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 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
705 for_each_sched_entity(se
) {
706 struct load_weight
*load
;
707 struct load_weight lw
;
709 cfs_rq
= cfs_rq_of(se
);
710 load
= &cfs_rq
->load
;
712 if (unlikely(!se
->on_rq
)) {
715 update_load_add(&lw
, se
->load
.weight
);
718 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
724 * We calculate the vruntime slice of a to-be-inserted task.
728 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
730 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
736 static int select_idle_sibling(struct task_struct
*p
, int prev_cpu
, int cpu
);
737 static unsigned long task_h_load(struct task_struct
*p
);
738 static unsigned long capacity_of(int cpu
);
740 /* Give new sched_entity start runnable values to heavy its load in infant time */
741 void init_entity_runnable_average(struct sched_entity
*se
)
743 struct sched_avg
*sa
= &se
->avg
;
745 memset(sa
, 0, sizeof(*sa
));
748 * Tasks are initialized with full load to be seen as heavy tasks until
749 * they get a chance to stabilize to their real load level.
750 * Group entities are initialized with zero load to reflect the fact that
751 * nothing has been attached to the task group yet.
753 if (entity_is_task(se
))
754 sa
->load_avg
= scale_load_down(se
->load
.weight
);
756 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
759 static void attach_entity_cfs_rq(struct sched_entity
*se
);
762 * With new tasks being created, their initial util_avgs are extrapolated
763 * based on the cfs_rq's current util_avg:
765 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
767 * However, in many cases, the above util_avg does not give a desired
768 * value. Moreover, the sum of the util_avgs may be divergent, such
769 * as when the series is a harmonic series.
771 * To solve this problem, we also cap the util_avg of successive tasks to
772 * only 1/2 of the left utilization budget:
774 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
776 * where n denotes the nth task and cpu_scale the CPU capacity.
778 * For example, for a CPU with 1024 of capacity, a simplest series from
779 * the beginning would be like:
781 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
782 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
784 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
785 * if util_avg > util_avg_cap.
787 void post_init_entity_util_avg(struct task_struct
*p
)
789 struct sched_entity
*se
= &p
->se
;
790 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
791 struct sched_avg
*sa
= &se
->avg
;
792 long cpu_scale
= arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq
)));
793 long cap
= (long)(cpu_scale
- cfs_rq
->avg
.util_avg
) / 2;
796 if (cfs_rq
->avg
.util_avg
!= 0) {
797 sa
->util_avg
= cfs_rq
->avg
.util_avg
* se
->load
.weight
;
798 sa
->util_avg
/= (cfs_rq
->avg
.load_avg
+ 1);
800 if (sa
->util_avg
> cap
)
807 sa
->runnable_avg
= sa
->util_avg
;
809 if (p
->sched_class
!= &fair_sched_class
) {
811 * For !fair tasks do:
813 update_cfs_rq_load_avg(now, cfs_rq);
814 attach_entity_load_avg(cfs_rq, se);
815 switched_from_fair(rq, p);
817 * such that the next switched_to_fair() has the
820 se
->avg
.last_update_time
= cfs_rq_clock_pelt(cfs_rq
);
824 attach_entity_cfs_rq(se
);
827 #else /* !CONFIG_SMP */
828 void init_entity_runnable_average(struct sched_entity
*se
)
831 void post_init_entity_util_avg(struct task_struct
*p
)
834 static void update_tg_load_avg(struct cfs_rq
*cfs_rq
)
837 #endif /* CONFIG_SMP */
840 * Update the current task's runtime statistics.
842 static void update_curr(struct cfs_rq
*cfs_rq
)
844 struct sched_entity
*curr
= cfs_rq
->curr
;
845 u64 now
= rq_clock_task(rq_of(cfs_rq
));
851 delta_exec
= now
- curr
->exec_start
;
852 if (unlikely((s64
)delta_exec
<= 0))
855 curr
->exec_start
= now
;
857 schedstat_set(curr
->statistics
.exec_max
,
858 max(delta_exec
, curr
->statistics
.exec_max
));
860 curr
->sum_exec_runtime
+= delta_exec
;
861 schedstat_add(cfs_rq
->exec_clock
, delta_exec
);
863 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
864 update_min_vruntime(cfs_rq
);
866 if (entity_is_task(curr
)) {
867 struct task_struct
*curtask
= task_of(curr
);
869 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
870 cgroup_account_cputime(curtask
, delta_exec
);
871 account_group_exec_runtime(curtask
, delta_exec
);
874 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
877 static void update_curr_fair(struct rq
*rq
)
879 update_curr(cfs_rq_of(&rq
->curr
->se
));
883 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
885 u64 wait_start
, prev_wait_start
;
887 if (!schedstat_enabled())
890 wait_start
= rq_clock(rq_of(cfs_rq
));
891 prev_wait_start
= schedstat_val(se
->statistics
.wait_start
);
893 if (entity_is_task(se
) && task_on_rq_migrating(task_of(se
)) &&
894 likely(wait_start
> prev_wait_start
))
895 wait_start
-= prev_wait_start
;
897 __schedstat_set(se
->statistics
.wait_start
, wait_start
);
901 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
903 struct task_struct
*p
;
906 if (!schedstat_enabled())
910 * When the sched_schedstat changes from 0 to 1, some sched se
911 * maybe already in the runqueue, the se->statistics.wait_start
912 * will be 0.So it will let the delta wrong. We need to avoid this
915 if (unlikely(!schedstat_val(se
->statistics
.wait_start
)))
918 delta
= rq_clock(rq_of(cfs_rq
)) - schedstat_val(se
->statistics
.wait_start
);
920 if (entity_is_task(se
)) {
922 if (task_on_rq_migrating(p
)) {
924 * Preserve migrating task's wait time so wait_start
925 * time stamp can be adjusted to accumulate wait time
926 * prior to migration.
928 __schedstat_set(se
->statistics
.wait_start
, delta
);
931 trace_sched_stat_wait(p
, delta
);
934 __schedstat_set(se
->statistics
.wait_max
,
935 max(schedstat_val(se
->statistics
.wait_max
), delta
));
936 __schedstat_inc(se
->statistics
.wait_count
);
937 __schedstat_add(se
->statistics
.wait_sum
, delta
);
938 __schedstat_set(se
->statistics
.wait_start
, 0);
942 update_stats_enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
944 struct task_struct
*tsk
= NULL
;
945 u64 sleep_start
, block_start
;
947 if (!schedstat_enabled())
950 sleep_start
= schedstat_val(se
->statistics
.sleep_start
);
951 block_start
= schedstat_val(se
->statistics
.block_start
);
953 if (entity_is_task(se
))
957 u64 delta
= rq_clock(rq_of(cfs_rq
)) - sleep_start
;
962 if (unlikely(delta
> schedstat_val(se
->statistics
.sleep_max
)))
963 __schedstat_set(se
->statistics
.sleep_max
, delta
);
965 __schedstat_set(se
->statistics
.sleep_start
, 0);
966 __schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
969 account_scheduler_latency(tsk
, delta
>> 10, 1);
970 trace_sched_stat_sleep(tsk
, delta
);
974 u64 delta
= rq_clock(rq_of(cfs_rq
)) - block_start
;
979 if (unlikely(delta
> schedstat_val(se
->statistics
.block_max
)))
980 __schedstat_set(se
->statistics
.block_max
, delta
);
982 __schedstat_set(se
->statistics
.block_start
, 0);
983 __schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
986 if (tsk
->in_iowait
) {
987 __schedstat_add(se
->statistics
.iowait_sum
, delta
);
988 __schedstat_inc(se
->statistics
.iowait_count
);
989 trace_sched_stat_iowait(tsk
, delta
);
992 trace_sched_stat_blocked(tsk
, delta
);
995 * Blocking time is in units of nanosecs, so shift by
996 * 20 to get a milliseconds-range estimation of the
997 * amount of time that the task spent sleeping:
999 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
1000 profile_hits(SLEEP_PROFILING
,
1001 (void *)get_wchan(tsk
),
1004 account_scheduler_latency(tsk
, delta
>> 10, 0);
1010 * Task is being enqueued - update stats:
1013 update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1015 if (!schedstat_enabled())
1019 * Are we enqueueing a waiting task? (for current tasks
1020 * a dequeue/enqueue event is a NOP)
1022 if (se
!= cfs_rq
->curr
)
1023 update_stats_wait_start(cfs_rq
, se
);
1025 if (flags
& ENQUEUE_WAKEUP
)
1026 update_stats_enqueue_sleeper(cfs_rq
, se
);
1030 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1033 if (!schedstat_enabled())
1037 * Mark the end of the wait period if dequeueing a
1040 if (se
!= cfs_rq
->curr
)
1041 update_stats_wait_end(cfs_rq
, se
);
1043 if ((flags
& DEQUEUE_SLEEP
) && entity_is_task(se
)) {
1044 struct task_struct
*tsk
= task_of(se
);
1046 if (tsk
->state
& TASK_INTERRUPTIBLE
)
1047 __schedstat_set(se
->statistics
.sleep_start
,
1048 rq_clock(rq_of(cfs_rq
)));
1049 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
1050 __schedstat_set(se
->statistics
.block_start
,
1051 rq_clock(rq_of(cfs_rq
)));
1056 * We are picking a new current task - update its stats:
1059 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1062 * We are starting a new run period:
1064 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
1067 /**************************************************
1068 * Scheduling class queueing methods:
1071 #ifdef CONFIG_NUMA_BALANCING
1073 * Approximate time to scan a full NUMA task in ms. The task scan period is
1074 * calculated based on the tasks virtual memory size and
1075 * numa_balancing_scan_size.
1077 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
1078 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
1080 /* Portion of address space to scan in MB */
1081 unsigned int sysctl_numa_balancing_scan_size
= 256;
1083 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1084 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
1087 refcount_t refcount
;
1089 spinlock_t lock
; /* nr_tasks, tasks */
1094 struct rcu_head rcu
;
1095 unsigned long total_faults
;
1096 unsigned long max_faults_cpu
;
1098 * Faults_cpu is used to decide whether memory should move
1099 * towards the CPU. As a consequence, these stats are weighted
1100 * more by CPU use than by memory faults.
1102 unsigned long *faults_cpu
;
1103 unsigned long faults
[];
1107 * For functions that can be called in multiple contexts that permit reading
1108 * ->numa_group (see struct task_struct for locking rules).
1110 static struct numa_group
*deref_task_numa_group(struct task_struct
*p
)
1112 return rcu_dereference_check(p
->numa_group
, p
== current
||
1113 (lockdep_is_held(&task_rq(p
)->lock
) && !READ_ONCE(p
->on_cpu
)));
1116 static struct numa_group
*deref_curr_numa_group(struct task_struct
*p
)
1118 return rcu_dereference_protected(p
->numa_group
, p
== current
);
1121 static inline unsigned long group_faults_priv(struct numa_group
*ng
);
1122 static inline unsigned long group_faults_shared(struct numa_group
*ng
);
1124 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
1126 unsigned long rss
= 0;
1127 unsigned long nr_scan_pages
;
1130 * Calculations based on RSS as non-present and empty pages are skipped
1131 * by the PTE scanner and NUMA hinting faults should be trapped based
1134 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
1135 rss
= get_mm_rss(p
->mm
);
1137 rss
= nr_scan_pages
;
1139 rss
= round_up(rss
, nr_scan_pages
);
1140 return rss
/ nr_scan_pages
;
1143 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1144 #define MAX_SCAN_WINDOW 2560
1146 static unsigned int task_scan_min(struct task_struct
*p
)
1148 unsigned int scan_size
= READ_ONCE(sysctl_numa_balancing_scan_size
);
1149 unsigned int scan
, floor
;
1150 unsigned int windows
= 1;
1152 if (scan_size
< MAX_SCAN_WINDOW
)
1153 windows
= MAX_SCAN_WINDOW
/ scan_size
;
1154 floor
= 1000 / windows
;
1156 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
1157 return max_t(unsigned int, floor
, scan
);
1160 static unsigned int task_scan_start(struct task_struct
*p
)
1162 unsigned long smin
= task_scan_min(p
);
1163 unsigned long period
= smin
;
1164 struct numa_group
*ng
;
1166 /* Scale the maximum scan period with the amount of shared memory. */
1168 ng
= rcu_dereference(p
->numa_group
);
1170 unsigned long shared
= group_faults_shared(ng
);
1171 unsigned long private = group_faults_priv(ng
);
1173 period
*= refcount_read(&ng
->refcount
);
1174 period
*= shared
+ 1;
1175 period
/= private + shared
+ 1;
1179 return max(smin
, period
);
1182 static unsigned int task_scan_max(struct task_struct
*p
)
1184 unsigned long smin
= task_scan_min(p
);
1186 struct numa_group
*ng
;
1188 /* Watch for min being lower than max due to floor calculations */
1189 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
1191 /* Scale the maximum scan period with the amount of shared memory. */
1192 ng
= deref_curr_numa_group(p
);
1194 unsigned long shared
= group_faults_shared(ng
);
1195 unsigned long private = group_faults_priv(ng
);
1196 unsigned long period
= smax
;
1198 period
*= refcount_read(&ng
->refcount
);
1199 period
*= shared
+ 1;
1200 period
/= private + shared
+ 1;
1202 smax
= max(smax
, period
);
1205 return max(smin
, smax
);
1208 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
1210 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= NUMA_NO_NODE
);
1211 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
1214 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
1216 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= NUMA_NO_NODE
);
1217 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
1220 /* Shared or private faults. */
1221 #define NR_NUMA_HINT_FAULT_TYPES 2
1223 /* Memory and CPU locality */
1224 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1226 /* Averaged statistics, and temporary buffers. */
1227 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1229 pid_t
task_numa_group_id(struct task_struct
*p
)
1231 struct numa_group
*ng
;
1235 ng
= rcu_dereference(p
->numa_group
);
1244 * The averaged statistics, shared & private, memory & CPU,
1245 * occupy the first half of the array. The second half of the
1246 * array is for current counters, which are averaged into the
1247 * first set by task_numa_placement.
1249 static inline int task_faults_idx(enum numa_faults_stats s
, int nid
, int priv
)
1251 return NR_NUMA_HINT_FAULT_TYPES
* (s
* nr_node_ids
+ nid
) + priv
;
1254 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
1256 if (!p
->numa_faults
)
1259 return p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1260 p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1263 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
1265 struct numa_group
*ng
= deref_task_numa_group(p
);
1270 return ng
->faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1271 ng
->faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1274 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
1276 return group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1277 group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1280 static inline unsigned long group_faults_priv(struct numa_group
*ng
)
1282 unsigned long faults
= 0;
1285 for_each_online_node(node
) {
1286 faults
+= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
1292 static inline unsigned long group_faults_shared(struct numa_group
*ng
)
1294 unsigned long faults
= 0;
1297 for_each_online_node(node
) {
1298 faults
+= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
1305 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1306 * considered part of a numa group's pseudo-interleaving set. Migrations
1307 * between these nodes are slowed down, to allow things to settle down.
1309 #define ACTIVE_NODE_FRACTION 3
1311 static bool numa_is_active_node(int nid
, struct numa_group
*ng
)
1313 return group_faults_cpu(ng
, nid
) * ACTIVE_NODE_FRACTION
> ng
->max_faults_cpu
;
1316 /* Handle placement on systems where not all nodes are directly connected. */
1317 static unsigned long score_nearby_nodes(struct task_struct
*p
, int nid
,
1318 int maxdist
, bool task
)
1320 unsigned long score
= 0;
1324 * All nodes are directly connected, and the same distance
1325 * from each other. No need for fancy placement algorithms.
1327 if (sched_numa_topology_type
== NUMA_DIRECT
)
1331 * This code is called for each node, introducing N^2 complexity,
1332 * which should be ok given the number of nodes rarely exceeds 8.
1334 for_each_online_node(node
) {
1335 unsigned long faults
;
1336 int dist
= node_distance(nid
, node
);
1339 * The furthest away nodes in the system are not interesting
1340 * for placement; nid was already counted.
1342 if (dist
== sched_max_numa_distance
|| node
== nid
)
1346 * On systems with a backplane NUMA topology, compare groups
1347 * of nodes, and move tasks towards the group with the most
1348 * memory accesses. When comparing two nodes at distance
1349 * "hoplimit", only nodes closer by than "hoplimit" are part
1350 * of each group. Skip other nodes.
1352 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1356 /* Add up the faults from nearby nodes. */
1358 faults
= task_faults(p
, node
);
1360 faults
= group_faults(p
, node
);
1363 * On systems with a glueless mesh NUMA topology, there are
1364 * no fixed "groups of nodes". Instead, nodes that are not
1365 * directly connected bounce traffic through intermediate
1366 * nodes; a numa_group can occupy any set of nodes.
1367 * The further away a node is, the less the faults count.
1368 * This seems to result in good task placement.
1370 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1371 faults
*= (sched_max_numa_distance
- dist
);
1372 faults
/= (sched_max_numa_distance
- LOCAL_DISTANCE
);
1382 * These return the fraction of accesses done by a particular task, or
1383 * task group, on a particular numa node. The group weight is given a
1384 * larger multiplier, in order to group tasks together that are almost
1385 * evenly spread out between numa nodes.
1387 static inline unsigned long task_weight(struct task_struct
*p
, int nid
,
1390 unsigned long faults
, total_faults
;
1392 if (!p
->numa_faults
)
1395 total_faults
= p
->total_numa_faults
;
1400 faults
= task_faults(p
, nid
);
1401 faults
+= score_nearby_nodes(p
, nid
, dist
, true);
1403 return 1000 * faults
/ total_faults
;
1406 static inline unsigned long group_weight(struct task_struct
*p
, int nid
,
1409 struct numa_group
*ng
= deref_task_numa_group(p
);
1410 unsigned long faults
, total_faults
;
1415 total_faults
= ng
->total_faults
;
1420 faults
= group_faults(p
, nid
);
1421 faults
+= score_nearby_nodes(p
, nid
, dist
, false);
1423 return 1000 * faults
/ total_faults
;
1426 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
1427 int src_nid
, int dst_cpu
)
1429 struct numa_group
*ng
= deref_curr_numa_group(p
);
1430 int dst_nid
= cpu_to_node(dst_cpu
);
1431 int last_cpupid
, this_cpupid
;
1433 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
1434 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
1437 * Allow first faults or private faults to migrate immediately early in
1438 * the lifetime of a task. The magic number 4 is based on waiting for
1439 * two full passes of the "multi-stage node selection" test that is
1442 if ((p
->numa_preferred_nid
== NUMA_NO_NODE
|| p
->numa_scan_seq
<= 4) &&
1443 (cpupid_pid_unset(last_cpupid
) || cpupid_match_pid(p
, last_cpupid
)))
1447 * Multi-stage node selection is used in conjunction with a periodic
1448 * migration fault to build a temporal task<->page relation. By using
1449 * a two-stage filter we remove short/unlikely relations.
1451 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1452 * a task's usage of a particular page (n_p) per total usage of this
1453 * page (n_t) (in a given time-span) to a probability.
1455 * Our periodic faults will sample this probability and getting the
1456 * same result twice in a row, given these samples are fully
1457 * independent, is then given by P(n)^2, provided our sample period
1458 * is sufficiently short compared to the usage pattern.
1460 * This quadric squishes small probabilities, making it less likely we
1461 * act on an unlikely task<->page relation.
1463 if (!cpupid_pid_unset(last_cpupid
) &&
1464 cpupid_to_nid(last_cpupid
) != dst_nid
)
1467 /* Always allow migrate on private faults */
1468 if (cpupid_match_pid(p
, last_cpupid
))
1471 /* A shared fault, but p->numa_group has not been set up yet. */
1476 * Destination node is much more heavily used than the source
1477 * node? Allow migration.
1479 if (group_faults_cpu(ng
, dst_nid
) > group_faults_cpu(ng
, src_nid
) *
1480 ACTIVE_NODE_FRACTION
)
1484 * Distribute memory according to CPU & memory use on each node,
1485 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1487 * faults_cpu(dst) 3 faults_cpu(src)
1488 * --------------- * - > ---------------
1489 * faults_mem(dst) 4 faults_mem(src)
1491 return group_faults_cpu(ng
, dst_nid
) * group_faults(p
, src_nid
) * 3 >
1492 group_faults_cpu(ng
, src_nid
) * group_faults(p
, dst_nid
) * 4;
1496 * 'numa_type' describes the node at the moment of load balancing.
1499 /* The node has spare capacity that can be used to run more tasks. */
1502 * The node is fully used and the tasks don't compete for more CPU
1503 * cycles. Nevertheless, some tasks might wait before running.
1507 * The node is overloaded and can't provide expected CPU cycles to all
1513 /* Cached statistics for all CPUs within a node */
1516 unsigned long runnable
;
1518 /* Total compute capacity of CPUs on a node */
1519 unsigned long compute_capacity
;
1520 unsigned int nr_running
;
1521 unsigned int weight
;
1522 enum numa_type node_type
;
1526 static inline bool is_core_idle(int cpu
)
1528 #ifdef CONFIG_SCHED_SMT
1531 for_each_cpu(sibling
, cpu_smt_mask(cpu
)) {
1543 struct task_numa_env
{
1544 struct task_struct
*p
;
1546 int src_cpu
, src_nid
;
1547 int dst_cpu
, dst_nid
;
1549 struct numa_stats src_stats
, dst_stats
;
1554 struct task_struct
*best_task
;
1559 static unsigned long cpu_load(struct rq
*rq
);
1560 static unsigned long cpu_runnable(struct rq
*rq
);
1561 static unsigned long cpu_util(int cpu
);
1562 static inline long adjust_numa_imbalance(int imbalance
, int nr_running
);
1565 numa_type
numa_classify(unsigned int imbalance_pct
,
1566 struct numa_stats
*ns
)
1568 if ((ns
->nr_running
> ns
->weight
) &&
1569 (((ns
->compute_capacity
* 100) < (ns
->util
* imbalance_pct
)) ||
1570 ((ns
->compute_capacity
* imbalance_pct
) < (ns
->runnable
* 100))))
1571 return node_overloaded
;
1573 if ((ns
->nr_running
< ns
->weight
) ||
1574 (((ns
->compute_capacity
* 100) > (ns
->util
* imbalance_pct
)) &&
1575 ((ns
->compute_capacity
* imbalance_pct
) > (ns
->runnable
* 100))))
1576 return node_has_spare
;
1578 return node_fully_busy
;
1581 #ifdef CONFIG_SCHED_SMT
1582 /* Forward declarations of select_idle_sibling helpers */
1583 static inline bool test_idle_cores(int cpu
, bool def
);
1584 static inline int numa_idle_core(int idle_core
, int cpu
)
1586 if (!static_branch_likely(&sched_smt_present
) ||
1587 idle_core
>= 0 || !test_idle_cores(cpu
, false))
1591 * Prefer cores instead of packing HT siblings
1592 * and triggering future load balancing.
1594 if (is_core_idle(cpu
))
1600 static inline int numa_idle_core(int idle_core
, int cpu
)
1607 * Gather all necessary information to make NUMA balancing placement
1608 * decisions that are compatible with standard load balancer. This
1609 * borrows code and logic from update_sg_lb_stats but sharing a
1610 * common implementation is impractical.
1612 static void update_numa_stats(struct task_numa_env
*env
,
1613 struct numa_stats
*ns
, int nid
,
1616 int cpu
, idle_core
= -1;
1618 memset(ns
, 0, sizeof(*ns
));
1622 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1623 struct rq
*rq
= cpu_rq(cpu
);
1625 ns
->load
+= cpu_load(rq
);
1626 ns
->runnable
+= cpu_runnable(rq
);
1627 ns
->util
+= cpu_util(cpu
);
1628 ns
->nr_running
+= rq
->cfs
.h_nr_running
;
1629 ns
->compute_capacity
+= capacity_of(cpu
);
1631 if (find_idle
&& !rq
->nr_running
&& idle_cpu(cpu
)) {
1632 if (READ_ONCE(rq
->numa_migrate_on
) ||
1633 !cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
))
1636 if (ns
->idle_cpu
== -1)
1639 idle_core
= numa_idle_core(idle_core
, cpu
);
1644 ns
->weight
= cpumask_weight(cpumask_of_node(nid
));
1646 ns
->node_type
= numa_classify(env
->imbalance_pct
, ns
);
1649 ns
->idle_cpu
= idle_core
;
1652 static void task_numa_assign(struct task_numa_env
*env
,
1653 struct task_struct
*p
, long imp
)
1655 struct rq
*rq
= cpu_rq(env
->dst_cpu
);
1657 /* Check if run-queue part of active NUMA balance. */
1658 if (env
->best_cpu
!= env
->dst_cpu
&& xchg(&rq
->numa_migrate_on
, 1)) {
1660 int start
= env
->dst_cpu
;
1662 /* Find alternative idle CPU. */
1663 for_each_cpu_wrap(cpu
, cpumask_of_node(env
->dst_nid
), start
) {
1664 if (cpu
== env
->best_cpu
|| !idle_cpu(cpu
) ||
1665 !cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
)) {
1670 rq
= cpu_rq(env
->dst_cpu
);
1671 if (!xchg(&rq
->numa_migrate_on
, 1))
1675 /* Failed to find an alternative idle CPU */
1681 * Clear previous best_cpu/rq numa-migrate flag, since task now
1682 * found a better CPU to move/swap.
1684 if (env
->best_cpu
!= -1 && env
->best_cpu
!= env
->dst_cpu
) {
1685 rq
= cpu_rq(env
->best_cpu
);
1686 WRITE_ONCE(rq
->numa_migrate_on
, 0);
1690 put_task_struct(env
->best_task
);
1695 env
->best_imp
= imp
;
1696 env
->best_cpu
= env
->dst_cpu
;
1699 static bool load_too_imbalanced(long src_load
, long dst_load
,
1700 struct task_numa_env
*env
)
1703 long orig_src_load
, orig_dst_load
;
1704 long src_capacity
, dst_capacity
;
1707 * The load is corrected for the CPU capacity available on each node.
1710 * ------------ vs ---------
1711 * src_capacity dst_capacity
1713 src_capacity
= env
->src_stats
.compute_capacity
;
1714 dst_capacity
= env
->dst_stats
.compute_capacity
;
1716 imb
= abs(dst_load
* src_capacity
- src_load
* dst_capacity
);
1718 orig_src_load
= env
->src_stats
.load
;
1719 orig_dst_load
= env
->dst_stats
.load
;
1721 old_imb
= abs(orig_dst_load
* src_capacity
- orig_src_load
* dst_capacity
);
1723 /* Would this change make things worse? */
1724 return (imb
> old_imb
);
1728 * Maximum NUMA importance can be 1998 (2*999);
1729 * SMALLIMP @ 30 would be close to 1998/64.
1730 * Used to deter task migration.
1735 * This checks if the overall compute and NUMA accesses of the system would
1736 * be improved if the source tasks was migrated to the target dst_cpu taking
1737 * into account that it might be best if task running on the dst_cpu should
1738 * be exchanged with the source task
1740 static bool task_numa_compare(struct task_numa_env
*env
,
1741 long taskimp
, long groupimp
, bool maymove
)
1743 struct numa_group
*cur_ng
, *p_ng
= deref_curr_numa_group(env
->p
);
1744 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1745 long imp
= p_ng
? groupimp
: taskimp
;
1746 struct task_struct
*cur
;
1747 long src_load
, dst_load
;
1748 int dist
= env
->dist
;
1751 bool stopsearch
= false;
1753 if (READ_ONCE(dst_rq
->numa_migrate_on
))
1757 cur
= rcu_dereference(dst_rq
->curr
);
1758 if (cur
&& ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
)))
1762 * Because we have preemption enabled we can get migrated around and
1763 * end try selecting ourselves (current == env->p) as a swap candidate.
1765 if (cur
== env
->p
) {
1771 if (maymove
&& moveimp
>= env
->best_imp
)
1777 /* Skip this swap candidate if cannot move to the source cpu. */
1778 if (!cpumask_test_cpu(env
->src_cpu
, cur
->cpus_ptr
))
1782 * Skip this swap candidate if it is not moving to its preferred
1783 * node and the best task is.
1785 if (env
->best_task
&&
1786 env
->best_task
->numa_preferred_nid
== env
->src_nid
&&
1787 cur
->numa_preferred_nid
!= env
->src_nid
) {
1792 * "imp" is the fault differential for the source task between the
1793 * source and destination node. Calculate the total differential for
1794 * the source task and potential destination task. The more negative
1795 * the value is, the more remote accesses that would be expected to
1796 * be incurred if the tasks were swapped.
1798 * If dst and source tasks are in the same NUMA group, or not
1799 * in any group then look only at task weights.
1801 cur_ng
= rcu_dereference(cur
->numa_group
);
1802 if (cur_ng
== p_ng
) {
1803 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1804 task_weight(cur
, env
->dst_nid
, dist
);
1806 * Add some hysteresis to prevent swapping the
1807 * tasks within a group over tiny differences.
1813 * Compare the group weights. If a task is all by itself
1814 * (not part of a group), use the task weight instead.
1817 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1818 group_weight(cur
, env
->dst_nid
, dist
);
1820 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1821 task_weight(cur
, env
->dst_nid
, dist
);
1824 /* Discourage picking a task already on its preferred node */
1825 if (cur
->numa_preferred_nid
== env
->dst_nid
)
1829 * Encourage picking a task that moves to its preferred node.
1830 * This potentially makes imp larger than it's maximum of
1831 * 1998 (see SMALLIMP and task_weight for why) but in this
1832 * case, it does not matter.
1834 if (cur
->numa_preferred_nid
== env
->src_nid
)
1837 if (maymove
&& moveimp
> imp
&& moveimp
> env
->best_imp
) {
1844 * Prefer swapping with a task moving to its preferred node over a
1847 if (env
->best_task
&& cur
->numa_preferred_nid
== env
->src_nid
&&
1848 env
->best_task
->numa_preferred_nid
!= env
->src_nid
) {
1853 * If the NUMA importance is less than SMALLIMP,
1854 * task migration might only result in ping pong
1855 * of tasks and also hurt performance due to cache
1858 if (imp
< SMALLIMP
|| imp
<= env
->best_imp
+ SMALLIMP
/ 2)
1862 * In the overloaded case, try and keep the load balanced.
1864 load
= task_h_load(env
->p
) - task_h_load(cur
);
1868 dst_load
= env
->dst_stats
.load
+ load
;
1869 src_load
= env
->src_stats
.load
- load
;
1871 if (load_too_imbalanced(src_load
, dst_load
, env
))
1875 /* Evaluate an idle CPU for a task numa move. */
1877 int cpu
= env
->dst_stats
.idle_cpu
;
1879 /* Nothing cached so current CPU went idle since the search. */
1884 * If the CPU is no longer truly idle and the previous best CPU
1885 * is, keep using it.
1887 if (!idle_cpu(cpu
) && env
->best_cpu
>= 0 &&
1888 idle_cpu(env
->best_cpu
)) {
1889 cpu
= env
->best_cpu
;
1895 task_numa_assign(env
, cur
, imp
);
1898 * If a move to idle is allowed because there is capacity or load
1899 * balance improves then stop the search. While a better swap
1900 * candidate may exist, a search is not free.
1902 if (maymove
&& !cur
&& env
->best_cpu
>= 0 && idle_cpu(env
->best_cpu
))
1906 * If a swap candidate must be identified and the current best task
1907 * moves its preferred node then stop the search.
1909 if (!maymove
&& env
->best_task
&&
1910 env
->best_task
->numa_preferred_nid
== env
->src_nid
) {
1919 static void task_numa_find_cpu(struct task_numa_env
*env
,
1920 long taskimp
, long groupimp
)
1922 bool maymove
= false;
1926 * If dst node has spare capacity, then check if there is an
1927 * imbalance that would be overruled by the load balancer.
1929 if (env
->dst_stats
.node_type
== node_has_spare
) {
1930 unsigned int imbalance
;
1931 int src_running
, dst_running
;
1934 * Would movement cause an imbalance? Note that if src has
1935 * more running tasks that the imbalance is ignored as the
1936 * move improves the imbalance from the perspective of the
1937 * CPU load balancer.
1939 src_running
= env
->src_stats
.nr_running
- 1;
1940 dst_running
= env
->dst_stats
.nr_running
+ 1;
1941 imbalance
= max(0, dst_running
- src_running
);
1942 imbalance
= adjust_numa_imbalance(imbalance
, dst_running
);
1944 /* Use idle CPU if there is no imbalance */
1947 if (env
->dst_stats
.idle_cpu
>= 0) {
1948 env
->dst_cpu
= env
->dst_stats
.idle_cpu
;
1949 task_numa_assign(env
, NULL
, 0);
1954 long src_load
, dst_load
, load
;
1956 * If the improvement from just moving env->p direction is better
1957 * than swapping tasks around, check if a move is possible.
1959 load
= task_h_load(env
->p
);
1960 dst_load
= env
->dst_stats
.load
+ load
;
1961 src_load
= env
->src_stats
.load
- load
;
1962 maymove
= !load_too_imbalanced(src_load
, dst_load
, env
);
1965 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1966 /* Skip this CPU if the source task cannot migrate */
1967 if (!cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
))
1971 if (task_numa_compare(env
, taskimp
, groupimp
, maymove
))
1976 static int task_numa_migrate(struct task_struct
*p
)
1978 struct task_numa_env env
= {
1981 .src_cpu
= task_cpu(p
),
1982 .src_nid
= task_node(p
),
1984 .imbalance_pct
= 112,
1990 unsigned long taskweight
, groupweight
;
1991 struct sched_domain
*sd
;
1992 long taskimp
, groupimp
;
1993 struct numa_group
*ng
;
1998 * Pick the lowest SD_NUMA domain, as that would have the smallest
1999 * imbalance and would be the first to start moving tasks about.
2001 * And we want to avoid any moving of tasks about, as that would create
2002 * random movement of tasks -- counter the numa conditions we're trying
2006 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
2008 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
2012 * Cpusets can break the scheduler domain tree into smaller
2013 * balance domains, some of which do not cross NUMA boundaries.
2014 * Tasks that are "trapped" in such domains cannot be migrated
2015 * elsewhere, so there is no point in (re)trying.
2017 if (unlikely(!sd
)) {
2018 sched_setnuma(p
, task_node(p
));
2022 env
.dst_nid
= p
->numa_preferred_nid
;
2023 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
2024 taskweight
= task_weight(p
, env
.src_nid
, dist
);
2025 groupweight
= group_weight(p
, env
.src_nid
, dist
);
2026 update_numa_stats(&env
, &env
.src_stats
, env
.src_nid
, false);
2027 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
2028 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
2029 update_numa_stats(&env
, &env
.dst_stats
, env
.dst_nid
, true);
2031 /* Try to find a spot on the preferred nid. */
2032 task_numa_find_cpu(&env
, taskimp
, groupimp
);
2035 * Look at other nodes in these cases:
2036 * - there is no space available on the preferred_nid
2037 * - the task is part of a numa_group that is interleaved across
2038 * multiple NUMA nodes; in order to better consolidate the group,
2039 * we need to check other locations.
2041 ng
= deref_curr_numa_group(p
);
2042 if (env
.best_cpu
== -1 || (ng
&& ng
->active_nodes
> 1)) {
2043 for_each_online_node(nid
) {
2044 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
2047 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
2048 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
2050 taskweight
= task_weight(p
, env
.src_nid
, dist
);
2051 groupweight
= group_weight(p
, env
.src_nid
, dist
);
2054 /* Only consider nodes where both task and groups benefit */
2055 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
2056 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
2057 if (taskimp
< 0 && groupimp
< 0)
2062 update_numa_stats(&env
, &env
.dst_stats
, env
.dst_nid
, true);
2063 task_numa_find_cpu(&env
, taskimp
, groupimp
);
2068 * If the task is part of a workload that spans multiple NUMA nodes,
2069 * and is migrating into one of the workload's active nodes, remember
2070 * this node as the task's preferred numa node, so the workload can
2072 * A task that migrated to a second choice node will be better off
2073 * trying for a better one later. Do not set the preferred node here.
2076 if (env
.best_cpu
== -1)
2079 nid
= cpu_to_node(env
.best_cpu
);
2081 if (nid
!= p
->numa_preferred_nid
)
2082 sched_setnuma(p
, nid
);
2085 /* No better CPU than the current one was found. */
2086 if (env
.best_cpu
== -1) {
2087 trace_sched_stick_numa(p
, env
.src_cpu
, NULL
, -1);
2091 best_rq
= cpu_rq(env
.best_cpu
);
2092 if (env
.best_task
== NULL
) {
2093 ret
= migrate_task_to(p
, env
.best_cpu
);
2094 WRITE_ONCE(best_rq
->numa_migrate_on
, 0);
2096 trace_sched_stick_numa(p
, env
.src_cpu
, NULL
, env
.best_cpu
);
2100 ret
= migrate_swap(p
, env
.best_task
, env
.best_cpu
, env
.src_cpu
);
2101 WRITE_ONCE(best_rq
->numa_migrate_on
, 0);
2104 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_task
, env
.best_cpu
);
2105 put_task_struct(env
.best_task
);
2109 /* Attempt to migrate a task to a CPU on the preferred node. */
2110 static void numa_migrate_preferred(struct task_struct
*p
)
2112 unsigned long interval
= HZ
;
2114 /* This task has no NUMA fault statistics yet */
2115 if (unlikely(p
->numa_preferred_nid
== NUMA_NO_NODE
|| !p
->numa_faults
))
2118 /* Periodically retry migrating the task to the preferred node */
2119 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
2120 p
->numa_migrate_retry
= jiffies
+ interval
;
2122 /* Success if task is already running on preferred CPU */
2123 if (task_node(p
) == p
->numa_preferred_nid
)
2126 /* Otherwise, try migrate to a CPU on the preferred node */
2127 task_numa_migrate(p
);
2131 * Find out how many nodes on the workload is actively running on. Do this by
2132 * tracking the nodes from which NUMA hinting faults are triggered. This can
2133 * be different from the set of nodes where the workload's memory is currently
2136 static void numa_group_count_active_nodes(struct numa_group
*numa_group
)
2138 unsigned long faults
, max_faults
= 0;
2139 int nid
, active_nodes
= 0;
2141 for_each_online_node(nid
) {
2142 faults
= group_faults_cpu(numa_group
, nid
);
2143 if (faults
> max_faults
)
2144 max_faults
= faults
;
2147 for_each_online_node(nid
) {
2148 faults
= group_faults_cpu(numa_group
, nid
);
2149 if (faults
* ACTIVE_NODE_FRACTION
> max_faults
)
2153 numa_group
->max_faults_cpu
= max_faults
;
2154 numa_group
->active_nodes
= active_nodes
;
2158 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2159 * increments. The more local the fault statistics are, the higher the scan
2160 * period will be for the next scan window. If local/(local+remote) ratio is
2161 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2162 * the scan period will decrease. Aim for 70% local accesses.
2164 #define NUMA_PERIOD_SLOTS 10
2165 #define NUMA_PERIOD_THRESHOLD 7
2168 * Increase the scan period (slow down scanning) if the majority of
2169 * our memory is already on our local node, or if the majority of
2170 * the page accesses are shared with other processes.
2171 * Otherwise, decrease the scan period.
2173 static void update_task_scan_period(struct task_struct
*p
,
2174 unsigned long shared
, unsigned long private)
2176 unsigned int period_slot
;
2177 int lr_ratio
, ps_ratio
;
2180 unsigned long remote
= p
->numa_faults_locality
[0];
2181 unsigned long local
= p
->numa_faults_locality
[1];
2184 * If there were no record hinting faults then either the task is
2185 * completely idle or all activity is areas that are not of interest
2186 * to automatic numa balancing. Related to that, if there were failed
2187 * migration then it implies we are migrating too quickly or the local
2188 * node is overloaded. In either case, scan slower
2190 if (local
+ shared
== 0 || p
->numa_faults_locality
[2]) {
2191 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
2192 p
->numa_scan_period
<< 1);
2194 p
->mm
->numa_next_scan
= jiffies
+
2195 msecs_to_jiffies(p
->numa_scan_period
);
2201 * Prepare to scale scan period relative to the current period.
2202 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2203 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2204 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2206 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
2207 lr_ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
2208 ps_ratio
= (private * NUMA_PERIOD_SLOTS
) / (private + shared
);
2210 if (ps_ratio
>= NUMA_PERIOD_THRESHOLD
) {
2212 * Most memory accesses are local. There is no need to
2213 * do fast NUMA scanning, since memory is already local.
2215 int slot
= ps_ratio
- NUMA_PERIOD_THRESHOLD
;
2218 diff
= slot
* period_slot
;
2219 } else if (lr_ratio
>= NUMA_PERIOD_THRESHOLD
) {
2221 * Most memory accesses are shared with other tasks.
2222 * There is no point in continuing fast NUMA scanning,
2223 * since other tasks may just move the memory elsewhere.
2225 int slot
= lr_ratio
- NUMA_PERIOD_THRESHOLD
;
2228 diff
= slot
* period_slot
;
2231 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2232 * yet they are not on the local NUMA node. Speed up
2233 * NUMA scanning to get the memory moved over.
2235 int ratio
= max(lr_ratio
, ps_ratio
);
2236 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
2239 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
2240 task_scan_min(p
), task_scan_max(p
));
2241 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2245 * Get the fraction of time the task has been running since the last
2246 * NUMA placement cycle. The scheduler keeps similar statistics, but
2247 * decays those on a 32ms period, which is orders of magnitude off
2248 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2249 * stats only if the task is so new there are no NUMA statistics yet.
2251 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
2253 u64 runtime
, delta
, now
;
2254 /* Use the start of this time slice to avoid calculations. */
2255 now
= p
->se
.exec_start
;
2256 runtime
= p
->se
.sum_exec_runtime
;
2258 if (p
->last_task_numa_placement
) {
2259 delta
= runtime
- p
->last_sum_exec_runtime
;
2260 *period
= now
- p
->last_task_numa_placement
;
2262 /* Avoid time going backwards, prevent potential divide error: */
2263 if (unlikely((s64
)*period
< 0))
2266 delta
= p
->se
.avg
.load_sum
;
2267 *period
= LOAD_AVG_MAX
;
2270 p
->last_sum_exec_runtime
= runtime
;
2271 p
->last_task_numa_placement
= now
;
2277 * Determine the preferred nid for a task in a numa_group. This needs to
2278 * be done in a way that produces consistent results with group_weight,
2279 * otherwise workloads might not converge.
2281 static int preferred_group_nid(struct task_struct
*p
, int nid
)
2286 /* Direct connections between all NUMA nodes. */
2287 if (sched_numa_topology_type
== NUMA_DIRECT
)
2291 * On a system with glueless mesh NUMA topology, group_weight
2292 * scores nodes according to the number of NUMA hinting faults on
2293 * both the node itself, and on nearby nodes.
2295 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
2296 unsigned long score
, max_score
= 0;
2297 int node
, max_node
= nid
;
2299 dist
= sched_max_numa_distance
;
2301 for_each_online_node(node
) {
2302 score
= group_weight(p
, node
, dist
);
2303 if (score
> max_score
) {
2312 * Finding the preferred nid in a system with NUMA backplane
2313 * interconnect topology is more involved. The goal is to locate
2314 * tasks from numa_groups near each other in the system, and
2315 * untangle workloads from different sides of the system. This requires
2316 * searching down the hierarchy of node groups, recursively searching
2317 * inside the highest scoring group of nodes. The nodemask tricks
2318 * keep the complexity of the search down.
2320 nodes
= node_online_map
;
2321 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
2322 unsigned long max_faults
= 0;
2323 nodemask_t max_group
= NODE_MASK_NONE
;
2326 /* Are there nodes at this distance from each other? */
2327 if (!find_numa_distance(dist
))
2330 for_each_node_mask(a
, nodes
) {
2331 unsigned long faults
= 0;
2332 nodemask_t this_group
;
2333 nodes_clear(this_group
);
2335 /* Sum group's NUMA faults; includes a==b case. */
2336 for_each_node_mask(b
, nodes
) {
2337 if (node_distance(a
, b
) < dist
) {
2338 faults
+= group_faults(p
, b
);
2339 node_set(b
, this_group
);
2340 node_clear(b
, nodes
);
2344 /* Remember the top group. */
2345 if (faults
> max_faults
) {
2346 max_faults
= faults
;
2347 max_group
= this_group
;
2349 * subtle: at the smallest distance there is
2350 * just one node left in each "group", the
2351 * winner is the preferred nid.
2356 /* Next round, evaluate the nodes within max_group. */
2364 static void task_numa_placement(struct task_struct
*p
)
2366 int seq
, nid
, max_nid
= NUMA_NO_NODE
;
2367 unsigned long max_faults
= 0;
2368 unsigned long fault_types
[2] = { 0, 0 };
2369 unsigned long total_faults
;
2370 u64 runtime
, period
;
2371 spinlock_t
*group_lock
= NULL
;
2372 struct numa_group
*ng
;
2375 * The p->mm->numa_scan_seq field gets updated without
2376 * exclusive access. Use READ_ONCE() here to ensure
2377 * that the field is read in a single access:
2379 seq
= READ_ONCE(p
->mm
->numa_scan_seq
);
2380 if (p
->numa_scan_seq
== seq
)
2382 p
->numa_scan_seq
= seq
;
2383 p
->numa_scan_period_max
= task_scan_max(p
);
2385 total_faults
= p
->numa_faults_locality
[0] +
2386 p
->numa_faults_locality
[1];
2387 runtime
= numa_get_avg_runtime(p
, &period
);
2389 /* If the task is part of a group prevent parallel updates to group stats */
2390 ng
= deref_curr_numa_group(p
);
2392 group_lock
= &ng
->lock
;
2393 spin_lock_irq(group_lock
);
2396 /* Find the node with the highest number of faults */
2397 for_each_online_node(nid
) {
2398 /* Keep track of the offsets in numa_faults array */
2399 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
2400 unsigned long faults
= 0, group_faults
= 0;
2403 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
2404 long diff
, f_diff
, f_weight
;
2406 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
2407 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
2408 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
2409 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
2411 /* Decay existing window, copy faults since last scan */
2412 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
2413 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
2414 p
->numa_faults
[membuf_idx
] = 0;
2417 * Normalize the faults_from, so all tasks in a group
2418 * count according to CPU use, instead of by the raw
2419 * number of faults. Tasks with little runtime have
2420 * little over-all impact on throughput, and thus their
2421 * faults are less important.
2423 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
2424 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
2426 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
2427 p
->numa_faults
[cpubuf_idx
] = 0;
2429 p
->numa_faults
[mem_idx
] += diff
;
2430 p
->numa_faults
[cpu_idx
] += f_diff
;
2431 faults
+= p
->numa_faults
[mem_idx
];
2432 p
->total_numa_faults
+= diff
;
2435 * safe because we can only change our own group
2437 * mem_idx represents the offset for a given
2438 * nid and priv in a specific region because it
2439 * is at the beginning of the numa_faults array.
2441 ng
->faults
[mem_idx
] += diff
;
2442 ng
->faults_cpu
[mem_idx
] += f_diff
;
2443 ng
->total_faults
+= diff
;
2444 group_faults
+= ng
->faults
[mem_idx
];
2449 if (faults
> max_faults
) {
2450 max_faults
= faults
;
2453 } else if (group_faults
> max_faults
) {
2454 max_faults
= group_faults
;
2460 numa_group_count_active_nodes(ng
);
2461 spin_unlock_irq(group_lock
);
2462 max_nid
= preferred_group_nid(p
, max_nid
);
2466 /* Set the new preferred node */
2467 if (max_nid
!= p
->numa_preferred_nid
)
2468 sched_setnuma(p
, max_nid
);
2471 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
2474 static inline int get_numa_group(struct numa_group
*grp
)
2476 return refcount_inc_not_zero(&grp
->refcount
);
2479 static inline void put_numa_group(struct numa_group
*grp
)
2481 if (refcount_dec_and_test(&grp
->refcount
))
2482 kfree_rcu(grp
, rcu
);
2485 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
2488 struct numa_group
*grp
, *my_grp
;
2489 struct task_struct
*tsk
;
2491 int cpu
= cpupid_to_cpu(cpupid
);
2494 if (unlikely(!deref_curr_numa_group(p
))) {
2495 unsigned int size
= sizeof(struct numa_group
) +
2496 4*nr_node_ids
*sizeof(unsigned long);
2498 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
2502 refcount_set(&grp
->refcount
, 1);
2503 grp
->active_nodes
= 1;
2504 grp
->max_faults_cpu
= 0;
2505 spin_lock_init(&grp
->lock
);
2507 /* Second half of the array tracks nids where faults happen */
2508 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
2511 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2512 grp
->faults
[i
] = p
->numa_faults
[i
];
2514 grp
->total_faults
= p
->total_numa_faults
;
2517 rcu_assign_pointer(p
->numa_group
, grp
);
2521 tsk
= READ_ONCE(cpu_rq(cpu
)->curr
);
2523 if (!cpupid_match_pid(tsk
, cpupid
))
2526 grp
= rcu_dereference(tsk
->numa_group
);
2530 my_grp
= deref_curr_numa_group(p
);
2535 * Only join the other group if its bigger; if we're the bigger group,
2536 * the other task will join us.
2538 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
2542 * Tie-break on the grp address.
2544 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
2547 /* Always join threads in the same process. */
2548 if (tsk
->mm
== current
->mm
)
2551 /* Simple filter to avoid false positives due to PID collisions */
2552 if (flags
& TNF_SHARED
)
2555 /* Update priv based on whether false sharing was detected */
2558 if (join
&& !get_numa_group(grp
))
2566 BUG_ON(irqs_disabled());
2567 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
2569 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
2570 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
2571 grp
->faults
[i
] += p
->numa_faults
[i
];
2573 my_grp
->total_faults
-= p
->total_numa_faults
;
2574 grp
->total_faults
+= p
->total_numa_faults
;
2579 spin_unlock(&my_grp
->lock
);
2580 spin_unlock_irq(&grp
->lock
);
2582 rcu_assign_pointer(p
->numa_group
, grp
);
2584 put_numa_group(my_grp
);
2593 * Get rid of NUMA staticstics associated with a task (either current or dead).
2594 * If @final is set, the task is dead and has reached refcount zero, so we can
2595 * safely free all relevant data structures. Otherwise, there might be
2596 * concurrent reads from places like load balancing and procfs, and we should
2597 * reset the data back to default state without freeing ->numa_faults.
2599 void task_numa_free(struct task_struct
*p
, bool final
)
2601 /* safe: p either is current or is being freed by current */
2602 struct numa_group
*grp
= rcu_dereference_raw(p
->numa_group
);
2603 unsigned long *numa_faults
= p
->numa_faults
;
2604 unsigned long flags
;
2611 spin_lock_irqsave(&grp
->lock
, flags
);
2612 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2613 grp
->faults
[i
] -= p
->numa_faults
[i
];
2614 grp
->total_faults
-= p
->total_numa_faults
;
2617 spin_unlock_irqrestore(&grp
->lock
, flags
);
2618 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2619 put_numa_group(grp
);
2623 p
->numa_faults
= NULL
;
2626 p
->total_numa_faults
= 0;
2627 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2633 * Got a PROT_NONE fault for a page on @node.
2635 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2637 struct task_struct
*p
= current
;
2638 bool migrated
= flags
& TNF_MIGRATED
;
2639 int cpu_node
= task_node(current
);
2640 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2641 struct numa_group
*ng
;
2644 if (!static_branch_likely(&sched_numa_balancing
))
2647 /* for example, ksmd faulting in a user's mm */
2651 /* Allocate buffer to track faults on a per-node basis */
2652 if (unlikely(!p
->numa_faults
)) {
2653 int size
= sizeof(*p
->numa_faults
) *
2654 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2656 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2657 if (!p
->numa_faults
)
2660 p
->total_numa_faults
= 0;
2661 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2665 * First accesses are treated as private, otherwise consider accesses
2666 * to be private if the accessing pid has not changed
2668 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2671 priv
= cpupid_match_pid(p
, last_cpupid
);
2672 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2673 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2677 * If a workload spans multiple NUMA nodes, a shared fault that
2678 * occurs wholly within the set of nodes that the workload is
2679 * actively using should be counted as local. This allows the
2680 * scan rate to slow down when a workload has settled down.
2682 ng
= deref_curr_numa_group(p
);
2683 if (!priv
&& !local
&& ng
&& ng
->active_nodes
> 1 &&
2684 numa_is_active_node(cpu_node
, ng
) &&
2685 numa_is_active_node(mem_node
, ng
))
2689 * Retry to migrate task to preferred node periodically, in case it
2690 * previously failed, or the scheduler moved us.
2692 if (time_after(jiffies
, p
->numa_migrate_retry
)) {
2693 task_numa_placement(p
);
2694 numa_migrate_preferred(p
);
2698 p
->numa_pages_migrated
+= pages
;
2699 if (flags
& TNF_MIGRATE_FAIL
)
2700 p
->numa_faults_locality
[2] += pages
;
2702 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2703 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2704 p
->numa_faults_locality
[local
] += pages
;
2707 static void reset_ptenuma_scan(struct task_struct
*p
)
2710 * We only did a read acquisition of the mmap sem, so
2711 * p->mm->numa_scan_seq is written to without exclusive access
2712 * and the update is not guaranteed to be atomic. That's not
2713 * much of an issue though, since this is just used for
2714 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2715 * expensive, to avoid any form of compiler optimizations:
2717 WRITE_ONCE(p
->mm
->numa_scan_seq
, READ_ONCE(p
->mm
->numa_scan_seq
) + 1);
2718 p
->mm
->numa_scan_offset
= 0;
2722 * The expensive part of numa migration is done from task_work context.
2723 * Triggered from task_tick_numa().
2725 static void task_numa_work(struct callback_head
*work
)
2727 unsigned long migrate
, next_scan
, now
= jiffies
;
2728 struct task_struct
*p
= current
;
2729 struct mm_struct
*mm
= p
->mm
;
2730 u64 runtime
= p
->se
.sum_exec_runtime
;
2731 struct vm_area_struct
*vma
;
2732 unsigned long start
, end
;
2733 unsigned long nr_pte_updates
= 0;
2734 long pages
, virtpages
;
2736 SCHED_WARN_ON(p
!= container_of(work
, struct task_struct
, numa_work
));
2740 * Who cares about NUMA placement when they're dying.
2742 * NOTE: make sure not to dereference p->mm before this check,
2743 * exit_task_work() happens _after_ exit_mm() so we could be called
2744 * without p->mm even though we still had it when we enqueued this
2747 if (p
->flags
& PF_EXITING
)
2750 if (!mm
->numa_next_scan
) {
2751 mm
->numa_next_scan
= now
+
2752 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2756 * Enforce maximal scan/migration frequency..
2758 migrate
= mm
->numa_next_scan
;
2759 if (time_before(now
, migrate
))
2762 if (p
->numa_scan_period
== 0) {
2763 p
->numa_scan_period_max
= task_scan_max(p
);
2764 p
->numa_scan_period
= task_scan_start(p
);
2767 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2768 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2772 * Delay this task enough that another task of this mm will likely win
2773 * the next time around.
2775 p
->node_stamp
+= 2 * TICK_NSEC
;
2777 start
= mm
->numa_scan_offset
;
2778 pages
= sysctl_numa_balancing_scan_size
;
2779 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2780 virtpages
= pages
* 8; /* Scan up to this much virtual space */
2785 if (!mmap_read_trylock(mm
))
2787 vma
= find_vma(mm
, start
);
2789 reset_ptenuma_scan(p
);
2793 for (; vma
; vma
= vma
->vm_next
) {
2794 if (!vma_migratable(vma
) || !vma_policy_mof(vma
) ||
2795 is_vm_hugetlb_page(vma
) || (vma
->vm_flags
& VM_MIXEDMAP
)) {
2800 * Shared library pages mapped by multiple processes are not
2801 * migrated as it is expected they are cache replicated. Avoid
2802 * hinting faults in read-only file-backed mappings or the vdso
2803 * as migrating the pages will be of marginal benefit.
2806 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2810 * Skip inaccessible VMAs to avoid any confusion between
2811 * PROT_NONE and NUMA hinting ptes
2813 if (!vma_is_accessible(vma
))
2817 start
= max(start
, vma
->vm_start
);
2818 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2819 end
= min(end
, vma
->vm_end
);
2820 nr_pte_updates
= change_prot_numa(vma
, start
, end
);
2823 * Try to scan sysctl_numa_balancing_size worth of
2824 * hpages that have at least one present PTE that
2825 * is not already pte-numa. If the VMA contains
2826 * areas that are unused or already full of prot_numa
2827 * PTEs, scan up to virtpages, to skip through those
2831 pages
-= (end
- start
) >> PAGE_SHIFT
;
2832 virtpages
-= (end
- start
) >> PAGE_SHIFT
;
2835 if (pages
<= 0 || virtpages
<= 0)
2839 } while (end
!= vma
->vm_end
);
2844 * It is possible to reach the end of the VMA list but the last few
2845 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2846 * would find the !migratable VMA on the next scan but not reset the
2847 * scanner to the start so check it now.
2850 mm
->numa_scan_offset
= start
;
2852 reset_ptenuma_scan(p
);
2853 mmap_read_unlock(mm
);
2856 * Make sure tasks use at least 32x as much time to run other code
2857 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2858 * Usually update_task_scan_period slows down scanning enough; on an
2859 * overloaded system we need to limit overhead on a per task basis.
2861 if (unlikely(p
->se
.sum_exec_runtime
!= runtime
)) {
2862 u64 diff
= p
->se
.sum_exec_runtime
- runtime
;
2863 p
->node_stamp
+= 32 * diff
;
2867 void init_numa_balancing(unsigned long clone_flags
, struct task_struct
*p
)
2870 struct mm_struct
*mm
= p
->mm
;
2873 mm_users
= atomic_read(&mm
->mm_users
);
2874 if (mm_users
== 1) {
2875 mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2876 mm
->numa_scan_seq
= 0;
2880 p
->numa_scan_seq
= mm
? mm
->numa_scan_seq
: 0;
2881 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2882 /* Protect against double add, see task_tick_numa and task_numa_work */
2883 p
->numa_work
.next
= &p
->numa_work
;
2884 p
->numa_faults
= NULL
;
2885 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2886 p
->last_task_numa_placement
= 0;
2887 p
->last_sum_exec_runtime
= 0;
2889 init_task_work(&p
->numa_work
, task_numa_work
);
2891 /* New address space, reset the preferred nid */
2892 if (!(clone_flags
& CLONE_VM
)) {
2893 p
->numa_preferred_nid
= NUMA_NO_NODE
;
2898 * New thread, keep existing numa_preferred_nid which should be copied
2899 * already by arch_dup_task_struct but stagger when scans start.
2904 delay
= min_t(unsigned int, task_scan_max(current
),
2905 current
->numa_scan_period
* mm_users
* NSEC_PER_MSEC
);
2906 delay
+= 2 * TICK_NSEC
;
2907 p
->node_stamp
= delay
;
2912 * Drive the periodic memory faults..
2914 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2916 struct callback_head
*work
= &curr
->numa_work
;
2920 * We don't care about NUMA placement if we don't have memory.
2922 if ((curr
->flags
& (PF_EXITING
| PF_KTHREAD
)) || work
->next
!= work
)
2926 * Using runtime rather than walltime has the dual advantage that
2927 * we (mostly) drive the selection from busy threads and that the
2928 * task needs to have done some actual work before we bother with
2931 now
= curr
->se
.sum_exec_runtime
;
2932 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2934 if (now
> curr
->node_stamp
+ period
) {
2935 if (!curr
->node_stamp
)
2936 curr
->numa_scan_period
= task_scan_start(curr
);
2937 curr
->node_stamp
+= period
;
2939 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
))
2940 task_work_add(curr
, work
, TWA_RESUME
);
2944 static void update_scan_period(struct task_struct
*p
, int new_cpu
)
2946 int src_nid
= cpu_to_node(task_cpu(p
));
2947 int dst_nid
= cpu_to_node(new_cpu
);
2949 if (!static_branch_likely(&sched_numa_balancing
))
2952 if (!p
->mm
|| !p
->numa_faults
|| (p
->flags
& PF_EXITING
))
2955 if (src_nid
== dst_nid
)
2959 * Allow resets if faults have been trapped before one scan
2960 * has completed. This is most likely due to a new task that
2961 * is pulled cross-node due to wakeups or load balancing.
2963 if (p
->numa_scan_seq
) {
2965 * Avoid scan adjustments if moving to the preferred
2966 * node or if the task was not previously running on
2967 * the preferred node.
2969 if (dst_nid
== p
->numa_preferred_nid
||
2970 (p
->numa_preferred_nid
!= NUMA_NO_NODE
&&
2971 src_nid
!= p
->numa_preferred_nid
))
2975 p
->numa_scan_period
= task_scan_start(p
);
2979 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2983 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2987 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2991 static inline void update_scan_period(struct task_struct
*p
, int new_cpu
)
2995 #endif /* CONFIG_NUMA_BALANCING */
2998 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3000 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
3002 if (entity_is_task(se
)) {
3003 struct rq
*rq
= rq_of(cfs_rq
);
3005 account_numa_enqueue(rq
, task_of(se
));
3006 list_add(&se
->group_node
, &rq
->cfs_tasks
);
3009 cfs_rq
->nr_running
++;
3013 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3015 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
3017 if (entity_is_task(se
)) {
3018 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
3019 list_del_init(&se
->group_node
);
3022 cfs_rq
->nr_running
--;
3026 * Signed add and clamp on underflow.
3028 * Explicitly do a load-store to ensure the intermediate value never hits
3029 * memory. This allows lockless observations without ever seeing the negative
3032 #define add_positive(_ptr, _val) do { \
3033 typeof(_ptr) ptr = (_ptr); \
3034 typeof(_val) val = (_val); \
3035 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3039 if (val < 0 && res > var) \
3042 WRITE_ONCE(*ptr, res); \
3046 * Unsigned subtract and clamp on underflow.
3048 * Explicitly do a load-store to ensure the intermediate value never hits
3049 * memory. This allows lockless observations without ever seeing the negative
3052 #define sub_positive(_ptr, _val) do { \
3053 typeof(_ptr) ptr = (_ptr); \
3054 typeof(*ptr) val = (_val); \
3055 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3059 WRITE_ONCE(*ptr, res); \
3063 * Remove and clamp on negative, from a local variable.
3065 * A variant of sub_positive(), which does not use explicit load-store
3066 * and is thus optimized for local variable updates.
3068 #define lsub_positive(_ptr, _val) do { \
3069 typeof(_ptr) ptr = (_ptr); \
3070 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3075 enqueue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3077 cfs_rq
->avg
.load_avg
+= se
->avg
.load_avg
;
3078 cfs_rq
->avg
.load_sum
+= se_weight(se
) * se
->avg
.load_sum
;
3082 dequeue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3084 sub_positive(&cfs_rq
->avg
.load_avg
, se
->avg
.load_avg
);
3085 sub_positive(&cfs_rq
->avg
.load_sum
, se_weight(se
) * se
->avg
.load_sum
);
3089 enqueue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
3091 dequeue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
3094 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
3095 unsigned long weight
)
3098 /* commit outstanding execution time */
3099 if (cfs_rq
->curr
== se
)
3100 update_curr(cfs_rq
);
3101 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
3103 dequeue_load_avg(cfs_rq
, se
);
3105 update_load_set(&se
->load
, weight
);
3109 u32 divider
= get_pelt_divider(&se
->avg
);
3111 se
->avg
.load_avg
= div_u64(se_weight(se
) * se
->avg
.load_sum
, divider
);
3115 enqueue_load_avg(cfs_rq
, se
);
3117 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
3121 void reweight_task(struct task_struct
*p
, int prio
)
3123 struct sched_entity
*se
= &p
->se
;
3124 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3125 struct load_weight
*load
= &se
->load
;
3126 unsigned long weight
= scale_load(sched_prio_to_weight
[prio
]);
3128 reweight_entity(cfs_rq
, se
, weight
);
3129 load
->inv_weight
= sched_prio_to_wmult
[prio
];
3132 #ifdef CONFIG_FAIR_GROUP_SCHED
3135 * All this does is approximate the hierarchical proportion which includes that
3136 * global sum we all love to hate.
3138 * That is, the weight of a group entity, is the proportional share of the
3139 * group weight based on the group runqueue weights. That is:
3141 * tg->weight * grq->load.weight
3142 * ge->load.weight = ----------------------------- (1)
3143 * \Sum grq->load.weight
3145 * Now, because computing that sum is prohibitively expensive to compute (been
3146 * there, done that) we approximate it with this average stuff. The average
3147 * moves slower and therefore the approximation is cheaper and more stable.
3149 * So instead of the above, we substitute:
3151 * grq->load.weight -> grq->avg.load_avg (2)
3153 * which yields the following:
3155 * tg->weight * grq->avg.load_avg
3156 * ge->load.weight = ------------------------------ (3)
3159 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3161 * That is shares_avg, and it is right (given the approximation (2)).
3163 * The problem with it is that because the average is slow -- it was designed
3164 * to be exactly that of course -- this leads to transients in boundary
3165 * conditions. In specific, the case where the group was idle and we start the
3166 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3167 * yielding bad latency etc..
3169 * Now, in that special case (1) reduces to:
3171 * tg->weight * grq->load.weight
3172 * ge->load.weight = ----------------------------- = tg->weight (4)
3175 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3177 * So what we do is modify our approximation (3) to approach (4) in the (near)
3182 * tg->weight * grq->load.weight
3183 * --------------------------------------------------- (5)
3184 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3186 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3187 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3190 * tg->weight * grq->load.weight
3191 * ge->load.weight = ----------------------------- (6)
3196 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3197 * max(grq->load.weight, grq->avg.load_avg)
3199 * And that is shares_weight and is icky. In the (near) UP case it approaches
3200 * (4) while in the normal case it approaches (3). It consistently
3201 * overestimates the ge->load.weight and therefore:
3203 * \Sum ge->load.weight >= tg->weight
3207 static long calc_group_shares(struct cfs_rq
*cfs_rq
)
3209 long tg_weight
, tg_shares
, load
, shares
;
3210 struct task_group
*tg
= cfs_rq
->tg
;
3212 tg_shares
= READ_ONCE(tg
->shares
);
3214 load
= max(scale_load_down(cfs_rq
->load
.weight
), cfs_rq
->avg
.load_avg
);
3216 tg_weight
= atomic_long_read(&tg
->load_avg
);
3218 /* Ensure tg_weight >= load */
3219 tg_weight
-= cfs_rq
->tg_load_avg_contrib
;
3222 shares
= (tg_shares
* load
);
3224 shares
/= tg_weight
;
3227 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3228 * of a group with small tg->shares value. It is a floor value which is
3229 * assigned as a minimum load.weight to the sched_entity representing
3230 * the group on a CPU.
3232 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3233 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3234 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3235 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3238 return clamp_t(long, shares
, MIN_SHARES
, tg_shares
);
3240 #endif /* CONFIG_SMP */
3242 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
3245 * Recomputes the group entity based on the current state of its group
3248 static void update_cfs_group(struct sched_entity
*se
)
3250 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3256 if (throttled_hierarchy(gcfs_rq
))
3260 shares
= READ_ONCE(gcfs_rq
->tg
->shares
);
3262 if (likely(se
->load
.weight
== shares
))
3265 shares
= calc_group_shares(gcfs_rq
);
3268 reweight_entity(cfs_rq_of(se
), se
, shares
);
3271 #else /* CONFIG_FAIR_GROUP_SCHED */
3272 static inline void update_cfs_group(struct sched_entity
*se
)
3275 #endif /* CONFIG_FAIR_GROUP_SCHED */
3277 static inline void cfs_rq_util_change(struct cfs_rq
*cfs_rq
, int flags
)
3279 struct rq
*rq
= rq_of(cfs_rq
);
3281 if (&rq
->cfs
== cfs_rq
) {
3283 * There are a few boundary cases this might miss but it should
3284 * get called often enough that that should (hopefully) not be
3287 * It will not get called when we go idle, because the idle
3288 * thread is a different class (!fair), nor will the utilization
3289 * number include things like RT tasks.
3291 * As is, the util number is not freq-invariant (we'd have to
3292 * implement arch_scale_freq_capacity() for that).
3296 cpufreq_update_util(rq
, flags
);
3301 #ifdef CONFIG_FAIR_GROUP_SCHED
3303 * update_tg_load_avg - update the tg's load avg
3304 * @cfs_rq: the cfs_rq whose avg changed
3306 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3307 * However, because tg->load_avg is a global value there are performance
3310 * In order to avoid having to look at the other cfs_rq's, we use a
3311 * differential update where we store the last value we propagated. This in
3312 * turn allows skipping updates if the differential is 'small'.
3314 * Updating tg's load_avg is necessary before update_cfs_share().
3316 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
)
3318 long delta
= cfs_rq
->avg
.load_avg
- cfs_rq
->tg_load_avg_contrib
;
3321 * No need to update load_avg for root_task_group as it is not used.
3323 if (cfs_rq
->tg
== &root_task_group
)
3326 if (abs(delta
) > cfs_rq
->tg_load_avg_contrib
/ 64) {
3327 atomic_long_add(delta
, &cfs_rq
->tg
->load_avg
);
3328 cfs_rq
->tg_load_avg_contrib
= cfs_rq
->avg
.load_avg
;
3333 * Called within set_task_rq() right before setting a task's CPU. The
3334 * caller only guarantees p->pi_lock is held; no other assumptions,
3335 * including the state of rq->lock, should be made.
3337 void set_task_rq_fair(struct sched_entity
*se
,
3338 struct cfs_rq
*prev
, struct cfs_rq
*next
)
3340 u64 p_last_update_time
;
3341 u64 n_last_update_time
;
3343 if (!sched_feat(ATTACH_AGE_LOAD
))
3347 * We are supposed to update the task to "current" time, then its up to
3348 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3349 * getting what current time is, so simply throw away the out-of-date
3350 * time. This will result in the wakee task is less decayed, but giving
3351 * the wakee more load sounds not bad.
3353 if (!(se
->avg
.last_update_time
&& prev
))
3356 #ifndef CONFIG_64BIT
3358 u64 p_last_update_time_copy
;
3359 u64 n_last_update_time_copy
;
3362 p_last_update_time_copy
= prev
->load_last_update_time_copy
;
3363 n_last_update_time_copy
= next
->load_last_update_time_copy
;
3367 p_last_update_time
= prev
->avg
.last_update_time
;
3368 n_last_update_time
= next
->avg
.last_update_time
;
3370 } while (p_last_update_time
!= p_last_update_time_copy
||
3371 n_last_update_time
!= n_last_update_time_copy
);
3374 p_last_update_time
= prev
->avg
.last_update_time
;
3375 n_last_update_time
= next
->avg
.last_update_time
;
3377 __update_load_avg_blocked_se(p_last_update_time
, se
);
3378 se
->avg
.last_update_time
= n_last_update_time
;
3383 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3384 * propagate its contribution. The key to this propagation is the invariant
3385 * that for each group:
3387 * ge->avg == grq->avg (1)
3389 * _IFF_ we look at the pure running and runnable sums. Because they
3390 * represent the very same entity, just at different points in the hierarchy.
3392 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3393 * and simply copies the running/runnable sum over (but still wrong, because
3394 * the group entity and group rq do not have their PELT windows aligned).
3396 * However, update_tg_cfs_load() is more complex. So we have:
3398 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3400 * And since, like util, the runnable part should be directly transferable,
3401 * the following would _appear_ to be the straight forward approach:
3403 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3405 * And per (1) we have:
3407 * ge->avg.runnable_avg == grq->avg.runnable_avg
3411 * ge->load.weight * grq->avg.load_avg
3412 * ge->avg.load_avg = ----------------------------------- (4)
3415 * Except that is wrong!
3417 * Because while for entities historical weight is not important and we
3418 * really only care about our future and therefore can consider a pure
3419 * runnable sum, runqueues can NOT do this.
3421 * We specifically want runqueues to have a load_avg that includes
3422 * historical weights. Those represent the blocked load, the load we expect
3423 * to (shortly) return to us. This only works by keeping the weights as
3424 * integral part of the sum. We therefore cannot decompose as per (3).
3426 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3427 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3428 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3429 * runnable section of these tasks overlap (or not). If they were to perfectly
3430 * align the rq as a whole would be runnable 2/3 of the time. If however we
3431 * always have at least 1 runnable task, the rq as a whole is always runnable.
3433 * So we'll have to approximate.. :/
3435 * Given the constraint:
3437 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3439 * We can construct a rule that adds runnable to a rq by assuming minimal
3442 * On removal, we'll assume each task is equally runnable; which yields:
3444 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3446 * XXX: only do this for the part of runnable > running ?
3451 update_tg_cfs_util(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3453 long delta
= gcfs_rq
->avg
.util_avg
- se
->avg
.util_avg
;
3456 /* Nothing to update */
3461 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3462 * See ___update_load_avg() for details.
3464 divider
= get_pelt_divider(&cfs_rq
->avg
);
3466 /* Set new sched_entity's utilization */
3467 se
->avg
.util_avg
= gcfs_rq
->avg
.util_avg
;
3468 se
->avg
.util_sum
= se
->avg
.util_avg
* divider
;
3470 /* Update parent cfs_rq utilization */
3471 add_positive(&cfs_rq
->avg
.util_avg
, delta
);
3472 cfs_rq
->avg
.util_sum
= cfs_rq
->avg
.util_avg
* divider
;
3476 update_tg_cfs_runnable(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3478 long delta
= gcfs_rq
->avg
.runnable_avg
- se
->avg
.runnable_avg
;
3481 /* Nothing to update */
3486 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3487 * See ___update_load_avg() for details.
3489 divider
= get_pelt_divider(&cfs_rq
->avg
);
3491 /* Set new sched_entity's runnable */
3492 se
->avg
.runnable_avg
= gcfs_rq
->avg
.runnable_avg
;
3493 se
->avg
.runnable_sum
= se
->avg
.runnable_avg
* divider
;
3495 /* Update parent cfs_rq runnable */
3496 add_positive(&cfs_rq
->avg
.runnable_avg
, delta
);
3497 cfs_rq
->avg
.runnable_sum
= cfs_rq
->avg
.runnable_avg
* divider
;
3501 update_tg_cfs_load(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3503 long delta_avg
, running_sum
, runnable_sum
= gcfs_rq
->prop_runnable_sum
;
3504 unsigned long load_avg
;
3512 gcfs_rq
->prop_runnable_sum
= 0;
3515 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3516 * See ___update_load_avg() for details.
3518 divider
= get_pelt_divider(&cfs_rq
->avg
);
3520 if (runnable_sum
>= 0) {
3522 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3523 * the CPU is saturated running == runnable.
3525 runnable_sum
+= se
->avg
.load_sum
;
3526 runnable_sum
= min_t(long, runnable_sum
, divider
);
3529 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3530 * assuming all tasks are equally runnable.
3532 if (scale_load_down(gcfs_rq
->load
.weight
)) {
3533 load_sum
= div_s64(gcfs_rq
->avg
.load_sum
,
3534 scale_load_down(gcfs_rq
->load
.weight
));
3537 /* But make sure to not inflate se's runnable */
3538 runnable_sum
= min(se
->avg
.load_sum
, load_sum
);
3542 * runnable_sum can't be lower than running_sum
3543 * Rescale running sum to be in the same range as runnable sum
3544 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3545 * runnable_sum is in [0 : LOAD_AVG_MAX]
3547 running_sum
= se
->avg
.util_sum
>> SCHED_CAPACITY_SHIFT
;
3548 runnable_sum
= max(runnable_sum
, running_sum
);
3550 load_sum
= (s64
)se_weight(se
) * runnable_sum
;
3551 load_avg
= div_s64(load_sum
, divider
);
3553 delta_sum
= load_sum
- (s64
)se_weight(se
) * se
->avg
.load_sum
;
3554 delta_avg
= load_avg
- se
->avg
.load_avg
;
3556 se
->avg
.load_sum
= runnable_sum
;
3557 se
->avg
.load_avg
= load_avg
;
3558 add_positive(&cfs_rq
->avg
.load_avg
, delta_avg
);
3559 add_positive(&cfs_rq
->avg
.load_sum
, delta_sum
);
3562 static inline void add_tg_cfs_propagate(struct cfs_rq
*cfs_rq
, long runnable_sum
)
3564 cfs_rq
->propagate
= 1;
3565 cfs_rq
->prop_runnable_sum
+= runnable_sum
;
3568 /* Update task and its cfs_rq load average */
3569 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3571 struct cfs_rq
*cfs_rq
, *gcfs_rq
;
3573 if (entity_is_task(se
))
3576 gcfs_rq
= group_cfs_rq(se
);
3577 if (!gcfs_rq
->propagate
)
3580 gcfs_rq
->propagate
= 0;
3582 cfs_rq
= cfs_rq_of(se
);
3584 add_tg_cfs_propagate(cfs_rq
, gcfs_rq
->prop_runnable_sum
);
3586 update_tg_cfs_util(cfs_rq
, se
, gcfs_rq
);
3587 update_tg_cfs_runnable(cfs_rq
, se
, gcfs_rq
);
3588 update_tg_cfs_load(cfs_rq
, se
, gcfs_rq
);
3590 trace_pelt_cfs_tp(cfs_rq
);
3591 trace_pelt_se_tp(se
);
3597 * Check if we need to update the load and the utilization of a blocked
3600 static inline bool skip_blocked_update(struct sched_entity
*se
)
3602 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3605 * If sched_entity still have not zero load or utilization, we have to
3608 if (se
->avg
.load_avg
|| se
->avg
.util_avg
)
3612 * If there is a pending propagation, we have to update the load and
3613 * the utilization of the sched_entity:
3615 if (gcfs_rq
->propagate
)
3619 * Otherwise, the load and the utilization of the sched_entity is
3620 * already zero and there is no pending propagation, so it will be a
3621 * waste of time to try to decay it:
3626 #else /* CONFIG_FAIR_GROUP_SCHED */
3628 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
) {}
3630 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3635 static inline void add_tg_cfs_propagate(struct cfs_rq
*cfs_rq
, long runnable_sum
) {}
3637 #endif /* CONFIG_FAIR_GROUP_SCHED */
3640 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3641 * @now: current time, as per cfs_rq_clock_pelt()
3642 * @cfs_rq: cfs_rq to update
3644 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3645 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3646 * post_init_entity_util_avg().
3648 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3650 * Returns true if the load decayed or we removed load.
3652 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3653 * call update_tg_load_avg() when this function returns true.
3656 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
)
3658 unsigned long removed_load
= 0, removed_util
= 0, removed_runnable
= 0;
3659 struct sched_avg
*sa
= &cfs_rq
->avg
;
3662 if (cfs_rq
->removed
.nr
) {
3664 u32 divider
= get_pelt_divider(&cfs_rq
->avg
);
3666 raw_spin_lock(&cfs_rq
->removed
.lock
);
3667 swap(cfs_rq
->removed
.util_avg
, removed_util
);
3668 swap(cfs_rq
->removed
.load_avg
, removed_load
);
3669 swap(cfs_rq
->removed
.runnable_avg
, removed_runnable
);
3670 cfs_rq
->removed
.nr
= 0;
3671 raw_spin_unlock(&cfs_rq
->removed
.lock
);
3674 sub_positive(&sa
->load_avg
, r
);
3675 sub_positive(&sa
->load_sum
, r
* divider
);
3678 sub_positive(&sa
->util_avg
, r
);
3679 sub_positive(&sa
->util_sum
, r
* divider
);
3681 r
= removed_runnable
;
3682 sub_positive(&sa
->runnable_avg
, r
);
3683 sub_positive(&sa
->runnable_sum
, r
* divider
);
3686 * removed_runnable is the unweighted version of removed_load so we
3687 * can use it to estimate removed_load_sum.
3689 add_tg_cfs_propagate(cfs_rq
,
3690 -(long)(removed_runnable
* divider
) >> SCHED_CAPACITY_SHIFT
);
3695 decayed
|= __update_load_avg_cfs_rq(now
, cfs_rq
);
3697 #ifndef CONFIG_64BIT
3699 cfs_rq
->load_last_update_time_copy
= sa
->last_update_time
;
3706 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3707 * @cfs_rq: cfs_rq to attach to
3708 * @se: sched_entity to attach
3710 * Must call update_cfs_rq_load_avg() before this, since we rely on
3711 * cfs_rq->avg.last_update_time being current.
3713 static void attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3716 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3717 * See ___update_load_avg() for details.
3719 u32 divider
= get_pelt_divider(&cfs_rq
->avg
);
3722 * When we attach the @se to the @cfs_rq, we must align the decay
3723 * window because without that, really weird and wonderful things can
3728 se
->avg
.last_update_time
= cfs_rq
->avg
.last_update_time
;
3729 se
->avg
.period_contrib
= cfs_rq
->avg
.period_contrib
;
3732 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3733 * period_contrib. This isn't strictly correct, but since we're
3734 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3737 se
->avg
.util_sum
= se
->avg
.util_avg
* divider
;
3739 se
->avg
.runnable_sum
= se
->avg
.runnable_avg
* divider
;
3741 se
->avg
.load_sum
= divider
;
3742 if (se_weight(se
)) {
3744 div_u64(se
->avg
.load_avg
* se
->avg
.load_sum
, se_weight(se
));
3747 enqueue_load_avg(cfs_rq
, se
);
3748 cfs_rq
->avg
.util_avg
+= se
->avg
.util_avg
;
3749 cfs_rq
->avg
.util_sum
+= se
->avg
.util_sum
;
3750 cfs_rq
->avg
.runnable_avg
+= se
->avg
.runnable_avg
;
3751 cfs_rq
->avg
.runnable_sum
+= se
->avg
.runnable_sum
;
3753 add_tg_cfs_propagate(cfs_rq
, se
->avg
.load_sum
);
3755 cfs_rq_util_change(cfs_rq
, 0);
3757 trace_pelt_cfs_tp(cfs_rq
);
3761 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3762 * @cfs_rq: cfs_rq to detach from
3763 * @se: sched_entity to detach
3765 * Must call update_cfs_rq_load_avg() before this, since we rely on
3766 * cfs_rq->avg.last_update_time being current.
3768 static void detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3770 dequeue_load_avg(cfs_rq
, se
);
3771 sub_positive(&cfs_rq
->avg
.util_avg
, se
->avg
.util_avg
);
3772 sub_positive(&cfs_rq
->avg
.util_sum
, se
->avg
.util_sum
);
3773 sub_positive(&cfs_rq
->avg
.runnable_avg
, se
->avg
.runnable_avg
);
3774 sub_positive(&cfs_rq
->avg
.runnable_sum
, se
->avg
.runnable_sum
);
3776 add_tg_cfs_propagate(cfs_rq
, -se
->avg
.load_sum
);
3778 cfs_rq_util_change(cfs_rq
, 0);
3780 trace_pelt_cfs_tp(cfs_rq
);
3784 * Optional action to be done while updating the load average
3786 #define UPDATE_TG 0x1
3787 #define SKIP_AGE_LOAD 0x2
3788 #define DO_ATTACH 0x4
3790 /* Update task and its cfs_rq load average */
3791 static inline void update_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3793 u64 now
= cfs_rq_clock_pelt(cfs_rq
);
3797 * Track task load average for carrying it to new CPU after migrated, and
3798 * track group sched_entity load average for task_h_load calc in migration
3800 if (se
->avg
.last_update_time
&& !(flags
& SKIP_AGE_LOAD
))
3801 __update_load_avg_se(now
, cfs_rq
, se
);
3803 decayed
= update_cfs_rq_load_avg(now
, cfs_rq
);
3804 decayed
|= propagate_entity_load_avg(se
);
3806 if (!se
->avg
.last_update_time
&& (flags
& DO_ATTACH
)) {
3809 * DO_ATTACH means we're here from enqueue_entity().
3810 * !last_update_time means we've passed through
3811 * migrate_task_rq_fair() indicating we migrated.
3813 * IOW we're enqueueing a task on a new CPU.
3815 attach_entity_load_avg(cfs_rq
, se
);
3816 update_tg_load_avg(cfs_rq
);
3818 } else if (decayed
) {
3819 cfs_rq_util_change(cfs_rq
, 0);
3821 if (flags
& UPDATE_TG
)
3822 update_tg_load_avg(cfs_rq
);
3826 #ifndef CONFIG_64BIT
3827 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3829 u64 last_update_time_copy
;
3830 u64 last_update_time
;
3833 last_update_time_copy
= cfs_rq
->load_last_update_time_copy
;
3835 last_update_time
= cfs_rq
->avg
.last_update_time
;
3836 } while (last_update_time
!= last_update_time_copy
);
3838 return last_update_time
;
3841 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3843 return cfs_rq
->avg
.last_update_time
;
3848 * Synchronize entity load avg of dequeued entity without locking
3851 static void sync_entity_load_avg(struct sched_entity
*se
)
3853 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3854 u64 last_update_time
;
3856 last_update_time
= cfs_rq_last_update_time(cfs_rq
);
3857 __update_load_avg_blocked_se(last_update_time
, se
);
3861 * Task first catches up with cfs_rq, and then subtract
3862 * itself from the cfs_rq (task must be off the queue now).
3864 static void remove_entity_load_avg(struct sched_entity
*se
)
3866 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3867 unsigned long flags
;
3870 * tasks cannot exit without having gone through wake_up_new_task() ->
3871 * post_init_entity_util_avg() which will have added things to the
3872 * cfs_rq, so we can remove unconditionally.
3875 sync_entity_load_avg(se
);
3877 raw_spin_lock_irqsave(&cfs_rq
->removed
.lock
, flags
);
3878 ++cfs_rq
->removed
.nr
;
3879 cfs_rq
->removed
.util_avg
+= se
->avg
.util_avg
;
3880 cfs_rq
->removed
.load_avg
+= se
->avg
.load_avg
;
3881 cfs_rq
->removed
.runnable_avg
+= se
->avg
.runnable_avg
;
3882 raw_spin_unlock_irqrestore(&cfs_rq
->removed
.lock
, flags
);
3885 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq
*cfs_rq
)
3887 return cfs_rq
->avg
.runnable_avg
;
3890 static inline unsigned long cfs_rq_load_avg(struct cfs_rq
*cfs_rq
)
3892 return cfs_rq
->avg
.load_avg
;
3895 static int newidle_balance(struct rq
*this_rq
, struct rq_flags
*rf
);
3897 static inline unsigned long task_util(struct task_struct
*p
)
3899 return READ_ONCE(p
->se
.avg
.util_avg
);
3902 static inline unsigned long _task_util_est(struct task_struct
*p
)
3904 struct util_est ue
= READ_ONCE(p
->se
.avg
.util_est
);
3906 return (max(ue
.ewma
, ue
.enqueued
) | UTIL_AVG_UNCHANGED
);
3909 static inline unsigned long task_util_est(struct task_struct
*p
)
3911 return max(task_util(p
), _task_util_est(p
));
3914 #ifdef CONFIG_UCLAMP_TASK
3915 static inline unsigned long uclamp_task_util(struct task_struct
*p
)
3917 return clamp(task_util_est(p
),
3918 uclamp_eff_value(p
, UCLAMP_MIN
),
3919 uclamp_eff_value(p
, UCLAMP_MAX
));
3922 static inline unsigned long uclamp_task_util(struct task_struct
*p
)
3924 return task_util_est(p
);
3928 static inline void util_est_enqueue(struct cfs_rq
*cfs_rq
,
3929 struct task_struct
*p
)
3931 unsigned int enqueued
;
3933 if (!sched_feat(UTIL_EST
))
3936 /* Update root cfs_rq's estimated utilization */
3937 enqueued
= cfs_rq
->avg
.util_est
.enqueued
;
3938 enqueued
+= _task_util_est(p
);
3939 WRITE_ONCE(cfs_rq
->avg
.util_est
.enqueued
, enqueued
);
3941 trace_sched_util_est_cfs_tp(cfs_rq
);
3945 * Check if a (signed) value is within a specified (unsigned) margin,
3946 * based on the observation that:
3948 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3950 * NOTE: this only works when value + maring < INT_MAX.
3952 static inline bool within_margin(int value
, int margin
)
3954 return ((unsigned int)(value
+ margin
- 1) < (2 * margin
- 1));
3958 util_est_dequeue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
, bool task_sleep
)
3960 long last_ewma_diff
;
3964 if (!sched_feat(UTIL_EST
))
3967 /* Update root cfs_rq's estimated utilization */
3968 ue
.enqueued
= cfs_rq
->avg
.util_est
.enqueued
;
3969 ue
.enqueued
-= min_t(unsigned int, ue
.enqueued
, _task_util_est(p
));
3970 WRITE_ONCE(cfs_rq
->avg
.util_est
.enqueued
, ue
.enqueued
);
3972 trace_sched_util_est_cfs_tp(cfs_rq
);
3975 * Skip update of task's estimated utilization when the task has not
3976 * yet completed an activation, e.g. being migrated.
3982 * If the PELT values haven't changed since enqueue time,
3983 * skip the util_est update.
3985 ue
= p
->se
.avg
.util_est
;
3986 if (ue
.enqueued
& UTIL_AVG_UNCHANGED
)
3990 * Reset EWMA on utilization increases, the moving average is used only
3991 * to smooth utilization decreases.
3993 ue
.enqueued
= (task_util(p
) | UTIL_AVG_UNCHANGED
);
3994 if (sched_feat(UTIL_EST_FASTUP
)) {
3995 if (ue
.ewma
< ue
.enqueued
) {
3996 ue
.ewma
= ue
.enqueued
;
4002 * Skip update of task's estimated utilization when its EWMA is
4003 * already ~1% close to its last activation value.
4005 last_ewma_diff
= ue
.enqueued
- ue
.ewma
;
4006 if (within_margin(last_ewma_diff
, (SCHED_CAPACITY_SCALE
/ 100)))
4010 * To avoid overestimation of actual task utilization, skip updates if
4011 * we cannot grant there is idle time in this CPU.
4013 cpu
= cpu_of(rq_of(cfs_rq
));
4014 if (task_util(p
) > capacity_orig_of(cpu
))
4018 * Update Task's estimated utilization
4020 * When *p completes an activation we can consolidate another sample
4021 * of the task size. This is done by storing the current PELT value
4022 * as ue.enqueued and by using this value to update the Exponential
4023 * Weighted Moving Average (EWMA):
4025 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4026 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4027 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4028 * = w * ( last_ewma_diff ) + ewma(t-1)
4029 * = w * (last_ewma_diff + ewma(t-1) / w)
4031 * Where 'w' is the weight of new samples, which is configured to be
4032 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4034 ue
.ewma
<<= UTIL_EST_WEIGHT_SHIFT
;
4035 ue
.ewma
+= last_ewma_diff
;
4036 ue
.ewma
>>= UTIL_EST_WEIGHT_SHIFT
;
4038 WRITE_ONCE(p
->se
.avg
.util_est
, ue
);
4040 trace_sched_util_est_se_tp(&p
->se
);
4043 static inline int task_fits_capacity(struct task_struct
*p
, long capacity
)
4045 return fits_capacity(uclamp_task_util(p
), capacity
);
4048 static inline void update_misfit_status(struct task_struct
*p
, struct rq
*rq
)
4050 if (!static_branch_unlikely(&sched_asym_cpucapacity
))
4054 rq
->misfit_task_load
= 0;
4058 if (task_fits_capacity(p
, capacity_of(cpu_of(rq
)))) {
4059 rq
->misfit_task_load
= 0;
4064 * Make sure that misfit_task_load will not be null even if
4065 * task_h_load() returns 0.
4067 rq
->misfit_task_load
= max_t(unsigned long, task_h_load(p
), 1);
4070 #else /* CONFIG_SMP */
4072 #define UPDATE_TG 0x0
4073 #define SKIP_AGE_LOAD 0x0
4074 #define DO_ATTACH 0x0
4076 static inline void update_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int not_used1
)
4078 cfs_rq_util_change(cfs_rq
, 0);
4081 static inline void remove_entity_load_avg(struct sched_entity
*se
) {}
4084 attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
4086 detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
4088 static inline int newidle_balance(struct rq
*rq
, struct rq_flags
*rf
)
4094 util_est_enqueue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
) {}
4097 util_est_dequeue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
,
4099 static inline void update_misfit_status(struct task_struct
*p
, struct rq
*rq
) {}
4101 #endif /* CONFIG_SMP */
4103 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4105 #ifdef CONFIG_SCHED_DEBUG
4106 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
4111 if (d
> 3*sysctl_sched_latency
)
4112 schedstat_inc(cfs_rq
->nr_spread_over
);
4117 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
4119 u64 vruntime
= cfs_rq
->min_vruntime
;
4122 * The 'current' period is already promised to the current tasks,
4123 * however the extra weight of the new task will slow them down a
4124 * little, place the new task so that it fits in the slot that
4125 * stays open at the end.
4127 if (initial
&& sched_feat(START_DEBIT
))
4128 vruntime
+= sched_vslice(cfs_rq
, se
);
4130 /* sleeps up to a single latency don't count. */
4132 unsigned long thresh
= sysctl_sched_latency
;
4135 * Halve their sleep time's effect, to allow
4136 * for a gentler effect of sleepers:
4138 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
4144 /* ensure we never gain time by being placed backwards. */
4145 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
4148 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
4150 static inline void check_schedstat_required(void)
4152 #ifdef CONFIG_SCHEDSTATS
4153 if (schedstat_enabled())
4156 /* Force schedstat enabled if a dependent tracepoint is active */
4157 if (trace_sched_stat_wait_enabled() ||
4158 trace_sched_stat_sleep_enabled() ||
4159 trace_sched_stat_iowait_enabled() ||
4160 trace_sched_stat_blocked_enabled() ||
4161 trace_sched_stat_runtime_enabled()) {
4162 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4163 "stat_blocked and stat_runtime require the "
4164 "kernel parameter schedstats=enable or "
4165 "kernel.sched_schedstats=1\n");
4170 static inline bool cfs_bandwidth_used(void);
4177 * update_min_vruntime()
4178 * vruntime -= min_vruntime
4182 * update_min_vruntime()
4183 * vruntime += min_vruntime
4185 * this way the vruntime transition between RQs is done when both
4186 * min_vruntime are up-to-date.
4190 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4191 * vruntime -= min_vruntime
4195 * update_min_vruntime()
4196 * vruntime += min_vruntime
4198 * this way we don't have the most up-to-date min_vruntime on the originating
4199 * CPU and an up-to-date min_vruntime on the destination CPU.
4203 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
4205 bool renorm
= !(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_MIGRATED
);
4206 bool curr
= cfs_rq
->curr
== se
;
4209 * If we're the current task, we must renormalise before calling
4213 se
->vruntime
+= cfs_rq
->min_vruntime
;
4215 update_curr(cfs_rq
);
4218 * Otherwise, renormalise after, such that we're placed at the current
4219 * moment in time, instead of some random moment in the past. Being
4220 * placed in the past could significantly boost this task to the
4221 * fairness detriment of existing tasks.
4223 if (renorm
&& !curr
)
4224 se
->vruntime
+= cfs_rq
->min_vruntime
;
4227 * When enqueuing a sched_entity, we must:
4228 * - Update loads to have both entity and cfs_rq synced with now.
4229 * - Add its load to cfs_rq->runnable_avg
4230 * - For group_entity, update its weight to reflect the new share of
4232 * - Add its new weight to cfs_rq->load.weight
4234 update_load_avg(cfs_rq
, se
, UPDATE_TG
| DO_ATTACH
);
4235 se_update_runnable(se
);
4236 update_cfs_group(se
);
4237 account_entity_enqueue(cfs_rq
, se
);
4239 if (flags
& ENQUEUE_WAKEUP
)
4240 place_entity(cfs_rq
, se
, 0);
4242 check_schedstat_required();
4243 update_stats_enqueue(cfs_rq
, se
, flags
);
4244 check_spread(cfs_rq
, se
);
4246 __enqueue_entity(cfs_rq
, se
);
4250 * When bandwidth control is enabled, cfs might have been removed
4251 * because of a parent been throttled but cfs->nr_running > 1. Try to
4252 * add it unconditionnally.
4254 if (cfs_rq
->nr_running
== 1 || cfs_bandwidth_used())
4255 list_add_leaf_cfs_rq(cfs_rq
);
4257 if (cfs_rq
->nr_running
== 1)
4258 check_enqueue_throttle(cfs_rq
);
4261 static void __clear_buddies_last(struct sched_entity
*se
)
4263 for_each_sched_entity(se
) {
4264 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4265 if (cfs_rq
->last
!= se
)
4268 cfs_rq
->last
= NULL
;
4272 static void __clear_buddies_next(struct sched_entity
*se
)
4274 for_each_sched_entity(se
) {
4275 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4276 if (cfs_rq
->next
!= se
)
4279 cfs_rq
->next
= NULL
;
4283 static void __clear_buddies_skip(struct sched_entity
*se
)
4285 for_each_sched_entity(se
) {
4286 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4287 if (cfs_rq
->skip
!= se
)
4290 cfs_rq
->skip
= NULL
;
4294 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4296 if (cfs_rq
->last
== se
)
4297 __clear_buddies_last(se
);
4299 if (cfs_rq
->next
== se
)
4300 __clear_buddies_next(se
);
4302 if (cfs_rq
->skip
== se
)
4303 __clear_buddies_skip(se
);
4306 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
4309 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
4312 * Update run-time statistics of the 'current'.
4314 update_curr(cfs_rq
);
4317 * When dequeuing a sched_entity, we must:
4318 * - Update loads to have both entity and cfs_rq synced with now.
4319 * - Subtract its load from the cfs_rq->runnable_avg.
4320 * - Subtract its previous weight from cfs_rq->load.weight.
4321 * - For group entity, update its weight to reflect the new share
4322 * of its group cfs_rq.
4324 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4325 se_update_runnable(se
);
4327 update_stats_dequeue(cfs_rq
, se
, flags
);
4329 clear_buddies(cfs_rq
, se
);
4331 if (se
!= cfs_rq
->curr
)
4332 __dequeue_entity(cfs_rq
, se
);
4334 account_entity_dequeue(cfs_rq
, se
);
4337 * Normalize after update_curr(); which will also have moved
4338 * min_vruntime if @se is the one holding it back. But before doing
4339 * update_min_vruntime() again, which will discount @se's position and
4340 * can move min_vruntime forward still more.
4342 if (!(flags
& DEQUEUE_SLEEP
))
4343 se
->vruntime
-= cfs_rq
->min_vruntime
;
4345 /* return excess runtime on last dequeue */
4346 return_cfs_rq_runtime(cfs_rq
);
4348 update_cfs_group(se
);
4351 * Now advance min_vruntime if @se was the entity holding it back,
4352 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4353 * put back on, and if we advance min_vruntime, we'll be placed back
4354 * further than we started -- ie. we'll be penalized.
4356 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) != DEQUEUE_SAVE
)
4357 update_min_vruntime(cfs_rq
);
4361 * Preempt the current task with a newly woken task if needed:
4364 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
4366 unsigned long ideal_runtime
, delta_exec
;
4367 struct sched_entity
*se
;
4370 ideal_runtime
= sched_slice(cfs_rq
, curr
);
4371 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
4372 if (delta_exec
> ideal_runtime
) {
4373 resched_curr(rq_of(cfs_rq
));
4375 * The current task ran long enough, ensure it doesn't get
4376 * re-elected due to buddy favours.
4378 clear_buddies(cfs_rq
, curr
);
4383 * Ensure that a task that missed wakeup preemption by a
4384 * narrow margin doesn't have to wait for a full slice.
4385 * This also mitigates buddy induced latencies under load.
4387 if (delta_exec
< sysctl_sched_min_granularity
)
4390 se
= __pick_first_entity(cfs_rq
);
4391 delta
= curr
->vruntime
- se
->vruntime
;
4396 if (delta
> ideal_runtime
)
4397 resched_curr(rq_of(cfs_rq
));
4401 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4403 /* 'current' is not kept within the tree. */
4406 * Any task has to be enqueued before it get to execute on
4407 * a CPU. So account for the time it spent waiting on the
4410 update_stats_wait_end(cfs_rq
, se
);
4411 __dequeue_entity(cfs_rq
, se
);
4412 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4415 update_stats_curr_start(cfs_rq
, se
);
4419 * Track our maximum slice length, if the CPU's load is at
4420 * least twice that of our own weight (i.e. dont track it
4421 * when there are only lesser-weight tasks around):
4423 if (schedstat_enabled() &&
4424 rq_of(cfs_rq
)->cfs
.load
.weight
>= 2*se
->load
.weight
) {
4425 schedstat_set(se
->statistics
.slice_max
,
4426 max((u64
)schedstat_val(se
->statistics
.slice_max
),
4427 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
));
4430 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
4434 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
4437 * Pick the next process, keeping these things in mind, in this order:
4438 * 1) keep things fair between processes/task groups
4439 * 2) pick the "next" process, since someone really wants that to run
4440 * 3) pick the "last" process, for cache locality
4441 * 4) do not run the "skip" process, if something else is available
4443 static struct sched_entity
*
4444 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
4446 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
4447 struct sched_entity
*se
;
4450 * If curr is set we have to see if its left of the leftmost entity
4451 * still in the tree, provided there was anything in the tree at all.
4453 if (!left
|| (curr
&& entity_before(curr
, left
)))
4456 se
= left
; /* ideally we run the leftmost entity */
4459 * Avoid running the skip buddy, if running something else can
4460 * be done without getting too unfair.
4462 if (cfs_rq
->skip
== se
) {
4463 struct sched_entity
*second
;
4466 second
= __pick_first_entity(cfs_rq
);
4468 second
= __pick_next_entity(se
);
4469 if (!second
|| (curr
&& entity_before(curr
, second
)))
4473 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
4477 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1) {
4479 * Someone really wants this to run. If it's not unfair, run it.
4482 } else if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1) {
4484 * Prefer last buddy, try to return the CPU to a preempted task.
4489 clear_buddies(cfs_rq
, se
);
4494 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
4496 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
4499 * If still on the runqueue then deactivate_task()
4500 * was not called and update_curr() has to be done:
4503 update_curr(cfs_rq
);
4505 /* throttle cfs_rqs exceeding runtime */
4506 check_cfs_rq_runtime(cfs_rq
);
4508 check_spread(cfs_rq
, prev
);
4511 update_stats_wait_start(cfs_rq
, prev
);
4512 /* Put 'current' back into the tree. */
4513 __enqueue_entity(cfs_rq
, prev
);
4514 /* in !on_rq case, update occurred at dequeue */
4515 update_load_avg(cfs_rq
, prev
, 0);
4517 cfs_rq
->curr
= NULL
;
4521 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
4524 * Update run-time statistics of the 'current'.
4526 update_curr(cfs_rq
);
4529 * Ensure that runnable average is periodically updated.
4531 update_load_avg(cfs_rq
, curr
, UPDATE_TG
);
4532 update_cfs_group(curr
);
4534 #ifdef CONFIG_SCHED_HRTICK
4536 * queued ticks are scheduled to match the slice, so don't bother
4537 * validating it and just reschedule.
4540 resched_curr(rq_of(cfs_rq
));
4544 * don't let the period tick interfere with the hrtick preemption
4546 if (!sched_feat(DOUBLE_TICK
) &&
4547 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
4551 if (cfs_rq
->nr_running
> 1)
4552 check_preempt_tick(cfs_rq
, curr
);
4556 /**************************************************
4557 * CFS bandwidth control machinery
4560 #ifdef CONFIG_CFS_BANDWIDTH
4562 #ifdef CONFIG_JUMP_LABEL
4563 static struct static_key __cfs_bandwidth_used
;
4565 static inline bool cfs_bandwidth_used(void)
4567 return static_key_false(&__cfs_bandwidth_used
);
4570 void cfs_bandwidth_usage_inc(void)
4572 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used
);
4575 void cfs_bandwidth_usage_dec(void)
4577 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used
);
4579 #else /* CONFIG_JUMP_LABEL */
4580 static bool cfs_bandwidth_used(void)
4585 void cfs_bandwidth_usage_inc(void) {}
4586 void cfs_bandwidth_usage_dec(void) {}
4587 #endif /* CONFIG_JUMP_LABEL */
4590 * default period for cfs group bandwidth.
4591 * default: 0.1s, units: nanoseconds
4593 static inline u64
default_cfs_period(void)
4595 return 100000000ULL;
4598 static inline u64
sched_cfs_bandwidth_slice(void)
4600 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
4604 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4605 * directly instead of rq->clock to avoid adding additional synchronization
4608 * requires cfs_b->lock
4610 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
4612 if (cfs_b
->quota
!= RUNTIME_INF
)
4613 cfs_b
->runtime
= cfs_b
->quota
;
4616 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4618 return &tg
->cfs_bandwidth
;
4621 /* returns 0 on failure to allocate runtime */
4622 static int __assign_cfs_rq_runtime(struct cfs_bandwidth
*cfs_b
,
4623 struct cfs_rq
*cfs_rq
, u64 target_runtime
)
4625 u64 min_amount
, amount
= 0;
4627 lockdep_assert_held(&cfs_b
->lock
);
4629 /* note: this is a positive sum as runtime_remaining <= 0 */
4630 min_amount
= target_runtime
- cfs_rq
->runtime_remaining
;
4632 if (cfs_b
->quota
== RUNTIME_INF
)
4633 amount
= min_amount
;
4635 start_cfs_bandwidth(cfs_b
);
4637 if (cfs_b
->runtime
> 0) {
4638 amount
= min(cfs_b
->runtime
, min_amount
);
4639 cfs_b
->runtime
-= amount
;
4644 cfs_rq
->runtime_remaining
+= amount
;
4646 return cfs_rq
->runtime_remaining
> 0;
4649 /* returns 0 on failure to allocate runtime */
4650 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4652 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4655 raw_spin_lock(&cfs_b
->lock
);
4656 ret
= __assign_cfs_rq_runtime(cfs_b
, cfs_rq
, sched_cfs_bandwidth_slice());
4657 raw_spin_unlock(&cfs_b
->lock
);
4662 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4664 /* dock delta_exec before expiring quota (as it could span periods) */
4665 cfs_rq
->runtime_remaining
-= delta_exec
;
4667 if (likely(cfs_rq
->runtime_remaining
> 0))
4670 if (cfs_rq
->throttled
)
4673 * if we're unable to extend our runtime we resched so that the active
4674 * hierarchy can be throttled
4676 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
4677 resched_curr(rq_of(cfs_rq
));
4680 static __always_inline
4681 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4683 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
4686 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
4689 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4691 return cfs_bandwidth_used() && cfs_rq
->throttled
;
4694 /* check whether cfs_rq, or any parent, is throttled */
4695 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4697 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
4701 * Ensure that neither of the group entities corresponding to src_cpu or
4702 * dest_cpu are members of a throttled hierarchy when performing group
4703 * load-balance operations.
4705 static inline int throttled_lb_pair(struct task_group
*tg
,
4706 int src_cpu
, int dest_cpu
)
4708 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
4710 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
4711 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
4713 return throttled_hierarchy(src_cfs_rq
) ||
4714 throttled_hierarchy(dest_cfs_rq
);
4717 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
4719 struct rq
*rq
= data
;
4720 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4722 cfs_rq
->throttle_count
--;
4723 if (!cfs_rq
->throttle_count
) {
4724 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
4725 cfs_rq
->throttled_clock_task
;
4727 /* Add cfs_rq with already running entity in the list */
4728 if (cfs_rq
->nr_running
>= 1)
4729 list_add_leaf_cfs_rq(cfs_rq
);
4735 static int tg_throttle_down(struct task_group
*tg
, void *data
)
4737 struct rq
*rq
= data
;
4738 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4740 /* group is entering throttled state, stop time */
4741 if (!cfs_rq
->throttle_count
) {
4742 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
4743 list_del_leaf_cfs_rq(cfs_rq
);
4745 cfs_rq
->throttle_count
++;
4750 static bool throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
4752 struct rq
*rq
= rq_of(cfs_rq
);
4753 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4754 struct sched_entity
*se
;
4755 long task_delta
, idle_task_delta
, dequeue
= 1;
4757 raw_spin_lock(&cfs_b
->lock
);
4758 /* This will start the period timer if necessary */
4759 if (__assign_cfs_rq_runtime(cfs_b
, cfs_rq
, 1)) {
4761 * We have raced with bandwidth becoming available, and if we
4762 * actually throttled the timer might not unthrottle us for an
4763 * entire period. We additionally needed to make sure that any
4764 * subsequent check_cfs_rq_runtime calls agree not to throttle
4765 * us, as we may commit to do cfs put_prev+pick_next, so we ask
4766 * for 1ns of runtime rather than just check cfs_b.
4770 list_add_tail_rcu(&cfs_rq
->throttled_list
,
4771 &cfs_b
->throttled_cfs_rq
);
4773 raw_spin_unlock(&cfs_b
->lock
);
4776 return false; /* Throttle no longer required. */
4778 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
4780 /* freeze hierarchy runnable averages while throttled */
4782 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
4785 task_delta
= cfs_rq
->h_nr_running
;
4786 idle_task_delta
= cfs_rq
->idle_h_nr_running
;
4787 for_each_sched_entity(se
) {
4788 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4789 /* throttled entity or throttle-on-deactivate */
4793 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
4795 qcfs_rq
->h_nr_running
-= task_delta
;
4796 qcfs_rq
->idle_h_nr_running
-= idle_task_delta
;
4798 if (qcfs_rq
->load
.weight
) {
4799 /* Avoid re-evaluating load for this entity: */
4800 se
= parent_entity(se
);
4805 for_each_sched_entity(se
) {
4806 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4807 /* throttled entity or throttle-on-deactivate */
4811 update_load_avg(qcfs_rq
, se
, 0);
4812 se_update_runnable(se
);
4814 qcfs_rq
->h_nr_running
-= task_delta
;
4815 qcfs_rq
->idle_h_nr_running
-= idle_task_delta
;
4818 /* At this point se is NULL and we are at root level*/
4819 sub_nr_running(rq
, task_delta
);
4823 * Note: distribution will already see us throttled via the
4824 * throttled-list. rq->lock protects completion.
4826 cfs_rq
->throttled
= 1;
4827 cfs_rq
->throttled_clock
= rq_clock(rq
);
4831 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
4833 struct rq
*rq
= rq_of(cfs_rq
);
4834 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4835 struct sched_entity
*se
;
4836 long task_delta
, idle_task_delta
;
4838 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
4840 cfs_rq
->throttled
= 0;
4842 update_rq_clock(rq
);
4844 raw_spin_lock(&cfs_b
->lock
);
4845 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
4846 list_del_rcu(&cfs_rq
->throttled_list
);
4847 raw_spin_unlock(&cfs_b
->lock
);
4849 /* update hierarchical throttle state */
4850 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
4852 if (!cfs_rq
->load
.weight
)
4855 task_delta
= cfs_rq
->h_nr_running
;
4856 idle_task_delta
= cfs_rq
->idle_h_nr_running
;
4857 for_each_sched_entity(se
) {
4860 cfs_rq
= cfs_rq_of(se
);
4861 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
4863 cfs_rq
->h_nr_running
+= task_delta
;
4864 cfs_rq
->idle_h_nr_running
+= idle_task_delta
;
4866 /* end evaluation on encountering a throttled cfs_rq */
4867 if (cfs_rq_throttled(cfs_rq
))
4868 goto unthrottle_throttle
;
4871 for_each_sched_entity(se
) {
4872 cfs_rq
= cfs_rq_of(se
);
4874 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4875 se_update_runnable(se
);
4877 cfs_rq
->h_nr_running
+= task_delta
;
4878 cfs_rq
->idle_h_nr_running
+= idle_task_delta
;
4881 /* end evaluation on encountering a throttled cfs_rq */
4882 if (cfs_rq_throttled(cfs_rq
))
4883 goto unthrottle_throttle
;
4886 * One parent has been throttled and cfs_rq removed from the
4887 * list. Add it back to not break the leaf list.
4889 if (throttled_hierarchy(cfs_rq
))
4890 list_add_leaf_cfs_rq(cfs_rq
);
4893 /* At this point se is NULL and we are at root level*/
4894 add_nr_running(rq
, task_delta
);
4896 unthrottle_throttle
:
4898 * The cfs_rq_throttled() breaks in the above iteration can result in
4899 * incomplete leaf list maintenance, resulting in triggering the
4902 for_each_sched_entity(se
) {
4903 cfs_rq
= cfs_rq_of(se
);
4905 if (list_add_leaf_cfs_rq(cfs_rq
))
4909 assert_list_leaf_cfs_rq(rq
);
4911 /* Determine whether we need to wake up potentially idle CPU: */
4912 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
4916 static void distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
)
4918 struct cfs_rq
*cfs_rq
;
4919 u64 runtime
, remaining
= 1;
4922 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
4924 struct rq
*rq
= rq_of(cfs_rq
);
4927 rq_lock_irqsave(rq
, &rf
);
4928 if (!cfs_rq_throttled(cfs_rq
))
4931 /* By the above check, this should never be true */
4932 SCHED_WARN_ON(cfs_rq
->runtime_remaining
> 0);
4934 raw_spin_lock(&cfs_b
->lock
);
4935 runtime
= -cfs_rq
->runtime_remaining
+ 1;
4936 if (runtime
> cfs_b
->runtime
)
4937 runtime
= cfs_b
->runtime
;
4938 cfs_b
->runtime
-= runtime
;
4939 remaining
= cfs_b
->runtime
;
4940 raw_spin_unlock(&cfs_b
->lock
);
4942 cfs_rq
->runtime_remaining
+= runtime
;
4944 /* we check whether we're throttled above */
4945 if (cfs_rq
->runtime_remaining
> 0)
4946 unthrottle_cfs_rq(cfs_rq
);
4949 rq_unlock_irqrestore(rq
, &rf
);
4958 * Responsible for refilling a task_group's bandwidth and unthrottling its
4959 * cfs_rqs as appropriate. If there has been no activity within the last
4960 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4961 * used to track this state.
4963 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
, unsigned long flags
)
4967 /* no need to continue the timer with no bandwidth constraint */
4968 if (cfs_b
->quota
== RUNTIME_INF
)
4969 goto out_deactivate
;
4971 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4972 cfs_b
->nr_periods
+= overrun
;
4975 * idle depends on !throttled (for the case of a large deficit), and if
4976 * we're going inactive then everything else can be deferred
4978 if (cfs_b
->idle
&& !throttled
)
4979 goto out_deactivate
;
4981 __refill_cfs_bandwidth_runtime(cfs_b
);
4984 /* mark as potentially idle for the upcoming period */
4989 /* account preceding periods in which throttling occurred */
4990 cfs_b
->nr_throttled
+= overrun
;
4993 * This check is repeated as we release cfs_b->lock while we unthrottle.
4995 while (throttled
&& cfs_b
->runtime
> 0) {
4996 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
4997 /* we can't nest cfs_b->lock while distributing bandwidth */
4998 distribute_cfs_runtime(cfs_b
);
4999 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5001 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
5005 * While we are ensured activity in the period following an
5006 * unthrottle, this also covers the case in which the new bandwidth is
5007 * insufficient to cover the existing bandwidth deficit. (Forcing the
5008 * timer to remain active while there are any throttled entities.)
5018 /* a cfs_rq won't donate quota below this amount */
5019 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
5020 /* minimum remaining period time to redistribute slack quota */
5021 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
5022 /* how long we wait to gather additional slack before distributing */
5023 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
5026 * Are we near the end of the current quota period?
5028 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5029 * hrtimer base being cleared by hrtimer_start. In the case of
5030 * migrate_hrtimers, base is never cleared, so we are fine.
5032 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
5034 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
5037 /* if the call-back is running a quota refresh is already occurring */
5038 if (hrtimer_callback_running(refresh_timer
))
5041 /* is a quota refresh about to occur? */
5042 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
5043 if (remaining
< min_expire
)
5049 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
5051 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
5053 /* if there's a quota refresh soon don't bother with slack */
5054 if (runtime_refresh_within(cfs_b
, min_left
))
5057 /* don't push forwards an existing deferred unthrottle */
5058 if (cfs_b
->slack_started
)
5060 cfs_b
->slack_started
= true;
5062 hrtimer_start(&cfs_b
->slack_timer
,
5063 ns_to_ktime(cfs_bandwidth_slack_period
),
5067 /* we know any runtime found here is valid as update_curr() precedes return */
5068 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5070 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
5071 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
5073 if (slack_runtime
<= 0)
5076 raw_spin_lock(&cfs_b
->lock
);
5077 if (cfs_b
->quota
!= RUNTIME_INF
) {
5078 cfs_b
->runtime
+= slack_runtime
;
5080 /* we are under rq->lock, defer unthrottling using a timer */
5081 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
5082 !list_empty(&cfs_b
->throttled_cfs_rq
))
5083 start_cfs_slack_bandwidth(cfs_b
);
5085 raw_spin_unlock(&cfs_b
->lock
);
5087 /* even if it's not valid for return we don't want to try again */
5088 cfs_rq
->runtime_remaining
-= slack_runtime
;
5091 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5093 if (!cfs_bandwidth_used())
5096 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
5099 __return_cfs_rq_runtime(cfs_rq
);
5103 * This is done with a timer (instead of inline with bandwidth return) since
5104 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5106 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
5108 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
5109 unsigned long flags
;
5111 /* confirm we're still not at a refresh boundary */
5112 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5113 cfs_b
->slack_started
= false;
5115 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
5116 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5120 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
5121 runtime
= cfs_b
->runtime
;
5123 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5128 distribute_cfs_runtime(cfs_b
);
5132 * When a group wakes up we want to make sure that its quota is not already
5133 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5134 * runtime as update_curr() throttling can not not trigger until it's on-rq.
5136 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
5138 if (!cfs_bandwidth_used())
5141 /* an active group must be handled by the update_curr()->put() path */
5142 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
5145 /* ensure the group is not already throttled */
5146 if (cfs_rq_throttled(cfs_rq
))
5149 /* update runtime allocation */
5150 account_cfs_rq_runtime(cfs_rq
, 0);
5151 if (cfs_rq
->runtime_remaining
<= 0)
5152 throttle_cfs_rq(cfs_rq
);
5155 static void sync_throttle(struct task_group
*tg
, int cpu
)
5157 struct cfs_rq
*pcfs_rq
, *cfs_rq
;
5159 if (!cfs_bandwidth_used())
5165 cfs_rq
= tg
->cfs_rq
[cpu
];
5166 pcfs_rq
= tg
->parent
->cfs_rq
[cpu
];
5168 cfs_rq
->throttle_count
= pcfs_rq
->throttle_count
;
5169 cfs_rq
->throttled_clock_task
= rq_clock_task(cpu_rq(cpu
));
5172 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5173 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5175 if (!cfs_bandwidth_used())
5178 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
5182 * it's possible for a throttled entity to be forced into a running
5183 * state (e.g. set_curr_task), in this case we're finished.
5185 if (cfs_rq_throttled(cfs_rq
))
5188 return throttle_cfs_rq(cfs_rq
);
5191 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
5193 struct cfs_bandwidth
*cfs_b
=
5194 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
5196 do_sched_cfs_slack_timer(cfs_b
);
5198 return HRTIMER_NORESTART
;
5201 extern const u64 max_cfs_quota_period
;
5203 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
5205 struct cfs_bandwidth
*cfs_b
=
5206 container_of(timer
, struct cfs_bandwidth
, period_timer
);
5207 unsigned long flags
;
5212 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5214 overrun
= hrtimer_forward_now(timer
, cfs_b
->period
);
5218 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
, flags
);
5221 u64
new, old
= ktime_to_ns(cfs_b
->period
);
5224 * Grow period by a factor of 2 to avoid losing precision.
5225 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5229 if (new < max_cfs_quota_period
) {
5230 cfs_b
->period
= ns_to_ktime(new);
5233 pr_warn_ratelimited(
5234 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5236 div_u64(new, NSEC_PER_USEC
),
5237 div_u64(cfs_b
->quota
, NSEC_PER_USEC
));
5239 pr_warn_ratelimited(
5240 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5242 div_u64(old
, NSEC_PER_USEC
),
5243 div_u64(cfs_b
->quota
, NSEC_PER_USEC
));
5246 /* reset count so we don't come right back in here */
5251 cfs_b
->period_active
= 0;
5252 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5254 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
5257 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5259 raw_spin_lock_init(&cfs_b
->lock
);
5261 cfs_b
->quota
= RUNTIME_INF
;
5262 cfs_b
->period
= ns_to_ktime(default_cfs_period());
5264 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
5265 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_PINNED
);
5266 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
5267 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5268 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
5269 cfs_b
->slack_started
= false;
5272 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5274 cfs_rq
->runtime_enabled
= 0;
5275 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
5278 void start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5280 lockdep_assert_held(&cfs_b
->lock
);
5282 if (cfs_b
->period_active
)
5285 cfs_b
->period_active
= 1;
5286 hrtimer_forward_now(&cfs_b
->period_timer
, cfs_b
->period
);
5287 hrtimer_start_expires(&cfs_b
->period_timer
, HRTIMER_MODE_ABS_PINNED
);
5290 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5292 /* init_cfs_bandwidth() was not called */
5293 if (!cfs_b
->throttled_cfs_rq
.next
)
5296 hrtimer_cancel(&cfs_b
->period_timer
);
5297 hrtimer_cancel(&cfs_b
->slack_timer
);
5301 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5303 * The race is harmless, since modifying bandwidth settings of unhooked group
5304 * bits doesn't do much.
5307 /* cpu online calback */
5308 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
5310 struct task_group
*tg
;
5312 lockdep_assert_held(&rq
->lock
);
5315 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
5316 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
5317 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
5319 raw_spin_lock(&cfs_b
->lock
);
5320 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
5321 raw_spin_unlock(&cfs_b
->lock
);
5326 /* cpu offline callback */
5327 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
5329 struct task_group
*tg
;
5331 lockdep_assert_held(&rq
->lock
);
5334 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
5335 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
5337 if (!cfs_rq
->runtime_enabled
)
5341 * clock_task is not advancing so we just need to make sure
5342 * there's some valid quota amount
5344 cfs_rq
->runtime_remaining
= 1;
5346 * Offline rq is schedulable till CPU is completely disabled
5347 * in take_cpu_down(), so we prevent new cfs throttling here.
5349 cfs_rq
->runtime_enabled
= 0;
5351 if (cfs_rq_throttled(cfs_rq
))
5352 unthrottle_cfs_rq(cfs_rq
);
5357 #else /* CONFIG_CFS_BANDWIDTH */
5359 static inline bool cfs_bandwidth_used(void)
5364 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
5365 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
5366 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
5367 static inline void sync_throttle(struct task_group
*tg
, int cpu
) {}
5368 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
5370 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
5375 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
5380 static inline int throttled_lb_pair(struct task_group
*tg
,
5381 int src_cpu
, int dest_cpu
)
5386 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
5388 #ifdef CONFIG_FAIR_GROUP_SCHED
5389 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
5392 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
5396 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
5397 static inline void update_runtime_enabled(struct rq
*rq
) {}
5398 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
5400 #endif /* CONFIG_CFS_BANDWIDTH */
5402 /**************************************************
5403 * CFS operations on tasks:
5406 #ifdef CONFIG_SCHED_HRTICK
5407 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
5409 struct sched_entity
*se
= &p
->se
;
5410 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5412 SCHED_WARN_ON(task_rq(p
) != rq
);
5414 if (rq
->cfs
.h_nr_running
> 1) {
5415 u64 slice
= sched_slice(cfs_rq
, se
);
5416 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
5417 s64 delta
= slice
- ran
;
5424 hrtick_start(rq
, delta
);
5429 * called from enqueue/dequeue and updates the hrtick when the
5430 * current task is from our class and nr_running is low enough
5433 static void hrtick_update(struct rq
*rq
)
5435 struct task_struct
*curr
= rq
->curr
;
5437 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
5440 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
5441 hrtick_start_fair(rq
, curr
);
5443 #else /* !CONFIG_SCHED_HRTICK */
5445 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
5449 static inline void hrtick_update(struct rq
*rq
)
5455 static inline unsigned long cpu_util(int cpu
);
5457 static inline bool cpu_overutilized(int cpu
)
5459 return !fits_capacity(cpu_util(cpu
), capacity_of(cpu
));
5462 static inline void update_overutilized_status(struct rq
*rq
)
5464 if (!READ_ONCE(rq
->rd
->overutilized
) && cpu_overutilized(rq
->cpu
)) {
5465 WRITE_ONCE(rq
->rd
->overutilized
, SG_OVERUTILIZED
);
5466 trace_sched_overutilized_tp(rq
->rd
, SG_OVERUTILIZED
);
5470 static inline void update_overutilized_status(struct rq
*rq
) { }
5473 /* Runqueue only has SCHED_IDLE tasks enqueued */
5474 static int sched_idle_rq(struct rq
*rq
)
5476 return unlikely(rq
->nr_running
== rq
->cfs
.idle_h_nr_running
&&
5481 static int sched_idle_cpu(int cpu
)
5483 return sched_idle_rq(cpu_rq(cpu
));
5488 * The enqueue_task method is called before nr_running is
5489 * increased. Here we update the fair scheduling stats and
5490 * then put the task into the rbtree:
5493 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
5495 struct cfs_rq
*cfs_rq
;
5496 struct sched_entity
*se
= &p
->se
;
5497 int idle_h_nr_running
= task_has_idle_policy(p
);
5500 * The code below (indirectly) updates schedutil which looks at
5501 * the cfs_rq utilization to select a frequency.
5502 * Let's add the task's estimated utilization to the cfs_rq's
5503 * estimated utilization, before we update schedutil.
5505 util_est_enqueue(&rq
->cfs
, p
);
5508 * If in_iowait is set, the code below may not trigger any cpufreq
5509 * utilization updates, so do it here explicitly with the IOWAIT flag
5513 cpufreq_update_util(rq
, SCHED_CPUFREQ_IOWAIT
);
5515 for_each_sched_entity(se
) {
5518 cfs_rq
= cfs_rq_of(se
);
5519 enqueue_entity(cfs_rq
, se
, flags
);
5521 cfs_rq
->h_nr_running
++;
5522 cfs_rq
->idle_h_nr_running
+= idle_h_nr_running
;
5524 /* end evaluation on encountering a throttled cfs_rq */
5525 if (cfs_rq_throttled(cfs_rq
))
5526 goto enqueue_throttle
;
5528 flags
= ENQUEUE_WAKEUP
;
5531 for_each_sched_entity(se
) {
5532 cfs_rq
= cfs_rq_of(se
);
5534 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
5535 se_update_runnable(se
);
5536 update_cfs_group(se
);
5538 cfs_rq
->h_nr_running
++;
5539 cfs_rq
->idle_h_nr_running
+= idle_h_nr_running
;
5541 /* end evaluation on encountering a throttled cfs_rq */
5542 if (cfs_rq_throttled(cfs_rq
))
5543 goto enqueue_throttle
;
5546 * One parent has been throttled and cfs_rq removed from the
5547 * list. Add it back to not break the leaf list.
5549 if (throttled_hierarchy(cfs_rq
))
5550 list_add_leaf_cfs_rq(cfs_rq
);
5553 /* At this point se is NULL and we are at root level*/
5554 add_nr_running(rq
, 1);
5557 * Since new tasks are assigned an initial util_avg equal to
5558 * half of the spare capacity of their CPU, tiny tasks have the
5559 * ability to cross the overutilized threshold, which will
5560 * result in the load balancer ruining all the task placement
5561 * done by EAS. As a way to mitigate that effect, do not account
5562 * for the first enqueue operation of new tasks during the
5563 * overutilized flag detection.
5565 * A better way of solving this problem would be to wait for
5566 * the PELT signals of tasks to converge before taking them
5567 * into account, but that is not straightforward to implement,
5568 * and the following generally works well enough in practice.
5570 if (flags
& ENQUEUE_WAKEUP
)
5571 update_overutilized_status(rq
);
5574 if (cfs_bandwidth_used()) {
5576 * When bandwidth control is enabled; the cfs_rq_throttled()
5577 * breaks in the above iteration can result in incomplete
5578 * leaf list maintenance, resulting in triggering the assertion
5581 for_each_sched_entity(se
) {
5582 cfs_rq
= cfs_rq_of(se
);
5584 if (list_add_leaf_cfs_rq(cfs_rq
))
5589 assert_list_leaf_cfs_rq(rq
);
5594 static void set_next_buddy(struct sched_entity
*se
);
5597 * The dequeue_task method is called before nr_running is
5598 * decreased. We remove the task from the rbtree and
5599 * update the fair scheduling stats:
5601 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
5603 struct cfs_rq
*cfs_rq
;
5604 struct sched_entity
*se
= &p
->se
;
5605 int task_sleep
= flags
& DEQUEUE_SLEEP
;
5606 int idle_h_nr_running
= task_has_idle_policy(p
);
5607 bool was_sched_idle
= sched_idle_rq(rq
);
5609 for_each_sched_entity(se
) {
5610 cfs_rq
= cfs_rq_of(se
);
5611 dequeue_entity(cfs_rq
, se
, flags
);
5613 cfs_rq
->h_nr_running
--;
5614 cfs_rq
->idle_h_nr_running
-= idle_h_nr_running
;
5616 /* end evaluation on encountering a throttled cfs_rq */
5617 if (cfs_rq_throttled(cfs_rq
))
5618 goto dequeue_throttle
;
5620 /* Don't dequeue parent if it has other entities besides us */
5621 if (cfs_rq
->load
.weight
) {
5622 /* Avoid re-evaluating load for this entity: */
5623 se
= parent_entity(se
);
5625 * Bias pick_next to pick a task from this cfs_rq, as
5626 * p is sleeping when it is within its sched_slice.
5628 if (task_sleep
&& se
&& !throttled_hierarchy(cfs_rq
))
5632 flags
|= DEQUEUE_SLEEP
;
5635 for_each_sched_entity(se
) {
5636 cfs_rq
= cfs_rq_of(se
);
5638 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
5639 se_update_runnable(se
);
5640 update_cfs_group(se
);
5642 cfs_rq
->h_nr_running
--;
5643 cfs_rq
->idle_h_nr_running
-= idle_h_nr_running
;
5645 /* end evaluation on encountering a throttled cfs_rq */
5646 if (cfs_rq_throttled(cfs_rq
))
5647 goto dequeue_throttle
;
5651 /* At this point se is NULL and we are at root level*/
5652 sub_nr_running(rq
, 1);
5654 /* balance early to pull high priority tasks */
5655 if (unlikely(!was_sched_idle
&& sched_idle_rq(rq
)))
5656 rq
->next_balance
= jiffies
;
5659 util_est_dequeue(&rq
->cfs
, p
, task_sleep
);
5665 /* Working cpumask for: load_balance, load_balance_newidle. */
5666 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5667 DEFINE_PER_CPU(cpumask_var_t
, select_idle_mask
);
5669 #ifdef CONFIG_NO_HZ_COMMON
5672 cpumask_var_t idle_cpus_mask
;
5674 int has_blocked
; /* Idle CPUS has blocked load */
5675 unsigned long next_balance
; /* in jiffy units */
5676 unsigned long next_blocked
; /* Next update of blocked load in jiffies */
5677 } nohz ____cacheline_aligned
;
5679 #endif /* CONFIG_NO_HZ_COMMON */
5681 static unsigned long cpu_load(struct rq
*rq
)
5683 return cfs_rq_load_avg(&rq
->cfs
);
5687 * cpu_load_without - compute CPU load without any contributions from *p
5688 * @cpu: the CPU which load is requested
5689 * @p: the task which load should be discounted
5691 * The load of a CPU is defined by the load of tasks currently enqueued on that
5692 * CPU as well as tasks which are currently sleeping after an execution on that
5695 * This method returns the load of the specified CPU by discounting the load of
5696 * the specified task, whenever the task is currently contributing to the CPU
5699 static unsigned long cpu_load_without(struct rq
*rq
, struct task_struct
*p
)
5701 struct cfs_rq
*cfs_rq
;
5704 /* Task has no contribution or is new */
5705 if (cpu_of(rq
) != task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
5706 return cpu_load(rq
);
5709 load
= READ_ONCE(cfs_rq
->avg
.load_avg
);
5711 /* Discount task's util from CPU's util */
5712 lsub_positive(&load
, task_h_load(p
));
5717 static unsigned long cpu_runnable(struct rq
*rq
)
5719 return cfs_rq_runnable_avg(&rq
->cfs
);
5722 static unsigned long cpu_runnable_without(struct rq
*rq
, struct task_struct
*p
)
5724 struct cfs_rq
*cfs_rq
;
5725 unsigned int runnable
;
5727 /* Task has no contribution or is new */
5728 if (cpu_of(rq
) != task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
5729 return cpu_runnable(rq
);
5732 runnable
= READ_ONCE(cfs_rq
->avg
.runnable_avg
);
5734 /* Discount task's runnable from CPU's runnable */
5735 lsub_positive(&runnable
, p
->se
.avg
.runnable_avg
);
5740 static unsigned long capacity_of(int cpu
)
5742 return cpu_rq(cpu
)->cpu_capacity
;
5745 static void record_wakee(struct task_struct
*p
)
5748 * Only decay a single time; tasks that have less then 1 wakeup per
5749 * jiffy will not have built up many flips.
5751 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
5752 current
->wakee_flips
>>= 1;
5753 current
->wakee_flip_decay_ts
= jiffies
;
5756 if (current
->last_wakee
!= p
) {
5757 current
->last_wakee
= p
;
5758 current
->wakee_flips
++;
5763 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5765 * A waker of many should wake a different task than the one last awakened
5766 * at a frequency roughly N times higher than one of its wakees.
5768 * In order to determine whether we should let the load spread vs consolidating
5769 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5770 * partner, and a factor of lls_size higher frequency in the other.
5772 * With both conditions met, we can be relatively sure that the relationship is
5773 * non-monogamous, with partner count exceeding socket size.
5775 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5776 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5779 static int wake_wide(struct task_struct
*p
)
5781 unsigned int master
= current
->wakee_flips
;
5782 unsigned int slave
= p
->wakee_flips
;
5783 int factor
= __this_cpu_read(sd_llc_size
);
5786 swap(master
, slave
);
5787 if (slave
< factor
|| master
< slave
* factor
)
5793 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5794 * soonest. For the purpose of speed we only consider the waking and previous
5797 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5798 * cache-affine and is (or will be) idle.
5800 * wake_affine_weight() - considers the weight to reflect the average
5801 * scheduling latency of the CPUs. This seems to work
5802 * for the overloaded case.
5805 wake_affine_idle(int this_cpu
, int prev_cpu
, int sync
)
5808 * If this_cpu is idle, it implies the wakeup is from interrupt
5809 * context. Only allow the move if cache is shared. Otherwise an
5810 * interrupt intensive workload could force all tasks onto one
5811 * node depending on the IO topology or IRQ affinity settings.
5813 * If the prev_cpu is idle and cache affine then avoid a migration.
5814 * There is no guarantee that the cache hot data from an interrupt
5815 * is more important than cache hot data on the prev_cpu and from
5816 * a cpufreq perspective, it's better to have higher utilisation
5819 if (available_idle_cpu(this_cpu
) && cpus_share_cache(this_cpu
, prev_cpu
))
5820 return available_idle_cpu(prev_cpu
) ? prev_cpu
: this_cpu
;
5822 if (sync
&& cpu_rq(this_cpu
)->nr_running
== 1)
5825 if (available_idle_cpu(prev_cpu
))
5828 return nr_cpumask_bits
;
5832 wake_affine_weight(struct sched_domain
*sd
, struct task_struct
*p
,
5833 int this_cpu
, int prev_cpu
, int sync
)
5835 s64 this_eff_load
, prev_eff_load
;
5836 unsigned long task_load
;
5838 this_eff_load
= cpu_load(cpu_rq(this_cpu
));
5841 unsigned long current_load
= task_h_load(current
);
5843 if (current_load
> this_eff_load
)
5846 this_eff_load
-= current_load
;
5849 task_load
= task_h_load(p
);
5851 this_eff_load
+= task_load
;
5852 if (sched_feat(WA_BIAS
))
5853 this_eff_load
*= 100;
5854 this_eff_load
*= capacity_of(prev_cpu
);
5856 prev_eff_load
= cpu_load(cpu_rq(prev_cpu
));
5857 prev_eff_load
-= task_load
;
5858 if (sched_feat(WA_BIAS
))
5859 prev_eff_load
*= 100 + (sd
->imbalance_pct
- 100) / 2;
5860 prev_eff_load
*= capacity_of(this_cpu
);
5863 * If sync, adjust the weight of prev_eff_load such that if
5864 * prev_eff == this_eff that select_idle_sibling() will consider
5865 * stacking the wakee on top of the waker if no other CPU is
5871 return this_eff_load
< prev_eff_load
? this_cpu
: nr_cpumask_bits
;
5874 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
,
5875 int this_cpu
, int prev_cpu
, int sync
)
5877 int target
= nr_cpumask_bits
;
5879 if (sched_feat(WA_IDLE
))
5880 target
= wake_affine_idle(this_cpu
, prev_cpu
, sync
);
5882 if (sched_feat(WA_WEIGHT
) && target
== nr_cpumask_bits
)
5883 target
= wake_affine_weight(sd
, p
, this_cpu
, prev_cpu
, sync
);
5885 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine_attempts
);
5886 if (target
== nr_cpumask_bits
)
5889 schedstat_inc(sd
->ttwu_move_affine
);
5890 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine
);
5894 static struct sched_group
*
5895 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
);
5898 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5901 find_idlest_group_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
5903 unsigned long load
, min_load
= ULONG_MAX
;
5904 unsigned int min_exit_latency
= UINT_MAX
;
5905 u64 latest_idle_timestamp
= 0;
5906 int least_loaded_cpu
= this_cpu
;
5907 int shallowest_idle_cpu
= -1;
5910 /* Check if we have any choice: */
5911 if (group
->group_weight
== 1)
5912 return cpumask_first(sched_group_span(group
));
5914 /* Traverse only the allowed CPUs */
5915 for_each_cpu_and(i
, sched_group_span(group
), p
->cpus_ptr
) {
5916 if (sched_idle_cpu(i
))
5919 if (available_idle_cpu(i
)) {
5920 struct rq
*rq
= cpu_rq(i
);
5921 struct cpuidle_state
*idle
= idle_get_state(rq
);
5922 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
5924 * We give priority to a CPU whose idle state
5925 * has the smallest exit latency irrespective
5926 * of any idle timestamp.
5928 min_exit_latency
= idle
->exit_latency
;
5929 latest_idle_timestamp
= rq
->idle_stamp
;
5930 shallowest_idle_cpu
= i
;
5931 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
5932 rq
->idle_stamp
> latest_idle_timestamp
) {
5934 * If equal or no active idle state, then
5935 * the most recently idled CPU might have
5938 latest_idle_timestamp
= rq
->idle_stamp
;
5939 shallowest_idle_cpu
= i
;
5941 } else if (shallowest_idle_cpu
== -1) {
5942 load
= cpu_load(cpu_rq(i
));
5943 if (load
< min_load
) {
5945 least_loaded_cpu
= i
;
5950 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
5953 static inline int find_idlest_cpu(struct sched_domain
*sd
, struct task_struct
*p
,
5954 int cpu
, int prev_cpu
, int sd_flag
)
5958 if (!cpumask_intersects(sched_domain_span(sd
), p
->cpus_ptr
))
5962 * We need task's util for cpu_util_without, sync it up to
5963 * prev_cpu's last_update_time.
5965 if (!(sd_flag
& SD_BALANCE_FORK
))
5966 sync_entity_load_avg(&p
->se
);
5969 struct sched_group
*group
;
5970 struct sched_domain
*tmp
;
5973 if (!(sd
->flags
& sd_flag
)) {
5978 group
= find_idlest_group(sd
, p
, cpu
);
5984 new_cpu
= find_idlest_group_cpu(group
, p
, cpu
);
5985 if (new_cpu
== cpu
) {
5986 /* Now try balancing at a lower domain level of 'cpu': */
5991 /* Now try balancing at a lower domain level of 'new_cpu': */
5993 weight
= sd
->span_weight
;
5995 for_each_domain(cpu
, tmp
) {
5996 if (weight
<= tmp
->span_weight
)
5998 if (tmp
->flags
& sd_flag
)
6006 #ifdef CONFIG_SCHED_SMT
6007 DEFINE_STATIC_KEY_FALSE(sched_smt_present
);
6008 EXPORT_SYMBOL_GPL(sched_smt_present
);
6010 static inline void set_idle_cores(int cpu
, int val
)
6012 struct sched_domain_shared
*sds
;
6014 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
6016 WRITE_ONCE(sds
->has_idle_cores
, val
);
6019 static inline bool test_idle_cores(int cpu
, bool def
)
6021 struct sched_domain_shared
*sds
;
6023 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
6025 return READ_ONCE(sds
->has_idle_cores
);
6031 * Scans the local SMT mask to see if the entire core is idle, and records this
6032 * information in sd_llc_shared->has_idle_cores.
6034 * Since SMT siblings share all cache levels, inspecting this limited remote
6035 * state should be fairly cheap.
6037 void __update_idle_core(struct rq
*rq
)
6039 int core
= cpu_of(rq
);
6043 if (test_idle_cores(core
, true))
6046 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
6050 if (!available_idle_cpu(cpu
))
6054 set_idle_cores(core
, 1);
6060 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6061 * there are no idle cores left in the system; tracked through
6062 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6064 static int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6066 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6069 if (!static_branch_likely(&sched_smt_present
))
6072 if (!test_idle_cores(target
, false))
6075 cpumask_and(cpus
, sched_domain_span(sd
), p
->cpus_ptr
);
6077 for_each_cpu_wrap(core
, cpus
, target
) {
6080 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
6081 if (!available_idle_cpu(cpu
)) {
6086 cpumask_andnot(cpus
, cpus
, cpu_smt_mask(core
));
6093 * Failed to find an idle core; stop looking for one.
6095 set_idle_cores(target
, 0);
6101 * Scan the local SMT mask for idle CPUs.
6103 static int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6107 if (!static_branch_likely(&sched_smt_present
))
6110 for_each_cpu(cpu
, cpu_smt_mask(target
)) {
6111 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
) ||
6112 !cpumask_test_cpu(cpu
, sched_domain_span(sd
)))
6114 if (available_idle_cpu(cpu
) || sched_idle_cpu(cpu
))
6121 #else /* CONFIG_SCHED_SMT */
6123 static inline int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6128 static inline int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6133 #endif /* CONFIG_SCHED_SMT */
6136 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6137 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6138 * average idle time for this rq (as found in rq->avg_idle).
6140 static int select_idle_cpu(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6142 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6143 struct sched_domain
*this_sd
;
6144 u64 avg_cost
, avg_idle
;
6146 int this = smp_processor_id();
6147 int cpu
, nr
= INT_MAX
;
6149 this_sd
= rcu_dereference(*this_cpu_ptr(&sd_llc
));
6154 * Due to large variance we need a large fuzz factor; hackbench in
6155 * particularly is sensitive here.
6157 avg_idle
= this_rq()->avg_idle
/ 512;
6158 avg_cost
= this_sd
->avg_scan_cost
+ 1;
6160 if (sched_feat(SIS_AVG_CPU
) && avg_idle
< avg_cost
)
6163 if (sched_feat(SIS_PROP
)) {
6164 u64 span_avg
= sd
->span_weight
* avg_idle
;
6165 if (span_avg
> 4*avg_cost
)
6166 nr
= div_u64(span_avg
, avg_cost
);
6171 time
= cpu_clock(this);
6173 cpumask_and(cpus
, sched_domain_span(sd
), p
->cpus_ptr
);
6175 for_each_cpu_wrap(cpu
, cpus
, target
) {
6178 if (available_idle_cpu(cpu
) || sched_idle_cpu(cpu
))
6182 time
= cpu_clock(this) - time
;
6183 update_avg(&this_sd
->avg_scan_cost
, time
);
6189 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6190 * the task fits. If no CPU is big enough, but there are idle ones, try to
6191 * maximize capacity.
6194 select_idle_capacity(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6196 unsigned long best_cap
= 0;
6197 int cpu
, best_cpu
= -1;
6198 struct cpumask
*cpus
;
6200 sync_entity_load_avg(&p
->se
);
6202 cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6203 cpumask_and(cpus
, sched_domain_span(sd
), p
->cpus_ptr
);
6205 for_each_cpu_wrap(cpu
, cpus
, target
) {
6206 unsigned long cpu_cap
= capacity_of(cpu
);
6208 if (!available_idle_cpu(cpu
) && !sched_idle_cpu(cpu
))
6210 if (task_fits_capacity(p
, cpu_cap
))
6213 if (cpu_cap
> best_cap
) {
6223 * Try and locate an idle core/thread in the LLC cache domain.
6225 static int select_idle_sibling(struct task_struct
*p
, int prev
, int target
)
6227 struct sched_domain
*sd
;
6228 int i
, recent_used_cpu
;
6231 * For asymmetric CPU capacity systems, our domain of interest is
6232 * sd_asym_cpucapacity rather than sd_llc.
6234 if (static_branch_unlikely(&sched_asym_cpucapacity
)) {
6235 sd
= rcu_dereference(per_cpu(sd_asym_cpucapacity
, target
));
6237 * On an asymmetric CPU capacity system where an exclusive
6238 * cpuset defines a symmetric island (i.e. one unique
6239 * capacity_orig value through the cpuset), the key will be set
6240 * but the CPUs within that cpuset will not have a domain with
6241 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6247 i
= select_idle_capacity(p
, sd
, target
);
6248 return ((unsigned)i
< nr_cpumask_bits
) ? i
: target
;
6252 if (available_idle_cpu(target
) || sched_idle_cpu(target
))
6256 * If the previous CPU is cache affine and idle, don't be stupid:
6258 if (prev
!= target
&& cpus_share_cache(prev
, target
) &&
6259 (available_idle_cpu(prev
) || sched_idle_cpu(prev
)))
6263 * Allow a per-cpu kthread to stack with the wakee if the
6264 * kworker thread and the tasks previous CPUs are the same.
6265 * The assumption is that the wakee queued work for the
6266 * per-cpu kthread that is now complete and the wakeup is
6267 * essentially a sync wakeup. An obvious example of this
6268 * pattern is IO completions.
6270 if (is_per_cpu_kthread(current
) &&
6271 prev
== smp_processor_id() &&
6272 this_rq()->nr_running
<= 1) {
6276 /* Check a recently used CPU as a potential idle candidate: */
6277 recent_used_cpu
= p
->recent_used_cpu
;
6278 if (recent_used_cpu
!= prev
&&
6279 recent_used_cpu
!= target
&&
6280 cpus_share_cache(recent_used_cpu
, target
) &&
6281 (available_idle_cpu(recent_used_cpu
) || sched_idle_cpu(recent_used_cpu
)) &&
6282 cpumask_test_cpu(p
->recent_used_cpu
, p
->cpus_ptr
)) {
6284 * Replace recent_used_cpu with prev as it is a potential
6285 * candidate for the next wake:
6287 p
->recent_used_cpu
= prev
;
6288 return recent_used_cpu
;
6291 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
6295 i
= select_idle_core(p
, sd
, target
);
6296 if ((unsigned)i
< nr_cpumask_bits
)
6299 i
= select_idle_cpu(p
, sd
, target
);
6300 if ((unsigned)i
< nr_cpumask_bits
)
6303 i
= select_idle_smt(p
, sd
, target
);
6304 if ((unsigned)i
< nr_cpumask_bits
)
6311 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6312 * @cpu: the CPU to get the utilization of
6314 * The unit of the return value must be the one of capacity so we can compare
6315 * the utilization with the capacity of the CPU that is available for CFS task
6316 * (ie cpu_capacity).
6318 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6319 * recent utilization of currently non-runnable tasks on a CPU. It represents
6320 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6321 * capacity_orig is the cpu_capacity available at the highest frequency
6322 * (arch_scale_freq_capacity()).
6323 * The utilization of a CPU converges towards a sum equal to or less than the
6324 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6325 * the running time on this CPU scaled by capacity_curr.
6327 * The estimated utilization of a CPU is defined to be the maximum between its
6328 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6329 * currently RUNNABLE on that CPU.
6330 * This allows to properly represent the expected utilization of a CPU which
6331 * has just got a big task running since a long sleep period. At the same time
6332 * however it preserves the benefits of the "blocked utilization" in
6333 * describing the potential for other tasks waking up on the same CPU.
6335 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6336 * higher than capacity_orig because of unfortunate rounding in
6337 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6338 * the average stabilizes with the new running time. We need to check that the
6339 * utilization stays within the range of [0..capacity_orig] and cap it if
6340 * necessary. Without utilization capping, a group could be seen as overloaded
6341 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6342 * available capacity. We allow utilization to overshoot capacity_curr (but not
6343 * capacity_orig) as it useful for predicting the capacity required after task
6344 * migrations (scheduler-driven DVFS).
6346 * Return: the (estimated) utilization for the specified CPU
6348 static inline unsigned long cpu_util(int cpu
)
6350 struct cfs_rq
*cfs_rq
;
6353 cfs_rq
= &cpu_rq(cpu
)->cfs
;
6354 util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6356 if (sched_feat(UTIL_EST
))
6357 util
= max(util
, READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
));
6359 return min_t(unsigned long, util
, capacity_orig_of(cpu
));
6363 * cpu_util_without: compute cpu utilization without any contributions from *p
6364 * @cpu: the CPU which utilization is requested
6365 * @p: the task which utilization should be discounted
6367 * The utilization of a CPU is defined by the utilization of tasks currently
6368 * enqueued on that CPU as well as tasks which are currently sleeping after an
6369 * execution on that CPU.
6371 * This method returns the utilization of the specified CPU by discounting the
6372 * utilization of the specified task, whenever the task is currently
6373 * contributing to the CPU utilization.
6375 static unsigned long cpu_util_without(int cpu
, struct task_struct
*p
)
6377 struct cfs_rq
*cfs_rq
;
6380 /* Task has no contribution or is new */
6381 if (cpu
!= task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
6382 return cpu_util(cpu
);
6384 cfs_rq
= &cpu_rq(cpu
)->cfs
;
6385 util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6387 /* Discount task's util from CPU's util */
6388 lsub_positive(&util
, task_util(p
));
6393 * a) if *p is the only task sleeping on this CPU, then:
6394 * cpu_util (== task_util) > util_est (== 0)
6395 * and thus we return:
6396 * cpu_util_without = (cpu_util - task_util) = 0
6398 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6400 * cpu_util >= task_util
6401 * cpu_util > util_est (== 0)
6402 * and thus we discount *p's blocked utilization to return:
6403 * cpu_util_without = (cpu_util - task_util) >= 0
6405 * c) if other tasks are RUNNABLE on that CPU and
6406 * util_est > cpu_util
6407 * then we use util_est since it returns a more restrictive
6408 * estimation of the spare capacity on that CPU, by just
6409 * considering the expected utilization of tasks already
6410 * runnable on that CPU.
6412 * Cases a) and b) are covered by the above code, while case c) is
6413 * covered by the following code when estimated utilization is
6416 if (sched_feat(UTIL_EST
)) {
6417 unsigned int estimated
=
6418 READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
);
6421 * Despite the following checks we still have a small window
6422 * for a possible race, when an execl's select_task_rq_fair()
6423 * races with LB's detach_task():
6426 * p->on_rq = TASK_ON_RQ_MIGRATING;
6427 * ---------------------------------- A
6428 * deactivate_task() \
6429 * dequeue_task() + RaceTime
6430 * util_est_dequeue() /
6431 * ---------------------------------- B
6433 * The additional check on "current == p" it's required to
6434 * properly fix the execl regression and it helps in further
6435 * reducing the chances for the above race.
6437 if (unlikely(task_on_rq_queued(p
) || current
== p
))
6438 lsub_positive(&estimated
, _task_util_est(p
));
6440 util
= max(util
, estimated
);
6444 * Utilization (estimated) can exceed the CPU capacity, thus let's
6445 * clamp to the maximum CPU capacity to ensure consistency with
6446 * the cpu_util call.
6448 return min_t(unsigned long, util
, capacity_orig_of(cpu
));
6452 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6455 static unsigned long cpu_util_next(int cpu
, struct task_struct
*p
, int dst_cpu
)
6457 struct cfs_rq
*cfs_rq
= &cpu_rq(cpu
)->cfs
;
6458 unsigned long util_est
, util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6461 * If @p migrates from @cpu to another, remove its contribution. Or,
6462 * if @p migrates from another CPU to @cpu, add its contribution. In
6463 * the other cases, @cpu is not impacted by the migration, so the
6464 * util_avg should already be correct.
6466 if (task_cpu(p
) == cpu
&& dst_cpu
!= cpu
)
6467 sub_positive(&util
, task_util(p
));
6468 else if (task_cpu(p
) != cpu
&& dst_cpu
== cpu
)
6469 util
+= task_util(p
);
6471 if (sched_feat(UTIL_EST
)) {
6472 util_est
= READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
);
6475 * During wake-up, the task isn't enqueued yet and doesn't
6476 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6477 * so just add it (if needed) to "simulate" what will be
6478 * cpu_util() after the task has been enqueued.
6481 util_est
+= _task_util_est(p
);
6483 util
= max(util
, util_est
);
6486 return min(util
, capacity_orig_of(cpu
));
6490 * compute_energy(): Estimates the energy that @pd would consume if @p was
6491 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6492 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6493 * to compute what would be the energy if we decided to actually migrate that
6497 compute_energy(struct task_struct
*p
, int dst_cpu
, struct perf_domain
*pd
)
6499 struct cpumask
*pd_mask
= perf_domain_span(pd
);
6500 unsigned long cpu_cap
= arch_scale_cpu_capacity(cpumask_first(pd_mask
));
6501 unsigned long max_util
= 0, sum_util
= 0;
6505 * The capacity state of CPUs of the current rd can be driven by CPUs
6506 * of another rd if they belong to the same pd. So, account for the
6507 * utilization of these CPUs too by masking pd with cpu_online_mask
6508 * instead of the rd span.
6510 * If an entire pd is outside of the current rd, it will not appear in
6511 * its pd list and will not be accounted by compute_energy().
6513 for_each_cpu_and(cpu
, pd_mask
, cpu_online_mask
) {
6514 unsigned long cpu_util
, util_cfs
= cpu_util_next(cpu
, p
, dst_cpu
);
6515 struct task_struct
*tsk
= cpu
== dst_cpu
? p
: NULL
;
6518 * Busy time computation: utilization clamping is not
6519 * required since the ratio (sum_util / cpu_capacity)
6520 * is already enough to scale the EM reported power
6521 * consumption at the (eventually clamped) cpu_capacity.
6523 sum_util
+= schedutil_cpu_util(cpu
, util_cfs
, cpu_cap
,
6527 * Performance domain frequency: utilization clamping
6528 * must be considered since it affects the selection
6529 * of the performance domain frequency.
6530 * NOTE: in case RT tasks are running, by default the
6531 * FREQUENCY_UTIL's utilization can be max OPP.
6533 cpu_util
= schedutil_cpu_util(cpu
, util_cfs
, cpu_cap
,
6534 FREQUENCY_UTIL
, tsk
);
6535 max_util
= max(max_util
, cpu_util
);
6538 return em_cpu_energy(pd
->em_pd
, max_util
, sum_util
);
6542 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6543 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6544 * spare capacity in each performance domain and uses it as a potential
6545 * candidate to execute the task. Then, it uses the Energy Model to figure
6546 * out which of the CPU candidates is the most energy-efficient.
6548 * The rationale for this heuristic is as follows. In a performance domain,
6549 * all the most energy efficient CPU candidates (according to the Energy
6550 * Model) are those for which we'll request a low frequency. When there are
6551 * several CPUs for which the frequency request will be the same, we don't
6552 * have enough data to break the tie between them, because the Energy Model
6553 * only includes active power costs. With this model, if we assume that
6554 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6555 * the maximum spare capacity in a performance domain is guaranteed to be among
6556 * the best candidates of the performance domain.
6558 * In practice, it could be preferable from an energy standpoint to pack
6559 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6560 * but that could also hurt our chances to go cluster idle, and we have no
6561 * ways to tell with the current Energy Model if this is actually a good
6562 * idea or not. So, find_energy_efficient_cpu() basically favors
6563 * cluster-packing, and spreading inside a cluster. That should at least be
6564 * a good thing for latency, and this is consistent with the idea that most
6565 * of the energy savings of EAS come from the asymmetry of the system, and
6566 * not so much from breaking the tie between identical CPUs. That's also the
6567 * reason why EAS is enabled in the topology code only for systems where
6568 * SD_ASYM_CPUCAPACITY is set.
6570 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6571 * they don't have any useful utilization data yet and it's not possible to
6572 * forecast their impact on energy consumption. Consequently, they will be
6573 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6574 * to be energy-inefficient in some use-cases. The alternative would be to
6575 * bias new tasks towards specific types of CPUs first, or to try to infer
6576 * their util_avg from the parent task, but those heuristics could hurt
6577 * other use-cases too. So, until someone finds a better way to solve this,
6578 * let's keep things simple by re-using the existing slow path.
6580 static int find_energy_efficient_cpu(struct task_struct
*p
, int prev_cpu
)
6582 unsigned long prev_delta
= ULONG_MAX
, best_delta
= ULONG_MAX
;
6583 struct root_domain
*rd
= cpu_rq(smp_processor_id())->rd
;
6584 unsigned long cpu_cap
, util
, base_energy
= 0;
6585 int cpu
, best_energy_cpu
= prev_cpu
;
6586 struct sched_domain
*sd
;
6587 struct perf_domain
*pd
;
6590 pd
= rcu_dereference(rd
->pd
);
6591 if (!pd
|| READ_ONCE(rd
->overutilized
))
6595 * Energy-aware wake-up happens on the lowest sched_domain starting
6596 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6598 sd
= rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity
));
6599 while (sd
&& !cpumask_test_cpu(prev_cpu
, sched_domain_span(sd
)))
6604 sync_entity_load_avg(&p
->se
);
6605 if (!task_util_est(p
))
6608 for (; pd
; pd
= pd
->next
) {
6609 unsigned long cur_delta
, spare_cap
, max_spare_cap
= 0;
6610 unsigned long base_energy_pd
;
6611 int max_spare_cap_cpu
= -1;
6613 /* Compute the 'base' energy of the pd, without @p */
6614 base_energy_pd
= compute_energy(p
, -1, pd
);
6615 base_energy
+= base_energy_pd
;
6617 for_each_cpu_and(cpu
, perf_domain_span(pd
), sched_domain_span(sd
)) {
6618 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
6621 util
= cpu_util_next(cpu
, p
, cpu
);
6622 cpu_cap
= capacity_of(cpu
);
6623 spare_cap
= cpu_cap
;
6624 lsub_positive(&spare_cap
, util
);
6627 * Skip CPUs that cannot satisfy the capacity request.
6628 * IOW, placing the task there would make the CPU
6629 * overutilized. Take uclamp into account to see how
6630 * much capacity we can get out of the CPU; this is
6631 * aligned with schedutil_cpu_util().
6633 util
= uclamp_rq_util_with(cpu_rq(cpu
), util
, p
);
6634 if (!fits_capacity(util
, cpu_cap
))
6637 /* Always use prev_cpu as a candidate. */
6638 if (cpu
== prev_cpu
) {
6639 prev_delta
= compute_energy(p
, prev_cpu
, pd
);
6640 prev_delta
-= base_energy_pd
;
6641 best_delta
= min(best_delta
, prev_delta
);
6645 * Find the CPU with the maximum spare capacity in
6646 * the performance domain
6648 if (spare_cap
> max_spare_cap
) {
6649 max_spare_cap
= spare_cap
;
6650 max_spare_cap_cpu
= cpu
;
6654 /* Evaluate the energy impact of using this CPU. */
6655 if (max_spare_cap_cpu
>= 0 && max_spare_cap_cpu
!= prev_cpu
) {
6656 cur_delta
= compute_energy(p
, max_spare_cap_cpu
, pd
);
6657 cur_delta
-= base_energy_pd
;
6658 if (cur_delta
< best_delta
) {
6659 best_delta
= cur_delta
;
6660 best_energy_cpu
= max_spare_cap_cpu
;
6668 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6669 * least 6% of the energy used by prev_cpu.
6671 if (prev_delta
== ULONG_MAX
)
6672 return best_energy_cpu
;
6674 if ((prev_delta
- best_delta
) > ((prev_delta
+ base_energy
) >> 4))
6675 return best_energy_cpu
;
6686 * select_task_rq_fair: Select target runqueue for the waking task in domains
6687 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
6688 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6690 * Balances load by selecting the idlest CPU in the idlest group, or under
6691 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6693 * Returns the target CPU number.
6695 * preempt must be disabled.
6698 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int wake_flags
)
6700 int sync
= (wake_flags
& WF_SYNC
) && !(current
->flags
& PF_EXITING
);
6701 struct sched_domain
*tmp
, *sd
= NULL
;
6702 int cpu
= smp_processor_id();
6703 int new_cpu
= prev_cpu
;
6704 int want_affine
= 0;
6705 /* SD_flags and WF_flags share the first nibble */
6706 int sd_flag
= wake_flags
& 0xF;
6708 if (wake_flags
& WF_TTWU
) {
6711 if (sched_energy_enabled()) {
6712 new_cpu
= find_energy_efficient_cpu(p
, prev_cpu
);
6718 want_affine
= !wake_wide(p
) && cpumask_test_cpu(cpu
, p
->cpus_ptr
);
6722 for_each_domain(cpu
, tmp
) {
6724 * If both 'cpu' and 'prev_cpu' are part of this domain,
6725 * cpu is a valid SD_WAKE_AFFINE target.
6727 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
6728 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
6729 if (cpu
!= prev_cpu
)
6730 new_cpu
= wake_affine(tmp
, p
, cpu
, prev_cpu
, sync
);
6732 sd
= NULL
; /* Prefer wake_affine over balance flags */
6736 if (tmp
->flags
& sd_flag
)
6738 else if (!want_affine
)
6744 new_cpu
= find_idlest_cpu(sd
, p
, cpu
, prev_cpu
, sd_flag
);
6745 } else if (wake_flags
& WF_TTWU
) { /* XXX always ? */
6747 new_cpu
= select_idle_sibling(p
, prev_cpu
, new_cpu
);
6750 current
->recent_used_cpu
= cpu
;
6757 static void detach_entity_cfs_rq(struct sched_entity
*se
);
6760 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6761 * cfs_rq_of(p) references at time of call are still valid and identify the
6762 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6764 static void migrate_task_rq_fair(struct task_struct
*p
, int new_cpu
)
6767 * As blocked tasks retain absolute vruntime the migration needs to
6768 * deal with this by subtracting the old and adding the new
6769 * min_vruntime -- the latter is done by enqueue_entity() when placing
6770 * the task on the new runqueue.
6772 if (p
->state
== TASK_WAKING
) {
6773 struct sched_entity
*se
= &p
->se
;
6774 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6777 #ifndef CONFIG_64BIT
6778 u64 min_vruntime_copy
;
6781 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
6783 min_vruntime
= cfs_rq
->min_vruntime
;
6784 } while (min_vruntime
!= min_vruntime_copy
);
6786 min_vruntime
= cfs_rq
->min_vruntime
;
6789 se
->vruntime
-= min_vruntime
;
6792 if (p
->on_rq
== TASK_ON_RQ_MIGRATING
) {
6794 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6795 * rq->lock and can modify state directly.
6797 lockdep_assert_held(&task_rq(p
)->lock
);
6798 detach_entity_cfs_rq(&p
->se
);
6802 * We are supposed to update the task to "current" time, then
6803 * its up to date and ready to go to new CPU/cfs_rq. But we
6804 * have difficulty in getting what current time is, so simply
6805 * throw away the out-of-date time. This will result in the
6806 * wakee task is less decayed, but giving the wakee more load
6809 remove_entity_load_avg(&p
->se
);
6812 /* Tell new CPU we are migrated */
6813 p
->se
.avg
.last_update_time
= 0;
6815 /* We have migrated, no longer consider this task hot */
6816 p
->se
.exec_start
= 0;
6818 update_scan_period(p
, new_cpu
);
6821 static void task_dead_fair(struct task_struct
*p
)
6823 remove_entity_load_avg(&p
->se
);
6827 balance_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
6832 return newidle_balance(rq
, rf
) != 0;
6834 #endif /* CONFIG_SMP */
6836 static unsigned long wakeup_gran(struct sched_entity
*se
)
6838 unsigned long gran
= sysctl_sched_wakeup_granularity
;
6841 * Since its curr running now, convert the gran from real-time
6842 * to virtual-time in his units.
6844 * By using 'se' instead of 'curr' we penalize light tasks, so
6845 * they get preempted easier. That is, if 'se' < 'curr' then
6846 * the resulting gran will be larger, therefore penalizing the
6847 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6848 * be smaller, again penalizing the lighter task.
6850 * This is especially important for buddies when the leftmost
6851 * task is higher priority than the buddy.
6853 return calc_delta_fair(gran
, se
);
6857 * Should 'se' preempt 'curr'.
6871 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
6873 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
6878 gran
= wakeup_gran(se
);
6885 static void set_last_buddy(struct sched_entity
*se
)
6887 if (entity_is_task(se
) && unlikely(task_has_idle_policy(task_of(se
))))
6890 for_each_sched_entity(se
) {
6891 if (SCHED_WARN_ON(!se
->on_rq
))
6893 cfs_rq_of(se
)->last
= se
;
6897 static void set_next_buddy(struct sched_entity
*se
)
6899 if (entity_is_task(se
) && unlikely(task_has_idle_policy(task_of(se
))))
6902 for_each_sched_entity(se
) {
6903 if (SCHED_WARN_ON(!se
->on_rq
))
6905 cfs_rq_of(se
)->next
= se
;
6909 static void set_skip_buddy(struct sched_entity
*se
)
6911 for_each_sched_entity(se
)
6912 cfs_rq_of(se
)->skip
= se
;
6916 * Preempt the current task with a newly woken task if needed:
6918 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
6920 struct task_struct
*curr
= rq
->curr
;
6921 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
6922 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
6923 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
6924 int next_buddy_marked
= 0;
6926 if (unlikely(se
== pse
))
6930 * This is possible from callers such as attach_tasks(), in which we
6931 * unconditionally check_prempt_curr() after an enqueue (which may have
6932 * lead to a throttle). This both saves work and prevents false
6933 * next-buddy nomination below.
6935 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
6938 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
6939 set_next_buddy(pse
);
6940 next_buddy_marked
= 1;
6944 * We can come here with TIF_NEED_RESCHED already set from new task
6947 * Note: this also catches the edge-case of curr being in a throttled
6948 * group (e.g. via set_curr_task), since update_curr() (in the
6949 * enqueue of curr) will have resulted in resched being set. This
6950 * prevents us from potentially nominating it as a false LAST_BUDDY
6953 if (test_tsk_need_resched(curr
))
6956 /* Idle tasks are by definition preempted by non-idle tasks. */
6957 if (unlikely(task_has_idle_policy(curr
)) &&
6958 likely(!task_has_idle_policy(p
)))
6962 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6963 * is driven by the tick):
6965 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
6968 find_matching_se(&se
, &pse
);
6969 update_curr(cfs_rq_of(se
));
6971 if (wakeup_preempt_entity(se
, pse
) == 1) {
6973 * Bias pick_next to pick the sched entity that is
6974 * triggering this preemption.
6976 if (!next_buddy_marked
)
6977 set_next_buddy(pse
);
6986 * Only set the backward buddy when the current task is still
6987 * on the rq. This can happen when a wakeup gets interleaved
6988 * with schedule on the ->pre_schedule() or idle_balance()
6989 * point, either of which can * drop the rq lock.
6991 * Also, during early boot the idle thread is in the fair class,
6992 * for obvious reasons its a bad idea to schedule back to it.
6994 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
6997 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
7001 struct task_struct
*
7002 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
7004 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
7005 struct sched_entity
*se
;
7006 struct task_struct
*p
;
7010 if (!sched_fair_runnable(rq
))
7013 #ifdef CONFIG_FAIR_GROUP_SCHED
7014 if (!prev
|| prev
->sched_class
!= &fair_sched_class
)
7018 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7019 * likely that a next task is from the same cgroup as the current.
7021 * Therefore attempt to avoid putting and setting the entire cgroup
7022 * hierarchy, only change the part that actually changes.
7026 struct sched_entity
*curr
= cfs_rq
->curr
;
7029 * Since we got here without doing put_prev_entity() we also
7030 * have to consider cfs_rq->curr. If it is still a runnable
7031 * entity, update_curr() will update its vruntime, otherwise
7032 * forget we've ever seen it.
7036 update_curr(cfs_rq
);
7041 * This call to check_cfs_rq_runtime() will do the
7042 * throttle and dequeue its entity in the parent(s).
7043 * Therefore the nr_running test will indeed
7046 if (unlikely(check_cfs_rq_runtime(cfs_rq
))) {
7049 if (!cfs_rq
->nr_running
)
7056 se
= pick_next_entity(cfs_rq
, curr
);
7057 cfs_rq
= group_cfs_rq(se
);
7063 * Since we haven't yet done put_prev_entity and if the selected task
7064 * is a different task than we started out with, try and touch the
7065 * least amount of cfs_rqs.
7068 struct sched_entity
*pse
= &prev
->se
;
7070 while (!(cfs_rq
= is_same_group(se
, pse
))) {
7071 int se_depth
= se
->depth
;
7072 int pse_depth
= pse
->depth
;
7074 if (se_depth
<= pse_depth
) {
7075 put_prev_entity(cfs_rq_of(pse
), pse
);
7076 pse
= parent_entity(pse
);
7078 if (se_depth
>= pse_depth
) {
7079 set_next_entity(cfs_rq_of(se
), se
);
7080 se
= parent_entity(se
);
7084 put_prev_entity(cfs_rq
, pse
);
7085 set_next_entity(cfs_rq
, se
);
7092 put_prev_task(rq
, prev
);
7095 se
= pick_next_entity(cfs_rq
, NULL
);
7096 set_next_entity(cfs_rq
, se
);
7097 cfs_rq
= group_cfs_rq(se
);
7102 done
: __maybe_unused
;
7105 * Move the next running task to the front of
7106 * the list, so our cfs_tasks list becomes MRU
7109 list_move(&p
->se
.group_node
, &rq
->cfs_tasks
);
7112 if (hrtick_enabled(rq
))
7113 hrtick_start_fair(rq
, p
);
7115 update_misfit_status(p
, rq
);
7123 new_tasks
= newidle_balance(rq
, rf
);
7126 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7127 * possible for any higher priority task to appear. In that case we
7128 * must re-start the pick_next_entity() loop.
7137 * rq is about to be idle, check if we need to update the
7138 * lost_idle_time of clock_pelt
7140 update_idle_rq_clock_pelt(rq
);
7145 static struct task_struct
*__pick_next_task_fair(struct rq
*rq
)
7147 return pick_next_task_fair(rq
, NULL
, NULL
);
7151 * Account for a descheduled task:
7153 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
7155 struct sched_entity
*se
= &prev
->se
;
7156 struct cfs_rq
*cfs_rq
;
7158 for_each_sched_entity(se
) {
7159 cfs_rq
= cfs_rq_of(se
);
7160 put_prev_entity(cfs_rq
, se
);
7165 * sched_yield() is very simple
7167 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7169 static void yield_task_fair(struct rq
*rq
)
7171 struct task_struct
*curr
= rq
->curr
;
7172 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
7173 struct sched_entity
*se
= &curr
->se
;
7176 * Are we the only task in the tree?
7178 if (unlikely(rq
->nr_running
== 1))
7181 clear_buddies(cfs_rq
, se
);
7183 if (curr
->policy
!= SCHED_BATCH
) {
7184 update_rq_clock(rq
);
7186 * Update run-time statistics of the 'current'.
7188 update_curr(cfs_rq
);
7190 * Tell update_rq_clock() that we've just updated,
7191 * so we don't do microscopic update in schedule()
7192 * and double the fastpath cost.
7194 rq_clock_skip_update(rq
);
7200 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
)
7202 struct sched_entity
*se
= &p
->se
;
7204 /* throttled hierarchies are not runnable */
7205 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
7208 /* Tell the scheduler that we'd really like pse to run next. */
7211 yield_task_fair(rq
);
7217 /**************************************************
7218 * Fair scheduling class load-balancing methods.
7222 * The purpose of load-balancing is to achieve the same basic fairness the
7223 * per-CPU scheduler provides, namely provide a proportional amount of compute
7224 * time to each task. This is expressed in the following equation:
7226 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7228 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7229 * W_i,0 is defined as:
7231 * W_i,0 = \Sum_j w_i,j (2)
7233 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7234 * is derived from the nice value as per sched_prio_to_weight[].
7236 * The weight average is an exponential decay average of the instantaneous
7239 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7241 * C_i is the compute capacity of CPU i, typically it is the
7242 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7243 * can also include other factors [XXX].
7245 * To achieve this balance we define a measure of imbalance which follows
7246 * directly from (1):
7248 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7250 * We them move tasks around to minimize the imbalance. In the continuous
7251 * function space it is obvious this converges, in the discrete case we get
7252 * a few fun cases generally called infeasible weight scenarios.
7255 * - infeasible weights;
7256 * - local vs global optima in the discrete case. ]
7261 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7262 * for all i,j solution, we create a tree of CPUs that follows the hardware
7263 * topology where each level pairs two lower groups (or better). This results
7264 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7265 * tree to only the first of the previous level and we decrease the frequency
7266 * of load-balance at each level inv. proportional to the number of CPUs in
7272 * \Sum { --- * --- * 2^i } = O(n) (5)
7274 * `- size of each group
7275 * | | `- number of CPUs doing load-balance
7277 * `- sum over all levels
7279 * Coupled with a limit on how many tasks we can migrate every balance pass,
7280 * this makes (5) the runtime complexity of the balancer.
7282 * An important property here is that each CPU is still (indirectly) connected
7283 * to every other CPU in at most O(log n) steps:
7285 * The adjacency matrix of the resulting graph is given by:
7288 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7291 * And you'll find that:
7293 * A^(log_2 n)_i,j != 0 for all i,j (7)
7295 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7296 * The task movement gives a factor of O(m), giving a convergence complexity
7299 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7304 * In order to avoid CPUs going idle while there's still work to do, new idle
7305 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7306 * tree itself instead of relying on other CPUs to bring it work.
7308 * This adds some complexity to both (5) and (8) but it reduces the total idle
7316 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7319 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7324 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7326 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7328 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7331 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7332 * rewrite all of this once again.]
7335 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
7337 enum fbq_type
{ regular
, remote
, all
};
7340 * 'group_type' describes the group of CPUs at the moment of load balancing.
7342 * The enum is ordered by pulling priority, with the group with lowest priority
7343 * first so the group_type can simply be compared when selecting the busiest
7344 * group. See update_sd_pick_busiest().
7347 /* The group has spare capacity that can be used to run more tasks. */
7348 group_has_spare
= 0,
7350 * The group is fully used and the tasks don't compete for more CPU
7351 * cycles. Nevertheless, some tasks might wait before running.
7355 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
7356 * and must be migrated to a more powerful CPU.
7360 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7361 * and the task should be migrated to it instead of running on the
7366 * The tasks' affinity constraints previously prevented the scheduler
7367 * from balancing the load across the system.
7371 * The CPU is overloaded and can't provide expected CPU cycles to all
7377 enum migration_type
{
7384 #define LBF_ALL_PINNED 0x01
7385 #define LBF_NEED_BREAK 0x02
7386 #define LBF_DST_PINNED 0x04
7387 #define LBF_SOME_PINNED 0x08
7388 #define LBF_NOHZ_STATS 0x10
7389 #define LBF_NOHZ_AGAIN 0x20
7392 struct sched_domain
*sd
;
7400 struct cpumask
*dst_grpmask
;
7402 enum cpu_idle_type idle
;
7404 /* The set of CPUs under consideration for load-balancing */
7405 struct cpumask
*cpus
;
7410 unsigned int loop_break
;
7411 unsigned int loop_max
;
7413 enum fbq_type fbq_type
;
7414 enum migration_type migration_type
;
7415 struct list_head tasks
;
7419 * Is this task likely cache-hot:
7421 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
7425 lockdep_assert_held(&env
->src_rq
->lock
);
7427 if (p
->sched_class
!= &fair_sched_class
)
7430 if (unlikely(task_has_idle_policy(p
)))
7433 /* SMT siblings share cache */
7434 if (env
->sd
->flags
& SD_SHARE_CPUCAPACITY
)
7438 * Buddy candidates are cache hot:
7440 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
7441 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
7442 &p
->se
== cfs_rq_of(&p
->se
)->last
))
7445 if (sysctl_sched_migration_cost
== -1)
7447 if (sysctl_sched_migration_cost
== 0)
7450 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
7452 return delta
< (s64
)sysctl_sched_migration_cost
;
7455 #ifdef CONFIG_NUMA_BALANCING
7457 * Returns 1, if task migration degrades locality
7458 * Returns 0, if task migration improves locality i.e migration preferred.
7459 * Returns -1, if task migration is not affected by locality.
7461 static int migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
7463 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
7464 unsigned long src_weight
, dst_weight
;
7465 int src_nid
, dst_nid
, dist
;
7467 if (!static_branch_likely(&sched_numa_balancing
))
7470 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
7473 src_nid
= cpu_to_node(env
->src_cpu
);
7474 dst_nid
= cpu_to_node(env
->dst_cpu
);
7476 if (src_nid
== dst_nid
)
7479 /* Migrating away from the preferred node is always bad. */
7480 if (src_nid
== p
->numa_preferred_nid
) {
7481 if (env
->src_rq
->nr_running
> env
->src_rq
->nr_preferred_running
)
7487 /* Encourage migration to the preferred node. */
7488 if (dst_nid
== p
->numa_preferred_nid
)
7491 /* Leaving a core idle is often worse than degrading locality. */
7492 if (env
->idle
== CPU_IDLE
)
7495 dist
= node_distance(src_nid
, dst_nid
);
7497 src_weight
= group_weight(p
, src_nid
, dist
);
7498 dst_weight
= group_weight(p
, dst_nid
, dist
);
7500 src_weight
= task_weight(p
, src_nid
, dist
);
7501 dst_weight
= task_weight(p
, dst_nid
, dist
);
7504 return dst_weight
< src_weight
;
7508 static inline int migrate_degrades_locality(struct task_struct
*p
,
7516 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7519 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
7523 lockdep_assert_held(&env
->src_rq
->lock
);
7526 * We do not migrate tasks that are:
7527 * 1) throttled_lb_pair, or
7528 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7529 * 3) running (obviously), or
7530 * 4) are cache-hot on their current CPU.
7532 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
7535 if (!cpumask_test_cpu(env
->dst_cpu
, p
->cpus_ptr
)) {
7538 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_affine
);
7540 env
->flags
|= LBF_SOME_PINNED
;
7543 * Remember if this task can be migrated to any other CPU in
7544 * our sched_group. We may want to revisit it if we couldn't
7545 * meet load balance goals by pulling other tasks on src_cpu.
7547 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7548 * already computed one in current iteration.
7550 if (env
->idle
== CPU_NEWLY_IDLE
|| (env
->flags
& LBF_DST_PINNED
))
7553 /* Prevent to re-select dst_cpu via env's CPUs: */
7554 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
7555 if (cpumask_test_cpu(cpu
, p
->cpus_ptr
)) {
7556 env
->flags
|= LBF_DST_PINNED
;
7557 env
->new_dst_cpu
= cpu
;
7565 /* Record that we found atleast one task that could run on dst_cpu */
7566 env
->flags
&= ~LBF_ALL_PINNED
;
7568 if (task_running(env
->src_rq
, p
)) {
7569 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_running
);
7574 * Aggressive migration if:
7575 * 1) destination numa is preferred
7576 * 2) task is cache cold, or
7577 * 3) too many balance attempts have failed.
7579 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
7580 if (tsk_cache_hot
== -1)
7581 tsk_cache_hot
= task_hot(p
, env
);
7583 if (tsk_cache_hot
<= 0 ||
7584 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
7585 if (tsk_cache_hot
== 1) {
7586 schedstat_inc(env
->sd
->lb_hot_gained
[env
->idle
]);
7587 schedstat_inc(p
->se
.statistics
.nr_forced_migrations
);
7592 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_hot
);
7597 * detach_task() -- detach the task for the migration specified in env
7599 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
7601 lockdep_assert_held(&env
->src_rq
->lock
);
7603 deactivate_task(env
->src_rq
, p
, DEQUEUE_NOCLOCK
);
7604 set_task_cpu(p
, env
->dst_cpu
);
7608 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7609 * part of active balancing operations within "domain".
7611 * Returns a task if successful and NULL otherwise.
7613 static struct task_struct
*detach_one_task(struct lb_env
*env
)
7615 struct task_struct
*p
;
7617 lockdep_assert_held(&env
->src_rq
->lock
);
7619 list_for_each_entry_reverse(p
,
7620 &env
->src_rq
->cfs_tasks
, se
.group_node
) {
7621 if (!can_migrate_task(p
, env
))
7624 detach_task(p
, env
);
7627 * Right now, this is only the second place where
7628 * lb_gained[env->idle] is updated (other is detach_tasks)
7629 * so we can safely collect stats here rather than
7630 * inside detach_tasks().
7632 schedstat_inc(env
->sd
->lb_gained
[env
->idle
]);
7638 static const unsigned int sched_nr_migrate_break
= 32;
7641 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7642 * busiest_rq, as part of a balancing operation within domain "sd".
7644 * Returns number of detached tasks if successful and 0 otherwise.
7646 static int detach_tasks(struct lb_env
*env
)
7648 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
7649 unsigned long util
, load
;
7650 struct task_struct
*p
;
7653 lockdep_assert_held(&env
->src_rq
->lock
);
7655 if (env
->imbalance
<= 0)
7658 while (!list_empty(tasks
)) {
7660 * We don't want to steal all, otherwise we may be treated likewise,
7661 * which could at worst lead to a livelock crash.
7663 if (env
->idle
!= CPU_NOT_IDLE
&& env
->src_rq
->nr_running
<= 1)
7666 p
= list_last_entry(tasks
, struct task_struct
, se
.group_node
);
7669 /* We've more or less seen every task there is, call it quits */
7670 if (env
->loop
> env
->loop_max
)
7673 /* take a breather every nr_migrate tasks */
7674 if (env
->loop
> env
->loop_break
) {
7675 env
->loop_break
+= sched_nr_migrate_break
;
7676 env
->flags
|= LBF_NEED_BREAK
;
7680 if (!can_migrate_task(p
, env
))
7683 switch (env
->migration_type
) {
7686 * Depending of the number of CPUs and tasks and the
7687 * cgroup hierarchy, task_h_load() can return a null
7688 * value. Make sure that env->imbalance decreases
7689 * otherwise detach_tasks() will stop only after
7690 * detaching up to loop_max tasks.
7692 load
= max_t(unsigned long, task_h_load(p
), 1);
7694 if (sched_feat(LB_MIN
) &&
7695 load
< 16 && !env
->sd
->nr_balance_failed
)
7699 * Make sure that we don't migrate too much load.
7700 * Nevertheless, let relax the constraint if
7701 * scheduler fails to find a good waiting task to
7705 if ((load
>> env
->sd
->nr_balance_failed
) > env
->imbalance
)
7708 env
->imbalance
-= load
;
7712 util
= task_util_est(p
);
7714 if (util
> env
->imbalance
)
7717 env
->imbalance
-= util
;
7724 case migrate_misfit
:
7725 /* This is not a misfit task */
7726 if (task_fits_capacity(p
, capacity_of(env
->src_cpu
)))
7733 detach_task(p
, env
);
7734 list_add(&p
->se
.group_node
, &env
->tasks
);
7738 #ifdef CONFIG_PREEMPTION
7740 * NEWIDLE balancing is a source of latency, so preemptible
7741 * kernels will stop after the first task is detached to minimize
7742 * the critical section.
7744 if (env
->idle
== CPU_NEWLY_IDLE
)
7749 * We only want to steal up to the prescribed amount of
7752 if (env
->imbalance
<= 0)
7757 list_move(&p
->se
.group_node
, tasks
);
7761 * Right now, this is one of only two places we collect this stat
7762 * so we can safely collect detach_one_task() stats here rather
7763 * than inside detach_one_task().
7765 schedstat_add(env
->sd
->lb_gained
[env
->idle
], detached
);
7771 * attach_task() -- attach the task detached by detach_task() to its new rq.
7773 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
7775 lockdep_assert_held(&rq
->lock
);
7777 BUG_ON(task_rq(p
) != rq
);
7778 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
7779 check_preempt_curr(rq
, p
, 0);
7783 * attach_one_task() -- attaches the task returned from detach_one_task() to
7786 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
7791 update_rq_clock(rq
);
7797 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7800 static void attach_tasks(struct lb_env
*env
)
7802 struct list_head
*tasks
= &env
->tasks
;
7803 struct task_struct
*p
;
7806 rq_lock(env
->dst_rq
, &rf
);
7807 update_rq_clock(env
->dst_rq
);
7809 while (!list_empty(tasks
)) {
7810 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
7811 list_del_init(&p
->se
.group_node
);
7813 attach_task(env
->dst_rq
, p
);
7816 rq_unlock(env
->dst_rq
, &rf
);
7819 #ifdef CONFIG_NO_HZ_COMMON
7820 static inline bool cfs_rq_has_blocked(struct cfs_rq
*cfs_rq
)
7822 if (cfs_rq
->avg
.load_avg
)
7825 if (cfs_rq
->avg
.util_avg
)
7831 static inline bool others_have_blocked(struct rq
*rq
)
7833 if (READ_ONCE(rq
->avg_rt
.util_avg
))
7836 if (READ_ONCE(rq
->avg_dl
.util_avg
))
7839 if (thermal_load_avg(rq
))
7842 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7843 if (READ_ONCE(rq
->avg_irq
.util_avg
))
7850 static inline void update_blocked_load_status(struct rq
*rq
, bool has_blocked
)
7852 rq
->last_blocked_load_update_tick
= jiffies
;
7855 rq
->has_blocked_load
= 0;
7858 static inline bool cfs_rq_has_blocked(struct cfs_rq
*cfs_rq
) { return false; }
7859 static inline bool others_have_blocked(struct rq
*rq
) { return false; }
7860 static inline void update_blocked_load_status(struct rq
*rq
, bool has_blocked
) {}
7863 static bool __update_blocked_others(struct rq
*rq
, bool *done
)
7865 const struct sched_class
*curr_class
;
7866 u64 now
= rq_clock_pelt(rq
);
7867 unsigned long thermal_pressure
;
7871 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
7872 * DL and IRQ signals have been updated before updating CFS.
7874 curr_class
= rq
->curr
->sched_class
;
7876 thermal_pressure
= arch_scale_thermal_pressure(cpu_of(rq
));
7878 decayed
= update_rt_rq_load_avg(now
, rq
, curr_class
== &rt_sched_class
) |
7879 update_dl_rq_load_avg(now
, rq
, curr_class
== &dl_sched_class
) |
7880 update_thermal_load_avg(rq_clock_thermal(rq
), rq
, thermal_pressure
) |
7881 update_irq_load_avg(rq
, 0);
7883 if (others_have_blocked(rq
))
7889 #ifdef CONFIG_FAIR_GROUP_SCHED
7891 static inline bool cfs_rq_is_decayed(struct cfs_rq
*cfs_rq
)
7893 if (cfs_rq
->load
.weight
)
7896 if (cfs_rq
->avg
.load_sum
)
7899 if (cfs_rq
->avg
.util_sum
)
7902 if (cfs_rq
->avg
.runnable_sum
)
7908 static bool __update_blocked_fair(struct rq
*rq
, bool *done
)
7910 struct cfs_rq
*cfs_rq
, *pos
;
7911 bool decayed
= false;
7912 int cpu
= cpu_of(rq
);
7915 * Iterates the task_group tree in a bottom up fashion, see
7916 * list_add_leaf_cfs_rq() for details.
7918 for_each_leaf_cfs_rq_safe(rq
, cfs_rq
, pos
) {
7919 struct sched_entity
*se
;
7921 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq
), cfs_rq
)) {
7922 update_tg_load_avg(cfs_rq
);
7924 if (cfs_rq
== &rq
->cfs
)
7928 /* Propagate pending load changes to the parent, if any: */
7929 se
= cfs_rq
->tg
->se
[cpu
];
7930 if (se
&& !skip_blocked_update(se
))
7931 update_load_avg(cfs_rq_of(se
), se
, 0);
7934 * There can be a lot of idle CPU cgroups. Don't let fully
7935 * decayed cfs_rqs linger on the list.
7937 if (cfs_rq_is_decayed(cfs_rq
))
7938 list_del_leaf_cfs_rq(cfs_rq
);
7940 /* Don't need periodic decay once load/util_avg are null */
7941 if (cfs_rq_has_blocked(cfs_rq
))
7949 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7950 * This needs to be done in a top-down fashion because the load of a child
7951 * group is a fraction of its parents load.
7953 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
7955 struct rq
*rq
= rq_of(cfs_rq
);
7956 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
7957 unsigned long now
= jiffies
;
7960 if (cfs_rq
->last_h_load_update
== now
)
7963 WRITE_ONCE(cfs_rq
->h_load_next
, NULL
);
7964 for_each_sched_entity(se
) {
7965 cfs_rq
= cfs_rq_of(se
);
7966 WRITE_ONCE(cfs_rq
->h_load_next
, se
);
7967 if (cfs_rq
->last_h_load_update
== now
)
7972 cfs_rq
->h_load
= cfs_rq_load_avg(cfs_rq
);
7973 cfs_rq
->last_h_load_update
= now
;
7976 while ((se
= READ_ONCE(cfs_rq
->h_load_next
)) != NULL
) {
7977 load
= cfs_rq
->h_load
;
7978 load
= div64_ul(load
* se
->avg
.load_avg
,
7979 cfs_rq_load_avg(cfs_rq
) + 1);
7980 cfs_rq
= group_cfs_rq(se
);
7981 cfs_rq
->h_load
= load
;
7982 cfs_rq
->last_h_load_update
= now
;
7986 static unsigned long task_h_load(struct task_struct
*p
)
7988 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
7990 update_cfs_rq_h_load(cfs_rq
);
7991 return div64_ul(p
->se
.avg
.load_avg
* cfs_rq
->h_load
,
7992 cfs_rq_load_avg(cfs_rq
) + 1);
7995 static bool __update_blocked_fair(struct rq
*rq
, bool *done
)
7997 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
8000 decayed
= update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq
), cfs_rq
);
8001 if (cfs_rq_has_blocked(cfs_rq
))
8007 static unsigned long task_h_load(struct task_struct
*p
)
8009 return p
->se
.avg
.load_avg
;
8013 static void update_blocked_averages(int cpu
)
8015 bool decayed
= false, done
= true;
8016 struct rq
*rq
= cpu_rq(cpu
);
8019 rq_lock_irqsave(rq
, &rf
);
8020 update_rq_clock(rq
);
8022 decayed
|= __update_blocked_others(rq
, &done
);
8023 decayed
|= __update_blocked_fair(rq
, &done
);
8025 update_blocked_load_status(rq
, !done
);
8027 cpufreq_update_util(rq
, 0);
8028 rq_unlock_irqrestore(rq
, &rf
);
8031 /********** Helpers for find_busiest_group ************************/
8034 * sg_lb_stats - stats of a sched_group required for load_balancing
8036 struct sg_lb_stats
{
8037 unsigned long avg_load
; /*Avg load across the CPUs of the group */
8038 unsigned long group_load
; /* Total load over the CPUs of the group */
8039 unsigned long group_capacity
;
8040 unsigned long group_util
; /* Total utilization over the CPUs of the group */
8041 unsigned long group_runnable
; /* Total runnable time over the CPUs of the group */
8042 unsigned int sum_nr_running
; /* Nr of tasks running in the group */
8043 unsigned int sum_h_nr_running
; /* Nr of CFS tasks running in the group */
8044 unsigned int idle_cpus
;
8045 unsigned int group_weight
;
8046 enum group_type group_type
;
8047 unsigned int group_asym_packing
; /* Tasks should be moved to preferred CPU */
8048 unsigned long group_misfit_task_load
; /* A CPU has a task too big for its capacity */
8049 #ifdef CONFIG_NUMA_BALANCING
8050 unsigned int nr_numa_running
;
8051 unsigned int nr_preferred_running
;
8056 * sd_lb_stats - Structure to store the statistics of a sched_domain
8057 * during load balancing.
8059 struct sd_lb_stats
{
8060 struct sched_group
*busiest
; /* Busiest group in this sd */
8061 struct sched_group
*local
; /* Local group in this sd */
8062 unsigned long total_load
; /* Total load of all groups in sd */
8063 unsigned long total_capacity
; /* Total capacity of all groups in sd */
8064 unsigned long avg_load
; /* Average load across all groups in sd */
8065 unsigned int prefer_sibling
; /* tasks should go to sibling first */
8067 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
8068 struct sg_lb_stats local_stat
; /* Statistics of the local group */
8071 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
8074 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8075 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8076 * We must however set busiest_stat::group_type and
8077 * busiest_stat::idle_cpus to the worst busiest group because
8078 * update_sd_pick_busiest() reads these before assignment.
8080 *sds
= (struct sd_lb_stats
){
8084 .total_capacity
= 0UL,
8086 .idle_cpus
= UINT_MAX
,
8087 .group_type
= group_has_spare
,
8092 static unsigned long scale_rt_capacity(int cpu
)
8094 struct rq
*rq
= cpu_rq(cpu
);
8095 unsigned long max
= arch_scale_cpu_capacity(cpu
);
8096 unsigned long used
, free
;
8099 irq
= cpu_util_irq(rq
);
8101 if (unlikely(irq
>= max
))
8105 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8106 * (running and not running) with weights 0 and 1024 respectively.
8107 * avg_thermal.load_avg tracks thermal pressure and the weighted
8108 * average uses the actual delta max capacity(load).
8110 used
= READ_ONCE(rq
->avg_rt
.util_avg
);
8111 used
+= READ_ONCE(rq
->avg_dl
.util_avg
);
8112 used
+= thermal_load_avg(rq
);
8114 if (unlikely(used
>= max
))
8119 return scale_irq_capacity(free
, irq
, max
);
8122 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
8124 unsigned long capacity
= scale_rt_capacity(cpu
);
8125 struct sched_group
*sdg
= sd
->groups
;
8127 cpu_rq(cpu
)->cpu_capacity_orig
= arch_scale_cpu_capacity(cpu
);
8132 cpu_rq(cpu
)->cpu_capacity
= capacity
;
8133 trace_sched_cpu_capacity_tp(cpu_rq(cpu
));
8135 sdg
->sgc
->capacity
= capacity
;
8136 sdg
->sgc
->min_capacity
= capacity
;
8137 sdg
->sgc
->max_capacity
= capacity
;
8140 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
8142 struct sched_domain
*child
= sd
->child
;
8143 struct sched_group
*group
, *sdg
= sd
->groups
;
8144 unsigned long capacity
, min_capacity
, max_capacity
;
8145 unsigned long interval
;
8147 interval
= msecs_to_jiffies(sd
->balance_interval
);
8148 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
8149 sdg
->sgc
->next_update
= jiffies
+ interval
;
8152 update_cpu_capacity(sd
, cpu
);
8157 min_capacity
= ULONG_MAX
;
8160 if (child
->flags
& SD_OVERLAP
) {
8162 * SD_OVERLAP domains cannot assume that child groups
8163 * span the current group.
8166 for_each_cpu(cpu
, sched_group_span(sdg
)) {
8167 unsigned long cpu_cap
= capacity_of(cpu
);
8169 capacity
+= cpu_cap
;
8170 min_capacity
= min(cpu_cap
, min_capacity
);
8171 max_capacity
= max(cpu_cap
, max_capacity
);
8175 * !SD_OVERLAP domains can assume that child groups
8176 * span the current group.
8179 group
= child
->groups
;
8181 struct sched_group_capacity
*sgc
= group
->sgc
;
8183 capacity
+= sgc
->capacity
;
8184 min_capacity
= min(sgc
->min_capacity
, min_capacity
);
8185 max_capacity
= max(sgc
->max_capacity
, max_capacity
);
8186 group
= group
->next
;
8187 } while (group
!= child
->groups
);
8190 sdg
->sgc
->capacity
= capacity
;
8191 sdg
->sgc
->min_capacity
= min_capacity
;
8192 sdg
->sgc
->max_capacity
= max_capacity
;
8196 * Check whether the capacity of the rq has been noticeably reduced by side
8197 * activity. The imbalance_pct is used for the threshold.
8198 * Return true is the capacity is reduced
8201 check_cpu_capacity(struct rq
*rq
, struct sched_domain
*sd
)
8203 return ((rq
->cpu_capacity
* sd
->imbalance_pct
) <
8204 (rq
->cpu_capacity_orig
* 100));
8208 * Check whether a rq has a misfit task and if it looks like we can actually
8209 * help that task: we can migrate the task to a CPU of higher capacity, or
8210 * the task's current CPU is heavily pressured.
8212 static inline int check_misfit_status(struct rq
*rq
, struct sched_domain
*sd
)
8214 return rq
->misfit_task_load
&&
8215 (rq
->cpu_capacity_orig
< rq
->rd
->max_cpu_capacity
||
8216 check_cpu_capacity(rq
, sd
));
8220 * Group imbalance indicates (and tries to solve) the problem where balancing
8221 * groups is inadequate due to ->cpus_ptr constraints.
8223 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8224 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8227 * { 0 1 2 3 } { 4 5 6 7 }
8230 * If we were to balance group-wise we'd place two tasks in the first group and
8231 * two tasks in the second group. Clearly this is undesired as it will overload
8232 * cpu 3 and leave one of the CPUs in the second group unused.
8234 * The current solution to this issue is detecting the skew in the first group
8235 * by noticing the lower domain failed to reach balance and had difficulty
8236 * moving tasks due to affinity constraints.
8238 * When this is so detected; this group becomes a candidate for busiest; see
8239 * update_sd_pick_busiest(). And calculate_imbalance() and
8240 * find_busiest_group() avoid some of the usual balance conditions to allow it
8241 * to create an effective group imbalance.
8243 * This is a somewhat tricky proposition since the next run might not find the
8244 * group imbalance and decide the groups need to be balanced again. A most
8245 * subtle and fragile situation.
8248 static inline int sg_imbalanced(struct sched_group
*group
)
8250 return group
->sgc
->imbalance
;
8254 * group_has_capacity returns true if the group has spare capacity that could
8255 * be used by some tasks.
8256 * We consider that a group has spare capacity if the * number of task is
8257 * smaller than the number of CPUs or if the utilization is lower than the
8258 * available capacity for CFS tasks.
8259 * For the latter, we use a threshold to stabilize the state, to take into
8260 * account the variance of the tasks' load and to return true if the available
8261 * capacity in meaningful for the load balancer.
8262 * As an example, an available capacity of 1% can appear but it doesn't make
8263 * any benefit for the load balance.
8266 group_has_capacity(unsigned int imbalance_pct
, struct sg_lb_stats
*sgs
)
8268 if (sgs
->sum_nr_running
< sgs
->group_weight
)
8271 if ((sgs
->group_capacity
* imbalance_pct
) <
8272 (sgs
->group_runnable
* 100))
8275 if ((sgs
->group_capacity
* 100) >
8276 (sgs
->group_util
* imbalance_pct
))
8283 * group_is_overloaded returns true if the group has more tasks than it can
8285 * group_is_overloaded is not equals to !group_has_capacity because a group
8286 * with the exact right number of tasks, has no more spare capacity but is not
8287 * overloaded so both group_has_capacity and group_is_overloaded return
8291 group_is_overloaded(unsigned int imbalance_pct
, struct sg_lb_stats
*sgs
)
8293 if (sgs
->sum_nr_running
<= sgs
->group_weight
)
8296 if ((sgs
->group_capacity
* 100) <
8297 (sgs
->group_util
* imbalance_pct
))
8300 if ((sgs
->group_capacity
* imbalance_pct
) <
8301 (sgs
->group_runnable
* 100))
8308 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
8309 * per-CPU capacity than sched_group ref.
8312 group_smaller_min_cpu_capacity(struct sched_group
*sg
, struct sched_group
*ref
)
8314 return fits_capacity(sg
->sgc
->min_capacity
, ref
->sgc
->min_capacity
);
8318 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
8319 * per-CPU capacity_orig than sched_group ref.
8322 group_smaller_max_cpu_capacity(struct sched_group
*sg
, struct sched_group
*ref
)
8324 return fits_capacity(sg
->sgc
->max_capacity
, ref
->sgc
->max_capacity
);
8328 group_type
group_classify(unsigned int imbalance_pct
,
8329 struct sched_group
*group
,
8330 struct sg_lb_stats
*sgs
)
8332 if (group_is_overloaded(imbalance_pct
, sgs
))
8333 return group_overloaded
;
8335 if (sg_imbalanced(group
))
8336 return group_imbalanced
;
8338 if (sgs
->group_asym_packing
)
8339 return group_asym_packing
;
8341 if (sgs
->group_misfit_task_load
)
8342 return group_misfit_task
;
8344 if (!group_has_capacity(imbalance_pct
, sgs
))
8345 return group_fully_busy
;
8347 return group_has_spare
;
8350 static bool update_nohz_stats(struct rq
*rq
, bool force
)
8352 #ifdef CONFIG_NO_HZ_COMMON
8353 unsigned int cpu
= rq
->cpu
;
8355 if (!rq
->has_blocked_load
)
8358 if (!cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))
8361 if (!force
&& !time_after(jiffies
, rq
->last_blocked_load_update_tick
))
8364 update_blocked_averages(cpu
);
8366 return rq
->has_blocked_load
;
8373 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8374 * @env: The load balancing environment.
8375 * @group: sched_group whose statistics are to be updated.
8376 * @sgs: variable to hold the statistics for this group.
8377 * @sg_status: Holds flag indicating the status of the sched_group
8379 static inline void update_sg_lb_stats(struct lb_env
*env
,
8380 struct sched_group
*group
,
8381 struct sg_lb_stats
*sgs
,
8384 int i
, nr_running
, local_group
;
8386 memset(sgs
, 0, sizeof(*sgs
));
8388 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_span(group
));
8390 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
8391 struct rq
*rq
= cpu_rq(i
);
8393 if ((env
->flags
& LBF_NOHZ_STATS
) && update_nohz_stats(rq
, false))
8394 env
->flags
|= LBF_NOHZ_AGAIN
;
8396 sgs
->group_load
+= cpu_load(rq
);
8397 sgs
->group_util
+= cpu_util(i
);
8398 sgs
->group_runnable
+= cpu_runnable(rq
);
8399 sgs
->sum_h_nr_running
+= rq
->cfs
.h_nr_running
;
8401 nr_running
= rq
->nr_running
;
8402 sgs
->sum_nr_running
+= nr_running
;
8405 *sg_status
|= SG_OVERLOAD
;
8407 if (cpu_overutilized(i
))
8408 *sg_status
|= SG_OVERUTILIZED
;
8410 #ifdef CONFIG_NUMA_BALANCING
8411 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
8412 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
8415 * No need to call idle_cpu() if nr_running is not 0
8417 if (!nr_running
&& idle_cpu(i
)) {
8419 /* Idle cpu can't have misfit task */
8426 /* Check for a misfit task on the cpu */
8427 if (env
->sd
->flags
& SD_ASYM_CPUCAPACITY
&&
8428 sgs
->group_misfit_task_load
< rq
->misfit_task_load
) {
8429 sgs
->group_misfit_task_load
= rq
->misfit_task_load
;
8430 *sg_status
|= SG_OVERLOAD
;
8434 /* Check if dst CPU is idle and preferred to this group */
8435 if (env
->sd
->flags
& SD_ASYM_PACKING
&&
8436 env
->idle
!= CPU_NOT_IDLE
&&
8437 sgs
->sum_h_nr_running
&&
8438 sched_asym_prefer(env
->dst_cpu
, group
->asym_prefer_cpu
)) {
8439 sgs
->group_asym_packing
= 1;
8442 sgs
->group_capacity
= group
->sgc
->capacity
;
8444 sgs
->group_weight
= group
->group_weight
;
8446 sgs
->group_type
= group_classify(env
->sd
->imbalance_pct
, group
, sgs
);
8448 /* Computing avg_load makes sense only when group is overloaded */
8449 if (sgs
->group_type
== group_overloaded
)
8450 sgs
->avg_load
= (sgs
->group_load
* SCHED_CAPACITY_SCALE
) /
8451 sgs
->group_capacity
;
8455 * update_sd_pick_busiest - return 1 on busiest group
8456 * @env: The load balancing environment.
8457 * @sds: sched_domain statistics
8458 * @sg: sched_group candidate to be checked for being the busiest
8459 * @sgs: sched_group statistics
8461 * Determine if @sg is a busier group than the previously selected
8464 * Return: %true if @sg is a busier group than the previously selected
8465 * busiest group. %false otherwise.
8467 static bool update_sd_pick_busiest(struct lb_env
*env
,
8468 struct sd_lb_stats
*sds
,
8469 struct sched_group
*sg
,
8470 struct sg_lb_stats
*sgs
)
8472 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
8474 /* Make sure that there is at least one task to pull */
8475 if (!sgs
->sum_h_nr_running
)
8479 * Don't try to pull misfit tasks we can't help.
8480 * We can use max_capacity here as reduction in capacity on some
8481 * CPUs in the group should either be possible to resolve
8482 * internally or be covered by avg_load imbalance (eventually).
8484 if (sgs
->group_type
== group_misfit_task
&&
8485 (!group_smaller_max_cpu_capacity(sg
, sds
->local
) ||
8486 sds
->local_stat
.group_type
!= group_has_spare
))
8489 if (sgs
->group_type
> busiest
->group_type
)
8492 if (sgs
->group_type
< busiest
->group_type
)
8496 * The candidate and the current busiest group are the same type of
8497 * group. Let check which one is the busiest according to the type.
8500 switch (sgs
->group_type
) {
8501 case group_overloaded
:
8502 /* Select the overloaded group with highest avg_load. */
8503 if (sgs
->avg_load
<= busiest
->avg_load
)
8507 case group_imbalanced
:
8509 * Select the 1st imbalanced group as we don't have any way to
8510 * choose one more than another.
8514 case group_asym_packing
:
8515 /* Prefer to move from lowest priority CPU's work */
8516 if (sched_asym_prefer(sg
->asym_prefer_cpu
, sds
->busiest
->asym_prefer_cpu
))
8520 case group_misfit_task
:
8522 * If we have more than one misfit sg go with the biggest
8525 if (sgs
->group_misfit_task_load
< busiest
->group_misfit_task_load
)
8529 case group_fully_busy
:
8531 * Select the fully busy group with highest avg_load. In
8532 * theory, there is no need to pull task from such kind of
8533 * group because tasks have all compute capacity that they need
8534 * but we can still improve the overall throughput by reducing
8535 * contention when accessing shared HW resources.
8537 * XXX for now avg_load is not computed and always 0 so we
8538 * select the 1st one.
8540 if (sgs
->avg_load
<= busiest
->avg_load
)
8544 case group_has_spare
:
8546 * Select not overloaded group with lowest number of idle cpus
8547 * and highest number of running tasks. We could also compare
8548 * the spare capacity which is more stable but it can end up
8549 * that the group has less spare capacity but finally more idle
8550 * CPUs which means less opportunity to pull tasks.
8552 if (sgs
->idle_cpus
> busiest
->idle_cpus
)
8554 else if ((sgs
->idle_cpus
== busiest
->idle_cpus
) &&
8555 (sgs
->sum_nr_running
<= busiest
->sum_nr_running
))
8562 * Candidate sg has no more than one task per CPU and has higher
8563 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8564 * throughput. Maximize throughput, power/energy consequences are not
8567 if ((env
->sd
->flags
& SD_ASYM_CPUCAPACITY
) &&
8568 (sgs
->group_type
<= group_fully_busy
) &&
8569 (group_smaller_min_cpu_capacity(sds
->local
, sg
)))
8575 #ifdef CONFIG_NUMA_BALANCING
8576 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
8578 if (sgs
->sum_h_nr_running
> sgs
->nr_numa_running
)
8580 if (sgs
->sum_h_nr_running
> sgs
->nr_preferred_running
)
8585 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
8587 if (rq
->nr_running
> rq
->nr_numa_running
)
8589 if (rq
->nr_running
> rq
->nr_preferred_running
)
8594 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
8599 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
8603 #endif /* CONFIG_NUMA_BALANCING */
8609 * task_running_on_cpu - return 1 if @p is running on @cpu.
8612 static unsigned int task_running_on_cpu(int cpu
, struct task_struct
*p
)
8614 /* Task has no contribution or is new */
8615 if (cpu
!= task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
8618 if (task_on_rq_queued(p
))
8625 * idle_cpu_without - would a given CPU be idle without p ?
8626 * @cpu: the processor on which idleness is tested.
8627 * @p: task which should be ignored.
8629 * Return: 1 if the CPU would be idle. 0 otherwise.
8631 static int idle_cpu_without(int cpu
, struct task_struct
*p
)
8633 struct rq
*rq
= cpu_rq(cpu
);
8635 if (rq
->curr
!= rq
->idle
&& rq
->curr
!= p
)
8639 * rq->nr_running can't be used but an updated version without the
8640 * impact of p on cpu must be used instead. The updated nr_running
8641 * be computed and tested before calling idle_cpu_without().
8645 if (rq
->ttwu_pending
)
8653 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8654 * @sd: The sched_domain level to look for idlest group.
8655 * @group: sched_group whose statistics are to be updated.
8656 * @sgs: variable to hold the statistics for this group.
8657 * @p: The task for which we look for the idlest group/CPU.
8659 static inline void update_sg_wakeup_stats(struct sched_domain
*sd
,
8660 struct sched_group
*group
,
8661 struct sg_lb_stats
*sgs
,
8662 struct task_struct
*p
)
8666 memset(sgs
, 0, sizeof(*sgs
));
8668 for_each_cpu(i
, sched_group_span(group
)) {
8669 struct rq
*rq
= cpu_rq(i
);
8672 sgs
->group_load
+= cpu_load_without(rq
, p
);
8673 sgs
->group_util
+= cpu_util_without(i
, p
);
8674 sgs
->group_runnable
+= cpu_runnable_without(rq
, p
);
8675 local
= task_running_on_cpu(i
, p
);
8676 sgs
->sum_h_nr_running
+= rq
->cfs
.h_nr_running
- local
;
8678 nr_running
= rq
->nr_running
- local
;
8679 sgs
->sum_nr_running
+= nr_running
;
8682 * No need to call idle_cpu_without() if nr_running is not 0
8684 if (!nr_running
&& idle_cpu_without(i
, p
))
8689 /* Check if task fits in the group */
8690 if (sd
->flags
& SD_ASYM_CPUCAPACITY
&&
8691 !task_fits_capacity(p
, group
->sgc
->max_capacity
)) {
8692 sgs
->group_misfit_task_load
= 1;
8695 sgs
->group_capacity
= group
->sgc
->capacity
;
8697 sgs
->group_weight
= group
->group_weight
;
8699 sgs
->group_type
= group_classify(sd
->imbalance_pct
, group
, sgs
);
8702 * Computing avg_load makes sense only when group is fully busy or
8705 if (sgs
->group_type
== group_fully_busy
||
8706 sgs
->group_type
== group_overloaded
)
8707 sgs
->avg_load
= (sgs
->group_load
* SCHED_CAPACITY_SCALE
) /
8708 sgs
->group_capacity
;
8711 static bool update_pick_idlest(struct sched_group
*idlest
,
8712 struct sg_lb_stats
*idlest_sgs
,
8713 struct sched_group
*group
,
8714 struct sg_lb_stats
*sgs
)
8716 if (sgs
->group_type
< idlest_sgs
->group_type
)
8719 if (sgs
->group_type
> idlest_sgs
->group_type
)
8723 * The candidate and the current idlest group are the same type of
8724 * group. Let check which one is the idlest according to the type.
8727 switch (sgs
->group_type
) {
8728 case group_overloaded
:
8729 case group_fully_busy
:
8730 /* Select the group with lowest avg_load. */
8731 if (idlest_sgs
->avg_load
<= sgs
->avg_load
)
8735 case group_imbalanced
:
8736 case group_asym_packing
:
8737 /* Those types are not used in the slow wakeup path */
8740 case group_misfit_task
:
8741 /* Select group with the highest max capacity */
8742 if (idlest
->sgc
->max_capacity
>= group
->sgc
->max_capacity
)
8746 case group_has_spare
:
8747 /* Select group with most idle CPUs */
8748 if (idlest_sgs
->idle_cpus
> sgs
->idle_cpus
)
8751 /* Select group with lowest group_util */
8752 if (idlest_sgs
->idle_cpus
== sgs
->idle_cpus
&&
8753 idlest_sgs
->group_util
<= sgs
->group_util
)
8763 * find_idlest_group() finds and returns the least busy CPU group within the
8766 * Assumes p is allowed on at least one CPU in sd.
8768 static struct sched_group
*
8769 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
8771 struct sched_group
*idlest
= NULL
, *local
= NULL
, *group
= sd
->groups
;
8772 struct sg_lb_stats local_sgs
, tmp_sgs
;
8773 struct sg_lb_stats
*sgs
;
8774 unsigned long imbalance
;
8775 struct sg_lb_stats idlest_sgs
= {
8776 .avg_load
= UINT_MAX
,
8777 .group_type
= group_overloaded
,
8780 imbalance
= scale_load_down(NICE_0_LOAD
) *
8781 (sd
->imbalance_pct
-100) / 100;
8786 /* Skip over this group if it has no CPUs allowed */
8787 if (!cpumask_intersects(sched_group_span(group
),
8791 local_group
= cpumask_test_cpu(this_cpu
,
8792 sched_group_span(group
));
8801 update_sg_wakeup_stats(sd
, group
, sgs
, p
);
8803 if (!local_group
&& update_pick_idlest(idlest
, &idlest_sgs
, group
, sgs
)) {
8808 } while (group
= group
->next
, group
!= sd
->groups
);
8811 /* There is no idlest group to push tasks to */
8815 /* The local group has been skipped because of CPU affinity */
8820 * If the local group is idler than the selected idlest group
8821 * don't try and push the task.
8823 if (local_sgs
.group_type
< idlest_sgs
.group_type
)
8827 * If the local group is busier than the selected idlest group
8828 * try and push the task.
8830 if (local_sgs
.group_type
> idlest_sgs
.group_type
)
8833 switch (local_sgs
.group_type
) {
8834 case group_overloaded
:
8835 case group_fully_busy
:
8837 * When comparing groups across NUMA domains, it's possible for
8838 * the local domain to be very lightly loaded relative to the
8839 * remote domains but "imbalance" skews the comparison making
8840 * remote CPUs look much more favourable. When considering
8841 * cross-domain, add imbalance to the load on the remote node
8842 * and consider staying local.
8845 if ((sd
->flags
& SD_NUMA
) &&
8846 ((idlest_sgs
.avg_load
+ imbalance
) >= local_sgs
.avg_load
))
8850 * If the local group is less loaded than the selected
8851 * idlest group don't try and push any tasks.
8853 if (idlest_sgs
.avg_load
>= (local_sgs
.avg_load
+ imbalance
))
8856 if (100 * local_sgs
.avg_load
<= sd
->imbalance_pct
* idlest_sgs
.avg_load
)
8860 case group_imbalanced
:
8861 case group_asym_packing
:
8862 /* Those type are not used in the slow wakeup path */
8865 case group_misfit_task
:
8866 /* Select group with the highest max capacity */
8867 if (local
->sgc
->max_capacity
>= idlest
->sgc
->max_capacity
)
8871 case group_has_spare
:
8872 if (sd
->flags
& SD_NUMA
) {
8873 #ifdef CONFIG_NUMA_BALANCING
8876 * If there is spare capacity at NUMA, try to select
8877 * the preferred node
8879 if (cpu_to_node(this_cpu
) == p
->numa_preferred_nid
)
8882 idlest_cpu
= cpumask_first(sched_group_span(idlest
));
8883 if (cpu_to_node(idlest_cpu
) == p
->numa_preferred_nid
)
8887 * Otherwise, keep the task on this node to stay close
8888 * its wakeup source and improve locality. If there is
8889 * a real need of migration, periodic load balance will
8892 if (local_sgs
.idle_cpus
)
8897 * Select group with highest number of idle CPUs. We could also
8898 * compare the utilization which is more stable but it can end
8899 * up that the group has less spare capacity but finally more
8900 * idle CPUs which means more opportunity to run task.
8902 if (local_sgs
.idle_cpus
>= idlest_sgs
.idle_cpus
)
8911 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8912 * @env: The load balancing environment.
8913 * @sds: variable to hold the statistics for this sched_domain.
8916 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
8918 struct sched_domain
*child
= env
->sd
->child
;
8919 struct sched_group
*sg
= env
->sd
->groups
;
8920 struct sg_lb_stats
*local
= &sds
->local_stat
;
8921 struct sg_lb_stats tmp_sgs
;
8924 #ifdef CONFIG_NO_HZ_COMMON
8925 if (env
->idle
== CPU_NEWLY_IDLE
&& READ_ONCE(nohz
.has_blocked
))
8926 env
->flags
|= LBF_NOHZ_STATS
;
8930 struct sg_lb_stats
*sgs
= &tmp_sgs
;
8933 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_span(sg
));
8938 if (env
->idle
!= CPU_NEWLY_IDLE
||
8939 time_after_eq(jiffies
, sg
->sgc
->next_update
))
8940 update_group_capacity(env
->sd
, env
->dst_cpu
);
8943 update_sg_lb_stats(env
, sg
, sgs
, &sg_status
);
8949 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
8951 sds
->busiest_stat
= *sgs
;
8955 /* Now, start updating sd_lb_stats */
8956 sds
->total_load
+= sgs
->group_load
;
8957 sds
->total_capacity
+= sgs
->group_capacity
;
8960 } while (sg
!= env
->sd
->groups
);
8962 /* Tag domain that child domain prefers tasks go to siblings first */
8963 sds
->prefer_sibling
= child
&& child
->flags
& SD_PREFER_SIBLING
;
8965 #ifdef CONFIG_NO_HZ_COMMON
8966 if ((env
->flags
& LBF_NOHZ_AGAIN
) &&
8967 cpumask_subset(nohz
.idle_cpus_mask
, sched_domain_span(env
->sd
))) {
8969 WRITE_ONCE(nohz
.next_blocked
,
8970 jiffies
+ msecs_to_jiffies(LOAD_AVG_PERIOD
));
8974 if (env
->sd
->flags
& SD_NUMA
)
8975 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
8977 if (!env
->sd
->parent
) {
8978 struct root_domain
*rd
= env
->dst_rq
->rd
;
8980 /* update overload indicator if we are at root domain */
8981 WRITE_ONCE(rd
->overload
, sg_status
& SG_OVERLOAD
);
8983 /* Update over-utilization (tipping point, U >= 0) indicator */
8984 WRITE_ONCE(rd
->overutilized
, sg_status
& SG_OVERUTILIZED
);
8985 trace_sched_overutilized_tp(rd
, sg_status
& SG_OVERUTILIZED
);
8986 } else if (sg_status
& SG_OVERUTILIZED
) {
8987 struct root_domain
*rd
= env
->dst_rq
->rd
;
8989 WRITE_ONCE(rd
->overutilized
, SG_OVERUTILIZED
);
8990 trace_sched_overutilized_tp(rd
, SG_OVERUTILIZED
);
8994 static inline long adjust_numa_imbalance(int imbalance
, int nr_running
)
8996 unsigned int imbalance_min
;
8999 * Allow a small imbalance based on a simple pair of communicating
9000 * tasks that remain local when the source domain is almost idle.
9003 if (nr_running
<= imbalance_min
)
9010 * calculate_imbalance - Calculate the amount of imbalance present within the
9011 * groups of a given sched_domain during load balance.
9012 * @env: load balance environment
9013 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
9015 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
9017 struct sg_lb_stats
*local
, *busiest
;
9019 local
= &sds
->local_stat
;
9020 busiest
= &sds
->busiest_stat
;
9022 if (busiest
->group_type
== group_misfit_task
) {
9023 /* Set imbalance to allow misfit tasks to be balanced. */
9024 env
->migration_type
= migrate_misfit
;
9029 if (busiest
->group_type
== group_asym_packing
) {
9031 * In case of asym capacity, we will try to migrate all load to
9032 * the preferred CPU.
9034 env
->migration_type
= migrate_task
;
9035 env
->imbalance
= busiest
->sum_h_nr_running
;
9039 if (busiest
->group_type
== group_imbalanced
) {
9041 * In the group_imb case we cannot rely on group-wide averages
9042 * to ensure CPU-load equilibrium, try to move any task to fix
9043 * the imbalance. The next load balance will take care of
9044 * balancing back the system.
9046 env
->migration_type
= migrate_task
;
9052 * Try to use spare capacity of local group without overloading it or
9055 if (local
->group_type
== group_has_spare
) {
9056 if (busiest
->group_type
> group_fully_busy
) {
9058 * If busiest is overloaded, try to fill spare
9059 * capacity. This might end up creating spare capacity
9060 * in busiest or busiest still being overloaded but
9061 * there is no simple way to directly compute the
9062 * amount of load to migrate in order to balance the
9065 env
->migration_type
= migrate_util
;
9066 env
->imbalance
= max(local
->group_capacity
, local
->group_util
) -
9070 * In some cases, the group's utilization is max or even
9071 * higher than capacity because of migrations but the
9072 * local CPU is (newly) idle. There is at least one
9073 * waiting task in this overloaded busiest group. Let's
9076 if (env
->idle
!= CPU_NOT_IDLE
&& env
->imbalance
== 0) {
9077 env
->migration_type
= migrate_task
;
9084 if (busiest
->group_weight
== 1 || sds
->prefer_sibling
) {
9085 unsigned int nr_diff
= busiest
->sum_nr_running
;
9087 * When prefer sibling, evenly spread running tasks on
9090 env
->migration_type
= migrate_task
;
9091 lsub_positive(&nr_diff
, local
->sum_nr_running
);
9092 env
->imbalance
= nr_diff
>> 1;
9096 * If there is no overload, we just want to even the number of
9099 env
->migration_type
= migrate_task
;
9100 env
->imbalance
= max_t(long, 0, (local
->idle_cpus
-
9101 busiest
->idle_cpus
) >> 1);
9104 /* Consider allowing a small imbalance between NUMA groups */
9105 if (env
->sd
->flags
& SD_NUMA
)
9106 env
->imbalance
= adjust_numa_imbalance(env
->imbalance
,
9107 busiest
->sum_nr_running
);
9113 * Local is fully busy but has to take more load to relieve the
9116 if (local
->group_type
< group_overloaded
) {
9118 * Local will become overloaded so the avg_load metrics are
9122 local
->avg_load
= (local
->group_load
* SCHED_CAPACITY_SCALE
) /
9123 local
->group_capacity
;
9125 sds
->avg_load
= (sds
->total_load
* SCHED_CAPACITY_SCALE
) /
9126 sds
->total_capacity
;
9128 * If the local group is more loaded than the selected
9129 * busiest group don't try to pull any tasks.
9131 if (local
->avg_load
>= busiest
->avg_load
) {
9138 * Both group are or will become overloaded and we're trying to get all
9139 * the CPUs to the average_load, so we don't want to push ourselves
9140 * above the average load, nor do we wish to reduce the max loaded CPU
9141 * below the average load. At the same time, we also don't want to
9142 * reduce the group load below the group capacity. Thus we look for
9143 * the minimum possible imbalance.
9145 env
->migration_type
= migrate_load
;
9146 env
->imbalance
= min(
9147 (busiest
->avg_load
- sds
->avg_load
) * busiest
->group_capacity
,
9148 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
9149 ) / SCHED_CAPACITY_SCALE
;
9152 /******* find_busiest_group() helpers end here *********************/
9155 * Decision matrix according to the local and busiest group type:
9157 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9158 * has_spare nr_idle balanced N/A N/A balanced balanced
9159 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9160 * misfit_task force N/A N/A N/A force force
9161 * asym_packing force force N/A N/A force force
9162 * imbalanced force force N/A N/A force force
9163 * overloaded force force N/A N/A force avg_load
9165 * N/A : Not Applicable because already filtered while updating
9167 * balanced : The system is balanced for these 2 groups.
9168 * force : Calculate the imbalance as load migration is probably needed.
9169 * avg_load : Only if imbalance is significant enough.
9170 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9171 * different in groups.
9175 * find_busiest_group - Returns the busiest group within the sched_domain
9176 * if there is an imbalance.
9178 * Also calculates the amount of runnable load which should be moved
9179 * to restore balance.
9181 * @env: The load balancing environment.
9183 * Return: - The busiest group if imbalance exists.
9185 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
9187 struct sg_lb_stats
*local
, *busiest
;
9188 struct sd_lb_stats sds
;
9190 init_sd_lb_stats(&sds
);
9193 * Compute the various statistics relevant for load balancing at
9196 update_sd_lb_stats(env
, &sds
);
9198 if (sched_energy_enabled()) {
9199 struct root_domain
*rd
= env
->dst_rq
->rd
;
9201 if (rcu_dereference(rd
->pd
) && !READ_ONCE(rd
->overutilized
))
9205 local
= &sds
.local_stat
;
9206 busiest
= &sds
.busiest_stat
;
9208 /* There is no busy sibling group to pull tasks from */
9212 /* Misfit tasks should be dealt with regardless of the avg load */
9213 if (busiest
->group_type
== group_misfit_task
)
9216 /* ASYM feature bypasses nice load balance check */
9217 if (busiest
->group_type
== group_asym_packing
)
9221 * If the busiest group is imbalanced the below checks don't
9222 * work because they assume all things are equal, which typically
9223 * isn't true due to cpus_ptr constraints and the like.
9225 if (busiest
->group_type
== group_imbalanced
)
9229 * If the local group is busier than the selected busiest group
9230 * don't try and pull any tasks.
9232 if (local
->group_type
> busiest
->group_type
)
9236 * When groups are overloaded, use the avg_load to ensure fairness
9239 if (local
->group_type
== group_overloaded
) {
9241 * If the local group is more loaded than the selected
9242 * busiest group don't try to pull any tasks.
9244 if (local
->avg_load
>= busiest
->avg_load
)
9247 /* XXX broken for overlapping NUMA groups */
9248 sds
.avg_load
= (sds
.total_load
* SCHED_CAPACITY_SCALE
) /
9252 * Don't pull any tasks if this group is already above the
9253 * domain average load.
9255 if (local
->avg_load
>= sds
.avg_load
)
9259 * If the busiest group is more loaded, use imbalance_pct to be
9262 if (100 * busiest
->avg_load
<=
9263 env
->sd
->imbalance_pct
* local
->avg_load
)
9267 /* Try to move all excess tasks to child's sibling domain */
9268 if (sds
.prefer_sibling
&& local
->group_type
== group_has_spare
&&
9269 busiest
->sum_nr_running
> local
->sum_nr_running
+ 1)
9272 if (busiest
->group_type
!= group_overloaded
) {
9273 if (env
->idle
== CPU_NOT_IDLE
)
9275 * If the busiest group is not overloaded (and as a
9276 * result the local one too) but this CPU is already
9277 * busy, let another idle CPU try to pull task.
9281 if (busiest
->group_weight
> 1 &&
9282 local
->idle_cpus
<= (busiest
->idle_cpus
+ 1))
9284 * If the busiest group is not overloaded
9285 * and there is no imbalance between this and busiest
9286 * group wrt idle CPUs, it is balanced. The imbalance
9287 * becomes significant if the diff is greater than 1
9288 * otherwise we might end up to just move the imbalance
9289 * on another group. Of course this applies only if
9290 * there is more than 1 CPU per group.
9294 if (busiest
->sum_h_nr_running
== 1)
9296 * busiest doesn't have any tasks waiting to run
9302 /* Looks like there is an imbalance. Compute it */
9303 calculate_imbalance(env
, &sds
);
9304 return env
->imbalance
? sds
.busiest
: NULL
;
9312 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9314 static struct rq
*find_busiest_queue(struct lb_env
*env
,
9315 struct sched_group
*group
)
9317 struct rq
*busiest
= NULL
, *rq
;
9318 unsigned long busiest_util
= 0, busiest_load
= 0, busiest_capacity
= 1;
9319 unsigned int busiest_nr
= 0;
9322 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
9323 unsigned long capacity
, load
, util
;
9324 unsigned int nr_running
;
9328 rt
= fbq_classify_rq(rq
);
9331 * We classify groups/runqueues into three groups:
9332 * - regular: there are !numa tasks
9333 * - remote: there are numa tasks that run on the 'wrong' node
9334 * - all: there is no distinction
9336 * In order to avoid migrating ideally placed numa tasks,
9337 * ignore those when there's better options.
9339 * If we ignore the actual busiest queue to migrate another
9340 * task, the next balance pass can still reduce the busiest
9341 * queue by moving tasks around inside the node.
9343 * If we cannot move enough load due to this classification
9344 * the next pass will adjust the group classification and
9345 * allow migration of more tasks.
9347 * Both cases only affect the total convergence complexity.
9349 if (rt
> env
->fbq_type
)
9352 capacity
= capacity_of(i
);
9353 nr_running
= rq
->cfs
.h_nr_running
;
9356 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9357 * eventually lead to active_balancing high->low capacity.
9358 * Higher per-CPU capacity is considered better than balancing
9361 if (env
->sd
->flags
& SD_ASYM_CPUCAPACITY
&&
9362 capacity_of(env
->dst_cpu
) < capacity
&&
9366 switch (env
->migration_type
) {
9369 * When comparing with load imbalance, use cpu_load()
9370 * which is not scaled with the CPU capacity.
9372 load
= cpu_load(rq
);
9374 if (nr_running
== 1 && load
> env
->imbalance
&&
9375 !check_cpu_capacity(rq
, env
->sd
))
9379 * For the load comparisons with the other CPUs,
9380 * consider the cpu_load() scaled with the CPU
9381 * capacity, so that the load can be moved away
9382 * from the CPU that is potentially running at a
9385 * Thus we're looking for max(load_i / capacity_i),
9386 * crosswise multiplication to rid ourselves of the
9387 * division works out to:
9388 * load_i * capacity_j > load_j * capacity_i;
9389 * where j is our previous maximum.
9391 if (load
* busiest_capacity
> busiest_load
* capacity
) {
9392 busiest_load
= load
;
9393 busiest_capacity
= capacity
;
9399 util
= cpu_util(cpu_of(rq
));
9402 * Don't try to pull utilization from a CPU with one
9403 * running task. Whatever its utilization, we will fail
9406 if (nr_running
<= 1)
9409 if (busiest_util
< util
) {
9410 busiest_util
= util
;
9416 if (busiest_nr
< nr_running
) {
9417 busiest_nr
= nr_running
;
9422 case migrate_misfit
:
9424 * For ASYM_CPUCAPACITY domains with misfit tasks we
9425 * simply seek the "biggest" misfit task.
9427 if (rq
->misfit_task_load
> busiest_load
) {
9428 busiest_load
= rq
->misfit_task_load
;
9441 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9442 * so long as it is large enough.
9444 #define MAX_PINNED_INTERVAL 512
9447 asym_active_balance(struct lb_env
*env
)
9450 * ASYM_PACKING needs to force migrate tasks from busy but
9451 * lower priority CPUs in order to pack all tasks in the
9452 * highest priority CPUs.
9454 return env
->idle
!= CPU_NOT_IDLE
&& (env
->sd
->flags
& SD_ASYM_PACKING
) &&
9455 sched_asym_prefer(env
->dst_cpu
, env
->src_cpu
);
9459 voluntary_active_balance(struct lb_env
*env
)
9461 struct sched_domain
*sd
= env
->sd
;
9463 if (asym_active_balance(env
))
9467 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
9468 * It's worth migrating the task if the src_cpu's capacity is reduced
9469 * because of other sched_class or IRQs if more capacity stays
9470 * available on dst_cpu.
9472 if ((env
->idle
!= CPU_NOT_IDLE
) &&
9473 (env
->src_rq
->cfs
.h_nr_running
== 1)) {
9474 if ((check_cpu_capacity(env
->src_rq
, sd
)) &&
9475 (capacity_of(env
->src_cpu
)*sd
->imbalance_pct
< capacity_of(env
->dst_cpu
)*100))
9479 if (env
->migration_type
== migrate_misfit
)
9485 static int need_active_balance(struct lb_env
*env
)
9487 struct sched_domain
*sd
= env
->sd
;
9489 if (voluntary_active_balance(env
))
9492 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
9495 static int active_load_balance_cpu_stop(void *data
);
9497 static int should_we_balance(struct lb_env
*env
)
9499 struct sched_group
*sg
= env
->sd
->groups
;
9503 * Ensure the balancing environment is consistent; can happen
9504 * when the softirq triggers 'during' hotplug.
9506 if (!cpumask_test_cpu(env
->dst_cpu
, env
->cpus
))
9510 * In the newly idle case, we will allow all the CPUs
9511 * to do the newly idle load balance.
9513 if (env
->idle
== CPU_NEWLY_IDLE
)
9516 /* Try to find first idle CPU */
9517 for_each_cpu_and(cpu
, group_balance_mask(sg
), env
->cpus
) {
9521 /* Are we the first idle CPU? */
9522 return cpu
== env
->dst_cpu
;
9525 /* Are we the first CPU of this group ? */
9526 return group_balance_cpu(sg
) == env
->dst_cpu
;
9530 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9531 * tasks if there is an imbalance.
9533 static int load_balance(int this_cpu
, struct rq
*this_rq
,
9534 struct sched_domain
*sd
, enum cpu_idle_type idle
,
9535 int *continue_balancing
)
9537 int ld_moved
, cur_ld_moved
, active_balance
= 0;
9538 struct sched_domain
*sd_parent
= sd
->parent
;
9539 struct sched_group
*group
;
9542 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
9544 struct lb_env env
= {
9546 .dst_cpu
= this_cpu
,
9548 .dst_grpmask
= sched_group_span(sd
->groups
),
9550 .loop_break
= sched_nr_migrate_break
,
9553 .tasks
= LIST_HEAD_INIT(env
.tasks
),
9556 cpumask_and(cpus
, sched_domain_span(sd
), cpu_active_mask
);
9558 schedstat_inc(sd
->lb_count
[idle
]);
9561 if (!should_we_balance(&env
)) {
9562 *continue_balancing
= 0;
9566 group
= find_busiest_group(&env
);
9568 schedstat_inc(sd
->lb_nobusyg
[idle
]);
9572 busiest
= find_busiest_queue(&env
, group
);
9574 schedstat_inc(sd
->lb_nobusyq
[idle
]);
9578 BUG_ON(busiest
== env
.dst_rq
);
9580 schedstat_add(sd
->lb_imbalance
[idle
], env
.imbalance
);
9582 env
.src_cpu
= busiest
->cpu
;
9583 env
.src_rq
= busiest
;
9586 if (busiest
->nr_running
> 1) {
9588 * Attempt to move tasks. If find_busiest_group has found
9589 * an imbalance but busiest->nr_running <= 1, the group is
9590 * still unbalanced. ld_moved simply stays zero, so it is
9591 * correctly treated as an imbalance.
9593 env
.flags
|= LBF_ALL_PINNED
;
9594 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
9597 rq_lock_irqsave(busiest
, &rf
);
9598 update_rq_clock(busiest
);
9601 * cur_ld_moved - load moved in current iteration
9602 * ld_moved - cumulative load moved across iterations
9604 cur_ld_moved
= detach_tasks(&env
);
9607 * We've detached some tasks from busiest_rq. Every
9608 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9609 * unlock busiest->lock, and we are able to be sure
9610 * that nobody can manipulate the tasks in parallel.
9611 * See task_rq_lock() family for the details.
9614 rq_unlock(busiest
, &rf
);
9618 ld_moved
+= cur_ld_moved
;
9621 local_irq_restore(rf
.flags
);
9623 if (env
.flags
& LBF_NEED_BREAK
) {
9624 env
.flags
&= ~LBF_NEED_BREAK
;
9629 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9630 * us and move them to an alternate dst_cpu in our sched_group
9631 * where they can run. The upper limit on how many times we
9632 * iterate on same src_cpu is dependent on number of CPUs in our
9635 * This changes load balance semantics a bit on who can move
9636 * load to a given_cpu. In addition to the given_cpu itself
9637 * (or a ilb_cpu acting on its behalf where given_cpu is
9638 * nohz-idle), we now have balance_cpu in a position to move
9639 * load to given_cpu. In rare situations, this may cause
9640 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9641 * _independently_ and at _same_ time to move some load to
9642 * given_cpu) causing exceess load to be moved to given_cpu.
9643 * This however should not happen so much in practice and
9644 * moreover subsequent load balance cycles should correct the
9645 * excess load moved.
9647 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
9649 /* Prevent to re-select dst_cpu via env's CPUs */
9650 __cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
9652 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
9653 env
.dst_cpu
= env
.new_dst_cpu
;
9654 env
.flags
&= ~LBF_DST_PINNED
;
9656 env
.loop_break
= sched_nr_migrate_break
;
9659 * Go back to "more_balance" rather than "redo" since we
9660 * need to continue with same src_cpu.
9666 * We failed to reach balance because of affinity.
9669 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
9671 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
9672 *group_imbalance
= 1;
9675 /* All tasks on this runqueue were pinned by CPU affinity */
9676 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
9677 __cpumask_clear_cpu(cpu_of(busiest
), cpus
);
9679 * Attempting to continue load balancing at the current
9680 * sched_domain level only makes sense if there are
9681 * active CPUs remaining as possible busiest CPUs to
9682 * pull load from which are not contained within the
9683 * destination group that is receiving any migrated
9686 if (!cpumask_subset(cpus
, env
.dst_grpmask
)) {
9688 env
.loop_break
= sched_nr_migrate_break
;
9691 goto out_all_pinned
;
9696 schedstat_inc(sd
->lb_failed
[idle
]);
9698 * Increment the failure counter only on periodic balance.
9699 * We do not want newidle balance, which can be very
9700 * frequent, pollute the failure counter causing
9701 * excessive cache_hot migrations and active balances.
9703 if (idle
!= CPU_NEWLY_IDLE
)
9704 sd
->nr_balance_failed
++;
9706 if (need_active_balance(&env
)) {
9707 unsigned long flags
;
9709 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
9712 * Don't kick the active_load_balance_cpu_stop,
9713 * if the curr task on busiest CPU can't be
9714 * moved to this_cpu:
9716 if (!cpumask_test_cpu(this_cpu
, busiest
->curr
->cpus_ptr
)) {
9717 raw_spin_unlock_irqrestore(&busiest
->lock
,
9719 env
.flags
|= LBF_ALL_PINNED
;
9720 goto out_one_pinned
;
9724 * ->active_balance synchronizes accesses to
9725 * ->active_balance_work. Once set, it's cleared
9726 * only after active load balance is finished.
9728 if (!busiest
->active_balance
) {
9729 busiest
->active_balance
= 1;
9730 busiest
->push_cpu
= this_cpu
;
9733 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
9735 if (active_balance
) {
9736 stop_one_cpu_nowait(cpu_of(busiest
),
9737 active_load_balance_cpu_stop
, busiest
,
9738 &busiest
->active_balance_work
);
9741 /* We've kicked active balancing, force task migration. */
9742 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
9745 sd
->nr_balance_failed
= 0;
9747 if (likely(!active_balance
) || voluntary_active_balance(&env
)) {
9748 /* We were unbalanced, so reset the balancing interval */
9749 sd
->balance_interval
= sd
->min_interval
;
9752 * If we've begun active balancing, start to back off. This
9753 * case may not be covered by the all_pinned logic if there
9754 * is only 1 task on the busy runqueue (because we don't call
9757 if (sd
->balance_interval
< sd
->max_interval
)
9758 sd
->balance_interval
*= 2;
9765 * We reach balance although we may have faced some affinity
9766 * constraints. Clear the imbalance flag only if other tasks got
9767 * a chance to move and fix the imbalance.
9769 if (sd_parent
&& !(env
.flags
& LBF_ALL_PINNED
)) {
9770 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
9772 if (*group_imbalance
)
9773 *group_imbalance
= 0;
9778 * We reach balance because all tasks are pinned at this level so
9779 * we can't migrate them. Let the imbalance flag set so parent level
9780 * can try to migrate them.
9782 schedstat_inc(sd
->lb_balanced
[idle
]);
9784 sd
->nr_balance_failed
= 0;
9790 * newidle_balance() disregards balance intervals, so we could
9791 * repeatedly reach this code, which would lead to balance_interval
9792 * skyrocketting in a short amount of time. Skip the balance_interval
9793 * increase logic to avoid that.
9795 if (env
.idle
== CPU_NEWLY_IDLE
)
9798 /* tune up the balancing interval */
9799 if ((env
.flags
& LBF_ALL_PINNED
&&
9800 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
9801 sd
->balance_interval
< sd
->max_interval
)
9802 sd
->balance_interval
*= 2;
9807 static inline unsigned long
9808 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
9810 unsigned long interval
= sd
->balance_interval
;
9813 interval
*= sd
->busy_factor
;
9815 /* scale ms to jiffies */
9816 interval
= msecs_to_jiffies(interval
);
9819 * Reduce likelihood of busy balancing at higher domains racing with
9820 * balancing at lower domains by preventing their balancing periods
9821 * from being multiples of each other.
9826 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
9832 update_next_balance(struct sched_domain
*sd
, unsigned long *next_balance
)
9834 unsigned long interval
, next
;
9836 /* used by idle balance, so cpu_busy = 0 */
9837 interval
= get_sd_balance_interval(sd
, 0);
9838 next
= sd
->last_balance
+ interval
;
9840 if (time_after(*next_balance
, next
))
9841 *next_balance
= next
;
9845 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9846 * running tasks off the busiest CPU onto idle CPUs. It requires at
9847 * least 1 task to be running on each physical CPU where possible, and
9848 * avoids physical / logical imbalances.
9850 static int active_load_balance_cpu_stop(void *data
)
9852 struct rq
*busiest_rq
= data
;
9853 int busiest_cpu
= cpu_of(busiest_rq
);
9854 int target_cpu
= busiest_rq
->push_cpu
;
9855 struct rq
*target_rq
= cpu_rq(target_cpu
);
9856 struct sched_domain
*sd
;
9857 struct task_struct
*p
= NULL
;
9860 rq_lock_irq(busiest_rq
, &rf
);
9862 * Between queueing the stop-work and running it is a hole in which
9863 * CPUs can become inactive. We should not move tasks from or to
9866 if (!cpu_active(busiest_cpu
) || !cpu_active(target_cpu
))
9869 /* Make sure the requested CPU hasn't gone down in the meantime: */
9870 if (unlikely(busiest_cpu
!= smp_processor_id() ||
9871 !busiest_rq
->active_balance
))
9874 /* Is there any task to move? */
9875 if (busiest_rq
->nr_running
<= 1)
9879 * This condition is "impossible", if it occurs
9880 * we need to fix it. Originally reported by
9881 * Bjorn Helgaas on a 128-CPU setup.
9883 BUG_ON(busiest_rq
== target_rq
);
9885 /* Search for an sd spanning us and the target CPU. */
9887 for_each_domain(target_cpu
, sd
) {
9888 if (cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
9893 struct lb_env env
= {
9895 .dst_cpu
= target_cpu
,
9896 .dst_rq
= target_rq
,
9897 .src_cpu
= busiest_rq
->cpu
,
9898 .src_rq
= busiest_rq
,
9901 * can_migrate_task() doesn't need to compute new_dst_cpu
9902 * for active balancing. Since we have CPU_IDLE, but no
9903 * @dst_grpmask we need to make that test go away with lying
9906 .flags
= LBF_DST_PINNED
,
9909 schedstat_inc(sd
->alb_count
);
9910 update_rq_clock(busiest_rq
);
9912 p
= detach_one_task(&env
);
9914 schedstat_inc(sd
->alb_pushed
);
9915 /* Active balancing done, reset the failure counter. */
9916 sd
->nr_balance_failed
= 0;
9918 schedstat_inc(sd
->alb_failed
);
9923 busiest_rq
->active_balance
= 0;
9924 rq_unlock(busiest_rq
, &rf
);
9927 attach_one_task(target_rq
, p
);
9934 static DEFINE_SPINLOCK(balancing
);
9937 * Scale the max load_balance interval with the number of CPUs in the system.
9938 * This trades load-balance latency on larger machines for less cross talk.
9940 void update_max_interval(void)
9942 max_load_balance_interval
= HZ
*num_online_cpus()/10;
9946 * It checks each scheduling domain to see if it is due to be balanced,
9947 * and initiates a balancing operation if so.
9949 * Balancing parameters are set up in init_sched_domains.
9951 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
9953 int continue_balancing
= 1;
9955 int busy
= idle
!= CPU_IDLE
&& !sched_idle_cpu(cpu
);
9956 unsigned long interval
;
9957 struct sched_domain
*sd
;
9958 /* Earliest time when we have to do rebalance again */
9959 unsigned long next_balance
= jiffies
+ 60*HZ
;
9960 int update_next_balance
= 0;
9961 int need_serialize
, need_decay
= 0;
9965 for_each_domain(cpu
, sd
) {
9967 * Decay the newidle max times here because this is a regular
9968 * visit to all the domains. Decay ~1% per second.
9970 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
9971 sd
->max_newidle_lb_cost
=
9972 (sd
->max_newidle_lb_cost
* 253) / 256;
9973 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
9976 max_cost
+= sd
->max_newidle_lb_cost
;
9979 * Stop the load balance at this level. There is another
9980 * CPU in our sched group which is doing load balancing more
9983 if (!continue_balancing
) {
9989 interval
= get_sd_balance_interval(sd
, busy
);
9991 need_serialize
= sd
->flags
& SD_SERIALIZE
;
9992 if (need_serialize
) {
9993 if (!spin_trylock(&balancing
))
9997 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
9998 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
10000 * The LBF_DST_PINNED logic could have changed
10001 * env->dst_cpu, so we can't know our idle
10002 * state even if we migrated tasks. Update it.
10004 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
10005 busy
= idle
!= CPU_IDLE
&& !sched_idle_cpu(cpu
);
10007 sd
->last_balance
= jiffies
;
10008 interval
= get_sd_balance_interval(sd
, busy
);
10010 if (need_serialize
)
10011 spin_unlock(&balancing
);
10013 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
10014 next_balance
= sd
->last_balance
+ interval
;
10015 update_next_balance
= 1;
10020 * Ensure the rq-wide value also decays but keep it at a
10021 * reasonable floor to avoid funnies with rq->avg_idle.
10023 rq
->max_idle_balance_cost
=
10024 max((u64
)sysctl_sched_migration_cost
, max_cost
);
10029 * next_balance will be updated only when there is a need.
10030 * When the cpu is attached to null domain for ex, it will not be
10033 if (likely(update_next_balance
)) {
10034 rq
->next_balance
= next_balance
;
10036 #ifdef CONFIG_NO_HZ_COMMON
10038 * If this CPU has been elected to perform the nohz idle
10039 * balance. Other idle CPUs have already rebalanced with
10040 * nohz_idle_balance() and nohz.next_balance has been
10041 * updated accordingly. This CPU is now running the idle load
10042 * balance for itself and we need to update the
10043 * nohz.next_balance accordingly.
10045 if ((idle
== CPU_IDLE
) && time_after(nohz
.next_balance
, rq
->next_balance
))
10046 nohz
.next_balance
= rq
->next_balance
;
10051 static inline int on_null_domain(struct rq
*rq
)
10053 return unlikely(!rcu_dereference_sched(rq
->sd
));
10056 #ifdef CONFIG_NO_HZ_COMMON
10058 * idle load balancing details
10059 * - When one of the busy CPUs notice that there may be an idle rebalancing
10060 * needed, they will kick the idle load balancer, which then does idle
10061 * load balancing for all the idle CPUs.
10062 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
10066 static inline int find_new_ilb(void)
10070 for_each_cpu_and(ilb
, nohz
.idle_cpus_mask
,
10071 housekeeping_cpumask(HK_FLAG_MISC
)) {
10073 if (ilb
== smp_processor_id())
10084 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
10085 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
10087 static void kick_ilb(unsigned int flags
)
10092 * Increase nohz.next_balance only when if full ilb is triggered but
10093 * not if we only update stats.
10095 if (flags
& NOHZ_BALANCE_KICK
)
10096 nohz
.next_balance
= jiffies
+1;
10098 ilb_cpu
= find_new_ilb();
10100 if (ilb_cpu
>= nr_cpu_ids
)
10104 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
10105 * the first flag owns it; cleared by nohz_csd_func().
10107 flags
= atomic_fetch_or(flags
, nohz_flags(ilb_cpu
));
10108 if (flags
& NOHZ_KICK_MASK
)
10112 * This way we generate an IPI on the target CPU which
10113 * is idle. And the softirq performing nohz idle load balance
10114 * will be run before returning from the IPI.
10116 smp_call_function_single_async(ilb_cpu
, &cpu_rq(ilb_cpu
)->nohz_csd
);
10120 * Current decision point for kicking the idle load balancer in the presence
10121 * of idle CPUs in the system.
10123 static void nohz_balancer_kick(struct rq
*rq
)
10125 unsigned long now
= jiffies
;
10126 struct sched_domain_shared
*sds
;
10127 struct sched_domain
*sd
;
10128 int nr_busy
, i
, cpu
= rq
->cpu
;
10129 unsigned int flags
= 0;
10131 if (unlikely(rq
->idle_balance
))
10135 * We may be recently in ticked or tickless idle mode. At the first
10136 * busy tick after returning from idle, we will update the busy stats.
10138 nohz_balance_exit_idle(rq
);
10141 * None are in tickless mode and hence no need for NOHZ idle load
10144 if (likely(!atomic_read(&nohz
.nr_cpus
)))
10147 if (READ_ONCE(nohz
.has_blocked
) &&
10148 time_after(now
, READ_ONCE(nohz
.next_blocked
)))
10149 flags
= NOHZ_STATS_KICK
;
10151 if (time_before(now
, nohz
.next_balance
))
10154 if (rq
->nr_running
>= 2) {
10155 flags
= NOHZ_KICK_MASK
;
10161 sd
= rcu_dereference(rq
->sd
);
10164 * If there's a CFS task and the current CPU has reduced
10165 * capacity; kick the ILB to see if there's a better CPU to run
10168 if (rq
->cfs
.h_nr_running
>= 1 && check_cpu_capacity(rq
, sd
)) {
10169 flags
= NOHZ_KICK_MASK
;
10174 sd
= rcu_dereference(per_cpu(sd_asym_packing
, cpu
));
10177 * When ASYM_PACKING; see if there's a more preferred CPU
10178 * currently idle; in which case, kick the ILB to move tasks
10181 for_each_cpu_and(i
, sched_domain_span(sd
), nohz
.idle_cpus_mask
) {
10182 if (sched_asym_prefer(i
, cpu
)) {
10183 flags
= NOHZ_KICK_MASK
;
10189 sd
= rcu_dereference(per_cpu(sd_asym_cpucapacity
, cpu
));
10192 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10193 * to run the misfit task on.
10195 if (check_misfit_status(rq
, sd
)) {
10196 flags
= NOHZ_KICK_MASK
;
10201 * For asymmetric systems, we do not want to nicely balance
10202 * cache use, instead we want to embrace asymmetry and only
10203 * ensure tasks have enough CPU capacity.
10205 * Skip the LLC logic because it's not relevant in that case.
10210 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
10213 * If there is an imbalance between LLC domains (IOW we could
10214 * increase the overall cache use), we need some less-loaded LLC
10215 * domain to pull some load. Likewise, we may need to spread
10216 * load within the current LLC domain (e.g. packed SMT cores but
10217 * other CPUs are idle). We can't really know from here how busy
10218 * the others are - so just get a nohz balance going if it looks
10219 * like this LLC domain has tasks we could move.
10221 nr_busy
= atomic_read(&sds
->nr_busy_cpus
);
10223 flags
= NOHZ_KICK_MASK
;
10234 static void set_cpu_sd_state_busy(int cpu
)
10236 struct sched_domain
*sd
;
10239 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
10241 if (!sd
|| !sd
->nohz_idle
)
10245 atomic_inc(&sd
->shared
->nr_busy_cpus
);
10250 void nohz_balance_exit_idle(struct rq
*rq
)
10252 SCHED_WARN_ON(rq
!= this_rq());
10254 if (likely(!rq
->nohz_tick_stopped
))
10257 rq
->nohz_tick_stopped
= 0;
10258 cpumask_clear_cpu(rq
->cpu
, nohz
.idle_cpus_mask
);
10259 atomic_dec(&nohz
.nr_cpus
);
10261 set_cpu_sd_state_busy(rq
->cpu
);
10264 static void set_cpu_sd_state_idle(int cpu
)
10266 struct sched_domain
*sd
;
10269 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
10271 if (!sd
|| sd
->nohz_idle
)
10275 atomic_dec(&sd
->shared
->nr_busy_cpus
);
10281 * This routine will record that the CPU is going idle with tick stopped.
10282 * This info will be used in performing idle load balancing in the future.
10284 void nohz_balance_enter_idle(int cpu
)
10286 struct rq
*rq
= cpu_rq(cpu
);
10288 SCHED_WARN_ON(cpu
!= smp_processor_id());
10290 /* If this CPU is going down, then nothing needs to be done: */
10291 if (!cpu_active(cpu
))
10294 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10295 if (!housekeeping_cpu(cpu
, HK_FLAG_SCHED
))
10299 * Can be set safely without rq->lock held
10300 * If a clear happens, it will have evaluated last additions because
10301 * rq->lock is held during the check and the clear
10303 rq
->has_blocked_load
= 1;
10306 * The tick is still stopped but load could have been added in the
10307 * meantime. We set the nohz.has_blocked flag to trig a check of the
10308 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10309 * of nohz.has_blocked can only happen after checking the new load
10311 if (rq
->nohz_tick_stopped
)
10314 /* If we're a completely isolated CPU, we don't play: */
10315 if (on_null_domain(rq
))
10318 rq
->nohz_tick_stopped
= 1;
10320 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
10321 atomic_inc(&nohz
.nr_cpus
);
10324 * Ensures that if nohz_idle_balance() fails to observe our
10325 * @idle_cpus_mask store, it must observe the @has_blocked
10328 smp_mb__after_atomic();
10330 set_cpu_sd_state_idle(cpu
);
10334 * Each time a cpu enter idle, we assume that it has blocked load and
10335 * enable the periodic update of the load of idle cpus
10337 WRITE_ONCE(nohz
.has_blocked
, 1);
10341 * Internal function that runs load balance for all idle cpus. The load balance
10342 * can be a simple update of blocked load or a complete load balance with
10343 * tasks movement depending of flags.
10344 * The function returns false if the loop has stopped before running
10345 * through all idle CPUs.
10347 static bool _nohz_idle_balance(struct rq
*this_rq
, unsigned int flags
,
10348 enum cpu_idle_type idle
)
10350 /* Earliest time when we have to do rebalance again */
10351 unsigned long now
= jiffies
;
10352 unsigned long next_balance
= now
+ 60*HZ
;
10353 bool has_blocked_load
= false;
10354 int update_next_balance
= 0;
10355 int this_cpu
= this_rq
->cpu
;
10360 SCHED_WARN_ON((flags
& NOHZ_KICK_MASK
) == NOHZ_BALANCE_KICK
);
10363 * We assume there will be no idle load after this update and clear
10364 * the has_blocked flag. If a cpu enters idle in the mean time, it will
10365 * set the has_blocked flag and trig another update of idle load.
10366 * Because a cpu that becomes idle, is added to idle_cpus_mask before
10367 * setting the flag, we are sure to not clear the state and not
10368 * check the load of an idle cpu.
10370 WRITE_ONCE(nohz
.has_blocked
, 0);
10373 * Ensures that if we miss the CPU, we must see the has_blocked
10374 * store from nohz_balance_enter_idle().
10378 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
10379 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
10383 * If this CPU gets work to do, stop the load balancing
10384 * work being done for other CPUs. Next load
10385 * balancing owner will pick it up.
10387 if (need_resched()) {
10388 has_blocked_load
= true;
10392 rq
= cpu_rq(balance_cpu
);
10394 has_blocked_load
|= update_nohz_stats(rq
, true);
10397 * If time for next balance is due,
10400 if (time_after_eq(jiffies
, rq
->next_balance
)) {
10401 struct rq_flags rf
;
10403 rq_lock_irqsave(rq
, &rf
);
10404 update_rq_clock(rq
);
10405 rq_unlock_irqrestore(rq
, &rf
);
10407 if (flags
& NOHZ_BALANCE_KICK
)
10408 rebalance_domains(rq
, CPU_IDLE
);
10411 if (time_after(next_balance
, rq
->next_balance
)) {
10412 next_balance
= rq
->next_balance
;
10413 update_next_balance
= 1;
10418 * next_balance will be updated only when there is a need.
10419 * When the CPU is attached to null domain for ex, it will not be
10422 if (likely(update_next_balance
))
10423 nohz
.next_balance
= next_balance
;
10425 /* Newly idle CPU doesn't need an update */
10426 if (idle
!= CPU_NEWLY_IDLE
) {
10427 update_blocked_averages(this_cpu
);
10428 has_blocked_load
|= this_rq
->has_blocked_load
;
10431 if (flags
& NOHZ_BALANCE_KICK
)
10432 rebalance_domains(this_rq
, CPU_IDLE
);
10434 WRITE_ONCE(nohz
.next_blocked
,
10435 now
+ msecs_to_jiffies(LOAD_AVG_PERIOD
));
10437 /* The full idle balance loop has been done */
10441 /* There is still blocked load, enable periodic update */
10442 if (has_blocked_load
)
10443 WRITE_ONCE(nohz
.has_blocked
, 1);
10449 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
10450 * rebalancing for all the cpus for whom scheduler ticks are stopped.
10452 static bool nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
10454 unsigned int flags
= this_rq
->nohz_idle_balance
;
10459 this_rq
->nohz_idle_balance
= 0;
10461 if (idle
!= CPU_IDLE
)
10464 _nohz_idle_balance(this_rq
, flags
, idle
);
10469 static void nohz_newidle_balance(struct rq
*this_rq
)
10471 int this_cpu
= this_rq
->cpu
;
10474 * This CPU doesn't want to be disturbed by scheduler
10477 if (!housekeeping_cpu(this_cpu
, HK_FLAG_SCHED
))
10480 /* Will wake up very soon. No time for doing anything else*/
10481 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
10484 /* Don't need to update blocked load of idle CPUs*/
10485 if (!READ_ONCE(nohz
.has_blocked
) ||
10486 time_before(jiffies
, READ_ONCE(nohz
.next_blocked
)))
10489 raw_spin_unlock(&this_rq
->lock
);
10491 * This CPU is going to be idle and blocked load of idle CPUs
10492 * need to be updated. Run the ilb locally as it is a good
10493 * candidate for ilb instead of waking up another idle CPU.
10494 * Kick an normal ilb if we failed to do the update.
10496 if (!_nohz_idle_balance(this_rq
, NOHZ_STATS_KICK
, CPU_NEWLY_IDLE
))
10497 kick_ilb(NOHZ_STATS_KICK
);
10498 raw_spin_lock(&this_rq
->lock
);
10501 #else /* !CONFIG_NO_HZ_COMMON */
10502 static inline void nohz_balancer_kick(struct rq
*rq
) { }
10504 static inline bool nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
10509 static inline void nohz_newidle_balance(struct rq
*this_rq
) { }
10510 #endif /* CONFIG_NO_HZ_COMMON */
10513 * idle_balance is called by schedule() if this_cpu is about to become
10514 * idle. Attempts to pull tasks from other CPUs.
10517 * < 0 - we released the lock and there are !fair tasks present
10518 * 0 - failed, no new tasks
10519 * > 0 - success, new (fair) tasks present
10521 static int newidle_balance(struct rq
*this_rq
, struct rq_flags
*rf
)
10523 unsigned long next_balance
= jiffies
+ HZ
;
10524 int this_cpu
= this_rq
->cpu
;
10525 struct sched_domain
*sd
;
10526 int pulled_task
= 0;
10529 update_misfit_status(NULL
, this_rq
);
10531 * We must set idle_stamp _before_ calling idle_balance(), such that we
10532 * measure the duration of idle_balance() as idle time.
10534 this_rq
->idle_stamp
= rq_clock(this_rq
);
10537 * Do not pull tasks towards !active CPUs...
10539 if (!cpu_active(this_cpu
))
10543 * This is OK, because current is on_cpu, which avoids it being picked
10544 * for load-balance and preemption/IRQs are still disabled avoiding
10545 * further scheduler activity on it and we're being very careful to
10546 * re-start the picking loop.
10548 rq_unpin_lock(this_rq
, rf
);
10550 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
10551 !READ_ONCE(this_rq
->rd
->overload
)) {
10554 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
10556 update_next_balance(sd
, &next_balance
);
10559 nohz_newidle_balance(this_rq
);
10564 raw_spin_unlock(&this_rq
->lock
);
10566 update_blocked_averages(this_cpu
);
10568 for_each_domain(this_cpu
, sd
) {
10569 int continue_balancing
= 1;
10570 u64 t0
, domain_cost
;
10572 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
10573 update_next_balance(sd
, &next_balance
);
10577 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
10578 t0
= sched_clock_cpu(this_cpu
);
10580 pulled_task
= load_balance(this_cpu
, this_rq
,
10581 sd
, CPU_NEWLY_IDLE
,
10582 &continue_balancing
);
10584 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
10585 if (domain_cost
> sd
->max_newidle_lb_cost
)
10586 sd
->max_newidle_lb_cost
= domain_cost
;
10588 curr_cost
+= domain_cost
;
10591 update_next_balance(sd
, &next_balance
);
10594 * Stop searching for tasks to pull if there are
10595 * now runnable tasks on this rq.
10597 if (pulled_task
|| this_rq
->nr_running
> 0)
10602 raw_spin_lock(&this_rq
->lock
);
10604 if (curr_cost
> this_rq
->max_idle_balance_cost
)
10605 this_rq
->max_idle_balance_cost
= curr_cost
;
10609 * While browsing the domains, we released the rq lock, a task could
10610 * have been enqueued in the meantime. Since we're not going idle,
10611 * pretend we pulled a task.
10613 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
10616 /* Move the next balance forward */
10617 if (time_after(this_rq
->next_balance
, next_balance
))
10618 this_rq
->next_balance
= next_balance
;
10620 /* Is there a task of a high priority class? */
10621 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
10625 this_rq
->idle_stamp
= 0;
10627 rq_repin_lock(this_rq
, rf
);
10629 return pulled_task
;
10633 * run_rebalance_domains is triggered when needed from the scheduler tick.
10634 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10636 static __latent_entropy
void run_rebalance_domains(struct softirq_action
*h
)
10638 struct rq
*this_rq
= this_rq();
10639 enum cpu_idle_type idle
= this_rq
->idle_balance
?
10640 CPU_IDLE
: CPU_NOT_IDLE
;
10643 * If this CPU has a pending nohz_balance_kick, then do the
10644 * balancing on behalf of the other idle CPUs whose ticks are
10645 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10646 * give the idle CPUs a chance to load balance. Else we may
10647 * load balance only within the local sched_domain hierarchy
10648 * and abort nohz_idle_balance altogether if we pull some load.
10650 if (nohz_idle_balance(this_rq
, idle
))
10653 /* normal load balance */
10654 update_blocked_averages(this_rq
->cpu
);
10655 rebalance_domains(this_rq
, idle
);
10659 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10661 void trigger_load_balance(struct rq
*rq
)
10663 /* Don't need to rebalance while attached to NULL domain */
10664 if (unlikely(on_null_domain(rq
)))
10667 if (time_after_eq(jiffies
, rq
->next_balance
))
10668 raise_softirq(SCHED_SOFTIRQ
);
10670 nohz_balancer_kick(rq
);
10673 static void rq_online_fair(struct rq
*rq
)
10677 update_runtime_enabled(rq
);
10680 static void rq_offline_fair(struct rq
*rq
)
10684 /* Ensure any throttled groups are reachable by pick_next_task */
10685 unthrottle_offline_cfs_rqs(rq
);
10688 #endif /* CONFIG_SMP */
10691 * scheduler tick hitting a task of our scheduling class.
10693 * NOTE: This function can be called remotely by the tick offload that
10694 * goes along full dynticks. Therefore no local assumption can be made
10695 * and everything must be accessed through the @rq and @curr passed in
10698 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
10700 struct cfs_rq
*cfs_rq
;
10701 struct sched_entity
*se
= &curr
->se
;
10703 for_each_sched_entity(se
) {
10704 cfs_rq
= cfs_rq_of(se
);
10705 entity_tick(cfs_rq
, se
, queued
);
10708 if (static_branch_unlikely(&sched_numa_balancing
))
10709 task_tick_numa(rq
, curr
);
10711 update_misfit_status(curr
, rq
);
10712 update_overutilized_status(task_rq(curr
));
10716 * called on fork with the child task as argument from the parent's context
10717 * - child not yet on the tasklist
10718 * - preemption disabled
10720 static void task_fork_fair(struct task_struct
*p
)
10722 struct cfs_rq
*cfs_rq
;
10723 struct sched_entity
*se
= &p
->se
, *curr
;
10724 struct rq
*rq
= this_rq();
10725 struct rq_flags rf
;
10728 update_rq_clock(rq
);
10730 cfs_rq
= task_cfs_rq(current
);
10731 curr
= cfs_rq
->curr
;
10733 update_curr(cfs_rq
);
10734 se
->vruntime
= curr
->vruntime
;
10736 place_entity(cfs_rq
, se
, 1);
10738 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
10740 * Upon rescheduling, sched_class::put_prev_task() will place
10741 * 'current' within the tree based on its new key value.
10743 swap(curr
->vruntime
, se
->vruntime
);
10747 se
->vruntime
-= cfs_rq
->min_vruntime
;
10748 rq_unlock(rq
, &rf
);
10752 * Priority of the task has changed. Check to see if we preempt
10753 * the current task.
10756 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
10758 if (!task_on_rq_queued(p
))
10761 if (rq
->cfs
.nr_running
== 1)
10765 * Reschedule if we are currently running on this runqueue and
10766 * our priority decreased, or if we are not currently running on
10767 * this runqueue and our priority is higher than the current's
10769 if (rq
->curr
== p
) {
10770 if (p
->prio
> oldprio
)
10773 check_preempt_curr(rq
, p
, 0);
10776 static inline bool vruntime_normalized(struct task_struct
*p
)
10778 struct sched_entity
*se
= &p
->se
;
10781 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10782 * the dequeue_entity(.flags=0) will already have normalized the
10789 * When !on_rq, vruntime of the task has usually NOT been normalized.
10790 * But there are some cases where it has already been normalized:
10792 * - A forked child which is waiting for being woken up by
10793 * wake_up_new_task().
10794 * - A task which has been woken up by try_to_wake_up() and
10795 * waiting for actually being woken up by sched_ttwu_pending().
10797 if (!se
->sum_exec_runtime
||
10798 (p
->state
== TASK_WAKING
&& p
->sched_remote_wakeup
))
10804 #ifdef CONFIG_FAIR_GROUP_SCHED
10806 * Propagate the changes of the sched_entity across the tg tree to make it
10807 * visible to the root
10809 static void propagate_entity_cfs_rq(struct sched_entity
*se
)
10811 struct cfs_rq
*cfs_rq
;
10813 /* Start to propagate at parent */
10816 for_each_sched_entity(se
) {
10817 cfs_rq
= cfs_rq_of(se
);
10819 if (cfs_rq_throttled(cfs_rq
))
10822 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
10826 static void propagate_entity_cfs_rq(struct sched_entity
*se
) { }
10829 static void detach_entity_cfs_rq(struct sched_entity
*se
)
10831 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10833 /* Catch up with the cfs_rq and remove our load when we leave */
10834 update_load_avg(cfs_rq
, se
, 0);
10835 detach_entity_load_avg(cfs_rq
, se
);
10836 update_tg_load_avg(cfs_rq
);
10837 propagate_entity_cfs_rq(se
);
10840 static void attach_entity_cfs_rq(struct sched_entity
*se
)
10842 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10844 #ifdef CONFIG_FAIR_GROUP_SCHED
10846 * Since the real-depth could have been changed (only FAIR
10847 * class maintain depth value), reset depth properly.
10849 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
10852 /* Synchronize entity with its cfs_rq */
10853 update_load_avg(cfs_rq
, se
, sched_feat(ATTACH_AGE_LOAD
) ? 0 : SKIP_AGE_LOAD
);
10854 attach_entity_load_avg(cfs_rq
, se
);
10855 update_tg_load_avg(cfs_rq
);
10856 propagate_entity_cfs_rq(se
);
10859 static void detach_task_cfs_rq(struct task_struct
*p
)
10861 struct sched_entity
*se
= &p
->se
;
10862 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10864 if (!vruntime_normalized(p
)) {
10866 * Fix up our vruntime so that the current sleep doesn't
10867 * cause 'unlimited' sleep bonus.
10869 place_entity(cfs_rq
, se
, 0);
10870 se
->vruntime
-= cfs_rq
->min_vruntime
;
10873 detach_entity_cfs_rq(se
);
10876 static void attach_task_cfs_rq(struct task_struct
*p
)
10878 struct sched_entity
*se
= &p
->se
;
10879 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10881 attach_entity_cfs_rq(se
);
10883 if (!vruntime_normalized(p
))
10884 se
->vruntime
+= cfs_rq
->min_vruntime
;
10887 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
10889 detach_task_cfs_rq(p
);
10892 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
10894 attach_task_cfs_rq(p
);
10896 if (task_on_rq_queued(p
)) {
10898 * We were most likely switched from sched_rt, so
10899 * kick off the schedule if running, otherwise just see
10900 * if we can still preempt the current task.
10905 check_preempt_curr(rq
, p
, 0);
10909 /* Account for a task changing its policy or group.
10911 * This routine is mostly called to set cfs_rq->curr field when a task
10912 * migrates between groups/classes.
10914 static void set_next_task_fair(struct rq
*rq
, struct task_struct
*p
, bool first
)
10916 struct sched_entity
*se
= &p
->se
;
10919 if (task_on_rq_queued(p
)) {
10921 * Move the next running task to the front of the list, so our
10922 * cfs_tasks list becomes MRU one.
10924 list_move(&se
->group_node
, &rq
->cfs_tasks
);
10928 for_each_sched_entity(se
) {
10929 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10931 set_next_entity(cfs_rq
, se
);
10932 /* ensure bandwidth has been allocated on our new cfs_rq */
10933 account_cfs_rq_runtime(cfs_rq
, 0);
10937 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
10939 cfs_rq
->tasks_timeline
= RB_ROOT_CACHED
;
10940 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
10941 #ifndef CONFIG_64BIT
10942 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
10945 raw_spin_lock_init(&cfs_rq
->removed
.lock
);
10949 #ifdef CONFIG_FAIR_GROUP_SCHED
10950 static void task_set_group_fair(struct task_struct
*p
)
10952 struct sched_entity
*se
= &p
->se
;
10954 set_task_rq(p
, task_cpu(p
));
10955 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
10958 static void task_move_group_fair(struct task_struct
*p
)
10960 detach_task_cfs_rq(p
);
10961 set_task_rq(p
, task_cpu(p
));
10964 /* Tell se's cfs_rq has been changed -- migrated */
10965 p
->se
.avg
.last_update_time
= 0;
10967 attach_task_cfs_rq(p
);
10970 static void task_change_group_fair(struct task_struct
*p
, int type
)
10973 case TASK_SET_GROUP
:
10974 task_set_group_fair(p
);
10977 case TASK_MOVE_GROUP
:
10978 task_move_group_fair(p
);
10983 void free_fair_sched_group(struct task_group
*tg
)
10987 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
10989 for_each_possible_cpu(i
) {
10991 kfree(tg
->cfs_rq
[i
]);
11000 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
11002 struct sched_entity
*se
;
11003 struct cfs_rq
*cfs_rq
;
11006 tg
->cfs_rq
= kcalloc(nr_cpu_ids
, sizeof(cfs_rq
), GFP_KERNEL
);
11009 tg
->se
= kcalloc(nr_cpu_ids
, sizeof(se
), GFP_KERNEL
);
11013 tg
->shares
= NICE_0_LOAD
;
11015 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
11017 for_each_possible_cpu(i
) {
11018 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
11019 GFP_KERNEL
, cpu_to_node(i
));
11023 se
= kzalloc_node(sizeof(struct sched_entity
),
11024 GFP_KERNEL
, cpu_to_node(i
));
11028 init_cfs_rq(cfs_rq
);
11029 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
11030 init_entity_runnable_average(se
);
11041 void online_fair_sched_group(struct task_group
*tg
)
11043 struct sched_entity
*se
;
11044 struct rq_flags rf
;
11048 for_each_possible_cpu(i
) {
11051 rq_lock_irq(rq
, &rf
);
11052 update_rq_clock(rq
);
11053 attach_entity_cfs_rq(se
);
11054 sync_throttle(tg
, i
);
11055 rq_unlock_irq(rq
, &rf
);
11059 void unregister_fair_sched_group(struct task_group
*tg
)
11061 unsigned long flags
;
11065 for_each_possible_cpu(cpu
) {
11067 remove_entity_load_avg(tg
->se
[cpu
]);
11070 * Only empty task groups can be destroyed; so we can speculatively
11071 * check on_list without danger of it being re-added.
11073 if (!tg
->cfs_rq
[cpu
]->on_list
)
11078 raw_spin_lock_irqsave(&rq
->lock
, flags
);
11079 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
11080 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
11084 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
11085 struct sched_entity
*se
, int cpu
,
11086 struct sched_entity
*parent
)
11088 struct rq
*rq
= cpu_rq(cpu
);
11092 init_cfs_rq_runtime(cfs_rq
);
11094 tg
->cfs_rq
[cpu
] = cfs_rq
;
11097 /* se could be NULL for root_task_group */
11102 se
->cfs_rq
= &rq
->cfs
;
11105 se
->cfs_rq
= parent
->my_q
;
11106 se
->depth
= parent
->depth
+ 1;
11110 /* guarantee group entities always have weight */
11111 update_load_set(&se
->load
, NICE_0_LOAD
);
11112 se
->parent
= parent
;
11115 static DEFINE_MUTEX(shares_mutex
);
11117 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
11122 * We can't change the weight of the root cgroup.
11127 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
11129 mutex_lock(&shares_mutex
);
11130 if (tg
->shares
== shares
)
11133 tg
->shares
= shares
;
11134 for_each_possible_cpu(i
) {
11135 struct rq
*rq
= cpu_rq(i
);
11136 struct sched_entity
*se
= tg
->se
[i
];
11137 struct rq_flags rf
;
11139 /* Propagate contribution to hierarchy */
11140 rq_lock_irqsave(rq
, &rf
);
11141 update_rq_clock(rq
);
11142 for_each_sched_entity(se
) {
11143 update_load_avg(cfs_rq_of(se
), se
, UPDATE_TG
);
11144 update_cfs_group(se
);
11146 rq_unlock_irqrestore(rq
, &rf
);
11150 mutex_unlock(&shares_mutex
);
11153 #else /* CONFIG_FAIR_GROUP_SCHED */
11155 void free_fair_sched_group(struct task_group
*tg
) { }
11157 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
11162 void online_fair_sched_group(struct task_group
*tg
) { }
11164 void unregister_fair_sched_group(struct task_group
*tg
) { }
11166 #endif /* CONFIG_FAIR_GROUP_SCHED */
11169 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
11171 struct sched_entity
*se
= &task
->se
;
11172 unsigned int rr_interval
= 0;
11175 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11178 if (rq
->cfs
.load
.weight
)
11179 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
11181 return rr_interval
;
11185 * All the scheduling class methods:
11187 DEFINE_SCHED_CLASS(fair
) = {
11189 .enqueue_task
= enqueue_task_fair
,
11190 .dequeue_task
= dequeue_task_fair
,
11191 .yield_task
= yield_task_fair
,
11192 .yield_to_task
= yield_to_task_fair
,
11194 .check_preempt_curr
= check_preempt_wakeup
,
11196 .pick_next_task
= __pick_next_task_fair
,
11197 .put_prev_task
= put_prev_task_fair
,
11198 .set_next_task
= set_next_task_fair
,
11201 .balance
= balance_fair
,
11202 .select_task_rq
= select_task_rq_fair
,
11203 .migrate_task_rq
= migrate_task_rq_fair
,
11205 .rq_online
= rq_online_fair
,
11206 .rq_offline
= rq_offline_fair
,
11208 .task_dead
= task_dead_fair
,
11209 .set_cpus_allowed
= set_cpus_allowed_common
,
11212 .task_tick
= task_tick_fair
,
11213 .task_fork
= task_fork_fair
,
11215 .prio_changed
= prio_changed_fair
,
11216 .switched_from
= switched_from_fair
,
11217 .switched_to
= switched_to_fair
,
11219 .get_rr_interval
= get_rr_interval_fair
,
11221 .update_curr
= update_curr_fair
,
11223 #ifdef CONFIG_FAIR_GROUP_SCHED
11224 .task_change_group
= task_change_group_fair
,
11227 #ifdef CONFIG_UCLAMP_TASK
11228 .uclamp_enabled
= 1,
11232 #ifdef CONFIG_SCHED_DEBUG
11233 void print_cfs_stats(struct seq_file
*m
, int cpu
)
11235 struct cfs_rq
*cfs_rq
, *pos
;
11238 for_each_leaf_cfs_rq_safe(cpu_rq(cpu
), cfs_rq
, pos
)
11239 print_cfs_rq(m
, cpu
, cfs_rq
);
11243 #ifdef CONFIG_NUMA_BALANCING
11244 void show_numa_stats(struct task_struct
*p
, struct seq_file
*m
)
11247 unsigned long tsf
= 0, tpf
= 0, gsf
= 0, gpf
= 0;
11248 struct numa_group
*ng
;
11251 ng
= rcu_dereference(p
->numa_group
);
11252 for_each_online_node(node
) {
11253 if (p
->numa_faults
) {
11254 tsf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
11255 tpf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
11258 gsf
= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)],
11259 gpf
= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
11261 print_numa_stats(m
, node
, tsf
, tpf
, gsf
, gpf
);
11265 #endif /* CONFIG_NUMA_BALANCING */
11266 #endif /* CONFIG_SCHED_DEBUG */
11268 __init
void init_sched_fair_class(void)
11271 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
11273 #ifdef CONFIG_NO_HZ_COMMON
11274 nohz
.next_balance
= jiffies
;
11275 nohz
.next_blocked
= jiffies
;
11276 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
11283 * Helper functions to facilitate extracting info from tracepoints.
11286 const struct sched_avg
*sched_trace_cfs_rq_avg(struct cfs_rq
*cfs_rq
)
11289 return cfs_rq
? &cfs_rq
->avg
: NULL
;
11294 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg
);
11296 char *sched_trace_cfs_rq_path(struct cfs_rq
*cfs_rq
, char *str
, int len
)
11300 strlcpy(str
, "(null)", len
);
11305 cfs_rq_tg_path(cfs_rq
, str
, len
);
11308 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path
);
11310 int sched_trace_cfs_rq_cpu(struct cfs_rq
*cfs_rq
)
11312 return cfs_rq
? cpu_of(rq_of(cfs_rq
)) : -1;
11314 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu
);
11316 const struct sched_avg
*sched_trace_rq_avg_rt(struct rq
*rq
)
11319 return rq
? &rq
->avg_rt
: NULL
;
11324 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt
);
11326 const struct sched_avg
*sched_trace_rq_avg_dl(struct rq
*rq
)
11329 return rq
? &rq
->avg_dl
: NULL
;
11334 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl
);
11336 const struct sched_avg
*sched_trace_rq_avg_irq(struct rq
*rq
)
11338 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
11339 return rq
? &rq
->avg_irq
: NULL
;
11344 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq
);
11346 int sched_trace_rq_cpu(struct rq
*rq
)
11348 return rq
? cpu_of(rq
) : -1;
11350 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu
);
11352 int sched_trace_rq_cpu_capacity(struct rq
*rq
)
11358 SCHED_CAPACITY_SCALE
11362 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity
);
11364 const struct cpumask
*sched_trace_rd_span(struct root_domain
*rd
)
11367 return rd
? rd
->span
: NULL
;
11372 EXPORT_SYMBOL_GPL(sched_trace_rd_span
);
11374 int sched_trace_rq_nr_running(struct rq
*rq
)
11376 return rq
? rq
->nr_running
: -1;
11378 EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running
);