
->numa_scan_offset
= 0;
2734 * The expensive part of numa migration is done from task_work context.
2735 * Triggered from task_tick_numa().
2737 static void task_numa_work(struct callback_head
*work
)
2739 unsigned long migrate
, next_scan
, now
= jiffies
;
2740 struct task_struct
*p
= current
;
2741 struct mm_struct
*mm
= p
->mm
;
2742 u64 runtime
= p
->se
.sum_exec_runtime
;
2743 struct vm_area_struct
*vma
;
2744 unsigned long start
, end
;
2745 unsigned long nr_pte_updates
= 0;
2746 long pages
, virtpages
;
2748 SCHED_WARN_ON(p
!= container_of(work
, struct task_struct
, numa_work
));
2752 * Who cares about NUMA placement when they're dying.
2754 * NOTE: make sure not to dereference p->mm before this check,
2755 * exit_task_work() happens _after_ exit_mm() so we could be called
2756 * without p->mm even though we still had it when we enqueued this
2759 if (p
->flags
& PF_EXITING
)
2762 if (!mm
->numa_next_scan
) {
2763 mm
->numa_next_scan
= now
+
2764 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2768 * Enforce maximal scan/migration frequency..
2770 migrate
= mm
->numa_next_scan
;
2771 if (time_before(now
, migrate
))
2774 if (p
->numa_scan_period
== 0) {
2775 p
->numa_scan_period_max
= task_scan_max(p
);
2776 p
->numa_scan_period
= task_scan_start(p
);
2779 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2780 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2784 * Delay this task enough that another task of this mm will likely win
2785 * the next time around.
2787 p
->node_stamp
+= 2 * TICK_NSEC
;
2789 start
= mm
->numa_scan_offset
;
2790 pages
= sysctl_numa_balancing_scan_size
;
2791 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2792 virtpages
= pages
* 8; /* Scan up to this much virtual space */
2797 if (!mmap_read_trylock(mm
))
2799 vma
= find_vma(mm
, start
);
2801 reset_ptenuma_scan(p
);
2805 for (; vma
; vma
= vma
->vm_next
) {
2806 if (!vma_migratable(vma
) || !vma_policy_mof(vma
) ||
2807 is_vm_hugetlb_page(vma
) || (vma
->vm_flags
& VM_MIXEDMAP
)) {
2812 * Shared library pages mapped by multiple processes are not
2813 * migrated as it is expected they are cache replicated. Avoid
2814 * hinting faults in read-only file-backed mappings or the vdso
2815 * as migrating the pages will be of marginal benefit.
2818 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2822 * Skip inaccessible VMAs to avoid any confusion between
2823 * PROT_NONE and NUMA hinting ptes
2825 if (!vma_is_accessible(vma
))
2829 start
= max(start
, vma
->vm_start
);
2830 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2831 end
= min(end
, vma
->vm_end
);
2832 nr_pte_updates
= change_prot_numa(vma
, start
, end
);
2835 * Try to scan sysctl_numa_balancing_size worth of
2836 * hpages that have at least one present PTE that
2837 * is not already pte-numa. If the VMA contains
2838 * areas that are unused or already full of prot_numa
2839 * PTEs, scan up to virtpages, to skip through those
2843 pages
-= (end
- start
) >> PAGE_SHIFT
;
2844 virtpages
-= (end
- start
) >> PAGE_SHIFT
;
2847 if (pages
<= 0 || virtpages
<= 0)
2851 } while (end
!= vma
->vm_end
);
2856 * It is possible to reach the end of the VMA list but the last few
2857 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2858 * would find the !migratable VMA on the next scan but not reset the
2859 * scanner to the start so check it now.
2862 mm
->numa_scan_offset
= start
;
2864 reset_ptenuma_scan(p
);
2865 mmap_read_unlock(mm
);
2868 * Make sure tasks use at least 32x as much time to run other code
2869 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2870 * Usually update_task_scan_period slows down scanning enough; on an
2871 * overloaded system we need to limit overhead on a per task basis.
2873 if (unlikely(p
->se
.sum_exec_runtime
!= runtime
)) {
2874 u64 diff
= p
->se
.sum_exec_runtime
- runtime
;
2875 p
->node_stamp
+= 32 * diff
;
2879 void init_numa_balancing(unsigned long clone_flags
, struct task_struct
*p
)
2882 struct mm_struct
*mm
= p
->mm
;
2885 mm_users
= atomic_read(&mm
->mm_users
);
2886 if (mm_users
== 1) {
2887 mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2888 mm
->numa_scan_seq
= 0;
2892 p
->numa_scan_seq
= mm
? mm
->numa_scan_seq
: 0;
2893 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2894 /* Protect against double add, see task_tick_numa and task_numa_work */
2895 p
->numa_work
.next
= &p
->numa_work
;
2896 p
->numa_faults
= NULL
;
2897 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2898 p
->last_task_numa_placement
= 0;
2899 p
->last_sum_exec_runtime
= 0;
2901 init_task_work(&p
->numa_work
, task_numa_work
);
2903 /* New address space, reset the preferred nid */
2904 if (!(clone_flags
& CLONE_VM
)) {
2905 p
->numa_preferred_nid
= NUMA_NO_NODE
;
2910 * New thread, keep existing numa_preferred_nid which should be copied
2911 * already by arch_dup_task_struct but stagger when scans start.
2916 delay
= min_t(unsigned int, task_scan_max(current
),
2917 current
->numa_scan_period
* mm_users
* NSEC_PER_MSEC
);
2918 delay
+= 2 * TICK_NSEC
;
2919 p
->node_stamp
= delay
;
2924 * Drive the periodic memory faults..
2926 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2928 struct callback_head
*work
= &curr
->numa_work
;
2932 * We don't care about NUMA placement if we don't have memory.
2934 if ((curr
->flags
& (PF_EXITING
| PF_KTHREAD
)) || work
->next
!= work
)
2938 * Using runtime rather than walltime has the dual advantage that
2939 * we (mostly) drive the selection from busy threads and that the
2940 * task needs to have done some actual work before we bother with
2943 now
= curr
->se
.sum_exec_runtime
;
2944 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2946 if (now
> curr
->node_stamp
+ period
) {
2947 if (!curr
->node_stamp
)
2948 curr
->numa_scan_period
= task_scan_start(curr
);
2949 curr
->node_stamp
+= period
;
2951 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
))
2952 task_work_add(curr
, work
, TWA_RESUME
);
2956 static void update_scan_period(struct task_struct
*p
, int new_cpu
)
2958 int src_nid
= cpu_to_node(task_cpu(p
));
2959 int dst_nid
= cpu_to_node(new_cpu
);
2961 if (!static_branch_likely(&sched_numa_balancing
))
2964 if (!p
->mm
|| !p
->numa_faults
|| (p
->flags
& PF_EXITING
))
2967 if (src_nid
== dst_nid
)
2971 * Allow resets if faults have been trapped before one scan
2972 * has completed. This is most likely due to a new task that
2973 * is pulled cross-node due to wakeups or load balancing.
2975 if (p
->numa_scan_seq
) {
2977 * Avoid scan adjustments if moving to the preferred
2978 * node or if the task was not previously running on
2979 * the preferred node.
2981 if (dst_nid
== p
->numa_preferred_nid
||
2982 (p
->numa_preferred_nid
!= NUMA_NO_NODE
&&
2983 src_nid
!= p
->numa_preferred_nid
))
2987 p
->numa_scan_period
= task_scan_start(p
);
2991 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2995 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2999 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
3003 static inline void update_scan_period(struct task_struct
*p
, int new_cpu
)
3007 #endif /* CONFIG_NUMA_BALANCING */
3010 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3012 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
3014 if (entity_is_task(se
)) {
3015 struct rq
*rq
= rq_of(cfs_rq
);
3017 account_numa_enqueue(rq
, task_of(se
));
3018 list_add(&se
->group_node
, &rq
->cfs_tasks
);
3021 cfs_rq
->nr_running
++;
3025 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3027 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
3029 if (entity_is_task(se
)) {
3030 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
3031 list_del_init(&se
->group_node
);
3034 cfs_rq
->nr_running
--;
3038 * Signed add and clamp on underflow.
3040 * Explicitly do a load-store to ensure the intermediate value never hits
3041 * memory. This allows lockless observations without ever seeing the negative
3044 #define add_positive(_ptr, _val) do { \
3045 typeof(_ptr) ptr = (_ptr); \
3046 typeof(_val) val = (_val); \
3047 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3051 if (val < 0 && res > var) \
3054 WRITE_ONCE(*ptr, res); \
3058 * Unsigned subtract and clamp on underflow.
3060 * Explicitly do a load-store to ensure the intermediate value never hits
3061 * memory. This allows lockless observations without ever seeing the negative
3064 #define sub_positive(_ptr, _val) do { \
3065 typeof(_ptr) ptr = (_ptr); \
3066 typeof(*ptr) val = (_val); \
3067 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3071 WRITE_ONCE(*ptr, res); \
3075 * Remove and clamp on negative, from a local variable.
3077 * A variant of sub_positive(), which does not use explicit load-store
3078 * and is thus optimized for local variable updates.
3080 #define lsub_positive(_ptr, _val) do { \
3081 typeof(_ptr) ptr = (_ptr); \
3082 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3087 enqueue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3089 cfs_rq
->avg
.load_avg
+= se
->avg
.load_avg
;
3090 cfs_rq
->avg
.load_sum
+= se_weight(se
) * se
->avg
.load_sum
;
3094 dequeue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3096 sub_positive(&cfs_rq
->avg
.load_avg
, se
->avg
.load_avg
);
3097 sub_positive(&cfs_rq
->avg
.load_sum
, se_weight(se
) * se
->avg
.load_sum
);
3101 enqueue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
3103 dequeue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
3106 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
3107 unsigned long weight
)
3110 /* commit outstanding execution time */
3111 if (cfs_rq
->curr
== se
)
3112 update_curr(cfs_rq
);
3113 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
3115 dequeue_load_avg(cfs_rq
, se
);
3117 update_load_set(&se
->load
, weight
);
3121 u32 divider
= get_pelt_divider(&se
->avg
);
3123 se
->avg
.load_avg
= div_u64(se_weight(se
) * se
->avg
.load_sum
, divider
);
3127 enqueue_load_avg(cfs_rq
, se
);
3129 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
3133 void reweight_task(struct task_struct
*p
, int prio
)
3135 struct sched_entity
*se
= &p
->se
;
3136 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3137 struct load_weight
*load
= &se
->load
;
3138 unsigned long weight
= scale_load(sched_prio_to_weight
[prio
]);
3140 reweight_entity(cfs_rq
, se
, weight
);
3141 load
->inv_weight
= sched_prio_to_wmult
[prio
];
3144 #ifdef CONFIG_FAIR_GROUP_SCHED
3147 * All this does is approximate the hierarchical proportion which includes that
3148 * global sum we all love to hate.
3150 * That is, the weight of a group entity, is the proportional share of the
3151 * group weight based on the group runqueue weights. That is:
3153 * tg->weight * grq->load.weight
3154 * ge->load.weight = ----------------------------- (1)
3155 * \Sum grq->load.weight
3157 * Now, because computing that sum is prohibitively expensive to compute (been
3158 * there, done that) we approximate it with this average stuff. The average
3159 * moves slower and therefore the approximation is cheaper and more stable.
3161 * So instead of the above, we substitute:
3163 * grq->load.weight -> grq->avg.load_avg (2)
3165 * which yields the following:
3167 * tg->weight * grq->avg.load_avg
3168 * ge->load.weight = ------------------------------ (3)
3171 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3173 * That is shares_avg, and it is right (given the approximation (2)).
3175 * The problem with it is that because the average is slow -- it was designed
3176 * to be exactly that of course -- this leads to transients in boundary
3177 * conditions. In specific, the case where the group was idle and we start the
3178 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3179 * yielding bad latency etc..
3181 * Now, in that special case (1) reduces to:
3183 * tg->weight * grq->load.weight
3184 * ge->load.weight = ----------------------------- = tg->weight (4)
3187 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3189 * So what we do is modify our approximation (3) to approach (4) in the (near)
3194 * tg->weight * grq->load.weight
3195 * --------------------------------------------------- (5)
3196 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3198 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3199 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3202 * tg->weight * grq->load.weight
3203 * ge->load.weight = ----------------------------- (6)
3208 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3209 * max(grq->load.weight, grq->avg.load_avg)
3211 * And that is shares_weight and is icky. In the (near) UP case it approaches
3212 * (4) while in the normal case it approaches (3). It consistently
3213 * overestimates the ge->load.weight and therefore:
3215 * \Sum ge->load.weight >= tg->weight
3219 static long calc_group_shares(struct cfs_rq
*cfs_rq
)
3221 long tg_weight
, tg_shares
, load
, shares
;
3222 struct task_group
*tg
= cfs_rq
->tg
;
3224 tg_shares
= READ_ONCE(tg
->shares
);
3226 load
= max(scale_load_down(cfs_rq
->load
.weight
), cfs_rq
->avg
.load_avg
);
3228 tg_weight
= atomic_long_read(&tg
->load_avg
);
3230 /* Ensure tg_weight >= load */
3231 tg_weight
-= cfs_rq
->tg_load_avg_contrib
;
3234 shares
= (tg_shares
* load
);
3236 shares
/= tg_weight
;
3239 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3240 * of a group with small tg->shares value. It is a floor value which is
3241 * assigned as a minimum load.weight to the sched_entity representing
3242 * the group on a CPU.
3244 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3245 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3246 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3247 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3250 return clamp_t(long, shares
, MIN_SHARES
, tg_shares
);
3252 #endif /* CONFIG_SMP */
3254 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
3257 * Recomputes the group entity based on the current state of its group
3260 static void update_cfs_group(struct sched_entity
*se
)
3262 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3268 if (throttled_hierarchy(gcfs_rq
))
3272 shares
= READ_ONCE(gcfs_rq
->tg
->shares
);
3274 if (likely(se
->load
.weight
== shares
))
3277 shares
= calc_group_shares(gcfs_rq
);
3280 reweight_entity(cfs_rq_of(se
), se
, shares
);
3283 #else /* CONFIG_FAIR_GROUP_SCHED */
3284 static inline void update_cfs_group(struct sched_entity
*se
)
3287 #endif /* CONFIG_FAIR_GROUP_SCHED */
3289 static inline void cfs_rq_util_change(struct cfs_rq
*cfs_rq
, int flags
)
3291 struct rq
*rq
= rq_of(cfs_rq
);
3293 if (&rq
->cfs
== cfs_rq
) {
3295 * There are a few boundary cases this might miss but it should
3296 * get called often enough that that should (hopefully) not be
3299 * It will not get called when we go idle, because the idle
3300 * thread is a different class (!fair), nor will the utilization
3301 * number include things like RT tasks.
3303 * As is, the util number is not freq-invariant (we'd have to
3304 * implement arch_scale_freq_capacity() for that).
3308 cpufreq_update_util(rq
, flags
);
3313 #ifdef CONFIG_FAIR_GROUP_SCHED
3315 * update_tg_load_avg - update the tg's load avg
3316 * @cfs_rq: the cfs_rq whose avg changed
3318 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3319 * However, because tg->load_avg is a global value there are performance
3322 * In order to avoid having to look at the other cfs_rq's, we use a
3323 * differential update where we store the last value we propagated. This in
3324 * turn allows skipping updates if the differential is 'small'.
3326 * Updating tg's load_avg is necessary before update_cfs_share().
3328 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
)
3330 long delta
= cfs_rq
->avg
.load_avg
- cfs_rq
->tg_load_avg_contrib
;
3333 * No need to update load_avg for root_task_group as it is not used.
3335 if (cfs_rq
->tg
== &root_task_group
)
3338 if (abs(delta
) > cfs_rq
->tg_load_avg_contrib
/ 64) {
3339 atomic_long_add(delta
, &cfs_rq
->tg
->load_avg
);
3340 cfs_rq
->tg_load_avg_contrib
= cfs_rq
->avg
.load_avg
;
3345 * Called within set_task_rq() right before setting a task's CPU. The
3346 * caller only guarantees p->pi_lock is held; no other assumptions,
3347 * including the state of rq->lock, should be made.
3349 void set_task_rq_fair(struct sched_entity
*se
,
3350 struct cfs_rq
*prev
, struct cfs_rq
*next
)
3352 u64 p_last_update_time
;
3353 u64 n_last_update_time
;
3355 if (!sched_feat(ATTACH_AGE_LOAD
))
3359 * We are supposed to update the task to "current" time, then its up to
3360 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3361 * getting what current time is, so simply throw away the out-of-date
3362 * time. This will result in the wakee task is less decayed, but giving
3363 * the wakee more load sounds not bad.
3365 if (!(se
->avg
.last_update_time
&& prev
))
3368 #ifndef CONFIG_64BIT
3370 u64 p_last_update_time_copy
;
3371 u64 n_last_update_time_copy
;
3374 p_last_update_time_copy
= prev
->load_last_update_time_copy
;
3375 n_last_update_time_copy
= next
->load_last_update_time_copy
;
3379 p_last_update_time
= prev
->avg
.last_update_time
;
3380 n_last_update_time
= next
->avg
.last_update_time
;
3382 } while (p_last_update_time
!= p_last_update_time_copy
||
3383 n_last_update_time
!= n_last_update_time_copy
);
3386 p_last_update_time
= prev
->avg
.last_update_time
;
3387 n_last_update_time
= next
->avg
.last_update_time
;
3389 __update_load_avg_blocked_se(p_last_update_time
, se
);
3390 se
->avg
.last_update_time
= n_last_update_time
;
3395 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3396 * propagate its contribution. The key to this propagation is the invariant
3397 * that for each group:
3399 * ge->avg == grq->avg (1)
3401 * _IFF_ we look at the pure running and runnable sums. Because they
3402 * represent the very same entity, just at different points in the hierarchy.
3404 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3405 * and simply copies the running/runnable sum over (but still wrong, because
3406 * the group entity and group rq do not have their PELT windows aligned).
3408 * However, update_tg_cfs_load() is more complex. So we have:
3410 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3412 * And since, like util, the runnable part should be directly transferable,
3413 * the following would _appear_ to be the straight forward approach:
3415 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3417 * And per (1) we have:
3419 * ge->avg.runnable_avg == grq->avg.runnable_avg
3423 * ge->load.weight * grq->avg.load_avg
3424 * ge->avg.load_avg = ----------------------------------- (4)
3427 * Except that is wrong!
3429 * Because while for entities historical weight is not important and we
3430 * really only care about our future and therefore can consider a pure
3431 * runnable sum, runqueues can NOT do this.
3433 * We specifically want runqueues to have a load_avg that includes
3434 * historical weights. Those represent the blocked load, the load we expect
3435 * to (shortly) return to us. This only works by keeping the weights as
3436 * integral part of the sum. We therefore cannot decompose as per (3).
3438 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3439 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3440 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3441 * runnable section of these tasks overlap (or not). If they were to perfectly
3442 * align the rq as a whole would be runnable 2/3 of the time. If however we
3443 * always have at least 1 runnable task, the rq as a whole is always runnable.
3445 * So we'll have to approximate.. :/
3447 * Given the constraint:
3449 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3451 * We can construct a rule that adds runnable to a rq by assuming minimal
3454 * On removal, we'll assume each task is equally runnable; which yields:
3456 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3458 * XXX: only do this for the part of runnable > running ?
3463 update_tg_cfs_util(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3465 long delta
= gcfs_rq
->avg
.util_avg
- se
->avg
.util_avg
;
3468 /* Nothing to update */
3473 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3474 * See ___update_load_avg() for details.
3476 divider
= get_pelt_divider(&cfs_rq
->avg
);
3478 /* Set new sched_entity's utilization */
3479 se
->avg
.util_avg
= gcfs_rq
->avg
.util_avg
;
3480 se
->avg
.util_sum
= se
->avg
.util_avg
* divider
;
3482 /* Update parent cfs_rq utilization */
3483 add_positive(&cfs_rq
->avg
.util_avg
, delta
);
3484 cfs_rq
->avg
.util_sum
= cfs_rq
->avg
.util_avg
* divider
;
3488 update_tg_cfs_runnable(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3490 long delta
= gcfs_rq
->avg
.runnable_avg
- se
->avg
.runnable_avg
;
3493 /* Nothing to update */
3498 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3499 * See ___update_load_avg() for details.
3501 divider
= get_pelt_divider(&cfs_rq
->avg
);
3503 /* Set new sched_entity's runnable */
3504 se
->avg
.runnable_avg
= gcfs_rq
->avg
.runnable_avg
;
3505 se
->avg
.runnable_sum
= se
->avg
.runnable_avg
* divider
;
3507 /* Update parent cfs_rq runnable */
3508 add_positive(&cfs_rq
->avg
.runnable_avg
, delta
);
3509 cfs_rq
->avg
.runnable_sum
= cfs_rq
->avg
.runnable_avg
* divider
;
3513 update_tg_cfs_load(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3515 long delta_avg
, running_sum
, runnable_sum
= gcfs_rq
->prop_runnable_sum
;
3516 unsigned long load_avg
;
3524 gcfs_rq
->prop_runnable_sum
= 0;
3527 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3528 * See ___update_load_avg() for details.
3530 divider
= get_pelt_divider(&cfs_rq
->avg
);
3532 if (runnable_sum
>= 0) {
3534 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3535 * the CPU is saturated running == runnable.
3537 runnable_sum
+= se
->avg
.load_sum
;
3538 runnable_sum
= min_t(long, runnable_sum
, divider
);
3541 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3542 * assuming all tasks are equally runnable.
3544 if (scale_load_down(gcfs_rq
->load
.weight
)) {
3545 load_sum
= div_s64(gcfs_rq
->avg
.load_sum
,
3546 scale_load_down(gcfs_rq
->load
.weight
));
3549 /* But make sure to not inflate se's runnable */
3550 runnable_sum
= min(se
->avg
.load_sum
, load_sum
);
3554 * runnable_sum can't be lower than running_sum
3555 * Rescale running sum to be in the same range as runnable sum
3556 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3557 * runnable_sum is in [0 : LOAD_AVG_MAX]
3559 running_sum
= se
->avg
.util_sum
>> SCHED_CAPACITY_SHIFT
;
3560 runnable_sum
= max(runnable_sum
, running_sum
);
3562 load_sum
= (s64
)se_weight(se
) * runnable_sum
;
3563 load_avg
= div_s64(load_sum
, divider
);
3565 delta_sum
= load_sum
- (s64
)se_weight(se
) * se
->avg
.load_sum
;
3566 delta_avg
= load_avg
- se
->avg
.load_avg
;
3568 se
->avg
.load_sum
= runnable_sum
;
3569 se
->avg
.load_avg
= load_avg
;
3570 add_positive(&cfs_rq
->avg
.load_avg
, delta_avg
);
3571 add_positive(&cfs_rq
->avg
.load_sum
, delta_sum
);
3574 static inline void add_tg_cfs_propagate(struct cfs_rq
*cfs_rq
, long runnable_sum
)
3576 cfs_rq
->propagate
= 1;
3577 cfs_rq
->prop_runnable_sum
+= runnable_sum
;
3580 /* Update task and its cfs_rq load average */
3581 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3583 struct cfs_rq
*cfs_rq
, *gcfs_rq
;
3585 if (entity_is_task(se
))
3588 gcfs_rq
= group_cfs_rq(se
);
3589 if (!gcfs_rq
->propagate
)
3592 gcfs_rq
->propagate
= 0;
3594 cfs_rq
= cfs_rq_of(se
);
3596 add_tg_cfs_propagate(cfs_rq
, gcfs_rq
->prop_runnable_sum
);
3598 update_tg_cfs_util(cfs_rq
, se
, gcfs_rq
);
3599 update_tg_cfs_runnable(cfs_rq
, se
, gcfs_rq
);
3600 update_tg_cfs_load(cfs_rq
, se
, gcfs_rq
);
3602 trace_pelt_cfs_tp(cfs_rq
);
3603 trace_pelt_se_tp(se
);
3609 * Check if we need to update the load and the utilization of a blocked
3612 static inline bool skip_blocked_update(struct sched_entity
*se
)
3614 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3617 * If sched_entity still have not zero load or utilization, we have to
3620 if (se
->avg
.load_avg
|| se
->avg
.util_avg
)
3624 * If there is a pending propagation, we have to update the load and
3625 * the utilization of the sched_entity:
3627 if (gcfs_rq
->propagate
)
3631 * Otherwise, the load and the utilization of the sched_entity is
3632 * already zero and there is no pending propagation, so it will be a
3633 * waste of time to try to decay it:
3638 #else /* CONFIG_FAIR_GROUP_SCHED */
3640 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
) {}
3642 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3647 static inline void add_tg_cfs_propagate(struct cfs_rq
*cfs_rq
, long runnable_sum
) {}
3649 #endif /* CONFIG_FAIR_GROUP_SCHED */
3652 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3653 * @now: current time, as per cfs_rq_clock_pelt()
3654 * @cfs_rq: cfs_rq to update
3656 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3657 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3658 * post_init_entity_util_avg().
3660 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3662 * Returns true if the load decayed or we removed load.
3664 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3665 * call update_tg_load_avg() when this function returns true.
3668 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
)
3670 unsigned long removed_load
= 0, removed_util
= 0, removed_runnable
= 0;
3671 struct sched_avg
*sa
= &cfs_rq
->avg
;
3674 if (cfs_rq
->removed
.nr
) {
3676 u32 divider
= get_pelt_divider(&cfs_rq
->avg
);
3678 raw_spin_lock(&cfs_rq
->removed
.lock
);
3679 swap(cfs_rq
->removed
.util_avg
, removed_util
);
3680 swap(cfs_rq
->removed
.load_avg
, removed_load
);
3681 swap(cfs_rq
->removed
.runnable_avg
, removed_runnable
);
3682 cfs_rq
->removed
.nr
= 0;
3683 raw_spin_unlock(&cfs_rq
->removed
.lock
);
3686 sub_positive(&sa
->load_avg
, r
);
3687 sub_positive(&sa
->load_sum
, r
* divider
);
3690 sub_positive(&sa
->util_avg
, r
);
3691 sub_positive(&sa
->util_sum
, r
* divider
);
3693 r
= removed_runnable
;
3694 sub_positive(&sa
->runnable_avg
, r
);
3695 sub_positive(&sa
->runnable_sum
, r
* divider
);
3698 * removed_runnable is the unweighted version of removed_load so we
3699 * can use it to estimate removed_load_sum.
3701 add_tg_cfs_propagate(cfs_rq
,
3702 -(long)(removed_runnable
* divider
) >> SCHED_CAPACITY_SHIFT
);
3707 decayed
|= __update_load_avg_cfs_rq(now
, cfs_rq
);
3709 #ifndef CONFIG_64BIT
3711 cfs_rq
->load_last_update_time_copy
= sa
->last_update_time
;
3718 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3719 * @cfs_rq: cfs_rq to attach to
3720 * @se: sched_entity to attach
3722 * Must call update_cfs_rq_load_avg() before this, since we rely on
3723 * cfs_rq->avg.last_update_time being current.
3725 static void attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3728 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3729 * See ___update_load_avg() for details.
3731 u32 divider
= get_pelt_divider(&cfs_rq
->avg
);
3734 * When we attach the @se to the @cfs_rq, we must align the decay
3735 * window because without that, really weird and wonderful things can
3740 se
->avg
.last_update_time
= cfs_rq
->avg
.last_update_time
;
3741 se
->avg
.period_contrib
= cfs_rq
->avg
.period_contrib
;
3744 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3745 * period_contrib. This isn't strictly correct, but since we're
3746 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3749 se
->avg
.util_sum
= se
->avg
.util_avg
* divider
;
3751 se
->avg
.runnable_sum
= se
->avg
.runnable_avg
* divider
;
3753 se
->avg
.load_sum
= divider
;
3754 if (se_weight(se
)) {
3756 div_u64(se
->avg
.load_avg
* se
->avg
.load_sum
, se_weight(se
));
3759 enqueue_load_avg(cfs_rq
, se
);
3760 cfs_rq
->avg
.util_avg
+= se
->avg
.util_avg
;
3761 cfs_rq
->avg
.util_sum
+= se
->avg
.util_sum
;
3762 cfs_rq
->avg
.runnable_avg
+= se
->avg
.runnable_avg
;
3763 cfs_rq
->avg
.runnable_sum
+= se
->avg
.runnable_sum
;
3765 add_tg_cfs_propagate(cfs_rq
, se
->avg
.load_sum
);
3767 cfs_rq_util_change(cfs_rq
, 0);
3769 trace_pelt_cfs_tp(cfs_rq
);
3773 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3774 * @cfs_rq: cfs_rq to detach from
3775 * @se: sched_entity to detach
3777 * Must call update_cfs_rq_load_avg() before this, since we rely on
3778 * cfs_rq->avg.last_update_time being current.
3780 static void detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3782 dequeue_load_avg(cfs_rq
, se
);
3783 sub_positive(&cfs_rq
->avg
.util_avg
, se
->avg
.util_avg
);
3784 sub_positive(&cfs_rq
->avg
.util_sum
, se
->avg
.util_sum
);
3785 sub_positive(&cfs_rq
->avg
.runnable_avg
, se
->avg
.runnable_avg
);
3786 sub_positive(&cfs_rq
->avg
.runnable_sum
, se
->avg
.runnable_sum
);
3788 add_tg_cfs_propagate(cfs_rq
, -se
->avg
.load_sum
);
3790 cfs_rq_util_change(cfs_rq
, 0);
3792 trace_pelt_cfs_tp(cfs_rq
);
3796 * Optional action to be done while updating the load average
3798 #define UPDATE_TG 0x1
3799 #define SKIP_AGE_LOAD 0x2
3800 #define DO_ATTACH 0x4
3802 /* Update task and its cfs_rq load average */
3803 static inline void update_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3805 u64 now
= cfs_rq_clock_pelt(cfs_rq
);
3809 * Track task load average for carrying it to new CPU after migrated, and
3810 * track group sched_entity load average for task_h_load calc in migration
3812 if (se
->avg
.last_update_time
&& !(flags
& SKIP_AGE_LOAD
))
3813 __update_load_avg_se(now
, cfs_rq
, se
);
3815 decayed
= update_cfs_rq_load_avg(now
, cfs_rq
);
3816 decayed
|= propagate_entity_load_avg(se
);
3818 if (!se
->avg
.last_update_time
&& (flags
& DO_ATTACH
)) {
3821 * DO_ATTACH means we're here from enqueue_entity().
3822 * !last_update_time means we've passed through
3823 * migrate_task_rq_fair() indicating we migrated.
3825 * IOW we're enqueueing a task on a new CPU.
3827 attach_entity_load_avg(cfs_rq
, se
);
3828 update_tg_load_avg(cfs_rq
);
3830 } else if (decayed
) {
3831 cfs_rq_util_change(cfs_rq
, 0);
3833 if (flags
& UPDATE_TG
)
3834 update_tg_load_avg(cfs_rq
);
3838 #ifndef CONFIG_64BIT
3839 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3841 u64 last_update_time_copy
;
3842 u64 last_update_time
;
3845 last_update_time_copy
= cfs_rq
->load_last_update_time_copy
;
3847 last_update_time
= cfs_rq
->avg
.last_update_time
;
3848 } while (last_update_time
!= last_update_time_copy
);
3850 return last_update_time
;
3853 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3855 return cfs_rq
->avg
.last_update_time
;
3860 * Synchronize entity load avg of dequeued entity without locking
3863 static void sync_entity_load_avg(struct sched_entity
*se
)
3865 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3866 u64 last_update_time
;
3868 last_update_time
= cfs_rq_last_update_time(cfs_rq
);
3869 __update_load_avg_blocked_se(last_update_time
, se
);
3873 * Task first catches up with cfs_rq, and then subtract
3874 * itself from the cfs_rq (task must be off the queue now).
3876 static void remove_entity_load_avg(struct sched_entity
*se
)
3878 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3879 unsigned long flags
;
3882 * tasks cannot exit without having gone through wake_up_new_task() ->
3883 * post_init_entity_util_avg() which will have added things to the
3884 * cfs_rq, so we can remove unconditionally.
3887 sync_entity_load_avg(se
);
3889 raw_spin_lock_irqsave(&cfs_rq
->removed
.lock
, flags
);
3890 ++cfs_rq
->removed
.nr
;
3891 cfs_rq
->removed
.util_avg
+= se
->avg
.util_avg
;
3892 cfs_rq
->removed
.load_avg
+= se
->avg
.load_avg
;
3893 cfs_rq
->removed
.runnable_avg
+= se
->avg
.runnable_avg
;
3894 raw_spin_unlock_irqrestore(&cfs_rq
->removed
.lock
, flags
);
3897 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq
*cfs_rq
)
3899 return cfs_rq
->avg
.runnable_avg
;
3902 static inline unsigned long cfs_rq_load_avg(struct cfs_rq
*cfs_rq
)
3904 return cfs_rq
->avg
.load_avg
;
3907 static int newidle_balance(struct rq
*this_rq
, struct rq_flags
*rf
);
3909 static inline unsigned long task_util(struct task_struct
*p
)
3911 return READ_ONCE(p
->se
.avg
.util_avg
);
3914 static inline unsigned long _task_util_est(struct task_struct
*p
)
3916 struct util_est ue
= READ_ONCE(p
->se
.avg
.util_est
);
3918 return (max(ue
.ewma
, ue
.enqueued
) | UTIL_AVG_UNCHANGED
);
3921 static inline unsigned long task_util_est(struct task_struct
*p
)
3923 return max(task_util(p
), _task_util_est(p
));
3926 #ifdef CONFIG_UCLAMP_TASK
3927 static inline unsigned long uclamp_task_util(struct task_struct
*p
)
3929 return clamp(task_util_est(p
),
3930 uclamp_eff_value(p
, UCLAMP_MIN
),
3931 uclamp_eff_value(p
, UCLAMP_MAX
));
3934 static inline unsigned long uclamp_task_util(struct task_struct
*p
)
3936 return task_util_est(p
);
3940 static inline void util_est_enqueue(struct cfs_rq
*cfs_rq
,
3941 struct task_struct
*p
)
3943 unsigned int enqueued
;
3945 if (!sched_feat(UTIL_EST
))
3948 /* Update root cfs_rq's estimated utilization */
3949 enqueued
= cfs_rq
->avg
.util_est
.enqueued
;
3950 enqueued
+= _task_util_est(p
);
3951 WRITE_ONCE(cfs_rq
->avg
.util_est
.enqueued
, enqueued
);
3953 trace_sched_util_est_cfs_tp(cfs_rq
);
3956 static inline void util_est_dequeue(struct cfs_rq
*cfs_rq
,
3957 struct task_struct
*p
)
3959 unsigned int enqueued
;
3961 if (!sched_feat(UTIL_EST
))
3964 /* Update root cfs_rq's estimated utilization */
3965 enqueued
= cfs_rq
->avg
.util_est
.enqueued
;
3966 enqueued
-= min_t(unsigned int, enqueued
, _task_util_est(p
));
3967 WRITE_ONCE(cfs_rq
->avg
.util_est
.enqueued
, enqueued
);
3969 trace_sched_util_est_cfs_tp(cfs_rq
);
3972 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
3975 * Check if a (signed) value is within a specified (unsigned) margin,
3976 * based on the observation that:
3978 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3980 * NOTE: this only works when value + maring < INT_MAX.
3982 static inline bool within_margin(int value
, int margin
)
3984 return ((unsigned int)(value
+ margin
- 1) < (2 * margin
- 1));
3987 static inline void util_est_update(struct cfs_rq
*cfs_rq
,
3988 struct task_struct
*p
,
3991 long last_ewma_diff
, last_enqueued_diff
;
3994 if (!sched_feat(UTIL_EST
))
3998 * Skip update of task's estimated utilization when the task has not
3999 * yet completed an activation, e.g. being migrated.
4005 * If the PELT values haven't changed since enqueue time,
4006 * skip the util_est update.
4008 ue
= p
->se
.avg
.util_est
;
4009 if (ue
.enqueued
& UTIL_AVG_UNCHANGED
)
4012 last_enqueued_diff
= ue
.enqueued
;
4015 * Reset EWMA on utilization increases, the moving average is used only
4016 * to smooth utilization decreases.
4018 ue
.enqueued
= (task_util(p
) | UTIL_AVG_UNCHANGED
);
4019 if (sched_feat(UTIL_EST_FASTUP
)) {
4020 if (ue
.ewma
< ue
.enqueued
) {
4021 ue
.ewma
= ue
.enqueued
;
4027 * Skip update of task's estimated utilization when its members are
4028 * already ~1% close to its last activation value.
4030 last_ewma_diff
= ue
.enqueued
- ue
.ewma
;
4031 last_enqueued_diff
-= ue
.enqueued
;
4032 if (within_margin(last_ewma_diff
, UTIL_EST_MARGIN
)) {
4033 if (!within_margin(last_enqueued_diff
, UTIL_EST_MARGIN
))
4040 * To avoid overestimation of actual task utilization, skip updates if
4041 * we cannot grant there is idle time in this CPU.
4043 if (task_util(p
) > capacity_orig_of(cpu_of(rq_of(cfs_rq
))))
4047 * Update Task's estimated utilization
4049 * When *p completes an activation we can consolidate another sample
4050 * of the task size. This is done by storing the current PELT value
4051 * as ue.enqueued and by using this value to update the Exponential
4052 * Weighted Moving Average (EWMA):
4054 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4055 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4056 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4057 * = w * ( last_ewma_diff ) + ewma(t-1)
4058 * = w * (last_ewma_diff + ewma(t-1) / w)
4060 * Where 'w' is the weight of new samples, which is configured to be
4061 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4063 ue
.ewma
<<= UTIL_EST_WEIGHT_SHIFT
;
4064 ue
.ewma
+= last_ewma_diff
;
4065 ue
.ewma
>>= UTIL_EST_WEIGHT_SHIFT
;
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
, long capacity
)
4074 return fits_capacity(uclamp_task_util(p
), capacity
);
4077 static inline void update_misfit_status(struct task_struct
*p
, struct rq
*rq
)
4079 if (!static_branch_unlikely(&sched_asym_cpucapacity
))
4082 if (!p
|| p
->nr_cpus_allowed
== 1) {
4083 rq
->misfit_task_load
= 0;
4087 if (task_fits_capacity(p
, capacity_of(cpu_of(rq
)))) {
4088 rq
->misfit_task_load
= 0;
4093 * Make sure that misfit_task_load will not be null even if
4094 * task_h_load() returns 0.
4096 rq
->misfit_task_load
= max_t(unsigned long, task_h_load(p
), 1);
4099 #else /* CONFIG_SMP */
4101 #define UPDATE_TG 0x0
4102 #define SKIP_AGE_LOAD 0x0
4103 #define DO_ATTACH 0x0
4105 static inline void update_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int not_used1
)
4107 cfs_rq_util_change(cfs_rq
, 0);
4110 static inline void remove_entity_load_avg(struct sched_entity
*se
) {}
4113 attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
4115 detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
4117 static inline int newidle_balance(struct rq
*rq
, struct rq_flags
*rf
)
4123 util_est_enqueue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
) {}
4126 util_est_dequeue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
) {}
4129 util_est_update(struct cfs_rq
*cfs_rq
, struct task_struct
*p
,
4131 static inline void update_misfit_status(struct task_struct
*p
, struct rq
*rq
) {}
4133 #endif /* CONFIG_SMP */
4135 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4137 #ifdef CONFIG_SCHED_DEBUG
4138 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
4143 if (d
> 3*sysctl_sched_latency
)
4144 schedstat_inc(cfs_rq
->nr_spread_over
);
4149 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
4151 u64 vruntime
= cfs_rq
->min_vruntime
;
4154 * The 'current' period is already promised to the current tasks,
4155 * however the extra weight of the new task will slow them down a
4156 * little, place the new task so that it fits in the slot that
4157 * stays open at the end.
4159 if (initial
&& sched_feat(START_DEBIT
))
4160 vruntime
+= sched_vslice(cfs_rq
, se
);
4162 /* sleeps up to a single latency don't count. */
4164 unsigned long thresh
= sysctl_sched_latency
;
4167 * Halve their sleep time's effect, to allow
4168 * for a gentler effect of sleepers:
4170 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
4176 /* ensure we never gain time by being placed backwards. */
4177 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
4180 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
4182 static inline void check_schedstat_required(void)
4184 #ifdef CONFIG_SCHEDSTATS
4185 if (schedstat_enabled())
4188 /* Force schedstat enabled if a dependent tracepoint is active */
4189 if (trace_sched_stat_wait_enabled() ||
4190 trace_sched_stat_sleep_enabled() ||
4191 trace_sched_stat_iowait_enabled() ||
4192 trace_sched_stat_blocked_enabled() ||
4193 trace_sched_stat_runtime_enabled()) {
4194 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4195 "stat_blocked and stat_runtime require the "
4196 "kernel parameter schedstats=enable or "
4197 "kernel.sched_schedstats=1\n");
4202 static inline bool cfs_bandwidth_used(void);
4209 * update_min_vruntime()
4210 * vruntime -= min_vruntime
4214 * update_min_vruntime()
4215 * vruntime += min_vruntime
4217 * this way the vruntime transition between RQs is done when both
4218 * min_vruntime are up-to-date.
4222 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4223 * vruntime -= min_vruntime
4227 * update_min_vruntime()
4228 * vruntime += min_vruntime
4230 * this way we don't have the most up-to-date min_vruntime on the originating
4231 * CPU and an up-to-date min_vruntime on the destination CPU.
4235 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
4237 bool renorm
= !(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_MIGRATED
);
4238 bool curr
= cfs_rq
->curr
== se
;
4241 * If we're the current task, we must renormalise before calling
4245 se
->vruntime
+= cfs_rq
->min_vruntime
;
4247 update_curr(cfs_rq
);
4250 * Otherwise, renormalise after, such that we're placed at the current
4251 * moment in time, instead of some random moment in the past. Being
4252 * placed in the past could significantly boost this task to the
4253 * fairness detriment of existing tasks.
4255 if (renorm
&& !curr
)
4256 se
->vruntime
+= cfs_rq
->min_vruntime
;
4259 * When enqueuing a sched_entity, we must:
4260 * - Update loads to have both entity and cfs_rq synced with now.
4261 * - Add its load to cfs_rq->runnable_avg
4262 * - For group_entity, update its weight to reflect the new share of
4264 * - Add its new weight to cfs_rq->load.weight
4266 update_load_avg(cfs_rq
, se
, UPDATE_TG
| DO_ATTACH
);
4267 se_update_runnable(se
);
4268 update_cfs_group(se
);
4269 account_entity_enqueue(cfs_rq
, se
);
4271 if (flags
& ENQUEUE_WAKEUP
)
4272 place_entity(cfs_rq
, se
, 0);
4274 check_schedstat_required();
4275 update_stats_enqueue(cfs_rq
, se
, flags
);
4276 check_spread(cfs_rq
, se
);
4278 __enqueue_entity(cfs_rq
, se
);
4282 * When bandwidth control is enabled, cfs might have been removed
4283 * because of a parent been throttled but cfs->nr_running > 1. Try to
4284 * add it unconditionnally.
4286 if (cfs_rq
->nr_running
== 1 || cfs_bandwidth_used())
4287 list_add_leaf_cfs_rq(cfs_rq
);
4289 if (cfs_rq
->nr_running
== 1)
4290 check_enqueue_throttle(cfs_rq
);
4293 static void __clear_buddies_last(struct sched_entity
*se
)
4295 for_each_sched_entity(se
) {
4296 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4297 if (cfs_rq
->last
!= se
)
4300 cfs_rq
->last
= NULL
;
4304 static void __clear_buddies_next(struct sched_entity
*se
)
4306 for_each_sched_entity(se
) {
4307 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4308 if (cfs_rq
->next
!= se
)
4311 cfs_rq
->next
= NULL
;
4315 static void __clear_buddies_skip(struct sched_entity
*se
)
4317 for_each_sched_entity(se
) {
4318 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4319 if (cfs_rq
->skip
!= se
)
4322 cfs_rq
->skip
= NULL
;
4326 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4328 if (cfs_rq
->last
== se
)
4329 __clear_buddies_last(se
);
4331 if (cfs_rq
->next
== se
)
4332 __clear_buddies_next(se
);
4334 if (cfs_rq
->skip
== se
)
4335 __clear_buddies_skip(se
);
4338 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
4341 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
4344 * Update run-time statistics of the 'current'.
4346 update_curr(cfs_rq
);
4349 * When dequeuing a sched_entity, we must:
4350 * - Update loads to have both entity and cfs_rq synced with now.
4351 * - Subtract its load from the cfs_rq->runnable_avg.
4352 * - Subtract its previous weight from cfs_rq->load.weight.
4353 * - For group entity, update its weight to reflect the new share
4354 * of its group cfs_rq.
4356 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4357 se_update_runnable(se
);
4359 update_stats_dequeue(cfs_rq
, se
, flags
);
4361 clear_buddies(cfs_rq
, se
);
4363 if (se
!= cfs_rq
->curr
)
4364 __dequeue_entity(cfs_rq
, se
);
4366 account_entity_dequeue(cfs_rq
, se
);
4369 * Normalize after update_curr(); which will also have moved
4370 * min_vruntime if @se is the one holding it back. But before doing
4371 * update_min_vruntime() again, which will discount @se's position and
4372 * can move min_vruntime forward still more.
4374 if (!(flags
& DEQUEUE_SLEEP
))
4375 se
->vruntime
-= cfs_rq
->min_vruntime
;
4377 /* return excess runtime on last dequeue */
4378 return_cfs_rq_runtime(cfs_rq
);
4380 update_cfs_group(se
);
4383 * Now advance min_vruntime if @se was the entity holding it back,
4384 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4385 * put back on, and if we advance min_vruntime, we'll be placed back
4386 * further than we started -- ie. we'll be penalized.
4388 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) != DEQUEUE_SAVE
)
4389 update_min_vruntime(cfs_rq
);
4393 * Preempt the current task with a newly woken task if needed:
4396 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
4398 unsigned long ideal_runtime
, delta_exec
;
4399 struct sched_entity
*se
;
4402 ideal_runtime
= sched_slice(cfs_rq
, curr
);
4403 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
4404 if (delta_exec
> ideal_runtime
) {
4405 resched_curr(rq_of(cfs_rq
));
4407 * The current task ran long enough, ensure it doesn't get
4408 * re-elected due to buddy favours.
4410 clear_buddies(cfs_rq
, curr
);
4415 * Ensure that a task that missed wakeup preemption by a
4416 * narrow margin doesn't have to wait for a full slice.
4417 * This also mitigates buddy induced latencies under load.
4419 if (delta_exec
< sysctl_sched_min_granularity
)
4422 se
= __pick_first_entity(cfs_rq
);
4423 delta
= curr
->vruntime
- se
->vruntime
;
4428 if (delta
> ideal_runtime
)
4429 resched_curr(rq_of(cfs_rq
));
4433 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4435 /* 'current' is not kept within the tree. */
4438 * Any task has to be enqueued before it get to execute on
4439 * a CPU. So account for the time it spent waiting on the
4442 update_stats_wait_end(cfs_rq
, se
);
4443 __dequeue_entity(cfs_rq
, se
);
4444 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4447 update_stats_curr_start(cfs_rq
, se
);
4451 * Track our maximum slice length, if the CPU's load is at
4452 * least twice that of our own weight (i.e. dont track it
4453 * when there are only lesser-weight tasks around):
4455 if (schedstat_enabled() &&
4456 rq_of(cfs_rq
)->cfs
.load
.weight
>= 2*se
->load
.weight
) {
4457 schedstat_set(se
->statistics
.slice_max
,
4458 max((u64
)schedstat_val(se
->statistics
.slice_max
),
4459 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
));
4462 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
4466 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
4469 * Pick the next process, keeping these things in mind, in this order:
4470 * 1) keep things fair between processes/task groups
4471 * 2) pick the "next" process, since someone really wants that to run
4472 * 3) pick the "last" process, for cache locality
4473 * 4) do not run the "skip" process, if something else is available
4475 static struct sched_entity
*
4476 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
4478 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
4479 struct sched_entity
*se
;
4482 * If curr is set we have to see if its left of the leftmost entity
4483 * still in the tree, provided there was anything in the tree at all.
4485 if (!left
|| (curr
&& entity_before(curr
, left
)))
4488 se
= left
; /* ideally we run the leftmost entity */
4491 * Avoid running the skip buddy, if running something else can
4492 * be done without getting too unfair.
4494 if (cfs_rq
->skip
== se
) {
4495 struct sched_entity
*second
;
4498 second
= __pick_first_entity(cfs_rq
);
4500 second
= __pick_next_entity(se
);
4501 if (!second
|| (curr
&& entity_before(curr
, second
)))
4505 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
4509 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1) {
4511 * Someone really wants this to run. If it's not unfair, run it.
4514 } else if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1) {
4516 * Prefer last buddy, try to return the CPU to a preempted task.
4521 clear_buddies(cfs_rq
, se
);
4526 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
4528 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
4531 * If still on the runqueue then deactivate_task()
4532 * was not called and update_curr() has to be done:
4535 update_curr(cfs_rq
);
4537 /* throttle cfs_rqs exceeding runtime */
4538 check_cfs_rq_runtime(cfs_rq
);
4540 check_spread(cfs_rq
, prev
);
4543 update_stats_wait_start(cfs_rq
, prev
);
4544 /* Put 'current' back into the tree. */
4545 __enqueue_entity(cfs_rq
, prev
);
4546 /* in !on_rq case, update occurred at dequeue */
4547 update_load_avg(cfs_rq
, prev
, 0);
4549 cfs_rq
->curr
= NULL
;
4553 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
4556 * Update run-time statistics of the 'current'.
4558 update_curr(cfs_rq
);
4561 * Ensure that runnable average is periodically updated.
4563 update_load_avg(cfs_rq
, curr
, UPDATE_TG
);
4564 update_cfs_group(curr
);
4566 #ifdef CONFIG_SCHED_HRTICK
4568 * queued ticks are scheduled to match the slice, so don't bother
4569 * validating it and just reschedule.
4572 resched_curr(rq_of(cfs_rq
));
4576 * don't let the period tick interfere with the hrtick preemption
4578 if (!sched_feat(DOUBLE_TICK
) &&
4579 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
4583 if (cfs_rq
->nr_running
> 1)
4584 check_preempt_tick(cfs_rq
, curr
);
4588 /**************************************************
4589 * CFS bandwidth control machinery
4592 #ifdef CONFIG_CFS_BANDWIDTH
4594 #ifdef CONFIG_JUMP_LABEL
4595 static struct static_key __cfs_bandwidth_used
;
4597 static inline bool cfs_bandwidth_used(void)
4599 return static_key_false(&__cfs_bandwidth_used
);
4602 void cfs_bandwidth_usage_inc(void)
4604 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used
);
4607 void cfs_bandwidth_usage_dec(void)
4609 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used
);
4611 #else /* CONFIG_JUMP_LABEL */
4612 static bool cfs_bandwidth_used(void)
4617 void cfs_bandwidth_usage_inc(void) {}
4618 void cfs_bandwidth_usage_dec(void) {}
4619 #endif /* CONFIG_JUMP_LABEL */
4622 * default period for cfs group bandwidth.
4623 * default: 0.1s, units: nanoseconds
4625 static inline u64
default_cfs_period(void)
4627 return 100000000ULL;
4630 static inline u64
sched_cfs_bandwidth_slice(void)
4632 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
4636 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4637 * directly instead of rq->clock to avoid adding additional synchronization
4640 * requires cfs_b->lock
4642 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
4644 if (cfs_b
->quota
!= RUNTIME_INF
)
4645 cfs_b
->runtime
= cfs_b
->quota
;
4648 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4650 return &tg
->cfs_bandwidth
;
4653 /* returns 0 on failure to allocate runtime */
4654 static int __assign_cfs_rq_runtime(struct cfs_bandwidth
*cfs_b
,
4655 struct cfs_rq
*cfs_rq
, u64 target_runtime
)
4657 u64 min_amount
, amount
= 0;
4659 lockdep_assert_held(&cfs_b
->lock
);
4661 /* note: this is a positive sum as runtime_remaining <= 0 */
4662 min_amount
= target_runtime
- cfs_rq
->runtime_remaining
;
4664 if (cfs_b
->quota
== RUNTIME_INF
)
4665 amount
= min_amount
;
4667 start_cfs_bandwidth(cfs_b
);
4669 if (cfs_b
->runtime
> 0) {
4670 amount
= min(cfs_b
->runtime
, min_amount
);
4671 cfs_b
->runtime
-= amount
;
4676 cfs_rq
->runtime_remaining
+= amount
;
4678 return cfs_rq
->runtime_remaining
> 0;
4681 /* returns 0 on failure to allocate runtime */
4682 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4684 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4687 raw_spin_lock(&cfs_b
->lock
);
4688 ret
= __assign_cfs_rq_runtime(cfs_b
, cfs_rq
, sched_cfs_bandwidth_slice());
4689 raw_spin_unlock(&cfs_b
->lock
);
4694 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4696 /* dock delta_exec before expiring quota (as it could span periods) */
4697 cfs_rq
->runtime_remaining
-= delta_exec
;
4699 if (likely(cfs_rq
->runtime_remaining
> 0))
4702 if (cfs_rq
->throttled
)
4705 * if we're unable to extend our runtime we resched so that the active
4706 * hierarchy can be throttled
4708 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
4709 resched_curr(rq_of(cfs_rq
));
4712 static __always_inline
4713 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4715 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
4718 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
4721 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4723 return cfs_bandwidth_used() && cfs_rq
->throttled
;
4726 /* check whether cfs_rq, or any parent, is throttled */
4727 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4729 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
4733 * Ensure that neither of the group entities corresponding to src_cpu or
4734 * dest_cpu are members of a throttled hierarchy when performing group
4735 * load-balance operations.
4737 static inline int throttled_lb_pair(struct task_group
*tg
,
4738 int src_cpu
, int dest_cpu
)
4740 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
4742 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
4743 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
4745 return throttled_hierarchy(src_cfs_rq
) ||
4746 throttled_hierarchy(dest_cfs_rq
);
4749 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
4751 struct rq
*rq
= data
;
4752 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4754 cfs_rq
->throttle_count
--;
4755 if (!cfs_rq
->throttle_count
) {
4756 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
4757 cfs_rq
->throttled_clock_task
;
4759 /* Add cfs_rq with already running entity in the list */
4760 if (cfs_rq
->nr_running
>= 1)
4761 list_add_leaf_cfs_rq(cfs_rq
);
4767 static int tg_throttle_down(struct task_group
*tg
, void *data
)
4769 struct rq
*rq
= data
;
4770 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4772 /* group is entering throttled state, stop time */
4773 if (!cfs_rq
->throttle_count
) {
4774 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
4775 list_del_leaf_cfs_rq(cfs_rq
);
4777 cfs_rq
->throttle_count
++;
4782 static bool throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
4784 struct rq
*rq
= rq_of(cfs_rq
);
4785 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4786 struct sched_entity
*se
;
4787 long task_delta
, idle_task_delta
, dequeue
= 1;
4789 raw_spin_lock(&cfs_b
->lock
);
4790 /* This will start the period timer if necessary */
4791 if (__assign_cfs_rq_runtime(cfs_b
, cfs_rq
, 1)) {
4793 * We have raced with bandwidth becoming available, and if we
4794 * actually throttled the timer might not unthrottle us for an
4795 * entire period. We additionally needed to make sure that any
4796 * subsequent check_cfs_rq_runtime calls agree not to throttle
4797 * us, as we may commit to do cfs put_prev+pick_next, so we ask
4798 * for 1ns of runtime rather than just check cfs_b.
4802 list_add_tail_rcu(&cfs_rq
->throttled_list
,
4803 &cfs_b
->throttled_cfs_rq
);
4805 raw_spin_unlock(&cfs_b
->lock
);
4808 return false; /* Throttle no longer required. */
4810 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
4812 /* freeze hierarchy runnable averages while throttled */
4814 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
4817 task_delta
= cfs_rq
->h_nr_running
;
4818 idle_task_delta
= cfs_rq
->idle_h_nr_running
;
4819 for_each_sched_entity(se
) {
4820 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4821 /* throttled entity or throttle-on-deactivate */
4825 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
4827 qcfs_rq
->h_nr_running
-= task_delta
;
4828 qcfs_rq
->idle_h_nr_running
-= idle_task_delta
;
4830 if (qcfs_rq
->load
.weight
) {
4831 /* Avoid re-evaluating load for this entity: */
4832 se
= parent_entity(se
);
4837 for_each_sched_entity(se
) {
4838 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4839 /* throttled entity or throttle-on-deactivate */
4843 update_load_avg(qcfs_rq
, se
, 0);
4844 se_update_runnable(se
);
4846 qcfs_rq
->h_nr_running
-= task_delta
;
4847 qcfs_rq
->idle_h_nr_running
-= idle_task_delta
;
4850 /* At this point se is NULL and we are at root level*/
4851 sub_nr_running(rq
, task_delta
);
4855 * Note: distribution will already see us throttled via the
4856 * throttled-list. rq->lock protects completion.
4858 cfs_rq
->throttled
= 1;
4859 cfs_rq
->throttled_clock
= rq_clock(rq
);
4863 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
4865 struct rq
*rq
= rq_of(cfs_rq
);
4866 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4867 struct sched_entity
*se
;
4868 long task_delta
, idle_task_delta
;
4870 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
4872 cfs_rq
->throttled
= 0;
4874 update_rq_clock(rq
);
4876 raw_spin_lock(&cfs_b
->lock
);
4877 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
4878 list_del_rcu(&cfs_rq
->throttled_list
);
4879 raw_spin_unlock(&cfs_b
->lock
);
4881 /* update hierarchical throttle state */
4882 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
4884 if (!cfs_rq
->load
.weight
)
4887 task_delta
= cfs_rq
->h_nr_running
;
4888 idle_task_delta
= cfs_rq
->idle_h_nr_running
;
4889 for_each_sched_entity(se
) {
4892 cfs_rq
= cfs_rq_of(se
);
4893 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
4895 cfs_rq
->h_nr_running
+= task_delta
;
4896 cfs_rq
->idle_h_nr_running
+= idle_task_delta
;
4898 /* end evaluation on encountering a throttled cfs_rq */
4899 if (cfs_rq_throttled(cfs_rq
))
4900 goto unthrottle_throttle
;
4903 for_each_sched_entity(se
) {
4904 cfs_rq
= cfs_rq_of(se
);
4906 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4907 se_update_runnable(se
);
4909 cfs_rq
->h_nr_running
+= task_delta
;
4910 cfs_rq
->idle_h_nr_running
+= idle_task_delta
;
4913 /* end evaluation on encountering a throttled cfs_rq */
4914 if (cfs_rq_throttled(cfs_rq
))
4915 goto unthrottle_throttle
;
4918 * One parent has been throttled and cfs_rq removed from the
4919 * list. Add it back to not break the leaf list.
4921 if (throttled_hierarchy(cfs_rq
))
4922 list_add_leaf_cfs_rq(cfs_rq
);
4925 /* At this point se is NULL and we are at root level*/
4926 add_nr_running(rq
, task_delta
);
4928 unthrottle_throttle
:
4930 * The cfs_rq_throttled() breaks in the above iteration can result in
4931 * incomplete leaf list maintenance, resulting in triggering the
4934 for_each_sched_entity(se
) {
4935 cfs_rq
= cfs_rq_of(se
);
4937 if (list_add_leaf_cfs_rq(cfs_rq
))
4941 assert_list_leaf_cfs_rq(rq
);
4943 /* Determine whether we need to wake up potentially idle CPU: */
4944 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
4948 static void distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
)
4950 struct cfs_rq
*cfs_rq
;
4951 u64 runtime
, remaining
= 1;
4954 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
4956 struct rq
*rq
= rq_of(cfs_rq
);
4959 rq_lock_irqsave(rq
, &rf
);
4960 if (!cfs_rq_throttled(cfs_rq
))
4963 /* By the above check, this should never be true */
4964 SCHED_WARN_ON(cfs_rq
->runtime_remaining
> 0);
4966 raw_spin_lock(&cfs_b
->lock
);
4967 runtime
= -cfs_rq
->runtime_remaining
+ 1;
4968 if (runtime
> cfs_b
->runtime
)
4969 runtime
= cfs_b
->runtime
;
4970 cfs_b
->runtime
-= runtime
;
4971 remaining
= cfs_b
->runtime
;
4972 raw_spin_unlock(&cfs_b
->lock
);
4974 cfs_rq
->runtime_remaining
+= runtime
;
4976 /* we check whether we're throttled above */
4977 if (cfs_rq
->runtime_remaining
> 0)
4978 unthrottle_cfs_rq(cfs_rq
);
4981 rq_unlock_irqrestore(rq
, &rf
);
4990 * Responsible for refilling a task_group's bandwidth and unthrottling its
4991 * cfs_rqs as appropriate. If there has been no activity within the last
4992 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4993 * used to track this state.
4995 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
, unsigned long flags
)
4999 /* no need to continue the timer with no bandwidth constraint */
5000 if (cfs_b
->quota
== RUNTIME_INF
)
5001 goto out_deactivate
;
5003 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
5004 cfs_b
->nr_periods
+= overrun
;
5007 * idle depends on !throttled (for the case of a large deficit), and if
5008 * we're going inactive then everything else can be deferred
5010 if (cfs_b
->idle
&& !throttled
)
5011 goto out_deactivate
;
5013 __refill_cfs_bandwidth_runtime(cfs_b
);
5016 /* mark as potentially idle for the upcoming period */
5021 /* account preceding periods in which throttling occurred */
5022 cfs_b
->nr_throttled
+= overrun
;
5025 * This check is repeated as we release cfs_b->lock while we unthrottle.
5027 while (throttled
&& cfs_b
->runtime
> 0) {
5028 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5029 /* we can't nest cfs_b->lock while distributing bandwidth */
5030 distribute_cfs_runtime(cfs_b
);
5031 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5033 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
5037 * While we are ensured activity in the period following an
5038 * unthrottle, this also covers the case in which the new bandwidth is
5039 * insufficient to cover the existing bandwidth deficit. (Forcing the
5040 * timer to remain active while there are any throttled entities.)
5050 /* a cfs_rq won't donate quota below this amount */
5051 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
5052 /* minimum remaining period time to redistribute slack quota */
5053 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
5054 /* how long we wait to gather additional slack before distributing */
5055 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
5058 * Are we near the end of the current quota period?
5060 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5061 * hrtimer base being cleared by hrtimer_start. In the case of
5062 * migrate_hrtimers, base is never cleared, so we are fine.
5064 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
5066 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
5069 /* if the call-back is running a quota refresh is already occurring */
5070 if (hrtimer_callback_running(refresh_timer
))
5073 /* is a quota refresh about to occur? */
5074 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
5075 if (remaining
< min_expire
)
5081 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
5083 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
5085 /* if there's a quota refresh soon don't bother with slack */
5086 if (runtime_refresh_within(cfs_b
, min_left
))
5089 /* don't push forwards an existing deferred unthrottle */
5090 if (cfs_b
->slack_started
)
5092 cfs_b
->slack_started
= true;
5094 hrtimer_start(&cfs_b
->slack_timer
,
5095 ns_to_ktime(cfs_bandwidth_slack_period
),
5099 /* we know any runtime found here is valid as update_curr() precedes return */
5100 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5102 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
5103 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
5105 if (slack_runtime
<= 0)
5108 raw_spin_lock(&cfs_b
->lock
);
5109 if (cfs_b
->quota
!= RUNTIME_INF
) {
5110 cfs_b
->runtime
+= slack_runtime
;
5112 /* we are under rq->lock, defer unthrottling using a timer */
5113 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
5114 !list_empty(&cfs_b
->throttled_cfs_rq
))
5115 start_cfs_slack_bandwidth(cfs_b
);
5117 raw_spin_unlock(&cfs_b
->lock
);
5119 /* even if it's not valid for return we don't want to try again */
5120 cfs_rq
->runtime_remaining
-= slack_runtime
;
5123 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5125 if (!cfs_bandwidth_used())
5128 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
5131 __return_cfs_rq_runtime(cfs_rq
);
5135 * This is done with a timer (instead of inline with bandwidth return) since
5136 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5138 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
5140 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
5141 unsigned long flags
;
5143 /* confirm we're still not at a refresh boundary */
5144 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5145 cfs_b
->slack_started
= false;
5147 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
5148 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5152 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
5153 runtime
= cfs_b
->runtime
;
5155 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5160 distribute_cfs_runtime(cfs_b
);
5164 * When a group wakes up we want to make sure that its quota is not already
5165 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5166 * runtime as update_curr() throttling can not not trigger until it's on-rq.
5168 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
5170 if (!cfs_bandwidth_used())
5173 /* an active group must be handled by the update_curr()->put() path */
5174 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
5177 /* ensure the group is not already throttled */
5178 if (cfs_rq_throttled(cfs_rq
))
5181 /* update runtime allocation */
5182 account_cfs_rq_runtime(cfs_rq
, 0);
5183 if (cfs_rq
->runtime_remaining
<= 0)
5184 throttle_cfs_rq(cfs_rq
);
5187 static void sync_throttle(struct task_group
*tg
, int cpu
)
5189 struct cfs_rq
*pcfs_rq
, *cfs_rq
;
5191 if (!cfs_bandwidth_used())
5197 cfs_rq
= tg
->cfs_rq
[cpu
];
5198 pcfs_rq
= tg
->parent
->cfs_rq
[cpu
];
5200 cfs_rq
->throttle_count
= pcfs_rq
->throttle_count
;
5201 cfs_rq
->throttled_clock_task
= rq_clock_task(cpu_rq(cpu
));
5204 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5205 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5207 if (!cfs_bandwidth_used())
5210 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
5214 * it's possible for a throttled entity to be forced into a running
5215 * state (e.g. set_curr_task), in this case we're finished.
5217 if (cfs_rq_throttled(cfs_rq
))
5220 return throttle_cfs_rq(cfs_rq
);
5223 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
5225 struct cfs_bandwidth
*cfs_b
=
5226 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
5228 do_sched_cfs_slack_timer(cfs_b
);
5230 return HRTIMER_NORESTART
;
5233 extern const u64 max_cfs_quota_period
;
5235 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
5237 struct cfs_bandwidth
*cfs_b
=
5238 container_of(timer
, struct cfs_bandwidth
, period_timer
);
5239 unsigned long flags
;
5244 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5246 overrun
= hrtimer_forward_now(timer
, cfs_b
->period
);
5250 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
, flags
);
5253 u64
new, old
= ktime_to_ns(cfs_b
->period
);
5256 * Grow period by a factor of 2 to avoid losing precision.
5257 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5261 if (new < max_cfs_quota_period
) {
5262 cfs_b
->period
= ns_to_ktime(new);
5265 pr_warn_ratelimited(
5266 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5268 div_u64(new, NSEC_PER_USEC
),
5269 div_u64(cfs_b
->quota
, NSEC_PER_USEC
));
5271 pr_warn_ratelimited(
5272 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5274 div_u64(old
, NSEC_PER_USEC
),
5275 div_u64(cfs_b
->quota
, NSEC_PER_USEC
));
5278 /* reset count so we don't come right back in here */
5283 cfs_b
->period_active
= 0;
5284 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5286 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
5289 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5291 raw_spin_lock_init(&cfs_b
->lock
);
5293 cfs_b
->quota
= RUNTIME_INF
;
5294 cfs_b
->period
= ns_to_ktime(default_cfs_period());
5296 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
5297 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_PINNED
);
5298 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
5299 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5300 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
5301 cfs_b
->slack_started
= false;
5304 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5306 cfs_rq
->runtime_enabled
= 0;
5307 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
5310 void start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5312 lockdep_assert_held(&cfs_b
->lock
);
5314 if (cfs_b
->period_active
)
5317 cfs_b
->period_active
= 1;
5318 hrtimer_forward_now(&cfs_b
->period_timer
, cfs_b
->period
);
5319 hrtimer_start_expires(&cfs_b
->period_timer
, HRTIMER_MODE_ABS_PINNED
);
5322 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5324 /* init_cfs_bandwidth() was not called */
5325 if (!cfs_b
->throttled_cfs_rq
.next
)
5328 hrtimer_cancel(&cfs_b
->period_timer
);
5329 hrtimer_cancel(&cfs_b
->slack_timer
);
5333 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5335 * The race is harmless, since modifying bandwidth settings of unhooked group
5336 * bits doesn't do much.
5339 /* cpu online calback */
5340 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
5342 struct task_group
*tg
;
5344 lockdep_assert_held(&rq
->lock
);
5347 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
5348 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
5349 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
5351 raw_spin_lock(&cfs_b
->lock
);
5352 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
5353 raw_spin_unlock(&cfs_b
->lock
);
5358 /* cpu offline callback */
5359 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
5361 struct task_group
*tg
;
5363 lockdep_assert_held(&rq
->lock
);
5366 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
5367 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
5369 if (!cfs_rq
->runtime_enabled
)
5373 * clock_task is not advancing so we just need to make sure
5374 * there's some valid quota amount
5376 cfs_rq
->runtime_remaining
= 1;
5378 * Offline rq is schedulable till CPU is completely disabled
5379 * in take_cpu_down(), so we prevent new cfs throttling here.
5381 cfs_rq
->runtime_enabled
= 0;
5383 if (cfs_rq_throttled(cfs_rq
))
5384 unthrottle_cfs_rq(cfs_rq
);
5389 #else /* CONFIG_CFS_BANDWIDTH */
5391 static inline bool cfs_bandwidth_used(void)
5396 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
5397 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
5398 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
5399 static inline void sync_throttle(struct task_group
*tg
, int cpu
) {}
5400 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
5402 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
5407 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
5412 static inline int throttled_lb_pair(struct task_group
*tg
,
5413 int src_cpu
, int dest_cpu
)
5418 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
5420 #ifdef CONFIG_FAIR_GROUP_SCHED
5421 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
5424 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
5428 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
5429 static inline void update_runtime_enabled(struct rq
*rq
) {}
5430 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
5432 #endif /* CONFIG_CFS_BANDWIDTH */
5434 /**************************************************
5435 * CFS operations on tasks:
5438 #ifdef CONFIG_SCHED_HRTICK
5439 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
5441 struct sched_entity
*se
= &p
->se
;
5442 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5444 SCHED_WARN_ON(task_rq(p
) != rq
);
5446 if (rq
->cfs
.h_nr_running
> 1) {
5447 u64 slice
= sched_slice(cfs_rq
, se
);
5448 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
5449 s64 delta
= slice
- ran
;
5456 hrtick_start(rq
, delta
);
5461 * called from enqueue/dequeue and updates the hrtick when the
5462 * current task is from our class and nr_running is low enough
5465 static void hrtick_update(struct rq
*rq
)
5467 struct task_struct
*curr
= rq
->curr
;
5469 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
5472 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
5473 hrtick_start_fair(rq
, curr
);
5475 #else /* !CONFIG_SCHED_HRTICK */
5477 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
5481 static inline void hrtick_update(struct rq
*rq
)
5487 static inline unsigned long cpu_util(int cpu
);
5489 static inline bool cpu_overutilized(int cpu
)
5491 return !fits_capacity(cpu_util(cpu
), capacity_of(cpu
));
5494 static inline void update_overutilized_status(struct rq
*rq
)
5496 if (!READ_ONCE(rq
->rd
->overutilized
) && cpu_overutilized(rq
->cpu
)) {
5497 WRITE_ONCE(rq
->rd
->overutilized
, SG_OVERUTILIZED
);
5498 trace_sched_overutilized_tp(rq
->rd
, SG_OVERUTILIZED
);
5502 static inline void update_overutilized_status(struct rq
*rq
) { }
5505 /* Runqueue only has SCHED_IDLE tasks enqueued */
5506 static int sched_idle_rq(struct rq
*rq
)
5508 return unlikely(rq
->nr_running
== rq
->cfs
.idle_h_nr_running
&&
5513 static int sched_idle_cpu(int cpu
)
5515 return sched_idle_rq(cpu_rq(cpu
));
5520 * The enqueue_task method is called before nr_running is
5521 * increased. Here we update the fair scheduling stats and
5522 * then put the task into the rbtree:
5525 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
5527 struct cfs_rq
*cfs_rq
;
5528 struct sched_entity
*se
= &p
->se
;
5529 int idle_h_nr_running
= task_has_idle_policy(p
);
5530 int task_new
= !(flags
& ENQUEUE_WAKEUP
);
5533 * The code below (indirectly) updates schedutil which looks at
5534 * the cfs_rq utilization to select a frequency.
5535 * Let's add the task's estimated utilization to the cfs_rq's
5536 * estimated utilization, before we update schedutil.
5538 util_est_enqueue(&rq
->cfs
, p
);
5541 * If in_iowait is set, the code below may not trigger any cpufreq
5542 * utilization updates, so do it here explicitly with the IOWAIT flag
5546 cpufreq_update_util(rq
, SCHED_CPUFREQ_IOWAIT
);
5548 for_each_sched_entity(se
) {
5551 cfs_rq
= cfs_rq_of(se
);
5552 enqueue_entity(cfs_rq
, se
, flags
);
5554 cfs_rq
->h_nr_running
++;
5555 cfs_rq
->idle_h_nr_running
+= idle_h_nr_running
;
5557 /* end evaluation on encountering a throttled cfs_rq */
5558 if (cfs_rq_throttled(cfs_rq
))
5559 goto enqueue_throttle
;
5561 flags
= ENQUEUE_WAKEUP
;
5564 for_each_sched_entity(se
) {
5565 cfs_rq
= cfs_rq_of(se
);
5567 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
5568 se_update_runnable(se
);
5569 update_cfs_group(se
);
5571 cfs_rq
->h_nr_running
++;
5572 cfs_rq
->idle_h_nr_running
+= idle_h_nr_running
;
5574 /* end evaluation on encountering a throttled cfs_rq */
5575 if (cfs_rq_throttled(cfs_rq
))
5576 goto enqueue_throttle
;
5579 * One parent has been throttled and cfs_rq removed from the
5580 * list. Add it back to not break the leaf list.
5582 if (throttled_hierarchy(cfs_rq
))
5583 list_add_leaf_cfs_rq(cfs_rq
);
5586 /* At this point se is NULL and we are at root level*/
5587 add_nr_running(rq
, 1);
5590 * Since new tasks are assigned an initial util_avg equal to
5591 * half of the spare capacity of their CPU, tiny tasks have the
5592 * ability to cross the overutilized threshold, which will
5593 * result in the load balancer ruining all the task placement
5594 * done by EAS. As a way to mitigate that effect, do not account
5595 * for the first enqueue operation of new tasks during the
5596 * overutilized flag detection.
5598 * A better way of solving this problem would be to wait for
5599 * the PELT signals of tasks to converge before taking them
5600 * into account, but that is not straightforward to implement,
5601 * and the following generally works well enough in practice.
5604 update_overutilized_status(rq
);
5607 if (cfs_bandwidth_used()) {
5609 * When bandwidth control is enabled; the cfs_rq_throttled()
5610 * breaks in the above iteration can result in incomplete
5611 * leaf list maintenance, resulting in triggering the assertion
5614 for_each_sched_entity(se
) {
5615 cfs_rq
= cfs_rq_of(se
);
5617 if (list_add_leaf_cfs_rq(cfs_rq
))
5622 assert_list_leaf_cfs_rq(rq
);
5627 static void set_next_buddy(struct sched_entity
*se
);
5630 * The dequeue_task method is called before nr_running is
5631 * decreased. We remove the task from the rbtree and
5632 * update the fair scheduling stats:
5634 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
5636 struct cfs_rq
*cfs_rq
;
5637 struct sched_entity
*se
= &p
->se
;
5638 int task_sleep
= flags
& DEQUEUE_SLEEP
;
5639 int idle_h_nr_running
= task_has_idle_policy(p
);
5640 bool was_sched_idle
= sched_idle_rq(rq
);
5642 util_est_dequeue(&rq
->cfs
, p
);
5644 for_each_sched_entity(se
) {
5645 cfs_rq
= cfs_rq_of(se
);
5646 dequeue_entity(cfs_rq
, se
, flags
);
5648 cfs_rq
->h_nr_running
--;
5649 cfs_rq
->idle_h_nr_running
-= idle_h_nr_running
;
5651 /* end evaluation on encountering a throttled cfs_rq */
5652 if (cfs_rq_throttled(cfs_rq
))
5653 goto dequeue_throttle
;
5655 /* Don't dequeue parent if it has other entities besides us */
5656 if (cfs_rq
->load
.weight
) {
5657 /* Avoid re-evaluating load for this entity: */
5658 se
= parent_entity(se
);
5660 * Bias pick_next to pick a task from this cfs_rq, as
5661 * p is sleeping when it is within its sched_slice.
5663 if (task_sleep
&& se
&& !throttled_hierarchy(cfs_rq
))
5667 flags
|= DEQUEUE_SLEEP
;
5670 for_each_sched_entity(se
) {
5671 cfs_rq
= cfs_rq_of(se
);
5673 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
5674 se_update_runnable(se
);
5675 update_cfs_group(se
);
5677 cfs_rq
->h_nr_running
--;
5678 cfs_rq
->idle_h_nr_running
-= idle_h_nr_running
;
5680 /* end evaluation on encountering a throttled cfs_rq */
5681 if (cfs_rq_throttled(cfs_rq
))
5682 goto dequeue_throttle
;
5686 /* At this point se is NULL and we are at root level*/
5687 sub_nr_running(rq
, 1);
5689 /* balance early to pull high priority tasks */
5690 if (unlikely(!was_sched_idle
&& sched_idle_rq(rq
)))
5691 rq
->next_balance
= jiffies
;
5694 util_est_update(&rq
->cfs
, p
, task_sleep
);
5700 /* Working cpumask for: load_balance, load_balance_newidle. */
5701 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5702 DEFINE_PER_CPU(cpumask_var_t
, select_idle_mask
);
5704 #ifdef CONFIG_NO_HZ_COMMON
5707 cpumask_var_t idle_cpus_mask
;
5709 int has_blocked
; /* Idle CPUS has blocked load */
5710 unsigned long next_balance
; /* in jiffy units */
5711 unsigned long next_blocked
; /* Next update of blocked load in jiffies */
5712 } nohz ____cacheline_aligned
;
5714 #endif /* CONFIG_NO_HZ_COMMON */
5716 static unsigned long cpu_load(struct rq
*rq
)
5718 return cfs_rq_load_avg(&rq
->cfs
);
5722 * cpu_load_without - compute CPU load without any contributions from *p
5723 * @cpu: the CPU which load is requested
5724 * @p: the task which load should be discounted
5726 * The load of a CPU is defined by the load of tasks currently enqueued on that
5727 * CPU as well as tasks which are currently sleeping after an execution on that
5730 * This method returns the load of the specified CPU by discounting the load of
5731 * the specified task, whenever the task is currently contributing to the CPU
5734 static unsigned long cpu_load_without(struct rq
*rq
, struct task_struct
*p
)
5736 struct cfs_rq
*cfs_rq
;
5739 /* Task has no contribution or is new */
5740 if (cpu_of(rq
) != task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
5741 return cpu_load(rq
);
5744 load
= READ_ONCE(cfs_rq
->avg
.load_avg
);
5746 /* Discount task's util from CPU's util */
5747 lsub_positive(&load
, task_h_load(p
));
5752 static unsigned long cpu_runnable(struct rq
*rq
)
5754 return cfs_rq_runnable_avg(&rq
->cfs
);
5757 static unsigned long cpu_runnable_without(struct rq
*rq
, struct task_struct
*p
)
5759 struct cfs_rq
*cfs_rq
;
5760 unsigned int runnable
;
5762 /* Task has no contribution or is new */
5763 if (cpu_of(rq
) != task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
5764 return cpu_runnable(rq
);
5767 runnable
= READ_ONCE(cfs_rq
->avg
.runnable_avg
);
5769 /* Discount task's runnable from CPU's runnable */
5770 lsub_positive(&runnable
, p
->se
.avg
.runnable_avg
);
5775 static unsigned long capacity_of(int cpu
)
5777 return cpu_rq(cpu
)->cpu_capacity
;
5780 static void record_wakee(struct task_struct
*p
)
5783 * Only decay a single time; tasks that have less then 1 wakeup per
5784 * jiffy will not have built up many flips.
5786 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
5787 current
->wakee_flips
>>= 1;
5788 current
->wakee_flip_decay_ts
= jiffies
;
5791 if (current
->last_wakee
!= p
) {
5792 current
->last_wakee
= p
;
5793 current
->wakee_flips
++;
5798 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5800 * A waker of many should wake a different task than the one last awakened
5801 * at a frequency roughly N times higher than one of its wakees.
5803 * In order to determine whether we should let the load spread vs consolidating
5804 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5805 * partner, and a factor of lls_size higher frequency in the other.
5807 * With both conditions met, we can be relatively sure that the relationship is
5808 * non-monogamous, with partner count exceeding socket size.
5810 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5811 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5814 static int wake_wide(struct task_struct
*p
)
5816 unsigned int master
= current
->wakee_flips
;
5817 unsigned int slave
= p
->wakee_flips
;
5818 int factor
= __this_cpu_read(sd_llc_size
);
5821 swap(master
, slave
);
5822 if (slave
< factor
|| master
< slave
* factor
)
5828 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5829 * soonest. For the purpose of speed we only consider the waking and previous
5832 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5833 * cache-affine and is (or will be) idle.
5835 * wake_affine_weight() - considers the weight to reflect the average
5836 * scheduling latency of the CPUs. This seems to work
5837 * for the overloaded case.
5840 wake_affine_idle(int this_cpu
, int prev_cpu
, int sync
)
5843 * If this_cpu is idle, it implies the wakeup is from interrupt
5844 * context. Only allow the move if cache is shared. Otherwise an
5845 * interrupt intensive workload could force all tasks onto one
5846 * node depending on the IO topology or IRQ affinity settings.
5848 * If the prev_cpu is idle and cache affine then avoid a migration.
5849 * There is no guarantee that the cache hot data from an interrupt
5850 * is more important than cache hot data on the prev_cpu and from
5851 * a cpufreq perspective, it's better to have higher utilisation
5854 if (available_idle_cpu(this_cpu
) && cpus_share_cache(this_cpu
, prev_cpu
))
5855 return available_idle_cpu(prev_cpu
) ? prev_cpu
: this_cpu
;
5857 if (sync
&& cpu_rq(this_cpu
)->nr_running
== 1)
5860 if (available_idle_cpu(prev_cpu
))
5863 return nr_cpumask_bits
;
5867 wake_affine_weight(struct sched_domain
*sd
, struct task_struct
*p
,
5868 int this_cpu
, int prev_cpu
, int sync
)
5870 s64 this_eff_load
, prev_eff_load
;
5871 unsigned long task_load
;
5873 this_eff_load
= cpu_load(cpu_rq(this_cpu
));
5876 unsigned long current_load
= task_h_load(current
);
5878 if (current_load
> this_eff_load
)
5881 this_eff_load
-= current_load
;
5884 task_load
= task_h_load(p
);
5886 this_eff_load
+= task_load
;
5887 if (sched_feat(WA_BIAS
))
5888 this_eff_load
*= 100;
5889 this_eff_load
*= capacity_of(prev_cpu
);
5891 prev_eff_load
= cpu_load(cpu_rq(prev_cpu
));
5892 prev_eff_load
-= task_load
;
5893 if (sched_feat(WA_BIAS
))
5894 prev_eff_load
*= 100 + (sd
->imbalance_pct
- 100) / 2;
5895 prev_eff_load
*= capacity_of(this_cpu
);
5898 * If sync, adjust the weight of prev_eff_load such that if
5899 * prev_eff == this_eff that select_idle_sibling() will consider
5900 * stacking the wakee on top of the waker if no other CPU is
5906 return this_eff_load
< prev_eff_load
? this_cpu
: nr_cpumask_bits
;
5909 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
,
5910 int this_cpu
, int prev_cpu
, int sync
)
5912 int target
= nr_cpumask_bits
;
5914 if (sched_feat(WA_IDLE
))
5915 target
= wake_affine_idle(this_cpu
, prev_cpu
, sync
);
5917 if (sched_feat(WA_WEIGHT
) && target
== nr_cpumask_bits
)
5918 target
= wake_affine_weight(sd
, p
, this_cpu
, prev_cpu
, sync
);
5920 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine_attempts
);
5921 if (target
== nr_cpumask_bits
)
5924 schedstat_inc(sd
->ttwu_move_affine
);
5925 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine
);
5929 static struct sched_group
*
5930 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
);
5933 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5936 find_idlest_group_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
5938 unsigned long load
, min_load
= ULONG_MAX
;
5939 unsigned int min_exit_latency
= UINT_MAX
;
5940 u64 latest_idle_timestamp
= 0;
5941 int least_loaded_cpu
= this_cpu
;
5942 int shallowest_idle_cpu
= -1;
5945 /* Check if we have any choice: */
5946 if (group
->group_weight
== 1)
5947 return cpumask_first(sched_group_span(group
));
5949 /* Traverse only the allowed CPUs */
5950 for_each_cpu_and(i
, sched_group_span(group
), p
->cpus_ptr
) {
5951 if (sched_idle_cpu(i
))
5954 if (available_idle_cpu(i
)) {
5955 struct rq
*rq
= cpu_rq(i
);
5956 struct cpuidle_state
*idle
= idle_get_state(rq
);
5957 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
5959 * We give priority to a CPU whose idle state
5960 * has the smallest exit latency irrespective
5961 * of any idle timestamp.
5963 min_exit_latency
= idle
->exit_latency
;
5964 latest_idle_timestamp
= rq
->idle_stamp
;
5965 shallowest_idle_cpu
= i
;
5966 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
5967 rq
->idle_stamp
> latest_idle_timestamp
) {
5969 * If equal or no active idle state, then
5970 * the most recently idled CPU might have
5973 latest_idle_timestamp
= rq
->idle_stamp
;
5974 shallowest_idle_cpu
= i
;
5976 } else if (shallowest_idle_cpu
== -1) {
5977 load
= cpu_load(cpu_rq(i
));
5978 if (load
< min_load
) {
5980 least_loaded_cpu
= i
;
5985 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
5988 static inline int find_idlest_cpu(struct sched_domain
*sd
, struct task_struct
*p
,
5989 int cpu
, int prev_cpu
, int sd_flag
)
5993 if (!cpumask_intersects(sched_domain_span(sd
), p
->cpus_ptr
))
5997 * We need task's util for cpu_util_without, sync it up to
5998 * prev_cpu's last_update_time.
6000 if (!(sd_flag
& SD_BALANCE_FORK
))
6001 sync_entity_load_avg(&p
->se
);
6004 struct sched_group
*group
;
6005 struct sched_domain
*tmp
;
6008 if (!(sd
->flags
& sd_flag
)) {
6013 group
= find_idlest_group(sd
, p
, cpu
);
6019 new_cpu
= find_idlest_group_cpu(group
, p
, cpu
);
6020 if (new_cpu
== cpu
) {
6021 /* Now try balancing at a lower domain level of 'cpu': */
6026 /* Now try balancing at a lower domain level of 'new_cpu': */
6028 weight
= sd
->span_weight
;
6030 for_each_domain(cpu
, tmp
) {
6031 if (weight
<= tmp
->span_weight
)
6033 if (tmp
->flags
& sd_flag
)
6041 #ifdef CONFIG_SCHED_SMT
6042 DEFINE_STATIC_KEY_FALSE(sched_smt_present
);
6043 EXPORT_SYMBOL_GPL(sched_smt_present
);
6045 static inline void set_idle_cores(int cpu
, int val
)
6047 struct sched_domain_shared
*sds
;
6049 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
6051 WRITE_ONCE(sds
->has_idle_cores
, val
);
6054 static inline bool test_idle_cores(int cpu
, bool def
)
6056 struct sched_domain_shared
*sds
;
6058 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
6060 return READ_ONCE(sds
->has_idle_cores
);
6066 * Scans the local SMT mask to see if the entire core is idle, and records this
6067 * information in sd_llc_shared->has_idle_cores.
6069 * Since SMT siblings share all cache levels, inspecting this limited remote
6070 * state should be fairly cheap.
6072 void __update_idle_core(struct rq
*rq
)
6074 int core
= cpu_of(rq
);
6078 if (test_idle_cores(core
, true))
6081 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
6085 if (!available_idle_cpu(cpu
))
6089 set_idle_cores(core
, 1);
6095 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6096 * there are no idle cores left in the system; tracked through
6097 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6099 static int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6101 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6104 if (!static_branch_likely(&sched_smt_present
))
6107 if (!test_idle_cores(target
, false))
6110 cpumask_and(cpus
, sched_domain_span(sd
), p
->cpus_ptr
);
6112 for_each_cpu_wrap(core
, cpus
, target
) {
6115 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
6116 if (!available_idle_cpu(cpu
)) {
6125 cpumask_andnot(cpus
, cpus
, cpu_smt_mask(core
));
6129 * Failed to find an idle core; stop looking for one.
6131 set_idle_cores(target
, 0);
6137 * Scan the local SMT mask for idle CPUs.
6139 static int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6143 if (!static_branch_likely(&sched_smt_present
))
6146 for_each_cpu(cpu
, cpu_smt_mask(target
)) {
6147 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
) ||
6148 !cpumask_test_cpu(cpu
, sched_domain_span(sd
)))
6150 if (available_idle_cpu(cpu
) || sched_idle_cpu(cpu
))
6157 #else /* CONFIG_SCHED_SMT */
6159 static inline int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6164 static inline int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6169 #endif /* CONFIG_SCHED_SMT */
6172 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6173 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6174 * average idle time for this rq (as found in rq->avg_idle).
6176 static int select_idle_cpu(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6178 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6179 struct sched_domain
*this_sd
;
6180 u64 avg_cost
, avg_idle
;
6182 int this = smp_processor_id();
6183 int cpu
, nr
= INT_MAX
;
6185 this_sd
= rcu_dereference(*this_cpu_ptr(&sd_llc
));
6190 * Due to large variance we need a large fuzz factor; hackbench in
6191 * particularly is sensitive here.
6193 avg_idle
= this_rq()->avg_idle
/ 512;
6194 avg_cost
= this_sd
->avg_scan_cost
+ 1;
6196 if (sched_feat(SIS_AVG_CPU
) && avg_idle
< avg_cost
)
6199 if (sched_feat(SIS_PROP
)) {
6200 u64 span_avg
= sd
->span_weight
* avg_idle
;
6201 if (span_avg
> 4*avg_cost
)
6202 nr
= div_u64(span_avg
, avg_cost
);
6207 time
= cpu_clock(this);
6209 cpumask_and(cpus
, sched_domain_span(sd
), p
->cpus_ptr
);
6211 for_each_cpu_wrap(cpu
, cpus
, target
) {
6214 if (available_idle_cpu(cpu
) || sched_idle_cpu(cpu
))
6218 time
= cpu_clock(this) - time
;
6219 update_avg(&this_sd
->avg_scan_cost
, time
);
6225 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6226 * the task fits. If no CPU is big enough, but there are idle ones, try to
6227 * maximize capacity.
6230 select_idle_capacity(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6232 unsigned long task_util
, best_cap
= 0;
6233 int cpu
, best_cpu
= -1;
6234 struct cpumask
*cpus
;
6236 cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6237 cpumask_and(cpus
, sched_domain_span(sd
), p
->cpus_ptr
);
6239 task_util
= uclamp_task_util(p
);
6241 for_each_cpu_wrap(cpu
, cpus
, target
) {
6242 unsigned long cpu_cap
= capacity_of(cpu
);
6244 if (!available_idle_cpu(cpu
) && !sched_idle_cpu(cpu
))
6246 if (fits_capacity(task_util
, cpu_cap
))
6249 if (cpu_cap
> best_cap
) {
6258 static inline bool asym_fits_capacity(int task_util
, int cpu
)
6260 if (static_branch_unlikely(&sched_asym_cpucapacity
))
6261 return fits_capacity(task_util
, capacity_of(cpu
));
6267 * Try and locate an idle core/thread in the LLC cache domain.
6269 static int select_idle_sibling(struct task_struct
*p
, int prev
, int target
)
6271 struct sched_domain
*sd
;
6272 unsigned long task_util
;
6273 int i
, recent_used_cpu
;
6276 * On asymmetric system, update task utilization because we will check
6277 * that the task fits with cpu's capacity.
6279 if (static_branch_unlikely(&sched_asym_cpucapacity
)) {
6280 sync_entity_load_avg(&p
->se
);
6281 task_util
= uclamp_task_util(p
);
6284 if ((available_idle_cpu(target
) || sched_idle_cpu(target
)) &&
6285 asym_fits_capacity(task_util
, target
))
6289 * If the previous CPU is cache affine and idle, don't be stupid:
6291 if (prev
!= target
&& cpus_share_cache(prev
, target
) &&
6292 (available_idle_cpu(prev
) || sched_idle_cpu(prev
)) &&
6293 asym_fits_capacity(task_util
, prev
))
6297 * Allow a per-cpu kthread to stack with the wakee if the
6298 * kworker thread and the tasks previous CPUs are the same.
6299 * The assumption is that the wakee queued work for the
6300 * per-cpu kthread that is now complete and the wakeup is
6301 * essentially a sync wakeup. An obvious example of this
6302 * pattern is IO completions.
6304 if (is_per_cpu_kthread(current
) &&
6305 prev
== smp_processor_id() &&
6306 this_rq()->nr_running
<= 1) {
6310 /* Check a recently used CPU as a potential idle candidate: */
6311 recent_used_cpu
= p
->recent_used_cpu
;
6312 if (recent_used_cpu
!= prev
&&
6313 recent_used_cpu
!= target
&&
6314 cpus_share_cache(recent_used_cpu
, target
) &&
6315 (available_idle_cpu(recent_used_cpu
) || sched_idle_cpu(recent_used_cpu
)) &&
6316 cpumask_test_cpu(p
->recent_used_cpu
, p
->cpus_ptr
) &&
6317 asym_fits_capacity(task_util
, recent_used_cpu
)) {
6319 * Replace recent_used_cpu with prev as it is a potential
6320 * candidate for the next wake:
6322 p
->recent_used_cpu
= prev
;
6323 return recent_used_cpu
;
6327 * For asymmetric CPU capacity systems, our domain of interest is
6328 * sd_asym_cpucapacity rather than sd_llc.
6330 if (static_branch_unlikely(&sched_asym_cpucapacity
)) {
6331 sd
= rcu_dereference(per_cpu(sd_asym_cpucapacity
, target
));
6333 * On an asymmetric CPU capacity system where an exclusive
6334 * cpuset defines a symmetric island (i.e. one unique
6335 * capacity_orig value through the cpuset), the key will be set
6336 * but the CPUs within that cpuset will not have a domain with
6337 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6341 i
= select_idle_capacity(p
, sd
, target
);
6342 return ((unsigned)i
< nr_cpumask_bits
) ? i
: target
;
6346 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
6350 i
= select_idle_core(p
, sd
, target
);
6351 if ((unsigned)i
< nr_cpumask_bits
)
6354 i
= select_idle_cpu(p
, sd
, target
);
6355 if ((unsigned)i
< nr_cpumask_bits
)
6358 i
= select_idle_smt(p
, sd
, target
);
6359 if ((unsigned)i
< nr_cpumask_bits
)
6366 * cpu_util - Estimates the amount of capacity of a CPU used by CFS tasks.
6367 * @cpu: the CPU to get the utilization of
6369 * The unit of the return value must be the one of capacity so we can compare
6370 * the utilization with the capacity of the CPU that is available for CFS task
6371 * (ie cpu_capacity).
6373 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6374 * recent utilization of currently non-runnable tasks on a CPU. It represents
6375 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6376 * capacity_orig is the cpu_capacity available at the highest frequency
6377 * (arch_scale_freq_capacity()).
6378 * The utilization of a CPU converges towards a sum equal to or less than the
6379 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6380 * the running time on this CPU scaled by capacity_curr.
6382 * The estimated utilization of a CPU is defined to be the maximum between its
6383 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6384 * currently RUNNABLE on that CPU.
6385 * This allows to properly represent the expected utilization of a CPU which
6386 * has just got a big task running since a long sleep period. At the same time
6387 * however it preserves the benefits of the "blocked utilization" in
6388 * describing the potential for other tasks waking up on the same CPU.
6390 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6391 * higher than capacity_orig because of unfortunate rounding in
6392 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6393 * the average stabilizes with the new running time. We need to check that the
6394 * utilization stays within the range of [0..capacity_orig] and cap it if
6395 * necessary. Without utilization capping, a group could be seen as overloaded
6396 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6397 * available capacity. We allow utilization to overshoot capacity_curr (but not
6398 * capacity_orig) as it useful for predicting the capacity required after task
6399 * migrations (scheduler-driven DVFS).
6401 * Return: the (estimated) utilization for the specified CPU
6403 static inline unsigned long cpu_util(int cpu
)
6405 struct cfs_rq
*cfs_rq
;
6408 cfs_rq
= &cpu_rq(cpu
)->cfs
;
6409 util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6411 if (sched_feat(UTIL_EST
))
6412 util
= max(util
, READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
));
6414 return min_t(unsigned long, util
, capacity_orig_of(cpu
));
6418 * cpu_util_without: compute cpu utilization without any contributions from *p
6419 * @cpu: the CPU which utilization is requested
6420 * @p: the task which utilization should be discounted
6422 * The utilization of a CPU is defined by the utilization of tasks currently
6423 * enqueued on that CPU as well as tasks which are currently sleeping after an
6424 * execution on that CPU.
6426 * This method returns the utilization of the specified CPU by discounting the
6427 * utilization of the specified task, whenever the task is currently
6428 * contributing to the CPU utilization.
6430 static unsigned long cpu_util_without(int cpu
, struct task_struct
*p
)
6432 struct cfs_rq
*cfs_rq
;
6435 /* Task has no contribution or is new */
6436 if (cpu
!= task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
6437 return cpu_util(cpu
);
6439 cfs_rq
= &cpu_rq(cpu
)->cfs
;
6440 util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6442 /* Discount task's util from CPU's util */
6443 lsub_positive(&util
, task_util(p
));
6448 * a) if *p is the only task sleeping on this CPU, then:
6449 * cpu_util (== task_util) > util_est (== 0)
6450 * and thus we return:
6451 * cpu_util_without = (cpu_util - task_util) = 0
6453 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6455 * cpu_util >= task_util
6456 * cpu_util > util_est (== 0)
6457 * and thus we discount *p's blocked utilization to return:
6458 * cpu_util_without = (cpu_util - task_util) >= 0
6460 * c) if other tasks are RUNNABLE on that CPU and
6461 * util_est > cpu_util
6462 * then we use util_est since it returns a more restrictive
6463 * estimation of the spare capacity on that CPU, by just
6464 * considering the expected utilization of tasks already
6465 * runnable on that CPU.
6467 * Cases a) and b) are covered by the above code, while case c) is
6468 * covered by the following code when estimated utilization is
6471 if (sched_feat(UTIL_EST
)) {
6472 unsigned int estimated
=
6473 READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
);
6476 * Despite the following checks we still have a small window
6477 * for a possible race, when an execl's select_task_rq_fair()
6478 * races with LB's detach_task():
6481 * p->on_rq = TASK_ON_RQ_MIGRATING;
6482 * ---------------------------------- A
6483 * deactivate_task() \
6484 * dequeue_task() + RaceTime
6485 * util_est_dequeue() /
6486 * ---------------------------------- B
6488 * The additional check on "current == p" it's required to
6489 * properly fix the execl regression and it helps in further
6490 * reducing the chances for the above race.
6492 if (unlikely(task_on_rq_queued(p
) || current
== p
))
6493 lsub_positive(&estimated
, _task_util_est(p
));
6495 util
= max(util
, estimated
);
6499 * Utilization (estimated) can exceed the CPU capacity, thus let's
6500 * clamp to the maximum CPU capacity to ensure consistency with
6501 * the cpu_util call.
6503 return min_t(unsigned long, util
, capacity_orig_of(cpu
));
6507 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6510 static unsigned long cpu_util_next(int cpu
, struct task_struct
*p
, int dst_cpu
)
6512 struct cfs_rq
*cfs_rq
= &cpu_rq(cpu
)->cfs
;
6513 unsigned long util_est
, util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6516 * If @p migrates from @cpu to another, remove its contribution. Or,
6517 * if @p migrates from another CPU to @cpu, add its contribution. In
6518 * the other cases, @cpu is not impacted by the migration, so the
6519 * util_avg should already be correct.
6521 if (task_cpu(p
) == cpu
&& dst_cpu
!= cpu
)
6522 sub_positive(&util
, task_util(p
));
6523 else if (task_cpu(p
) != cpu
&& dst_cpu
== cpu
)
6524 util
+= task_util(p
);
6526 if (sched_feat(UTIL_EST
)) {
6527 util_est
= READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
);
6530 * During wake-up, the task isn't enqueued yet and doesn't
6531 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6532 * so just add it (if needed) to "simulate" what will be
6533 * cpu_util() after the task has been enqueued.
6536 util_est
+= _task_util_est(p
);
6538 util
= max(util
, util_est
);
6541 return min(util
, capacity_orig_of(cpu
));
6545 * compute_energy(): Estimates the energy that @pd would consume if @p was
6546 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6547 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6548 * to compute what would be the energy if we decided to actually migrate that
6552 compute_energy(struct task_struct
*p
, int dst_cpu
, struct perf_domain
*pd
)
6554 struct cpumask
*pd_mask
= perf_domain_span(pd
);
6555 unsigned long cpu_cap
= arch_scale_cpu_capacity(cpumask_first(pd_mask
));
6556 unsigned long max_util
= 0, sum_util
= 0;
6560 * The capacity state of CPUs of the current rd can be driven by CPUs
6561 * of another rd if they belong to the same pd. So, account for the
6562 * utilization of these CPUs too by masking pd with cpu_online_mask
6563 * instead of the rd span.
6565 * If an entire pd is outside of the current rd, it will not appear in
6566 * its pd list and will not be accounted by compute_energy().
6568 for_each_cpu_and(cpu
, pd_mask
, cpu_online_mask
) {
6569 unsigned long cpu_util
, util_cfs
= cpu_util_next(cpu
, p
, dst_cpu
);
6570 struct task_struct
*tsk
= cpu
== dst_cpu
? p
: NULL
;
6573 * Busy time computation: utilization clamping is not
6574 * required since the ratio (sum_util / cpu_capacity)
6575 * is already enough to scale the EM reported power
6576 * consumption at the (eventually clamped) cpu_capacity.
6578 sum_util
+= schedutil_cpu_util(cpu
, util_cfs
, cpu_cap
,
6582 * Performance domain frequency: utilization clamping
6583 * must be considered since it affects the selection
6584 * of the performance domain frequency.
6585 * NOTE: in case RT tasks are running, by default the
6586 * FREQUENCY_UTIL's utilization can be max OPP.
6588 cpu_util
= schedutil_cpu_util(cpu
, util_cfs
, cpu_cap
,
6589 FREQUENCY_UTIL
, tsk
);
6590 max_util
= max(max_util
, cpu_util
);
6593 return em_cpu_energy(pd
->em_pd
, max_util
, sum_util
);
6597 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6598 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6599 * spare capacity in each performance domain and uses it as a potential
6600 * candidate to execute the task. Then, it uses the Energy Model to figure
6601 * out which of the CPU candidates is the most energy-efficient.
6603 * The rationale for this heuristic is as follows. In a performance domain,
6604 * all the most energy efficient CPU candidates (according to the Energy
6605 * Model) are those for which we'll request a low frequency. When there are
6606 * several CPUs for which the frequency request will be the same, we don't
6607 * have enough data to break the tie between them, because the Energy Model
6608 * only includes active power costs. With this model, if we assume that
6609 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6610 * the maximum spare capacity in a performance domain is guaranteed to be among
6611 * the best candidates of the performance domain.
6613 * In practice, it could be preferable from an energy standpoint to pack
6614 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6615 * but that could also hurt our chances to go cluster idle, and we have no
6616 * ways to tell with the current Energy Model if this is actually a good
6617 * idea or not. So, find_energy_efficient_cpu() basically favors
6618 * cluster-packing, and spreading inside a cluster. That should at least be
6619 * a good thing for latency, and this is consistent with the idea that most
6620 * of the energy savings of EAS come from the asymmetry of the system, and
6621 * not so much from breaking the tie between identical CPUs. That's also the
6622 * reason why EAS is enabled in the topology code only for systems where
6623 * SD_ASYM_CPUCAPACITY is set.
6625 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6626 * they don't have any useful utilization data yet and it's not possible to
6627 * forecast their impact on energy consumption. Consequently, they will be
6628 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6629 * to be energy-inefficient in some use-cases. The alternative would be to
6630 * bias new tasks towards specific types of CPUs first, or to try to infer
6631 * their util_avg from the parent task, but those heuristics could hurt
6632 * other use-cases too. So, until someone finds a better way to solve this,
6633 * let's keep things simple by re-using the existing slow path.
6635 static int find_energy_efficient_cpu(struct task_struct
*p
, int prev_cpu
)
6637 unsigned long prev_delta
= ULONG_MAX
, best_delta
= ULONG_MAX
;
6638 struct root_domain
*rd
= cpu_rq(smp_processor_id())->rd
;
6639 unsigned long cpu_cap
, util
, base_energy
= 0;
6640 int cpu
, best_energy_cpu
= prev_cpu
;
6641 struct sched_domain
*sd
;
6642 struct perf_domain
*pd
;
6645 pd
= rcu_dereference(rd
->pd
);
6646 if (!pd
|| READ_ONCE(rd
->overutilized
))
6650 * Energy-aware wake-up happens on the lowest sched_domain starting
6651 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6653 sd
= rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity
));
6654 while (sd
&& !cpumask_test_cpu(prev_cpu
, sched_domain_span(sd
)))
6659 sync_entity_load_avg(&p
->se
);
6660 if (!task_util_est(p
))
6663 for (; pd
; pd
= pd
->next
) {
6664 unsigned long cur_delta
, spare_cap
, max_spare_cap
= 0;
6665 unsigned long base_energy_pd
;
6666 int max_spare_cap_cpu
= -1;
6668 /* Compute the 'base' energy of the pd, without @p */
6669 base_energy_pd
= compute_energy(p
, -1, pd
);
6670 base_energy
+= base_energy_pd
;
6672 for_each_cpu_and(cpu
, perf_domain_span(pd
), sched_domain_span(sd
)) {
6673 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
6676 util
= cpu_util_next(cpu
, p
, cpu
);
6677 cpu_cap
= capacity_of(cpu
);
6678 spare_cap
= cpu_cap
;
6679 lsub_positive(&spare_cap
, util
);
6682 * Skip CPUs that cannot satisfy the capacity request.
6683 * IOW, placing the task there would make the CPU
6684 * overutilized. Take uclamp into account to see how
6685 * much capacity we can get out of the CPU; this is
6686 * aligned with schedutil_cpu_util().
6688 util
= uclamp_rq_util_with(cpu_rq(cpu
), util
, p
);
6689 if (!fits_capacity(util
, cpu_cap
))
6692 /* Always use prev_cpu as a candidate. */
6693 if (cpu
== prev_cpu
) {
6694 prev_delta
= compute_energy(p
, prev_cpu
, pd
);
6695 prev_delta
-= base_energy_pd
;
6696 best_delta
= min(best_delta
, prev_delta
);
6700 * Find the CPU with the maximum spare capacity in
6701 * the performance domain
6703 if (spare_cap
> max_spare_cap
) {
6704 max_spare_cap
= spare_cap
;
6705 max_spare_cap_cpu
= cpu
;
6709 /* Evaluate the energy impact of using this CPU. */
6710 if (max_spare_cap_cpu
>= 0 && max_spare_cap_cpu
!= prev_cpu
) {
6711 cur_delta
= compute_energy(p
, max_spare_cap_cpu
, pd
);
6712 cur_delta
-= base_energy_pd
;
6713 if (cur_delta
< best_delta
) {
6714 best_delta
= cur_delta
;
6715 best_energy_cpu
= max_spare_cap_cpu
;
6723 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6724 * least 6% of the energy used by prev_cpu.
6726 if (prev_delta
== ULONG_MAX
)
6727 return best_energy_cpu
;
6729 if ((prev_delta
- best_delta
) > ((prev_delta
+ base_energy
) >> 4))
6730 return best_energy_cpu
;
6741 * select_task_rq_fair: Select target runqueue for the waking task in domains
6742 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
6743 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6745 * Balances load by selecting the idlest CPU in the idlest group, or under
6746 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6748 * Returns the target CPU number.
6750 * preempt must be disabled.
6753 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int wake_flags
)
6755 int sync
= (wake_flags
& WF_SYNC
) && !(current
->flags
& PF_EXITING
);
6756 struct sched_domain
*tmp
, *sd
= NULL
;
6757 int cpu
= smp_processor_id();
6758 int new_cpu
= prev_cpu
;
6759 int want_affine
= 0;
6760 /* SD_flags and WF_flags share the first nibble */
6761 int sd_flag
= wake_flags
& 0xF;
6763 if (wake_flags
& WF_TTWU
) {
6766 if (sched_energy_enabled()) {
6767 new_cpu
= find_energy_efficient_cpu(p
, prev_cpu
);
6773 want_affine
= !wake_wide(p
) && cpumask_test_cpu(cpu
, p
->cpus_ptr
);
6777 for_each_domain(cpu
, tmp
) {
6779 * If both 'cpu' and 'prev_cpu' are part of this domain,
6780 * cpu is a valid SD_WAKE_AFFINE target.
6782 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
6783 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
6784 if (cpu
!= prev_cpu
)
6785 new_cpu
= wake_affine(tmp
, p
, cpu
, prev_cpu
, sync
);
6787 sd
= NULL
; /* Prefer wake_affine over balance flags */
6791 if (tmp
->flags
& sd_flag
)
6793 else if (!want_affine
)
6799 new_cpu
= find_idlest_cpu(sd
, p
, cpu
, prev_cpu
, sd_flag
);
6800 } else if (wake_flags
& WF_TTWU
) { /* XXX always ? */
6802 new_cpu
= select_idle_sibling(p
, prev_cpu
, new_cpu
);
6805 current
->recent_used_cpu
= cpu
;
6812 static void detach_entity_cfs_rq(struct sched_entity
*se
);
6815 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6816 * cfs_rq_of(p) references at time of call are still valid and identify the
6817 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6819 static void migrate_task_rq_fair(struct task_struct
*p
, int new_cpu
)
6822 * As blocked tasks retain absolute vruntime the migration needs to
6823 * deal with this by subtracting the old and adding the new
6824 * min_vruntime -- the latter is done by enqueue_entity() when placing
6825 * the task on the new runqueue.
6827 if (p
->state
== TASK_WAKING
) {
6828 struct sched_entity
*se
= &p
->se
;
6829 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6832 #ifndef CONFIG_64BIT
6833 u64 min_vruntime_copy
;
6836 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
6838 min_vruntime
= cfs_rq
->min_vruntime
;
6839 } while (min_vruntime
!= min_vruntime_copy
);
6841 min_vruntime
= cfs_rq
->min_vruntime
;
6844 se
->vruntime
-= min_vruntime
;
6847 if (p
->on_rq
== TASK_ON_RQ_MIGRATING
) {
6849 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6850 * rq->lock and can modify state directly.
6852 lockdep_assert_held(&task_rq(p
)->lock
);
6853 detach_entity_cfs_rq(&p
->se
);
6857 * We are supposed to update the task to "current" time, then
6858 * its up to date and ready to go to new CPU/cfs_rq. But we
6859 * have difficulty in getting what current time is, so simply
6860 * throw away the out-of-date time. This will result in the
6861 * wakee task is less decayed, but giving the wakee more load
6864 remove_entity_load_avg(&p
->se
);
6867 /* Tell new CPU we are migrated */
6868 p
->se
.avg
.last_update_time
= 0;
6870 /* We have migrated, no longer consider this task hot */
6871 p
->se
.exec_start
= 0;
6873 update_scan_period(p
, new_cpu
);
6876 static void task_dead_fair(struct task_struct
*p
)
6878 remove_entity_load_avg(&p
->se
);
6882 balance_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
6887 return newidle_balance(rq
, rf
) != 0;
6889 #endif /* CONFIG_SMP */
6891 static unsigned long wakeup_gran(struct sched_entity
*se
)
6893 unsigned long gran
= sysctl_sched_wakeup_granularity
;
6896 * Since its curr running now, convert the gran from real-time
6897 * to virtual-time in his units.
6899 * By using 'se' instead of 'curr' we penalize light tasks, so
6900 * they get preempted easier. That is, if 'se' < 'curr' then
6901 * the resulting gran will be larger, therefore penalizing the
6902 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6903 * be smaller, again penalizing the lighter task.
6905 * This is especially important for buddies when the leftmost
6906 * task is higher priority than the buddy.
6908 return calc_delta_fair(gran
, se
);
6912 * Should 'se' preempt 'curr'.
6926 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
6928 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
6933 gran
= wakeup_gran(se
);
6940 static void set_last_buddy(struct sched_entity
*se
)
6942 if (entity_is_task(se
) && unlikely(task_has_idle_policy(task_of(se
))))
6945 for_each_sched_entity(se
) {
6946 if (SCHED_WARN_ON(!se
->on_rq
))
6948 cfs_rq_of(se
)->last
= se
;
6952 static void set_next_buddy(struct sched_entity
*se
)
6954 if (entity_is_task(se
) && unlikely(task_has_idle_policy(task_of(se
))))
6957 for_each_sched_entity(se
) {
6958 if (SCHED_WARN_ON(!se
->on_rq
))
6960 cfs_rq_of(se
)->next
= se
;
6964 static void set_skip_buddy(struct sched_entity
*se
)
6966 for_each_sched_entity(se
)
6967 cfs_rq_of(se
)->skip
= se
;
6971 * Preempt the current task with a newly woken task if needed:
6973 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
6975 struct task_struct
*curr
= rq
->curr
;
6976 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
6977 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
6978 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
6979 int next_buddy_marked
= 0;
6981 if (unlikely(se
== pse
))
6985 * This is possible from callers such as attach_tasks(), in which we
6986 * unconditionally check_prempt_curr() after an enqueue (which may have
6987 * lead to a throttle). This both saves work and prevents false
6988 * next-buddy nomination below.
6990 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
6993 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
6994 set_next_buddy(pse
);
6995 next_buddy_marked
= 1;
6999 * We can come here with TIF_NEED_RESCHED already set from new task
7002 * Note: this also catches the edge-case of curr being in a throttled
7003 * group (e.g. via set_curr_task), since update_curr() (in the
7004 * enqueue of curr) will have resulted in resched being set. This
7005 * prevents us from potentially nominating it as a false LAST_BUDDY
7008 if (test_tsk_need_resched(curr
))
7011 /* Idle tasks are by definition preempted by non-idle tasks. */
7012 if (unlikely(task_has_idle_policy(curr
)) &&
7013 likely(!task_has_idle_policy(p
)))
7017 * Batch and idle tasks do not preempt non-idle tasks (their preemption
7018 * is driven by the tick):
7020 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
7023 find_matching_se(&se
, &pse
);
7024 update_curr(cfs_rq_of(se
));
7026 if (wakeup_preempt_entity(se
, pse
) == 1) {
7028 * Bias pick_next to pick the sched entity that is
7029 * triggering this preemption.
7031 if (!next_buddy_marked
)
7032 set_next_buddy(pse
);
7041 * Only set the backward buddy when the current task is still
7042 * on the rq. This can happen when a wakeup gets interleaved
7043 * with schedule on the ->pre_schedule() or idle_balance()
7044 * point, either of which can * drop the rq lock.
7046 * Also, during early boot the idle thread is in the fair class,
7047 * for obvious reasons its a bad idea to schedule back to it.
7049 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
7052 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
7056 struct task_struct
*
7057 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
7059 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
7060 struct sched_entity
*se
;
7061 struct task_struct
*p
;
7065 if (!sched_fair_runnable(rq
))
7068 #ifdef CONFIG_FAIR_GROUP_SCHED
7069 if (!prev
|| prev
->sched_class
!= &fair_sched_class
)
7073 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7074 * likely that a next task is from the same cgroup as the current.
7076 * Therefore attempt to avoid putting and setting the entire cgroup
7077 * hierarchy, only change the part that actually changes.
7081 struct sched_entity
*curr
= cfs_rq
->curr
;
7084 * Since we got here without doing put_prev_entity() we also
7085 * have to consider cfs_rq->curr. If it is still a runnable
7086 * entity, update_curr() will update its vruntime, otherwise
7087 * forget we've ever seen it.
7091 update_curr(cfs_rq
);
7096 * This call to check_cfs_rq_runtime() will do the
7097 * throttle and dequeue its entity in the parent(s).
7098 * Therefore the nr_running test will indeed
7101 if (unlikely(check_cfs_rq_runtime(cfs_rq
))) {
7104 if (!cfs_rq
->nr_running
)
7111 se
= pick_next_entity(cfs_rq
, curr
);
7112 cfs_rq
= group_cfs_rq(se
);
7118 * Since we haven't yet done put_prev_entity and if the selected task
7119 * is a different task than we started out with, try and touch the
7120 * least amount of cfs_rqs.
7123 struct sched_entity
*pse
= &prev
->se
;
7125 while (!(cfs_rq
= is_same_group(se
, pse
))) {
7126 int se_depth
= se
->depth
;
7127 int pse_depth
= pse
->depth
;
7129 if (se_depth
<= pse_depth
) {
7130 put_prev_entity(cfs_rq_of(pse
), pse
);
7131 pse
= parent_entity(pse
);
7133 if (se_depth
>= pse_depth
) {
7134 set_next_entity(cfs_rq_of(se
), se
);
7135 se
= parent_entity(se
);
7139 put_prev_entity(cfs_rq
, pse
);
7140 set_next_entity(cfs_rq
, se
);
7147 put_prev_task(rq
, prev
);
7150 se
= pick_next_entity(cfs_rq
, NULL
);
7151 set_next_entity(cfs_rq
, se
);
7152 cfs_rq
= group_cfs_rq(se
);
7157 done
: __maybe_unused
;
7160 * Move the next running task to the front of
7161 * the list, so our cfs_tasks list becomes MRU
7164 list_move(&p
->se
.group_node
, &rq
->cfs_tasks
);
7167 if (hrtick_enabled(rq
))
7168 hrtick_start_fair(rq
, p
);
7170 update_misfit_status(p
, rq
);
7178 new_tasks
= newidle_balance(rq
, rf
);
7181 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7182 * possible for any higher priority task to appear. In that case we
7183 * must re-start the pick_next_entity() loop.
7192 * rq is about to be idle, check if we need to update the
7193 * lost_idle_time of clock_pelt
7195 update_idle_rq_clock_pelt(rq
);
7200 static struct task_struct
*__pick_next_task_fair(struct rq
*rq
)
7202 return pick_next_task_fair(rq
, NULL
, NULL
);
7206 * Account for a descheduled task:
7208 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
7210 struct sched_entity
*se
= &prev
->se
;
7211 struct cfs_rq
*cfs_rq
;
7213 for_each_sched_entity(se
) {
7214 cfs_rq
= cfs_rq_of(se
);
7215 put_prev_entity(cfs_rq
, se
);
7220 * sched_yield() is very simple
7222 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7224 static void yield_task_fair(struct rq
*rq
)
7226 struct task_struct
*curr
= rq
->curr
;
7227 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
7228 struct sched_entity
*se
= &curr
->se
;
7231 * Are we the only task in the tree?
7233 if (unlikely(rq
->nr_running
== 1))
7236 clear_buddies(cfs_rq
, se
);
7238 if (curr
->policy
!= SCHED_BATCH
) {
7239 update_rq_clock(rq
);
7241 * Update run-time statistics of the 'current'.
7243 update_curr(cfs_rq
);
7245 * Tell update_rq_clock() that we've just updated,
7246 * so we don't do microscopic update in schedule()
7247 * and double the fastpath cost.
7249 rq_clock_skip_update(rq
);
7255 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
)
7257 struct sched_entity
*se
= &p
->se
;
7259 /* throttled hierarchies are not runnable */
7260 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
7263 /* Tell the scheduler that we'd really like pse to run next. */
7266 yield_task_fair(rq
);
7272 /**************************************************
7273 * Fair scheduling class load-balancing methods.
7277 * The purpose of load-balancing is to achieve the same basic fairness the
7278 * per-CPU scheduler provides, namely provide a proportional amount of compute
7279 * time to each task. This is expressed in the following equation:
7281 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7283 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7284 * W_i,0 is defined as:
7286 * W_i,0 = \Sum_j w_i,j (2)
7288 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7289 * is derived from the nice value as per sched_prio_to_weight[].
7291 * The weight average is an exponential decay average of the instantaneous
7294 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7296 * C_i is the compute capacity of CPU i, typically it is the
7297 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7298 * can also include other factors [XXX].
7300 * To achieve this balance we define a measure of imbalance which follows
7301 * directly from (1):
7303 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7305 * We them move tasks around to minimize the imbalance. In the continuous
7306 * function space it is obvious this converges, in the discrete case we get
7307 * a few fun cases generally called infeasible weight scenarios.
7310 * - infeasible weights;
7311 * - local vs global optima in the discrete case. ]
7316 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7317 * for all i,j solution, we create a tree of CPUs that follows the hardware
7318 * topology where each level pairs two lower groups (or better). This results
7319 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7320 * tree to only the first of the previous level and we decrease the frequency
7321 * of load-balance at each level inv. proportional to the number of CPUs in
7327 * \Sum { --- * --- * 2^i } = O(n) (5)
7329 * `- size of each group
7330 * | | `- number of CPUs doing load-balance
7332 * `- sum over all levels
7334 * Coupled with a limit on how many tasks we can migrate every balance pass,
7335 * this makes (5) the runtime complexity of the balancer.
7337 * An important property here is that each CPU is still (indirectly) connected
7338 * to every other CPU in at most O(log n) steps:
7340 * The adjacency matrix of the resulting graph is given by:
7343 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7346 * And you'll find that:
7348 * A^(log_2 n)_i,j != 0 for all i,j (7)
7350 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7351 * The task movement gives a factor of O(m), giving a convergence complexity
7354 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7359 * In order to avoid CPUs going idle while there's still work to do, new idle
7360 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7361 * tree itself instead of relying on other CPUs to bring it work.
7363 * This adds some complexity to both (5) and (8) but it reduces the total idle
7371 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7374 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7379 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7381 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7383 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7386 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7387 * rewrite all of this once again.]
7390 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
7392 enum fbq_type
{ regular
, remote
, all
};
7395 * 'group_type' describes the group of CPUs at the moment of load balancing.
7397 * The enum is ordered by pulling priority, with the group with lowest priority
7398 * first so the group_type can simply be compared when selecting the busiest
7399 * group. See update_sd_pick_busiest().
7402 /* The group has spare capacity that can be used to run more tasks. */
7403 group_has_spare
= 0,
7405 * The group is fully used and the tasks don't compete for more CPU
7406 * cycles. Nevertheless, some tasks might wait before running.
7410 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
7411 * and must be migrated to a more powerful CPU.
7415 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7416 * and the task should be migrated to it instead of running on the
7421 * The tasks' affinity constraints previously prevented the scheduler
7422 * from balancing the load across the system.
7426 * The CPU is overloaded and can't provide expected CPU cycles to all
7432 enum migration_type
{
7439 #define LBF_ALL_PINNED 0x01
7440 #define LBF_NEED_BREAK 0x02
7441 #define LBF_DST_PINNED 0x04
7442 #define LBF_SOME_PINNED 0x08
7443 #define LBF_NOHZ_STATS 0x10
7444 #define LBF_NOHZ_AGAIN 0x20
7447 struct sched_domain
*sd
;
7455 struct cpumask
*dst_grpmask
;
7457 enum cpu_idle_type idle
;
7459 /* The set of CPUs under consideration for load-balancing */
7460 struct cpumask
*cpus
;
7465 unsigned int loop_break
;
7466 unsigned int loop_max
;
7468 enum fbq_type fbq_type
;
7469 enum migration_type migration_type
;
7470 struct list_head tasks
;
7474 * Is this task likely cache-hot:
7476 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
7480 lockdep_assert_held(&env
->src_rq
->lock
);
7482 if (p
->sched_class
!= &fair_sched_class
)
7485 if (unlikely(task_has_idle_policy(p
)))
7488 /* SMT siblings share cache */
7489 if (env
->sd
->flags
& SD_SHARE_CPUCAPACITY
)
7493 * Buddy candidates are cache hot:
7495 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
7496 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
7497 &p
->se
== cfs_rq_of(&p
->se
)->last
))
7500 if (sysctl_sched_migration_cost
== -1)
7502 if (sysctl_sched_migration_cost
== 0)
7505 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
7507 return delta
< (s64
)sysctl_sched_migration_cost
;
7510 #ifdef CONFIG_NUMA_BALANCING
7512 * Returns 1, if task migration degrades locality
7513 * Returns 0, if task migration improves locality i.e migration preferred.
7514 * Returns -1, if task migration is not affected by locality.
7516 static int migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
7518 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
7519 unsigned long src_weight
, dst_weight
;
7520 int src_nid
, dst_nid
, dist
;
7522 if (!static_branch_likely(&sched_numa_balancing
))
7525 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
7528 src_nid
= cpu_to_node(env
->src_cpu
);
7529 dst_nid
= cpu_to_node(env
->dst_cpu
);
7531 if (src_nid
== dst_nid
)
7534 /* Migrating away from the preferred node is always bad. */
7535 if (src_nid
== p
->numa_preferred_nid
) {
7536 if (env
->src_rq
->nr_running
> env
->src_rq
->nr_preferred_running
)
7542 /* Encourage migration to the preferred node. */
7543 if (dst_nid
== p
->numa_preferred_nid
)
7546 /* Leaving a core idle is often worse than degrading locality. */
7547 if (env
->idle
== CPU_IDLE
)
7550 dist
= node_distance(src_nid
, dst_nid
);
7552 src_weight
= group_weight(p
, src_nid
, dist
);
7553 dst_weight
= group_weight(p
, dst_nid
, dist
);
7555 src_weight
= task_weight(p
, src_nid
, dist
);
7556 dst_weight
= task_weight(p
, dst_nid
, dist
);
7559 return dst_weight
< src_weight
;
7563 static inline int migrate_degrades_locality(struct task_struct
*p
,
7571 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7574 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
7578 lockdep_assert_held(&env
->src_rq
->lock
);
7581 * We do not migrate tasks that are:
7582 * 1) throttled_lb_pair, or
7583 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7584 * 3) running (obviously), or
7585 * 4) are cache-hot on their current CPU.
7587 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
7590 /* Disregard pcpu kthreads; they are where they need to be. */
7591 if (kthread_is_per_cpu(p
))
7594 if (!cpumask_test_cpu(env
->dst_cpu
, p
->cpus_ptr
)) {
7597 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_affine
);
7599 env
->flags
|= LBF_SOME_PINNED
;
7602 * Remember if this task can be migrated to any other CPU in
7603 * our sched_group. We may want to revisit it if we couldn't
7604 * meet load balance goals by pulling other tasks on src_cpu.
7606 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7607 * already computed one in current iteration.
7609 if (env
->idle
== CPU_NEWLY_IDLE
|| (env
->flags
& LBF_DST_PINNED
))
7612 /* Prevent to re-select dst_cpu via env's CPUs: */
7613 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
7614 if (cpumask_test_cpu(cpu
, p
->cpus_ptr
)) {
7615 env
->flags
|= LBF_DST_PINNED
;
7616 env
->new_dst_cpu
= cpu
;
7624 /* Record that we found atleast one task that could run on dst_cpu */
7625 env
->flags
&= ~LBF_ALL_PINNED
;
7627 if (task_running(env
->src_rq
, p
)) {
7628 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_running
);
7633 * Aggressive migration if:
7634 * 1) destination numa is preferred
7635 * 2) task is cache cold, or
7636 * 3) too many balance attempts have failed.
7638 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
7639 if (tsk_cache_hot
== -1)
7640 tsk_cache_hot
= task_hot(p
, env
);
7642 if (tsk_cache_hot
<= 0 ||
7643 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
7644 if (tsk_cache_hot
== 1) {
7645 schedstat_inc(env
->sd
->lb_hot_gained
[env
->idle
]);
7646 schedstat_inc(p
->se
.statistics
.nr_forced_migrations
);
7651 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_hot
);
7656 * detach_task() -- detach the task for the migration specified in env
7658 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
7660 lockdep_assert_held(&env
->src_rq
->lock
);
7662 deactivate_task(env
->src_rq
, p
, DEQUEUE_NOCLOCK
);
7663 set_task_cpu(p
, env
->dst_cpu
);
7667 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7668 * part of active balancing operations within "domain".
7670 * Returns a task if successful and NULL otherwise.
7672 static struct task_struct
*detach_one_task(struct lb_env
*env
)
7674 struct task_struct
*p
;
7676 lockdep_assert_held(&env
->src_rq
->lock
);
7678 list_for_each_entry_reverse(p
,
7679 &env
->src_rq
->cfs_tasks
, se
.group_node
) {
7680 if (!can_migrate_task(p
, env
))
7683 detach_task(p
, env
);
7686 * Right now, this is only the second place where
7687 * lb_gained[env->idle] is updated (other is detach_tasks)
7688 * so we can safely collect stats here rather than
7689 * inside detach_tasks().
7691 schedstat_inc(env
->sd
->lb_gained
[env
->idle
]);
7697 static const unsigned int sched_nr_migrate_break
= 32;
7700 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7701 * busiest_rq, as part of a balancing operation within domain "sd".
7703 * Returns number of detached tasks if successful and 0 otherwise.
7705 static int detach_tasks(struct lb_env
*env
)
7707 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
7708 unsigned long util
, load
;
7709 struct task_struct
*p
;
7712 lockdep_assert_held(&env
->src_rq
->lock
);
7714 if (env
->imbalance
<= 0)
7717 while (!list_empty(tasks
)) {
7719 * We don't want to steal all, otherwise we may be treated likewise,
7720 * which could at worst lead to a livelock crash.
7722 if (env
->idle
!= CPU_NOT_IDLE
&& env
->src_rq
->nr_running
<= 1)
7725 p
= list_last_entry(tasks
, struct task_struct
, se
.group_node
);
7728 /* We've more or less seen every task there is, call it quits */
7729 if (env
->loop
> env
->loop_max
)
7732 /* take a breather every nr_migrate tasks */
7733 if (env
->loop
> env
->loop_break
) {
7734 env
->loop_break
+= sched_nr_migrate_break
;
7735 env
->flags
|= LBF_NEED_BREAK
;
7739 if (!can_migrate_task(p
, env
))
7742 switch (env
->migration_type
) {
7745 * Depending of the number of CPUs and tasks and the
7746 * cgroup hierarchy, task_h_load() can return a null
7747 * value. Make sure that env->imbalance decreases
7748 * otherwise detach_tasks() will stop only after
7749 * detaching up to loop_max tasks.
7751 load
= max_t(unsigned long, task_h_load(p
), 1);
7753 if (sched_feat(LB_MIN
) &&
7754 load
< 16 && !env
->sd
->nr_balance_failed
)
7758 * Make sure that we don't migrate too much load.
7759 * Nevertheless, let relax the constraint if
7760 * scheduler fails to find a good waiting task to
7763 if (shr_bound(load
, env
->sd
->nr_balance_failed
) > env
->imbalance
)
7766 env
->imbalance
-= load
;
7770 util
= task_util_est(p
);
7772 if (util
> env
->imbalance
)
7775 env
->imbalance
-= util
;
7782 case migrate_misfit
:
7783 /* This is not a misfit task */
7784 if (task_fits_capacity(p
, capacity_of(env
->src_cpu
)))
7791 detach_task(p
, env
);
7792 list_add(&p
->se
.group_node
, &env
->tasks
);
7796 #ifdef CONFIG_PREEMPTION
7798 * NEWIDLE balancing is a source of latency, so preemptible
7799 * kernels will stop after the first task is detached to minimize
7800 * the critical section.
7802 if (env
->idle
== CPU_NEWLY_IDLE
)
7807 * We only want to steal up to the prescribed amount of
7810 if (env
->imbalance
<= 0)
7815 list_move(&p
->se
.group_node
, tasks
);
7819 * Right now, this is one of only two places we collect this stat
7820 * so we can safely collect detach_one_task() stats here rather
7821 * than inside detach_one_task().
7823 schedstat_add(env
->sd
->lb_gained
[env
->idle
], detached
);
7829 * attach_task() -- attach the task detached by detach_task() to its new rq.
7831 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
7833 lockdep_assert_held(&rq
->lock
);
7835 BUG_ON(task_rq(p
) != rq
);
7836 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
7837 check_preempt_curr(rq
, p
, 0);
7841 * attach_one_task() -- attaches the task returned from detach_one_task() to
7844 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
7849 update_rq_clock(rq
);
7855 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7858 static void attach_tasks(struct lb_env
*env
)
7860 struct list_head
*tasks
= &env
->tasks
;
7861 struct task_struct
*p
;
7864 rq_lock(env
->dst_rq
, &rf
);
7865 update_rq_clock(env
->dst_rq
);
7867 while (!list_empty(tasks
)) {
7868 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
7869 list_del_init(&p
->se
.group_node
);
7871 attach_task(env
->dst_rq
, p
);
7874 rq_unlock(env
->dst_rq
, &rf
);
7877 #ifdef CONFIG_NO_HZ_COMMON
7878 static inline bool cfs_rq_has_blocked(struct cfs_rq
*cfs_rq
)
7880 if (cfs_rq
->avg
.load_avg
)
7883 if (cfs_rq
->avg
.util_avg
)
7889 static inline bool others_have_blocked(struct rq
*rq
)
7891 if (READ_ONCE(rq
->avg_rt
.util_avg
))
7894 if (READ_ONCE(rq
->avg_dl
.util_avg
))
7897 if (thermal_load_avg(rq
))
7900 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7901 if (READ_ONCE(rq
->avg_irq
.util_avg
))
7908 static inline void update_blocked_load_status(struct rq
*rq
, bool has_blocked
)
7910 rq
->last_blocked_load_update_tick
= jiffies
;
7913 rq
->has_blocked_load
= 0;
7916 static inline bool cfs_rq_has_blocked(struct cfs_rq
*cfs_rq
) { return false; }
7917 static inline bool others_have_blocked(struct rq
*rq
) { return false; }
7918 static inline void update_blocked_load_status(struct rq
*rq
, bool has_blocked
) {}
7921 static bool __update_blocked_others(struct rq
*rq
, bool *done
)
7923 const struct sched_class
*curr_class
;
7924 u64 now
= rq_clock_pelt(rq
);
7925 unsigned long thermal_pressure
;
7929 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
7930 * DL and IRQ signals have been updated before updating CFS.
7932 curr_class
= rq
->curr
->sched_class
;
7934 thermal_pressure
= arch_scale_thermal_pressure(cpu_of(rq
));
7936 decayed
= update_rt_rq_load_avg(now
, rq
, curr_class
== &rt_sched_class
) |
7937 update_dl_rq_load_avg(now
, rq
, curr_class
== &dl_sched_class
) |
7938 update_thermal_load_avg(rq_clock_thermal(rq
), rq
, thermal_pressure
) |
7939 update_irq_load_avg(rq
, 0);
7941 if (others_have_blocked(rq
))
7947 #ifdef CONFIG_FAIR_GROUP_SCHED
7949 static inline bool cfs_rq_is_decayed(struct cfs_rq
*cfs_rq
)
7951 if (cfs_rq
->load
.weight
)
7954 if (cfs_rq
->avg
.load_sum
)
7957 if (cfs_rq
->avg
.util_sum
)
7960 if (cfs_rq
->avg
.runnable_sum
)
7966 static bool __update_blocked_fair(struct rq
*rq
, bool *done
)
7968 struct cfs_rq
*cfs_rq
, *pos
;
7969 bool decayed
= false;
7970 int cpu
= cpu_of(rq
);
7973 * Iterates the task_group tree in a bottom up fashion, see
7974 * list_add_leaf_cfs_rq() for details.
7976 for_each_leaf_cfs_rq_safe(rq
, cfs_rq
, pos
) {
7977 struct sched_entity
*se
;
7979 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq
), cfs_rq
)) {
7980 update_tg_load_avg(cfs_rq
);
7982 if (cfs_rq
== &rq
->cfs
)
7986 /* Propagate pending load changes to the parent, if any: */
7987 se
= cfs_rq
->tg
->se
[cpu
];
7988 if (se
&& !skip_blocked_update(se
))
7989 update_load_avg(cfs_rq_of(se
), se
, 0);
7992 * There can be a lot of idle CPU cgroups. Don't let fully
7993 * decayed cfs_rqs linger on the list.
7995 if (cfs_rq_is_decayed(cfs_rq
))
7996 list_del_leaf_cfs_rq(cfs_rq
);
7998 /* Don't need periodic decay once load/util_avg are null */
7999 if (cfs_rq_has_blocked(cfs_rq
))
8007 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
8008 * This needs to be done in a top-down fashion because the load of a child
8009 * group is a fraction of its parents load.
8011 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
8013 struct rq
*rq
= rq_of(cfs_rq
);
8014 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
8015 unsigned long now
= jiffies
;
8018 if (cfs_rq
->last_h_load_update
== now
)
8021 WRITE_ONCE(cfs_rq
->h_load_next
, NULL
);
8022 for_each_sched_entity(se
) {
8023 cfs_rq
= cfs_rq_of(se
);
8024 WRITE_ONCE(cfs_rq
->h_load_next
, se
);
8025 if (cfs_rq
->last_h_load_update
== now
)
8030 cfs_rq
->h_load
= cfs_rq_load_avg(cfs_rq
);
8031 cfs_rq
->last_h_load_update
= now
;
8034 while ((se
= READ_ONCE(cfs_rq
->h_load_next
)) != NULL
) {
8035 load
= cfs_rq
->h_load
;
8036 load
= div64_ul(load
* se
->avg
.load_avg
,
8037 cfs_rq_load_avg(cfs_rq
) + 1);
8038 cfs_rq
= group_cfs_rq(se
);
8039 cfs_rq
->h_load
= load
;
8040 cfs_rq
->last_h_load_update
= now
;
8044 static unsigned long task_h_load(struct task_struct
*p
)
8046 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
8048 update_cfs_rq_h_load(cfs_rq
);
8049 return div64_ul(p
->se
.avg
.load_avg
* cfs_rq
->h_load
,
8050 cfs_rq_load_avg(cfs_rq
) + 1);
8053 static bool __update_blocked_fair(struct rq
*rq
, bool *done
)
8055 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
8058 decayed
= update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq
), cfs_rq
);
8059 if (cfs_rq_has_blocked(cfs_rq
))
8065 static unsigned long task_h_load(struct task_struct
*p
)
8067 return p
->se
.avg
.load_avg
;
8071 static void update_blocked_averages(int cpu
)
8073 bool decayed
= false, done
= true;
8074 struct rq
*rq
= cpu_rq(cpu
);
8077 rq_lock_irqsave(rq
, &rf
);
8078 update_rq_clock(rq
);
8080 decayed
|= __update_blocked_others(rq
, &done
);
8081 decayed
|= __update_blocked_fair(rq
, &done
);
8083 update_blocked_load_status(rq
, !done
);
8085 cpufreq_update_util(rq
, 0);
8086 rq_unlock_irqrestore(rq
, &rf
);
8089 /********** Helpers for find_busiest_group ************************/
8092 * sg_lb_stats - stats of a sched_group required for load_balancing
8094 struct sg_lb_stats
{
8095 unsigned long avg_load
; /*Avg load across the CPUs of the group */
8096 unsigned long group_load
; /* Total load over the CPUs of the group */
8097 unsigned long group_capacity
;
8098 unsigned long group_util
; /* Total utilization over the CPUs of the group */
8099 unsigned long group_runnable
; /* Total runnable time over the CPUs of the group */
8100 unsigned int sum_nr_running
; /* Nr of tasks running in the group */
8101 unsigned int sum_h_nr_running
; /* Nr of CFS tasks running in the group */
8102 unsigned int idle_cpus
;
8103 unsigned int group_weight
;
8104 enum group_type group_type
;
8105 unsigned int group_asym_packing
; /* Tasks should be moved to preferred CPU */
8106 unsigned long group_misfit_task_load
; /* A CPU has a task too big for its capacity */
8107 #ifdef CONFIG_NUMA_BALANCING
8108 unsigned int nr_numa_running
;
8109 unsigned int nr_preferred_running
;
8114 * sd_lb_stats - Structure to store the statistics of a sched_domain
8115 * during load balancing.
8117 struct sd_lb_stats
{
8118 struct sched_group
*busiest
; /* Busiest group in this sd */
8119 struct sched_group
*local
; /* Local group in this sd */
8120 unsigned long total_load
; /* Total load of all groups in sd */
8121 unsigned long total_capacity
; /* Total capacity of all groups in sd */
8122 unsigned long avg_load
; /* Average load across all groups in sd */
8123 unsigned int prefer_sibling
; /* tasks should go to sibling first */
8125 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
8126 struct sg_lb_stats local_stat
; /* Statistics of the local group */
8129 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
8132 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8133 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8134 * We must however set busiest_stat::group_type and
8135 * busiest_stat::idle_cpus to the worst busiest group because
8136 * update_sd_pick_busiest() reads these before assignment.
8138 *sds
= (struct sd_lb_stats
){
8142 .total_capacity
= 0UL,
8144 .idle_cpus
= UINT_MAX
,
8145 .group_type
= group_has_spare
,
8150 static unsigned long scale_rt_capacity(int cpu
)
8152 struct rq
*rq
= cpu_rq(cpu
);
8153 unsigned long max
= arch_scale_cpu_capacity(cpu
);
8154 unsigned long used
, free
;
8157 irq
= cpu_util_irq(rq
);
8159 if (unlikely(irq
>= max
))
8163 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8164 * (running and not running) with weights 0 and 1024 respectively.
8165 * avg_thermal.load_avg tracks thermal pressure and the weighted
8166 * average uses the actual delta max capacity(load).
8168 used
= READ_ONCE(rq
->avg_rt
.util_avg
);
8169 used
+= READ_ONCE(rq
->avg_dl
.util_avg
);
8170 used
+= thermal_load_avg(rq
);
8172 if (unlikely(used
>= max
))
8177 return scale_irq_capacity(free
, irq
, max
);
8180 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
8182 unsigned long capacity
= scale_rt_capacity(cpu
);
8183 struct sched_group
*sdg
= sd
->groups
;
8185 cpu_rq(cpu
)->cpu_capacity_orig
= arch_scale_cpu_capacity(cpu
);
8190 cpu_rq(cpu
)->cpu_capacity
= capacity
;
8191 trace_sched_cpu_capacity_tp(cpu_rq(cpu
));
8193 sdg
->sgc
->capacity
= capacity
;
8194 sdg
->sgc
->min_capacity
= capacity
;
8195 sdg
->sgc
->max_capacity
= capacity
;
8198 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
8200 struct sched_domain
*child
= sd
->child
;
8201 struct sched_group
*group
, *sdg
= sd
->groups
;
8202 unsigned long capacity
, min_capacity
, max_capacity
;
8203 unsigned long interval
;
8205 interval
= msecs_to_jiffies(sd
->balance_interval
);
8206 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
8207 sdg
->sgc
->next_update
= jiffies
+ interval
;
8210 update_cpu_capacity(sd
, cpu
);
8215 min_capacity
= ULONG_MAX
;
8218 if (child
->flags
& SD_OVERLAP
) {
8220 * SD_OVERLAP domains cannot assume that child groups
8221 * span the current group.
8224 for_each_cpu(cpu
, sched_group_span(sdg
)) {
8225 unsigned long cpu_cap
= capacity_of(cpu
);
8227 capacity
+= cpu_cap
;
8228 min_capacity
= min(cpu_cap
, min_capacity
);
8229 max_capacity
= max(cpu_cap
, max_capacity
);
8233 * !SD_OVERLAP domains can assume that child groups
8234 * span the current group.
8237 group
= child
->groups
;
8239 struct sched_group_capacity
*sgc
= group
->sgc
;
8241 capacity
+= sgc
->capacity
;
8242 min_capacity
= min(sgc
->min_capacity
, min_capacity
);
8243 max_capacity
= max(sgc
->max_capacity
, max_capacity
);
8244 group
= group
->next
;
8245 } while (group
!= child
->groups
);
8248 sdg
->sgc
->capacity
= capacity
;
8249 sdg
->sgc
->min_capacity
= min_capacity
;
8250 sdg
->sgc
->max_capacity
= max_capacity
;
8254 * Check whether the capacity of the rq has been noticeably reduced by side
8255 * activity. The imbalance_pct is used for the threshold.
8256 * Return true is the capacity is reduced
8259 check_cpu_capacity(struct rq
*rq
, struct sched_domain
*sd
)
8261 return ((rq
->cpu_capacity
* sd
->imbalance_pct
) <
8262 (rq
->cpu_capacity_orig
* 100));
8266 * Check whether a rq has a misfit task and if it looks like we can actually
8267 * help that task: we can migrate the task to a CPU of higher capacity, or
8268 * the task's current CPU is heavily pressured.
8270 static inline int check_misfit_status(struct rq
*rq
, struct sched_domain
*sd
)
8272 return rq
->misfit_task_load
&&
8273 (rq
->cpu_capacity_orig
< rq
->rd
->max_cpu_capacity
||
8274 check_cpu_capacity(rq
, sd
));
8278 * Group imbalance indicates (and tries to solve) the problem where balancing
8279 * groups is inadequate due to ->cpus_ptr constraints.
8281 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8282 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8285 * { 0 1 2 3 } { 4 5 6 7 }
8288 * If we were to balance group-wise we'd place two tasks in the first group and
8289 * two tasks in the second group. Clearly this is undesired as it will overload
8290 * cpu 3 and leave one of the CPUs in the second group unused.
8292 * The current solution to this issue is detecting the skew in the first group
8293 * by noticing the lower domain failed to reach balance and had difficulty
8294 * moving tasks due to affinity constraints.
8296 * When this is so detected; this group becomes a candidate for busiest; see
8297 * update_sd_pick_busiest(). And calculate_imbalance() and
8298 * find_busiest_group() avoid some of the usual balance conditions to allow it
8299 * to create an effective group imbalance.
8301 * This is a somewhat tricky proposition since the next run might not find the
8302 * group imbalance and decide the groups need to be balanced again. A most
8303 * subtle and fragile situation.
8306 static inline int sg_imbalanced(struct sched_group
*group
)
8308 return group
->sgc
->imbalance
;
8312 * group_has_capacity returns true if the group has spare capacity that could
8313 * be used by some tasks.
8314 * We consider that a group has spare capacity if the * number of task is
8315 * smaller than the number of CPUs or if the utilization is lower than the
8316 * available capacity for CFS tasks.
8317 * For the latter, we use a threshold to stabilize the state, to take into
8318 * account the variance of the tasks' load and to return true if the available
8319 * capacity in meaningful for the load balancer.
8320 * As an example, an available capacity of 1% can appear but it doesn't make
8321 * any benefit for the load balance.
8324 group_has_capacity(unsigned int imbalance_pct
, struct sg_lb_stats
*sgs
)
8326 if (sgs
->sum_nr_running
< sgs
->group_weight
)
8329 if ((sgs
->group_capacity
* imbalance_pct
) <
8330 (sgs
->group_runnable
* 100))
8333 if ((sgs
->group_capacity
* 100) >
8334 (sgs
->group_util
* imbalance_pct
))
8341 * group_is_overloaded returns true if the group has more tasks than it can
8343 * group_is_overloaded is not equals to !group_has_capacity because a group
8344 * with the exact right number of tasks, has no more spare capacity but is not
8345 * overloaded so both group_has_capacity and group_is_overloaded return
8349 group_is_overloaded(unsigned int imbalance_pct
, struct sg_lb_stats
*sgs
)
8351 if (sgs
->sum_nr_running
<= sgs
->group_weight
)
8354 if ((sgs
->group_capacity
* 100) <
8355 (sgs
->group_util
* imbalance_pct
))
8358 if ((sgs
->group_capacity
* imbalance_pct
) <
8359 (sgs
->group_runnable
* 100))
8366 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
8367 * per-CPU capacity than sched_group ref.
8370 group_smaller_min_cpu_capacity(struct sched_group
*sg
, struct sched_group
*ref
)
8372 return fits_capacity(sg
->sgc
->min_capacity
, ref
->sgc
->min_capacity
);
8376 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
8377 * per-CPU capacity_orig than sched_group ref.
8380 group_smaller_max_cpu_capacity(struct sched_group
*sg
, struct sched_group
*ref
)
8382 return fits_capacity(sg
->sgc
->max_capacity
, ref
->sgc
->max_capacity
);
8386 group_type
group_classify(unsigned int imbalance_pct
,
8387 struct sched_group
*group
,
8388 struct sg_lb_stats
*sgs
)
8390 if (group_is_overloaded(imbalance_pct
, sgs
))
8391 return group_overloaded
;
8393 if (sg_imbalanced(group
))
8394 return group_imbalanced
;
8396 if (sgs
->group_asym_packing
)
8397 return group_asym_packing
;
8399 if (sgs
->group_misfit_task_load
)
8400 return group_misfit_task
;
8402 if (!group_has_capacity(imbalance_pct
, sgs
))
8403 return group_fully_busy
;
8405 return group_has_spare
;
8408 static bool update_nohz_stats(struct rq
*rq
, bool force
)
8410 #ifdef CONFIG_NO_HZ_COMMON
8411 unsigned int cpu
= rq
->cpu
;
8413 if (!rq
->has_blocked_load
)
8416 if (!cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))
8419 if (!force
&& !time_after(jiffies
, rq
->last_blocked_load_update_tick
))
8422 update_blocked_averages(cpu
);
8424 return rq
->has_blocked_load
;
8431 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8432 * @env: The load balancing environment.
8433 * @group: sched_group whose statistics are to be updated.
8434 * @sgs: variable to hold the statistics for this group.
8435 * @sg_status: Holds flag indicating the status of the sched_group
8437 static inline void update_sg_lb_stats(struct lb_env
*env
,
8438 struct sched_group
*group
,
8439 struct sg_lb_stats
*sgs
,
8442 int i
, nr_running
, local_group
;
8444 memset(sgs
, 0, sizeof(*sgs
));
8446 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_span(group
));
8448 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
8449 struct rq
*rq
= cpu_rq(i
);
8451 if ((env
->flags
& LBF_NOHZ_STATS
) && update_nohz_stats(rq
, false))
8452 env
->flags
|= LBF_NOHZ_AGAIN
;
8454 sgs
->group_load
+= cpu_load(rq
);
8455 sgs
->group_util
+= cpu_util(i
);
8456 sgs
->group_runnable
+= cpu_runnable(rq
);
8457 sgs
->sum_h_nr_running
+= rq
->cfs
.h_nr_running
;
8459 nr_running
= rq
->nr_running
;
8460 sgs
->sum_nr_running
+= nr_running
;
8463 *sg_status
|= SG_OVERLOAD
;
8465 if (cpu_overutilized(i
))
8466 *sg_status
|= SG_OVERUTILIZED
;
8468 #ifdef CONFIG_NUMA_BALANCING
8469 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
8470 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
8473 * No need to call idle_cpu() if nr_running is not 0
8475 if (!nr_running
&& idle_cpu(i
)) {
8477 /* Idle cpu can't have misfit task */
8484 /* Check for a misfit task on the cpu */
8485 if (env
->sd
->flags
& SD_ASYM_CPUCAPACITY
&&
8486 sgs
->group_misfit_task_load
< rq
->misfit_task_load
) {
8487 sgs
->group_misfit_task_load
= rq
->misfit_task_load
;
8488 *sg_status
|= SG_OVERLOAD
;
8492 /* Check if dst CPU is idle and preferred to this group */
8493 if (env
->sd
->flags
& SD_ASYM_PACKING
&&
8494 env
->idle
!= CPU_NOT_IDLE
&&
8495 sgs
->sum_h_nr_running
&&
8496 sched_asym_prefer(env
->dst_cpu
, group
->asym_prefer_cpu
)) {
8497 sgs
->group_asym_packing
= 1;
8500 sgs
->group_capacity
= group
->sgc
->capacity
;
8502 sgs
->group_weight
= group
->group_weight
;
8504 sgs
->group_type
= group_classify(env
->sd
->imbalance_pct
, group
, sgs
);
8506 /* Computing avg_load makes sense only when group is overloaded */
8507 if (sgs
->group_type
== group_overloaded
)
8508 sgs
->avg_load
= (sgs
->group_load
* SCHED_CAPACITY_SCALE
) /
8509 sgs
->group_capacity
;
8513 * update_sd_pick_busiest - return 1 on busiest group
8514 * @env: The load balancing environment.
8515 * @sds: sched_domain statistics
8516 * @sg: sched_group candidate to be checked for being the busiest
8517 * @sgs: sched_group statistics
8519 * Determine if @sg is a busier group than the previously selected
8522 * Return: %true if @sg is a busier group than the previously selected
8523 * busiest group. %false otherwise.
8525 static bool update_sd_pick_busiest(struct lb_env
*env
,
8526 struct sd_lb_stats
*sds
,
8527 struct sched_group
*sg
,
8528 struct sg_lb_stats
*sgs
)
8530 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
8532 /* Make sure that there is at least one task to pull */
8533 if (!sgs
->sum_h_nr_running
)
8537 * Don't try to pull misfit tasks we can't help.
8538 * We can use max_capacity here as reduction in capacity on some
8539 * CPUs in the group should either be possible to resolve
8540 * internally or be covered by avg_load imbalance (eventually).
8542 if (sgs
->group_type
== group_misfit_task
&&
8543 (!group_smaller_max_cpu_capacity(sg
, sds
->local
) ||
8544 sds
->local_stat
.group_type
!= group_has_spare
))
8547 if (sgs
->group_type
> busiest
->group_type
)
8550 if (sgs
->group_type
< busiest
->group_type
)
8554 * The candidate and the current busiest group are the same type of
8555 * group. Let check which one is the busiest according to the type.
8558 switch (sgs
->group_type
) {
8559 case group_overloaded
:
8560 /* Select the overloaded group with highest avg_load. */
8561 if (sgs
->avg_load
<= busiest
->avg_load
)
8565 case group_imbalanced
:
8567 * Select the 1st imbalanced group as we don't have any way to
8568 * choose one more than another.
8572 case group_asym_packing
:
8573 /* Prefer to move from lowest priority CPU's work */
8574 if (sched_asym_prefer(sg
->asym_prefer_cpu
, sds
->busiest
->asym_prefer_cpu
))
8578 case group_misfit_task
:
8580 * If we have more than one misfit sg go with the biggest
8583 if (sgs
->group_misfit_task_load
< busiest
->group_misfit_task_load
)
8587 case group_fully_busy
:
8589 * Select the fully busy group with highest avg_load. In
8590 * theory, there is no need to pull task from such kind of
8591 * group because tasks have all compute capacity that they need
8592 * but we can still improve the overall throughput by reducing
8593 * contention when accessing shared HW resources.
8595 * XXX for now avg_load is not computed and always 0 so we
8596 * select the 1st one.
8598 if (sgs
->avg_load
<= busiest
->avg_load
)
8602 case group_has_spare
:
8604 * Select not overloaded group with lowest number of idle cpus
8605 * and highest number of running tasks. We could also compare
8606 * the spare capacity which is more stable but it can end up
8607 * that the group has less spare capacity but finally more idle
8608 * CPUs which means less opportunity to pull tasks.
8610 if (sgs
->idle_cpus
> busiest
->idle_cpus
)
8612 else if ((sgs
->idle_cpus
== busiest
->idle_cpus
) &&
8613 (sgs
->sum_nr_running
<= busiest
->sum_nr_running
))
8620 * Candidate sg has no more than one task per CPU and has higher
8621 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8622 * throughput. Maximize throughput, power/energy consequences are not
8625 if ((env
->sd
->flags
& SD_ASYM_CPUCAPACITY
) &&
8626 (sgs
->group_type
<= group_fully_busy
) &&
8627 (group_smaller_min_cpu_capacity(sds
->local
, sg
)))
8633 #ifdef CONFIG_NUMA_BALANCING
8634 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
8636 if (sgs
->sum_h_nr_running
> sgs
->nr_numa_running
)
8638 if (sgs
->sum_h_nr_running
> sgs
->nr_preferred_running
)
8643 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
8645 if (rq
->nr_running
> rq
->nr_numa_running
)
8647 if (rq
->nr_running
> rq
->nr_preferred_running
)
8652 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
8657 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
8661 #endif /* CONFIG_NUMA_BALANCING */
8667 * task_running_on_cpu - return 1 if @p is running on @cpu.
8670 static unsigned int task_running_on_cpu(int cpu
, struct task_struct
*p
)
8672 /* Task has no contribution or is new */
8673 if (cpu
!= task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
8676 if (task_on_rq_queued(p
))
8683 * idle_cpu_without - would a given CPU be idle without p ?
8684 * @cpu: the processor on which idleness is tested.
8685 * @p: task which should be ignored.
8687 * Return: 1 if the CPU would be idle. 0 otherwise.
8689 static int idle_cpu_without(int cpu
, struct task_struct
*p
)
8691 struct rq
*rq
= cpu_rq(cpu
);
8693 if (rq
->curr
!= rq
->idle
&& rq
->curr
!= p
)
8697 * rq->nr_running can't be used but an updated version without the
8698 * impact of p on cpu must be used instead. The updated nr_running
8699 * be computed and tested before calling idle_cpu_without().
8703 if (rq
->ttwu_pending
)
8711 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8712 * @sd: The sched_domain level to look for idlest group.
8713 * @group: sched_group whose statistics are to be updated.
8714 * @sgs: variable to hold the statistics for this group.
8715 * @p: The task for which we look for the idlest group/CPU.
8717 static inline void update_sg_wakeup_stats(struct sched_domain
*sd
,
8718 struct sched_group
*group
,
8719 struct sg_lb_stats
*sgs
,
8720 struct task_struct
*p
)
8724 memset(sgs
, 0, sizeof(*sgs
));
8726 for_each_cpu(i
, sched_group_span(group
)) {
8727 struct rq
*rq
= cpu_rq(i
);
8730 sgs
->group_load
+= cpu_load_without(rq
, p
);
8731 sgs
->group_util
+= cpu_util_without(i
, p
);
8732 sgs
->group_runnable
+= cpu_runnable_without(rq
, p
);
8733 local
= task_running_on_cpu(i
, p
);
8734 sgs
->sum_h_nr_running
+= rq
->cfs
.h_nr_running
- local
;
8736 nr_running
= rq
->nr_running
- local
;
8737 sgs
->sum_nr_running
+= nr_running
;
8740 * No need to call idle_cpu_without() if nr_running is not 0
8742 if (!nr_running
&& idle_cpu_without(i
, p
))
8747 /* Check if task fits in the group */
8748 if (sd
->flags
& SD_ASYM_CPUCAPACITY
&&
8749 !task_fits_capacity(p
, group
->sgc
->max_capacity
)) {
8750 sgs
->group_misfit_task_load
= 1;
8753 sgs
->group_capacity
= group
->sgc
->capacity
;
8755 sgs
->group_weight
= group
->group_weight
;
8757 sgs
->group_type
= group_classify(sd
->imbalance_pct
, group
, sgs
);
8760 * Computing avg_load makes sense only when group is fully busy or
8763 if (sgs
->group_type
== group_fully_busy
||
8764 sgs
->group_type
== group_overloaded
)
8765 sgs
->avg_load
= (sgs
->group_load
* SCHED_CAPACITY_SCALE
) /
8766 sgs
->group_capacity
;
8769 static bool update_pick_idlest(struct sched_group
*idlest
,
8770 struct sg_lb_stats
*idlest_sgs
,
8771 struct sched_group
*group
,
8772 struct sg_lb_stats
*sgs
)
8774 if (sgs
->group_type
< idlest_sgs
->group_type
)
8777 if (sgs
->group_type
> idlest_sgs
->group_type
)
8781 * The candidate and the current idlest group are the same type of
8782 * group. Let check which one is the idlest according to the type.
8785 switch (sgs
->group_type
) {
8786 case group_overloaded
:
8787 case group_fully_busy
:
8788 /* Select the group with lowest avg_load. */
8789 if (idlest_sgs
->avg_load
<= sgs
->avg_load
)
8793 case group_imbalanced
:
8794 case group_asym_packing
:
8795 /* Those types are not used in the slow wakeup path */
8798 case group_misfit_task
:
8799 /* Select group with the highest max capacity */
8800 if (idlest
->sgc
->max_capacity
>= group
->sgc
->max_capacity
)
8804 case group_has_spare
:
8805 /* Select group with most idle CPUs */
8806 if (idlest_sgs
->idle_cpus
> sgs
->idle_cpus
)
8809 /* Select group with lowest group_util */
8810 if (idlest_sgs
->idle_cpus
== sgs
->idle_cpus
&&
8811 idlest_sgs
->group_util
<= sgs
->group_util
)
8821 * Allow a NUMA imbalance if busy CPUs is less than 25% of the domain.
8822 * This is an approximation as the number of running tasks may not be
8823 * related to the number of busy CPUs due to sched_setaffinity.
8825 static inline bool allow_numa_imbalance(int dst_running
, int dst_weight
)
8827 return (dst_running
< (dst_weight
>> 2));
8831 * find_idlest_group() finds and returns the least busy CPU group within the
8834 * Assumes p is allowed on at least one CPU in sd.
8836 static struct sched_group
*
8837 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
8839 struct sched_group
*idlest
= NULL
, *local
= NULL
, *group
= sd
->groups
;
8840 struct sg_lb_stats local_sgs
, tmp_sgs
;
8841 struct sg_lb_stats
*sgs
;
8842 unsigned long imbalance
;
8843 struct sg_lb_stats idlest_sgs
= {
8844 .avg_load
= UINT_MAX
,
8845 .group_type
= group_overloaded
,
8851 /* Skip over this group if it has no CPUs allowed */
8852 if (!cpumask_intersects(sched_group_span(group
),
8856 local_group
= cpumask_test_cpu(this_cpu
,
8857 sched_group_span(group
));
8866 update_sg_wakeup_stats(sd
, group
, sgs
, p
);
8868 if (!local_group
&& update_pick_idlest(idlest
, &idlest_sgs
, group
, sgs
)) {
8873 } while (group
= group
->next
, group
!= sd
->groups
);
8876 /* There is no idlest group to push tasks to */
8880 /* The local group has been skipped because of CPU affinity */
8885 * If the local group is idler than the selected idlest group
8886 * don't try and push the task.
8888 if (local_sgs
.group_type
< idlest_sgs
.group_type
)
8892 * If the local group is busier than the selected idlest group
8893 * try and push the task.
8895 if (local_sgs
.group_type
> idlest_sgs
.group_type
)
8898 switch (local_sgs
.group_type
) {
8899 case group_overloaded
:
8900 case group_fully_busy
:
8902 /* Calculate allowed imbalance based on load */
8903 imbalance
= scale_load_down(NICE_0_LOAD
) *
8904 (sd
->imbalance_pct
-100) / 100;
8907 * When comparing groups across NUMA domains, it's possible for
8908 * the local domain to be very lightly loaded relative to the
8909 * remote domains but "imbalance" skews the comparison making
8910 * remote CPUs look much more favourable. When considering
8911 * cross-domain, add imbalance to the load on the remote node
8912 * and consider staying local.
8915 if ((sd
->flags
& SD_NUMA
) &&
8916 ((idlest_sgs
.avg_load
+ imbalance
) >= local_sgs
.avg_load
))
8920 * If the local group is less loaded than the selected
8921 * idlest group don't try and push any tasks.
8923 if (idlest_sgs
.avg_load
>= (local_sgs
.avg_load
+ imbalance
))
8926 if (100 * local_sgs
.avg_load
<= sd
->imbalance_pct
* idlest_sgs
.avg_load
)
8930 case group_imbalanced
:
8931 case group_asym_packing
:
8932 /* Those type are not used in the slow wakeup path */
8935 case group_misfit_task
:
8936 /* Select group with the highest max capacity */
8937 if (local
->sgc
->max_capacity
>= idlest
->sgc
->max_capacity
)
8941 case group_has_spare
:
8942 if (sd
->flags
& SD_NUMA
) {
8943 #ifdef CONFIG_NUMA_BALANCING
8946 * If there is spare capacity at NUMA, try to select
8947 * the preferred node
8949 if (cpu_to_node(this_cpu
) == p
->numa_preferred_nid
)
8952 idlest_cpu
= cpumask_first(sched_group_span(idlest
));
8953 if (cpu_to_node(idlest_cpu
) == p
->numa_preferred_nid
)
8957 * Otherwise, keep the task on this node to stay close
8958 * its wakeup source and improve locality. If there is
8959 * a real need of migration, periodic load balance will
8962 if (allow_numa_imbalance(local_sgs
.sum_nr_running
, sd
->span_weight
))
8967 * Select group with highest number of idle CPUs. We could also
8968 * compare the utilization which is more stable but it can end
8969 * up that the group has less spare capacity but finally more
8970 * idle CPUs which means more opportunity to run task.
8972 if (local_sgs
.idle_cpus
>= idlest_sgs
.idle_cpus
)
8981 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8982 * @env: The load balancing environment.
8983 * @sds: variable to hold the statistics for this sched_domain.
8986 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
8988 struct sched_domain
*child
= env
->sd
->child
;
8989 struct sched_group
*sg
= env
->sd
->groups
;
8990 struct sg_lb_stats
*local
= &sds
->local_stat
;
8991 struct sg_lb_stats tmp_sgs
;
8994 #ifdef CONFIG_NO_HZ_COMMON
8995 if (env
->idle
== CPU_NEWLY_IDLE
&& READ_ONCE(nohz
.has_blocked
))
8996 env
->flags
|= LBF_NOHZ_STATS
;
9000 struct sg_lb_stats
*sgs
= &tmp_sgs
;
9003 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_span(sg
));
9008 if (env
->idle
!= CPU_NEWLY_IDLE
||
9009 time_after_eq(jiffies
, sg
->sgc
->next_update
))
9010 update_group_capacity(env
->sd
, env
->dst_cpu
);
9013 update_sg_lb_stats(env
, sg
, sgs
, &sg_status
);
9019 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
9021 sds
->busiest_stat
= *sgs
;
9025 /* Now, start updating sd_lb_stats */
9026 sds
->total_load
+= sgs
->group_load
;
9027 sds
->total_capacity
+= sgs
->group_capacity
;
9030 } while (sg
!= env
->sd
->groups
);
9032 /* Tag domain that child domain prefers tasks go to siblings first */
9033 sds
->prefer_sibling
= child
&& child
->flags
& SD_PREFER_SIBLING
;
9035 #ifdef CONFIG_NO_HZ_COMMON
9036 if ((env
->flags
& LBF_NOHZ_AGAIN
) &&
9037 cpumask_subset(nohz
.idle_cpus_mask
, sched_domain_span(env
->sd
))) {
9039 WRITE_ONCE(nohz
.next_blocked
,
9040 jiffies
+ msecs_to_jiffies(LOAD_AVG_PERIOD
));
9044 if (env
->sd
->flags
& SD_NUMA
)
9045 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
9047 if (!env
->sd
->parent
) {
9048 struct root_domain
*rd
= env
->dst_rq
->rd
;
9050 /* update overload indicator if we are at root domain */
9051 WRITE_ONCE(rd
->overload
, sg_status
& SG_OVERLOAD
);
9053 /* Update over-utilization (tipping point, U >= 0) indicator */
9054 WRITE_ONCE(rd
->overutilized
, sg_status
& SG_OVERUTILIZED
);
9055 trace_sched_overutilized_tp(rd
, sg_status
& SG_OVERUTILIZED
);
9056 } else if (sg_status
& SG_OVERUTILIZED
) {
9057 struct root_domain
*rd
= env
->dst_rq
->rd
;
9059 WRITE_ONCE(rd
->overutilized
, SG_OVERUTILIZED
);
9060 trace_sched_overutilized_tp(rd
, SG_OVERUTILIZED
);
9064 #define NUMA_IMBALANCE_MIN 2
9066 static inline long adjust_numa_imbalance(int imbalance
,
9067 int dst_running
, int dst_weight
)
9069 if (!allow_numa_imbalance(dst_running
, dst_weight
))
9073 * Allow a small imbalance based on a simple pair of communicating
9074 * tasks that remain local when the destination is lightly loaded.
9076 if (imbalance
<= NUMA_IMBALANCE_MIN
)
9083 * calculate_imbalance - Calculate the amount of imbalance present within the
9084 * groups of a given sched_domain during load balance.
9085 * @env: load balance environment
9086 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
9088 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
9090 struct sg_lb_stats
*local
, *busiest
;
9092 local
= &sds
->local_stat
;
9093 busiest
= &sds
->busiest_stat
;
9095 if (busiest
->group_type
== group_misfit_task
) {
9096 /* Set imbalance to allow misfit tasks to be balanced. */
9097 env
->migration_type
= migrate_misfit
;
9102 if (busiest
->group_type
== group_asym_packing
) {
9104 * In case of asym capacity, we will try to migrate all load to
9105 * the preferred CPU.
9107 env
->migration_type
= migrate_task
;
9108 env
->imbalance
= busiest
->sum_h_nr_running
;
9112 if (busiest
->group_type
== group_imbalanced
) {
9114 * In the group_imb case we cannot rely on group-wide averages
9115 * to ensure CPU-load equilibrium, try to move any task to fix
9116 * the imbalance. The next load balance will take care of
9117 * balancing back the system.
9119 env
->migration_type
= migrate_task
;
9125 * Try to use spare capacity of local group without overloading it or
9128 if (local
->group_type
== group_has_spare
) {
9129 if ((busiest
->group_type
> group_fully_busy
) &&
9130 !(env
->sd
->flags
& SD_SHARE_PKG_RESOURCES
)) {
9132 * If busiest is overloaded, try to fill spare
9133 * capacity. This might end up creating spare capacity
9134 * in busiest or busiest still being overloaded but
9135 * there is no simple way to directly compute the
9136 * amount of load to migrate in order to balance the
9139 env
->migration_type
= migrate_util
;
9140 env
->imbalance
= max(local
->group_capacity
, local
->group_util
) -
9144 * In some cases, the group's utilization is max or even
9145 * higher than capacity because of migrations but the
9146 * local CPU is (newly) idle. There is at least one
9147 * waiting task in this overloaded busiest group. Let's
9150 if (env
->idle
!= CPU_NOT_IDLE
&& env
->imbalance
== 0) {
9151 env
->migration_type
= migrate_task
;
9158 if (busiest
->group_weight
== 1 || sds
->prefer_sibling
) {
9159 unsigned int nr_diff
= busiest
->sum_nr_running
;
9161 * When prefer sibling, evenly spread running tasks on
9164 env
->migration_type
= migrate_task
;
9165 lsub_positive(&nr_diff
, local
->sum_nr_running
);
9166 env
->imbalance
= nr_diff
>> 1;
9170 * If there is no overload, we just want to even the number of
9173 env
->migration_type
= migrate_task
;
9174 env
->imbalance
= max_t(long, 0, (local
->idle_cpus
-
9175 busiest
->idle_cpus
) >> 1);
9178 /* Consider allowing a small imbalance between NUMA groups */
9179 if (env
->sd
->flags
& SD_NUMA
) {
9180 env
->imbalance
= adjust_numa_imbalance(env
->imbalance
,
9181 busiest
->sum_nr_running
, busiest
->group_weight
);
9188 * Local is fully busy but has to take more load to relieve the
9191 if (local
->group_type
< group_overloaded
) {
9193 * Local will become overloaded so the avg_load metrics are
9197 local
->avg_load
= (local
->group_load
* SCHED_CAPACITY_SCALE
) /
9198 local
->group_capacity
;
9200 sds
->avg_load
= (sds
->total_load
* SCHED_CAPACITY_SCALE
) /
9201 sds
->total_capacity
;
9203 * If the local group is more loaded than the selected
9204 * busiest group don't try to pull any tasks.
9206 if (local
->avg_load
>= busiest
->avg_load
) {
9213 * Both group are or will become overloaded and we're trying to get all
9214 * the CPUs to the average_load, so we don't want to push ourselves
9215 * above the average load, nor do we wish to reduce the max loaded CPU
9216 * below the average load. At the same time, we also don't want to
9217 * reduce the group load below the group capacity. Thus we look for
9218 * the minimum possible imbalance.
9220 env
->migration_type
= migrate_load
;
9221 env
->imbalance
= min(
9222 (busiest
->avg_load
- sds
->avg_load
) * busiest
->group_capacity
,
9223 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
9224 ) / SCHED_CAPACITY_SCALE
;
9227 /******* find_busiest_group() helpers end here *********************/
9230 * Decision matrix according to the local and busiest group type:
9232 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9233 * has_spare nr_idle balanced N/A N/A balanced balanced
9234 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9235 * misfit_task force N/A N/A N/A force force
9236 * asym_packing force force N/A N/A force force
9237 * imbalanced force force N/A N/A force force
9238 * overloaded force force N/A N/A force avg_load
9240 * N/A : Not Applicable because already filtered while updating
9242 * balanced : The system is balanced for these 2 groups.
9243 * force : Calculate the imbalance as load migration is probably needed.
9244 * avg_load : Only if imbalance is significant enough.
9245 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9246 * different in groups.
9250 * find_busiest_group - Returns the busiest group within the sched_domain
9251 * if there is an imbalance.
9253 * Also calculates the amount of runnable load which should be moved
9254 * to restore balance.
9256 * @env: The load balancing environment.
9258 * Return: - The busiest group if imbalance exists.
9260 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
9262 struct sg_lb_stats
*local
, *busiest
;
9263 struct sd_lb_stats sds
;
9265 init_sd_lb_stats(&sds
);
9268 * Compute the various statistics relevant for load balancing at
9271 update_sd_lb_stats(env
, &sds
);
9273 if (sched_energy_enabled()) {
9274 struct root_domain
*rd
= env
->dst_rq
->rd
;
9276 if (rcu_dereference(rd
->pd
) && !READ_ONCE(rd
->overutilized
))
9280 local
= &sds
.local_stat
;
9281 busiest
= &sds
.busiest_stat
;
9283 /* There is no busy sibling group to pull tasks from */
9287 /* Misfit tasks should be dealt with regardless of the avg load */
9288 if (busiest
->group_type
== group_misfit_task
)
9291 /* ASYM feature bypasses nice load balance check */
9292 if (busiest
->group_type
== group_asym_packing
)
9296 * If the busiest group is imbalanced the below checks don't
9297 * work because they assume all things are equal, which typically
9298 * isn't true due to cpus_ptr constraints and the like.
9300 if (busiest
->group_type
== group_imbalanced
)
9304 * If the local group is busier than the selected busiest group
9305 * don't try and pull any tasks.
9307 if (local
->group_type
> busiest
->group_type
)
9311 * When groups are overloaded, use the avg_load to ensure fairness
9314 if (local
->group_type
== group_overloaded
) {
9316 * If the local group is more loaded than the selected
9317 * busiest group don't try to pull any tasks.
9319 if (local
->avg_load
>= busiest
->avg_load
)
9322 /* XXX broken for overlapping NUMA groups */
9323 sds
.avg_load
= (sds
.total_load
* SCHED_CAPACITY_SCALE
) /
9327 * Don't pull any tasks if this group is already above the
9328 * domain average load.
9330 if (local
->avg_load
>= sds
.avg_load
)
9334 * If the busiest group is more loaded, use imbalance_pct to be
9337 if (100 * busiest
->avg_load
<=
9338 env
->sd
->imbalance_pct
* local
->avg_load
)
9342 /* Try to move all excess tasks to child's sibling domain */
9343 if (sds
.prefer_sibling
&& local
->group_type
== group_has_spare
&&
9344 busiest
->sum_nr_running
> local
->sum_nr_running
+ 1)
9347 if (busiest
->group_type
!= group_overloaded
) {
9348 if (env
->idle
== CPU_NOT_IDLE
)
9350 * If the busiest group is not overloaded (and as a
9351 * result the local one too) but this CPU is already
9352 * busy, let another idle CPU try to pull task.
9356 if (busiest
->group_weight
> 1 &&
9357 local
->idle_cpus
<= (busiest
->idle_cpus
+ 1))
9359 * If the busiest group is not overloaded
9360 * and there is no imbalance between this and busiest
9361 * group wrt idle CPUs, it is balanced. The imbalance
9362 * becomes significant if the diff is greater than 1
9363 * otherwise we might end up to just move the imbalance
9364 * on another group. Of course this applies only if
9365 * there is more than 1 CPU per group.
9369 if (busiest
->sum_h_nr_running
== 1)
9371 * busiest doesn't have any tasks waiting to run
9377 /* Looks like there is an imbalance. Compute it */
9378 calculate_imbalance(env
, &sds
);
9379 return env
->imbalance
? sds
.busiest
: NULL
;
9387 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9389 static struct rq
*find_busiest_queue(struct lb_env
*env
,
9390 struct sched_group
*group
)
9392 struct rq
*busiest
= NULL
, *rq
;
9393 unsigned long busiest_util
= 0, busiest_load
= 0, busiest_capacity
= 1;
9394 unsigned int busiest_nr
= 0;
9397 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
9398 unsigned long capacity
, load
, util
;
9399 unsigned int nr_running
;
9403 rt
= fbq_classify_rq(rq
);
9406 * We classify groups/runqueues into three groups:
9407 * - regular: there are !numa tasks
9408 * - remote: there are numa tasks that run on the 'wrong' node
9409 * - all: there is no distinction
9411 * In order to avoid migrating ideally placed numa tasks,
9412 * ignore those when there's better options.
9414 * If we ignore the actual busiest queue to migrate another
9415 * task, the next balance pass can still reduce the busiest
9416 * queue by moving tasks around inside the node.
9418 * If we cannot move enough load due to this classification
9419 * the next pass will adjust the group classification and
9420 * allow migration of more tasks.
9422 * Both cases only affect the total convergence complexity.
9424 if (rt
> env
->fbq_type
)
9427 capacity
= capacity_of(i
);
9428 nr_running
= rq
->cfs
.h_nr_running
;
9431 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9432 * eventually lead to active_balancing high->low capacity.
9433 * Higher per-CPU capacity is considered better than balancing
9436 if (env
->sd
->flags
& SD_ASYM_CPUCAPACITY
&&
9437 capacity_of(env
->dst_cpu
) < capacity
&&
9441 switch (env
->migration_type
) {
9444 * When comparing with load imbalance, use cpu_load()
9445 * which is not scaled with the CPU capacity.
9447 load
= cpu_load(rq
);
9449 if (nr_running
== 1 && load
> env
->imbalance
&&
9450 !check_cpu_capacity(rq
, env
->sd
))
9454 * For the load comparisons with the other CPUs,
9455 * consider the cpu_load() scaled with the CPU
9456 * capacity, so that the load can be moved away
9457 * from the CPU that is potentially running at a
9460 * Thus we're looking for max(load_i / capacity_i),
9461 * crosswise multiplication to rid ourselves of the
9462 * division works out to:
9463 * load_i * capacity_j > load_j * capacity_i;
9464 * where j is our previous maximum.
9466 if (load
* busiest_capacity
> busiest_load
* capacity
) {
9467 busiest_load
= load
;
9468 busiest_capacity
= capacity
;
9474 util
= cpu_util(cpu_of(rq
));
9477 * Don't try to pull utilization from a CPU with one
9478 * running task. Whatever its utilization, we will fail
9481 if (nr_running
<= 1)
9484 if (busiest_util
< util
) {
9485 busiest_util
= util
;
9491 if (busiest_nr
< nr_running
) {
9492 busiest_nr
= nr_running
;
9497 case migrate_misfit
:
9499 * For ASYM_CPUCAPACITY domains with misfit tasks we
9500 * simply seek the "biggest" misfit task.
9502 if (rq
->misfit_task_load
> busiest_load
) {
9503 busiest_load
= rq
->misfit_task_load
;
9516 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9517 * so long as it is large enough.
9519 #define MAX_PINNED_INTERVAL 512
9522 asym_active_balance(struct lb_env
*env
)
9525 * ASYM_PACKING needs to force migrate tasks from busy but
9526 * lower priority CPUs in order to pack all tasks in the
9527 * highest priority CPUs.
9529 return env
->idle
!= CPU_NOT_IDLE
&& (env
->sd
->flags
& SD_ASYM_PACKING
) &&
9530 sched_asym_prefer(env
->dst_cpu
, env
->src_cpu
);
9534 voluntary_active_balance(struct lb_env
*env
)
9536 struct sched_domain
*sd
= env
->sd
;
9538 if (asym_active_balance(env
))
9542 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
9543 * It's worth migrating the task if the src_cpu's capacity is reduced
9544 * because of other sched_class or IRQs if more capacity stays
9545 * available on dst_cpu.
9547 if ((env
->idle
!= CPU_NOT_IDLE
) &&
9548 (env
->src_rq
->cfs
.h_nr_running
== 1)) {
9549 if ((check_cpu_capacity(env
->src_rq
, sd
)) &&
9550 (capacity_of(env
->src_cpu
)*sd
->imbalance_pct
< capacity_of(env
->dst_cpu
)*100))
9554 if (env
->migration_type
== migrate_misfit
)
9560 static int need_active_balance(struct lb_env
*env
)
9562 struct sched_domain
*sd
= env
->sd
;
9564 if (voluntary_active_balance(env
))
9567 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
9570 static int active_load_balance_cpu_stop(void *data
);
9572 static int should_we_balance(struct lb_env
*env
)
9574 struct sched_group
*sg
= env
->sd
->groups
;
9578 * Ensure the balancing environment is consistent; can happen
9579 * when the softirq triggers 'during' hotplug.
9581 if (!cpumask_test_cpu(env
->dst_cpu
, env
->cpus
))
9585 * In the newly idle case, we will allow all the CPUs
9586 * to do the newly idle load balance.
9588 if (env
->idle
== CPU_NEWLY_IDLE
)
9591 /* Try to find first idle CPU */
9592 for_each_cpu_and(cpu
, group_balance_mask(sg
), env
->cpus
) {
9596 /* Are we the first idle CPU? */
9597 return cpu
== env
->dst_cpu
;
9600 /* Are we the first CPU of this group ? */
9601 return group_balance_cpu(sg
) == env
->dst_cpu
;
9605 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9606 * tasks if there is an imbalance.
9608 static int load_balance(int this_cpu
, struct rq
*this_rq
,
9609 struct sched_domain
*sd
, enum cpu_idle_type idle
,
9610 int *continue_balancing
)
9612 int ld_moved
, cur_ld_moved
, active_balance
= 0;
9613 struct sched_domain
*sd_parent
= sd
->parent
;
9614 struct sched_group
*group
;
9617 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
9619 struct lb_env env
= {
9621 .dst_cpu
= this_cpu
,
9623 .dst_grpmask
= sched_group_span(sd
->groups
),
9625 .loop_break
= sched_nr_migrate_break
,
9628 .tasks
= LIST_HEAD_INIT(env
.tasks
),
9631 cpumask_and(cpus
, sched_domain_span(sd
), cpu_active_mask
);
9633 schedstat_inc(sd
->lb_count
[idle
]);
9636 if (!should_we_balance(&env
)) {
9637 *continue_balancing
= 0;
9641 group
= find_busiest_group(&env
);
9643 schedstat_inc(sd
->lb_nobusyg
[idle
]);
9647 busiest
= find_busiest_queue(&env
, group
);
9649 schedstat_inc(sd
->lb_nobusyq
[idle
]);
9653 BUG_ON(busiest
== env
.dst_rq
);
9655 schedstat_add(sd
->lb_imbalance
[idle
], env
.imbalance
);
9657 env
.src_cpu
= busiest
->cpu
;
9658 env
.src_rq
= busiest
;
9661 if (busiest
->nr_running
> 1) {
9663 * Attempt to move tasks. If find_busiest_group has found
9664 * an imbalance but busiest->nr_running <= 1, the group is
9665 * still unbalanced. ld_moved simply stays zero, so it is
9666 * correctly treated as an imbalance.
9668 env
.flags
|= LBF_ALL_PINNED
;
9669 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
9672 rq_lock_irqsave(busiest
, &rf
);
9673 update_rq_clock(busiest
);
9676 * cur_ld_moved - load moved in current iteration
9677 * ld_moved - cumulative load moved across iterations
9679 cur_ld_moved
= detach_tasks(&env
);
9682 * We've detached some tasks from busiest_rq. Every
9683 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9684 * unlock busiest->lock, and we are able to be sure
9685 * that nobody can manipulate the tasks in parallel.
9686 * See task_rq_lock() family for the details.
9689 rq_unlock(busiest
, &rf
);
9693 ld_moved
+= cur_ld_moved
;
9696 local_irq_restore(rf
.flags
);
9698 if (env
.flags
& LBF_NEED_BREAK
) {
9699 env
.flags
&= ~LBF_NEED_BREAK
;
9704 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9705 * us and move them to an alternate dst_cpu in our sched_group
9706 * where they can run. The upper limit on how many times we
9707 * iterate on same src_cpu is dependent on number of CPUs in our
9710 * This changes load balance semantics a bit on who can move
9711 * load to a given_cpu. In addition to the given_cpu itself
9712 * (or a ilb_cpu acting on its behalf where given_cpu is
9713 * nohz-idle), we now have balance_cpu in a position to move
9714 * load to given_cpu. In rare situations, this may cause
9715 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9716 * _independently_ and at _same_ time to move some load to
9717 * given_cpu) causing exceess load to be moved to given_cpu.
9718 * This however should not happen so much in practice and
9719 * moreover subsequent load balance cycles should correct the
9720 * excess load moved.
9722 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
9724 /* Prevent to re-select dst_cpu via env's CPUs */
9725 __cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
9727 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
9728 env
.dst_cpu
= env
.new_dst_cpu
;
9729 env
.flags
&= ~LBF_DST_PINNED
;
9731 env
.loop_break
= sched_nr_migrate_break
;
9734 * Go back to "more_balance" rather than "redo" since we
9735 * need to continue with same src_cpu.
9741 * We failed to reach balance because of affinity.
9744 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
9746 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
9747 *group_imbalance
= 1;
9750 /* All tasks on this runqueue were pinned by CPU affinity */
9751 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
9752 __cpumask_clear_cpu(cpu_of(busiest
), cpus
);
9754 * Attempting to continue load balancing at the current
9755 * sched_domain level only makes sense if there are
9756 * active CPUs remaining as possible busiest CPUs to
9757 * pull load from which are not contained within the
9758 * destination group that is receiving any migrated
9761 if (!cpumask_subset(cpus
, env
.dst_grpmask
)) {
9763 env
.loop_break
= sched_nr_migrate_break
;
9766 goto out_all_pinned
;
9771 schedstat_inc(sd
->lb_failed
[idle
]);
9773 * Increment the failure counter only on periodic balance.
9774 * We do not want newidle balance, which can be very
9775 * frequent, pollute the failure counter causing
9776 * excessive cache_hot migrations and active balances.
9778 if (idle
!= CPU_NEWLY_IDLE
)
9779 sd
->nr_balance_failed
++;
9781 if (need_active_balance(&env
)) {
9782 unsigned long flags
;
9784 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
9787 * Don't kick the active_load_balance_cpu_stop,
9788 * if the curr task on busiest CPU can't be
9789 * moved to this_cpu:
9791 if (!cpumask_test_cpu(this_cpu
, busiest
->curr
->cpus_ptr
)) {
9792 raw_spin_unlock_irqrestore(&busiest
->lock
,
9794 env
.flags
|= LBF_ALL_PINNED
;
9795 goto out_one_pinned
;
9799 * ->active_balance synchronizes accesses to
9800 * ->active_balance_work. Once set, it's cleared
9801 * only after active load balance is finished.
9803 if (!busiest
->active_balance
) {
9804 busiest
->active_balance
= 1;
9805 busiest
->push_cpu
= this_cpu
;
9808 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
9810 if (active_balance
) {
9811 stop_one_cpu_nowait(cpu_of(busiest
),
9812 active_load_balance_cpu_stop
, busiest
,
9813 &busiest
->active_balance_work
);
9816 /* We've kicked active balancing, force task migration. */
9817 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
9820 sd
->nr_balance_failed
= 0;
9822 if (likely(!active_balance
) || voluntary_active_balance(&env
)) {
9823 /* We were unbalanced, so reset the balancing interval */
9824 sd
->balance_interval
= sd
->min_interval
;
9827 * If we've begun active balancing, start to back off. This
9828 * case may not be covered by the all_pinned logic if there
9829 * is only 1 task on the busy runqueue (because we don't call
9832 if (sd
->balance_interval
< sd
->max_interval
)
9833 sd
->balance_interval
*= 2;
9840 * We reach balance although we may have faced some affinity
9841 * constraints. Clear the imbalance flag only if other tasks got
9842 * a chance to move and fix the imbalance.
9844 if (sd_parent
&& !(env
.flags
& LBF_ALL_PINNED
)) {
9845 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
9847 if (*group_imbalance
)
9848 *group_imbalance
= 0;
9853 * We reach balance because all tasks are pinned at this level so
9854 * we can't migrate them. Let the imbalance flag set so parent level
9855 * can try to migrate them.
9857 schedstat_inc(sd
->lb_balanced
[idle
]);
9859 sd
->nr_balance_failed
= 0;
9865 * newidle_balance() disregards balance intervals, so we could
9866 * repeatedly reach this code, which would lead to balance_interval
9867 * skyrocketting in a short amount of time. Skip the balance_interval
9868 * increase logic to avoid that.
9870 if (env
.idle
== CPU_NEWLY_IDLE
)
9873 /* tune up the balancing interval */
9874 if ((env
.flags
& LBF_ALL_PINNED
&&
9875 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
9876 sd
->balance_interval
< sd
->max_interval
)
9877 sd
->balance_interval
*= 2;
9882 static inline unsigned long
9883 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
9885 unsigned long interval
= sd
->balance_interval
;
9888 interval
*= sd
->busy_factor
;
9890 /* scale ms to jiffies */
9891 interval
= msecs_to_jiffies(interval
);
9894 * Reduce likelihood of busy balancing at higher domains racing with
9895 * balancing at lower domains by preventing their balancing periods
9896 * from being multiples of each other.
9901 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
9907 update_next_balance(struct sched_domain
*sd
, unsigned long *next_balance
)
9909 unsigned long interval
, next
;
9911 /* used by idle balance, so cpu_busy = 0 */
9912 interval
= get_sd_balance_interval(sd
, 0);
9913 next
= sd
->last_balance
+ interval
;
9915 if (time_after(*next_balance
, next
))
9916 *next_balance
= next
;
9920 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9921 * running tasks off the busiest CPU onto idle CPUs. It requires at
9922 * least 1 task to be running on each physical CPU where possible, and
9923 * avoids physical / logical imbalances.
9925 static int active_load_balance_cpu_stop(void *data
)
9927 struct rq
*busiest_rq
= data
;
9928 int busiest_cpu
= cpu_of(busiest_rq
);
9929 int target_cpu
= busiest_rq
->push_cpu
;
9930 struct rq
*target_rq
= cpu_rq(target_cpu
);
9931 struct sched_domain
*sd
;
9932 struct task_struct
*p
= NULL
;
9935 rq_lock_irq(busiest_rq
, &rf
);
9937 * Between queueing the stop-work and running it is a hole in which
9938 * CPUs can become inactive. We should not move tasks from or to
9941 if (!cpu_active(busiest_cpu
) || !cpu_active(target_cpu
))
9944 /* Make sure the requested CPU hasn't gone down in the meantime: */
9945 if (unlikely(busiest_cpu
!= smp_processor_id() ||
9946 !busiest_rq
->active_balance
))
9949 /* Is there any task to move? */
9950 if (busiest_rq
->nr_running
<= 1)
9954 * This condition is "impossible", if it occurs
9955 * we need to fix it. Originally reported by
9956 * Bjorn Helgaas on a 128-CPU setup.
9958 BUG_ON(busiest_rq
== target_rq
);
9960 /* Search for an sd spanning us and the target CPU. */
9962 for_each_domain(target_cpu
, sd
) {
9963 if (cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
9968 struct lb_env env
= {
9970 .dst_cpu
= target_cpu
,
9971 .dst_rq
= target_rq
,
9972 .src_cpu
= busiest_rq
->cpu
,
9973 .src_rq
= busiest_rq
,
9976 * can_migrate_task() doesn't need to compute new_dst_cpu
9977 * for active balancing. Since we have CPU_IDLE, but no
9978 * @dst_grpmask we need to make that test go away with lying
9981 .flags
= LBF_DST_PINNED
,
9984 schedstat_inc(sd
->alb_count
);
9985 update_rq_clock(busiest_rq
);
9987 p
= detach_one_task(&env
);
9989 schedstat_inc(sd
->alb_pushed
);
9990 /* Active balancing done, reset the failure counter. */
9991 sd
->nr_balance_failed
= 0;
9993 schedstat_inc(sd
->alb_failed
);
9998 busiest_rq
->active_balance
= 0;
9999 rq_unlock(busiest_rq
, &rf
);
10002 attach_one_task(target_rq
, p
);
10004 local_irq_enable();
10009 static DEFINE_SPINLOCK(balancing
);
10012 * Scale the max load_balance interval with the number of CPUs in the system.
10013 * This trades load-balance latency on larger machines for less cross talk.
10015 void update_max_interval(void)
10017 max_load_balance_interval
= HZ
*num_online_cpus()/10;
10021 * It checks each scheduling domain to see if it is due to be balanced,
10022 * and initiates a balancing operation if so.
10024 * Balancing parameters are set up in init_sched_domains.
10026 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
10028 int continue_balancing
= 1;
10030 int busy
= idle
!= CPU_IDLE
&& !sched_idle_cpu(cpu
);
10031 unsigned long interval
;
10032 struct sched_domain
*sd
;
10033 /* Earliest time when we have to do rebalance again */
10034 unsigned long next_balance
= jiffies
+ 60*HZ
;
10035 int update_next_balance
= 0;
10036 int need_serialize
, need_decay
= 0;
10040 for_each_domain(cpu
, sd
) {
10042 * Decay the newidle max times here because this is a regular
10043 * visit to all the domains. Decay ~1% per second.
10045 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
10046 sd
->max_newidle_lb_cost
=
10047 (sd
->max_newidle_lb_cost
* 253) / 256;
10048 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
10051 max_cost
+= sd
->max_newidle_lb_cost
;
10054 * Stop the load balance at this level. There is another
10055 * CPU in our sched group which is doing load balancing more
10058 if (!continue_balancing
) {
10064 interval
= get_sd_balance_interval(sd
, busy
);
10066 need_serialize
= sd
->flags
& SD_SERIALIZE
;
10067 if (need_serialize
) {
10068 if (!spin_trylock(&balancing
))
10072 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
10073 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
10075 * The LBF_DST_PINNED logic could have changed
10076 * env->dst_cpu, so we can't know our idle
10077 * state even if we migrated tasks. Update it.
10079 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
10080 busy
= idle
!= CPU_IDLE
&& !sched_idle_cpu(cpu
);
10082 sd
->last_balance
= jiffies
;
10083 interval
= get_sd_balance_interval(sd
, busy
);
10085 if (need_serialize
)
10086 spin_unlock(&balancing
);
10088 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
10089 next_balance
= sd
->last_balance
+ interval
;
10090 update_next_balance
= 1;
10095 * Ensure the rq-wide value also decays but keep it at a
10096 * reasonable floor to avoid funnies with rq->avg_idle.
10098 rq
->max_idle_balance_cost
=
10099 max((u64
)sysctl_sched_migration_cost
, max_cost
);
10104 * next_balance will be updated only when there is a need.
10105 * When the cpu is attached to null domain for ex, it will not be
10108 if (likely(update_next_balance
)) {
10109 rq
->next_balance
= next_balance
;
10111 #ifdef CONFIG_NO_HZ_COMMON
10113 * If this CPU has been elected to perform the nohz idle
10114 * balance. Other idle CPUs have already rebalanced with
10115 * nohz_idle_balance() and nohz.next_balance has been
10116 * updated accordingly. This CPU is now running the idle load
10117 * balance for itself and we need to update the
10118 * nohz.next_balance accordingly.
10120 if ((idle
== CPU_IDLE
) && time_after(nohz
.next_balance
, rq
->next_balance
))
10121 nohz
.next_balance
= rq
->next_balance
;
10126 static inline int on_null_domain(struct rq
*rq
)
10128 return unlikely(!rcu_dereference_sched(rq
->sd
));
10131 #ifdef CONFIG_NO_HZ_COMMON
10133 * idle load balancing details
10134 * - When one of the busy CPUs notice that there may be an idle rebalancing
10135 * needed, they will kick the idle load balancer, which then does idle
10136 * load balancing for all the idle CPUs.
10137 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
10141 static inline int find_new_ilb(void)
10145 for_each_cpu_and(ilb
, nohz
.idle_cpus_mask
,
10146 housekeeping_cpumask(HK_FLAG_MISC
)) {
10148 if (ilb
== smp_processor_id())
10159 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
10160 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
10162 static void kick_ilb(unsigned int flags
)
10167 * Increase nohz.next_balance only when if full ilb is triggered but
10168 * not if we only update stats.
10170 if (flags
& NOHZ_BALANCE_KICK
)
10171 nohz
.next_balance
= jiffies
+1;
10173 ilb_cpu
= find_new_ilb();
10175 if (ilb_cpu
>= nr_cpu_ids
)
10179 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
10180 * the first flag owns it; cleared by nohz_csd_func().
10182 flags
= atomic_fetch_or(flags
, nohz_flags(ilb_cpu
));
10183 if (flags
& NOHZ_KICK_MASK
)
10187 * This way we generate an IPI on the target CPU which
10188 * is idle. And the softirq performing nohz idle load balance
10189 * will be run before returning from the IPI.
10191 smp_call_function_single_async(ilb_cpu
, &cpu_rq(ilb_cpu
)->nohz_csd
);
10195 * Current decision point for kicking the idle load balancer in the presence
10196 * of idle CPUs in the system.
10198 static void nohz_balancer_kick(struct rq
*rq
)
10200 unsigned long now
= jiffies
;
10201 struct sched_domain_shared
*sds
;
10202 struct sched_domain
*sd
;
10203 int nr_busy
, i
, cpu
= rq
->cpu
;
10204 unsigned int flags
= 0;
10206 if (unlikely(rq
->idle_balance
))
10210 * We may be recently in ticked or tickless idle mode. At the first
10211 * busy tick after returning from idle, we will update the busy stats.
10213 nohz_balance_exit_idle(rq
);
10216 * None are in tickless mode and hence no need for NOHZ idle load
10219 if (likely(!atomic_read(&nohz
.nr_cpus
)))
10222 if (READ_ONCE(nohz
.has_blocked
) &&
10223 time_after(now
, READ_ONCE(nohz
.next_blocked
)))
10224 flags
= NOHZ_STATS_KICK
;
10226 if (time_before(now
, nohz
.next_balance
))
10229 if (rq
->nr_running
>= 2) {
10230 flags
= NOHZ_KICK_MASK
;
10236 sd
= rcu_dereference(rq
->sd
);
10239 * If there's a CFS task and the current CPU has reduced
10240 * capacity; kick the ILB to see if there's a better CPU to run
10243 if (rq
->cfs
.h_nr_running
>= 1 && check_cpu_capacity(rq
, sd
)) {
10244 flags
= NOHZ_KICK_MASK
;
10249 sd
= rcu_dereference(per_cpu(sd_asym_packing
, cpu
));
10252 * When ASYM_PACKING; see if there's a more preferred CPU
10253 * currently idle; in which case, kick the ILB to move tasks
10256 for_each_cpu_and(i
, sched_domain_span(sd
), nohz
.idle_cpus_mask
) {
10257 if (sched_asym_prefer(i
, cpu
)) {
10258 flags
= NOHZ_KICK_MASK
;
10264 sd
= rcu_dereference(per_cpu(sd_asym_cpucapacity
, cpu
));
10267 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10268 * to run the misfit task on.
10270 if (check_misfit_status(rq
, sd
)) {
10271 flags
= NOHZ_KICK_MASK
;
10276 * For asymmetric systems, we do not want to nicely balance
10277 * cache use, instead we want to embrace asymmetry and only
10278 * ensure tasks have enough CPU capacity.
10280 * Skip the LLC logic because it's not relevant in that case.
10285 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
10288 * If there is an imbalance between LLC domains (IOW we could
10289 * increase the overall cache use), we need some less-loaded LLC
10290 * domain to pull some load. Likewise, we may need to spread
10291 * load within the current LLC domain (e.g. packed SMT cores but
10292 * other CPUs are idle). We can't really know from here how busy
10293 * the others are - so just get a nohz balance going if it looks
10294 * like this LLC domain has tasks we could move.
10296 nr_busy
= atomic_read(&sds
->nr_busy_cpus
);
10298 flags
= NOHZ_KICK_MASK
;
10309 static void set_cpu_sd_state_busy(int cpu
)
10311 struct sched_domain
*sd
;
10314 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
10316 if (!sd
|| !sd
->nohz_idle
)
10320 atomic_inc(&sd
->shared
->nr_busy_cpus
);
10325 void nohz_balance_exit_idle(struct rq
*rq
)
10327 SCHED_WARN_ON(rq
!= this_rq());
10329 if (likely(!rq
->nohz_tick_stopped
))
10332 rq
->nohz_tick_stopped
= 0;
10333 cpumask_clear_cpu(rq
->cpu
, nohz
.idle_cpus_mask
);
10334 atomic_dec(&nohz
.nr_cpus
);
10336 set_cpu_sd_state_busy(rq
->cpu
);
10339 static void set_cpu_sd_state_idle(int cpu
)
10341 struct sched_domain
*sd
;
10344 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
10346 if (!sd
|| sd
->nohz_idle
)
10350 atomic_dec(&sd
->shared
->nr_busy_cpus
);
10356 * This routine will record that the CPU is going idle with tick stopped.
10357 * This info will be used in performing idle load balancing in the future.
10359 void nohz_balance_enter_idle(int cpu
)
10361 struct rq
*rq
= cpu_rq(cpu
);
10363 SCHED_WARN_ON(cpu
!= smp_processor_id());
10365 /* If this CPU is going down, then nothing needs to be done: */
10366 if (!cpu_active(cpu
))
10369 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10370 if (!housekeeping_cpu(cpu
, HK_FLAG_SCHED
))
10374 * Can be set safely without rq->lock held
10375 * If a clear happens, it will have evaluated last additions because
10376 * rq->lock is held during the check and the clear
10378 rq
->has_blocked_load
= 1;
10381 * The tick is still stopped but load could have been added in the
10382 * meantime. We set the nohz.has_blocked flag to trig a check of the
10383 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10384 * of nohz.has_blocked can only happen after checking the new load
10386 if (rq
->nohz_tick_stopped
)
10389 /* If we're a completely isolated CPU, we don't play: */
10390 if (on_null_domain(rq
))
10393 rq
->nohz_tick_stopped
= 1;
10395 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
10396 atomic_inc(&nohz
.nr_cpus
);
10399 * Ensures that if nohz_idle_balance() fails to observe our
10400 * @idle_cpus_mask store, it must observe the @has_blocked
10403 smp_mb__after_atomic();
10405 set_cpu_sd_state_idle(cpu
);
10409 * Each time a cpu enter idle, we assume that it has blocked load and
10410 * enable the periodic update of the load of idle cpus
10412 WRITE_ONCE(nohz
.has_blocked
, 1);
10416 * Internal function that runs load balance for all idle cpus. The load balance
10417 * can be a simple update of blocked load or a complete load balance with
10418 * tasks movement depending of flags.
10419 * The function returns false if the loop has stopped before running
10420 * through all idle CPUs.
10422 static bool _nohz_idle_balance(struct rq
*this_rq
, unsigned int flags
,
10423 enum cpu_idle_type idle
)
10425 /* Earliest time when we have to do rebalance again */
10426 unsigned long now
= jiffies
;
10427 unsigned long next_balance
= now
+ 60*HZ
;
10428 bool has_blocked_load
= false;
10429 int update_next_balance
= 0;
10430 int this_cpu
= this_rq
->cpu
;
10435 SCHED_WARN_ON((flags
& NOHZ_KICK_MASK
) == NOHZ_BALANCE_KICK
);
10438 * We assume there will be no idle load after this update and clear
10439 * the has_blocked flag. If a cpu enters idle in the mean time, it will
10440 * set the has_blocked flag and trig another update of idle load.
10441 * Because a cpu that becomes idle, is added to idle_cpus_mask before
10442 * setting the flag, we are sure to not clear the state and not
10443 * check the load of an idle cpu.
10445 WRITE_ONCE(nohz
.has_blocked
, 0);
10448 * Ensures that if we miss the CPU, we must see the has_blocked
10449 * store from nohz_balance_enter_idle().
10453 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
10454 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
10458 * If this CPU gets work to do, stop the load balancing
10459 * work being done for other CPUs. Next load
10460 * balancing owner will pick it up.
10462 if (need_resched()) {
10463 has_blocked_load
= true;
10467 rq
= cpu_rq(balance_cpu
);
10469 has_blocked_load
|= update_nohz_stats(rq
, true);
10472 * If time for next balance is due,
10475 if (time_after_eq(jiffies
, rq
->next_balance
)) {
10476 struct rq_flags rf
;
10478 rq_lock_irqsave(rq
, &rf
);
10479 update_rq_clock(rq
);
10480 rq_unlock_irqrestore(rq
, &rf
);
10482 if (flags
& NOHZ_BALANCE_KICK
)
10483 rebalance_domains(rq
, CPU_IDLE
);
10486 if (time_after(next_balance
, rq
->next_balance
)) {
10487 next_balance
= rq
->next_balance
;
10488 update_next_balance
= 1;
10493 * next_balance will be updated only when there is a need.
10494 * When the CPU is attached to null domain for ex, it will not be
10497 if (likely(update_next_balance
))
10498 nohz
.next_balance
= next_balance
;
10500 /* Newly idle CPU doesn't need an update */
10501 if (idle
!= CPU_NEWLY_IDLE
) {
10502 update_blocked_averages(this_cpu
);
10503 has_blocked_load
|= this_rq
->has_blocked_load
;
10506 if (flags
& NOHZ_BALANCE_KICK
)
10507 rebalance_domains(this_rq
, CPU_IDLE
);
10509 WRITE_ONCE(nohz
.next_blocked
,
10510 now
+ msecs_to_jiffies(LOAD_AVG_PERIOD
));
10512 /* The full idle balance loop has been done */
10516 /* There is still blocked load, enable periodic update */
10517 if (has_blocked_load
)
10518 WRITE_ONCE(nohz
.has_blocked
, 1);
10524 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
10525 * rebalancing for all the cpus for whom scheduler ticks are stopped.
10527 static bool nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
10529 unsigned int flags
= this_rq
->nohz_idle_balance
;
10534 this_rq
->nohz_idle_balance
= 0;
10536 if (idle
!= CPU_IDLE
)
10539 _nohz_idle_balance(this_rq
, flags
, idle
);
10544 static void nohz_newidle_balance(struct rq
*this_rq
)
10546 int this_cpu
= this_rq
->cpu
;
10549 * This CPU doesn't want to be disturbed by scheduler
10552 if (!housekeeping_cpu(this_cpu
, HK_FLAG_SCHED
))
10555 /* Will wake up very soon. No time for doing anything else*/
10556 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
10559 /* Don't need to update blocked load of idle CPUs*/
10560 if (!READ_ONCE(nohz
.has_blocked
) ||
10561 time_before(jiffies
, READ_ONCE(nohz
.next_blocked
)))
10564 raw_spin_unlock(&this_rq
->lock
);
10566 * This CPU is going to be idle and blocked load of idle CPUs
10567 * need to be updated. Run the ilb locally as it is a good
10568 * candidate for ilb instead of waking up another idle CPU.
10569 * Kick an normal ilb if we failed to do the update.
10571 if (!_nohz_idle_balance(this_rq
, NOHZ_STATS_KICK
, CPU_NEWLY_IDLE
))
10572 kick_ilb(NOHZ_STATS_KICK
);
10573 raw_spin_lock(&this_rq
->lock
);
10576 #else /* !CONFIG_NO_HZ_COMMON */
10577 static inline void nohz_balancer_kick(struct rq
*rq
) { }
10579 static inline bool nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
10584 static inline void nohz_newidle_balance(struct rq
*this_rq
) { }
10585 #endif /* CONFIG_NO_HZ_COMMON */
10588 * newidle_balance is called by schedule() if this_cpu is about to become
10589 * idle. Attempts to pull tasks from other CPUs.
10592 * < 0 - we released the lock and there are !fair tasks present
10593 * 0 - failed, no new tasks
10594 * > 0 - success, new (fair) tasks present
10596 static int newidle_balance(struct rq
*this_rq
, struct rq_flags
*rf
)
10598 unsigned long next_balance
= jiffies
+ HZ
;
10599 int this_cpu
= this_rq
->cpu
;
10600 struct sched_domain
*sd
;
10601 int pulled_task
= 0;
10604 update_misfit_status(NULL
, this_rq
);
10606 * We must set idle_stamp _before_ calling idle_balance(), such that we
10607 * measure the duration of idle_balance() as idle time.
10609 this_rq
->idle_stamp
= rq_clock(this_rq
);
10612 * Do not pull tasks towards !active CPUs...
10614 if (!cpu_active(this_cpu
))
10618 * This is OK, because current is on_cpu, which avoids it being picked
10619 * for load-balance and preemption/IRQs are still disabled avoiding
10620 * further scheduler activity on it and we're being very careful to
10621 * re-start the picking loop.
10623 rq_unpin_lock(this_rq
, rf
);
10625 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
10626 !READ_ONCE(this_rq
->rd
->overload
)) {
10629 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
10631 update_next_balance(sd
, &next_balance
);
10634 nohz_newidle_balance(this_rq
);
10639 raw_spin_unlock(&this_rq
->lock
);
10641 update_blocked_averages(this_cpu
);
10643 for_each_domain(this_cpu
, sd
) {
10644 int continue_balancing
= 1;
10645 u64 t0
, domain_cost
;
10647 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
10648 update_next_balance(sd
, &next_balance
);
10652 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
10653 t0
= sched_clock_cpu(this_cpu
);
10655 pulled_task
= load_balance(this_cpu
, this_rq
,
10656 sd
, CPU_NEWLY_IDLE
,
10657 &continue_balancing
);
10659 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
10660 if (domain_cost
> sd
->max_newidle_lb_cost
)
10661 sd
->max_newidle_lb_cost
= domain_cost
;
10663 curr_cost
+= domain_cost
;
10666 update_next_balance(sd
, &next_balance
);
10669 * Stop searching for tasks to pull if there are
10670 * now runnable tasks on this rq.
10672 if (pulled_task
|| this_rq
->nr_running
> 0)
10677 raw_spin_lock(&this_rq
->lock
);
10679 if (curr_cost
> this_rq
->max_idle_balance_cost
)
10680 this_rq
->max_idle_balance_cost
= curr_cost
;
10684 * While browsing the domains, we released the rq lock, a task could
10685 * have been enqueued in the meantime. Since we're not going idle,
10686 * pretend we pulled a task.
10688 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
10691 /* Move the next balance forward */
10692 if (time_after(this_rq
->next_balance
, next_balance
))
10693 this_rq
->next_balance
= next_balance
;
10695 /* Is there a task of a high priority class? */
10696 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
10700 this_rq
->idle_stamp
= 0;
10702 rq_repin_lock(this_rq
, rf
);
10704 return pulled_task
;
10708 * run_rebalance_domains is triggered when needed from the scheduler tick.
10709 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10711 static __latent_entropy
void run_rebalance_domains(struct softirq_action
*h
)
10713 struct rq
*this_rq
= this_rq();
10714 enum cpu_idle_type idle
= this_rq
->idle_balance
?
10715 CPU_IDLE
: CPU_NOT_IDLE
;
10718 * If this CPU has a pending nohz_balance_kick, then do the
10719 * balancing on behalf of the other idle CPUs whose ticks are
10720 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10721 * give the idle CPUs a chance to load balance. Else we may
10722 * load balance only within the local sched_domain hierarchy
10723 * and abort nohz_idle_balance altogether if we pull some load.
10725 if (nohz_idle_balance(this_rq
, idle
))
10728 /* normal load balance */
10729 update_blocked_averages(this_rq
->cpu
);
10730 rebalance_domains(this_rq
, idle
);
10734 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10736 void trigger_load_balance(struct rq
*rq
)
10738 /* Don't need to rebalance while attached to NULL domain */
10739 if (unlikely(on_null_domain(rq
)))
10742 if (time_after_eq(jiffies
, rq
->next_balance
))
10743 raise_softirq(SCHED_SOFTIRQ
);
10745 nohz_balancer_kick(rq
);
10748 static void rq_online_fair(struct rq
*rq
)
10752 update_runtime_enabled(rq
);
10755 static void rq_offline_fair(struct rq
*rq
)
10759 /* Ensure any throttled groups are reachable by pick_next_task */
10760 unthrottle_offline_cfs_rqs(rq
);
10763 #endif /* CONFIG_SMP */
10766 * scheduler tick hitting a task of our scheduling class.
10768 * NOTE: This function can be called remotely by the tick offload that
10769 * goes along full dynticks. Therefore no local assumption can be made
10770 * and everything must be accessed through the @rq and @curr passed in
10773 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
10775 struct cfs_rq
*cfs_rq
;
10776 struct sched_entity
*se
= &curr
->se
;
10778 for_each_sched_entity(se
) {
10779 cfs_rq
= cfs_rq_of(se
);
10780 entity_tick(cfs_rq
, se
, queued
);
10783 if (static_branch_unlikely(&sched_numa_balancing
))
10784 task_tick_numa(rq
, curr
);
10786 update_misfit_status(curr
, rq
);
10787 update_overutilized_status(task_rq(curr
));
10791 * called on fork with the child task as argument from the parent's context
10792 * - child not yet on the tasklist
10793 * - preemption disabled
10795 static void task_fork_fair(struct task_struct
*p
)
10797 struct cfs_rq
*cfs_rq
;
10798 struct sched_entity
*se
= &p
->se
, *curr
;
10799 struct rq
*rq
= this_rq();
10800 struct rq_flags rf
;
10803 update_rq_clock(rq
);
10805 cfs_rq
= task_cfs_rq(current
);
10806 curr
= cfs_rq
->curr
;
10808 update_curr(cfs_rq
);
10809 se
->vruntime
= curr
->vruntime
;
10811 place_entity(cfs_rq
, se
, 1);
10813 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
10815 * Upon rescheduling, sched_class::put_prev_task() will place
10816 * 'current' within the tree based on its new key value.
10818 swap(curr
->vruntime
, se
->vruntime
);
10822 se
->vruntime
-= cfs_rq
->min_vruntime
;
10823 rq_unlock(rq
, &rf
);
10827 * Priority of the task has changed. Check to see if we preempt
10828 * the current task.
10831 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
10833 if (!task_on_rq_queued(p
))
10836 if (rq
->cfs
.nr_running
== 1)
10840 * Reschedule if we are currently running on this runqueue and
10841 * our priority decreased, or if we are not currently running on
10842 * this runqueue and our priority is higher than the current's
10844 if (rq
->curr
== p
) {
10845 if (p
->prio
> oldprio
)
10848 check_preempt_curr(rq
, p
, 0);
10851 static inline bool vruntime_normalized(struct task_struct
*p
)
10853 struct sched_entity
*se
= &p
->se
;
10856 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10857 * the dequeue_entity(.flags=0) will already have normalized the
10864 * When !on_rq, vruntime of the task has usually NOT been normalized.
10865 * But there are some cases where it has already been normalized:
10867 * - A forked child which is waiting for being woken up by
10868 * wake_up_new_task().
10869 * - A task which has been woken up by try_to_wake_up() and
10870 * waiting for actually being woken up by sched_ttwu_pending().
10872 if (!se
->sum_exec_runtime
||
10873 (p
->state
== TASK_WAKING
&& p
->sched_remote_wakeup
))
10879 #ifdef CONFIG_FAIR_GROUP_SCHED
10881 * Propagate the changes of the sched_entity across the tg tree to make it
10882 * visible to the root
10884 static void propagate_entity_cfs_rq(struct sched_entity
*se
)
10886 struct cfs_rq
*cfs_rq
;
10888 /* Start to propagate at parent */
10891 for_each_sched_entity(se
) {
10892 cfs_rq
= cfs_rq_of(se
);
10894 if (cfs_rq_throttled(cfs_rq
))
10897 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
10901 static void propagate_entity_cfs_rq(struct sched_entity
*se
) { }
10904 static void detach_entity_cfs_rq(struct sched_entity
*se
)
10906 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10908 /* Catch up with the cfs_rq and remove our load when we leave */
10909 update_load_avg(cfs_rq
, se
, 0);
10910 detach_entity_load_avg(cfs_rq
, se
);
10911 update_tg_load_avg(cfs_rq
);
10912 propagate_entity_cfs_rq(se
);
10915 static void attach_entity_cfs_rq(struct sched_entity
*se
)
10917 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10919 #ifdef CONFIG_FAIR_GROUP_SCHED
10921 * Since the real-depth could have been changed (only FAIR
10922 * class maintain depth value), reset depth properly.
10924 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
10927 /* Synchronize entity with its cfs_rq */
10928 update_load_avg(cfs_rq
, se
, sched_feat(ATTACH_AGE_LOAD
) ? 0 : SKIP_AGE_LOAD
);
10929 attach_entity_load_avg(cfs_rq
, se
);
10930 update_tg_load_avg(cfs_rq
);
10931 propagate_entity_cfs_rq(se
);
10934 static void detach_task_cfs_rq(struct task_struct
*p
)
10936 struct sched_entity
*se
= &p
->se
;
10937 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10939 if (!vruntime_normalized(p
)) {
10941 * Fix up our vruntime so that the current sleep doesn't
10942 * cause 'unlimited' sleep bonus.
10944 place_entity(cfs_rq
, se
, 0);
10945 se
->vruntime
-= cfs_rq
->min_vruntime
;
10948 detach_entity_cfs_rq(se
);
10951 static void attach_task_cfs_rq(struct task_struct
*p
)
10953 struct sched_entity
*se
= &p
->se
;
10954 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10956 attach_entity_cfs_rq(se
);
10958 if (!vruntime_normalized(p
))
10959 se
->vruntime
+= cfs_rq
->min_vruntime
;
10962 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
10964 detach_task_cfs_rq(p
);
10967 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
10969 attach_task_cfs_rq(p
);
10971 if (task_on_rq_queued(p
)) {
10973 * We were most likely switched from sched_rt, so
10974 * kick off the schedule if running, otherwise just see
10975 * if we can still preempt the current task.
10980 check_preempt_curr(rq
, p
, 0);
10984 /* Account for a task changing its policy or group.
10986 * This routine is mostly called to set cfs_rq->curr field when a task
10987 * migrates between groups/classes.
10989 static void set_next_task_fair(struct rq
*rq
, struct task_struct
*p
, bool first
)
10991 struct sched_entity
*se
= &p
->se
;
10994 if (task_on_rq_queued(p
)) {
10996 * Move the next running task to the front of the list, so our
10997 * cfs_tasks list becomes MRU one.
10999 list_move(&se
->group_node
, &rq
->cfs_tasks
);
11003 for_each_sched_entity(se
) {
11004 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
11006 set_next_entity(cfs_rq
, se
);
11007 /* ensure bandwidth has been allocated on our new cfs_rq */
11008 account_cfs_rq_runtime(cfs_rq
, 0);
11012 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
11014 cfs_rq
->tasks_timeline
= RB_ROOT_CACHED
;
11015 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
11016 #ifndef CONFIG_64BIT
11017 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
11020 raw_spin_lock_init(&cfs_rq
->removed
.lock
);
11024 #ifdef CONFIG_FAIR_GROUP_SCHED
11025 static void task_set_group_fair(struct task_struct
*p
)
11027 struct sched_entity
*se
= &p
->se
;
11029 set_task_rq(p
, task_cpu(p
));
11030 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
11033 static void task_move_group_fair(struct task_struct
*p
)
11035 detach_task_cfs_rq(p
);
11036 set_task_rq(p
, task_cpu(p
));
11039 /* Tell se's cfs_rq has been changed -- migrated */
11040 p
->se
.avg
.last_update_time
= 0;
11042 attach_task_cfs_rq(p
);
11045 static void task_change_group_fair(struct task_struct
*p
, int type
)
11048 case TASK_SET_GROUP
:
11049 task_set_group_fair(p
);
11052 case TASK_MOVE_GROUP
:
11053 task_move_group_fair(p
);
11058 void free_fair_sched_group(struct task_group
*tg
)
11062 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
11064 for_each_possible_cpu(i
) {
11066 kfree(tg
->cfs_rq
[i
]);
11075 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
11077 struct sched_entity
*se
;
11078 struct cfs_rq
*cfs_rq
;
11081 tg
->cfs_rq
= kcalloc(nr_cpu_ids
, sizeof(cfs_rq
), GFP_KERNEL
);
11084 tg
->se
= kcalloc(nr_cpu_ids
, sizeof(se
), GFP_KERNEL
);
11088 tg
->shares
= NICE_0_LOAD
;
11090 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
11092 for_each_possible_cpu(i
) {
11093 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
11094 GFP_KERNEL
, cpu_to_node(i
));
11098 se
= kzalloc_node(sizeof(struct sched_entity
),
11099 GFP_KERNEL
, cpu_to_node(i
));
11103 init_cfs_rq(cfs_rq
);
11104 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
11105 init_entity_runnable_average(se
);
11116 void online_fair_sched_group(struct task_group
*tg
)
11118 struct sched_entity
*se
;
11119 struct rq_flags rf
;
11123 for_each_possible_cpu(i
) {
11126 rq_lock_irq(rq
, &rf
);
11127 update_rq_clock(rq
);
11128 attach_entity_cfs_rq(se
);
11129 sync_throttle(tg
, i
);
11130 rq_unlock_irq(rq
, &rf
);
11134 void unregister_fair_sched_group(struct task_group
*tg
)
11136 unsigned long flags
;
11140 for_each_possible_cpu(cpu
) {
11142 remove_entity_load_avg(tg
->se
[cpu
]);
11145 * Only empty task groups can be destroyed; so we can speculatively
11146 * check on_list without danger of it being re-added.
11148 if (!tg
->cfs_rq
[cpu
]->on_list
)
11153 raw_spin_lock_irqsave(&rq
->lock
, flags
);
11154 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
11155 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
11159 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
11160 struct sched_entity
*se
, int cpu
,
11161 struct sched_entity
*parent
)
11163 struct rq
*rq
= cpu_rq(cpu
);
11167 init_cfs_rq_runtime(cfs_rq
);
11169 tg
->cfs_rq
[cpu
] = cfs_rq
;
11172 /* se could be NULL for root_task_group */
11177 se
->cfs_rq
= &rq
->cfs
;
11180 se
->cfs_rq
= parent
->my_q
;
11181 se
->depth
= parent
->depth
+ 1;
11185 /* guarantee group entities always have weight */
11186 update_load_set(&se
->load
, NICE_0_LOAD
);
11187 se
->parent
= parent
;
11190 static DEFINE_MUTEX(shares_mutex
);
11192 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
11197 * We can't change the weight of the root cgroup.
11202 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
11204 mutex_lock(&shares_mutex
);
11205 if (tg
->shares
== shares
)
11208 tg
->shares
= shares
;
11209 for_each_possible_cpu(i
) {
11210 struct rq
*rq
= cpu_rq(i
);
11211 struct sched_entity
*se
= tg
->se
[i
];
11212 struct rq_flags rf
;
11214 /* Propagate contribution to hierarchy */
11215 rq_lock_irqsave(rq
, &rf
);
11216 update_rq_clock(rq
);
11217 for_each_sched_entity(se
) {
11218 update_load_avg(cfs_rq_of(se
), se
, UPDATE_TG
);
11219 update_cfs_group(se
);
11221 rq_unlock_irqrestore(rq
, &rf
);
11225 mutex_unlock(&shares_mutex
);
11228 #else /* CONFIG_FAIR_GROUP_SCHED */
11230 void free_fair_sched_group(struct task_group
*tg
) { }
11232 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
11237 void online_fair_sched_group(struct task_group
*tg
) { }
11239 void unregister_fair_sched_group(struct task_group
*tg
) { }
11241 #endif /* CONFIG_FAIR_GROUP_SCHED */
11244 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
11246 struct sched_entity
*se
= &task
->se
;
11247 unsigned int rr_interval
= 0;
11250 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11253 if (rq
->cfs
.load
.weight
)
11254 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
11256 return rr_interval
;
11260 * All the scheduling class methods:
11262 DEFINE_SCHED_CLASS(fair
) = {
11264 .enqueue_task
= enqueue_task_fair
,
11265 .dequeue_task
= dequeue_task_fair
,
11266 .yield_task
= yield_task_fair
,
11267 .yield_to_task
= yield_to_task_fair
,
11269 .check_preempt_curr
= check_preempt_wakeup
,
11271 .pick_next_task
= __pick_next_task_fair
,
11272 .put_prev_task
= put_prev_task_fair
,
11273 .set_next_task
= set_next_task_fair
,
11276 .balance
= balance_fair
,
11277 .select_task_rq
= select_task_rq_fair
,
11278 .migrate_task_rq
= migrate_task_rq_fair
,
11280 .rq_online
= rq_online_fair
,
11281 .rq_offline
= rq_offline_fair
,
11283 .task_dead
= task_dead_fair
,
11284 .set_cpus_allowed
= set_cpus_allowed_common
,
11287 .task_tick
= task_tick_fair
,
11288 .task_fork
= task_fork_fair
,
11290 .prio_changed
= prio_changed_fair
,
11291 .switched_from
= switched_from_fair
,
11292 .switched_to
= switched_to_fair
,
11294 .get_rr_interval
= get_rr_interval_fair
,
11296 .update_curr
= update_curr_fair
,
11298 #ifdef CONFIG_FAIR_GROUP_SCHED
11299 .task_change_group
= task_change_group_fair
,
11302 #ifdef CONFIG_UCLAMP_TASK
11303 .uclamp_enabled
= 1,
11307 #ifdef CONFIG_SCHED_DEBUG
11308 void print_cfs_stats(struct seq_file
*m
, int cpu
)
11310 struct cfs_rq
*cfs_rq
, *pos
;
11313 for_each_leaf_cfs_rq_safe(cpu_rq(cpu
), cfs_rq
, pos
)
11314 print_cfs_rq(m
, cpu
, cfs_rq
);
11318 #ifdef CONFIG_NUMA_BALANCING
11319 void show_numa_stats(struct task_struct
*p
, struct seq_file
*m
)
11322 unsigned long tsf
= 0, tpf
= 0, gsf
= 0, gpf
= 0;
11323 struct numa_group
*ng
;
11326 ng
= rcu_dereference(p
->numa_group
);
11327 for_each_online_node(node
) {
11328 if (p
->numa_faults
) {
11329 tsf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
11330 tpf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
11333 gsf
= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)],
11334 gpf
= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
11336 print_numa_stats(m
, node
, tsf
, tpf
, gsf
, gpf
);
11340 #endif /* CONFIG_NUMA_BALANCING */
11341 #endif /* CONFIG_SCHED_DEBUG */
11343 __init
void init_sched_fair_class(void)
11346 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
11348 #ifdef CONFIG_NO_HZ_COMMON
11349 nohz
.next_balance
= jiffies
;
11350 nohz
.next_blocked
= jiffies
;
11351 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
11358 * Helper functions to facilitate extracting info from tracepoints.
11361 const struct sched_avg
*sched_trace_cfs_rq_avg(struct cfs_rq
*cfs_rq
)
11364 return cfs_rq
? &cfs_rq
->avg
: NULL
;
11369 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg
);
11371 char *sched_trace_cfs_rq_path(struct cfs_rq
*cfs_rq
, char *str
, int len
)
11375 strlcpy(str
, "(null)", len
);
11380 cfs_rq_tg_path(cfs_rq
, str
, len
);
11383 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path
);
11385 int sched_trace_cfs_rq_cpu(struct cfs_rq
*cfs_rq
)
11387 return cfs_rq
? cpu_of(rq_of(cfs_rq
)) : -1;
11389 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu
);
11391 const struct sched_avg
*sched_trace_rq_avg_rt(struct rq
*rq
)
11394 return rq
? &rq
->avg_rt
: NULL
;
11399 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt
);
11401 const struct sched_avg
*sched_trace_rq_avg_dl(struct rq
*rq
)
11404 return rq
? &rq
->avg_dl
: NULL
;
11409 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl
);
11411 const struct sched_avg
*sched_trace_rq_avg_irq(struct rq
*rq
)
11413 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
11414 return rq
? &rq
->avg_irq
: NULL
;
11419 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq
);
11421 int sched_trace_rq_cpu(struct rq
*rq
)
11423 return rq
? cpu_of(rq
) : -1;
11425 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu
);
11427 int sched_trace_rq_cpu_capacity(struct rq
*rq
)
11433 SCHED_CAPACITY_SCALE
11437 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity
);
11439 const struct cpumask
*sched_trace_rd_span(struct root_domain
*rd
)
11442 return rd
? rd
->span
: NULL
;
11447 EXPORT_SYMBOL_GPL(sched_trace_rd_span
);
11449 int sched_trace_rq_nr_running(struct rq
*rq
)
11451 return rq
? rq
->nr_running
: -1;
11453 EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running
);