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
, int force
)
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())
909 delta
= rq_clock(rq_of(cfs_rq
)) - schedstat_val(se
->statistics
.wait_start
);
911 if (entity_is_task(se
)) {
913 if (task_on_rq_migrating(p
)) {
915 * Preserve migrating task's wait time so wait_start
916 * time stamp can be adjusted to accumulate wait time
917 * prior to migration.
919 __schedstat_set(se
->statistics
.wait_start
, delta
);
922 trace_sched_stat_wait(p
, delta
);
925 __schedstat_set(se
->statistics
.wait_max
,
926 max(schedstat_val(se
->statistics
.wait_max
), delta
));
927 __schedstat_inc(se
->statistics
.wait_count
);
928 __schedstat_add(se
->statistics
.wait_sum
, delta
);
929 __schedstat_set(se
->statistics
.wait_start
, 0);
933 update_stats_enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
935 struct task_struct
*tsk
= NULL
;
936 u64 sleep_start
, block_start
;
938 if (!schedstat_enabled())
941 sleep_start
= schedstat_val(se
->statistics
.sleep_start
);
942 block_start
= schedstat_val(se
->statistics
.block_start
);
944 if (entity_is_task(se
))
948 u64 delta
= rq_clock(rq_of(cfs_rq
)) - sleep_start
;
953 if (unlikely(delta
> schedstat_val(se
->statistics
.sleep_max
)))
954 __schedstat_set(se
->statistics
.sleep_max
, delta
);
956 __schedstat_set(se
->statistics
.sleep_start
, 0);
957 __schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
960 account_scheduler_latency(tsk
, delta
>> 10, 1);
961 trace_sched_stat_sleep(tsk
, delta
);
965 u64 delta
= rq_clock(rq_of(cfs_rq
)) - block_start
;
970 if (unlikely(delta
> schedstat_val(se
->statistics
.block_max
)))
971 __schedstat_set(se
->statistics
.block_max
, delta
);
973 __schedstat_set(se
->statistics
.block_start
, 0);
974 __schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
977 if (tsk
->in_iowait
) {
978 __schedstat_add(se
->statistics
.iowait_sum
, delta
);
979 __schedstat_inc(se
->statistics
.iowait_count
);
980 trace_sched_stat_iowait(tsk
, delta
);
983 trace_sched_stat_blocked(tsk
, delta
);
986 * Blocking time is in units of nanosecs, so shift by
987 * 20 to get a milliseconds-range estimation of the
988 * amount of time that the task spent sleeping:
990 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
991 profile_hits(SLEEP_PROFILING
,
992 (void *)get_wchan(tsk
),
995 account_scheduler_latency(tsk
, delta
>> 10, 0);
1001 * Task is being enqueued - update stats:
1004 update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1006 if (!schedstat_enabled())
1010 * Are we enqueueing a waiting task? (for current tasks
1011 * a dequeue/enqueue event is a NOP)
1013 if (se
!= cfs_rq
->curr
)
1014 update_stats_wait_start(cfs_rq
, se
);
1016 if (flags
& ENQUEUE_WAKEUP
)
1017 update_stats_enqueue_sleeper(cfs_rq
, se
);
1021 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1024 if (!schedstat_enabled())
1028 * Mark the end of the wait period if dequeueing a
1031 if (se
!= cfs_rq
->curr
)
1032 update_stats_wait_end(cfs_rq
, se
);
1034 if ((flags
& DEQUEUE_SLEEP
) && entity_is_task(se
)) {
1035 struct task_struct
*tsk
= task_of(se
);
1037 if (tsk
->state
& TASK_INTERRUPTIBLE
)
1038 __schedstat_set(se
->statistics
.sleep_start
,
1039 rq_clock(rq_of(cfs_rq
)));
1040 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
1041 __schedstat_set(se
->statistics
.block_start
,
1042 rq_clock(rq_of(cfs_rq
)));
1047 * We are picking a new current task - update its stats:
1050 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1053 * We are starting a new run period:
1055 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
1058 /**************************************************
1059 * Scheduling class queueing methods:
1062 #ifdef CONFIG_NUMA_BALANCING
1064 * Approximate time to scan a full NUMA task in ms. The task scan period is
1065 * calculated based on the tasks virtual memory size and
1066 * numa_balancing_scan_size.
1068 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
1069 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
1071 /* Portion of address space to scan in MB */
1072 unsigned int sysctl_numa_balancing_scan_size
= 256;
1074 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1075 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
1078 refcount_t refcount
;
1080 spinlock_t lock
; /* nr_tasks, tasks */
1085 struct rcu_head rcu
;
1086 unsigned long total_faults
;
1087 unsigned long max_faults_cpu
;
1089 * Faults_cpu is used to decide whether memory should move
1090 * towards the CPU. As a consequence, these stats are weighted
1091 * more by CPU use than by memory faults.
1093 unsigned long *faults_cpu
;
1094 unsigned long faults
[];
1098 * For functions that can be called in multiple contexts that permit reading
1099 * ->numa_group (see struct task_struct for locking rules).
1101 static struct numa_group
*deref_task_numa_group(struct task_struct
*p
)
1103 return rcu_dereference_check(p
->numa_group
, p
== current
||
1104 (lockdep_is_held(&task_rq(p
)->lock
) && !READ_ONCE(p
->on_cpu
)));
1107 static struct numa_group
*deref_curr_numa_group(struct task_struct
*p
)
1109 return rcu_dereference_protected(p
->numa_group
, p
== current
);
1112 static inline unsigned long group_faults_priv(struct numa_group
*ng
);
1113 static inline unsigned long group_faults_shared(struct numa_group
*ng
);
1115 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
1117 unsigned long rss
= 0;
1118 unsigned long nr_scan_pages
;
1121 * Calculations based on RSS as non-present and empty pages are skipped
1122 * by the PTE scanner and NUMA hinting faults should be trapped based
1125 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
1126 rss
= get_mm_rss(p
->mm
);
1128 rss
= nr_scan_pages
;
1130 rss
= round_up(rss
, nr_scan_pages
);
1131 return rss
/ nr_scan_pages
;
1134 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1135 #define MAX_SCAN_WINDOW 2560
1137 static unsigned int task_scan_min(struct task_struct
*p
)
1139 unsigned int scan_size
= READ_ONCE(sysctl_numa_balancing_scan_size
);
1140 unsigned int scan
, floor
;
1141 unsigned int windows
= 1;
1143 if (scan_size
< MAX_SCAN_WINDOW
)
1144 windows
= MAX_SCAN_WINDOW
/ scan_size
;
1145 floor
= 1000 / windows
;
1147 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
1148 return max_t(unsigned int, floor
, scan
);
1151 static unsigned int task_scan_start(struct task_struct
*p
)
1153 unsigned long smin
= task_scan_min(p
);
1154 unsigned long period
= smin
;
1155 struct numa_group
*ng
;
1157 /* Scale the maximum scan period with the amount of shared memory. */
1159 ng
= rcu_dereference(p
->numa_group
);
1161 unsigned long shared
= group_faults_shared(ng
);
1162 unsigned long private = group_faults_priv(ng
);
1164 period
*= refcount_read(&ng
->refcount
);
1165 period
*= shared
+ 1;
1166 period
/= private + shared
+ 1;
1170 return max(smin
, period
);
1173 static unsigned int task_scan_max(struct task_struct
*p
)
1175 unsigned long smin
= task_scan_min(p
);
1177 struct numa_group
*ng
;
1179 /* Watch for min being lower than max due to floor calculations */
1180 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
1182 /* Scale the maximum scan period with the amount of shared memory. */
1183 ng
= deref_curr_numa_group(p
);
1185 unsigned long shared
= group_faults_shared(ng
);
1186 unsigned long private = group_faults_priv(ng
);
1187 unsigned long period
= smax
;
1189 period
*= refcount_read(&ng
->refcount
);
1190 period
*= shared
+ 1;
1191 period
/= private + shared
+ 1;
1193 smax
= max(smax
, period
);
1196 return max(smin
, smax
);
1199 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
1201 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= NUMA_NO_NODE
);
1202 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
1205 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
1207 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= NUMA_NO_NODE
);
1208 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
1211 /* Shared or private faults. */
1212 #define NR_NUMA_HINT_FAULT_TYPES 2
1214 /* Memory and CPU locality */
1215 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1217 /* Averaged statistics, and temporary buffers. */
1218 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1220 pid_t
task_numa_group_id(struct task_struct
*p
)
1222 struct numa_group
*ng
;
1226 ng
= rcu_dereference(p
->numa_group
);
1235 * The averaged statistics, shared & private, memory & CPU,
1236 * occupy the first half of the array. The second half of the
1237 * array is for current counters, which are averaged into the
1238 * first set by task_numa_placement.
1240 static inline int task_faults_idx(enum numa_faults_stats s
, int nid
, int priv
)
1242 return NR_NUMA_HINT_FAULT_TYPES
* (s
* nr_node_ids
+ nid
) + priv
;
1245 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
1247 if (!p
->numa_faults
)
1250 return p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1251 p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1254 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
1256 struct numa_group
*ng
= deref_task_numa_group(p
);
1261 return ng
->faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1262 ng
->faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1265 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
1267 return group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1268 group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1271 static inline unsigned long group_faults_priv(struct numa_group
*ng
)
1273 unsigned long faults
= 0;
1276 for_each_online_node(node
) {
1277 faults
+= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
1283 static inline unsigned long group_faults_shared(struct numa_group
*ng
)
1285 unsigned long faults
= 0;
1288 for_each_online_node(node
) {
1289 faults
+= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
1296 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1297 * considered part of a numa group's pseudo-interleaving set. Migrations
1298 * between these nodes are slowed down, to allow things to settle down.
1300 #define ACTIVE_NODE_FRACTION 3
1302 static bool numa_is_active_node(int nid
, struct numa_group
*ng
)
1304 return group_faults_cpu(ng
, nid
) * ACTIVE_NODE_FRACTION
> ng
->max_faults_cpu
;
1307 /* Handle placement on systems where not all nodes are directly connected. */
1308 static unsigned long score_nearby_nodes(struct task_struct
*p
, int nid
,
1309 int maxdist
, bool task
)
1311 unsigned long score
= 0;
1315 * All nodes are directly connected, and the same distance
1316 * from each other. No need for fancy placement algorithms.
1318 if (sched_numa_topology_type
== NUMA_DIRECT
)
1322 * This code is called for each node, introducing N^2 complexity,
1323 * which should be ok given the number of nodes rarely exceeds 8.
1325 for_each_online_node(node
) {
1326 unsigned long faults
;
1327 int dist
= node_distance(nid
, node
);
1330 * The furthest away nodes in the system are not interesting
1331 * for placement; nid was already counted.
1333 if (dist
== sched_max_numa_distance
|| node
== nid
)
1337 * On systems with a backplane NUMA topology, compare groups
1338 * of nodes, and move tasks towards the group with the most
1339 * memory accesses. When comparing two nodes at distance
1340 * "hoplimit", only nodes closer by than "hoplimit" are part
1341 * of each group. Skip other nodes.
1343 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1347 /* Add up the faults from nearby nodes. */
1349 faults
= task_faults(p
, node
);
1351 faults
= group_faults(p
, node
);
1354 * On systems with a glueless mesh NUMA topology, there are
1355 * no fixed "groups of nodes". Instead, nodes that are not
1356 * directly connected bounce traffic through intermediate
1357 * nodes; a numa_group can occupy any set of nodes.
1358 * The further away a node is, the less the faults count.
1359 * This seems to result in good task placement.
1361 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1362 faults
*= (sched_max_numa_distance
- dist
);
1363 faults
/= (sched_max_numa_distance
- LOCAL_DISTANCE
);
1373 * These return the fraction of accesses done by a particular task, or
1374 * task group, on a particular numa node. The group weight is given a
1375 * larger multiplier, in order to group tasks together that are almost
1376 * evenly spread out between numa nodes.
1378 static inline unsigned long task_weight(struct task_struct
*p
, int nid
,
1381 unsigned long faults
, total_faults
;
1383 if (!p
->numa_faults
)
1386 total_faults
= p
->total_numa_faults
;
1391 faults
= task_faults(p
, nid
);
1392 faults
+= score_nearby_nodes(p
, nid
, dist
, true);
1394 return 1000 * faults
/ total_faults
;
1397 static inline unsigned long group_weight(struct task_struct
*p
, int nid
,
1400 struct numa_group
*ng
= deref_task_numa_group(p
);
1401 unsigned long faults
, total_faults
;
1406 total_faults
= ng
->total_faults
;
1411 faults
= group_faults(p
, nid
);
1412 faults
+= score_nearby_nodes(p
, nid
, dist
, false);
1414 return 1000 * faults
/ total_faults
;
1417 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
1418 int src_nid
, int dst_cpu
)
1420 struct numa_group
*ng
= deref_curr_numa_group(p
);
1421 int dst_nid
= cpu_to_node(dst_cpu
);
1422 int last_cpupid
, this_cpupid
;
1424 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
1425 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
1428 * Allow first faults or private faults to migrate immediately early in
1429 * the lifetime of a task. The magic number 4 is based on waiting for
1430 * two full passes of the "multi-stage node selection" test that is
1433 if ((p
->numa_preferred_nid
== NUMA_NO_NODE
|| p
->numa_scan_seq
<= 4) &&
1434 (cpupid_pid_unset(last_cpupid
) || cpupid_match_pid(p
, last_cpupid
)))
1438 * Multi-stage node selection is used in conjunction with a periodic
1439 * migration fault to build a temporal task<->page relation. By using
1440 * a two-stage filter we remove short/unlikely relations.
1442 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1443 * a task's usage of a particular page (n_p) per total usage of this
1444 * page (n_t) (in a given time-span) to a probability.
1446 * Our periodic faults will sample this probability and getting the
1447 * same result twice in a row, given these samples are fully
1448 * independent, is then given by P(n)^2, provided our sample period
1449 * is sufficiently short compared to the usage pattern.
1451 * This quadric squishes small probabilities, making it less likely we
1452 * act on an unlikely task<->page relation.
1454 if (!cpupid_pid_unset(last_cpupid
) &&
1455 cpupid_to_nid(last_cpupid
) != dst_nid
)
1458 /* Always allow migrate on private faults */
1459 if (cpupid_match_pid(p
, last_cpupid
))
1462 /* A shared fault, but p->numa_group has not been set up yet. */
1467 * Destination node is much more heavily used than the source
1468 * node? Allow migration.
1470 if (group_faults_cpu(ng
, dst_nid
) > group_faults_cpu(ng
, src_nid
) *
1471 ACTIVE_NODE_FRACTION
)
1475 * Distribute memory according to CPU & memory use on each node,
1476 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1478 * faults_cpu(dst) 3 faults_cpu(src)
1479 * --------------- * - > ---------------
1480 * faults_mem(dst) 4 faults_mem(src)
1482 return group_faults_cpu(ng
, dst_nid
) * group_faults(p
, src_nid
) * 3 >
1483 group_faults_cpu(ng
, src_nid
) * group_faults(p
, dst_nid
) * 4;
1487 * 'numa_type' describes the node at the moment of load balancing.
1490 /* The node has spare capacity that can be used to run more tasks. */
1493 * The node is fully used and the tasks don't compete for more CPU
1494 * cycles. Nevertheless, some tasks might wait before running.
1498 * The node is overloaded and can't provide expected CPU cycles to all
1504 /* Cached statistics for all CPUs within a node */
1508 /* Total compute capacity of CPUs on a node */
1509 unsigned long compute_capacity
;
1510 unsigned int nr_running
;
1511 unsigned int weight
;
1512 enum numa_type node_type
;
1516 static inline bool is_core_idle(int cpu
)
1518 #ifdef CONFIG_SCHED_SMT
1521 for_each_cpu(sibling
, cpu_smt_mask(cpu
)) {
1533 struct task_numa_env
{
1534 struct task_struct
*p
;
1536 int src_cpu
, src_nid
;
1537 int dst_cpu
, dst_nid
;
1539 struct numa_stats src_stats
, dst_stats
;
1544 struct task_struct
*best_task
;
1549 static unsigned long cpu_load(struct rq
*rq
);
1550 static unsigned long cpu_util(int cpu
);
1551 static inline long adjust_numa_imbalance(int imbalance
, int src_nr_running
);
1554 numa_type
numa_classify(unsigned int imbalance_pct
,
1555 struct numa_stats
*ns
)
1557 if ((ns
->nr_running
> ns
->weight
) &&
1558 ((ns
->compute_capacity
* 100) < (ns
->util
* imbalance_pct
)))
1559 return node_overloaded
;
1561 if ((ns
->nr_running
< ns
->weight
) ||
1562 ((ns
->compute_capacity
* 100) > (ns
->util
* imbalance_pct
)))
1563 return node_has_spare
;
1565 return node_fully_busy
;
1568 #ifdef CONFIG_SCHED_SMT
1569 /* Forward declarations of select_idle_sibling helpers */
1570 static inline bool test_idle_cores(int cpu
, bool def
);
1571 static inline int numa_idle_core(int idle_core
, int cpu
)
1573 if (!static_branch_likely(&sched_smt_present
) ||
1574 idle_core
>= 0 || !test_idle_cores(cpu
, false))
1578 * Prefer cores instead of packing HT siblings
1579 * and triggering future load balancing.
1581 if (is_core_idle(cpu
))
1587 static inline int numa_idle_core(int idle_core
, int cpu
)
1594 * Gather all necessary information to make NUMA balancing placement
1595 * decisions that are compatible with standard load balancer. This
1596 * borrows code and logic from update_sg_lb_stats but sharing a
1597 * common implementation is impractical.
1599 static void update_numa_stats(struct task_numa_env
*env
,
1600 struct numa_stats
*ns
, int nid
,
1603 int cpu
, idle_core
= -1;
1605 memset(ns
, 0, sizeof(*ns
));
1609 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1610 struct rq
*rq
= cpu_rq(cpu
);
1612 ns
->load
+= cpu_load(rq
);
1613 ns
->util
+= cpu_util(cpu
);
1614 ns
->nr_running
+= rq
->cfs
.h_nr_running
;
1615 ns
->compute_capacity
+= capacity_of(cpu
);
1617 if (find_idle
&& !rq
->nr_running
&& idle_cpu(cpu
)) {
1618 if (READ_ONCE(rq
->numa_migrate_on
) ||
1619 !cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
))
1622 if (ns
->idle_cpu
== -1)
1625 idle_core
= numa_idle_core(idle_core
, cpu
);
1630 ns
->weight
= cpumask_weight(cpumask_of_node(nid
));
1632 ns
->node_type
= numa_classify(env
->imbalance_pct
, ns
);
1635 ns
->idle_cpu
= idle_core
;
1638 static void task_numa_assign(struct task_numa_env
*env
,
1639 struct task_struct
*p
, long imp
)
1641 struct rq
*rq
= cpu_rq(env
->dst_cpu
);
1643 /* Check if run-queue part of active NUMA balance. */
1644 if (env
->best_cpu
!= env
->dst_cpu
&& xchg(&rq
->numa_migrate_on
, 1)) {
1646 int start
= env
->dst_cpu
;
1648 /* Find alternative idle CPU. */
1649 for_each_cpu_wrap(cpu
, cpumask_of_node(env
->dst_nid
), start
) {
1650 if (cpu
== env
->best_cpu
|| !idle_cpu(cpu
) ||
1651 !cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
)) {
1656 rq
= cpu_rq(env
->dst_cpu
);
1657 if (!xchg(&rq
->numa_migrate_on
, 1))
1661 /* Failed to find an alternative idle CPU */
1667 * Clear previous best_cpu/rq numa-migrate flag, since task now
1668 * found a better CPU to move/swap.
1670 if (env
->best_cpu
!= -1 && env
->best_cpu
!= env
->dst_cpu
) {
1671 rq
= cpu_rq(env
->best_cpu
);
1672 WRITE_ONCE(rq
->numa_migrate_on
, 0);
1676 put_task_struct(env
->best_task
);
1681 env
->best_imp
= imp
;
1682 env
->best_cpu
= env
->dst_cpu
;
1685 static bool load_too_imbalanced(long src_load
, long dst_load
,
1686 struct task_numa_env
*env
)
1689 long orig_src_load
, orig_dst_load
;
1690 long src_capacity
, dst_capacity
;
1693 * The load is corrected for the CPU capacity available on each node.
1696 * ------------ vs ---------
1697 * src_capacity dst_capacity
1699 src_capacity
= env
->src_stats
.compute_capacity
;
1700 dst_capacity
= env
->dst_stats
.compute_capacity
;
1702 imb
= abs(dst_load
* src_capacity
- src_load
* dst_capacity
);
1704 orig_src_load
= env
->src_stats
.load
;
1705 orig_dst_load
= env
->dst_stats
.load
;
1707 old_imb
= abs(orig_dst_load
* src_capacity
- orig_src_load
* dst_capacity
);
1709 /* Would this change make things worse? */
1710 return (imb
> old_imb
);
1714 * Maximum NUMA importance can be 1998 (2*999);
1715 * SMALLIMP @ 30 would be close to 1998/64.
1716 * Used to deter task migration.
1721 * This checks if the overall compute and NUMA accesses of the system would
1722 * be improved if the source tasks was migrated to the target dst_cpu taking
1723 * into account that it might be best if task running on the dst_cpu should
1724 * be exchanged with the source task
1726 static bool task_numa_compare(struct task_numa_env
*env
,
1727 long taskimp
, long groupimp
, bool maymove
)
1729 struct numa_group
*cur_ng
, *p_ng
= deref_curr_numa_group(env
->p
);
1730 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1731 long imp
= p_ng
? groupimp
: taskimp
;
1732 struct task_struct
*cur
;
1733 long src_load
, dst_load
;
1734 int dist
= env
->dist
;
1737 bool stopsearch
= false;
1739 if (READ_ONCE(dst_rq
->numa_migrate_on
))
1743 cur
= rcu_dereference(dst_rq
->curr
);
1744 if (cur
&& ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
)))
1748 * Because we have preemption enabled we can get migrated around and
1749 * end try selecting ourselves (current == env->p) as a swap candidate.
1751 if (cur
== env
->p
) {
1757 if (maymove
&& moveimp
>= env
->best_imp
)
1763 /* Skip this swap candidate if cannot move to the source cpu. */
1764 if (!cpumask_test_cpu(env
->src_cpu
, cur
->cpus_ptr
))
1768 * Skip this swap candidate if it is not moving to its preferred
1769 * node and the best task is.
1771 if (env
->best_task
&&
1772 env
->best_task
->numa_preferred_nid
== env
->src_nid
&&
1773 cur
->numa_preferred_nid
!= env
->src_nid
) {
1778 * "imp" is the fault differential for the source task between the
1779 * source and destination node. Calculate the total differential for
1780 * the source task and potential destination task. The more negative
1781 * the value is, the more remote accesses that would be expected to
1782 * be incurred if the tasks were swapped.
1784 * If dst and source tasks are in the same NUMA group, or not
1785 * in any group then look only at task weights.
1787 cur_ng
= rcu_dereference(cur
->numa_group
);
1788 if (cur_ng
== p_ng
) {
1789 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1790 task_weight(cur
, env
->dst_nid
, dist
);
1792 * Add some hysteresis to prevent swapping the
1793 * tasks within a group over tiny differences.
1799 * Compare the group weights. If a task is all by itself
1800 * (not part of a group), use the task weight instead.
1803 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1804 group_weight(cur
, env
->dst_nid
, dist
);
1806 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1807 task_weight(cur
, env
->dst_nid
, dist
);
1810 /* Discourage picking a task already on its preferred node */
1811 if (cur
->numa_preferred_nid
== env
->dst_nid
)
1815 * Encourage picking a task that moves to its preferred node.
1816 * This potentially makes imp larger than it's maximum of
1817 * 1998 (see SMALLIMP and task_weight for why) but in this
1818 * case, it does not matter.
1820 if (cur
->numa_preferred_nid
== env
->src_nid
)
1823 if (maymove
&& moveimp
> imp
&& moveimp
> env
->best_imp
) {
1830 * Prefer swapping with a task moving to its preferred node over a
1833 if (env
->best_task
&& cur
->numa_preferred_nid
== env
->src_nid
&&
1834 env
->best_task
->numa_preferred_nid
!= env
->src_nid
) {
1839 * If the NUMA importance is less than SMALLIMP,
1840 * task migration might only result in ping pong
1841 * of tasks and also hurt performance due to cache
1844 if (imp
< SMALLIMP
|| imp
<= env
->best_imp
+ SMALLIMP
/ 2)
1848 * In the overloaded case, try and keep the load balanced.
1850 load
= task_h_load(env
->p
) - task_h_load(cur
);
1854 dst_load
= env
->dst_stats
.load
+ load
;
1855 src_load
= env
->src_stats
.load
- load
;
1857 if (load_too_imbalanced(src_load
, dst_load
, env
))
1861 /* Evaluate an idle CPU for a task numa move. */
1863 int cpu
= env
->dst_stats
.idle_cpu
;
1865 /* Nothing cached so current CPU went idle since the search. */
1870 * If the CPU is no longer truly idle and the previous best CPU
1871 * is, keep using it.
1873 if (!idle_cpu(cpu
) && env
->best_cpu
>= 0 &&
1874 idle_cpu(env
->best_cpu
)) {
1875 cpu
= env
->best_cpu
;
1881 task_numa_assign(env
, cur
, imp
);
1884 * If a move to idle is allowed because there is capacity or load
1885 * balance improves then stop the search. While a better swap
1886 * candidate may exist, a search is not free.
1888 if (maymove
&& !cur
&& env
->best_cpu
>= 0 && idle_cpu(env
->best_cpu
))
1892 * If a swap candidate must be identified and the current best task
1893 * moves its preferred node then stop the search.
1895 if (!maymove
&& env
->best_task
&&
1896 env
->best_task
->numa_preferred_nid
== env
->src_nid
) {
1905 static void task_numa_find_cpu(struct task_numa_env
*env
,
1906 long taskimp
, long groupimp
)
1908 bool maymove
= false;
1912 * If dst node has spare capacity, then check if there is an
1913 * imbalance that would be overruled by the load balancer.
1915 if (env
->dst_stats
.node_type
== node_has_spare
) {
1916 unsigned int imbalance
;
1917 int src_running
, dst_running
;
1920 * Would movement cause an imbalance? Note that if src has
1921 * more running tasks that the imbalance is ignored as the
1922 * move improves the imbalance from the perspective of the
1923 * CPU load balancer.
1925 src_running
= env
->src_stats
.nr_running
- 1;
1926 dst_running
= env
->dst_stats
.nr_running
+ 1;
1927 imbalance
= max(0, dst_running
- src_running
);
1928 imbalance
= adjust_numa_imbalance(imbalance
, src_running
);
1930 /* Use idle CPU if there is no imbalance */
1933 if (env
->dst_stats
.idle_cpu
>= 0) {
1934 env
->dst_cpu
= env
->dst_stats
.idle_cpu
;
1935 task_numa_assign(env
, NULL
, 0);
1940 long src_load
, dst_load
, load
;
1942 * If the improvement from just moving env->p direction is better
1943 * than swapping tasks around, check if a move is possible.
1945 load
= task_h_load(env
->p
);
1946 dst_load
= env
->dst_stats
.load
+ load
;
1947 src_load
= env
->src_stats
.load
- load
;
1948 maymove
= !load_too_imbalanced(src_load
, dst_load
, env
);
1951 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1952 /* Skip this CPU if the source task cannot migrate */
1953 if (!cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
))
1957 if (task_numa_compare(env
, taskimp
, groupimp
, maymove
))
1962 static int task_numa_migrate(struct task_struct
*p
)
1964 struct task_numa_env env
= {
1967 .src_cpu
= task_cpu(p
),
1968 .src_nid
= task_node(p
),
1970 .imbalance_pct
= 112,
1976 unsigned long taskweight
, groupweight
;
1977 struct sched_domain
*sd
;
1978 long taskimp
, groupimp
;
1979 struct numa_group
*ng
;
1984 * Pick the lowest SD_NUMA domain, as that would have the smallest
1985 * imbalance and would be the first to start moving tasks about.
1987 * And we want to avoid any moving of tasks about, as that would create
1988 * random movement of tasks -- counter the numa conditions we're trying
1992 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1994 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1998 * Cpusets can break the scheduler domain tree into smaller
1999 * balance domains, some of which do not cross NUMA boundaries.
2000 * Tasks that are "trapped" in such domains cannot be migrated
2001 * elsewhere, so there is no point in (re)trying.
2003 if (unlikely(!sd
)) {
2004 sched_setnuma(p
, task_node(p
));
2008 env
.dst_nid
= p
->numa_preferred_nid
;
2009 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
2010 taskweight
= task_weight(p
, env
.src_nid
, dist
);
2011 groupweight
= group_weight(p
, env
.src_nid
, dist
);
2012 update_numa_stats(&env
, &env
.src_stats
, env
.src_nid
, false);
2013 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
2014 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
2015 update_numa_stats(&env
, &env
.dst_stats
, env
.dst_nid
, true);
2017 /* Try to find a spot on the preferred nid. */
2018 task_numa_find_cpu(&env
, taskimp
, groupimp
);
2021 * Look at other nodes in these cases:
2022 * - there is no space available on the preferred_nid
2023 * - the task is part of a numa_group that is interleaved across
2024 * multiple NUMA nodes; in order to better consolidate the group,
2025 * we need to check other locations.
2027 ng
= deref_curr_numa_group(p
);
2028 if (env
.best_cpu
== -1 || (ng
&& ng
->active_nodes
> 1)) {
2029 for_each_online_node(nid
) {
2030 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
2033 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
2034 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
2036 taskweight
= task_weight(p
, env
.src_nid
, dist
);
2037 groupweight
= group_weight(p
, env
.src_nid
, dist
);
2040 /* Only consider nodes where both task and groups benefit */
2041 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
2042 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
2043 if (taskimp
< 0 && groupimp
< 0)
2048 update_numa_stats(&env
, &env
.dst_stats
, env
.dst_nid
, true);
2049 task_numa_find_cpu(&env
, taskimp
, groupimp
);
2054 * If the task is part of a workload that spans multiple NUMA nodes,
2055 * and is migrating into one of the workload's active nodes, remember
2056 * this node as the task's preferred numa node, so the workload can
2058 * A task that migrated to a second choice node will be better off
2059 * trying for a better one later. Do not set the preferred node here.
2062 if (env
.best_cpu
== -1)
2065 nid
= cpu_to_node(env
.best_cpu
);
2067 if (nid
!= p
->numa_preferred_nid
)
2068 sched_setnuma(p
, nid
);
2071 /* No better CPU than the current one was found. */
2072 if (env
.best_cpu
== -1) {
2073 trace_sched_stick_numa(p
, env
.src_cpu
, NULL
, -1);
2077 best_rq
= cpu_rq(env
.best_cpu
);
2078 if (env
.best_task
== NULL
) {
2079 ret
= migrate_task_to(p
, env
.best_cpu
);
2080 WRITE_ONCE(best_rq
->numa_migrate_on
, 0);
2082 trace_sched_stick_numa(p
, env
.src_cpu
, NULL
, env
.best_cpu
);
2086 ret
= migrate_swap(p
, env
.best_task
, env
.best_cpu
, env
.src_cpu
);
2087 WRITE_ONCE(best_rq
->numa_migrate_on
, 0);
2090 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_task
, env
.best_cpu
);
2091 put_task_struct(env
.best_task
);
2095 /* Attempt to migrate a task to a CPU on the preferred node. */
2096 static void numa_migrate_preferred(struct task_struct
*p
)
2098 unsigned long interval
= HZ
;
2100 /* This task has no NUMA fault statistics yet */
2101 if (unlikely(p
->numa_preferred_nid
== NUMA_NO_NODE
|| !p
->numa_faults
))
2104 /* Periodically retry migrating the task to the preferred node */
2105 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
2106 p
->numa_migrate_retry
= jiffies
+ interval
;
2108 /* Success if task is already running on preferred CPU */
2109 if (task_node(p
) == p
->numa_preferred_nid
)
2112 /* Otherwise, try migrate to a CPU on the preferred node */
2113 task_numa_migrate(p
);
2117 * Find out how many nodes on the workload is actively running on. Do this by
2118 * tracking the nodes from which NUMA hinting faults are triggered. This can
2119 * be different from the set of nodes where the workload's memory is currently
2122 static void numa_group_count_active_nodes(struct numa_group
*numa_group
)
2124 unsigned long faults
, max_faults
= 0;
2125 int nid
, active_nodes
= 0;
2127 for_each_online_node(nid
) {
2128 faults
= group_faults_cpu(numa_group
, nid
);
2129 if (faults
> max_faults
)
2130 max_faults
= faults
;
2133 for_each_online_node(nid
) {
2134 faults
= group_faults_cpu(numa_group
, nid
);
2135 if (faults
* ACTIVE_NODE_FRACTION
> max_faults
)
2139 numa_group
->max_faults_cpu
= max_faults
;
2140 numa_group
->active_nodes
= active_nodes
;
2144 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2145 * increments. The more local the fault statistics are, the higher the scan
2146 * period will be for the next scan window. If local/(local+remote) ratio is
2147 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2148 * the scan period will decrease. Aim for 70% local accesses.
2150 #define NUMA_PERIOD_SLOTS 10
2151 #define NUMA_PERIOD_THRESHOLD 7
2154 * Increase the scan period (slow down scanning) if the majority of
2155 * our memory is already on our local node, or if the majority of
2156 * the page accesses are shared with other processes.
2157 * Otherwise, decrease the scan period.
2159 static void update_task_scan_period(struct task_struct
*p
,
2160 unsigned long shared
, unsigned long private)
2162 unsigned int period_slot
;
2163 int lr_ratio
, ps_ratio
;
2166 unsigned long remote
= p
->numa_faults_locality
[0];
2167 unsigned long local
= p
->numa_faults_locality
[1];
2170 * If there were no record hinting faults then either the task is
2171 * completely idle or all activity is areas that are not of interest
2172 * to automatic numa balancing. Related to that, if there were failed
2173 * migration then it implies we are migrating too quickly or the local
2174 * node is overloaded. In either case, scan slower
2176 if (local
+ shared
== 0 || p
->numa_faults_locality
[2]) {
2177 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
2178 p
->numa_scan_period
<< 1);
2180 p
->mm
->numa_next_scan
= jiffies
+
2181 msecs_to_jiffies(p
->numa_scan_period
);
2187 * Prepare to scale scan period relative to the current period.
2188 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2189 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2190 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2192 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
2193 lr_ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
2194 ps_ratio
= (private * NUMA_PERIOD_SLOTS
) / (private + shared
);
2196 if (ps_ratio
>= NUMA_PERIOD_THRESHOLD
) {
2198 * Most memory accesses are local. There is no need to
2199 * do fast NUMA scanning, since memory is already local.
2201 int slot
= ps_ratio
- NUMA_PERIOD_THRESHOLD
;
2204 diff
= slot
* period_slot
;
2205 } else if (lr_ratio
>= NUMA_PERIOD_THRESHOLD
) {
2207 * Most memory accesses are shared with other tasks.
2208 * There is no point in continuing fast NUMA scanning,
2209 * since other tasks may just move the memory elsewhere.
2211 int slot
= lr_ratio
- NUMA_PERIOD_THRESHOLD
;
2214 diff
= slot
* period_slot
;
2217 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2218 * yet they are not on the local NUMA node. Speed up
2219 * NUMA scanning to get the memory moved over.
2221 int ratio
= max(lr_ratio
, ps_ratio
);
2222 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
2225 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
2226 task_scan_min(p
), task_scan_max(p
));
2227 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2231 * Get the fraction of time the task has been running since the last
2232 * NUMA placement cycle. The scheduler keeps similar statistics, but
2233 * decays those on a 32ms period, which is orders of magnitude off
2234 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2235 * stats only if the task is so new there are no NUMA statistics yet.
2237 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
2239 u64 runtime
, delta
, now
;
2240 /* Use the start of this time slice to avoid calculations. */
2241 now
= p
->se
.exec_start
;
2242 runtime
= p
->se
.sum_exec_runtime
;
2244 if (p
->last_task_numa_placement
) {
2245 delta
= runtime
- p
->last_sum_exec_runtime
;
2246 *period
= now
- p
->last_task_numa_placement
;
2248 /* Avoid time going backwards, prevent potential divide error: */
2249 if (unlikely((s64
)*period
< 0))
2252 delta
= p
->se
.avg
.load_sum
;
2253 *period
= LOAD_AVG_MAX
;
2256 p
->last_sum_exec_runtime
= runtime
;
2257 p
->last_task_numa_placement
= now
;
2263 * Determine the preferred nid for a task in a numa_group. This needs to
2264 * be done in a way that produces consistent results with group_weight,
2265 * otherwise workloads might not converge.
2267 static int preferred_group_nid(struct task_struct
*p
, int nid
)
2272 /* Direct connections between all NUMA nodes. */
2273 if (sched_numa_topology_type
== NUMA_DIRECT
)
2277 * On a system with glueless mesh NUMA topology, group_weight
2278 * scores nodes according to the number of NUMA hinting faults on
2279 * both the node itself, and on nearby nodes.
2281 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
2282 unsigned long score
, max_score
= 0;
2283 int node
, max_node
= nid
;
2285 dist
= sched_max_numa_distance
;
2287 for_each_online_node(node
) {
2288 score
= group_weight(p
, node
, dist
);
2289 if (score
> max_score
) {
2298 * Finding the preferred nid in a system with NUMA backplane
2299 * interconnect topology is more involved. The goal is to locate
2300 * tasks from numa_groups near each other in the system, and
2301 * untangle workloads from different sides of the system. This requires
2302 * searching down the hierarchy of node groups, recursively searching
2303 * inside the highest scoring group of nodes. The nodemask tricks
2304 * keep the complexity of the search down.
2306 nodes
= node_online_map
;
2307 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
2308 unsigned long max_faults
= 0;
2309 nodemask_t max_group
= NODE_MASK_NONE
;
2312 /* Are there nodes at this distance from each other? */
2313 if (!find_numa_distance(dist
))
2316 for_each_node_mask(a
, nodes
) {
2317 unsigned long faults
= 0;
2318 nodemask_t this_group
;
2319 nodes_clear(this_group
);
2321 /* Sum group's NUMA faults; includes a==b case. */
2322 for_each_node_mask(b
, nodes
) {
2323 if (node_distance(a
, b
) < dist
) {
2324 faults
+= group_faults(p
, b
);
2325 node_set(b
, this_group
);
2326 node_clear(b
, nodes
);
2330 /* Remember the top group. */
2331 if (faults
> max_faults
) {
2332 max_faults
= faults
;
2333 max_group
= this_group
;
2335 * subtle: at the smallest distance there is
2336 * just one node left in each "group", the
2337 * winner is the preferred nid.
2342 /* Next round, evaluate the nodes within max_group. */
2350 static void task_numa_placement(struct task_struct
*p
)
2352 int seq
, nid
, max_nid
= NUMA_NO_NODE
;
2353 unsigned long max_faults
= 0;
2354 unsigned long fault_types
[2] = { 0, 0 };
2355 unsigned long total_faults
;
2356 u64 runtime
, period
;
2357 spinlock_t
*group_lock
= NULL
;
2358 struct numa_group
*ng
;
2361 * The p->mm->numa_scan_seq field gets updated without
2362 * exclusive access. Use READ_ONCE() here to ensure
2363 * that the field is read in a single access:
2365 seq
= READ_ONCE(p
->mm
->numa_scan_seq
);
2366 if (p
->numa_scan_seq
== seq
)
2368 p
->numa_scan_seq
= seq
;
2369 p
->numa_scan_period_max
= task_scan_max(p
);
2371 total_faults
= p
->numa_faults_locality
[0] +
2372 p
->numa_faults_locality
[1];
2373 runtime
= numa_get_avg_runtime(p
, &period
);
2375 /* If the task is part of a group prevent parallel updates to group stats */
2376 ng
= deref_curr_numa_group(p
);
2378 group_lock
= &ng
->lock
;
2379 spin_lock_irq(group_lock
);
2382 /* Find the node with the highest number of faults */
2383 for_each_online_node(nid
) {
2384 /* Keep track of the offsets in numa_faults array */
2385 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
2386 unsigned long faults
= 0, group_faults
= 0;
2389 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
2390 long diff
, f_diff
, f_weight
;
2392 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
2393 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
2394 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
2395 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
2397 /* Decay existing window, copy faults since last scan */
2398 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
2399 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
2400 p
->numa_faults
[membuf_idx
] = 0;
2403 * Normalize the faults_from, so all tasks in a group
2404 * count according to CPU use, instead of by the raw
2405 * number of faults. Tasks with little runtime have
2406 * little over-all impact on throughput, and thus their
2407 * faults are less important.
2409 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
2410 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
2412 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
2413 p
->numa_faults
[cpubuf_idx
] = 0;
2415 p
->numa_faults
[mem_idx
] += diff
;
2416 p
->numa_faults
[cpu_idx
] += f_diff
;
2417 faults
+= p
->numa_faults
[mem_idx
];
2418 p
->total_numa_faults
+= diff
;
2421 * safe because we can only change our own group
2423 * mem_idx represents the offset for a given
2424 * nid and priv in a specific region because it
2425 * is at the beginning of the numa_faults array.
2427 ng
->faults
[mem_idx
] += diff
;
2428 ng
->faults_cpu
[mem_idx
] += f_diff
;
2429 ng
->total_faults
+= diff
;
2430 group_faults
+= ng
->faults
[mem_idx
];
2435 if (faults
> max_faults
) {
2436 max_faults
= faults
;
2439 } else if (group_faults
> max_faults
) {
2440 max_faults
= group_faults
;
2446 numa_group_count_active_nodes(ng
);
2447 spin_unlock_irq(group_lock
);
2448 max_nid
= preferred_group_nid(p
, max_nid
);
2452 /* Set the new preferred node */
2453 if (max_nid
!= p
->numa_preferred_nid
)
2454 sched_setnuma(p
, max_nid
);
2457 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
2460 static inline int get_numa_group(struct numa_group
*grp
)
2462 return refcount_inc_not_zero(&grp
->refcount
);
2465 static inline void put_numa_group(struct numa_group
*grp
)
2467 if (refcount_dec_and_test(&grp
->refcount
))
2468 kfree_rcu(grp
, rcu
);
2471 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
2474 struct numa_group
*grp
, *my_grp
;
2475 struct task_struct
*tsk
;
2477 int cpu
= cpupid_to_cpu(cpupid
);
2480 if (unlikely(!deref_curr_numa_group(p
))) {
2481 unsigned int size
= sizeof(struct numa_group
) +
2482 4*nr_node_ids
*sizeof(unsigned long);
2484 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
2488 refcount_set(&grp
->refcount
, 1);
2489 grp
->active_nodes
= 1;
2490 grp
->max_faults_cpu
= 0;
2491 spin_lock_init(&grp
->lock
);
2493 /* Second half of the array tracks nids where faults happen */
2494 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
2497 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2498 grp
->faults
[i
] = p
->numa_faults
[i
];
2500 grp
->total_faults
= p
->total_numa_faults
;
2503 rcu_assign_pointer(p
->numa_group
, grp
);
2507 tsk
= READ_ONCE(cpu_rq(cpu
)->curr
);
2509 if (!cpupid_match_pid(tsk
, cpupid
))
2512 grp
= rcu_dereference(tsk
->numa_group
);
2516 my_grp
= deref_curr_numa_group(p
);
2521 * Only join the other group if its bigger; if we're the bigger group,
2522 * the other task will join us.
2524 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
2528 * Tie-break on the grp address.
2530 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
2533 /* Always join threads in the same process. */
2534 if (tsk
->mm
== current
->mm
)
2537 /* Simple filter to avoid false positives due to PID collisions */
2538 if (flags
& TNF_SHARED
)
2541 /* Update priv based on whether false sharing was detected */
2544 if (join
&& !get_numa_group(grp
))
2552 BUG_ON(irqs_disabled());
2553 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
2555 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
2556 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
2557 grp
->faults
[i
] += p
->numa_faults
[i
];
2559 my_grp
->total_faults
-= p
->total_numa_faults
;
2560 grp
->total_faults
+= p
->total_numa_faults
;
2565 spin_unlock(&my_grp
->lock
);
2566 spin_unlock_irq(&grp
->lock
);
2568 rcu_assign_pointer(p
->numa_group
, grp
);
2570 put_numa_group(my_grp
);
2579 * Get rid of NUMA staticstics associated with a task (either current or dead).
2580 * If @final is set, the task is dead and has reached refcount zero, so we can
2581 * safely free all relevant data structures. Otherwise, there might be
2582 * concurrent reads from places like load balancing and procfs, and we should
2583 * reset the data back to default state without freeing ->numa_faults.
2585 void task_numa_free(struct task_struct
*p
, bool final
)
2587 /* safe: p either is current or is being freed by current */
2588 struct numa_group
*grp
= rcu_dereference_raw(p
->numa_group
);
2589 unsigned long *numa_faults
= p
->numa_faults
;
2590 unsigned long flags
;
2597 spin_lock_irqsave(&grp
->lock
, flags
);
2598 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2599 grp
->faults
[i
] -= p
->numa_faults
[i
];
2600 grp
->total_faults
-= p
->total_numa_faults
;
2603 spin_unlock_irqrestore(&grp
->lock
, flags
);
2604 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2605 put_numa_group(grp
);
2609 p
->numa_faults
= NULL
;
2612 p
->total_numa_faults
= 0;
2613 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2619 * Got a PROT_NONE fault for a page on @node.
2621 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2623 struct task_struct
*p
= current
;
2624 bool migrated
= flags
& TNF_MIGRATED
;
2625 int cpu_node
= task_node(current
);
2626 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2627 struct numa_group
*ng
;
2630 if (!static_branch_likely(&sched_numa_balancing
))
2633 /* for example, ksmd faulting in a user's mm */
2637 /* Allocate buffer to track faults on a per-node basis */
2638 if (unlikely(!p
->numa_faults
)) {
2639 int size
= sizeof(*p
->numa_faults
) *
2640 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2642 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2643 if (!p
->numa_faults
)
2646 p
->total_numa_faults
= 0;
2647 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2651 * First accesses are treated as private, otherwise consider accesses
2652 * to be private if the accessing pid has not changed
2654 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2657 priv
= cpupid_match_pid(p
, last_cpupid
);
2658 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2659 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2663 * If a workload spans multiple NUMA nodes, a shared fault that
2664 * occurs wholly within the set of nodes that the workload is
2665 * actively using should be counted as local. This allows the
2666 * scan rate to slow down when a workload has settled down.
2668 ng
= deref_curr_numa_group(p
);
2669 if (!priv
&& !local
&& ng
&& ng
->active_nodes
> 1 &&
2670 numa_is_active_node(cpu_node
, ng
) &&
2671 numa_is_active_node(mem_node
, ng
))
2675 * Retry to migrate task to preferred node periodically, in case it
2676 * previously failed, or the scheduler moved us.
2678 if (time_after(jiffies
, p
->numa_migrate_retry
)) {
2679 task_numa_placement(p
);
2680 numa_migrate_preferred(p
);
2684 p
->numa_pages_migrated
+= pages
;
2685 if (flags
& TNF_MIGRATE_FAIL
)
2686 p
->numa_faults_locality
[2] += pages
;
2688 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2689 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2690 p
->numa_faults_locality
[local
] += pages
;
2693 static void reset_ptenuma_scan(struct task_struct
*p
)
2696 * We only did a read acquisition of the mmap sem, so
2697 * p->mm->numa_scan_seq is written to without exclusive access
2698 * and the update is not guaranteed to be atomic. That's not
2699 * much of an issue though, since this is just used for
2700 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2701 * expensive, to avoid any form of compiler optimizations:
2703 WRITE_ONCE(p
->mm
->numa_scan_seq
, READ_ONCE(p
->mm
->numa_scan_seq
) + 1);
2704 p
->mm
->numa_scan_offset
= 0;
2708 * The expensive part of numa migration is done from task_work context.
2709 * Triggered from task_tick_numa().
2711 static void task_numa_work(struct callback_head
*work
)
2713 unsigned long migrate
, next_scan
, now
= jiffies
;
2714 struct task_struct
*p
= current
;
2715 struct mm_struct
*mm
= p
->mm
;
2716 u64 runtime
= p
->se
.sum_exec_runtime
;
2717 struct vm_area_struct
*vma
;
2718 unsigned long start
, end
;
2719 unsigned long nr_pte_updates
= 0;
2720 long pages
, virtpages
;
2722 SCHED_WARN_ON(p
!= container_of(work
, struct task_struct
, numa_work
));
2726 * Who cares about NUMA placement when they're dying.
2728 * NOTE: make sure not to dereference p->mm before this check,
2729 * exit_task_work() happens _after_ exit_mm() so we could be called
2730 * without p->mm even though we still had it when we enqueued this
2733 if (p
->flags
& PF_EXITING
)
2736 if (!mm
->numa_next_scan
) {
2737 mm
->numa_next_scan
= now
+
2738 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2742 * Enforce maximal scan/migration frequency..
2744 migrate
= mm
->numa_next_scan
;
2745 if (time_before(now
, migrate
))
2748 if (p
->numa_scan_period
== 0) {
2749 p
->numa_scan_period_max
= task_scan_max(p
);
2750 p
->numa_scan_period
= task_scan_start(p
);
2753 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2754 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2758 * Delay this task enough that another task of this mm will likely win
2759 * the next time around.
2761 p
->node_stamp
+= 2 * TICK_NSEC
;
2763 start
= mm
->numa_scan_offset
;
2764 pages
= sysctl_numa_balancing_scan_size
;
2765 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2766 virtpages
= pages
* 8; /* Scan up to this much virtual space */
2771 if (!mmap_read_trylock(mm
))
2773 vma
= find_vma(mm
, start
);
2775 reset_ptenuma_scan(p
);
2779 for (; vma
; vma
= vma
->vm_next
) {
2780 if (!vma_migratable(vma
) || !vma_policy_mof(vma
) ||
2781 is_vm_hugetlb_page(vma
) || (vma
->vm_flags
& VM_MIXEDMAP
)) {
2786 * Shared library pages mapped by multiple processes are not
2787 * migrated as it is expected they are cache replicated. Avoid
2788 * hinting faults in read-only file-backed mappings or the vdso
2789 * as migrating the pages will be of marginal benefit.
2792 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2796 * Skip inaccessible VMAs to avoid any confusion between
2797 * PROT_NONE and NUMA hinting ptes
2799 if (!vma_is_accessible(vma
))
2803 start
= max(start
, vma
->vm_start
);
2804 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2805 end
= min(end
, vma
->vm_end
);
2806 nr_pte_updates
= change_prot_numa(vma
, start
, end
);
2809 * Try to scan sysctl_numa_balancing_size worth of
2810 * hpages that have at least one present PTE that
2811 * is not already pte-numa. If the VMA contains
2812 * areas that are unused or already full of prot_numa
2813 * PTEs, scan up to virtpages, to skip through those
2817 pages
-= (end
- start
) >> PAGE_SHIFT
;
2818 virtpages
-= (end
- start
) >> PAGE_SHIFT
;
2821 if (pages
<= 0 || virtpages
<= 0)
2825 } while (end
!= vma
->vm_end
);
2830 * It is possible to reach the end of the VMA list but the last few
2831 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2832 * would find the !migratable VMA on the next scan but not reset the
2833 * scanner to the start so check it now.
2836 mm
->numa_scan_offset
= start
;
2838 reset_ptenuma_scan(p
);
2839 mmap_read_unlock(mm
);
2842 * Make sure tasks use at least 32x as much time to run other code
2843 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2844 * Usually update_task_scan_period slows down scanning enough; on an
2845 * overloaded system we need to limit overhead on a per task basis.
2847 if (unlikely(p
->se
.sum_exec_runtime
!= runtime
)) {
2848 u64 diff
= p
->se
.sum_exec_runtime
- runtime
;
2849 p
->node_stamp
+= 32 * diff
;
2853 void init_numa_balancing(unsigned long clone_flags
, struct task_struct
*p
)
2856 struct mm_struct
*mm
= p
->mm
;
2859 mm_users
= atomic_read(&mm
->mm_users
);
2860 if (mm_users
== 1) {
2861 mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2862 mm
->numa_scan_seq
= 0;
2866 p
->numa_scan_seq
= mm
? mm
->numa_scan_seq
: 0;
2867 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2868 /* Protect against double add, see task_tick_numa and task_numa_work */
2869 p
->numa_work
.next
= &p
->numa_work
;
2870 p
->numa_faults
= NULL
;
2871 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2872 p
->last_task_numa_placement
= 0;
2873 p
->last_sum_exec_runtime
= 0;
2875 init_task_work(&p
->numa_work
, task_numa_work
);
2877 /* New address space, reset the preferred nid */
2878 if (!(clone_flags
& CLONE_VM
)) {
2879 p
->numa_preferred_nid
= NUMA_NO_NODE
;
2884 * New thread, keep existing numa_preferred_nid which should be copied
2885 * already by arch_dup_task_struct but stagger when scans start.
2890 delay
= min_t(unsigned int, task_scan_max(current
),
2891 current
->numa_scan_period
* mm_users
* NSEC_PER_MSEC
);
2892 delay
+= 2 * TICK_NSEC
;
2893 p
->node_stamp
= delay
;
2898 * Drive the periodic memory faults..
2900 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2902 struct callback_head
*work
= &curr
->numa_work
;
2906 * We don't care about NUMA placement if we don't have memory.
2908 if ((curr
->flags
& (PF_EXITING
| PF_KTHREAD
)) || work
->next
!= work
)
2912 * Using runtime rather than walltime has the dual advantage that
2913 * we (mostly) drive the selection from busy threads and that the
2914 * task needs to have done some actual work before we bother with
2917 now
= curr
->se
.sum_exec_runtime
;
2918 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2920 if (now
> curr
->node_stamp
+ period
) {
2921 if (!curr
->node_stamp
)
2922 curr
->numa_scan_period
= task_scan_start(curr
);
2923 curr
->node_stamp
+= period
;
2925 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
))
2926 task_work_add(curr
, work
, true);
2930 static void update_scan_period(struct task_struct
*p
, int new_cpu
)
2932 int src_nid
= cpu_to_node(task_cpu(p
));
2933 int dst_nid
= cpu_to_node(new_cpu
);
2935 if (!static_branch_likely(&sched_numa_balancing
))
2938 if (!p
->mm
|| !p
->numa_faults
|| (p
->flags
& PF_EXITING
))
2941 if (src_nid
== dst_nid
)
2945 * Allow resets if faults have been trapped before one scan
2946 * has completed. This is most likely due to a new task that
2947 * is pulled cross-node due to wakeups or load balancing.
2949 if (p
->numa_scan_seq
) {
2951 * Avoid scan adjustments if moving to the preferred
2952 * node or if the task was not previously running on
2953 * the preferred node.
2955 if (dst_nid
== p
->numa_preferred_nid
||
2956 (p
->numa_preferred_nid
!= NUMA_NO_NODE
&&
2957 src_nid
!= p
->numa_preferred_nid
))
2961 p
->numa_scan_period
= task_scan_start(p
);
2965 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2969 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2973 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2977 static inline void update_scan_period(struct task_struct
*p
, int new_cpu
)
2981 #endif /* CONFIG_NUMA_BALANCING */
2984 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2986 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2988 if (entity_is_task(se
)) {
2989 struct rq
*rq
= rq_of(cfs_rq
);
2991 account_numa_enqueue(rq
, task_of(se
));
2992 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2995 cfs_rq
->nr_running
++;
2999 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3001 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
3003 if (entity_is_task(se
)) {
3004 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
3005 list_del_init(&se
->group_node
);
3008 cfs_rq
->nr_running
--;
3012 * Signed add and clamp on underflow.
3014 * Explicitly do a load-store to ensure the intermediate value never hits
3015 * memory. This allows lockless observations without ever seeing the negative
3018 #define add_positive(_ptr, _val) do { \
3019 typeof(_ptr) ptr = (_ptr); \
3020 typeof(_val) val = (_val); \
3021 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3025 if (val < 0 && res > var) \
3028 WRITE_ONCE(*ptr, res); \
3032 * Unsigned subtract and clamp on underflow.
3034 * Explicitly do a load-store to ensure the intermediate value never hits
3035 * memory. This allows lockless observations without ever seeing the negative
3038 #define sub_positive(_ptr, _val) do { \
3039 typeof(_ptr) ptr = (_ptr); \
3040 typeof(*ptr) val = (_val); \
3041 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3045 WRITE_ONCE(*ptr, res); \
3049 * Remove and clamp on negative, from a local variable.
3051 * A variant of sub_positive(), which does not use explicit load-store
3052 * and is thus optimized for local variable updates.
3054 #define lsub_positive(_ptr, _val) do { \
3055 typeof(_ptr) ptr = (_ptr); \
3056 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3061 enqueue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3063 cfs_rq
->avg
.load_avg
+= se
->avg
.load_avg
;
3064 cfs_rq
->avg
.load_sum
+= se_weight(se
) * se
->avg
.load_sum
;
3068 dequeue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3070 sub_positive(&cfs_rq
->avg
.load_avg
, se
->avg
.load_avg
);
3071 sub_positive(&cfs_rq
->avg
.load_sum
, se_weight(se
) * se
->avg
.load_sum
);
3075 enqueue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
3077 dequeue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
3080 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
3081 unsigned long weight
)
3084 /* commit outstanding execution time */
3085 if (cfs_rq
->curr
== se
)
3086 update_curr(cfs_rq
);
3087 account_entity_dequeue(cfs_rq
, se
);
3089 dequeue_load_avg(cfs_rq
, se
);
3091 update_load_set(&se
->load
, weight
);
3095 u32 divider
= get_pelt_divider(&se
->avg
);
3097 se
->avg
.load_avg
= div_u64(se_weight(se
) * se
->avg
.load_sum
, divider
);
3101 enqueue_load_avg(cfs_rq
, se
);
3103 account_entity_enqueue(cfs_rq
, se
);
3107 void reweight_task(struct task_struct
*p
, int prio
)
3109 struct sched_entity
*se
= &p
->se
;
3110 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3111 struct load_weight
*load
= &se
->load
;
3112 unsigned long weight
= scale_load(sched_prio_to_weight
[prio
]);
3114 reweight_entity(cfs_rq
, se
, weight
);
3115 load
->inv_weight
= sched_prio_to_wmult
[prio
];
3118 #ifdef CONFIG_FAIR_GROUP_SCHED
3121 * All this does is approximate the hierarchical proportion which includes that
3122 * global sum we all love to hate.
3124 * That is, the weight of a group entity, is the proportional share of the
3125 * group weight based on the group runqueue weights. That is:
3127 * tg->weight * grq->load.weight
3128 * ge->load.weight = ----------------------------- (1)
3129 * \Sum grq->load.weight
3131 * Now, because computing that sum is prohibitively expensive to compute (been
3132 * there, done that) we approximate it with this average stuff. The average
3133 * moves slower and therefore the approximation is cheaper and more stable.
3135 * So instead of the above, we substitute:
3137 * grq->load.weight -> grq->avg.load_avg (2)
3139 * which yields the following:
3141 * tg->weight * grq->avg.load_avg
3142 * ge->load.weight = ------------------------------ (3)
3145 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3147 * That is shares_avg, and it is right (given the approximation (2)).
3149 * The problem with it is that because the average is slow -- it was designed
3150 * to be exactly that of course -- this leads to transients in boundary
3151 * conditions. In specific, the case where the group was idle and we start the
3152 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3153 * yielding bad latency etc..
3155 * Now, in that special case (1) reduces to:
3157 * tg->weight * grq->load.weight
3158 * ge->load.weight = ----------------------------- = tg->weight (4)
3161 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3163 * So what we do is modify our approximation (3) to approach (4) in the (near)
3168 * tg->weight * grq->load.weight
3169 * --------------------------------------------------- (5)
3170 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3172 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3173 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3176 * tg->weight * grq->load.weight
3177 * ge->load.weight = ----------------------------- (6)
3182 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3183 * max(grq->load.weight, grq->avg.load_avg)
3185 * And that is shares_weight and is icky. In the (near) UP case it approaches
3186 * (4) while in the normal case it approaches (3). It consistently
3187 * overestimates the ge->load.weight and therefore:
3189 * \Sum ge->load.weight >= tg->weight
3193 static long calc_group_shares(struct cfs_rq
*cfs_rq
)
3195 long tg_weight
, tg_shares
, load
, shares
;
3196 struct task_group
*tg
= cfs_rq
->tg
;
3198 tg_shares
= READ_ONCE(tg
->shares
);
3200 load
= max(scale_load_down(cfs_rq
->load
.weight
), cfs_rq
->avg
.load_avg
);
3202 tg_weight
= atomic_long_read(&tg
->load_avg
);
3204 /* Ensure tg_weight >= load */
3205 tg_weight
-= cfs_rq
->tg_load_avg_contrib
;
3208 shares
= (tg_shares
* load
);
3210 shares
/= tg_weight
;
3213 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3214 * of a group with small tg->shares value. It is a floor value which is
3215 * assigned as a minimum load.weight to the sched_entity representing
3216 * the group on a CPU.
3218 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3219 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3220 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3221 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3224 return clamp_t(long, shares
, MIN_SHARES
, tg_shares
);
3226 #endif /* CONFIG_SMP */
3228 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
3231 * Recomputes the group entity based on the current state of its group
3234 static void update_cfs_group(struct sched_entity
*se
)
3236 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3242 if (throttled_hierarchy(gcfs_rq
))
3246 shares
= READ_ONCE(gcfs_rq
->tg
->shares
);
3248 if (likely(se
->load
.weight
== shares
))
3251 shares
= calc_group_shares(gcfs_rq
);
3254 reweight_entity(cfs_rq_of(se
), se
, shares
);
3257 #else /* CONFIG_FAIR_GROUP_SCHED */
3258 static inline void update_cfs_group(struct sched_entity
*se
)
3261 #endif /* CONFIG_FAIR_GROUP_SCHED */
3263 static inline void cfs_rq_util_change(struct cfs_rq
*cfs_rq
, int flags
)
3265 struct rq
*rq
= rq_of(cfs_rq
);
3267 if (&rq
->cfs
== cfs_rq
) {
3269 * There are a few boundary cases this might miss but it should
3270 * get called often enough that that should (hopefully) not be
3273 * It will not get called when we go idle, because the idle
3274 * thread is a different class (!fair), nor will the utilization
3275 * number include things like RT tasks.
3277 * As is, the util number is not freq-invariant (we'd have to
3278 * implement arch_scale_freq_capacity() for that).
3282 cpufreq_update_util(rq
, flags
);
3287 #ifdef CONFIG_FAIR_GROUP_SCHED
3289 * update_tg_load_avg - update the tg's load avg
3290 * @cfs_rq: the cfs_rq whose avg changed
3291 * @force: update regardless of how small the difference
3293 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3294 * However, because tg->load_avg is a global value there are performance
3297 * In order to avoid having to look at the other cfs_rq's, we use a
3298 * differential update where we store the last value we propagated. This in
3299 * turn allows skipping updates if the differential is 'small'.
3301 * Updating tg's load_avg is necessary before update_cfs_share().
3303 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
)
3305 long delta
= cfs_rq
->avg
.load_avg
- cfs_rq
->tg_load_avg_contrib
;
3308 * No need to update load_avg for root_task_group as it is not used.
3310 if (cfs_rq
->tg
== &root_task_group
)
3313 if (force
|| abs(delta
) > cfs_rq
->tg_load_avg_contrib
/ 64) {
3314 atomic_long_add(delta
, &cfs_rq
->tg
->load_avg
);
3315 cfs_rq
->tg_load_avg_contrib
= cfs_rq
->avg
.load_avg
;
3320 * Called within set_task_rq() right before setting a task's CPU. The
3321 * caller only guarantees p->pi_lock is held; no other assumptions,
3322 * including the state of rq->lock, should be made.
3324 void set_task_rq_fair(struct sched_entity
*se
,
3325 struct cfs_rq
*prev
, struct cfs_rq
*next
)
3327 u64 p_last_update_time
;
3328 u64 n_last_update_time
;
3330 if (!sched_feat(ATTACH_AGE_LOAD
))
3334 * We are supposed to update the task to "current" time, then its up to
3335 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3336 * getting what current time is, so simply throw away the out-of-date
3337 * time. This will result in the wakee task is less decayed, but giving
3338 * the wakee more load sounds not bad.
3340 if (!(se
->avg
.last_update_time
&& prev
))
3343 #ifndef CONFIG_64BIT
3345 u64 p_last_update_time_copy
;
3346 u64 n_last_update_time_copy
;
3349 p_last_update_time_copy
= prev
->load_last_update_time_copy
;
3350 n_last_update_time_copy
= next
->load_last_update_time_copy
;
3354 p_last_update_time
= prev
->avg
.last_update_time
;
3355 n_last_update_time
= next
->avg
.last_update_time
;
3357 } while (p_last_update_time
!= p_last_update_time_copy
||
3358 n_last_update_time
!= n_last_update_time_copy
);
3361 p_last_update_time
= prev
->avg
.last_update_time
;
3362 n_last_update_time
= next
->avg
.last_update_time
;
3364 __update_load_avg_blocked_se(p_last_update_time
, se
);
3365 se
->avg
.last_update_time
= n_last_update_time
;
3370 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3371 * propagate its contribution. The key to this propagation is the invariant
3372 * that for each group:
3374 * ge->avg == grq->avg (1)
3376 * _IFF_ we look at the pure running and runnable sums. Because they
3377 * represent the very same entity, just at different points in the hierarchy.
3379 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3380 * and simply copies the running/runnable sum over (but still wrong, because
3381 * the group entity and group rq do not have their PELT windows aligned).
3383 * However, update_tg_cfs_load() is more complex. So we have:
3385 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3387 * And since, like util, the runnable part should be directly transferable,
3388 * the following would _appear_ to be the straight forward approach:
3390 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3392 * And per (1) we have:
3394 * ge->avg.runnable_avg == grq->avg.runnable_avg
3398 * ge->load.weight * grq->avg.load_avg
3399 * ge->avg.load_avg = ----------------------------------- (4)
3402 * Except that is wrong!
3404 * Because while for entities historical weight is not important and we
3405 * really only care about our future and therefore can consider a pure
3406 * runnable sum, runqueues can NOT do this.
3408 * We specifically want runqueues to have a load_avg that includes
3409 * historical weights. Those represent the blocked load, the load we expect
3410 * to (shortly) return to us. This only works by keeping the weights as
3411 * integral part of the sum. We therefore cannot decompose as per (3).
3413 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3414 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3415 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3416 * runnable section of these tasks overlap (or not). If they were to perfectly
3417 * align the rq as a whole would be runnable 2/3 of the time. If however we
3418 * always have at least 1 runnable task, the rq as a whole is always runnable.
3420 * So we'll have to approximate.. :/
3422 * Given the constraint:
3424 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3426 * We can construct a rule that adds runnable to a rq by assuming minimal
3429 * On removal, we'll assume each task is equally runnable; which yields:
3431 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3433 * XXX: only do this for the part of runnable > running ?
3438 update_tg_cfs_util(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3440 long delta
= gcfs_rq
->avg
.util_avg
- se
->avg
.util_avg
;
3443 /* Nothing to update */
3448 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3449 * See ___update_load_avg() for details.
3451 divider
= get_pelt_divider(&cfs_rq
->avg
);
3453 /* Set new sched_entity's utilization */
3454 se
->avg
.util_avg
= gcfs_rq
->avg
.util_avg
;
3455 se
->avg
.util_sum
= se
->avg
.util_avg
* divider
;
3457 /* Update parent cfs_rq utilization */
3458 add_positive(&cfs_rq
->avg
.util_avg
, delta
);
3459 cfs_rq
->avg
.util_sum
= cfs_rq
->avg
.util_avg
* divider
;
3463 update_tg_cfs_runnable(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3465 long delta
= gcfs_rq
->avg
.runnable_avg
- se
->avg
.runnable_avg
;
3468 /* Nothing to update */
3473 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3474 * See ___update_load_avg() for details.
3476 divider
= get_pelt_divider(&cfs_rq
->avg
);
3478 /* Set new sched_entity's runnable */
3479 se
->avg
.runnable_avg
= gcfs_rq
->avg
.runnable_avg
;
3480 se
->avg
.runnable_sum
= se
->avg
.runnable_avg
* divider
;
3482 /* Update parent cfs_rq runnable */
3483 add_positive(&cfs_rq
->avg
.runnable_avg
, delta
);
3484 cfs_rq
->avg
.runnable_sum
= cfs_rq
->avg
.runnable_avg
* divider
;
3488 update_tg_cfs_load(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3490 long delta_avg
, running_sum
, runnable_sum
= gcfs_rq
->prop_runnable_sum
;
3491 unsigned long load_avg
;
3499 gcfs_rq
->prop_runnable_sum
= 0;
3502 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3503 * See ___update_load_avg() for details.
3505 divider
= get_pelt_divider(&cfs_rq
->avg
);
3507 if (runnable_sum
>= 0) {
3509 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3510 * the CPU is saturated running == runnable.
3512 runnable_sum
+= se
->avg
.load_sum
;
3513 runnable_sum
= min_t(long, runnable_sum
, divider
);
3516 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3517 * assuming all tasks are equally runnable.
3519 if (scale_load_down(gcfs_rq
->load
.weight
)) {
3520 load_sum
= div_s64(gcfs_rq
->avg
.load_sum
,
3521 scale_load_down(gcfs_rq
->load
.weight
));
3524 /* But make sure to not inflate se's runnable */
3525 runnable_sum
= min(se
->avg
.load_sum
, load_sum
);
3529 * runnable_sum can't be lower than running_sum
3530 * Rescale running sum to be in the same range as runnable sum
3531 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3532 * runnable_sum is in [0 : LOAD_AVG_MAX]
3534 running_sum
= se
->avg
.util_sum
>> SCHED_CAPACITY_SHIFT
;
3535 runnable_sum
= max(runnable_sum
, running_sum
);
3537 load_sum
= (s64
)se_weight(se
) * runnable_sum
;
3538 load_avg
= div_s64(load_sum
, divider
);
3540 delta_sum
= load_sum
- (s64
)se_weight(se
) * se
->avg
.load_sum
;
3541 delta_avg
= load_avg
- se
->avg
.load_avg
;
3543 se
->avg
.load_sum
= runnable_sum
;
3544 se
->avg
.load_avg
= load_avg
;
3545 add_positive(&cfs_rq
->avg
.load_avg
, delta_avg
);
3546 add_positive(&cfs_rq
->avg
.load_sum
, delta_sum
);
3549 static inline void add_tg_cfs_propagate(struct cfs_rq
*cfs_rq
, long runnable_sum
)
3551 cfs_rq
->propagate
= 1;
3552 cfs_rq
->prop_runnable_sum
+= runnable_sum
;
3555 /* Update task and its cfs_rq load average */
3556 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3558 struct cfs_rq
*cfs_rq
, *gcfs_rq
;
3560 if (entity_is_task(se
))
3563 gcfs_rq
= group_cfs_rq(se
);
3564 if (!gcfs_rq
->propagate
)
3567 gcfs_rq
->propagate
= 0;
3569 cfs_rq
= cfs_rq_of(se
);
3571 add_tg_cfs_propagate(cfs_rq
, gcfs_rq
->prop_runnable_sum
);
3573 update_tg_cfs_util(cfs_rq
, se
, gcfs_rq
);
3574 update_tg_cfs_runnable(cfs_rq
, se
, gcfs_rq
);
3575 update_tg_cfs_load(cfs_rq
, se
, gcfs_rq
);
3577 trace_pelt_cfs_tp(cfs_rq
);
3578 trace_pelt_se_tp(se
);
3584 * Check if we need to update the load and the utilization of a blocked
3587 static inline bool skip_blocked_update(struct sched_entity
*se
)
3589 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3592 * If sched_entity still have not zero load or utilization, we have to
3595 if (se
->avg
.load_avg
|| se
->avg
.util_avg
)
3599 * If there is a pending propagation, we have to update the load and
3600 * the utilization of the sched_entity:
3602 if (gcfs_rq
->propagate
)
3606 * Otherwise, the load and the utilization of the sched_entity is
3607 * already zero and there is no pending propagation, so it will be a
3608 * waste of time to try to decay it:
3613 #else /* CONFIG_FAIR_GROUP_SCHED */
3615 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
) {}
3617 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3622 static inline void add_tg_cfs_propagate(struct cfs_rq
*cfs_rq
, long runnable_sum
) {}
3624 #endif /* CONFIG_FAIR_GROUP_SCHED */
3627 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3628 * @now: current time, as per cfs_rq_clock_pelt()
3629 * @cfs_rq: cfs_rq to update
3631 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3632 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3633 * post_init_entity_util_avg().
3635 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3637 * Returns true if the load decayed or we removed load.
3639 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3640 * call update_tg_load_avg() when this function returns true.
3643 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
)
3645 unsigned long removed_load
= 0, removed_util
= 0, removed_runnable
= 0;
3646 struct sched_avg
*sa
= &cfs_rq
->avg
;
3649 if (cfs_rq
->removed
.nr
) {
3651 u32 divider
= get_pelt_divider(&cfs_rq
->avg
);
3653 raw_spin_lock(&cfs_rq
->removed
.lock
);
3654 swap(cfs_rq
->removed
.util_avg
, removed_util
);
3655 swap(cfs_rq
->removed
.load_avg
, removed_load
);
3656 swap(cfs_rq
->removed
.runnable_avg
, removed_runnable
);
3657 cfs_rq
->removed
.nr
= 0;
3658 raw_spin_unlock(&cfs_rq
->removed
.lock
);
3661 sub_positive(&sa
->load_avg
, r
);
3662 sub_positive(&sa
->load_sum
, r
* divider
);
3665 sub_positive(&sa
->util_avg
, r
);
3666 sub_positive(&sa
->util_sum
, r
* divider
);
3668 r
= removed_runnable
;
3669 sub_positive(&sa
->runnable_avg
, r
);
3670 sub_positive(&sa
->runnable_sum
, r
* divider
);
3673 * removed_runnable is the unweighted version of removed_load so we
3674 * can use it to estimate removed_load_sum.
3676 add_tg_cfs_propagate(cfs_rq
,
3677 -(long)(removed_runnable
* divider
) >> SCHED_CAPACITY_SHIFT
);
3682 decayed
|= __update_load_avg_cfs_rq(now
, cfs_rq
);
3684 #ifndef CONFIG_64BIT
3686 cfs_rq
->load_last_update_time_copy
= sa
->last_update_time
;
3693 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3694 * @cfs_rq: cfs_rq to attach to
3695 * @se: sched_entity to attach
3697 * Must call update_cfs_rq_load_avg() before this, since we rely on
3698 * cfs_rq->avg.last_update_time being current.
3700 static void attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3703 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3704 * See ___update_load_avg() for details.
3706 u32 divider
= get_pelt_divider(&cfs_rq
->avg
);
3709 * When we attach the @se to the @cfs_rq, we must align the decay
3710 * window because without that, really weird and wonderful things can
3715 se
->avg
.last_update_time
= cfs_rq
->avg
.last_update_time
;
3716 se
->avg
.period_contrib
= cfs_rq
->avg
.period_contrib
;
3719 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3720 * period_contrib. This isn't strictly correct, but since we're
3721 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3724 se
->avg
.util_sum
= se
->avg
.util_avg
* divider
;
3726 se
->avg
.runnable_sum
= se
->avg
.runnable_avg
* divider
;
3728 se
->avg
.load_sum
= divider
;
3729 if (se_weight(se
)) {
3731 div_u64(se
->avg
.load_avg
* se
->avg
.load_sum
, se_weight(se
));
3734 enqueue_load_avg(cfs_rq
, se
);
3735 cfs_rq
->avg
.util_avg
+= se
->avg
.util_avg
;
3736 cfs_rq
->avg
.util_sum
+= se
->avg
.util_sum
;
3737 cfs_rq
->avg
.runnable_avg
+= se
->avg
.runnable_avg
;
3738 cfs_rq
->avg
.runnable_sum
+= se
->avg
.runnable_sum
;
3740 add_tg_cfs_propagate(cfs_rq
, se
->avg
.load_sum
);
3742 cfs_rq_util_change(cfs_rq
, 0);
3744 trace_pelt_cfs_tp(cfs_rq
);
3748 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3749 * @cfs_rq: cfs_rq to detach from
3750 * @se: sched_entity to detach
3752 * Must call update_cfs_rq_load_avg() before this, since we rely on
3753 * cfs_rq->avg.last_update_time being current.
3755 static void detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3757 dequeue_load_avg(cfs_rq
, se
);
3758 sub_positive(&cfs_rq
->avg
.util_avg
, se
->avg
.util_avg
);
3759 sub_positive(&cfs_rq
->avg
.util_sum
, se
->avg
.util_sum
);
3760 sub_positive(&cfs_rq
->avg
.runnable_avg
, se
->avg
.runnable_avg
);
3761 sub_positive(&cfs_rq
->avg
.runnable_sum
, se
->avg
.runnable_sum
);
3763 add_tg_cfs_propagate(cfs_rq
, -se
->avg
.load_sum
);
3765 cfs_rq_util_change(cfs_rq
, 0);
3767 trace_pelt_cfs_tp(cfs_rq
);
3771 * Optional action to be done while updating the load average
3773 #define UPDATE_TG 0x1
3774 #define SKIP_AGE_LOAD 0x2
3775 #define DO_ATTACH 0x4
3777 /* Update task and its cfs_rq load average */
3778 static inline void update_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3780 u64 now
= cfs_rq_clock_pelt(cfs_rq
);
3784 * Track task load average for carrying it to new CPU after migrated, and
3785 * track group sched_entity load average for task_h_load calc in migration
3787 if (se
->avg
.last_update_time
&& !(flags
& SKIP_AGE_LOAD
))
3788 __update_load_avg_se(now
, cfs_rq
, se
);
3790 decayed
= update_cfs_rq_load_avg(now
, cfs_rq
);
3791 decayed
|= propagate_entity_load_avg(se
);
3793 if (!se
->avg
.last_update_time
&& (flags
& DO_ATTACH
)) {
3796 * DO_ATTACH means we're here from enqueue_entity().
3797 * !last_update_time means we've passed through
3798 * migrate_task_rq_fair() indicating we migrated.
3800 * IOW we're enqueueing a task on a new CPU.
3802 attach_entity_load_avg(cfs_rq
, se
);
3803 update_tg_load_avg(cfs_rq
, 0);
3805 } else if (decayed
) {
3806 cfs_rq_util_change(cfs_rq
, 0);
3808 if (flags
& UPDATE_TG
)
3809 update_tg_load_avg(cfs_rq
, 0);
3813 #ifndef CONFIG_64BIT
3814 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3816 u64 last_update_time_copy
;
3817 u64 last_update_time
;
3820 last_update_time_copy
= cfs_rq
->load_last_update_time_copy
;
3822 last_update_time
= cfs_rq
->avg
.last_update_time
;
3823 } while (last_update_time
!= last_update_time_copy
);
3825 return last_update_time
;
3828 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3830 return cfs_rq
->avg
.last_update_time
;
3835 * Synchronize entity load avg of dequeued entity without locking
3838 static void sync_entity_load_avg(struct sched_entity
*se
)
3840 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3841 u64 last_update_time
;
3843 last_update_time
= cfs_rq_last_update_time(cfs_rq
);
3844 __update_load_avg_blocked_se(last_update_time
, se
);
3848 * Task first catches up with cfs_rq, and then subtract
3849 * itself from the cfs_rq (task must be off the queue now).
3851 static void remove_entity_load_avg(struct sched_entity
*se
)
3853 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3854 unsigned long flags
;
3857 * tasks cannot exit without having gone through wake_up_new_task() ->
3858 * post_init_entity_util_avg() which will have added things to the
3859 * cfs_rq, so we can remove unconditionally.
3862 sync_entity_load_avg(se
);
3864 raw_spin_lock_irqsave(&cfs_rq
->removed
.lock
, flags
);
3865 ++cfs_rq
->removed
.nr
;
3866 cfs_rq
->removed
.util_avg
+= se
->avg
.util_avg
;
3867 cfs_rq
->removed
.load_avg
+= se
->avg
.load_avg
;
3868 cfs_rq
->removed
.runnable_avg
+= se
->avg
.runnable_avg
;
3869 raw_spin_unlock_irqrestore(&cfs_rq
->removed
.lock
, flags
);
3872 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq
*cfs_rq
)
3874 return cfs_rq
->avg
.runnable_avg
;
3877 static inline unsigned long cfs_rq_load_avg(struct cfs_rq
*cfs_rq
)
3879 return cfs_rq
->avg
.load_avg
;
3882 static int newidle_balance(struct rq
*this_rq
, struct rq_flags
*rf
);
3884 static inline unsigned long task_util(struct task_struct
*p
)
3886 return READ_ONCE(p
->se
.avg
.util_avg
);
3889 static inline unsigned long _task_util_est(struct task_struct
*p
)
3891 struct util_est ue
= READ_ONCE(p
->se
.avg
.util_est
);
3893 return (max(ue
.ewma
, ue
.enqueued
) | UTIL_AVG_UNCHANGED
);
3896 static inline unsigned long task_util_est(struct task_struct
*p
)
3898 return max(task_util(p
), _task_util_est(p
));
3901 #ifdef CONFIG_UCLAMP_TASK
3902 static inline unsigned long uclamp_task_util(struct task_struct
*p
)
3904 return clamp(task_util_est(p
),
3905 uclamp_eff_value(p
, UCLAMP_MIN
),
3906 uclamp_eff_value(p
, UCLAMP_MAX
));
3909 static inline unsigned long uclamp_task_util(struct task_struct
*p
)
3911 return task_util_est(p
);
3915 static inline void util_est_enqueue(struct cfs_rq
*cfs_rq
,
3916 struct task_struct
*p
)
3918 unsigned int enqueued
;
3920 if (!sched_feat(UTIL_EST
))
3923 /* Update root cfs_rq's estimated utilization */
3924 enqueued
= cfs_rq
->avg
.util_est
.enqueued
;
3925 enqueued
+= _task_util_est(p
);
3926 WRITE_ONCE(cfs_rq
->avg
.util_est
.enqueued
, enqueued
);
3928 trace_sched_util_est_cfs_tp(cfs_rq
);
3932 * Check if a (signed) value is within a specified (unsigned) margin,
3933 * based on the observation that:
3935 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3937 * NOTE: this only works when value + maring < INT_MAX.
3939 static inline bool within_margin(int value
, int margin
)
3941 return ((unsigned int)(value
+ margin
- 1) < (2 * margin
- 1));
3945 util_est_dequeue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
, bool task_sleep
)
3947 long last_ewma_diff
;
3951 if (!sched_feat(UTIL_EST
))
3954 /* Update root cfs_rq's estimated utilization */
3955 ue
.enqueued
= cfs_rq
->avg
.util_est
.enqueued
;
3956 ue
.enqueued
-= min_t(unsigned int, ue
.enqueued
, _task_util_est(p
));
3957 WRITE_ONCE(cfs_rq
->avg
.util_est
.enqueued
, ue
.enqueued
);
3959 trace_sched_util_est_cfs_tp(cfs_rq
);
3962 * Skip update of task's estimated utilization when the task has not
3963 * yet completed an activation, e.g. being migrated.
3969 * If the PELT values haven't changed since enqueue time,
3970 * skip the util_est update.
3972 ue
= p
->se
.avg
.util_est
;
3973 if (ue
.enqueued
& UTIL_AVG_UNCHANGED
)
3977 * Reset EWMA on utilization increases, the moving average is used only
3978 * to smooth utilization decreases.
3980 ue
.enqueued
= (task_util(p
) | UTIL_AVG_UNCHANGED
);
3981 if (sched_feat(UTIL_EST_FASTUP
)) {
3982 if (ue
.ewma
< ue
.enqueued
) {
3983 ue
.ewma
= ue
.enqueued
;
3989 * Skip update of task's estimated utilization when its EWMA is
3990 * already ~1% close to its last activation value.
3992 last_ewma_diff
= ue
.enqueued
- ue
.ewma
;
3993 if (within_margin(last_ewma_diff
, (SCHED_CAPACITY_SCALE
/ 100)))
3997 * To avoid overestimation of actual task utilization, skip updates if
3998 * we cannot grant there is idle time in this CPU.
4000 cpu
= cpu_of(rq_of(cfs_rq
));
4001 if (task_util(p
) > capacity_orig_of(cpu
))
4005 * Update Task's estimated utilization
4007 * When *p completes an activation we can consolidate another sample
4008 * of the task size. This is done by storing the current PELT value
4009 * as ue.enqueued and by using this value to update the Exponential
4010 * Weighted Moving Average (EWMA):
4012 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4013 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4014 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4015 * = w * ( last_ewma_diff ) + ewma(t-1)
4016 * = w * (last_ewma_diff + ewma(t-1) / w)
4018 * Where 'w' is the weight of new samples, which is configured to be
4019 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4021 ue
.ewma
<<= UTIL_EST_WEIGHT_SHIFT
;
4022 ue
.ewma
+= last_ewma_diff
;
4023 ue
.ewma
>>= UTIL_EST_WEIGHT_SHIFT
;
4025 WRITE_ONCE(p
->se
.avg
.util_est
, ue
);
4027 trace_sched_util_est_se_tp(&p
->se
);
4030 static inline int task_fits_capacity(struct task_struct
*p
, long capacity
)
4032 return fits_capacity(uclamp_task_util(p
), capacity
);
4035 static inline void update_misfit_status(struct task_struct
*p
, struct rq
*rq
)
4037 if (!static_branch_unlikely(&sched_asym_cpucapacity
))
4041 rq
->misfit_task_load
= 0;
4045 if (task_fits_capacity(p
, capacity_of(cpu_of(rq
)))) {
4046 rq
->misfit_task_load
= 0;
4051 * Make sure that misfit_task_load will not be null even if
4052 * task_h_load() returns 0.
4054 rq
->misfit_task_load
= max_t(unsigned long, task_h_load(p
), 1);
4057 #else /* CONFIG_SMP */
4059 #define UPDATE_TG 0x0
4060 #define SKIP_AGE_LOAD 0x0
4061 #define DO_ATTACH 0x0
4063 static inline void update_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int not_used1
)
4065 cfs_rq_util_change(cfs_rq
, 0);
4068 static inline void remove_entity_load_avg(struct sched_entity
*se
) {}
4071 attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
4073 detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
4075 static inline int newidle_balance(struct rq
*rq
, struct rq_flags
*rf
)
4081 util_est_enqueue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
) {}
4084 util_est_dequeue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
,
4086 static inline void update_misfit_status(struct task_struct
*p
, struct rq
*rq
) {}
4088 #endif /* CONFIG_SMP */
4090 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4092 #ifdef CONFIG_SCHED_DEBUG
4093 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
4098 if (d
> 3*sysctl_sched_latency
)
4099 schedstat_inc(cfs_rq
->nr_spread_over
);
4104 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
4106 u64 vruntime
= cfs_rq
->min_vruntime
;
4109 * The 'current' period is already promised to the current tasks,
4110 * however the extra weight of the new task will slow them down a
4111 * little, place the new task so that it fits in the slot that
4112 * stays open at the end.
4114 if (initial
&& sched_feat(START_DEBIT
))
4115 vruntime
+= sched_vslice(cfs_rq
, se
);
4117 /* sleeps up to a single latency don't count. */
4119 unsigned long thresh
= sysctl_sched_latency
;
4122 * Halve their sleep time's effect, to allow
4123 * for a gentler effect of sleepers:
4125 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
4131 /* ensure we never gain time by being placed backwards. */
4132 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
4135 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
4137 static inline void check_schedstat_required(void)
4139 #ifdef CONFIG_SCHEDSTATS
4140 if (schedstat_enabled())
4143 /* Force schedstat enabled if a dependent tracepoint is active */
4144 if (trace_sched_stat_wait_enabled() ||
4145 trace_sched_stat_sleep_enabled() ||
4146 trace_sched_stat_iowait_enabled() ||
4147 trace_sched_stat_blocked_enabled() ||
4148 trace_sched_stat_runtime_enabled()) {
4149 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4150 "stat_blocked and stat_runtime require the "
4151 "kernel parameter schedstats=enable or "
4152 "kernel.sched_schedstats=1\n");
4157 static inline bool cfs_bandwidth_used(void);
4164 * update_min_vruntime()
4165 * vruntime -= min_vruntime
4169 * update_min_vruntime()
4170 * vruntime += min_vruntime
4172 * this way the vruntime transition between RQs is done when both
4173 * min_vruntime are up-to-date.
4177 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4178 * vruntime -= min_vruntime
4182 * update_min_vruntime()
4183 * vruntime += min_vruntime
4185 * this way we don't have the most up-to-date min_vruntime on the originating
4186 * CPU and an up-to-date min_vruntime on the destination CPU.
4190 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
4192 bool renorm
= !(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_MIGRATED
);
4193 bool curr
= cfs_rq
->curr
== se
;
4196 * If we're the current task, we must renormalise before calling
4200 se
->vruntime
+= cfs_rq
->min_vruntime
;
4202 update_curr(cfs_rq
);
4205 * Otherwise, renormalise after, such that we're placed at the current
4206 * moment in time, instead of some random moment in the past. Being
4207 * placed in the past could significantly boost this task to the
4208 * fairness detriment of existing tasks.
4210 if (renorm
&& !curr
)
4211 se
->vruntime
+= cfs_rq
->min_vruntime
;
4214 * When enqueuing a sched_entity, we must:
4215 * - Update loads to have both entity and cfs_rq synced with now.
4216 * - Add its load to cfs_rq->runnable_avg
4217 * - For group_entity, update its weight to reflect the new share of
4219 * - Add its new weight to cfs_rq->load.weight
4221 update_load_avg(cfs_rq
, se
, UPDATE_TG
| DO_ATTACH
);
4222 se_update_runnable(se
);
4223 update_cfs_group(se
);
4224 account_entity_enqueue(cfs_rq
, se
);
4226 if (flags
& ENQUEUE_WAKEUP
)
4227 place_entity(cfs_rq
, se
, 0);
4229 check_schedstat_required();
4230 update_stats_enqueue(cfs_rq
, se
, flags
);
4231 check_spread(cfs_rq
, se
);
4233 __enqueue_entity(cfs_rq
, se
);
4237 * When bandwidth control is enabled, cfs might have been removed
4238 * because of a parent been throttled but cfs->nr_running > 1. Try to
4239 * add it unconditionnally.
4241 if (cfs_rq
->nr_running
== 1 || cfs_bandwidth_used())
4242 list_add_leaf_cfs_rq(cfs_rq
);
4244 if (cfs_rq
->nr_running
== 1)
4245 check_enqueue_throttle(cfs_rq
);
4248 static void __clear_buddies_last(struct sched_entity
*se
)
4250 for_each_sched_entity(se
) {
4251 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4252 if (cfs_rq
->last
!= se
)
4255 cfs_rq
->last
= NULL
;
4259 static void __clear_buddies_next(struct sched_entity
*se
)
4261 for_each_sched_entity(se
) {
4262 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4263 if (cfs_rq
->next
!= se
)
4266 cfs_rq
->next
= NULL
;
4270 static void __clear_buddies_skip(struct sched_entity
*se
)
4272 for_each_sched_entity(se
) {
4273 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4274 if (cfs_rq
->skip
!= se
)
4277 cfs_rq
->skip
= NULL
;
4281 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4283 if (cfs_rq
->last
== se
)
4284 __clear_buddies_last(se
);
4286 if (cfs_rq
->next
== se
)
4287 __clear_buddies_next(se
);
4289 if (cfs_rq
->skip
== se
)
4290 __clear_buddies_skip(se
);
4293 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
4296 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
4299 * Update run-time statistics of the 'current'.
4301 update_curr(cfs_rq
);
4304 * When dequeuing a sched_entity, we must:
4305 * - Update loads to have both entity and cfs_rq synced with now.
4306 * - Subtract its load from the cfs_rq->runnable_avg.
4307 * - Subtract its previous weight from cfs_rq->load.weight.
4308 * - For group entity, update its weight to reflect the new share
4309 * of its group cfs_rq.
4311 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4312 se_update_runnable(se
);
4314 update_stats_dequeue(cfs_rq
, se
, flags
);
4316 clear_buddies(cfs_rq
, se
);
4318 if (se
!= cfs_rq
->curr
)
4319 __dequeue_entity(cfs_rq
, se
);
4321 account_entity_dequeue(cfs_rq
, se
);
4324 * Normalize after update_curr(); which will also have moved
4325 * min_vruntime if @se is the one holding it back. But before doing
4326 * update_min_vruntime() again, which will discount @se's position and
4327 * can move min_vruntime forward still more.
4329 if (!(flags
& DEQUEUE_SLEEP
))
4330 se
->vruntime
-= cfs_rq
->min_vruntime
;
4332 /* return excess runtime on last dequeue */
4333 return_cfs_rq_runtime(cfs_rq
);
4335 update_cfs_group(se
);
4338 * Now advance min_vruntime if @se was the entity holding it back,
4339 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4340 * put back on, and if we advance min_vruntime, we'll be placed back
4341 * further than we started -- ie. we'll be penalized.
4343 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) != DEQUEUE_SAVE
)
4344 update_min_vruntime(cfs_rq
);
4348 * Preempt the current task with a newly woken task if needed:
4351 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
4353 unsigned long ideal_runtime
, delta_exec
;
4354 struct sched_entity
*se
;
4357 ideal_runtime
= sched_slice(cfs_rq
, curr
);
4358 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
4359 if (delta_exec
> ideal_runtime
) {
4360 resched_curr(rq_of(cfs_rq
));
4362 * The current task ran long enough, ensure it doesn't get
4363 * re-elected due to buddy favours.
4365 clear_buddies(cfs_rq
, curr
);
4370 * Ensure that a task that missed wakeup preemption by a
4371 * narrow margin doesn't have to wait for a full slice.
4372 * This also mitigates buddy induced latencies under load.
4374 if (delta_exec
< sysctl_sched_min_granularity
)
4377 se
= __pick_first_entity(cfs_rq
);
4378 delta
= curr
->vruntime
- se
->vruntime
;
4383 if (delta
> ideal_runtime
)
4384 resched_curr(rq_of(cfs_rq
));
4388 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4390 /* 'current' is not kept within the tree. */
4393 * Any task has to be enqueued before it get to execute on
4394 * a CPU. So account for the time it spent waiting on the
4397 update_stats_wait_end(cfs_rq
, se
);
4398 __dequeue_entity(cfs_rq
, se
);
4399 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4402 update_stats_curr_start(cfs_rq
, se
);
4406 * Track our maximum slice length, if the CPU's load is at
4407 * least twice that of our own weight (i.e. dont track it
4408 * when there are only lesser-weight tasks around):
4410 if (schedstat_enabled() &&
4411 rq_of(cfs_rq
)->cfs
.load
.weight
>= 2*se
->load
.weight
) {
4412 schedstat_set(se
->statistics
.slice_max
,
4413 max((u64
)schedstat_val(se
->statistics
.slice_max
),
4414 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
));
4417 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
4421 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
4424 * Pick the next process, keeping these things in mind, in this order:
4425 * 1) keep things fair between processes/task groups
4426 * 2) pick the "next" process, since someone really wants that to run
4427 * 3) pick the "last" process, for cache locality
4428 * 4) do not run the "skip" process, if something else is available
4430 static struct sched_entity
*
4431 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
4433 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
4434 struct sched_entity
*se
;
4437 * If curr is set we have to see if its left of the leftmost entity
4438 * still in the tree, provided there was anything in the tree at all.
4440 if (!left
|| (curr
&& entity_before(curr
, left
)))
4443 se
= left
; /* ideally we run the leftmost entity */
4446 * Avoid running the skip buddy, if running something else can
4447 * be done without getting too unfair.
4449 if (cfs_rq
->skip
== se
) {
4450 struct sched_entity
*second
;
4453 second
= __pick_first_entity(cfs_rq
);
4455 second
= __pick_next_entity(se
);
4456 if (!second
|| (curr
&& entity_before(curr
, second
)))
4460 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
4465 * Prefer last buddy, try to return the CPU to a preempted task.
4467 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
4471 * Someone really wants this to run. If it's not unfair, run it.
4473 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
4476 clear_buddies(cfs_rq
, se
);
4481 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
4483 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
4486 * If still on the runqueue then deactivate_task()
4487 * was not called and update_curr() has to be done:
4490 update_curr(cfs_rq
);
4492 /* throttle cfs_rqs exceeding runtime */
4493 check_cfs_rq_runtime(cfs_rq
);
4495 check_spread(cfs_rq
, prev
);
4498 update_stats_wait_start(cfs_rq
, prev
);
4499 /* Put 'current' back into the tree. */
4500 __enqueue_entity(cfs_rq
, prev
);
4501 /* in !on_rq case, update occurred at dequeue */
4502 update_load_avg(cfs_rq
, prev
, 0);
4504 cfs_rq
->curr
= NULL
;
4508 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
4511 * Update run-time statistics of the 'current'.
4513 update_curr(cfs_rq
);
4516 * Ensure that runnable average is periodically updated.
4518 update_load_avg(cfs_rq
, curr
, UPDATE_TG
);
4519 update_cfs_group(curr
);
4521 #ifdef CONFIG_SCHED_HRTICK
4523 * queued ticks are scheduled to match the slice, so don't bother
4524 * validating it and just reschedule.
4527 resched_curr(rq_of(cfs_rq
));
4531 * don't let the period tick interfere with the hrtick preemption
4533 if (!sched_feat(DOUBLE_TICK
) &&
4534 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
4538 if (cfs_rq
->nr_running
> 1)
4539 check_preempt_tick(cfs_rq
, curr
);
4543 /**************************************************
4544 * CFS bandwidth control machinery
4547 #ifdef CONFIG_CFS_BANDWIDTH
4549 #ifdef CONFIG_JUMP_LABEL
4550 static struct static_key __cfs_bandwidth_used
;
4552 static inline bool cfs_bandwidth_used(void)
4554 return static_key_false(&__cfs_bandwidth_used
);
4557 void cfs_bandwidth_usage_inc(void)
4559 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used
);
4562 void cfs_bandwidth_usage_dec(void)
4564 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used
);
4566 #else /* CONFIG_JUMP_LABEL */
4567 static bool cfs_bandwidth_used(void)
4572 void cfs_bandwidth_usage_inc(void) {}
4573 void cfs_bandwidth_usage_dec(void) {}
4574 #endif /* CONFIG_JUMP_LABEL */
4577 * default period for cfs group bandwidth.
4578 * default: 0.1s, units: nanoseconds
4580 static inline u64
default_cfs_period(void)
4582 return 100000000ULL;
4585 static inline u64
sched_cfs_bandwidth_slice(void)
4587 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
4591 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4592 * directly instead of rq->clock to avoid adding additional synchronization
4595 * requires cfs_b->lock
4597 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
4599 if (cfs_b
->quota
!= RUNTIME_INF
)
4600 cfs_b
->runtime
= cfs_b
->quota
;
4603 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4605 return &tg
->cfs_bandwidth
;
4608 /* returns 0 on failure to allocate runtime */
4609 static int __assign_cfs_rq_runtime(struct cfs_bandwidth
*cfs_b
,
4610 struct cfs_rq
*cfs_rq
, u64 target_runtime
)
4612 u64 min_amount
, amount
= 0;
4614 lockdep_assert_held(&cfs_b
->lock
);
4616 /* note: this is a positive sum as runtime_remaining <= 0 */
4617 min_amount
= target_runtime
- cfs_rq
->runtime_remaining
;
4619 if (cfs_b
->quota
== RUNTIME_INF
)
4620 amount
= min_amount
;
4622 start_cfs_bandwidth(cfs_b
);
4624 if (cfs_b
->runtime
> 0) {
4625 amount
= min(cfs_b
->runtime
, min_amount
);
4626 cfs_b
->runtime
-= amount
;
4631 cfs_rq
->runtime_remaining
+= amount
;
4633 return cfs_rq
->runtime_remaining
> 0;
4636 /* returns 0 on failure to allocate runtime */
4637 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4639 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4642 raw_spin_lock(&cfs_b
->lock
);
4643 ret
= __assign_cfs_rq_runtime(cfs_b
, cfs_rq
, sched_cfs_bandwidth_slice());
4644 raw_spin_unlock(&cfs_b
->lock
);
4649 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4651 /* dock delta_exec before expiring quota (as it could span periods) */
4652 cfs_rq
->runtime_remaining
-= delta_exec
;
4654 if (likely(cfs_rq
->runtime_remaining
> 0))
4657 if (cfs_rq
->throttled
)
4660 * if we're unable to extend our runtime we resched so that the active
4661 * hierarchy can be throttled
4663 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
4664 resched_curr(rq_of(cfs_rq
));
4667 static __always_inline
4668 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4670 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
4673 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
4676 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4678 return cfs_bandwidth_used() && cfs_rq
->throttled
;
4681 /* check whether cfs_rq, or any parent, is throttled */
4682 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4684 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
4688 * Ensure that neither of the group entities corresponding to src_cpu or
4689 * dest_cpu are members of a throttled hierarchy when performing group
4690 * load-balance operations.
4692 static inline int throttled_lb_pair(struct task_group
*tg
,
4693 int src_cpu
, int dest_cpu
)
4695 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
4697 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
4698 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
4700 return throttled_hierarchy(src_cfs_rq
) ||
4701 throttled_hierarchy(dest_cfs_rq
);
4704 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
4706 struct rq
*rq
= data
;
4707 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4709 cfs_rq
->throttle_count
--;
4710 if (!cfs_rq
->throttle_count
) {
4711 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
4712 cfs_rq
->throttled_clock_task
;
4714 /* Add cfs_rq with already running entity in the list */
4715 if (cfs_rq
->nr_running
>= 1)
4716 list_add_leaf_cfs_rq(cfs_rq
);
4722 static int tg_throttle_down(struct task_group
*tg
, void *data
)
4724 struct rq
*rq
= data
;
4725 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4727 /* group is entering throttled state, stop time */
4728 if (!cfs_rq
->throttle_count
) {
4729 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
4730 list_del_leaf_cfs_rq(cfs_rq
);
4732 cfs_rq
->throttle_count
++;
4737 static bool throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
4739 struct rq
*rq
= rq_of(cfs_rq
);
4740 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4741 struct sched_entity
*se
;
4742 long task_delta
, idle_task_delta
, dequeue
= 1;
4744 raw_spin_lock(&cfs_b
->lock
);
4745 /* This will start the period timer if necessary */
4746 if (__assign_cfs_rq_runtime(cfs_b
, cfs_rq
, 1)) {
4748 * We have raced with bandwidth becoming available, and if we
4749 * actually throttled the timer might not unthrottle us for an
4750 * entire period. We additionally needed to make sure that any
4751 * subsequent check_cfs_rq_runtime calls agree not to throttle
4752 * us, as we may commit to do cfs put_prev+pick_next, so we ask
4753 * for 1ns of runtime rather than just check cfs_b.
4757 list_add_tail_rcu(&cfs_rq
->throttled_list
,
4758 &cfs_b
->throttled_cfs_rq
);
4760 raw_spin_unlock(&cfs_b
->lock
);
4763 return false; /* Throttle no longer required. */
4765 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
4767 /* freeze hierarchy runnable averages while throttled */
4769 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
4772 task_delta
= cfs_rq
->h_nr_running
;
4773 idle_task_delta
= cfs_rq
->idle_h_nr_running
;
4774 for_each_sched_entity(se
) {
4775 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4776 /* throttled entity or throttle-on-deactivate */
4781 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
4783 update_load_avg(qcfs_rq
, se
, 0);
4784 se_update_runnable(se
);
4787 qcfs_rq
->h_nr_running
-= task_delta
;
4788 qcfs_rq
->idle_h_nr_running
-= idle_task_delta
;
4790 if (qcfs_rq
->load
.weight
)
4795 sub_nr_running(rq
, task_delta
);
4798 * Note: distribution will already see us throttled via the
4799 * throttled-list. rq->lock protects completion.
4801 cfs_rq
->throttled
= 1;
4802 cfs_rq
->throttled_clock
= rq_clock(rq
);
4806 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
4808 struct rq
*rq
= rq_of(cfs_rq
);
4809 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4810 struct sched_entity
*se
;
4811 long task_delta
, idle_task_delta
;
4813 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
4815 cfs_rq
->throttled
= 0;
4817 update_rq_clock(rq
);
4819 raw_spin_lock(&cfs_b
->lock
);
4820 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
4821 list_del_rcu(&cfs_rq
->throttled_list
);
4822 raw_spin_unlock(&cfs_b
->lock
);
4824 /* update hierarchical throttle state */
4825 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
4827 if (!cfs_rq
->load
.weight
)
4830 task_delta
= cfs_rq
->h_nr_running
;
4831 idle_task_delta
= cfs_rq
->idle_h_nr_running
;
4832 for_each_sched_entity(se
) {
4835 cfs_rq
= cfs_rq_of(se
);
4836 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
4838 cfs_rq
->h_nr_running
+= task_delta
;
4839 cfs_rq
->idle_h_nr_running
+= idle_task_delta
;
4841 /* end evaluation on encountering a throttled cfs_rq */
4842 if (cfs_rq_throttled(cfs_rq
))
4843 goto unthrottle_throttle
;
4846 for_each_sched_entity(se
) {
4847 cfs_rq
= cfs_rq_of(se
);
4849 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4850 se_update_runnable(se
);
4852 cfs_rq
->h_nr_running
+= task_delta
;
4853 cfs_rq
->idle_h_nr_running
+= idle_task_delta
;
4856 /* end evaluation on encountering a throttled cfs_rq */
4857 if (cfs_rq_throttled(cfs_rq
))
4858 goto unthrottle_throttle
;
4861 * One parent has been throttled and cfs_rq removed from the
4862 * list. Add it back to not break the leaf list.
4864 if (throttled_hierarchy(cfs_rq
))
4865 list_add_leaf_cfs_rq(cfs_rq
);
4868 /* At this point se is NULL and we are at root level*/
4869 add_nr_running(rq
, task_delta
);
4871 unthrottle_throttle
:
4873 * The cfs_rq_throttled() breaks in the above iteration can result in
4874 * incomplete leaf list maintenance, resulting in triggering the
4877 for_each_sched_entity(se
) {
4878 cfs_rq
= cfs_rq_of(se
);
4880 if (list_add_leaf_cfs_rq(cfs_rq
))
4884 assert_list_leaf_cfs_rq(rq
);
4886 /* Determine whether we need to wake up potentially idle CPU: */
4887 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
4891 static void distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
)
4893 struct cfs_rq
*cfs_rq
;
4894 u64 runtime
, remaining
= 1;
4897 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
4899 struct rq
*rq
= rq_of(cfs_rq
);
4902 rq_lock_irqsave(rq
, &rf
);
4903 if (!cfs_rq_throttled(cfs_rq
))
4906 /* By the above check, this should never be true */
4907 SCHED_WARN_ON(cfs_rq
->runtime_remaining
> 0);
4909 raw_spin_lock(&cfs_b
->lock
);
4910 runtime
= -cfs_rq
->runtime_remaining
+ 1;
4911 if (runtime
> cfs_b
->runtime
)
4912 runtime
= cfs_b
->runtime
;
4913 cfs_b
->runtime
-= runtime
;
4914 remaining
= cfs_b
->runtime
;
4915 raw_spin_unlock(&cfs_b
->lock
);
4917 cfs_rq
->runtime_remaining
+= runtime
;
4919 /* we check whether we're throttled above */
4920 if (cfs_rq
->runtime_remaining
> 0)
4921 unthrottle_cfs_rq(cfs_rq
);
4924 rq_unlock_irqrestore(rq
, &rf
);
4933 * Responsible for refilling a task_group's bandwidth and unthrottling its
4934 * cfs_rqs as appropriate. If there has been no activity within the last
4935 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4936 * used to track this state.
4938 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
, unsigned long flags
)
4942 /* no need to continue the timer with no bandwidth constraint */
4943 if (cfs_b
->quota
== RUNTIME_INF
)
4944 goto out_deactivate
;
4946 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4947 cfs_b
->nr_periods
+= overrun
;
4950 * idle depends on !throttled (for the case of a large deficit), and if
4951 * we're going inactive then everything else can be deferred
4953 if (cfs_b
->idle
&& !throttled
)
4954 goto out_deactivate
;
4956 __refill_cfs_bandwidth_runtime(cfs_b
);
4959 /* mark as potentially idle for the upcoming period */
4964 /* account preceding periods in which throttling occurred */
4965 cfs_b
->nr_throttled
+= overrun
;
4968 * This check is repeated as we release cfs_b->lock while we unthrottle.
4970 while (throttled
&& cfs_b
->runtime
> 0) {
4971 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
4972 /* we can't nest cfs_b->lock while distributing bandwidth */
4973 distribute_cfs_runtime(cfs_b
);
4974 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
4976 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4980 * While we are ensured activity in the period following an
4981 * unthrottle, this also covers the case in which the new bandwidth is
4982 * insufficient to cover the existing bandwidth deficit. (Forcing the
4983 * timer to remain active while there are any throttled entities.)
4993 /* a cfs_rq won't donate quota below this amount */
4994 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
4995 /* minimum remaining period time to redistribute slack quota */
4996 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
4997 /* how long we wait to gather additional slack before distributing */
4998 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
5001 * Are we near the end of the current quota period?
5003 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5004 * hrtimer base being cleared by hrtimer_start. In the case of
5005 * migrate_hrtimers, base is never cleared, so we are fine.
5007 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
5009 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
5012 /* if the call-back is running a quota refresh is already occurring */
5013 if (hrtimer_callback_running(refresh_timer
))
5016 /* is a quota refresh about to occur? */
5017 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
5018 if (remaining
< min_expire
)
5024 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
5026 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
5028 /* if there's a quota refresh soon don't bother with slack */
5029 if (runtime_refresh_within(cfs_b
, min_left
))
5032 /* don't push forwards an existing deferred unthrottle */
5033 if (cfs_b
->slack_started
)
5035 cfs_b
->slack_started
= true;
5037 hrtimer_start(&cfs_b
->slack_timer
,
5038 ns_to_ktime(cfs_bandwidth_slack_period
),
5042 /* we know any runtime found here is valid as update_curr() precedes return */
5043 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5045 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
5046 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
5048 if (slack_runtime
<= 0)
5051 raw_spin_lock(&cfs_b
->lock
);
5052 if (cfs_b
->quota
!= RUNTIME_INF
) {
5053 cfs_b
->runtime
+= slack_runtime
;
5055 /* we are under rq->lock, defer unthrottling using a timer */
5056 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
5057 !list_empty(&cfs_b
->throttled_cfs_rq
))
5058 start_cfs_slack_bandwidth(cfs_b
);
5060 raw_spin_unlock(&cfs_b
->lock
);
5062 /* even if it's not valid for return we don't want to try again */
5063 cfs_rq
->runtime_remaining
-= slack_runtime
;
5066 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5068 if (!cfs_bandwidth_used())
5071 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
5074 __return_cfs_rq_runtime(cfs_rq
);
5078 * This is done with a timer (instead of inline with bandwidth return) since
5079 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5081 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
5083 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
5084 unsigned long flags
;
5086 /* confirm we're still not at a refresh boundary */
5087 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5088 cfs_b
->slack_started
= false;
5090 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
5091 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5095 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
5096 runtime
= cfs_b
->runtime
;
5098 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5103 distribute_cfs_runtime(cfs_b
);
5105 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5106 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5110 * When a group wakes up we want to make sure that its quota is not already
5111 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5112 * runtime as update_curr() throttling can not not trigger until it's on-rq.
5114 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
5116 if (!cfs_bandwidth_used())
5119 /* an active group must be handled by the update_curr()->put() path */
5120 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
5123 /* ensure the group is not already throttled */
5124 if (cfs_rq_throttled(cfs_rq
))
5127 /* update runtime allocation */
5128 account_cfs_rq_runtime(cfs_rq
, 0);
5129 if (cfs_rq
->runtime_remaining
<= 0)
5130 throttle_cfs_rq(cfs_rq
);
5133 static void sync_throttle(struct task_group
*tg
, int cpu
)
5135 struct cfs_rq
*pcfs_rq
, *cfs_rq
;
5137 if (!cfs_bandwidth_used())
5143 cfs_rq
= tg
->cfs_rq
[cpu
];
5144 pcfs_rq
= tg
->parent
->cfs_rq
[cpu
];
5146 cfs_rq
->throttle_count
= pcfs_rq
->throttle_count
;
5147 cfs_rq
->throttled_clock_task
= rq_clock_task(cpu_rq(cpu
));
5150 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5151 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5153 if (!cfs_bandwidth_used())
5156 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
5160 * it's possible for a throttled entity to be forced into a running
5161 * state (e.g. set_curr_task), in this case we're finished.
5163 if (cfs_rq_throttled(cfs_rq
))
5166 return throttle_cfs_rq(cfs_rq
);
5169 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
5171 struct cfs_bandwidth
*cfs_b
=
5172 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
5174 do_sched_cfs_slack_timer(cfs_b
);
5176 return HRTIMER_NORESTART
;
5179 extern const u64 max_cfs_quota_period
;
5181 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
5183 struct cfs_bandwidth
*cfs_b
=
5184 container_of(timer
, struct cfs_bandwidth
, period_timer
);
5185 unsigned long flags
;
5190 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5192 overrun
= hrtimer_forward_now(timer
, cfs_b
->period
);
5196 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
, flags
);
5199 u64
new, old
= ktime_to_ns(cfs_b
->period
);
5202 * Grow period by a factor of 2 to avoid losing precision.
5203 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5207 if (new < max_cfs_quota_period
) {
5208 cfs_b
->period
= ns_to_ktime(new);
5211 pr_warn_ratelimited(
5212 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5214 div_u64(new, NSEC_PER_USEC
),
5215 div_u64(cfs_b
->quota
, NSEC_PER_USEC
));
5217 pr_warn_ratelimited(
5218 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5220 div_u64(old
, NSEC_PER_USEC
),
5221 div_u64(cfs_b
->quota
, NSEC_PER_USEC
));
5224 /* reset count so we don't come right back in here */
5229 cfs_b
->period_active
= 0;
5230 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5232 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
5235 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5237 raw_spin_lock_init(&cfs_b
->lock
);
5239 cfs_b
->quota
= RUNTIME_INF
;
5240 cfs_b
->period
= ns_to_ktime(default_cfs_period());
5242 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
5243 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_PINNED
);
5244 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
5245 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5246 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
5247 cfs_b
->slack_started
= false;
5250 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5252 cfs_rq
->runtime_enabled
= 0;
5253 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
5256 void start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5258 lockdep_assert_held(&cfs_b
->lock
);
5260 if (cfs_b
->period_active
)
5263 cfs_b
->period_active
= 1;
5264 hrtimer_forward_now(&cfs_b
->period_timer
, cfs_b
->period
);
5265 hrtimer_start_expires(&cfs_b
->period_timer
, HRTIMER_MODE_ABS_PINNED
);
5268 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5270 /* init_cfs_bandwidth() was not called */
5271 if (!cfs_b
->throttled_cfs_rq
.next
)
5274 hrtimer_cancel(&cfs_b
->period_timer
);
5275 hrtimer_cancel(&cfs_b
->slack_timer
);
5279 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5281 * The race is harmless, since modifying bandwidth settings of unhooked group
5282 * bits doesn't do much.
5285 /* cpu online calback */
5286 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
5288 struct task_group
*tg
;
5290 lockdep_assert_held(&rq
->lock
);
5293 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
5294 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
5295 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
5297 raw_spin_lock(&cfs_b
->lock
);
5298 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
5299 raw_spin_unlock(&cfs_b
->lock
);
5304 /* cpu offline callback */
5305 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
5307 struct task_group
*tg
;
5309 lockdep_assert_held(&rq
->lock
);
5312 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
5313 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
5315 if (!cfs_rq
->runtime_enabled
)
5319 * clock_task is not advancing so we just need to make sure
5320 * there's some valid quota amount
5322 cfs_rq
->runtime_remaining
= 1;
5324 * Offline rq is schedulable till CPU is completely disabled
5325 * in take_cpu_down(), so we prevent new cfs throttling here.
5327 cfs_rq
->runtime_enabled
= 0;
5329 if (cfs_rq_throttled(cfs_rq
))
5330 unthrottle_cfs_rq(cfs_rq
);
5335 #else /* CONFIG_CFS_BANDWIDTH */
5337 static inline bool cfs_bandwidth_used(void)
5342 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
5343 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
5344 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
5345 static inline void sync_throttle(struct task_group
*tg
, int cpu
) {}
5346 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
5348 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
5353 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
5358 static inline int throttled_lb_pair(struct task_group
*tg
,
5359 int src_cpu
, int dest_cpu
)
5364 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
5366 #ifdef CONFIG_FAIR_GROUP_SCHED
5367 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
5370 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
5374 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
5375 static inline void update_runtime_enabled(struct rq
*rq
) {}
5376 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
5378 #endif /* CONFIG_CFS_BANDWIDTH */
5380 /**************************************************
5381 * CFS operations on tasks:
5384 #ifdef CONFIG_SCHED_HRTICK
5385 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
5387 struct sched_entity
*se
= &p
->se
;
5388 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5390 SCHED_WARN_ON(task_rq(p
) != rq
);
5392 if (rq
->cfs
.h_nr_running
> 1) {
5393 u64 slice
= sched_slice(cfs_rq
, se
);
5394 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
5395 s64 delta
= slice
- ran
;
5402 hrtick_start(rq
, delta
);
5407 * called from enqueue/dequeue and updates the hrtick when the
5408 * current task is from our class and nr_running is low enough
5411 static void hrtick_update(struct rq
*rq
)
5413 struct task_struct
*curr
= rq
->curr
;
5415 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
5418 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
5419 hrtick_start_fair(rq
, curr
);
5421 #else /* !CONFIG_SCHED_HRTICK */
5423 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
5427 static inline void hrtick_update(struct rq
*rq
)
5433 static inline unsigned long cpu_util(int cpu
);
5435 static inline bool cpu_overutilized(int cpu
)
5437 return !fits_capacity(cpu_util(cpu
), capacity_of(cpu
));
5440 static inline void update_overutilized_status(struct rq
*rq
)
5442 if (!READ_ONCE(rq
->rd
->overutilized
) && cpu_overutilized(rq
->cpu
)) {
5443 WRITE_ONCE(rq
->rd
->overutilized
, SG_OVERUTILIZED
);
5444 trace_sched_overutilized_tp(rq
->rd
, SG_OVERUTILIZED
);
5448 static inline void update_overutilized_status(struct rq
*rq
) { }
5451 /* Runqueue only has SCHED_IDLE tasks enqueued */
5452 static int sched_idle_rq(struct rq
*rq
)
5454 return unlikely(rq
->nr_running
== rq
->cfs
.idle_h_nr_running
&&
5459 static int sched_idle_cpu(int cpu
)
5461 return sched_idle_rq(cpu_rq(cpu
));
5466 * The enqueue_task method is called before nr_running is
5467 * increased. Here we update the fair scheduling stats and
5468 * then put the task into the rbtree:
5471 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
5473 struct cfs_rq
*cfs_rq
;
5474 struct sched_entity
*se
= &p
->se
;
5475 int idle_h_nr_running
= task_has_idle_policy(p
);
5478 * The code below (indirectly) updates schedutil which looks at
5479 * the cfs_rq utilization to select a frequency.
5480 * Let's add the task's estimated utilization to the cfs_rq's
5481 * estimated utilization, before we update schedutil.
5483 util_est_enqueue(&rq
->cfs
, p
);
5486 * If in_iowait is set, the code below may not trigger any cpufreq
5487 * utilization updates, so do it here explicitly with the IOWAIT flag
5491 cpufreq_update_util(rq
, SCHED_CPUFREQ_IOWAIT
);
5493 for_each_sched_entity(se
) {
5496 cfs_rq
= cfs_rq_of(se
);
5497 enqueue_entity(cfs_rq
, se
, flags
);
5499 cfs_rq
->h_nr_running
++;
5500 cfs_rq
->idle_h_nr_running
+= idle_h_nr_running
;
5502 /* end evaluation on encountering a throttled cfs_rq */
5503 if (cfs_rq_throttled(cfs_rq
))
5504 goto enqueue_throttle
;
5506 flags
= ENQUEUE_WAKEUP
;
5509 for_each_sched_entity(se
) {
5510 cfs_rq
= cfs_rq_of(se
);
5512 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
5513 se_update_runnable(se
);
5514 update_cfs_group(se
);
5516 cfs_rq
->h_nr_running
++;
5517 cfs_rq
->idle_h_nr_running
+= idle_h_nr_running
;
5519 /* end evaluation on encountering a throttled cfs_rq */
5520 if (cfs_rq_throttled(cfs_rq
))
5521 goto enqueue_throttle
;
5524 * One parent has been throttled and cfs_rq removed from the
5525 * list. Add it back to not break the leaf list.
5527 if (throttled_hierarchy(cfs_rq
))
5528 list_add_leaf_cfs_rq(cfs_rq
);
5531 /* At this point se is NULL and we are at root level*/
5532 add_nr_running(rq
, 1);
5535 * Since new tasks are assigned an initial util_avg equal to
5536 * half of the spare capacity of their CPU, tiny tasks have the
5537 * ability to cross the overutilized threshold, which will
5538 * result in the load balancer ruining all the task placement
5539 * done by EAS. As a way to mitigate that effect, do not account
5540 * for the first enqueue operation of new tasks during the
5541 * overutilized flag detection.
5543 * A better way of solving this problem would be to wait for
5544 * the PELT signals of tasks to converge before taking them
5545 * into account, but that is not straightforward to implement,
5546 * and the following generally works well enough in practice.
5548 if (flags
& ENQUEUE_WAKEUP
)
5549 update_overutilized_status(rq
);
5552 if (cfs_bandwidth_used()) {
5554 * When bandwidth control is enabled; the cfs_rq_throttled()
5555 * breaks in the above iteration can result in incomplete
5556 * leaf list maintenance, resulting in triggering the assertion
5559 for_each_sched_entity(se
) {
5560 cfs_rq
= cfs_rq_of(se
);
5562 if (list_add_leaf_cfs_rq(cfs_rq
))
5567 assert_list_leaf_cfs_rq(rq
);
5572 static void set_next_buddy(struct sched_entity
*se
);
5575 * The dequeue_task method is called before nr_running is
5576 * decreased. We remove the task from the rbtree and
5577 * update the fair scheduling stats:
5579 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
5581 struct cfs_rq
*cfs_rq
;
5582 struct sched_entity
*se
= &p
->se
;
5583 int task_sleep
= flags
& DEQUEUE_SLEEP
;
5584 int idle_h_nr_running
= task_has_idle_policy(p
);
5585 bool was_sched_idle
= sched_idle_rq(rq
);
5587 for_each_sched_entity(se
) {
5588 cfs_rq
= cfs_rq_of(se
);
5589 dequeue_entity(cfs_rq
, se
, flags
);
5591 cfs_rq
->h_nr_running
--;
5592 cfs_rq
->idle_h_nr_running
-= idle_h_nr_running
;
5594 /* end evaluation on encountering a throttled cfs_rq */
5595 if (cfs_rq_throttled(cfs_rq
))
5596 goto dequeue_throttle
;
5598 /* Don't dequeue parent if it has other entities besides us */
5599 if (cfs_rq
->load
.weight
) {
5600 /* Avoid re-evaluating load for this entity: */
5601 se
= parent_entity(se
);
5603 * Bias pick_next to pick a task from this cfs_rq, as
5604 * p is sleeping when it is within its sched_slice.
5606 if (task_sleep
&& se
&& !throttled_hierarchy(cfs_rq
))
5610 flags
|= DEQUEUE_SLEEP
;
5613 for_each_sched_entity(se
) {
5614 cfs_rq
= cfs_rq_of(se
);
5616 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
5617 se_update_runnable(se
);
5618 update_cfs_group(se
);
5620 cfs_rq
->h_nr_running
--;
5621 cfs_rq
->idle_h_nr_running
-= idle_h_nr_running
;
5623 /* end evaluation on encountering a throttled cfs_rq */
5624 if (cfs_rq_throttled(cfs_rq
))
5625 goto dequeue_throttle
;
5629 /* At this point se is NULL and we are at root level*/
5630 sub_nr_running(rq
, 1);
5632 /* balance early to pull high priority tasks */
5633 if (unlikely(!was_sched_idle
&& sched_idle_rq(rq
)))
5634 rq
->next_balance
= jiffies
;
5637 util_est_dequeue(&rq
->cfs
, p
, task_sleep
);
5643 /* Working cpumask for: load_balance, load_balance_newidle. */
5644 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5645 DEFINE_PER_CPU(cpumask_var_t
, select_idle_mask
);
5647 #ifdef CONFIG_NO_HZ_COMMON
5650 cpumask_var_t idle_cpus_mask
;
5652 int has_blocked
; /* Idle CPUS has blocked load */
5653 unsigned long next_balance
; /* in jiffy units */
5654 unsigned long next_blocked
; /* Next update of blocked load in jiffies */
5655 } nohz ____cacheline_aligned
;
5657 #endif /* CONFIG_NO_HZ_COMMON */
5659 static unsigned long cpu_load(struct rq
*rq
)
5661 return cfs_rq_load_avg(&rq
->cfs
);
5665 * cpu_load_without - compute CPU load without any contributions from *p
5666 * @cpu: the CPU which load is requested
5667 * @p: the task which load should be discounted
5669 * The load of a CPU is defined by the load of tasks currently enqueued on that
5670 * CPU as well as tasks which are currently sleeping after an execution on that
5673 * This method returns the load of the specified CPU by discounting the load of
5674 * the specified task, whenever the task is currently contributing to the CPU
5677 static unsigned long cpu_load_without(struct rq
*rq
, struct task_struct
*p
)
5679 struct cfs_rq
*cfs_rq
;
5682 /* Task has no contribution or is new */
5683 if (cpu_of(rq
) != task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
5684 return cpu_load(rq
);
5687 load
= READ_ONCE(cfs_rq
->avg
.load_avg
);
5689 /* Discount task's util from CPU's util */
5690 lsub_positive(&load
, task_h_load(p
));
5695 static unsigned long cpu_runnable(struct rq
*rq
)
5697 return cfs_rq_runnable_avg(&rq
->cfs
);
5700 static unsigned long cpu_runnable_without(struct rq
*rq
, struct task_struct
*p
)
5702 struct cfs_rq
*cfs_rq
;
5703 unsigned int runnable
;
5705 /* Task has no contribution or is new */
5706 if (cpu_of(rq
) != task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
5707 return cpu_runnable(rq
);
5710 runnable
= READ_ONCE(cfs_rq
->avg
.runnable_avg
);
5712 /* Discount task's runnable from CPU's runnable */
5713 lsub_positive(&runnable
, p
->se
.avg
.runnable_avg
);
5718 static unsigned long capacity_of(int cpu
)
5720 return cpu_rq(cpu
)->cpu_capacity
;
5723 static void record_wakee(struct task_struct
*p
)
5726 * Only decay a single time; tasks that have less then 1 wakeup per
5727 * jiffy will not have built up many flips.
5729 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
5730 current
->wakee_flips
>>= 1;
5731 current
->wakee_flip_decay_ts
= jiffies
;
5734 if (current
->last_wakee
!= p
) {
5735 current
->last_wakee
= p
;
5736 current
->wakee_flips
++;
5741 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5743 * A waker of many should wake a different task than the one last awakened
5744 * at a frequency roughly N times higher than one of its wakees.
5746 * In order to determine whether we should let the load spread vs consolidating
5747 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5748 * partner, and a factor of lls_size higher frequency in the other.
5750 * With both conditions met, we can be relatively sure that the relationship is
5751 * non-monogamous, with partner count exceeding socket size.
5753 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5754 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5757 static int wake_wide(struct task_struct
*p
)
5759 unsigned int master
= current
->wakee_flips
;
5760 unsigned int slave
= p
->wakee_flips
;
5761 int factor
= __this_cpu_read(sd_llc_size
);
5764 swap(master
, slave
);
5765 if (slave
< factor
|| master
< slave
* factor
)
5771 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5772 * soonest. For the purpose of speed we only consider the waking and previous
5775 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5776 * cache-affine and is (or will be) idle.
5778 * wake_affine_weight() - considers the weight to reflect the average
5779 * scheduling latency of the CPUs. This seems to work
5780 * for the overloaded case.
5783 wake_affine_idle(int this_cpu
, int prev_cpu
, int sync
)
5786 * If this_cpu is idle, it implies the wakeup is from interrupt
5787 * context. Only allow the move if cache is shared. Otherwise an
5788 * interrupt intensive workload could force all tasks onto one
5789 * node depending on the IO topology or IRQ affinity settings.
5791 * If the prev_cpu is idle and cache affine then avoid a migration.
5792 * There is no guarantee that the cache hot data from an interrupt
5793 * is more important than cache hot data on the prev_cpu and from
5794 * a cpufreq perspective, it's better to have higher utilisation
5797 if (available_idle_cpu(this_cpu
) && cpus_share_cache(this_cpu
, prev_cpu
))
5798 return available_idle_cpu(prev_cpu
) ? prev_cpu
: this_cpu
;
5800 if (sync
&& cpu_rq(this_cpu
)->nr_running
== 1)
5803 return nr_cpumask_bits
;
5807 wake_affine_weight(struct sched_domain
*sd
, struct task_struct
*p
,
5808 int this_cpu
, int prev_cpu
, int sync
)
5810 s64 this_eff_load
, prev_eff_load
;
5811 unsigned long task_load
;
5813 this_eff_load
= cpu_load(cpu_rq(this_cpu
));
5816 unsigned long current_load
= task_h_load(current
);
5818 if (current_load
> this_eff_load
)
5821 this_eff_load
-= current_load
;
5824 task_load
= task_h_load(p
);
5826 this_eff_load
+= task_load
;
5827 if (sched_feat(WA_BIAS
))
5828 this_eff_load
*= 100;
5829 this_eff_load
*= capacity_of(prev_cpu
);
5831 prev_eff_load
= cpu_load(cpu_rq(prev_cpu
));
5832 prev_eff_load
-= task_load
;
5833 if (sched_feat(WA_BIAS
))
5834 prev_eff_load
*= 100 + (sd
->imbalance_pct
- 100) / 2;
5835 prev_eff_load
*= capacity_of(this_cpu
);
5838 * If sync, adjust the weight of prev_eff_load such that if
5839 * prev_eff == this_eff that select_idle_sibling() will consider
5840 * stacking the wakee on top of the waker if no other CPU is
5846 return this_eff_load
< prev_eff_load
? this_cpu
: nr_cpumask_bits
;
5849 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
,
5850 int this_cpu
, int prev_cpu
, int sync
)
5852 int target
= nr_cpumask_bits
;
5854 if (sched_feat(WA_IDLE
))
5855 target
= wake_affine_idle(this_cpu
, prev_cpu
, sync
);
5857 if (sched_feat(WA_WEIGHT
) && target
== nr_cpumask_bits
)
5858 target
= wake_affine_weight(sd
, p
, this_cpu
, prev_cpu
, sync
);
5860 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine_attempts
);
5861 if (target
== nr_cpumask_bits
)
5864 schedstat_inc(sd
->ttwu_move_affine
);
5865 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine
);
5869 static struct sched_group
*
5870 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
);
5873 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5876 find_idlest_group_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
5878 unsigned long load
, min_load
= ULONG_MAX
;
5879 unsigned int min_exit_latency
= UINT_MAX
;
5880 u64 latest_idle_timestamp
= 0;
5881 int least_loaded_cpu
= this_cpu
;
5882 int shallowest_idle_cpu
= -1;
5885 /* Check if we have any choice: */
5886 if (group
->group_weight
== 1)
5887 return cpumask_first(sched_group_span(group
));
5889 /* Traverse only the allowed CPUs */
5890 for_each_cpu_and(i
, sched_group_span(group
), p
->cpus_ptr
) {
5891 if (sched_idle_cpu(i
))
5894 if (available_idle_cpu(i
)) {
5895 struct rq
*rq
= cpu_rq(i
);
5896 struct cpuidle_state
*idle
= idle_get_state(rq
);
5897 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
5899 * We give priority to a CPU whose idle state
5900 * has the smallest exit latency irrespective
5901 * of any idle timestamp.
5903 min_exit_latency
= idle
->exit_latency
;
5904 latest_idle_timestamp
= rq
->idle_stamp
;
5905 shallowest_idle_cpu
= i
;
5906 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
5907 rq
->idle_stamp
> latest_idle_timestamp
) {
5909 * If equal or no active idle state, then
5910 * the most recently idled CPU might have
5913 latest_idle_timestamp
= rq
->idle_stamp
;
5914 shallowest_idle_cpu
= i
;
5916 } else if (shallowest_idle_cpu
== -1) {
5917 load
= cpu_load(cpu_rq(i
));
5918 if (load
< min_load
) {
5920 least_loaded_cpu
= i
;
5925 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
5928 static inline int find_idlest_cpu(struct sched_domain
*sd
, struct task_struct
*p
,
5929 int cpu
, int prev_cpu
, int sd_flag
)
5933 if (!cpumask_intersects(sched_domain_span(sd
), p
->cpus_ptr
))
5937 * We need task's util for cpu_util_without, sync it up to
5938 * prev_cpu's last_update_time.
5940 if (!(sd_flag
& SD_BALANCE_FORK
))
5941 sync_entity_load_avg(&p
->se
);
5944 struct sched_group
*group
;
5945 struct sched_domain
*tmp
;
5948 if (!(sd
->flags
& sd_flag
)) {
5953 group
= find_idlest_group(sd
, p
, cpu
);
5959 new_cpu
= find_idlest_group_cpu(group
, p
, cpu
);
5960 if (new_cpu
== cpu
) {
5961 /* Now try balancing at a lower domain level of 'cpu': */
5966 /* Now try balancing at a lower domain level of 'new_cpu': */
5968 weight
= sd
->span_weight
;
5970 for_each_domain(cpu
, tmp
) {
5971 if (weight
<= tmp
->span_weight
)
5973 if (tmp
->flags
& sd_flag
)
5981 #ifdef CONFIG_SCHED_SMT
5982 DEFINE_STATIC_KEY_FALSE(sched_smt_present
);
5983 EXPORT_SYMBOL_GPL(sched_smt_present
);
5985 static inline void set_idle_cores(int cpu
, int val
)
5987 struct sched_domain_shared
*sds
;
5989 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
5991 WRITE_ONCE(sds
->has_idle_cores
, val
);
5994 static inline bool test_idle_cores(int cpu
, bool def
)
5996 struct sched_domain_shared
*sds
;
5998 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
6000 return READ_ONCE(sds
->has_idle_cores
);
6006 * Scans the local SMT mask to see if the entire core is idle, and records this
6007 * information in sd_llc_shared->has_idle_cores.
6009 * Since SMT siblings share all cache levels, inspecting this limited remote
6010 * state should be fairly cheap.
6012 void __update_idle_core(struct rq
*rq
)
6014 int core
= cpu_of(rq
);
6018 if (test_idle_cores(core
, true))
6021 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
6025 if (!available_idle_cpu(cpu
))
6029 set_idle_cores(core
, 1);
6035 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6036 * there are no idle cores left in the system; tracked through
6037 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6039 static int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6041 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6044 if (!static_branch_likely(&sched_smt_present
))
6047 if (!test_idle_cores(target
, false))
6050 cpumask_and(cpus
, sched_domain_span(sd
), p
->cpus_ptr
);
6052 for_each_cpu_wrap(core
, cpus
, target
) {
6055 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
6056 if (!available_idle_cpu(cpu
)) {
6061 cpumask_andnot(cpus
, cpus
, cpu_smt_mask(core
));
6068 * Failed to find an idle core; stop looking for one.
6070 set_idle_cores(target
, 0);
6076 * Scan the local SMT mask for idle CPUs.
6078 static int select_idle_smt(struct task_struct
*p
, int target
)
6082 if (!static_branch_likely(&sched_smt_present
))
6085 for_each_cpu(cpu
, cpu_smt_mask(target
)) {
6086 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
6088 if (available_idle_cpu(cpu
) || sched_idle_cpu(cpu
))
6095 #else /* CONFIG_SCHED_SMT */
6097 static inline int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6102 static inline int select_idle_smt(struct task_struct
*p
, int target
)
6107 #endif /* CONFIG_SCHED_SMT */
6110 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6111 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6112 * average idle time for this rq (as found in rq->avg_idle).
6114 static int select_idle_cpu(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6116 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6117 struct sched_domain
*this_sd
;
6118 u64 avg_cost
, avg_idle
;
6120 int this = smp_processor_id();
6121 int cpu
, nr
= INT_MAX
;
6123 this_sd
= rcu_dereference(*this_cpu_ptr(&sd_llc
));
6128 * Due to large variance we need a large fuzz factor; hackbench in
6129 * particularly is sensitive here.
6131 avg_idle
= this_rq()->avg_idle
/ 512;
6132 avg_cost
= this_sd
->avg_scan_cost
+ 1;
6134 if (sched_feat(SIS_AVG_CPU
) && avg_idle
< avg_cost
)
6137 if (sched_feat(SIS_PROP
)) {
6138 u64 span_avg
= sd
->span_weight
* avg_idle
;
6139 if (span_avg
> 4*avg_cost
)
6140 nr
= div_u64(span_avg
, avg_cost
);
6145 time
= cpu_clock(this);
6147 cpumask_and(cpus
, sched_domain_span(sd
), p
->cpus_ptr
);
6149 for_each_cpu_wrap(cpu
, cpus
, target
) {
6152 if (available_idle_cpu(cpu
) || sched_idle_cpu(cpu
))
6156 time
= cpu_clock(this) - time
;
6157 update_avg(&this_sd
->avg_scan_cost
, time
);
6163 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6164 * the task fits. If no CPU is big enough, but there are idle ones, try to
6165 * maximize capacity.
6168 select_idle_capacity(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6170 unsigned long best_cap
= 0;
6171 int cpu
, best_cpu
= -1;
6172 struct cpumask
*cpus
;
6174 sync_entity_load_avg(&p
->se
);
6176 cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6177 cpumask_and(cpus
, sched_domain_span(sd
), p
->cpus_ptr
);
6179 for_each_cpu_wrap(cpu
, cpus
, target
) {
6180 unsigned long cpu_cap
= capacity_of(cpu
);
6182 if (!available_idle_cpu(cpu
) && !sched_idle_cpu(cpu
))
6184 if (task_fits_capacity(p
, cpu_cap
))
6187 if (cpu_cap
> best_cap
) {
6197 * Try and locate an idle core/thread in the LLC cache domain.
6199 static int select_idle_sibling(struct task_struct
*p
, int prev
, int target
)
6201 struct sched_domain
*sd
;
6202 int i
, recent_used_cpu
;
6205 * For asymmetric CPU capacity systems, our domain of interest is
6206 * sd_asym_cpucapacity rather than sd_llc.
6208 if (static_branch_unlikely(&sched_asym_cpucapacity
)) {
6209 sd
= rcu_dereference(per_cpu(sd_asym_cpucapacity
, target
));
6211 * On an asymmetric CPU capacity system where an exclusive
6212 * cpuset defines a symmetric island (i.e. one unique
6213 * capacity_orig value through the cpuset), the key will be set
6214 * but the CPUs within that cpuset will not have a domain with
6215 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6221 i
= select_idle_capacity(p
, sd
, target
);
6222 return ((unsigned)i
< nr_cpumask_bits
) ? i
: target
;
6226 if (available_idle_cpu(target
) || sched_idle_cpu(target
))
6230 * If the previous CPU is cache affine and idle, don't be stupid:
6232 if (prev
!= target
&& cpus_share_cache(prev
, target
) &&
6233 (available_idle_cpu(prev
) || sched_idle_cpu(prev
)))
6237 * Allow a per-cpu kthread to stack with the wakee if the
6238 * kworker thread and the tasks previous CPUs are the same.
6239 * The assumption is that the wakee queued work for the
6240 * per-cpu kthread that is now complete and the wakeup is
6241 * essentially a sync wakeup. An obvious example of this
6242 * pattern is IO completions.
6244 if (is_per_cpu_kthread(current
) &&
6245 prev
== smp_processor_id() &&
6246 this_rq()->nr_running
<= 1) {
6250 /* Check a recently used CPU as a potential idle candidate: */
6251 recent_used_cpu
= p
->recent_used_cpu
;
6252 if (recent_used_cpu
!= prev
&&
6253 recent_used_cpu
!= target
&&
6254 cpus_share_cache(recent_used_cpu
, target
) &&
6255 (available_idle_cpu(recent_used_cpu
) || sched_idle_cpu(recent_used_cpu
)) &&
6256 cpumask_test_cpu(p
->recent_used_cpu
, p
->cpus_ptr
)) {
6258 * Replace recent_used_cpu with prev as it is a potential
6259 * candidate for the next wake:
6261 p
->recent_used_cpu
= prev
;
6262 return recent_used_cpu
;
6265 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
6269 i
= select_idle_core(p
, sd
, target
);
6270 if ((unsigned)i
< nr_cpumask_bits
)
6273 i
= select_idle_cpu(p
, sd
, target
);
6274 if ((unsigned)i
< nr_cpumask_bits
)
6277 i
= select_idle_smt(p
, target
);
6278 if ((unsigned)i
< nr_cpumask_bits
)
6285 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6286 * @cpu: the CPU to get the utilization of
6288 * The unit of the return value must be the one of capacity so we can compare
6289 * the utilization with the capacity of the CPU that is available for CFS task
6290 * (ie cpu_capacity).
6292 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6293 * recent utilization of currently non-runnable tasks on a CPU. It represents
6294 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6295 * capacity_orig is the cpu_capacity available at the highest frequency
6296 * (arch_scale_freq_capacity()).
6297 * The utilization of a CPU converges towards a sum equal to or less than the
6298 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6299 * the running time on this CPU scaled by capacity_curr.
6301 * The estimated utilization of a CPU is defined to be the maximum between its
6302 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6303 * currently RUNNABLE on that CPU.
6304 * This allows to properly represent the expected utilization of a CPU which
6305 * has just got a big task running since a long sleep period. At the same time
6306 * however it preserves the benefits of the "blocked utilization" in
6307 * describing the potential for other tasks waking up on the same CPU.
6309 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6310 * higher than capacity_orig because of unfortunate rounding in
6311 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6312 * the average stabilizes with the new running time. We need to check that the
6313 * utilization stays within the range of [0..capacity_orig] and cap it if
6314 * necessary. Without utilization capping, a group could be seen as overloaded
6315 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6316 * available capacity. We allow utilization to overshoot capacity_curr (but not
6317 * capacity_orig) as it useful for predicting the capacity required after task
6318 * migrations (scheduler-driven DVFS).
6320 * Return: the (estimated) utilization for the specified CPU
6322 static inline unsigned long cpu_util(int cpu
)
6324 struct cfs_rq
*cfs_rq
;
6327 cfs_rq
= &cpu_rq(cpu
)->cfs
;
6328 util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6330 if (sched_feat(UTIL_EST
))
6331 util
= max(util
, READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
));
6333 return min_t(unsigned long, util
, capacity_orig_of(cpu
));
6337 * cpu_util_without: compute cpu utilization without any contributions from *p
6338 * @cpu: the CPU which utilization is requested
6339 * @p: the task which utilization should be discounted
6341 * The utilization of a CPU is defined by the utilization of tasks currently
6342 * enqueued on that CPU as well as tasks which are currently sleeping after an
6343 * execution on that CPU.
6345 * This method returns the utilization of the specified CPU by discounting the
6346 * utilization of the specified task, whenever the task is currently
6347 * contributing to the CPU utilization.
6349 static unsigned long cpu_util_without(int cpu
, struct task_struct
*p
)
6351 struct cfs_rq
*cfs_rq
;
6354 /* Task has no contribution or is new */
6355 if (cpu
!= task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
6356 return cpu_util(cpu
);
6358 cfs_rq
= &cpu_rq(cpu
)->cfs
;
6359 util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6361 /* Discount task's util from CPU's util */
6362 lsub_positive(&util
, task_util(p
));
6367 * a) if *p is the only task sleeping on this CPU, then:
6368 * cpu_util (== task_util) > util_est (== 0)
6369 * and thus we return:
6370 * cpu_util_without = (cpu_util - task_util) = 0
6372 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6374 * cpu_util >= task_util
6375 * cpu_util > util_est (== 0)
6376 * and thus we discount *p's blocked utilization to return:
6377 * cpu_util_without = (cpu_util - task_util) >= 0
6379 * c) if other tasks are RUNNABLE on that CPU and
6380 * util_est > cpu_util
6381 * then we use util_est since it returns a more restrictive
6382 * estimation of the spare capacity on that CPU, by just
6383 * considering the expected utilization of tasks already
6384 * runnable on that CPU.
6386 * Cases a) and b) are covered by the above code, while case c) is
6387 * covered by the following code when estimated utilization is
6390 if (sched_feat(UTIL_EST
)) {
6391 unsigned int estimated
=
6392 READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
);
6395 * Despite the following checks we still have a small window
6396 * for a possible race, when an execl's select_task_rq_fair()
6397 * races with LB's detach_task():
6400 * p->on_rq = TASK_ON_RQ_MIGRATING;
6401 * ---------------------------------- A
6402 * deactivate_task() \
6403 * dequeue_task() + RaceTime
6404 * util_est_dequeue() /
6405 * ---------------------------------- B
6407 * The additional check on "current == p" it's required to
6408 * properly fix the execl regression and it helps in further
6409 * reducing the chances for the above race.
6411 if (unlikely(task_on_rq_queued(p
) || current
== p
))
6412 lsub_positive(&estimated
, _task_util_est(p
));
6414 util
= max(util
, estimated
);
6418 * Utilization (estimated) can exceed the CPU capacity, thus let's
6419 * clamp to the maximum CPU capacity to ensure consistency with
6420 * the cpu_util call.
6422 return min_t(unsigned long, util
, capacity_orig_of(cpu
));
6426 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6429 static unsigned long cpu_util_next(int cpu
, struct task_struct
*p
, int dst_cpu
)
6431 struct cfs_rq
*cfs_rq
= &cpu_rq(cpu
)->cfs
;
6432 unsigned long util_est
, util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6435 * If @p migrates from @cpu to another, remove its contribution. Or,
6436 * if @p migrates from another CPU to @cpu, add its contribution. In
6437 * the other cases, @cpu is not impacted by the migration, so the
6438 * util_avg should already be correct.
6440 if (task_cpu(p
) == cpu
&& dst_cpu
!= cpu
)
6441 sub_positive(&util
, task_util(p
));
6442 else if (task_cpu(p
) != cpu
&& dst_cpu
== cpu
)
6443 util
+= task_util(p
);
6445 if (sched_feat(UTIL_EST
)) {
6446 util_est
= READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
);
6449 * During wake-up, the task isn't enqueued yet and doesn't
6450 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6451 * so just add it (if needed) to "simulate" what will be
6452 * cpu_util() after the task has been enqueued.
6455 util_est
+= _task_util_est(p
);
6457 util
= max(util
, util_est
);
6460 return min(util
, capacity_orig_of(cpu
));
6464 * compute_energy(): Estimates the energy that @pd would consume if @p was
6465 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6466 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6467 * to compute what would be the energy if we decided to actually migrate that
6471 compute_energy(struct task_struct
*p
, int dst_cpu
, struct perf_domain
*pd
)
6473 struct cpumask
*pd_mask
= perf_domain_span(pd
);
6474 unsigned long cpu_cap
= arch_scale_cpu_capacity(cpumask_first(pd_mask
));
6475 unsigned long max_util
= 0, sum_util
= 0;
6479 * The capacity state of CPUs of the current rd can be driven by CPUs
6480 * of another rd if they belong to the same pd. So, account for the
6481 * utilization of these CPUs too by masking pd with cpu_online_mask
6482 * instead of the rd span.
6484 * If an entire pd is outside of the current rd, it will not appear in
6485 * its pd list and will not be accounted by compute_energy().
6487 for_each_cpu_and(cpu
, pd_mask
, cpu_online_mask
) {
6488 unsigned long cpu_util
, util_cfs
= cpu_util_next(cpu
, p
, dst_cpu
);
6489 struct task_struct
*tsk
= cpu
== dst_cpu
? p
: NULL
;
6492 * Busy time computation: utilization clamping is not
6493 * required since the ratio (sum_util / cpu_capacity)
6494 * is already enough to scale the EM reported power
6495 * consumption at the (eventually clamped) cpu_capacity.
6497 sum_util
+= schedutil_cpu_util(cpu
, util_cfs
, cpu_cap
,
6501 * Performance domain frequency: utilization clamping
6502 * must be considered since it affects the selection
6503 * of the performance domain frequency.
6504 * NOTE: in case RT tasks are running, by default the
6505 * FREQUENCY_UTIL's utilization can be max OPP.
6507 cpu_util
= schedutil_cpu_util(cpu
, util_cfs
, cpu_cap
,
6508 FREQUENCY_UTIL
, tsk
);
6509 max_util
= max(max_util
, cpu_util
);
6512 return em_cpu_energy(pd
->em_pd
, max_util
, sum_util
);
6516 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6517 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6518 * spare capacity in each performance domain and uses it as a potential
6519 * candidate to execute the task. Then, it uses the Energy Model to figure
6520 * out which of the CPU candidates is the most energy-efficient.
6522 * The rationale for this heuristic is as follows. In a performance domain,
6523 * all the most energy efficient CPU candidates (according to the Energy
6524 * Model) are those for which we'll request a low frequency. When there are
6525 * several CPUs for which the frequency request will be the same, we don't
6526 * have enough data to break the tie between them, because the Energy Model
6527 * only includes active power costs. With this model, if we assume that
6528 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6529 * the maximum spare capacity in a performance domain is guaranteed to be among
6530 * the best candidates of the performance domain.
6532 * In practice, it could be preferable from an energy standpoint to pack
6533 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6534 * but that could also hurt our chances to go cluster idle, and we have no
6535 * ways to tell with the current Energy Model if this is actually a good
6536 * idea or not. So, find_energy_efficient_cpu() basically favors
6537 * cluster-packing, and spreading inside a cluster. That should at least be
6538 * a good thing for latency, and this is consistent with the idea that most
6539 * of the energy savings of EAS come from the asymmetry of the system, and
6540 * not so much from breaking the tie between identical CPUs. That's also the
6541 * reason why EAS is enabled in the topology code only for systems where
6542 * SD_ASYM_CPUCAPACITY is set.
6544 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6545 * they don't have any useful utilization data yet and it's not possible to
6546 * forecast their impact on energy consumption. Consequently, they will be
6547 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6548 * to be energy-inefficient in some use-cases. The alternative would be to
6549 * bias new tasks towards specific types of CPUs first, or to try to infer
6550 * their util_avg from the parent task, but those heuristics could hurt
6551 * other use-cases too. So, until someone finds a better way to solve this,
6552 * let's keep things simple by re-using the existing slow path.
6554 static int find_energy_efficient_cpu(struct task_struct
*p
, int prev_cpu
)
6556 unsigned long prev_delta
= ULONG_MAX
, best_delta
= ULONG_MAX
;
6557 struct root_domain
*rd
= cpu_rq(smp_processor_id())->rd
;
6558 unsigned long cpu_cap
, util
, base_energy
= 0;
6559 int cpu
, best_energy_cpu
= prev_cpu
;
6560 struct sched_domain
*sd
;
6561 struct perf_domain
*pd
;
6564 pd
= rcu_dereference(rd
->pd
);
6565 if (!pd
|| READ_ONCE(rd
->overutilized
))
6569 * Energy-aware wake-up happens on the lowest sched_domain starting
6570 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6572 sd
= rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity
));
6573 while (sd
&& !cpumask_test_cpu(prev_cpu
, sched_domain_span(sd
)))
6578 sync_entity_load_avg(&p
->se
);
6579 if (!task_util_est(p
))
6582 for (; pd
; pd
= pd
->next
) {
6583 unsigned long cur_delta
, spare_cap
, max_spare_cap
= 0;
6584 unsigned long base_energy_pd
;
6585 int max_spare_cap_cpu
= -1;
6587 /* Compute the 'base' energy of the pd, without @p */
6588 base_energy_pd
= compute_energy(p
, -1, pd
);
6589 base_energy
+= base_energy_pd
;
6591 for_each_cpu_and(cpu
, perf_domain_span(pd
), sched_domain_span(sd
)) {
6592 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
6595 util
= cpu_util_next(cpu
, p
, cpu
);
6596 cpu_cap
= capacity_of(cpu
);
6597 spare_cap
= cpu_cap
- util
;
6600 * Skip CPUs that cannot satisfy the capacity request.
6601 * IOW, placing the task there would make the CPU
6602 * overutilized. Take uclamp into account to see how
6603 * much capacity we can get out of the CPU; this is
6604 * aligned with schedutil_cpu_util().
6606 util
= uclamp_rq_util_with(cpu_rq(cpu
), util
, p
);
6607 if (!fits_capacity(util
, cpu_cap
))
6610 /* Always use prev_cpu as a candidate. */
6611 if (cpu
== prev_cpu
) {
6612 prev_delta
= compute_energy(p
, prev_cpu
, pd
);
6613 prev_delta
-= base_energy_pd
;
6614 best_delta
= min(best_delta
, prev_delta
);
6618 * Find the CPU with the maximum spare capacity in
6619 * the performance domain
6621 if (spare_cap
> max_spare_cap
) {
6622 max_spare_cap
= spare_cap
;
6623 max_spare_cap_cpu
= cpu
;
6627 /* Evaluate the energy impact of using this CPU. */
6628 if (max_spare_cap_cpu
>= 0 && max_spare_cap_cpu
!= prev_cpu
) {
6629 cur_delta
= compute_energy(p
, max_spare_cap_cpu
, pd
);
6630 cur_delta
-= base_energy_pd
;
6631 if (cur_delta
< best_delta
) {
6632 best_delta
= cur_delta
;
6633 best_energy_cpu
= max_spare_cap_cpu
;
6641 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6642 * least 6% of the energy used by prev_cpu.
6644 if (prev_delta
== ULONG_MAX
)
6645 return best_energy_cpu
;
6647 if ((prev_delta
- best_delta
) > ((prev_delta
+ base_energy
) >> 4))
6648 return best_energy_cpu
;
6659 * select_task_rq_fair: Select target runqueue for the waking task in domains
6660 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6661 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6663 * Balances load by selecting the idlest CPU in the idlest group, or under
6664 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6666 * Returns the target CPU number.
6668 * preempt must be disabled.
6671 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
6673 struct sched_domain
*tmp
, *sd
= NULL
;
6674 int cpu
= smp_processor_id();
6675 int new_cpu
= prev_cpu
;
6676 int want_affine
= 0;
6677 int sync
= (wake_flags
& WF_SYNC
) && !(current
->flags
& PF_EXITING
);
6679 if (sd_flag
& SD_BALANCE_WAKE
) {
6682 if (sched_energy_enabled()) {
6683 new_cpu
= find_energy_efficient_cpu(p
, prev_cpu
);
6689 want_affine
= !wake_wide(p
) && cpumask_test_cpu(cpu
, p
->cpus_ptr
);
6693 for_each_domain(cpu
, tmp
) {
6695 * If both 'cpu' and 'prev_cpu' are part of this domain,
6696 * cpu is a valid SD_WAKE_AFFINE target.
6698 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
6699 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
6700 if (cpu
!= prev_cpu
)
6701 new_cpu
= wake_affine(tmp
, p
, cpu
, prev_cpu
, sync
);
6703 sd
= NULL
; /* Prefer wake_affine over balance flags */
6707 if (tmp
->flags
& sd_flag
)
6709 else if (!want_affine
)
6715 new_cpu
= find_idlest_cpu(sd
, p
, cpu
, prev_cpu
, sd_flag
);
6716 } else if (sd_flag
& SD_BALANCE_WAKE
) { /* XXX always ? */
6719 new_cpu
= select_idle_sibling(p
, prev_cpu
, new_cpu
);
6722 current
->recent_used_cpu
= cpu
;
6729 static void detach_entity_cfs_rq(struct sched_entity
*se
);
6732 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6733 * cfs_rq_of(p) references at time of call are still valid and identify the
6734 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6736 static void migrate_task_rq_fair(struct task_struct
*p
, int new_cpu
)
6739 * As blocked tasks retain absolute vruntime the migration needs to
6740 * deal with this by subtracting the old and adding the new
6741 * min_vruntime -- the latter is done by enqueue_entity() when placing
6742 * the task on the new runqueue.
6744 if (p
->state
== TASK_WAKING
) {
6745 struct sched_entity
*se
= &p
->se
;
6746 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6749 #ifndef CONFIG_64BIT
6750 u64 min_vruntime_copy
;
6753 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
6755 min_vruntime
= cfs_rq
->min_vruntime
;
6756 } while (min_vruntime
!= min_vruntime_copy
);
6758 min_vruntime
= cfs_rq
->min_vruntime
;
6761 se
->vruntime
-= min_vruntime
;
6764 if (p
->on_rq
== TASK_ON_RQ_MIGRATING
) {
6766 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6767 * rq->lock and can modify state directly.
6769 lockdep_assert_held(&task_rq(p
)->lock
);
6770 detach_entity_cfs_rq(&p
->se
);
6774 * We are supposed to update the task to "current" time, then
6775 * its up to date and ready to go to new CPU/cfs_rq. But we
6776 * have difficulty in getting what current time is, so simply
6777 * throw away the out-of-date time. This will result in the
6778 * wakee task is less decayed, but giving the wakee more load
6781 remove_entity_load_avg(&p
->se
);
6784 /* Tell new CPU we are migrated */
6785 p
->se
.avg
.last_update_time
= 0;
6787 /* We have migrated, no longer consider this task hot */
6788 p
->se
.exec_start
= 0;
6790 update_scan_period(p
, new_cpu
);
6793 static void task_dead_fair(struct task_struct
*p
)
6795 remove_entity_load_avg(&p
->se
);
6799 balance_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
6804 return newidle_balance(rq
, rf
) != 0;
6806 #endif /* CONFIG_SMP */
6808 static unsigned long wakeup_gran(struct sched_entity
*se
)
6810 unsigned long gran
= sysctl_sched_wakeup_granularity
;
6813 * Since its curr running now, convert the gran from real-time
6814 * to virtual-time in his units.
6816 * By using 'se' instead of 'curr' we penalize light tasks, so
6817 * they get preempted easier. That is, if 'se' < 'curr' then
6818 * the resulting gran will be larger, therefore penalizing the
6819 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6820 * be smaller, again penalizing the lighter task.
6822 * This is especially important for buddies when the leftmost
6823 * task is higher priority than the buddy.
6825 return calc_delta_fair(gran
, se
);
6829 * Should 'se' preempt 'curr'.
6843 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
6845 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
6850 gran
= wakeup_gran(se
);
6857 static void set_last_buddy(struct sched_entity
*se
)
6859 if (entity_is_task(se
) && unlikely(task_has_idle_policy(task_of(se
))))
6862 for_each_sched_entity(se
) {
6863 if (SCHED_WARN_ON(!se
->on_rq
))
6865 cfs_rq_of(se
)->last
= se
;
6869 static void set_next_buddy(struct sched_entity
*se
)
6871 if (entity_is_task(se
) && unlikely(task_has_idle_policy(task_of(se
))))
6874 for_each_sched_entity(se
) {
6875 if (SCHED_WARN_ON(!se
->on_rq
))
6877 cfs_rq_of(se
)->next
= se
;
6881 static void set_skip_buddy(struct sched_entity
*se
)
6883 for_each_sched_entity(se
)
6884 cfs_rq_of(se
)->skip
= se
;
6888 * Preempt the current task with a newly woken task if needed:
6890 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
6892 struct task_struct
*curr
= rq
->curr
;
6893 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
6894 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
6895 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
6896 int next_buddy_marked
= 0;
6898 if (unlikely(se
== pse
))
6902 * This is possible from callers such as attach_tasks(), in which we
6903 * unconditionally check_prempt_curr() after an enqueue (which may have
6904 * lead to a throttle). This both saves work and prevents false
6905 * next-buddy nomination below.
6907 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
6910 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
6911 set_next_buddy(pse
);
6912 next_buddy_marked
= 1;
6916 * We can come here with TIF_NEED_RESCHED already set from new task
6919 * Note: this also catches the edge-case of curr being in a throttled
6920 * group (e.g. via set_curr_task), since update_curr() (in the
6921 * enqueue of curr) will have resulted in resched being set. This
6922 * prevents us from potentially nominating it as a false LAST_BUDDY
6925 if (test_tsk_need_resched(curr
))
6928 /* Idle tasks are by definition preempted by non-idle tasks. */
6929 if (unlikely(task_has_idle_policy(curr
)) &&
6930 likely(!task_has_idle_policy(p
)))
6934 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6935 * is driven by the tick):
6937 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
6940 find_matching_se(&se
, &pse
);
6941 update_curr(cfs_rq_of(se
));
6943 if (wakeup_preempt_entity(se
, pse
) == 1) {
6945 * Bias pick_next to pick the sched entity that is
6946 * triggering this preemption.
6948 if (!next_buddy_marked
)
6949 set_next_buddy(pse
);
6958 * Only set the backward buddy when the current task is still
6959 * on the rq. This can happen when a wakeup gets interleaved
6960 * with schedule on the ->pre_schedule() or idle_balance()
6961 * point, either of which can * drop the rq lock.
6963 * Also, during early boot the idle thread is in the fair class,
6964 * for obvious reasons its a bad idea to schedule back to it.
6966 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
6969 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
6973 struct task_struct
*
6974 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
6976 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
6977 struct sched_entity
*se
;
6978 struct task_struct
*p
;
6982 if (!sched_fair_runnable(rq
))
6985 #ifdef CONFIG_FAIR_GROUP_SCHED
6986 if (!prev
|| prev
->sched_class
!= &fair_sched_class
)
6990 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6991 * likely that a next task is from the same cgroup as the current.
6993 * Therefore attempt to avoid putting and setting the entire cgroup
6994 * hierarchy, only change the part that actually changes.
6998 struct sched_entity
*curr
= cfs_rq
->curr
;
7001 * Since we got here without doing put_prev_entity() we also
7002 * have to consider cfs_rq->curr. If it is still a runnable
7003 * entity, update_curr() will update its vruntime, otherwise
7004 * forget we've ever seen it.
7008 update_curr(cfs_rq
);
7013 * This call to check_cfs_rq_runtime() will do the
7014 * throttle and dequeue its entity in the parent(s).
7015 * Therefore the nr_running test will indeed
7018 if (unlikely(check_cfs_rq_runtime(cfs_rq
))) {
7021 if (!cfs_rq
->nr_running
)
7028 se
= pick_next_entity(cfs_rq
, curr
);
7029 cfs_rq
= group_cfs_rq(se
);
7035 * Since we haven't yet done put_prev_entity and if the selected task
7036 * is a different task than we started out with, try and touch the
7037 * least amount of cfs_rqs.
7040 struct sched_entity
*pse
= &prev
->se
;
7042 while (!(cfs_rq
= is_same_group(se
, pse
))) {
7043 int se_depth
= se
->depth
;
7044 int pse_depth
= pse
->depth
;
7046 if (se_depth
<= pse_depth
) {
7047 put_prev_entity(cfs_rq_of(pse
), pse
);
7048 pse
= parent_entity(pse
);
7050 if (se_depth
>= pse_depth
) {
7051 set_next_entity(cfs_rq_of(se
), se
);
7052 se
= parent_entity(se
);
7056 put_prev_entity(cfs_rq
, pse
);
7057 set_next_entity(cfs_rq
, se
);
7064 put_prev_task(rq
, prev
);
7067 se
= pick_next_entity(cfs_rq
, NULL
);
7068 set_next_entity(cfs_rq
, se
);
7069 cfs_rq
= group_cfs_rq(se
);
7074 done
: __maybe_unused
;
7077 * Move the next running task to the front of
7078 * the list, so our cfs_tasks list becomes MRU
7081 list_move(&p
->se
.group_node
, &rq
->cfs_tasks
);
7084 if (hrtick_enabled(rq
))
7085 hrtick_start_fair(rq
, p
);
7087 update_misfit_status(p
, rq
);
7095 new_tasks
= newidle_balance(rq
, rf
);
7098 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7099 * possible for any higher priority task to appear. In that case we
7100 * must re-start the pick_next_entity() loop.
7109 * rq is about to be idle, check if we need to update the
7110 * lost_idle_time of clock_pelt
7112 update_idle_rq_clock_pelt(rq
);
7117 static struct task_struct
*__pick_next_task_fair(struct rq
*rq
)
7119 return pick_next_task_fair(rq
, NULL
, NULL
);
7123 * Account for a descheduled task:
7125 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
7127 struct sched_entity
*se
= &prev
->se
;
7128 struct cfs_rq
*cfs_rq
;
7130 for_each_sched_entity(se
) {
7131 cfs_rq
= cfs_rq_of(se
);
7132 put_prev_entity(cfs_rq
, se
);
7137 * sched_yield() is very simple
7139 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7141 static void yield_task_fair(struct rq
*rq
)
7143 struct task_struct
*curr
= rq
->curr
;
7144 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
7145 struct sched_entity
*se
= &curr
->se
;
7148 * Are we the only task in the tree?
7150 if (unlikely(rq
->nr_running
== 1))
7153 clear_buddies(cfs_rq
, se
);
7155 if (curr
->policy
!= SCHED_BATCH
) {
7156 update_rq_clock(rq
);
7158 * Update run-time statistics of the 'current'.
7160 update_curr(cfs_rq
);
7162 * Tell update_rq_clock() that we've just updated,
7163 * so we don't do microscopic update in schedule()
7164 * and double the fastpath cost.
7166 rq_clock_skip_update(rq
);
7172 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
)
7174 struct sched_entity
*se
= &p
->se
;
7176 /* throttled hierarchies are not runnable */
7177 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
7180 /* Tell the scheduler that we'd really like pse to run next. */
7183 yield_task_fair(rq
);
7189 /**************************************************
7190 * Fair scheduling class load-balancing methods.
7194 * The purpose of load-balancing is to achieve the same basic fairness the
7195 * per-CPU scheduler provides, namely provide a proportional amount of compute
7196 * time to each task. This is expressed in the following equation:
7198 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7200 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7201 * W_i,0 is defined as:
7203 * W_i,0 = \Sum_j w_i,j (2)
7205 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7206 * is derived from the nice value as per sched_prio_to_weight[].
7208 * The weight average is an exponential decay average of the instantaneous
7211 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7213 * C_i is the compute capacity of CPU i, typically it is the
7214 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7215 * can also include other factors [XXX].
7217 * To achieve this balance we define a measure of imbalance which follows
7218 * directly from (1):
7220 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7222 * We them move tasks around to minimize the imbalance. In the continuous
7223 * function space it is obvious this converges, in the discrete case we get
7224 * a few fun cases generally called infeasible weight scenarios.
7227 * - infeasible weights;
7228 * - local vs global optima in the discrete case. ]
7233 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7234 * for all i,j solution, we create a tree of CPUs that follows the hardware
7235 * topology where each level pairs two lower groups (or better). This results
7236 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7237 * tree to only the first of the previous level and we decrease the frequency
7238 * of load-balance at each level inv. proportional to the number of CPUs in
7244 * \Sum { --- * --- * 2^i } = O(n) (5)
7246 * `- size of each group
7247 * | | `- number of CPUs doing load-balance
7249 * `- sum over all levels
7251 * Coupled with a limit on how many tasks we can migrate every balance pass,
7252 * this makes (5) the runtime complexity of the balancer.
7254 * An important property here is that each CPU is still (indirectly) connected
7255 * to every other CPU in at most O(log n) steps:
7257 * The adjacency matrix of the resulting graph is given by:
7260 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7263 * And you'll find that:
7265 * A^(log_2 n)_i,j != 0 for all i,j (7)
7267 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7268 * The task movement gives a factor of O(m), giving a convergence complexity
7271 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7276 * In order to avoid CPUs going idle while there's still work to do, new idle
7277 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7278 * tree itself instead of relying on other CPUs to bring it work.
7280 * This adds some complexity to both (5) and (8) but it reduces the total idle
7288 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7291 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7296 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7298 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7300 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7303 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7304 * rewrite all of this once again.]
7307 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
7309 enum fbq_type
{ regular
, remote
, all
};
7312 * 'group_type' describes the group of CPUs at the moment of load balancing.
7314 * The enum is ordered by pulling priority, with the group with lowest priority
7315 * first so the group_type can simply be compared when selecting the busiest
7316 * group. See update_sd_pick_busiest().
7319 /* The group has spare capacity that can be used to run more tasks. */
7320 group_has_spare
= 0,
7322 * The group is fully used and the tasks don't compete for more CPU
7323 * cycles. Nevertheless, some tasks might wait before running.
7327 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
7328 * and must be migrated to a more powerful CPU.
7332 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7333 * and the task should be migrated to it instead of running on the
7338 * The tasks' affinity constraints previously prevented the scheduler
7339 * from balancing the load across the system.
7343 * The CPU is overloaded and can't provide expected CPU cycles to all
7349 enum migration_type
{
7356 #define LBF_ALL_PINNED 0x01
7357 #define LBF_NEED_BREAK 0x02
7358 #define LBF_DST_PINNED 0x04
7359 #define LBF_SOME_PINNED 0x08
7360 #define LBF_NOHZ_STATS 0x10
7361 #define LBF_NOHZ_AGAIN 0x20
7364 struct sched_domain
*sd
;
7372 struct cpumask
*dst_grpmask
;
7374 enum cpu_idle_type idle
;
7376 /* The set of CPUs under consideration for load-balancing */
7377 struct cpumask
*cpus
;
7382 unsigned int loop_break
;
7383 unsigned int loop_max
;
7385 enum fbq_type fbq_type
;
7386 enum migration_type migration_type
;
7387 struct list_head tasks
;
7391 * Is this task likely cache-hot:
7393 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
7397 lockdep_assert_held(&env
->src_rq
->lock
);
7399 if (p
->sched_class
!= &fair_sched_class
)
7402 if (unlikely(task_has_idle_policy(p
)))
7406 * Buddy candidates are cache hot:
7408 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
7409 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
7410 &p
->se
== cfs_rq_of(&p
->se
)->last
))
7413 if (sysctl_sched_migration_cost
== -1)
7415 if (sysctl_sched_migration_cost
== 0)
7418 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
7420 return delta
< (s64
)sysctl_sched_migration_cost
;
7423 #ifdef CONFIG_NUMA_BALANCING
7425 * Returns 1, if task migration degrades locality
7426 * Returns 0, if task migration improves locality i.e migration preferred.
7427 * Returns -1, if task migration is not affected by locality.
7429 static int migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
7431 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
7432 unsigned long src_weight
, dst_weight
;
7433 int src_nid
, dst_nid
, dist
;
7435 if (!static_branch_likely(&sched_numa_balancing
))
7438 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
7441 src_nid
= cpu_to_node(env
->src_cpu
);
7442 dst_nid
= cpu_to_node(env
->dst_cpu
);
7444 if (src_nid
== dst_nid
)
7447 /* Migrating away from the preferred node is always bad. */
7448 if (src_nid
== p
->numa_preferred_nid
) {
7449 if (env
->src_rq
->nr_running
> env
->src_rq
->nr_preferred_running
)
7455 /* Encourage migration to the preferred node. */
7456 if (dst_nid
== p
->numa_preferred_nid
)
7459 /* Leaving a core idle is often worse than degrading locality. */
7460 if (env
->idle
== CPU_IDLE
)
7463 dist
= node_distance(src_nid
, dst_nid
);
7465 src_weight
= group_weight(p
, src_nid
, dist
);
7466 dst_weight
= group_weight(p
, dst_nid
, dist
);
7468 src_weight
= task_weight(p
, src_nid
, dist
);
7469 dst_weight
= task_weight(p
, dst_nid
, dist
);
7472 return dst_weight
< src_weight
;
7476 static inline int migrate_degrades_locality(struct task_struct
*p
,
7484 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7487 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
7491 lockdep_assert_held(&env
->src_rq
->lock
);
7494 * We do not migrate tasks that are:
7495 * 1) throttled_lb_pair, or
7496 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7497 * 3) running (obviously), or
7498 * 4) are cache-hot on their current CPU.
7500 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
7503 if (!cpumask_test_cpu(env
->dst_cpu
, p
->cpus_ptr
)) {
7506 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_affine
);
7508 env
->flags
|= LBF_SOME_PINNED
;
7511 * Remember if this task can be migrated to any other CPU in
7512 * our sched_group. We may want to revisit it if we couldn't
7513 * meet load balance goals by pulling other tasks on src_cpu.
7515 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7516 * already computed one in current iteration.
7518 if (env
->idle
== CPU_NEWLY_IDLE
|| (env
->flags
& LBF_DST_PINNED
))
7521 /* Prevent to re-select dst_cpu via env's CPUs: */
7522 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
7523 if (cpumask_test_cpu(cpu
, p
->cpus_ptr
)) {
7524 env
->flags
|= LBF_DST_PINNED
;
7525 env
->new_dst_cpu
= cpu
;
7533 /* Record that we found atleast one task that could run on dst_cpu */
7534 env
->flags
&= ~LBF_ALL_PINNED
;
7536 if (task_running(env
->src_rq
, p
)) {
7537 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_running
);
7542 * Aggressive migration if:
7543 * 1) destination numa is preferred
7544 * 2) task is cache cold, or
7545 * 3) too many balance attempts have failed.
7547 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
7548 if (tsk_cache_hot
== -1)
7549 tsk_cache_hot
= task_hot(p
, env
);
7551 if (tsk_cache_hot
<= 0 ||
7552 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
7553 if (tsk_cache_hot
== 1) {
7554 schedstat_inc(env
->sd
->lb_hot_gained
[env
->idle
]);
7555 schedstat_inc(p
->se
.statistics
.nr_forced_migrations
);
7560 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_hot
);
7565 * detach_task() -- detach the task for the migration specified in env
7567 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
7569 lockdep_assert_held(&env
->src_rq
->lock
);
7571 deactivate_task(env
->src_rq
, p
, DEQUEUE_NOCLOCK
);
7572 set_task_cpu(p
, env
->dst_cpu
);
7576 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7577 * part of active balancing operations within "domain".
7579 * Returns a task if successful and NULL otherwise.
7581 static struct task_struct
*detach_one_task(struct lb_env
*env
)
7583 struct task_struct
*p
;
7585 lockdep_assert_held(&env
->src_rq
->lock
);
7587 list_for_each_entry_reverse(p
,
7588 &env
->src_rq
->cfs_tasks
, se
.group_node
) {
7589 if (!can_migrate_task(p
, env
))
7592 detach_task(p
, env
);
7595 * Right now, this is only the second place where
7596 * lb_gained[env->idle] is updated (other is detach_tasks)
7597 * so we can safely collect stats here rather than
7598 * inside detach_tasks().
7600 schedstat_inc(env
->sd
->lb_gained
[env
->idle
]);
7606 static const unsigned int sched_nr_migrate_break
= 32;
7609 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7610 * busiest_rq, as part of a balancing operation within domain "sd".
7612 * Returns number of detached tasks if successful and 0 otherwise.
7614 static int detach_tasks(struct lb_env
*env
)
7616 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
7617 unsigned long util
, load
;
7618 struct task_struct
*p
;
7621 lockdep_assert_held(&env
->src_rq
->lock
);
7623 if (env
->imbalance
<= 0)
7626 while (!list_empty(tasks
)) {
7628 * We don't want to steal all, otherwise we may be treated likewise,
7629 * which could at worst lead to a livelock crash.
7631 if (env
->idle
!= CPU_NOT_IDLE
&& env
->src_rq
->nr_running
<= 1)
7634 p
= list_last_entry(tasks
, struct task_struct
, se
.group_node
);
7637 /* We've more or less seen every task there is, call it quits */
7638 if (env
->loop
> env
->loop_max
)
7641 /* take a breather every nr_migrate tasks */
7642 if (env
->loop
> env
->loop_break
) {
7643 env
->loop_break
+= sched_nr_migrate_break
;
7644 env
->flags
|= LBF_NEED_BREAK
;
7648 if (!can_migrate_task(p
, env
))
7651 switch (env
->migration_type
) {
7654 * Depending of the number of CPUs and tasks and the
7655 * cgroup hierarchy, task_h_load() can return a null
7656 * value. Make sure that env->imbalance decreases
7657 * otherwise detach_tasks() will stop only after
7658 * detaching up to loop_max tasks.
7660 load
= max_t(unsigned long, task_h_load(p
), 1);
7662 if (sched_feat(LB_MIN
) &&
7663 load
< 16 && !env
->sd
->nr_balance_failed
)
7667 * Make sure that we don't migrate too much load.
7668 * Nevertheless, let relax the constraint if
7669 * scheduler fails to find a good waiting task to
7672 if (load
/2 > env
->imbalance
&&
7673 env
->sd
->nr_balance_failed
<= env
->sd
->cache_nice_tries
)
7676 env
->imbalance
-= load
;
7680 util
= task_util_est(p
);
7682 if (util
> env
->imbalance
)
7685 env
->imbalance
-= util
;
7692 case migrate_misfit
:
7693 /* This is not a misfit task */
7694 if (task_fits_capacity(p
, capacity_of(env
->src_cpu
)))
7701 detach_task(p
, env
);
7702 list_add(&p
->se
.group_node
, &env
->tasks
);
7706 #ifdef CONFIG_PREEMPTION
7708 * NEWIDLE balancing is a source of latency, so preemptible
7709 * kernels will stop after the first task is detached to minimize
7710 * the critical section.
7712 if (env
->idle
== CPU_NEWLY_IDLE
)
7717 * We only want to steal up to the prescribed amount of
7720 if (env
->imbalance
<= 0)
7725 list_move(&p
->se
.group_node
, tasks
);
7729 * Right now, this is one of only two places we collect this stat
7730 * so we can safely collect detach_one_task() stats here rather
7731 * than inside detach_one_task().
7733 schedstat_add(env
->sd
->lb_gained
[env
->idle
], detached
);
7739 * attach_task() -- attach the task detached by detach_task() to its new rq.
7741 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
7743 lockdep_assert_held(&rq
->lock
);
7745 BUG_ON(task_rq(p
) != rq
);
7746 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
7747 check_preempt_curr(rq
, p
, 0);
7751 * attach_one_task() -- attaches the task returned from detach_one_task() to
7754 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
7759 update_rq_clock(rq
);
7765 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7768 static void attach_tasks(struct lb_env
*env
)
7770 struct list_head
*tasks
= &env
->tasks
;
7771 struct task_struct
*p
;
7774 rq_lock(env
->dst_rq
, &rf
);
7775 update_rq_clock(env
->dst_rq
);
7777 while (!list_empty(tasks
)) {
7778 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
7779 list_del_init(&p
->se
.group_node
);
7781 attach_task(env
->dst_rq
, p
);
7784 rq_unlock(env
->dst_rq
, &rf
);
7787 #ifdef CONFIG_NO_HZ_COMMON
7788 static inline bool cfs_rq_has_blocked(struct cfs_rq
*cfs_rq
)
7790 if (cfs_rq
->avg
.load_avg
)
7793 if (cfs_rq
->avg
.util_avg
)
7799 static inline bool others_have_blocked(struct rq
*rq
)
7801 if (READ_ONCE(rq
->avg_rt
.util_avg
))
7804 if (READ_ONCE(rq
->avg_dl
.util_avg
))
7807 if (thermal_load_avg(rq
))
7810 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7811 if (READ_ONCE(rq
->avg_irq
.util_avg
))
7818 static inline void update_blocked_load_status(struct rq
*rq
, bool has_blocked
)
7820 rq
->last_blocked_load_update_tick
= jiffies
;
7823 rq
->has_blocked_load
= 0;
7826 static inline bool cfs_rq_has_blocked(struct cfs_rq
*cfs_rq
) { return false; }
7827 static inline bool others_have_blocked(struct rq
*rq
) { return false; }
7828 static inline void update_blocked_load_status(struct rq
*rq
, bool has_blocked
) {}
7831 static bool __update_blocked_others(struct rq
*rq
, bool *done
)
7833 const struct sched_class
*curr_class
;
7834 u64 now
= rq_clock_pelt(rq
);
7835 unsigned long thermal_pressure
;
7839 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
7840 * DL and IRQ signals have been updated before updating CFS.
7842 curr_class
= rq
->curr
->sched_class
;
7844 thermal_pressure
= arch_scale_thermal_pressure(cpu_of(rq
));
7846 decayed
= update_rt_rq_load_avg(now
, rq
, curr_class
== &rt_sched_class
) |
7847 update_dl_rq_load_avg(now
, rq
, curr_class
== &dl_sched_class
) |
7848 update_thermal_load_avg(rq_clock_thermal(rq
), rq
, thermal_pressure
) |
7849 update_irq_load_avg(rq
, 0);
7851 if (others_have_blocked(rq
))
7857 #ifdef CONFIG_FAIR_GROUP_SCHED
7859 static inline bool cfs_rq_is_decayed(struct cfs_rq
*cfs_rq
)
7861 if (cfs_rq
->load
.weight
)
7864 if (cfs_rq
->avg
.load_sum
)
7867 if (cfs_rq
->avg
.util_sum
)
7870 if (cfs_rq
->avg
.runnable_sum
)
7876 static bool __update_blocked_fair(struct rq
*rq
, bool *done
)
7878 struct cfs_rq
*cfs_rq
, *pos
;
7879 bool decayed
= false;
7880 int cpu
= cpu_of(rq
);
7883 * Iterates the task_group tree in a bottom up fashion, see
7884 * list_add_leaf_cfs_rq() for details.
7886 for_each_leaf_cfs_rq_safe(rq
, cfs_rq
, pos
) {
7887 struct sched_entity
*se
;
7889 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq
), cfs_rq
)) {
7890 update_tg_load_avg(cfs_rq
, 0);
7892 if (cfs_rq
== &rq
->cfs
)
7896 /* Propagate pending load changes to the parent, if any: */
7897 se
= cfs_rq
->tg
->se
[cpu
];
7898 if (se
&& !skip_blocked_update(se
))
7899 update_load_avg(cfs_rq_of(se
), se
, 0);
7902 * There can be a lot of idle CPU cgroups. Don't let fully
7903 * decayed cfs_rqs linger on the list.
7905 if (cfs_rq_is_decayed(cfs_rq
))
7906 list_del_leaf_cfs_rq(cfs_rq
);
7908 /* Don't need periodic decay once load/util_avg are null */
7909 if (cfs_rq_has_blocked(cfs_rq
))
7917 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7918 * This needs to be done in a top-down fashion because the load of a child
7919 * group is a fraction of its parents load.
7921 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
7923 struct rq
*rq
= rq_of(cfs_rq
);
7924 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
7925 unsigned long now
= jiffies
;
7928 if (cfs_rq
->last_h_load_update
== now
)
7931 WRITE_ONCE(cfs_rq
->h_load_next
, NULL
);
7932 for_each_sched_entity(se
) {
7933 cfs_rq
= cfs_rq_of(se
);
7934 WRITE_ONCE(cfs_rq
->h_load_next
, se
);
7935 if (cfs_rq
->last_h_load_update
== now
)
7940 cfs_rq
->h_load
= cfs_rq_load_avg(cfs_rq
);
7941 cfs_rq
->last_h_load_update
= now
;
7944 while ((se
= READ_ONCE(cfs_rq
->h_load_next
)) != NULL
) {
7945 load
= cfs_rq
->h_load
;
7946 load
= div64_ul(load
* se
->avg
.load_avg
,
7947 cfs_rq_load_avg(cfs_rq
) + 1);
7948 cfs_rq
= group_cfs_rq(se
);
7949 cfs_rq
->h_load
= load
;
7950 cfs_rq
->last_h_load_update
= now
;
7954 static unsigned long task_h_load(struct task_struct
*p
)
7956 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
7958 update_cfs_rq_h_load(cfs_rq
);
7959 return div64_ul(p
->se
.avg
.load_avg
* cfs_rq
->h_load
,
7960 cfs_rq_load_avg(cfs_rq
) + 1);
7963 static bool __update_blocked_fair(struct rq
*rq
, bool *done
)
7965 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
7968 decayed
= update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq
), cfs_rq
);
7969 if (cfs_rq_has_blocked(cfs_rq
))
7975 static unsigned long task_h_load(struct task_struct
*p
)
7977 return p
->se
.avg
.load_avg
;
7981 static void update_blocked_averages(int cpu
)
7983 bool decayed
= false, done
= true;
7984 struct rq
*rq
= cpu_rq(cpu
);
7987 rq_lock_irqsave(rq
, &rf
);
7988 update_rq_clock(rq
);
7990 decayed
|= __update_blocked_others(rq
, &done
);
7991 decayed
|= __update_blocked_fair(rq
, &done
);
7993 update_blocked_load_status(rq
, !done
);
7995 cpufreq_update_util(rq
, 0);
7996 rq_unlock_irqrestore(rq
, &rf
);
7999 /********** Helpers for find_busiest_group ************************/
8002 * sg_lb_stats - stats of a sched_group required for load_balancing
8004 struct sg_lb_stats
{
8005 unsigned long avg_load
; /*Avg load across the CPUs of the group */
8006 unsigned long group_load
; /* Total load over the CPUs of the group */
8007 unsigned long group_capacity
;
8008 unsigned long group_util
; /* Total utilization over the CPUs of the group */
8009 unsigned long group_runnable
; /* Total runnable time over the CPUs of the group */
8010 unsigned int sum_nr_running
; /* Nr of tasks running in the group */
8011 unsigned int sum_h_nr_running
; /* Nr of CFS tasks running in the group */
8012 unsigned int idle_cpus
;
8013 unsigned int group_weight
;
8014 enum group_type group_type
;
8015 unsigned int group_asym_packing
; /* Tasks should be moved to preferred CPU */
8016 unsigned long group_misfit_task_load
; /* A CPU has a task too big for its capacity */
8017 #ifdef CONFIG_NUMA_BALANCING
8018 unsigned int nr_numa_running
;
8019 unsigned int nr_preferred_running
;
8024 * sd_lb_stats - Structure to store the statistics of a sched_domain
8025 * during load balancing.
8027 struct sd_lb_stats
{
8028 struct sched_group
*busiest
; /* Busiest group in this sd */
8029 struct sched_group
*local
; /* Local group in this sd */
8030 unsigned long total_load
; /* Total load of all groups in sd */
8031 unsigned long total_capacity
; /* Total capacity of all groups in sd */
8032 unsigned long avg_load
; /* Average load across all groups in sd */
8033 unsigned int prefer_sibling
; /* tasks should go to sibling first */
8035 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
8036 struct sg_lb_stats local_stat
; /* Statistics of the local group */
8039 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
8042 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8043 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8044 * We must however set busiest_stat::group_type and
8045 * busiest_stat::idle_cpus to the worst busiest group because
8046 * update_sd_pick_busiest() reads these before assignment.
8048 *sds
= (struct sd_lb_stats
){
8052 .total_capacity
= 0UL,
8054 .idle_cpus
= UINT_MAX
,
8055 .group_type
= group_has_spare
,
8060 static unsigned long scale_rt_capacity(int cpu
)
8062 struct rq
*rq
= cpu_rq(cpu
);
8063 unsigned long max
= arch_scale_cpu_capacity(cpu
);
8064 unsigned long used
, free
;
8067 irq
= cpu_util_irq(rq
);
8069 if (unlikely(irq
>= max
))
8073 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8074 * (running and not running) with weights 0 and 1024 respectively.
8075 * avg_thermal.load_avg tracks thermal pressure and the weighted
8076 * average uses the actual delta max capacity(load).
8078 used
= READ_ONCE(rq
->avg_rt
.util_avg
);
8079 used
+= READ_ONCE(rq
->avg_dl
.util_avg
);
8080 used
+= thermal_load_avg(rq
);
8082 if (unlikely(used
>= max
))
8087 return scale_irq_capacity(free
, irq
, max
);
8090 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
8092 unsigned long capacity
= scale_rt_capacity(cpu
);
8093 struct sched_group
*sdg
= sd
->groups
;
8095 cpu_rq(cpu
)->cpu_capacity_orig
= arch_scale_cpu_capacity(cpu
);
8100 cpu_rq(cpu
)->cpu_capacity
= capacity
;
8101 sdg
->sgc
->capacity
= capacity
;
8102 sdg
->sgc
->min_capacity
= capacity
;
8103 sdg
->sgc
->max_capacity
= capacity
;
8106 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
8108 struct sched_domain
*child
= sd
->child
;
8109 struct sched_group
*group
, *sdg
= sd
->groups
;
8110 unsigned long capacity
, min_capacity
, max_capacity
;
8111 unsigned long interval
;
8113 interval
= msecs_to_jiffies(sd
->balance_interval
);
8114 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
8115 sdg
->sgc
->next_update
= jiffies
+ interval
;
8118 update_cpu_capacity(sd
, cpu
);
8123 min_capacity
= ULONG_MAX
;
8126 if (child
->flags
& SD_OVERLAP
) {
8128 * SD_OVERLAP domains cannot assume that child groups
8129 * span the current group.
8132 for_each_cpu(cpu
, sched_group_span(sdg
)) {
8133 unsigned long cpu_cap
= capacity_of(cpu
);
8135 capacity
+= cpu_cap
;
8136 min_capacity
= min(cpu_cap
, min_capacity
);
8137 max_capacity
= max(cpu_cap
, max_capacity
);
8141 * !SD_OVERLAP domains can assume that child groups
8142 * span the current group.
8145 group
= child
->groups
;
8147 struct sched_group_capacity
*sgc
= group
->sgc
;
8149 capacity
+= sgc
->capacity
;
8150 min_capacity
= min(sgc
->min_capacity
, min_capacity
);
8151 max_capacity
= max(sgc
->max_capacity
, max_capacity
);
8152 group
= group
->next
;
8153 } while (group
!= child
->groups
);
8156 sdg
->sgc
->capacity
= capacity
;
8157 sdg
->sgc
->min_capacity
= min_capacity
;
8158 sdg
->sgc
->max_capacity
= max_capacity
;
8162 * Check whether the capacity of the rq has been noticeably reduced by side
8163 * activity. The imbalance_pct is used for the threshold.
8164 * Return true is the capacity is reduced
8167 check_cpu_capacity(struct rq
*rq
, struct sched_domain
*sd
)
8169 return ((rq
->cpu_capacity
* sd
->imbalance_pct
) <
8170 (rq
->cpu_capacity_orig
* 100));
8174 * Check whether a rq has a misfit task and if it looks like we can actually
8175 * help that task: we can migrate the task to a CPU of higher capacity, or
8176 * the task's current CPU is heavily pressured.
8178 static inline int check_misfit_status(struct rq
*rq
, struct sched_domain
*sd
)
8180 return rq
->misfit_task_load
&&
8181 (rq
->cpu_capacity_orig
< rq
->rd
->max_cpu_capacity
||
8182 check_cpu_capacity(rq
, sd
));
8186 * Group imbalance indicates (and tries to solve) the problem where balancing
8187 * groups is inadequate due to ->cpus_ptr constraints.
8189 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8190 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8193 * { 0 1 2 3 } { 4 5 6 7 }
8196 * If we were to balance group-wise we'd place two tasks in the first group and
8197 * two tasks in the second group. Clearly this is undesired as it will overload
8198 * cpu 3 and leave one of the CPUs in the second group unused.
8200 * The current solution to this issue is detecting the skew in the first group
8201 * by noticing the lower domain failed to reach balance and had difficulty
8202 * moving tasks due to affinity constraints.
8204 * When this is so detected; this group becomes a candidate for busiest; see
8205 * update_sd_pick_busiest(). And calculate_imbalance() and
8206 * find_busiest_group() avoid some of the usual balance conditions to allow it
8207 * to create an effective group imbalance.
8209 * This is a somewhat tricky proposition since the next run might not find the
8210 * group imbalance and decide the groups need to be balanced again. A most
8211 * subtle and fragile situation.
8214 static inline int sg_imbalanced(struct sched_group
*group
)
8216 return group
->sgc
->imbalance
;
8220 * group_has_capacity returns true if the group has spare capacity that could
8221 * be used by some tasks.
8222 * We consider that a group has spare capacity if the * number of task is
8223 * smaller than the number of CPUs or if the utilization is lower than the
8224 * available capacity for CFS tasks.
8225 * For the latter, we use a threshold to stabilize the state, to take into
8226 * account the variance of the tasks' load and to return true if the available
8227 * capacity in meaningful for the load balancer.
8228 * As an example, an available capacity of 1% can appear but it doesn't make
8229 * any benefit for the load balance.
8232 group_has_capacity(unsigned int imbalance_pct
, struct sg_lb_stats
*sgs
)
8234 if (sgs
->sum_nr_running
< sgs
->group_weight
)
8237 if ((sgs
->group_capacity
* imbalance_pct
) <
8238 (sgs
->group_runnable
* 100))
8241 if ((sgs
->group_capacity
* 100) >
8242 (sgs
->group_util
* imbalance_pct
))
8249 * group_is_overloaded returns true if the group has more tasks than it can
8251 * group_is_overloaded is not equals to !group_has_capacity because a group
8252 * with the exact right number of tasks, has no more spare capacity but is not
8253 * overloaded so both group_has_capacity and group_is_overloaded return
8257 group_is_overloaded(unsigned int imbalance_pct
, struct sg_lb_stats
*sgs
)
8259 if (sgs
->sum_nr_running
<= sgs
->group_weight
)
8262 if ((sgs
->group_capacity
* 100) <
8263 (sgs
->group_util
* imbalance_pct
))
8266 if ((sgs
->group_capacity
* imbalance_pct
) <
8267 (sgs
->group_runnable
* 100))
8274 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
8275 * per-CPU capacity than sched_group ref.
8278 group_smaller_min_cpu_capacity(struct sched_group
*sg
, struct sched_group
*ref
)
8280 return fits_capacity(sg
->sgc
->min_capacity
, ref
->sgc
->min_capacity
);
8284 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
8285 * per-CPU capacity_orig than sched_group ref.
8288 group_smaller_max_cpu_capacity(struct sched_group
*sg
, struct sched_group
*ref
)
8290 return fits_capacity(sg
->sgc
->max_capacity
, ref
->sgc
->max_capacity
);
8294 group_type
group_classify(unsigned int imbalance_pct
,
8295 struct sched_group
*group
,
8296 struct sg_lb_stats
*sgs
)
8298 if (group_is_overloaded(imbalance_pct
, sgs
))
8299 return group_overloaded
;
8301 if (sg_imbalanced(group
))
8302 return group_imbalanced
;
8304 if (sgs
->group_asym_packing
)
8305 return group_asym_packing
;
8307 if (sgs
->group_misfit_task_load
)
8308 return group_misfit_task
;
8310 if (!group_has_capacity(imbalance_pct
, sgs
))
8311 return group_fully_busy
;
8313 return group_has_spare
;
8316 static bool update_nohz_stats(struct rq
*rq
, bool force
)
8318 #ifdef CONFIG_NO_HZ_COMMON
8319 unsigned int cpu
= rq
->cpu
;
8321 if (!rq
->has_blocked_load
)
8324 if (!cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))
8327 if (!force
&& !time_after(jiffies
, rq
->last_blocked_load_update_tick
))
8330 update_blocked_averages(cpu
);
8332 return rq
->has_blocked_load
;
8339 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8340 * @env: The load balancing environment.
8341 * @group: sched_group whose statistics are to be updated.
8342 * @sgs: variable to hold the statistics for this group.
8343 * @sg_status: Holds flag indicating the status of the sched_group
8345 static inline void update_sg_lb_stats(struct lb_env
*env
,
8346 struct sched_group
*group
,
8347 struct sg_lb_stats
*sgs
,
8350 int i
, nr_running
, local_group
;
8352 memset(sgs
, 0, sizeof(*sgs
));
8354 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_span(group
));
8356 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
8357 struct rq
*rq
= cpu_rq(i
);
8359 if ((env
->flags
& LBF_NOHZ_STATS
) && update_nohz_stats(rq
, false))
8360 env
->flags
|= LBF_NOHZ_AGAIN
;
8362 sgs
->group_load
+= cpu_load(rq
);
8363 sgs
->group_util
+= cpu_util(i
);
8364 sgs
->group_runnable
+= cpu_runnable(rq
);
8365 sgs
->sum_h_nr_running
+= rq
->cfs
.h_nr_running
;
8367 nr_running
= rq
->nr_running
;
8368 sgs
->sum_nr_running
+= nr_running
;
8371 *sg_status
|= SG_OVERLOAD
;
8373 if (cpu_overutilized(i
))
8374 *sg_status
|= SG_OVERUTILIZED
;
8376 #ifdef CONFIG_NUMA_BALANCING
8377 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
8378 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
8381 * No need to call idle_cpu() if nr_running is not 0
8383 if (!nr_running
&& idle_cpu(i
)) {
8385 /* Idle cpu can't have misfit task */
8392 /* Check for a misfit task on the cpu */
8393 if (env
->sd
->flags
& SD_ASYM_CPUCAPACITY
&&
8394 sgs
->group_misfit_task_load
< rq
->misfit_task_load
) {
8395 sgs
->group_misfit_task_load
= rq
->misfit_task_load
;
8396 *sg_status
|= SG_OVERLOAD
;
8400 /* Check if dst CPU is idle and preferred to this group */
8401 if (env
->sd
->flags
& SD_ASYM_PACKING
&&
8402 env
->idle
!= CPU_NOT_IDLE
&&
8403 sgs
->sum_h_nr_running
&&
8404 sched_asym_prefer(env
->dst_cpu
, group
->asym_prefer_cpu
)) {
8405 sgs
->group_asym_packing
= 1;
8408 sgs
->group_capacity
= group
->sgc
->capacity
;
8410 sgs
->group_weight
= group
->group_weight
;
8412 sgs
->group_type
= group_classify(env
->sd
->imbalance_pct
, group
, sgs
);
8414 /* Computing avg_load makes sense only when group is overloaded */
8415 if (sgs
->group_type
== group_overloaded
)
8416 sgs
->avg_load
= (sgs
->group_load
* SCHED_CAPACITY_SCALE
) /
8417 sgs
->group_capacity
;
8421 * update_sd_pick_busiest - return 1 on busiest group
8422 * @env: The load balancing environment.
8423 * @sds: sched_domain statistics
8424 * @sg: sched_group candidate to be checked for being the busiest
8425 * @sgs: sched_group statistics
8427 * Determine if @sg is a busier group than the previously selected
8430 * Return: %true if @sg is a busier group than the previously selected
8431 * busiest group. %false otherwise.
8433 static bool update_sd_pick_busiest(struct lb_env
*env
,
8434 struct sd_lb_stats
*sds
,
8435 struct sched_group
*sg
,
8436 struct sg_lb_stats
*sgs
)
8438 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
8440 /* Make sure that there is at least one task to pull */
8441 if (!sgs
->sum_h_nr_running
)
8445 * Don't try to pull misfit tasks we can't help.
8446 * We can use max_capacity here as reduction in capacity on some
8447 * CPUs in the group should either be possible to resolve
8448 * internally or be covered by avg_load imbalance (eventually).
8450 if (sgs
->group_type
== group_misfit_task
&&
8451 (!group_smaller_max_cpu_capacity(sg
, sds
->local
) ||
8452 sds
->local_stat
.group_type
!= group_has_spare
))
8455 if (sgs
->group_type
> busiest
->group_type
)
8458 if (sgs
->group_type
< busiest
->group_type
)
8462 * The candidate and the current busiest group are the same type of
8463 * group. Let check which one is the busiest according to the type.
8466 switch (sgs
->group_type
) {
8467 case group_overloaded
:
8468 /* Select the overloaded group with highest avg_load. */
8469 if (sgs
->avg_load
<= busiest
->avg_load
)
8473 case group_imbalanced
:
8475 * Select the 1st imbalanced group as we don't have any way to
8476 * choose one more than another.
8480 case group_asym_packing
:
8481 /* Prefer to move from lowest priority CPU's work */
8482 if (sched_asym_prefer(sg
->asym_prefer_cpu
, sds
->busiest
->asym_prefer_cpu
))
8486 case group_misfit_task
:
8488 * If we have more than one misfit sg go with the biggest
8491 if (sgs
->group_misfit_task_load
< busiest
->group_misfit_task_load
)
8495 case group_fully_busy
:
8497 * Select the fully busy group with highest avg_load. In
8498 * theory, there is no need to pull task from such kind of
8499 * group because tasks have all compute capacity that they need
8500 * but we can still improve the overall throughput by reducing
8501 * contention when accessing shared HW resources.
8503 * XXX for now avg_load is not computed and always 0 so we
8504 * select the 1st one.
8506 if (sgs
->avg_load
<= busiest
->avg_load
)
8510 case group_has_spare
:
8512 * Select not overloaded group with lowest number of idle cpus
8513 * and highest number of running tasks. We could also compare
8514 * the spare capacity which is more stable but it can end up
8515 * that the group has less spare capacity but finally more idle
8516 * CPUs which means less opportunity to pull tasks.
8518 if (sgs
->idle_cpus
> busiest
->idle_cpus
)
8520 else if ((sgs
->idle_cpus
== busiest
->idle_cpus
) &&
8521 (sgs
->sum_nr_running
<= busiest
->sum_nr_running
))
8528 * Candidate sg has no more than one task per CPU and has higher
8529 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8530 * throughput. Maximize throughput, power/energy consequences are not
8533 if ((env
->sd
->flags
& SD_ASYM_CPUCAPACITY
) &&
8534 (sgs
->group_type
<= group_fully_busy
) &&
8535 (group_smaller_min_cpu_capacity(sds
->local
, sg
)))
8541 #ifdef CONFIG_NUMA_BALANCING
8542 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
8544 if (sgs
->sum_h_nr_running
> sgs
->nr_numa_running
)
8546 if (sgs
->sum_h_nr_running
> sgs
->nr_preferred_running
)
8551 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
8553 if (rq
->nr_running
> rq
->nr_numa_running
)
8555 if (rq
->nr_running
> rq
->nr_preferred_running
)
8560 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
8565 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
8569 #endif /* CONFIG_NUMA_BALANCING */
8575 * task_running_on_cpu - return 1 if @p is running on @cpu.
8578 static unsigned int task_running_on_cpu(int cpu
, struct task_struct
*p
)
8580 /* Task has no contribution or is new */
8581 if (cpu
!= task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
8584 if (task_on_rq_queued(p
))
8591 * idle_cpu_without - would a given CPU be idle without p ?
8592 * @cpu: the processor on which idleness is tested.
8593 * @p: task which should be ignored.
8595 * Return: 1 if the CPU would be idle. 0 otherwise.
8597 static int idle_cpu_without(int cpu
, struct task_struct
*p
)
8599 struct rq
*rq
= cpu_rq(cpu
);
8601 if (rq
->curr
!= rq
->idle
&& rq
->curr
!= p
)
8605 * rq->nr_running can't be used but an updated version without the
8606 * impact of p on cpu must be used instead. The updated nr_running
8607 * be computed and tested before calling idle_cpu_without().
8611 if (rq
->ttwu_pending
)
8619 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8620 * @sd: The sched_domain level to look for idlest group.
8621 * @group: sched_group whose statistics are to be updated.
8622 * @sgs: variable to hold the statistics for this group.
8623 * @p: The task for which we look for the idlest group/CPU.
8625 static inline void update_sg_wakeup_stats(struct sched_domain
*sd
,
8626 struct sched_group
*group
,
8627 struct sg_lb_stats
*sgs
,
8628 struct task_struct
*p
)
8632 memset(sgs
, 0, sizeof(*sgs
));
8634 for_each_cpu(i
, sched_group_span(group
)) {
8635 struct rq
*rq
= cpu_rq(i
);
8638 sgs
->group_load
+= cpu_load_without(rq
, p
);
8639 sgs
->group_util
+= cpu_util_without(i
, p
);
8640 sgs
->group_runnable
+= cpu_runnable_without(rq
, p
);
8641 local
= task_running_on_cpu(i
, p
);
8642 sgs
->sum_h_nr_running
+= rq
->cfs
.h_nr_running
- local
;
8644 nr_running
= rq
->nr_running
- local
;
8645 sgs
->sum_nr_running
+= nr_running
;
8648 * No need to call idle_cpu_without() if nr_running is not 0
8650 if (!nr_running
&& idle_cpu_without(i
, p
))
8655 /* Check if task fits in the group */
8656 if (sd
->flags
& SD_ASYM_CPUCAPACITY
&&
8657 !task_fits_capacity(p
, group
->sgc
->max_capacity
)) {
8658 sgs
->group_misfit_task_load
= 1;
8661 sgs
->group_capacity
= group
->sgc
->capacity
;
8663 sgs
->group_weight
= group
->group_weight
;
8665 sgs
->group_type
= group_classify(sd
->imbalance_pct
, group
, sgs
);
8668 * Computing avg_load makes sense only when group is fully busy or
8671 if (sgs
->group_type
== group_fully_busy
||
8672 sgs
->group_type
== group_overloaded
)
8673 sgs
->avg_load
= (sgs
->group_load
* SCHED_CAPACITY_SCALE
) /
8674 sgs
->group_capacity
;
8677 static bool update_pick_idlest(struct sched_group
*idlest
,
8678 struct sg_lb_stats
*idlest_sgs
,
8679 struct sched_group
*group
,
8680 struct sg_lb_stats
*sgs
)
8682 if (sgs
->group_type
< idlest_sgs
->group_type
)
8685 if (sgs
->group_type
> idlest_sgs
->group_type
)
8689 * The candidate and the current idlest group are the same type of
8690 * group. Let check which one is the idlest according to the type.
8693 switch (sgs
->group_type
) {
8694 case group_overloaded
:
8695 case group_fully_busy
:
8696 /* Select the group with lowest avg_load. */
8697 if (idlest_sgs
->avg_load
<= sgs
->avg_load
)
8701 case group_imbalanced
:
8702 case group_asym_packing
:
8703 /* Those types are not used in the slow wakeup path */
8706 case group_misfit_task
:
8707 /* Select group with the highest max capacity */
8708 if (idlest
->sgc
->max_capacity
>= group
->sgc
->max_capacity
)
8712 case group_has_spare
:
8713 /* Select group with most idle CPUs */
8714 if (idlest_sgs
->idle_cpus
> sgs
->idle_cpus
)
8717 /* Select group with lowest group_util */
8718 if (idlest_sgs
->idle_cpus
== sgs
->idle_cpus
&&
8719 idlest_sgs
->group_util
<= sgs
->group_util
)
8729 * find_idlest_group() finds and returns the least busy CPU group within the
8732 * Assumes p is allowed on at least one CPU in sd.
8734 static struct sched_group
*
8735 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
8737 struct sched_group
*idlest
= NULL
, *local
= NULL
, *group
= sd
->groups
;
8738 struct sg_lb_stats local_sgs
, tmp_sgs
;
8739 struct sg_lb_stats
*sgs
;
8740 unsigned long imbalance
;
8741 struct sg_lb_stats idlest_sgs
= {
8742 .avg_load
= UINT_MAX
,
8743 .group_type
= group_overloaded
,
8746 imbalance
= scale_load_down(NICE_0_LOAD
) *
8747 (sd
->imbalance_pct
-100) / 100;
8752 /* Skip over this group if it has no CPUs allowed */
8753 if (!cpumask_intersects(sched_group_span(group
),
8757 local_group
= cpumask_test_cpu(this_cpu
,
8758 sched_group_span(group
));
8767 update_sg_wakeup_stats(sd
, group
, sgs
, p
);
8769 if (!local_group
&& update_pick_idlest(idlest
, &idlest_sgs
, group
, sgs
)) {
8774 } while (group
= group
->next
, group
!= sd
->groups
);
8777 /* There is no idlest group to push tasks to */
8781 /* The local group has been skipped because of CPU affinity */
8786 * If the local group is idler than the selected idlest group
8787 * don't try and push the task.
8789 if (local_sgs
.group_type
< idlest_sgs
.group_type
)
8793 * If the local group is busier than the selected idlest group
8794 * try and push the task.
8796 if (local_sgs
.group_type
> idlest_sgs
.group_type
)
8799 switch (local_sgs
.group_type
) {
8800 case group_overloaded
:
8801 case group_fully_busy
:
8803 * When comparing groups across NUMA domains, it's possible for
8804 * the local domain to be very lightly loaded relative to the
8805 * remote domains but "imbalance" skews the comparison making
8806 * remote CPUs look much more favourable. When considering
8807 * cross-domain, add imbalance to the load on the remote node
8808 * and consider staying local.
8811 if ((sd
->flags
& SD_NUMA
) &&
8812 ((idlest_sgs
.avg_load
+ imbalance
) >= local_sgs
.avg_load
))
8816 * If the local group is less loaded than the selected
8817 * idlest group don't try and push any tasks.
8819 if (idlest_sgs
.avg_load
>= (local_sgs
.avg_load
+ imbalance
))
8822 if (100 * local_sgs
.avg_load
<= sd
->imbalance_pct
* idlest_sgs
.avg_load
)
8826 case group_imbalanced
:
8827 case group_asym_packing
:
8828 /* Those type are not used in the slow wakeup path */
8831 case group_misfit_task
:
8832 /* Select group with the highest max capacity */
8833 if (local
->sgc
->max_capacity
>= idlest
->sgc
->max_capacity
)
8837 case group_has_spare
:
8838 if (sd
->flags
& SD_NUMA
) {
8839 #ifdef CONFIG_NUMA_BALANCING
8842 * If there is spare capacity at NUMA, try to select
8843 * the preferred node
8845 if (cpu_to_node(this_cpu
) == p
->numa_preferred_nid
)
8848 idlest_cpu
= cpumask_first(sched_group_span(idlest
));
8849 if (cpu_to_node(idlest_cpu
) == p
->numa_preferred_nid
)
8853 * Otherwise, keep the task on this node to stay close
8854 * its wakeup source and improve locality. If there is
8855 * a real need of migration, periodic load balance will
8858 if (local_sgs
.idle_cpus
)
8863 * Select group with highest number of idle CPUs. We could also
8864 * compare the utilization which is more stable but it can end
8865 * up that the group has less spare capacity but finally more
8866 * idle CPUs which means more opportunity to run task.
8868 if (local_sgs
.idle_cpus
>= idlest_sgs
.idle_cpus
)
8877 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8878 * @env: The load balancing environment.
8879 * @sds: variable to hold the statistics for this sched_domain.
8882 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
8884 struct sched_domain
*child
= env
->sd
->child
;
8885 struct sched_group
*sg
= env
->sd
->groups
;
8886 struct sg_lb_stats
*local
= &sds
->local_stat
;
8887 struct sg_lb_stats tmp_sgs
;
8890 #ifdef CONFIG_NO_HZ_COMMON
8891 if (env
->idle
== CPU_NEWLY_IDLE
&& READ_ONCE(nohz
.has_blocked
))
8892 env
->flags
|= LBF_NOHZ_STATS
;
8896 struct sg_lb_stats
*sgs
= &tmp_sgs
;
8899 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_span(sg
));
8904 if (env
->idle
!= CPU_NEWLY_IDLE
||
8905 time_after_eq(jiffies
, sg
->sgc
->next_update
))
8906 update_group_capacity(env
->sd
, env
->dst_cpu
);
8909 update_sg_lb_stats(env
, sg
, sgs
, &sg_status
);
8915 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
8917 sds
->busiest_stat
= *sgs
;
8921 /* Now, start updating sd_lb_stats */
8922 sds
->total_load
+= sgs
->group_load
;
8923 sds
->total_capacity
+= sgs
->group_capacity
;
8926 } while (sg
!= env
->sd
->groups
);
8928 /* Tag domain that child domain prefers tasks go to siblings first */
8929 sds
->prefer_sibling
= child
&& child
->flags
& SD_PREFER_SIBLING
;
8931 #ifdef CONFIG_NO_HZ_COMMON
8932 if ((env
->flags
& LBF_NOHZ_AGAIN
) &&
8933 cpumask_subset(nohz
.idle_cpus_mask
, sched_domain_span(env
->sd
))) {
8935 WRITE_ONCE(nohz
.next_blocked
,
8936 jiffies
+ msecs_to_jiffies(LOAD_AVG_PERIOD
));
8940 if (env
->sd
->flags
& SD_NUMA
)
8941 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
8943 if (!env
->sd
->parent
) {
8944 struct root_domain
*rd
= env
->dst_rq
->rd
;
8946 /* update overload indicator if we are at root domain */
8947 WRITE_ONCE(rd
->overload
, sg_status
& SG_OVERLOAD
);
8949 /* Update over-utilization (tipping point, U >= 0) indicator */
8950 WRITE_ONCE(rd
->overutilized
, sg_status
& SG_OVERUTILIZED
);
8951 trace_sched_overutilized_tp(rd
, sg_status
& SG_OVERUTILIZED
);
8952 } else if (sg_status
& SG_OVERUTILIZED
) {
8953 struct root_domain
*rd
= env
->dst_rq
->rd
;
8955 WRITE_ONCE(rd
->overutilized
, SG_OVERUTILIZED
);
8956 trace_sched_overutilized_tp(rd
, SG_OVERUTILIZED
);
8960 static inline long adjust_numa_imbalance(int imbalance
, int src_nr_running
)
8962 unsigned int imbalance_min
;
8965 * Allow a small imbalance based on a simple pair of communicating
8966 * tasks that remain local when the source domain is almost idle.
8969 if (src_nr_running
<= imbalance_min
)
8976 * calculate_imbalance - Calculate the amount of imbalance present within the
8977 * groups of a given sched_domain during load balance.
8978 * @env: load balance environment
8979 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8981 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
8983 struct sg_lb_stats
*local
, *busiest
;
8985 local
= &sds
->local_stat
;
8986 busiest
= &sds
->busiest_stat
;
8988 if (busiest
->group_type
== group_misfit_task
) {
8989 /* Set imbalance to allow misfit tasks to be balanced. */
8990 env
->migration_type
= migrate_misfit
;
8995 if (busiest
->group_type
== group_asym_packing
) {
8997 * In case of asym capacity, we will try to migrate all load to
8998 * the preferred CPU.
9000 env
->migration_type
= migrate_task
;
9001 env
->imbalance
= busiest
->sum_h_nr_running
;
9005 if (busiest
->group_type
== group_imbalanced
) {
9007 * In the group_imb case we cannot rely on group-wide averages
9008 * to ensure CPU-load equilibrium, try to move any task to fix
9009 * the imbalance. The next load balance will take care of
9010 * balancing back the system.
9012 env
->migration_type
= migrate_task
;
9018 * Try to use spare capacity of local group without overloading it or
9021 if (local
->group_type
== group_has_spare
) {
9022 if (busiest
->group_type
> group_fully_busy
) {
9024 * If busiest is overloaded, try to fill spare
9025 * capacity. This might end up creating spare capacity
9026 * in busiest or busiest still being overloaded but
9027 * there is no simple way to directly compute the
9028 * amount of load to migrate in order to balance the
9031 env
->migration_type
= migrate_util
;
9032 env
->imbalance
= max(local
->group_capacity
, local
->group_util
) -
9036 * In some cases, the group's utilization is max or even
9037 * higher than capacity because of migrations but the
9038 * local CPU is (newly) idle. There is at least one
9039 * waiting task in this overloaded busiest group. Let's
9042 if (env
->idle
!= CPU_NOT_IDLE
&& env
->imbalance
== 0) {
9043 env
->migration_type
= migrate_task
;
9050 if (busiest
->group_weight
== 1 || sds
->prefer_sibling
) {
9051 unsigned int nr_diff
= busiest
->sum_nr_running
;
9053 * When prefer sibling, evenly spread running tasks on
9056 env
->migration_type
= migrate_task
;
9057 lsub_positive(&nr_diff
, local
->sum_nr_running
);
9058 env
->imbalance
= nr_diff
>> 1;
9062 * If there is no overload, we just want to even the number of
9065 env
->migration_type
= migrate_task
;
9066 env
->imbalance
= max_t(long, 0, (local
->idle_cpus
-
9067 busiest
->idle_cpus
) >> 1);
9070 /* Consider allowing a small imbalance between NUMA groups */
9071 if (env
->sd
->flags
& SD_NUMA
)
9072 env
->imbalance
= adjust_numa_imbalance(env
->imbalance
,
9073 busiest
->sum_nr_running
);
9079 * Local is fully busy but has to take more load to relieve the
9082 if (local
->group_type
< group_overloaded
) {
9084 * Local will become overloaded so the avg_load metrics are
9088 local
->avg_load
= (local
->group_load
* SCHED_CAPACITY_SCALE
) /
9089 local
->group_capacity
;
9091 sds
->avg_load
= (sds
->total_load
* SCHED_CAPACITY_SCALE
) /
9092 sds
->total_capacity
;
9094 * If the local group is more loaded than the selected
9095 * busiest group don't try to pull any tasks.
9097 if (local
->avg_load
>= busiest
->avg_load
) {
9104 * Both group are or will become overloaded and we're trying to get all
9105 * the CPUs to the average_load, so we don't want to push ourselves
9106 * above the average load, nor do we wish to reduce the max loaded CPU
9107 * below the average load. At the same time, we also don't want to
9108 * reduce the group load below the group capacity. Thus we look for
9109 * the minimum possible imbalance.
9111 env
->migration_type
= migrate_load
;
9112 env
->imbalance
= min(
9113 (busiest
->avg_load
- sds
->avg_load
) * busiest
->group_capacity
,
9114 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
9115 ) / SCHED_CAPACITY_SCALE
;
9118 /******* find_busiest_group() helpers end here *********************/
9121 * Decision matrix according to the local and busiest group type:
9123 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9124 * has_spare nr_idle balanced N/A N/A balanced balanced
9125 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9126 * misfit_task force N/A N/A N/A force force
9127 * asym_packing force force N/A N/A force force
9128 * imbalanced force force N/A N/A force force
9129 * overloaded force force N/A N/A force avg_load
9131 * N/A : Not Applicable because already filtered while updating
9133 * balanced : The system is balanced for these 2 groups.
9134 * force : Calculate the imbalance as load migration is probably needed.
9135 * avg_load : Only if imbalance is significant enough.
9136 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9137 * different in groups.
9141 * find_busiest_group - Returns the busiest group within the sched_domain
9142 * if there is an imbalance.
9144 * Also calculates the amount of runnable load which should be moved
9145 * to restore balance.
9147 * @env: The load balancing environment.
9149 * Return: - The busiest group if imbalance exists.
9151 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
9153 struct sg_lb_stats
*local
, *busiest
;
9154 struct sd_lb_stats sds
;
9156 init_sd_lb_stats(&sds
);
9159 * Compute the various statistics relevant for load balancing at
9162 update_sd_lb_stats(env
, &sds
);
9164 if (sched_energy_enabled()) {
9165 struct root_domain
*rd
= env
->dst_rq
->rd
;
9167 if (rcu_dereference(rd
->pd
) && !READ_ONCE(rd
->overutilized
))
9171 local
= &sds
.local_stat
;
9172 busiest
= &sds
.busiest_stat
;
9174 /* There is no busy sibling group to pull tasks from */
9178 /* Misfit tasks should be dealt with regardless of the avg load */
9179 if (busiest
->group_type
== group_misfit_task
)
9182 /* ASYM feature bypasses nice load balance check */
9183 if (busiest
->group_type
== group_asym_packing
)
9187 * If the busiest group is imbalanced the below checks don't
9188 * work because they assume all things are equal, which typically
9189 * isn't true due to cpus_ptr constraints and the like.
9191 if (busiest
->group_type
== group_imbalanced
)
9195 * If the local group is busier than the selected busiest group
9196 * don't try and pull any tasks.
9198 if (local
->group_type
> busiest
->group_type
)
9202 * When groups are overloaded, use the avg_load to ensure fairness
9205 if (local
->group_type
== group_overloaded
) {
9207 * If the local group is more loaded than the selected
9208 * busiest group don't try to pull any tasks.
9210 if (local
->avg_load
>= busiest
->avg_load
)
9213 /* XXX broken for overlapping NUMA groups */
9214 sds
.avg_load
= (sds
.total_load
* SCHED_CAPACITY_SCALE
) /
9218 * Don't pull any tasks if this group is already above the
9219 * domain average load.
9221 if (local
->avg_load
>= sds
.avg_load
)
9225 * If the busiest group is more loaded, use imbalance_pct to be
9228 if (100 * busiest
->avg_load
<=
9229 env
->sd
->imbalance_pct
* local
->avg_load
)
9233 /* Try to move all excess tasks to child's sibling domain */
9234 if (sds
.prefer_sibling
&& local
->group_type
== group_has_spare
&&
9235 busiest
->sum_nr_running
> local
->sum_nr_running
+ 1)
9238 if (busiest
->group_type
!= group_overloaded
) {
9239 if (env
->idle
== CPU_NOT_IDLE
)
9241 * If the busiest group is not overloaded (and as a
9242 * result the local one too) but this CPU is already
9243 * busy, let another idle CPU try to pull task.
9247 if (busiest
->group_weight
> 1 &&
9248 local
->idle_cpus
<= (busiest
->idle_cpus
+ 1))
9250 * If the busiest group is not overloaded
9251 * and there is no imbalance between this and busiest
9252 * group wrt idle CPUs, it is balanced. The imbalance
9253 * becomes significant if the diff is greater than 1
9254 * otherwise we might end up to just move the imbalance
9255 * on another group. Of course this applies only if
9256 * there is more than 1 CPU per group.
9260 if (busiest
->sum_h_nr_running
== 1)
9262 * busiest doesn't have any tasks waiting to run
9268 /* Looks like there is an imbalance. Compute it */
9269 calculate_imbalance(env
, &sds
);
9270 return env
->imbalance
? sds
.busiest
: NULL
;
9278 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9280 static struct rq
*find_busiest_queue(struct lb_env
*env
,
9281 struct sched_group
*group
)
9283 struct rq
*busiest
= NULL
, *rq
;
9284 unsigned long busiest_util
= 0, busiest_load
= 0, busiest_capacity
= 1;
9285 unsigned int busiest_nr
= 0;
9288 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
9289 unsigned long capacity
, load
, util
;
9290 unsigned int nr_running
;
9294 rt
= fbq_classify_rq(rq
);
9297 * We classify groups/runqueues into three groups:
9298 * - regular: there are !numa tasks
9299 * - remote: there are numa tasks that run on the 'wrong' node
9300 * - all: there is no distinction
9302 * In order to avoid migrating ideally placed numa tasks,
9303 * ignore those when there's better options.
9305 * If we ignore the actual busiest queue to migrate another
9306 * task, the next balance pass can still reduce the busiest
9307 * queue by moving tasks around inside the node.
9309 * If we cannot move enough load due to this classification
9310 * the next pass will adjust the group classification and
9311 * allow migration of more tasks.
9313 * Both cases only affect the total convergence complexity.
9315 if (rt
> env
->fbq_type
)
9318 capacity
= capacity_of(i
);
9319 nr_running
= rq
->cfs
.h_nr_running
;
9322 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9323 * eventually lead to active_balancing high->low capacity.
9324 * Higher per-CPU capacity is considered better than balancing
9327 if (env
->sd
->flags
& SD_ASYM_CPUCAPACITY
&&
9328 capacity_of(env
->dst_cpu
) < capacity
&&
9332 switch (env
->migration_type
) {
9335 * When comparing with load imbalance, use cpu_load()
9336 * which is not scaled with the CPU capacity.
9338 load
= cpu_load(rq
);
9340 if (nr_running
== 1 && load
> env
->imbalance
&&
9341 !check_cpu_capacity(rq
, env
->sd
))
9345 * For the load comparisons with the other CPUs,
9346 * consider the cpu_load() scaled with the CPU
9347 * capacity, so that the load can be moved away
9348 * from the CPU that is potentially running at a
9351 * Thus we're looking for max(load_i / capacity_i),
9352 * crosswise multiplication to rid ourselves of the
9353 * division works out to:
9354 * load_i * capacity_j > load_j * capacity_i;
9355 * where j is our previous maximum.
9357 if (load
* busiest_capacity
> busiest_load
* capacity
) {
9358 busiest_load
= load
;
9359 busiest_capacity
= capacity
;
9365 util
= cpu_util(cpu_of(rq
));
9368 * Don't try to pull utilization from a CPU with one
9369 * running task. Whatever its utilization, we will fail
9372 if (nr_running
<= 1)
9375 if (busiest_util
< util
) {
9376 busiest_util
= util
;
9382 if (busiest_nr
< nr_running
) {
9383 busiest_nr
= nr_running
;
9388 case migrate_misfit
:
9390 * For ASYM_CPUCAPACITY domains with misfit tasks we
9391 * simply seek the "biggest" misfit task.
9393 if (rq
->misfit_task_load
> busiest_load
) {
9394 busiest_load
= rq
->misfit_task_load
;
9407 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9408 * so long as it is large enough.
9410 #define MAX_PINNED_INTERVAL 512
9413 asym_active_balance(struct lb_env
*env
)
9416 * ASYM_PACKING needs to force migrate tasks from busy but
9417 * lower priority CPUs in order to pack all tasks in the
9418 * highest priority CPUs.
9420 return env
->idle
!= CPU_NOT_IDLE
&& (env
->sd
->flags
& SD_ASYM_PACKING
) &&
9421 sched_asym_prefer(env
->dst_cpu
, env
->src_cpu
);
9425 voluntary_active_balance(struct lb_env
*env
)
9427 struct sched_domain
*sd
= env
->sd
;
9429 if (asym_active_balance(env
))
9433 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
9434 * It's worth migrating the task if the src_cpu's capacity is reduced
9435 * because of other sched_class or IRQs if more capacity stays
9436 * available on dst_cpu.
9438 if ((env
->idle
!= CPU_NOT_IDLE
) &&
9439 (env
->src_rq
->cfs
.h_nr_running
== 1)) {
9440 if ((check_cpu_capacity(env
->src_rq
, sd
)) &&
9441 (capacity_of(env
->src_cpu
)*sd
->imbalance_pct
< capacity_of(env
->dst_cpu
)*100))
9445 if (env
->migration_type
== migrate_misfit
)
9451 static int need_active_balance(struct lb_env
*env
)
9453 struct sched_domain
*sd
= env
->sd
;
9455 if (voluntary_active_balance(env
))
9458 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
9461 static int active_load_balance_cpu_stop(void *data
);
9463 static int should_we_balance(struct lb_env
*env
)
9465 struct sched_group
*sg
= env
->sd
->groups
;
9469 * Ensure the balancing environment is consistent; can happen
9470 * when the softirq triggers 'during' hotplug.
9472 if (!cpumask_test_cpu(env
->dst_cpu
, env
->cpus
))
9476 * In the newly idle case, we will allow all the CPUs
9477 * to do the newly idle load balance.
9479 if (env
->idle
== CPU_NEWLY_IDLE
)
9482 /* Try to find first idle CPU */
9483 for_each_cpu_and(cpu
, group_balance_mask(sg
), env
->cpus
) {
9487 /* Are we the first idle CPU? */
9488 return cpu
== env
->dst_cpu
;
9491 /* Are we the first CPU of this group ? */
9492 return group_balance_cpu(sg
) == env
->dst_cpu
;
9496 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9497 * tasks if there is an imbalance.
9499 static int load_balance(int this_cpu
, struct rq
*this_rq
,
9500 struct sched_domain
*sd
, enum cpu_idle_type idle
,
9501 int *continue_balancing
)
9503 int ld_moved
, cur_ld_moved
, active_balance
= 0;
9504 struct sched_domain
*sd_parent
= sd
->parent
;
9505 struct sched_group
*group
;
9508 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
9510 struct lb_env env
= {
9512 .dst_cpu
= this_cpu
,
9514 .dst_grpmask
= sched_group_span(sd
->groups
),
9516 .loop_break
= sched_nr_migrate_break
,
9519 .tasks
= LIST_HEAD_INIT(env
.tasks
),
9522 cpumask_and(cpus
, sched_domain_span(sd
), cpu_active_mask
);
9524 schedstat_inc(sd
->lb_count
[idle
]);
9527 if (!should_we_balance(&env
)) {
9528 *continue_balancing
= 0;
9532 group
= find_busiest_group(&env
);
9534 schedstat_inc(sd
->lb_nobusyg
[idle
]);
9538 busiest
= find_busiest_queue(&env
, group
);
9540 schedstat_inc(sd
->lb_nobusyq
[idle
]);
9544 BUG_ON(busiest
== env
.dst_rq
);
9546 schedstat_add(sd
->lb_imbalance
[idle
], env
.imbalance
);
9548 env
.src_cpu
= busiest
->cpu
;
9549 env
.src_rq
= busiest
;
9552 if (busiest
->nr_running
> 1) {
9554 * Attempt to move tasks. If find_busiest_group has found
9555 * an imbalance but busiest->nr_running <= 1, the group is
9556 * still unbalanced. ld_moved simply stays zero, so it is
9557 * correctly treated as an imbalance.
9559 env
.flags
|= LBF_ALL_PINNED
;
9560 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
9563 rq_lock_irqsave(busiest
, &rf
);
9564 update_rq_clock(busiest
);
9567 * cur_ld_moved - load moved in current iteration
9568 * ld_moved - cumulative load moved across iterations
9570 cur_ld_moved
= detach_tasks(&env
);
9573 * We've detached some tasks from busiest_rq. Every
9574 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9575 * unlock busiest->lock, and we are able to be sure
9576 * that nobody can manipulate the tasks in parallel.
9577 * See task_rq_lock() family for the details.
9580 rq_unlock(busiest
, &rf
);
9584 ld_moved
+= cur_ld_moved
;
9587 local_irq_restore(rf
.flags
);
9589 if (env
.flags
& LBF_NEED_BREAK
) {
9590 env
.flags
&= ~LBF_NEED_BREAK
;
9595 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9596 * us and move them to an alternate dst_cpu in our sched_group
9597 * where they can run. The upper limit on how many times we
9598 * iterate on same src_cpu is dependent on number of CPUs in our
9601 * This changes load balance semantics a bit on who can move
9602 * load to a given_cpu. In addition to the given_cpu itself
9603 * (or a ilb_cpu acting on its behalf where given_cpu is
9604 * nohz-idle), we now have balance_cpu in a position to move
9605 * load to given_cpu. In rare situations, this may cause
9606 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9607 * _independently_ and at _same_ time to move some load to
9608 * given_cpu) causing exceess load to be moved to given_cpu.
9609 * This however should not happen so much in practice and
9610 * moreover subsequent load balance cycles should correct the
9611 * excess load moved.
9613 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
9615 /* Prevent to re-select dst_cpu via env's CPUs */
9616 __cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
9618 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
9619 env
.dst_cpu
= env
.new_dst_cpu
;
9620 env
.flags
&= ~LBF_DST_PINNED
;
9622 env
.loop_break
= sched_nr_migrate_break
;
9625 * Go back to "more_balance" rather than "redo" since we
9626 * need to continue with same src_cpu.
9632 * We failed to reach balance because of affinity.
9635 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
9637 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
9638 *group_imbalance
= 1;
9641 /* All tasks on this runqueue were pinned by CPU affinity */
9642 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
9643 __cpumask_clear_cpu(cpu_of(busiest
), cpus
);
9645 * Attempting to continue load balancing at the current
9646 * sched_domain level only makes sense if there are
9647 * active CPUs remaining as possible busiest CPUs to
9648 * pull load from which are not contained within the
9649 * destination group that is receiving any migrated
9652 if (!cpumask_subset(cpus
, env
.dst_grpmask
)) {
9654 env
.loop_break
= sched_nr_migrate_break
;
9657 goto out_all_pinned
;
9662 schedstat_inc(sd
->lb_failed
[idle
]);
9664 * Increment the failure counter only on periodic balance.
9665 * We do not want newidle balance, which can be very
9666 * frequent, pollute the failure counter causing
9667 * excessive cache_hot migrations and active balances.
9669 if (idle
!= CPU_NEWLY_IDLE
)
9670 sd
->nr_balance_failed
++;
9672 if (need_active_balance(&env
)) {
9673 unsigned long flags
;
9675 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
9678 * Don't kick the active_load_balance_cpu_stop,
9679 * if the curr task on busiest CPU can't be
9680 * moved to this_cpu:
9682 if (!cpumask_test_cpu(this_cpu
, busiest
->curr
->cpus_ptr
)) {
9683 raw_spin_unlock_irqrestore(&busiest
->lock
,
9685 env
.flags
|= LBF_ALL_PINNED
;
9686 goto out_one_pinned
;
9690 * ->active_balance synchronizes accesses to
9691 * ->active_balance_work. Once set, it's cleared
9692 * only after active load balance is finished.
9694 if (!busiest
->active_balance
) {
9695 busiest
->active_balance
= 1;
9696 busiest
->push_cpu
= this_cpu
;
9699 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
9701 if (active_balance
) {
9702 stop_one_cpu_nowait(cpu_of(busiest
),
9703 active_load_balance_cpu_stop
, busiest
,
9704 &busiest
->active_balance_work
);
9707 /* We've kicked active balancing, force task migration. */
9708 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
9711 sd
->nr_balance_failed
= 0;
9713 if (likely(!active_balance
) || voluntary_active_balance(&env
)) {
9714 /* We were unbalanced, so reset the balancing interval */
9715 sd
->balance_interval
= sd
->min_interval
;
9718 * If we've begun active balancing, start to back off. This
9719 * case may not be covered by the all_pinned logic if there
9720 * is only 1 task on the busy runqueue (because we don't call
9723 if (sd
->balance_interval
< sd
->max_interval
)
9724 sd
->balance_interval
*= 2;
9731 * We reach balance although we may have faced some affinity
9732 * constraints. Clear the imbalance flag only if other tasks got
9733 * a chance to move and fix the imbalance.
9735 if (sd_parent
&& !(env
.flags
& LBF_ALL_PINNED
)) {
9736 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
9738 if (*group_imbalance
)
9739 *group_imbalance
= 0;
9744 * We reach balance because all tasks are pinned at this level so
9745 * we can't migrate them. Let the imbalance flag set so parent level
9746 * can try to migrate them.
9748 schedstat_inc(sd
->lb_balanced
[idle
]);
9750 sd
->nr_balance_failed
= 0;
9756 * newidle_balance() disregards balance intervals, so we could
9757 * repeatedly reach this code, which would lead to balance_interval
9758 * skyrocketting in a short amount of time. Skip the balance_interval
9759 * increase logic to avoid that.
9761 if (env
.idle
== CPU_NEWLY_IDLE
)
9764 /* tune up the balancing interval */
9765 if ((env
.flags
& LBF_ALL_PINNED
&&
9766 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
9767 sd
->balance_interval
< sd
->max_interval
)
9768 sd
->balance_interval
*= 2;
9773 static inline unsigned long
9774 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
9776 unsigned long interval
= sd
->balance_interval
;
9779 interval
*= sd
->busy_factor
;
9781 /* scale ms to jiffies */
9782 interval
= msecs_to_jiffies(interval
);
9783 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
9789 update_next_balance(struct sched_domain
*sd
, unsigned long *next_balance
)
9791 unsigned long interval
, next
;
9793 /* used by idle balance, so cpu_busy = 0 */
9794 interval
= get_sd_balance_interval(sd
, 0);
9795 next
= sd
->last_balance
+ interval
;
9797 if (time_after(*next_balance
, next
))
9798 *next_balance
= next
;
9802 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9803 * running tasks off the busiest CPU onto idle CPUs. It requires at
9804 * least 1 task to be running on each physical CPU where possible, and
9805 * avoids physical / logical imbalances.
9807 static int active_load_balance_cpu_stop(void *data
)
9809 struct rq
*busiest_rq
= data
;
9810 int busiest_cpu
= cpu_of(busiest_rq
);
9811 int target_cpu
= busiest_rq
->push_cpu
;
9812 struct rq
*target_rq
= cpu_rq(target_cpu
);
9813 struct sched_domain
*sd
;
9814 struct task_struct
*p
= NULL
;
9817 rq_lock_irq(busiest_rq
, &rf
);
9819 * Between queueing the stop-work and running it is a hole in which
9820 * CPUs can become inactive. We should not move tasks from or to
9823 if (!cpu_active(busiest_cpu
) || !cpu_active(target_cpu
))
9826 /* Make sure the requested CPU hasn't gone down in the meantime: */
9827 if (unlikely(busiest_cpu
!= smp_processor_id() ||
9828 !busiest_rq
->active_balance
))
9831 /* Is there any task to move? */
9832 if (busiest_rq
->nr_running
<= 1)
9836 * This condition is "impossible", if it occurs
9837 * we need to fix it. Originally reported by
9838 * Bjorn Helgaas on a 128-CPU setup.
9840 BUG_ON(busiest_rq
== target_rq
);
9842 /* Search for an sd spanning us and the target CPU. */
9844 for_each_domain(target_cpu
, sd
) {
9845 if (cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
9850 struct lb_env env
= {
9852 .dst_cpu
= target_cpu
,
9853 .dst_rq
= target_rq
,
9854 .src_cpu
= busiest_rq
->cpu
,
9855 .src_rq
= busiest_rq
,
9858 * can_migrate_task() doesn't need to compute new_dst_cpu
9859 * for active balancing. Since we have CPU_IDLE, but no
9860 * @dst_grpmask we need to make that test go away with lying
9863 .flags
= LBF_DST_PINNED
,
9866 schedstat_inc(sd
->alb_count
);
9867 update_rq_clock(busiest_rq
);
9869 p
= detach_one_task(&env
);
9871 schedstat_inc(sd
->alb_pushed
);
9872 /* Active balancing done, reset the failure counter. */
9873 sd
->nr_balance_failed
= 0;
9875 schedstat_inc(sd
->alb_failed
);
9880 busiest_rq
->active_balance
= 0;
9881 rq_unlock(busiest_rq
, &rf
);
9884 attach_one_task(target_rq
, p
);
9891 static DEFINE_SPINLOCK(balancing
);
9894 * Scale the max load_balance interval with the number of CPUs in the system.
9895 * This trades load-balance latency on larger machines for less cross talk.
9897 void update_max_interval(void)
9899 max_load_balance_interval
= HZ
*num_online_cpus()/10;
9903 * It checks each scheduling domain to see if it is due to be balanced,
9904 * and initiates a balancing operation if so.
9906 * Balancing parameters are set up in init_sched_domains.
9908 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
9910 int continue_balancing
= 1;
9912 int busy
= idle
!= CPU_IDLE
&& !sched_idle_cpu(cpu
);
9913 unsigned long interval
;
9914 struct sched_domain
*sd
;
9915 /* Earliest time when we have to do rebalance again */
9916 unsigned long next_balance
= jiffies
+ 60*HZ
;
9917 int update_next_balance
= 0;
9918 int need_serialize
, need_decay
= 0;
9922 for_each_domain(cpu
, sd
) {
9924 * Decay the newidle max times here because this is a regular
9925 * visit to all the domains. Decay ~1% per second.
9927 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
9928 sd
->max_newidle_lb_cost
=
9929 (sd
->max_newidle_lb_cost
* 253) / 256;
9930 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
9933 max_cost
+= sd
->max_newidle_lb_cost
;
9936 * Stop the load balance at this level. There is another
9937 * CPU in our sched group which is doing load balancing more
9940 if (!continue_balancing
) {
9946 interval
= get_sd_balance_interval(sd
, busy
);
9948 need_serialize
= sd
->flags
& SD_SERIALIZE
;
9949 if (need_serialize
) {
9950 if (!spin_trylock(&balancing
))
9954 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
9955 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
9957 * The LBF_DST_PINNED logic could have changed
9958 * env->dst_cpu, so we can't know our idle
9959 * state even if we migrated tasks. Update it.
9961 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
9962 busy
= idle
!= CPU_IDLE
&& !sched_idle_cpu(cpu
);
9964 sd
->last_balance
= jiffies
;
9965 interval
= get_sd_balance_interval(sd
, busy
);
9968 spin_unlock(&balancing
);
9970 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
9971 next_balance
= sd
->last_balance
+ interval
;
9972 update_next_balance
= 1;
9977 * Ensure the rq-wide value also decays but keep it at a
9978 * reasonable floor to avoid funnies with rq->avg_idle.
9980 rq
->max_idle_balance_cost
=
9981 max((u64
)sysctl_sched_migration_cost
, max_cost
);
9986 * next_balance will be updated only when there is a need.
9987 * When the cpu is attached to null domain for ex, it will not be
9990 if (likely(update_next_balance
)) {
9991 rq
->next_balance
= next_balance
;
9993 #ifdef CONFIG_NO_HZ_COMMON
9995 * If this CPU has been elected to perform the nohz idle
9996 * balance. Other idle CPUs have already rebalanced with
9997 * nohz_idle_balance() and nohz.next_balance has been
9998 * updated accordingly. This CPU is now running the idle load
9999 * balance for itself and we need to update the
10000 * nohz.next_balance accordingly.
10002 if ((idle
== CPU_IDLE
) && time_after(nohz
.next_balance
, rq
->next_balance
))
10003 nohz
.next_balance
= rq
->next_balance
;
10008 static inline int on_null_domain(struct rq
*rq
)
10010 return unlikely(!rcu_dereference_sched(rq
->sd
));
10013 #ifdef CONFIG_NO_HZ_COMMON
10015 * idle load balancing details
10016 * - When one of the busy CPUs notice that there may be an idle rebalancing
10017 * needed, they will kick the idle load balancer, which then does idle
10018 * load balancing for all the idle CPUs.
10019 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
10023 static inline int find_new_ilb(void)
10027 for_each_cpu_and(ilb
, nohz
.idle_cpus_mask
,
10028 housekeeping_cpumask(HK_FLAG_MISC
)) {
10037 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
10038 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
10040 static void kick_ilb(unsigned int flags
)
10045 * Increase nohz.next_balance only when if full ilb is triggered but
10046 * not if we only update stats.
10048 if (flags
& NOHZ_BALANCE_KICK
)
10049 nohz
.next_balance
= jiffies
+1;
10051 ilb_cpu
= find_new_ilb();
10053 if (ilb_cpu
>= nr_cpu_ids
)
10057 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
10058 * the first flag owns it; cleared by nohz_csd_func().
10060 flags
= atomic_fetch_or(flags
, nohz_flags(ilb_cpu
));
10061 if (flags
& NOHZ_KICK_MASK
)
10065 * This way we generate an IPI on the target CPU which
10066 * is idle. And the softirq performing nohz idle load balance
10067 * will be run before returning from the IPI.
10069 smp_call_function_single_async(ilb_cpu
, &cpu_rq(ilb_cpu
)->nohz_csd
);
10073 * Current decision point for kicking the idle load balancer in the presence
10074 * of idle CPUs in the system.
10076 static void nohz_balancer_kick(struct rq
*rq
)
10078 unsigned long now
= jiffies
;
10079 struct sched_domain_shared
*sds
;
10080 struct sched_domain
*sd
;
10081 int nr_busy
, i
, cpu
= rq
->cpu
;
10082 unsigned int flags
= 0;
10084 if (unlikely(rq
->idle_balance
))
10088 * We may be recently in ticked or tickless idle mode. At the first
10089 * busy tick after returning from idle, we will update the busy stats.
10091 nohz_balance_exit_idle(rq
);
10094 * None are in tickless mode and hence no need for NOHZ idle load
10097 if (likely(!atomic_read(&nohz
.nr_cpus
)))
10100 if (READ_ONCE(nohz
.has_blocked
) &&
10101 time_after(now
, READ_ONCE(nohz
.next_blocked
)))
10102 flags
= NOHZ_STATS_KICK
;
10104 if (time_before(now
, nohz
.next_balance
))
10107 if (rq
->nr_running
>= 2) {
10108 flags
= NOHZ_KICK_MASK
;
10114 sd
= rcu_dereference(rq
->sd
);
10117 * If there's a CFS task and the current CPU has reduced
10118 * capacity; kick the ILB to see if there's a better CPU to run
10121 if (rq
->cfs
.h_nr_running
>= 1 && check_cpu_capacity(rq
, sd
)) {
10122 flags
= NOHZ_KICK_MASK
;
10127 sd
= rcu_dereference(per_cpu(sd_asym_packing
, cpu
));
10130 * When ASYM_PACKING; see if there's a more preferred CPU
10131 * currently idle; in which case, kick the ILB to move tasks
10134 for_each_cpu_and(i
, sched_domain_span(sd
), nohz
.idle_cpus_mask
) {
10135 if (sched_asym_prefer(i
, cpu
)) {
10136 flags
= NOHZ_KICK_MASK
;
10142 sd
= rcu_dereference(per_cpu(sd_asym_cpucapacity
, cpu
));
10145 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10146 * to run the misfit task on.
10148 if (check_misfit_status(rq
, sd
)) {
10149 flags
= NOHZ_KICK_MASK
;
10154 * For asymmetric systems, we do not want to nicely balance
10155 * cache use, instead we want to embrace asymmetry and only
10156 * ensure tasks have enough CPU capacity.
10158 * Skip the LLC logic because it's not relevant in that case.
10163 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
10166 * If there is an imbalance between LLC domains (IOW we could
10167 * increase the overall cache use), we need some less-loaded LLC
10168 * domain to pull some load. Likewise, we may need to spread
10169 * load within the current LLC domain (e.g. packed SMT cores but
10170 * other CPUs are idle). We can't really know from here how busy
10171 * the others are - so just get a nohz balance going if it looks
10172 * like this LLC domain has tasks we could move.
10174 nr_busy
= atomic_read(&sds
->nr_busy_cpus
);
10176 flags
= NOHZ_KICK_MASK
;
10187 static void set_cpu_sd_state_busy(int cpu
)
10189 struct sched_domain
*sd
;
10192 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
10194 if (!sd
|| !sd
->nohz_idle
)
10198 atomic_inc(&sd
->shared
->nr_busy_cpus
);
10203 void nohz_balance_exit_idle(struct rq
*rq
)
10205 SCHED_WARN_ON(rq
!= this_rq());
10207 if (likely(!rq
->nohz_tick_stopped
))
10210 rq
->nohz_tick_stopped
= 0;
10211 cpumask_clear_cpu(rq
->cpu
, nohz
.idle_cpus_mask
);
10212 atomic_dec(&nohz
.nr_cpus
);
10214 set_cpu_sd_state_busy(rq
->cpu
);
10217 static void set_cpu_sd_state_idle(int cpu
)
10219 struct sched_domain
*sd
;
10222 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
10224 if (!sd
|| sd
->nohz_idle
)
10228 atomic_dec(&sd
->shared
->nr_busy_cpus
);
10234 * This routine will record that the CPU is going idle with tick stopped.
10235 * This info will be used in performing idle load balancing in the future.
10237 void nohz_balance_enter_idle(int cpu
)
10239 struct rq
*rq
= cpu_rq(cpu
);
10241 SCHED_WARN_ON(cpu
!= smp_processor_id());
10243 /* If this CPU is going down, then nothing needs to be done: */
10244 if (!cpu_active(cpu
))
10247 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10248 if (!housekeeping_cpu(cpu
, HK_FLAG_SCHED
))
10252 * Can be set safely without rq->lock held
10253 * If a clear happens, it will have evaluated last additions because
10254 * rq->lock is held during the check and the clear
10256 rq
->has_blocked_load
= 1;
10259 * The tick is still stopped but load could have been added in the
10260 * meantime. We set the nohz.has_blocked flag to trig a check of the
10261 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10262 * of nohz.has_blocked can only happen after checking the new load
10264 if (rq
->nohz_tick_stopped
)
10267 /* If we're a completely isolated CPU, we don't play: */
10268 if (on_null_domain(rq
))
10271 rq
->nohz_tick_stopped
= 1;
10273 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
10274 atomic_inc(&nohz
.nr_cpus
);
10277 * Ensures that if nohz_idle_balance() fails to observe our
10278 * @idle_cpus_mask store, it must observe the @has_blocked
10281 smp_mb__after_atomic();
10283 set_cpu_sd_state_idle(cpu
);
10287 * Each time a cpu enter idle, we assume that it has blocked load and
10288 * enable the periodic update of the load of idle cpus
10290 WRITE_ONCE(nohz
.has_blocked
, 1);
10294 * Internal function that runs load balance for all idle cpus. The load balance
10295 * can be a simple update of blocked load or a complete load balance with
10296 * tasks movement depending of flags.
10297 * The function returns false if the loop has stopped before running
10298 * through all idle CPUs.
10300 static bool _nohz_idle_balance(struct rq
*this_rq
, unsigned int flags
,
10301 enum cpu_idle_type idle
)
10303 /* Earliest time when we have to do rebalance again */
10304 unsigned long now
= jiffies
;
10305 unsigned long next_balance
= now
+ 60*HZ
;
10306 bool has_blocked_load
= false;
10307 int update_next_balance
= 0;
10308 int this_cpu
= this_rq
->cpu
;
10313 SCHED_WARN_ON((flags
& NOHZ_KICK_MASK
) == NOHZ_BALANCE_KICK
);
10316 * We assume there will be no idle load after this update and clear
10317 * the has_blocked flag. If a cpu enters idle in the mean time, it will
10318 * set the has_blocked flag and trig another update of idle load.
10319 * Because a cpu that becomes idle, is added to idle_cpus_mask before
10320 * setting the flag, we are sure to not clear the state and not
10321 * check the load of an idle cpu.
10323 WRITE_ONCE(nohz
.has_blocked
, 0);
10326 * Ensures that if we miss the CPU, we must see the has_blocked
10327 * store from nohz_balance_enter_idle().
10331 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
10332 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
10336 * If this CPU gets work to do, stop the load balancing
10337 * work being done for other CPUs. Next load
10338 * balancing owner will pick it up.
10340 if (need_resched()) {
10341 has_blocked_load
= true;
10345 rq
= cpu_rq(balance_cpu
);
10347 has_blocked_load
|= update_nohz_stats(rq
, true);
10350 * If time for next balance is due,
10353 if (time_after_eq(jiffies
, rq
->next_balance
)) {
10354 struct rq_flags rf
;
10356 rq_lock_irqsave(rq
, &rf
);
10357 update_rq_clock(rq
);
10358 rq_unlock_irqrestore(rq
, &rf
);
10360 if (flags
& NOHZ_BALANCE_KICK
)
10361 rebalance_domains(rq
, CPU_IDLE
);
10364 if (time_after(next_balance
, rq
->next_balance
)) {
10365 next_balance
= rq
->next_balance
;
10366 update_next_balance
= 1;
10371 * next_balance will be updated only when there is a need.
10372 * When the CPU is attached to null domain for ex, it will not be
10375 if (likely(update_next_balance
))
10376 nohz
.next_balance
= next_balance
;
10378 /* Newly idle CPU doesn't need an update */
10379 if (idle
!= CPU_NEWLY_IDLE
) {
10380 update_blocked_averages(this_cpu
);
10381 has_blocked_load
|= this_rq
->has_blocked_load
;
10384 if (flags
& NOHZ_BALANCE_KICK
)
10385 rebalance_domains(this_rq
, CPU_IDLE
);
10387 WRITE_ONCE(nohz
.next_blocked
,
10388 now
+ msecs_to_jiffies(LOAD_AVG_PERIOD
));
10390 /* The full idle balance loop has been done */
10394 /* There is still blocked load, enable periodic update */
10395 if (has_blocked_load
)
10396 WRITE_ONCE(nohz
.has_blocked
, 1);
10402 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
10403 * rebalancing for all the cpus for whom scheduler ticks are stopped.
10405 static bool nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
10407 unsigned int flags
= this_rq
->nohz_idle_balance
;
10412 this_rq
->nohz_idle_balance
= 0;
10414 if (idle
!= CPU_IDLE
)
10417 _nohz_idle_balance(this_rq
, flags
, idle
);
10422 static void nohz_newidle_balance(struct rq
*this_rq
)
10424 int this_cpu
= this_rq
->cpu
;
10427 * This CPU doesn't want to be disturbed by scheduler
10430 if (!housekeeping_cpu(this_cpu
, HK_FLAG_SCHED
))
10433 /* Will wake up very soon. No time for doing anything else*/
10434 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
10437 /* Don't need to update blocked load of idle CPUs*/
10438 if (!READ_ONCE(nohz
.has_blocked
) ||
10439 time_before(jiffies
, READ_ONCE(nohz
.next_blocked
)))
10442 raw_spin_unlock(&this_rq
->lock
);
10444 * This CPU is going to be idle and blocked load of idle CPUs
10445 * need to be updated. Run the ilb locally as it is a good
10446 * candidate for ilb instead of waking up another idle CPU.
10447 * Kick an normal ilb if we failed to do the update.
10449 if (!_nohz_idle_balance(this_rq
, NOHZ_STATS_KICK
, CPU_NEWLY_IDLE
))
10450 kick_ilb(NOHZ_STATS_KICK
);
10451 raw_spin_lock(&this_rq
->lock
);
10454 #else /* !CONFIG_NO_HZ_COMMON */
10455 static inline void nohz_balancer_kick(struct rq
*rq
) { }
10457 static inline bool nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
10462 static inline void nohz_newidle_balance(struct rq
*this_rq
) { }
10463 #endif /* CONFIG_NO_HZ_COMMON */
10466 * idle_balance is called by schedule() if this_cpu is about to become
10467 * idle. Attempts to pull tasks from other CPUs.
10470 * < 0 - we released the lock and there are !fair tasks present
10471 * 0 - failed, no new tasks
10472 * > 0 - success, new (fair) tasks present
10474 static int newidle_balance(struct rq
*this_rq
, struct rq_flags
*rf
)
10476 unsigned long next_balance
= jiffies
+ HZ
;
10477 int this_cpu
= this_rq
->cpu
;
10478 struct sched_domain
*sd
;
10479 int pulled_task
= 0;
10482 update_misfit_status(NULL
, this_rq
);
10484 * We must set idle_stamp _before_ calling idle_balance(), such that we
10485 * measure the duration of idle_balance() as idle time.
10487 this_rq
->idle_stamp
= rq_clock(this_rq
);
10490 * Do not pull tasks towards !active CPUs...
10492 if (!cpu_active(this_cpu
))
10496 * This is OK, because current is on_cpu, which avoids it being picked
10497 * for load-balance and preemption/IRQs are still disabled avoiding
10498 * further scheduler activity on it and we're being very careful to
10499 * re-start the picking loop.
10501 rq_unpin_lock(this_rq
, rf
);
10503 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
10504 !READ_ONCE(this_rq
->rd
->overload
)) {
10507 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
10509 update_next_balance(sd
, &next_balance
);
10512 nohz_newidle_balance(this_rq
);
10517 raw_spin_unlock(&this_rq
->lock
);
10519 update_blocked_averages(this_cpu
);
10521 for_each_domain(this_cpu
, sd
) {
10522 int continue_balancing
= 1;
10523 u64 t0
, domain_cost
;
10525 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
10526 update_next_balance(sd
, &next_balance
);
10530 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
10531 t0
= sched_clock_cpu(this_cpu
);
10533 pulled_task
= load_balance(this_cpu
, this_rq
,
10534 sd
, CPU_NEWLY_IDLE
,
10535 &continue_balancing
);
10537 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
10538 if (domain_cost
> sd
->max_newidle_lb_cost
)
10539 sd
->max_newidle_lb_cost
= domain_cost
;
10541 curr_cost
+= domain_cost
;
10544 update_next_balance(sd
, &next_balance
);
10547 * Stop searching for tasks to pull if there are
10548 * now runnable tasks on this rq.
10550 if (pulled_task
|| this_rq
->nr_running
> 0)
10555 raw_spin_lock(&this_rq
->lock
);
10557 if (curr_cost
> this_rq
->max_idle_balance_cost
)
10558 this_rq
->max_idle_balance_cost
= curr_cost
;
10562 * While browsing the domains, we released the rq lock, a task could
10563 * have been enqueued in the meantime. Since we're not going idle,
10564 * pretend we pulled a task.
10566 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
10569 /* Move the next balance forward */
10570 if (time_after(this_rq
->next_balance
, next_balance
))
10571 this_rq
->next_balance
= next_balance
;
10573 /* Is there a task of a high priority class? */
10574 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
10578 this_rq
->idle_stamp
= 0;
10580 rq_repin_lock(this_rq
, rf
);
10582 return pulled_task
;
10586 * run_rebalance_domains is triggered when needed from the scheduler tick.
10587 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10589 static __latent_entropy
void run_rebalance_domains(struct softirq_action
*h
)
10591 struct rq
*this_rq
= this_rq();
10592 enum cpu_idle_type idle
= this_rq
->idle_balance
?
10593 CPU_IDLE
: CPU_NOT_IDLE
;
10596 * If this CPU has a pending nohz_balance_kick, then do the
10597 * balancing on behalf of the other idle CPUs whose ticks are
10598 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10599 * give the idle CPUs a chance to load balance. Else we may
10600 * load balance only within the local sched_domain hierarchy
10601 * and abort nohz_idle_balance altogether if we pull some load.
10603 if (nohz_idle_balance(this_rq
, idle
))
10606 /* normal load balance */
10607 update_blocked_averages(this_rq
->cpu
);
10608 rebalance_domains(this_rq
, idle
);
10612 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10614 void trigger_load_balance(struct rq
*rq
)
10616 /* Don't need to rebalance while attached to NULL domain */
10617 if (unlikely(on_null_domain(rq
)))
10620 if (time_after_eq(jiffies
, rq
->next_balance
))
10621 raise_softirq(SCHED_SOFTIRQ
);
10623 nohz_balancer_kick(rq
);
10626 static void rq_online_fair(struct rq
*rq
)
10630 update_runtime_enabled(rq
);
10633 static void rq_offline_fair(struct rq
*rq
)
10637 /* Ensure any throttled groups are reachable by pick_next_task */
10638 unthrottle_offline_cfs_rqs(rq
);
10641 #endif /* CONFIG_SMP */
10644 * scheduler tick hitting a task of our scheduling class.
10646 * NOTE: This function can be called remotely by the tick offload that
10647 * goes along full dynticks. Therefore no local assumption can be made
10648 * and everything must be accessed through the @rq and @curr passed in
10651 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
10653 struct cfs_rq
*cfs_rq
;
10654 struct sched_entity
*se
= &curr
->se
;
10656 for_each_sched_entity(se
) {
10657 cfs_rq
= cfs_rq_of(se
);
10658 entity_tick(cfs_rq
, se
, queued
);
10661 if (static_branch_unlikely(&sched_numa_balancing
))
10662 task_tick_numa(rq
, curr
);
10664 update_misfit_status(curr
, rq
);
10665 update_overutilized_status(task_rq(curr
));
10669 * called on fork with the child task as argument from the parent's context
10670 * - child not yet on the tasklist
10671 * - preemption disabled
10673 static void task_fork_fair(struct task_struct
*p
)
10675 struct cfs_rq
*cfs_rq
;
10676 struct sched_entity
*se
= &p
->se
, *curr
;
10677 struct rq
*rq
= this_rq();
10678 struct rq_flags rf
;
10681 update_rq_clock(rq
);
10683 cfs_rq
= task_cfs_rq(current
);
10684 curr
= cfs_rq
->curr
;
10686 update_curr(cfs_rq
);
10687 se
->vruntime
= curr
->vruntime
;
10689 place_entity(cfs_rq
, se
, 1);
10691 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
10693 * Upon rescheduling, sched_class::put_prev_task() will place
10694 * 'current' within the tree based on its new key value.
10696 swap(curr
->vruntime
, se
->vruntime
);
10700 se
->vruntime
-= cfs_rq
->min_vruntime
;
10701 rq_unlock(rq
, &rf
);
10705 * Priority of the task has changed. Check to see if we preempt
10706 * the current task.
10709 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
10711 if (!task_on_rq_queued(p
))
10714 if (rq
->cfs
.nr_running
== 1)
10718 * Reschedule if we are currently running on this runqueue and
10719 * our priority decreased, or if we are not currently running on
10720 * this runqueue and our priority is higher than the current's
10722 if (rq
->curr
== p
) {
10723 if (p
->prio
> oldprio
)
10726 check_preempt_curr(rq
, p
, 0);
10729 static inline bool vruntime_normalized(struct task_struct
*p
)
10731 struct sched_entity
*se
= &p
->se
;
10734 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10735 * the dequeue_entity(.flags=0) will already have normalized the
10742 * When !on_rq, vruntime of the task has usually NOT been normalized.
10743 * But there are some cases where it has already been normalized:
10745 * - A forked child which is waiting for being woken up by
10746 * wake_up_new_task().
10747 * - A task which has been woken up by try_to_wake_up() and
10748 * waiting for actually being woken up by sched_ttwu_pending().
10750 if (!se
->sum_exec_runtime
||
10751 (p
->state
== TASK_WAKING
&& p
->sched_remote_wakeup
))
10757 #ifdef CONFIG_FAIR_GROUP_SCHED
10759 * Propagate the changes of the sched_entity across the tg tree to make it
10760 * visible to the root
10762 static void propagate_entity_cfs_rq(struct sched_entity
*se
)
10764 struct cfs_rq
*cfs_rq
;
10766 /* Start to propagate at parent */
10769 for_each_sched_entity(se
) {
10770 cfs_rq
= cfs_rq_of(se
);
10772 if (cfs_rq_throttled(cfs_rq
))
10775 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
10779 static void propagate_entity_cfs_rq(struct sched_entity
*se
) { }
10782 static void detach_entity_cfs_rq(struct sched_entity
*se
)
10784 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10786 /* Catch up with the cfs_rq and remove our load when we leave */
10787 update_load_avg(cfs_rq
, se
, 0);
10788 detach_entity_load_avg(cfs_rq
, se
);
10789 update_tg_load_avg(cfs_rq
, false);
10790 propagate_entity_cfs_rq(se
);
10793 static void attach_entity_cfs_rq(struct sched_entity
*se
)
10795 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10797 #ifdef CONFIG_FAIR_GROUP_SCHED
10799 * Since the real-depth could have been changed (only FAIR
10800 * class maintain depth value), reset depth properly.
10802 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
10805 /* Synchronize entity with its cfs_rq */
10806 update_load_avg(cfs_rq
, se
, sched_feat(ATTACH_AGE_LOAD
) ? 0 : SKIP_AGE_LOAD
);
10807 attach_entity_load_avg(cfs_rq
, se
);
10808 update_tg_load_avg(cfs_rq
, false);
10809 propagate_entity_cfs_rq(se
);
10812 static void detach_task_cfs_rq(struct task_struct
*p
)
10814 struct sched_entity
*se
= &p
->se
;
10815 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10817 if (!vruntime_normalized(p
)) {
10819 * Fix up our vruntime so that the current sleep doesn't
10820 * cause 'unlimited' sleep bonus.
10822 place_entity(cfs_rq
, se
, 0);
10823 se
->vruntime
-= cfs_rq
->min_vruntime
;
10826 detach_entity_cfs_rq(se
);
10829 static void attach_task_cfs_rq(struct task_struct
*p
)
10831 struct sched_entity
*se
= &p
->se
;
10832 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10834 attach_entity_cfs_rq(se
);
10836 if (!vruntime_normalized(p
))
10837 se
->vruntime
+= cfs_rq
->min_vruntime
;
10840 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
10842 detach_task_cfs_rq(p
);
10845 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
10847 attach_task_cfs_rq(p
);
10849 if (task_on_rq_queued(p
)) {
10851 * We were most likely switched from sched_rt, so
10852 * kick off the schedule if running, otherwise just see
10853 * if we can still preempt the current task.
10858 check_preempt_curr(rq
, p
, 0);
10862 /* Account for a task changing its policy or group.
10864 * This routine is mostly called to set cfs_rq->curr field when a task
10865 * migrates between groups/classes.
10867 static void set_next_task_fair(struct rq
*rq
, struct task_struct
*p
, bool first
)
10869 struct sched_entity
*se
= &p
->se
;
10872 if (task_on_rq_queued(p
)) {
10874 * Move the next running task to the front of the list, so our
10875 * cfs_tasks list becomes MRU one.
10877 list_move(&se
->group_node
, &rq
->cfs_tasks
);
10881 for_each_sched_entity(se
) {
10882 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10884 set_next_entity(cfs_rq
, se
);
10885 /* ensure bandwidth has been allocated on our new cfs_rq */
10886 account_cfs_rq_runtime(cfs_rq
, 0);
10890 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
10892 cfs_rq
->tasks_timeline
= RB_ROOT_CACHED
;
10893 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
10894 #ifndef CONFIG_64BIT
10895 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
10898 raw_spin_lock_init(&cfs_rq
->removed
.lock
);
10902 #ifdef CONFIG_FAIR_GROUP_SCHED
10903 static void task_set_group_fair(struct task_struct
*p
)
10905 struct sched_entity
*se
= &p
->se
;
10907 set_task_rq(p
, task_cpu(p
));
10908 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
10911 static void task_move_group_fair(struct task_struct
*p
)
10913 detach_task_cfs_rq(p
);
10914 set_task_rq(p
, task_cpu(p
));
10917 /* Tell se's cfs_rq has been changed -- migrated */
10918 p
->se
.avg
.last_update_time
= 0;
10920 attach_task_cfs_rq(p
);
10923 static void task_change_group_fair(struct task_struct
*p
, int type
)
10926 case TASK_SET_GROUP
:
10927 task_set_group_fair(p
);
10930 case TASK_MOVE_GROUP
:
10931 task_move_group_fair(p
);
10936 void free_fair_sched_group(struct task_group
*tg
)
10940 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
10942 for_each_possible_cpu(i
) {
10944 kfree(tg
->cfs_rq
[i
]);
10953 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
10955 struct sched_entity
*se
;
10956 struct cfs_rq
*cfs_rq
;
10959 tg
->cfs_rq
= kcalloc(nr_cpu_ids
, sizeof(cfs_rq
), GFP_KERNEL
);
10962 tg
->se
= kcalloc(nr_cpu_ids
, sizeof(se
), GFP_KERNEL
);
10966 tg
->shares
= NICE_0_LOAD
;
10968 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
10970 for_each_possible_cpu(i
) {
10971 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
10972 GFP_KERNEL
, cpu_to_node(i
));
10976 se
= kzalloc_node(sizeof(struct sched_entity
),
10977 GFP_KERNEL
, cpu_to_node(i
));
10981 init_cfs_rq(cfs_rq
);
10982 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
10983 init_entity_runnable_average(se
);
10994 void online_fair_sched_group(struct task_group
*tg
)
10996 struct sched_entity
*se
;
10997 struct rq_flags rf
;
11001 for_each_possible_cpu(i
) {
11004 rq_lock_irq(rq
, &rf
);
11005 update_rq_clock(rq
);
11006 attach_entity_cfs_rq(se
);
11007 sync_throttle(tg
, i
);
11008 rq_unlock_irq(rq
, &rf
);
11012 void unregister_fair_sched_group(struct task_group
*tg
)
11014 unsigned long flags
;
11018 for_each_possible_cpu(cpu
) {
11020 remove_entity_load_avg(tg
->se
[cpu
]);
11023 * Only empty task groups can be destroyed; so we can speculatively
11024 * check on_list without danger of it being re-added.
11026 if (!tg
->cfs_rq
[cpu
]->on_list
)
11031 raw_spin_lock_irqsave(&rq
->lock
, flags
);
11032 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
11033 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
11037 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
11038 struct sched_entity
*se
, int cpu
,
11039 struct sched_entity
*parent
)
11041 struct rq
*rq
= cpu_rq(cpu
);
11045 init_cfs_rq_runtime(cfs_rq
);
11047 tg
->cfs_rq
[cpu
] = cfs_rq
;
11050 /* se could be NULL for root_task_group */
11055 se
->cfs_rq
= &rq
->cfs
;
11058 se
->cfs_rq
= parent
->my_q
;
11059 se
->depth
= parent
->depth
+ 1;
11063 /* guarantee group entities always have weight */
11064 update_load_set(&se
->load
, NICE_0_LOAD
);
11065 se
->parent
= parent
;
11068 static DEFINE_MUTEX(shares_mutex
);
11070 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
11075 * We can't change the weight of the root cgroup.
11080 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
11082 mutex_lock(&shares_mutex
);
11083 if (tg
->shares
== shares
)
11086 tg
->shares
= shares
;
11087 for_each_possible_cpu(i
) {
11088 struct rq
*rq
= cpu_rq(i
);
11089 struct sched_entity
*se
= tg
->se
[i
];
11090 struct rq_flags rf
;
11092 /* Propagate contribution to hierarchy */
11093 rq_lock_irqsave(rq
, &rf
);
11094 update_rq_clock(rq
);
11095 for_each_sched_entity(se
) {
11096 update_load_avg(cfs_rq_of(se
), se
, UPDATE_TG
);
11097 update_cfs_group(se
);
11099 rq_unlock_irqrestore(rq
, &rf
);
11103 mutex_unlock(&shares_mutex
);
11106 #else /* CONFIG_FAIR_GROUP_SCHED */
11108 void free_fair_sched_group(struct task_group
*tg
) { }
11110 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
11115 void online_fair_sched_group(struct task_group
*tg
) { }
11117 void unregister_fair_sched_group(struct task_group
*tg
) { }
11119 #endif /* CONFIG_FAIR_GROUP_SCHED */
11122 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
11124 struct sched_entity
*se
= &task
->se
;
11125 unsigned int rr_interval
= 0;
11128 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11131 if (rq
->cfs
.load
.weight
)
11132 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
11134 return rr_interval
;
11138 * All the scheduling class methods:
11140 const struct sched_class fair_sched_class
11141 __attribute__((section("__fair_sched_class"))) = {
11142 .enqueue_task
= enqueue_task_fair
,
11143 .dequeue_task
= dequeue_task_fair
,
11144 .yield_task
= yield_task_fair
,
11145 .yield_to_task
= yield_to_task_fair
,
11147 .check_preempt_curr
= check_preempt_wakeup
,
11149 .pick_next_task
= __pick_next_task_fair
,
11150 .put_prev_task
= put_prev_task_fair
,
11151 .set_next_task
= set_next_task_fair
,
11154 .balance
= balance_fair
,
11155 .select_task_rq
= select_task_rq_fair
,
11156 .migrate_task_rq
= migrate_task_rq_fair
,
11158 .rq_online
= rq_online_fair
,
11159 .rq_offline
= rq_offline_fair
,
11161 .task_dead
= task_dead_fair
,
11162 .set_cpus_allowed
= set_cpus_allowed_common
,
11165 .task_tick
= task_tick_fair
,
11166 .task_fork
= task_fork_fair
,
11168 .prio_changed
= prio_changed_fair
,
11169 .switched_from
= switched_from_fair
,
11170 .switched_to
= switched_to_fair
,
11172 .get_rr_interval
= get_rr_interval_fair
,
11174 .update_curr
= update_curr_fair
,
11176 #ifdef CONFIG_FAIR_GROUP_SCHED
11177 .task_change_group
= task_change_group_fair
,
11180 #ifdef CONFIG_UCLAMP_TASK
11181 .uclamp_enabled
= 1,
11185 #ifdef CONFIG_SCHED_DEBUG
11186 void print_cfs_stats(struct seq_file
*m
, int cpu
)
11188 struct cfs_rq
*cfs_rq
, *pos
;
11191 for_each_leaf_cfs_rq_safe(cpu_rq(cpu
), cfs_rq
, pos
)
11192 print_cfs_rq(m
, cpu
, cfs_rq
);
11196 #ifdef CONFIG_NUMA_BALANCING
11197 void show_numa_stats(struct task_struct
*p
, struct seq_file
*m
)
11200 unsigned long tsf
= 0, tpf
= 0, gsf
= 0, gpf
= 0;
11201 struct numa_group
*ng
;
11204 ng
= rcu_dereference(p
->numa_group
);
11205 for_each_online_node(node
) {
11206 if (p
->numa_faults
) {
11207 tsf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
11208 tpf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
11211 gsf
= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)],
11212 gpf
= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
11214 print_numa_stats(m
, node
, tsf
, tpf
, gsf
, gpf
);
11218 #endif /* CONFIG_NUMA_BALANCING */
11219 #endif /* CONFIG_SCHED_DEBUG */
11221 __init
void init_sched_fair_class(void)
11224 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
11226 #ifdef CONFIG_NO_HZ_COMMON
11227 nohz
.next_balance
= jiffies
;
11228 nohz
.next_blocked
= jiffies
;
11229 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
11236 * Helper functions to facilitate extracting info from tracepoints.
11239 const struct sched_avg
*sched_trace_cfs_rq_avg(struct cfs_rq
*cfs_rq
)
11242 return cfs_rq
? &cfs_rq
->avg
: NULL
;
11247 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg
);
11249 char *sched_trace_cfs_rq_path(struct cfs_rq
*cfs_rq
, char *str
, int len
)
11253 strlcpy(str
, "(null)", len
);
11258 cfs_rq_tg_path(cfs_rq
, str
, len
);
11261 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path
);
11263 int sched_trace_cfs_rq_cpu(struct cfs_rq
*cfs_rq
)
11265 return cfs_rq
? cpu_of(rq_of(cfs_rq
)) : -1;
11267 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu
);
11269 const struct sched_avg
*sched_trace_rq_avg_rt(struct rq
*rq
)
11272 return rq
? &rq
->avg_rt
: NULL
;
11277 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt
);
11279 const struct sched_avg
*sched_trace_rq_avg_dl(struct rq
*rq
)
11282 return rq
? &rq
->avg_dl
: NULL
;
11287 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl
);
11289 const struct sched_avg
*sched_trace_rq_avg_irq(struct rq
*rq
)
11291 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
11292 return rq
? &rq
->avg_irq
: NULL
;
11297 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq
);
11299 int sched_trace_rq_cpu(struct rq
*rq
)
11301 return rq
? cpu_of(rq
) : -1;
11303 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu
);
11305 const struct cpumask
*sched_trace_rd_span(struct root_domain
*rd
)
11308 return rd
? rd
->span
: NULL
;
11313 EXPORT_SYMBOL_GPL(sched_trace_rd_span
);
11315 int sched_trace_rq_nr_running(struct rq
*rq
)
11317 return rq
? rq
->nr_running
: -1;
11319 EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running
);