1 // SPDX-License-Identifier: GPL-2.0
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
26 * Targeted preemption latency for CPU-bound tasks:
28 * NOTE: this latency value is not the same as the concept of
29 * 'timeslice length' - timeslices in CFS are of variable length
30 * and have no persistent notion like in traditional, time-slice
31 * based scheduling concepts.
33 * (to see the precise effective timeslice length of your workload,
34 * run vmstat and monitor the context-switches (cs) field)
36 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
38 unsigned int sysctl_sched_latency
= 6000000ULL;
39 static unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
42 * The initial- and re-scaling of tunables is configurable
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
52 enum sched_tunable_scaling sysctl_sched_tunable_scaling
= SCHED_TUNABLESCALING_LOG
;
55 * Minimal preemption granularity for CPU-bound tasks:
57 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
59 unsigned int sysctl_sched_min_granularity
= 750000ULL;
60 static unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
63 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
65 static unsigned int sched_nr_latency
= 8;
68 * After fork, child runs first. If set to 0 (default) then
69 * parent will (try to) run first.
71 unsigned int sysctl_sched_child_runs_first __read_mostly
;
74 * SCHED_OTHER wake-up granularity.
76 * This option delays the preemption effects of decoupled workloads
77 * and reduces their over-scheduling. Synchronous workloads will still
78 * have immediate wakeup/sleep latencies.
80 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
82 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
83 static unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
85 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
87 int sched_thermal_decay_shift
;
88 static int __init
setup_sched_thermal_decay_shift(char *str
)
92 if (kstrtoint(str
, 0, &_shift
))
93 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
95 sched_thermal_decay_shift
= clamp(_shift
, 0, 10);
98 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift
);
102 * For asym packing, by default the lower numbered CPU has higher priority.
104 int __weak
arch_asym_cpu_priority(int cpu
)
110 * The margin used when comparing utilization with CPU capacity.
114 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
118 #ifdef CONFIG_CFS_BANDWIDTH
120 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
121 * each time a cfs_rq requests quota.
123 * Note: in the case that the slice exceeds the runtime remaining (either due
124 * to consumption or the quota being specified to be smaller than the slice)
125 * we will always only issue the remaining available time.
127 * (default: 5 msec, units: microseconds)
129 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
132 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
138 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
144 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
151 * Increase the granularity value when there are more CPUs,
152 * because with more CPUs the 'effective latency' as visible
153 * to users decreases. But the relationship is not linear,
154 * so pick a second-best guess by going with the log2 of the
157 * This idea comes from the SD scheduler of Con Kolivas:
159 static unsigned int get_update_sysctl_factor(void)
161 unsigned int cpus
= min_t(unsigned int, num_online_cpus(), 8);
164 switch (sysctl_sched_tunable_scaling
) {
165 case SCHED_TUNABLESCALING_NONE
:
168 case SCHED_TUNABLESCALING_LINEAR
:
171 case SCHED_TUNABLESCALING_LOG
:
173 factor
= 1 + ilog2(cpus
);
180 static void update_sysctl(void)
182 unsigned int factor
= get_update_sysctl_factor();
184 #define SET_SYSCTL(name) \
185 (sysctl_##name = (factor) * normalized_sysctl_##name)
186 SET_SYSCTL(sched_min_granularity
);
187 SET_SYSCTL(sched_latency
);
188 SET_SYSCTL(sched_wakeup_granularity
);
192 void __init
sched_init_granularity(void)
197 #define WMULT_CONST (~0U)
198 #define WMULT_SHIFT 32
200 static void __update_inv_weight(struct load_weight
*lw
)
204 if (likely(lw
->inv_weight
))
207 w
= scale_load_down(lw
->weight
);
209 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
211 else if (unlikely(!w
))
212 lw
->inv_weight
= WMULT_CONST
;
214 lw
->inv_weight
= WMULT_CONST
/ w
;
218 * delta_exec * weight / lw.weight
220 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
222 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
223 * we're guaranteed shift stays positive because inv_weight is guaranteed to
224 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
226 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
227 * weight/lw.weight <= 1, and therefore our shift will also be positive.
229 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
231 u64 fact
= scale_load_down(weight
);
232 int shift
= WMULT_SHIFT
;
234 __update_inv_weight(lw
);
236 if (unlikely(fact
>> 32)) {
243 fact
= mul_u32_u32(fact
, lw
->inv_weight
);
250 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
254 const struct sched_class fair_sched_class
;
256 /**************************************************************
257 * CFS operations on generic schedulable entities:
260 #ifdef CONFIG_FAIR_GROUP_SCHED
261 static inline struct task_struct
*task_of(struct sched_entity
*se
)
263 SCHED_WARN_ON(!entity_is_task(se
));
264 return container_of(se
, struct task_struct
, se
);
267 /* Walk up scheduling entities hierarchy */
268 #define for_each_sched_entity(se) \
269 for (; se; se = se->parent)
271 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
276 /* runqueue on which this entity is (to be) queued */
277 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
282 /* runqueue "owned" by this group */
283 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
288 static inline void cfs_rq_tg_path(struct cfs_rq
*cfs_rq
, char *path
, int len
)
293 if (cfs_rq
&& task_group_is_autogroup(cfs_rq
->tg
))
294 autogroup_path(cfs_rq
->tg
, path
, len
);
295 else if (cfs_rq
&& cfs_rq
->tg
->css
.cgroup
)
296 cgroup_path(cfs_rq
->tg
->css
.cgroup
, path
, len
);
298 strlcpy(path
, "(null)", len
);
301 static inline bool list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
303 struct rq
*rq
= rq_of(cfs_rq
);
304 int cpu
= cpu_of(rq
);
307 return rq
->tmp_alone_branch
== &rq
->leaf_cfs_rq_list
;
312 * Ensure we either appear before our parent (if already
313 * enqueued) or force our parent to appear after us when it is
314 * enqueued. The fact that we always enqueue bottom-up
315 * reduces this to two cases and a special case for the root
316 * cfs_rq. Furthermore, it also means that we will always reset
317 * tmp_alone_branch either when the branch is connected
318 * to a tree or when we reach the top of the tree
320 if (cfs_rq
->tg
->parent
&&
321 cfs_rq
->tg
->parent
->cfs_rq
[cpu
]->on_list
) {
323 * If parent is already on the list, we add the child
324 * just before. Thanks to circular linked property of
325 * the list, this means to put the child at the tail
326 * of the list that starts by parent.
328 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
329 &(cfs_rq
->tg
->parent
->cfs_rq
[cpu
]->leaf_cfs_rq_list
));
331 * The branch is now connected to its tree so we can
332 * reset tmp_alone_branch to the beginning of the
335 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
339 if (!cfs_rq
->tg
->parent
) {
341 * cfs rq without parent should be put
342 * at the tail of the list.
344 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
345 &rq
->leaf_cfs_rq_list
);
347 * We have reach the top of a tree so we can reset
348 * tmp_alone_branch to the beginning of the list.
350 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
355 * The parent has not already been added so we want to
356 * make sure that it will be put after us.
357 * tmp_alone_branch points to the begin of the branch
358 * where we will add parent.
360 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, rq
->tmp_alone_branch
);
362 * update tmp_alone_branch to points to the new begin
365 rq
->tmp_alone_branch
= &cfs_rq
->leaf_cfs_rq_list
;
369 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
371 if (cfs_rq
->on_list
) {
372 struct rq
*rq
= rq_of(cfs_rq
);
375 * With cfs_rq being unthrottled/throttled during an enqueue,
376 * it can happen the tmp_alone_branch points the a leaf that
377 * we finally want to del. In this case, tmp_alone_branch moves
378 * to the prev element but it will point to rq->leaf_cfs_rq_list
379 * at the end of the enqueue.
381 if (rq
->tmp_alone_branch
== &cfs_rq
->leaf_cfs_rq_list
)
382 rq
->tmp_alone_branch
= cfs_rq
->leaf_cfs_rq_list
.prev
;
384 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
389 static inline void assert_list_leaf_cfs_rq(struct rq
*rq
)
391 SCHED_WARN_ON(rq
->tmp_alone_branch
!= &rq
->leaf_cfs_rq_list
);
394 /* Iterate thr' all leaf cfs_rq's on a runqueue */
395 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
396 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
399 /* Do the two (enqueued) entities belong to the same group ? */
400 static inline struct cfs_rq
*
401 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
403 if (se
->cfs_rq
== pse
->cfs_rq
)
409 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
415 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
417 int se_depth
, pse_depth
;
420 * preemption test can be made between sibling entities who are in the
421 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
422 * both tasks until we find their ancestors who are siblings of common
426 /* First walk up until both entities are at same depth */
427 se_depth
= (*se
)->depth
;
428 pse_depth
= (*pse
)->depth
;
430 while (se_depth
> pse_depth
) {
432 *se
= parent_entity(*se
);
435 while (pse_depth
> se_depth
) {
437 *pse
= parent_entity(*pse
);
440 while (!is_same_group(*se
, *pse
)) {
441 *se
= parent_entity(*se
);
442 *pse
= parent_entity(*pse
);
446 #else /* !CONFIG_FAIR_GROUP_SCHED */
448 static inline struct task_struct
*task_of(struct sched_entity
*se
)
450 return container_of(se
, struct task_struct
, se
);
453 #define for_each_sched_entity(se) \
454 for (; se; se = NULL)
456 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
458 return &task_rq(p
)->cfs
;
461 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
463 struct task_struct
*p
= task_of(se
);
464 struct rq
*rq
= task_rq(p
);
469 /* runqueue "owned" by this group */
470 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
475 static inline void cfs_rq_tg_path(struct cfs_rq
*cfs_rq
, char *path
, int len
)
478 strlcpy(path
, "(null)", len
);
481 static inline bool list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
486 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
490 static inline void assert_list_leaf_cfs_rq(struct rq
*rq
)
494 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
495 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
497 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
503 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
507 #endif /* CONFIG_FAIR_GROUP_SCHED */
509 static __always_inline
510 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
512 /**************************************************************
513 * Scheduling class tree data structure manipulation methods:
516 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
518 s64 delta
= (s64
)(vruntime
- max_vruntime
);
520 max_vruntime
= vruntime
;
525 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
527 s64 delta
= (s64
)(vruntime
- min_vruntime
);
529 min_vruntime
= vruntime
;
534 static inline int entity_before(struct sched_entity
*a
,
535 struct sched_entity
*b
)
537 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
540 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
542 struct sched_entity
*curr
= cfs_rq
->curr
;
543 struct rb_node
*leftmost
= rb_first_cached(&cfs_rq
->tasks_timeline
);
545 u64 vruntime
= cfs_rq
->min_vruntime
;
549 vruntime
= curr
->vruntime
;
554 if (leftmost
) { /* non-empty tree */
555 struct sched_entity
*se
;
556 se
= rb_entry(leftmost
, struct sched_entity
, run_node
);
559 vruntime
= se
->vruntime
;
561 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
564 /* ensure we never gain time by being placed backwards. */
565 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
568 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
573 * Enqueue an entity into the rb-tree:
575 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
577 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_root
.rb_node
;
578 struct rb_node
*parent
= NULL
;
579 struct sched_entity
*entry
;
580 bool leftmost
= true;
583 * Find the right place in the rbtree:
587 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
589 * We dont care about collisions. Nodes with
590 * the same key stay together.
592 if (entity_before(se
, entry
)) {
593 link
= &parent
->rb_left
;
595 link
= &parent
->rb_right
;
600 rb_link_node(&se
->run_node
, parent
, link
);
601 rb_insert_color_cached(&se
->run_node
,
602 &cfs_rq
->tasks_timeline
, leftmost
);
605 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
607 rb_erase_cached(&se
->run_node
, &cfs_rq
->tasks_timeline
);
610 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
612 struct rb_node
*left
= rb_first_cached(&cfs_rq
->tasks_timeline
);
617 return rb_entry(left
, struct sched_entity
, run_node
);
620 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
622 struct rb_node
*next
= rb_next(&se
->run_node
);
627 return rb_entry(next
, struct sched_entity
, run_node
);
630 #ifdef CONFIG_SCHED_DEBUG
631 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
633 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
.rb_root
);
638 return rb_entry(last
, struct sched_entity
, run_node
);
641 /**************************************************************
642 * Scheduling class statistics methods:
645 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
646 void *buffer
, size_t *lenp
, loff_t
*ppos
)
648 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
649 unsigned int factor
= get_update_sysctl_factor();
654 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
655 sysctl_sched_min_granularity
);
657 #define WRT_SYSCTL(name) \
658 (normalized_sysctl_##name = sysctl_##name / (factor))
659 WRT_SYSCTL(sched_min_granularity
);
660 WRT_SYSCTL(sched_latency
);
661 WRT_SYSCTL(sched_wakeup_granularity
);
671 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
673 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
674 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
680 * The idea is to set a period in which each task runs once.
682 * When there are too many tasks (sched_nr_latency) we have to stretch
683 * this period because otherwise the slices get too small.
685 * p = (nr <= nl) ? l : l*nr/nl
687 static u64
__sched_period(unsigned long nr_running
)
689 if (unlikely(nr_running
> sched_nr_latency
))
690 return nr_running
* sysctl_sched_min_granularity
;
692 return sysctl_sched_latency
;
696 * We calculate the wall-time slice from the period by taking a part
697 * proportional to the weight.
701 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
703 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
705 for_each_sched_entity(se
) {
706 struct load_weight
*load
;
707 struct load_weight lw
;
709 cfs_rq
= cfs_rq_of(se
);
710 load
= &cfs_rq
->load
;
712 if (unlikely(!se
->on_rq
)) {
715 update_load_add(&lw
, se
->load
.weight
);
718 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
724 * We calculate the vruntime slice of a to-be-inserted task.
728 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
730 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
736 static int select_idle_sibling(struct task_struct
*p
, int prev_cpu
, int cpu
);
737 static unsigned long task_h_load(struct task_struct
*p
);
738 static unsigned long capacity_of(int cpu
);
740 /* Give new sched_entity start runnable values to heavy its load in infant time */
741 void init_entity_runnable_average(struct sched_entity
*se
)
743 struct sched_avg
*sa
= &se
->avg
;
745 memset(sa
, 0, sizeof(*sa
));
748 * Tasks are initialized with full load to be seen as heavy tasks until
749 * they get a chance to stabilize to their real load level.
750 * Group entities are initialized with zero load to reflect the fact that
751 * nothing has been attached to the task group yet.
753 if (entity_is_task(se
))
754 sa
->load_avg
= scale_load_down(se
->load
.weight
);
756 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
759 static void attach_entity_cfs_rq(struct sched_entity
*se
);
762 * With new tasks being created, their initial util_avgs are extrapolated
763 * based on the cfs_rq's current util_avg:
765 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
767 * However, in many cases, the above util_avg does not give a desired
768 * value. Moreover, the sum of the util_avgs may be divergent, such
769 * as when the series is a harmonic series.
771 * To solve this problem, we also cap the util_avg of successive tasks to
772 * only 1/2 of the left utilization budget:
774 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
776 * where n denotes the nth task and cpu_scale the CPU capacity.
778 * For example, for a CPU with 1024 of capacity, a simplest series from
779 * the beginning would be like:
781 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
782 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
784 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
785 * if util_avg > util_avg_cap.
787 void post_init_entity_util_avg(struct task_struct
*p
)
789 struct sched_entity
*se
= &p
->se
;
790 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
791 struct sched_avg
*sa
= &se
->avg
;
792 long cpu_scale
= arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq
)));
793 long cap
= (long)(cpu_scale
- cfs_rq
->avg
.util_avg
) / 2;
796 if (cfs_rq
->avg
.util_avg
!= 0) {
797 sa
->util_avg
= cfs_rq
->avg
.util_avg
* se
->load
.weight
;
798 sa
->util_avg
/= (cfs_rq
->avg
.load_avg
+ 1);
800 if (sa
->util_avg
> cap
)
807 sa
->runnable_avg
= sa
->util_avg
;
809 if (p
->sched_class
!= &fair_sched_class
) {
811 * For !fair tasks do:
813 update_cfs_rq_load_avg(now, cfs_rq);
814 attach_entity_load_avg(cfs_rq, se);
815 switched_from_fair(rq, p);
817 * such that the next switched_to_fair() has the
820 se
->avg
.last_update_time
= cfs_rq_clock_pelt(cfs_rq
);
824 attach_entity_cfs_rq(se
);
827 #else /* !CONFIG_SMP */
828 void init_entity_runnable_average(struct sched_entity
*se
)
831 void post_init_entity_util_avg(struct task_struct
*p
)
834 static void update_tg_load_avg(struct cfs_rq
*cfs_rq
)
837 #endif /* CONFIG_SMP */
840 * Update the current task's runtime statistics.
842 static void update_curr(struct cfs_rq
*cfs_rq
)
844 struct sched_entity
*curr
= cfs_rq
->curr
;
845 u64 now
= rq_clock_task(rq_of(cfs_rq
));
851 delta_exec
= now
- curr
->exec_start
;
852 if (unlikely((s64
)delta_exec
<= 0))
855 curr
->exec_start
= now
;
857 schedstat_set(curr
->statistics
.exec_max
,
858 max(delta_exec
, curr
->statistics
.exec_max
));
860 curr
->sum_exec_runtime
+= delta_exec
;
861 schedstat_add(cfs_rq
->exec_clock
, delta_exec
);
863 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
864 update_min_vruntime(cfs_rq
);
866 if (entity_is_task(curr
)) {
867 struct task_struct
*curtask
= task_of(curr
);
869 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
870 cgroup_account_cputime(curtask
, delta_exec
);
871 account_group_exec_runtime(curtask
, delta_exec
);
874 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
877 static void update_curr_fair(struct rq
*rq
)
879 update_curr(cfs_rq_of(&rq
->curr
->se
));
883 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
885 u64 wait_start
, prev_wait_start
;
887 if (!schedstat_enabled())
890 wait_start
= rq_clock(rq_of(cfs_rq
));
891 prev_wait_start
= schedstat_val(se
->statistics
.wait_start
);
893 if (entity_is_task(se
) && task_on_rq_migrating(task_of(se
)) &&
894 likely(wait_start
> prev_wait_start
))
895 wait_start
-= prev_wait_start
;
897 __schedstat_set(se
->statistics
.wait_start
, wait_start
);
901 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
903 struct task_struct
*p
;
906 if (!schedstat_enabled())
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 */
1507 unsigned long runnable
;
1509 /* Total compute capacity of CPUs on a node */
1510 unsigned long compute_capacity
;
1511 unsigned int nr_running
;
1512 unsigned int weight
;
1513 enum numa_type node_type
;
1517 static inline bool is_core_idle(int cpu
)
1519 #ifdef CONFIG_SCHED_SMT
1522 for_each_cpu(sibling
, cpu_smt_mask(cpu
)) {
1534 struct task_numa_env
{
1535 struct task_struct
*p
;
1537 int src_cpu
, src_nid
;
1538 int dst_cpu
, dst_nid
;
1540 struct numa_stats src_stats
, dst_stats
;
1545 struct task_struct
*best_task
;
1550 static unsigned long cpu_load(struct rq
*rq
);
1551 static unsigned long cpu_runnable(struct rq
*rq
);
1552 static unsigned long cpu_util(int cpu
);
1553 static inline long adjust_numa_imbalance(int imbalance
, int nr_running
);
1556 numa_type
numa_classify(unsigned int imbalance_pct
,
1557 struct numa_stats
*ns
)
1559 if ((ns
->nr_running
> ns
->weight
) &&
1560 (((ns
->compute_capacity
* 100) < (ns
->util
* imbalance_pct
)) ||
1561 ((ns
->compute_capacity
* imbalance_pct
) < (ns
->runnable
* 100))))
1562 return node_overloaded
;
1564 if ((ns
->nr_running
< ns
->weight
) ||
1565 (((ns
->compute_capacity
* 100) > (ns
->util
* imbalance_pct
)) &&
1566 ((ns
->compute_capacity
* imbalance_pct
) > (ns
->runnable
* 100))))
1567 return node_has_spare
;
1569 return node_fully_busy
;
1572 #ifdef CONFIG_SCHED_SMT
1573 /* Forward declarations of select_idle_sibling helpers */
1574 static inline bool test_idle_cores(int cpu
, bool def
);
1575 static inline int numa_idle_core(int idle_core
, int cpu
)
1577 if (!static_branch_likely(&sched_smt_present
) ||
1578 idle_core
>= 0 || !test_idle_cores(cpu
, false))
1582 * Prefer cores instead of packing HT siblings
1583 * and triggering future load balancing.
1585 if (is_core_idle(cpu
))
1591 static inline int numa_idle_core(int idle_core
, int cpu
)
1598 * Gather all necessary information to make NUMA balancing placement
1599 * decisions that are compatible with standard load balancer. This
1600 * borrows code and logic from update_sg_lb_stats but sharing a
1601 * common implementation is impractical.
1603 static void update_numa_stats(struct task_numa_env
*env
,
1604 struct numa_stats
*ns
, int nid
,
1607 int cpu
, idle_core
= -1;
1609 memset(ns
, 0, sizeof(*ns
));
1613 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1614 struct rq
*rq
= cpu_rq(cpu
);
1616 ns
->load
+= cpu_load(rq
);
1617 ns
->runnable
+= cpu_runnable(rq
);
1618 ns
->util
+= cpu_util(cpu
);
1619 ns
->nr_running
+= rq
->cfs
.h_nr_running
;
1620 ns
->compute_capacity
+= capacity_of(cpu
);
1622 if (find_idle
&& !rq
->nr_running
&& idle_cpu(cpu
)) {
1623 if (READ_ONCE(rq
->numa_migrate_on
) ||
1624 !cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
))
1627 if (ns
->idle_cpu
== -1)
1630 idle_core
= numa_idle_core(idle_core
, cpu
);
1635 ns
->weight
= cpumask_weight(cpumask_of_node(nid
));
1637 ns
->node_type
= numa_classify(env
->imbalance_pct
, ns
);
1640 ns
->idle_cpu
= idle_core
;
1643 static void task_numa_assign(struct task_numa_env
*env
,
1644 struct task_struct
*p
, long imp
)
1646 struct rq
*rq
= cpu_rq(env
->dst_cpu
);
1648 /* Check if run-queue part of active NUMA balance. */
1649 if (env
->best_cpu
!= env
->dst_cpu
&& xchg(&rq
->numa_migrate_on
, 1)) {
1651 int start
= env
->dst_cpu
;
1653 /* Find alternative idle CPU. */
1654 for_each_cpu_wrap(cpu
, cpumask_of_node(env
->dst_nid
), start
) {
1655 if (cpu
== env
->best_cpu
|| !idle_cpu(cpu
) ||
1656 !cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
)) {
1661 rq
= cpu_rq(env
->dst_cpu
);
1662 if (!xchg(&rq
->numa_migrate_on
, 1))
1666 /* Failed to find an alternative idle CPU */
1672 * Clear previous best_cpu/rq numa-migrate flag, since task now
1673 * found a better CPU to move/swap.
1675 if (env
->best_cpu
!= -1 && env
->best_cpu
!= env
->dst_cpu
) {
1676 rq
= cpu_rq(env
->best_cpu
);
1677 WRITE_ONCE(rq
->numa_migrate_on
, 0);
1681 put_task_struct(env
->best_task
);
1686 env
->best_imp
= imp
;
1687 env
->best_cpu
= env
->dst_cpu
;
1690 static bool load_too_imbalanced(long src_load
, long dst_load
,
1691 struct task_numa_env
*env
)
1694 long orig_src_load
, orig_dst_load
;
1695 long src_capacity
, dst_capacity
;
1698 * The load is corrected for the CPU capacity available on each node.
1701 * ------------ vs ---------
1702 * src_capacity dst_capacity
1704 src_capacity
= env
->src_stats
.compute_capacity
;
1705 dst_capacity
= env
->dst_stats
.compute_capacity
;
1707 imb
= abs(dst_load
* src_capacity
- src_load
* dst_capacity
);
1709 orig_src_load
= env
->src_stats
.load
;
1710 orig_dst_load
= env
->dst_stats
.load
;
1712 old_imb
= abs(orig_dst_load
* src_capacity
- orig_src_load
* dst_capacity
);
1714 /* Would this change make things worse? */
1715 return (imb
> old_imb
);
1719 * Maximum NUMA importance can be 1998 (2*999);
1720 * SMALLIMP @ 30 would be close to 1998/64.
1721 * Used to deter task migration.
1726 * This checks if the overall compute and NUMA accesses of the system would
1727 * be improved if the source tasks was migrated to the target dst_cpu taking
1728 * into account that it might be best if task running on the dst_cpu should
1729 * be exchanged with the source task
1731 static bool task_numa_compare(struct task_numa_env
*env
,
1732 long taskimp
, long groupimp
, bool maymove
)
1734 struct numa_group
*cur_ng
, *p_ng
= deref_curr_numa_group(env
->p
);
1735 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1736 long imp
= p_ng
? groupimp
: taskimp
;
1737 struct task_struct
*cur
;
1738 long src_load
, dst_load
;
1739 int dist
= env
->dist
;
1742 bool stopsearch
= false;
1744 if (READ_ONCE(dst_rq
->numa_migrate_on
))
1748 cur
= rcu_dereference(dst_rq
->curr
);
1749 if (cur
&& ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
)))
1753 * Because we have preemption enabled we can get migrated around and
1754 * end try selecting ourselves (current == env->p) as a swap candidate.
1756 if (cur
== env
->p
) {
1762 if (maymove
&& moveimp
>= env
->best_imp
)
1768 /* Skip this swap candidate if cannot move to the source cpu. */
1769 if (!cpumask_test_cpu(env
->src_cpu
, cur
->cpus_ptr
))
1773 * Skip this swap candidate if it is not moving to its preferred
1774 * node and the best task is.
1776 if (env
->best_task
&&
1777 env
->best_task
->numa_preferred_nid
== env
->src_nid
&&
1778 cur
->numa_preferred_nid
!= env
->src_nid
) {
1783 * "imp" is the fault differential for the source task between the
1784 * source and destination node. Calculate the total differential for
1785 * the source task and potential destination task. The more negative
1786 * the value is, the more remote accesses that would be expected to
1787 * be incurred if the tasks were swapped.
1789 * If dst and source tasks are in the same NUMA group, or not
1790 * in any group then look only at task weights.
1792 cur_ng
= rcu_dereference(cur
->numa_group
);
1793 if (cur_ng
== p_ng
) {
1794 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1795 task_weight(cur
, env
->dst_nid
, dist
);
1797 * Add some hysteresis to prevent swapping the
1798 * tasks within a group over tiny differences.
1804 * Compare the group weights. If a task is all by itself
1805 * (not part of a group), use the task weight instead.
1808 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1809 group_weight(cur
, env
->dst_nid
, dist
);
1811 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1812 task_weight(cur
, env
->dst_nid
, dist
);
1815 /* Discourage picking a task already on its preferred node */
1816 if (cur
->numa_preferred_nid
== env
->dst_nid
)
1820 * Encourage picking a task that moves to its preferred node.
1821 * This potentially makes imp larger than it's maximum of
1822 * 1998 (see SMALLIMP and task_weight for why) but in this
1823 * case, it does not matter.
1825 if (cur
->numa_preferred_nid
== env
->src_nid
)
1828 if (maymove
&& moveimp
> imp
&& moveimp
> env
->best_imp
) {
1835 * Prefer swapping with a task moving to its preferred node over a
1838 if (env
->best_task
&& cur
->numa_preferred_nid
== env
->src_nid
&&
1839 env
->best_task
->numa_preferred_nid
!= env
->src_nid
) {
1844 * If the NUMA importance is less than SMALLIMP,
1845 * task migration might only result in ping pong
1846 * of tasks and also hurt performance due to cache
1849 if (imp
< SMALLIMP
|| imp
<= env
->best_imp
+ SMALLIMP
/ 2)
1853 * In the overloaded case, try and keep the load balanced.
1855 load
= task_h_load(env
->p
) - task_h_load(cur
);
1859 dst_load
= env
->dst_stats
.load
+ load
;
1860 src_load
= env
->src_stats
.load
- load
;
1862 if (load_too_imbalanced(src_load
, dst_load
, env
))
1866 /* Evaluate an idle CPU for a task numa move. */
1868 int cpu
= env
->dst_stats
.idle_cpu
;
1870 /* Nothing cached so current CPU went idle since the search. */
1875 * If the CPU is no longer truly idle and the previous best CPU
1876 * is, keep using it.
1878 if (!idle_cpu(cpu
) && env
->best_cpu
>= 0 &&
1879 idle_cpu(env
->best_cpu
)) {
1880 cpu
= env
->best_cpu
;
1886 task_numa_assign(env
, cur
, imp
);
1889 * If a move to idle is allowed because there is capacity or load
1890 * balance improves then stop the search. While a better swap
1891 * candidate may exist, a search is not free.
1893 if (maymove
&& !cur
&& env
->best_cpu
>= 0 && idle_cpu(env
->best_cpu
))
1897 * If a swap candidate must be identified and the current best task
1898 * moves its preferred node then stop the search.
1900 if (!maymove
&& env
->best_task
&&
1901 env
->best_task
->numa_preferred_nid
== env
->src_nid
) {
1910 static void task_numa_find_cpu(struct task_numa_env
*env
,
1911 long taskimp
, long groupimp
)
1913 bool maymove
= false;
1917 * If dst node has spare capacity, then check if there is an
1918 * imbalance that would be overruled by the load balancer.
1920 if (env
->dst_stats
.node_type
== node_has_spare
) {
1921 unsigned int imbalance
;
1922 int src_running
, dst_running
;
1925 * Would movement cause an imbalance? Note that if src has
1926 * more running tasks that the imbalance is ignored as the
1927 * move improves the imbalance from the perspective of the
1928 * CPU load balancer.
1930 src_running
= env
->src_stats
.nr_running
- 1;
1931 dst_running
= env
->dst_stats
.nr_running
+ 1;
1932 imbalance
= max(0, dst_running
- src_running
);
1933 imbalance
= adjust_numa_imbalance(imbalance
, dst_running
);
1935 /* Use idle CPU if there is no imbalance */
1938 if (env
->dst_stats
.idle_cpu
>= 0) {
1939 env
->dst_cpu
= env
->dst_stats
.idle_cpu
;
1940 task_numa_assign(env
, NULL
, 0);
1945 long src_load
, dst_load
, load
;
1947 * If the improvement from just moving env->p direction is better
1948 * than swapping tasks around, check if a move is possible.
1950 load
= task_h_load(env
->p
);
1951 dst_load
= env
->dst_stats
.load
+ load
;
1952 src_load
= env
->src_stats
.load
- load
;
1953 maymove
= !load_too_imbalanced(src_load
, dst_load
, env
);
1956 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1957 /* Skip this CPU if the source task cannot migrate */
1958 if (!cpumask_test_cpu(cpu
, env
->p
->cpus_ptr
))
1962 if (task_numa_compare(env
, taskimp
, groupimp
, maymove
))
1967 static int task_numa_migrate(struct task_struct
*p
)
1969 struct task_numa_env env
= {
1972 .src_cpu
= task_cpu(p
),
1973 .src_nid
= task_node(p
),
1975 .imbalance_pct
= 112,
1981 unsigned long taskweight
, groupweight
;
1982 struct sched_domain
*sd
;
1983 long taskimp
, groupimp
;
1984 struct numa_group
*ng
;
1989 * Pick the lowest SD_NUMA domain, as that would have the smallest
1990 * imbalance and would be the first to start moving tasks about.
1992 * And we want to avoid any moving of tasks about, as that would create
1993 * random movement of tasks -- counter the numa conditions we're trying
1997 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1999 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
2003 * Cpusets can break the scheduler domain tree into smaller
2004 * balance domains, some of which do not cross NUMA boundaries.
2005 * Tasks that are "trapped" in such domains cannot be migrated
2006 * elsewhere, so there is no point in (re)trying.
2008 if (unlikely(!sd
)) {
2009 sched_setnuma(p
, task_node(p
));
2013 env
.dst_nid
= p
->numa_preferred_nid
;
2014 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
2015 taskweight
= task_weight(p
, env
.src_nid
, dist
);
2016 groupweight
= group_weight(p
, env
.src_nid
, dist
);
2017 update_numa_stats(&env
, &env
.src_stats
, env
.src_nid
, false);
2018 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
2019 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
2020 update_numa_stats(&env
, &env
.dst_stats
, env
.dst_nid
, true);
2022 /* Try to find a spot on the preferred nid. */
2023 task_numa_find_cpu(&env
, taskimp
, groupimp
);
2026 * Look at other nodes in these cases:
2027 * - there is no space available on the preferred_nid
2028 * - the task is part of a numa_group that is interleaved across
2029 * multiple NUMA nodes; in order to better consolidate the group,
2030 * we need to check other locations.
2032 ng
= deref_curr_numa_group(p
);
2033 if (env
.best_cpu
== -1 || (ng
&& ng
->active_nodes
> 1)) {
2034 for_each_online_node(nid
) {
2035 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
2038 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
2039 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
2041 taskweight
= task_weight(p
, env
.src_nid
, dist
);
2042 groupweight
= group_weight(p
, env
.src_nid
, dist
);
2045 /* Only consider nodes where both task and groups benefit */
2046 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
2047 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
2048 if (taskimp
< 0 && groupimp
< 0)
2053 update_numa_stats(&env
, &env
.dst_stats
, env
.dst_nid
, true);
2054 task_numa_find_cpu(&env
, taskimp
, groupimp
);
2059 * If the task is part of a workload that spans multiple NUMA nodes,
2060 * and is migrating into one of the workload's active nodes, remember
2061 * this node as the task's preferred numa node, so the workload can
2063 * A task that migrated to a second choice node will be better off
2064 * trying for a better one later. Do not set the preferred node here.
2067 if (env
.best_cpu
== -1)
2070 nid
= cpu_to_node(env
.best_cpu
);
2072 if (nid
!= p
->numa_preferred_nid
)
2073 sched_setnuma(p
, nid
);
2076 /* No better CPU than the current one was found. */
2077 if (env
.best_cpu
== -1) {
2078 trace_sched_stick_numa(p
, env
.src_cpu
, NULL
, -1);
2082 best_rq
= cpu_rq(env
.best_cpu
);
2083 if (env
.best_task
== NULL
) {
2084 ret
= migrate_task_to(p
, env
.best_cpu
);
2085 WRITE_ONCE(best_rq
->numa_migrate_on
, 0);
2087 trace_sched_stick_numa(p
, env
.src_cpu
, NULL
, env
.best_cpu
);
2091 ret
= migrate_swap(p
, env
.best_task
, env
.best_cpu
, env
.src_cpu
);
2092 WRITE_ONCE(best_rq
->numa_migrate_on
, 0);
2095 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_task
, env
.best_cpu
);
2096 put_task_struct(env
.best_task
);
2100 /* Attempt to migrate a task to a CPU on the preferred node. */
2101 static void numa_migrate_preferred(struct task_struct
*p
)
2103 unsigned long interval
= HZ
;
2105 /* This task has no NUMA fault statistics yet */
2106 if (unlikely(p
->numa_preferred_nid
== NUMA_NO_NODE
|| !p
->numa_faults
))
2109 /* Periodically retry migrating the task to the preferred node */
2110 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
2111 p
->numa_migrate_retry
= jiffies
+ interval
;
2113 /* Success if task is already running on preferred CPU */
2114 if (task_node(p
) == p
->numa_preferred_nid
)
2117 /* Otherwise, try migrate to a CPU on the preferred node */
2118 task_numa_migrate(p
);
2122 * Find out how many nodes on the workload is actively running on. Do this by
2123 * tracking the nodes from which NUMA hinting faults are triggered. This can
2124 * be different from the set of nodes where the workload's memory is currently
2127 static void numa_group_count_active_nodes(struct numa_group
*numa_group
)
2129 unsigned long faults
, max_faults
= 0;
2130 int nid
, active_nodes
= 0;
2132 for_each_online_node(nid
) {
2133 faults
= group_faults_cpu(numa_group
, nid
);
2134 if (faults
> max_faults
)
2135 max_faults
= faults
;
2138 for_each_online_node(nid
) {
2139 faults
= group_faults_cpu(numa_group
, nid
);
2140 if (faults
* ACTIVE_NODE_FRACTION
> max_faults
)
2144 numa_group
->max_faults_cpu
= max_faults
;
2145 numa_group
->active_nodes
= active_nodes
;
2149 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2150 * increments. The more local the fault statistics are, the higher the scan
2151 * period will be for the next scan window. If local/(local+remote) ratio is
2152 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2153 * the scan period will decrease. Aim for 70% local accesses.
2155 #define NUMA_PERIOD_SLOTS 10
2156 #define NUMA_PERIOD_THRESHOLD 7
2159 * Increase the scan period (slow down scanning) if the majority of
2160 * our memory is already on our local node, or if the majority of
2161 * the page accesses are shared with other processes.
2162 * Otherwise, decrease the scan period.
2164 static void update_task_scan_period(struct task_struct
*p
,
2165 unsigned long shared
, unsigned long private)
2167 unsigned int period_slot
;
2168 int lr_ratio
, ps_ratio
;
2171 unsigned long remote
= p
->numa_faults_locality
[0];
2172 unsigned long local
= p
->numa_faults_locality
[1];
2175 * If there were no record hinting faults then either the task is
2176 * completely idle or all activity is areas that are not of interest
2177 * to automatic numa balancing. Related to that, if there were failed
2178 * migration then it implies we are migrating too quickly or the local
2179 * node is overloaded. In either case, scan slower
2181 if (local
+ shared
== 0 || p
->numa_faults_locality
[2]) {
2182 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
2183 p
->numa_scan_period
<< 1);
2185 p
->mm
->numa_next_scan
= jiffies
+
2186 msecs_to_jiffies(p
->numa_scan_period
);
2192 * Prepare to scale scan period relative to the current period.
2193 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2194 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2195 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2197 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
2198 lr_ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
2199 ps_ratio
= (private * NUMA_PERIOD_SLOTS
) / (private + shared
);
2201 if (ps_ratio
>= NUMA_PERIOD_THRESHOLD
) {
2203 * Most memory accesses are local. There is no need to
2204 * do fast NUMA scanning, since memory is already local.
2206 int slot
= ps_ratio
- NUMA_PERIOD_THRESHOLD
;
2209 diff
= slot
* period_slot
;
2210 } else if (lr_ratio
>= NUMA_PERIOD_THRESHOLD
) {
2212 * Most memory accesses are shared with other tasks.
2213 * There is no point in continuing fast NUMA scanning,
2214 * since other tasks may just move the memory elsewhere.
2216 int slot
= lr_ratio
- NUMA_PERIOD_THRESHOLD
;
2219 diff
= slot
* period_slot
;
2222 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2223 * yet they are not on the local NUMA node. Speed up
2224 * NUMA scanning to get the memory moved over.
2226 int ratio
= max(lr_ratio
, ps_ratio
);
2227 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
2230 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
2231 task_scan_min(p
), task_scan_max(p
));
2232 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2236 * Get the fraction of time the task has been running since the last
2237 * NUMA placement cycle. The scheduler keeps similar statistics, but
2238 * decays those on a 32ms period, which is orders of magnitude off
2239 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2240 * stats only if the task is so new there are no NUMA statistics yet.
2242 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
2244 u64 runtime
, delta
, now
;
2245 /* Use the start of this time slice to avoid calculations. */
2246 now
= p
->se
.exec_start
;
2247 runtime
= p
->se
.sum_exec_runtime
;
2249 if (p
->last_task_numa_placement
) {
2250 delta
= runtime
- p
->last_sum_exec_runtime
;
2251 *period
= now
- p
->last_task_numa_placement
;
2253 /* Avoid time going backwards, prevent potential divide error: */
2254 if (unlikely((s64
)*period
< 0))
2257 delta
= p
->se
.avg
.load_sum
;
2258 *period
= LOAD_AVG_MAX
;
2261 p
->last_sum_exec_runtime
= runtime
;
2262 p
->last_task_numa_placement
= now
;
2268 * Determine the preferred nid for a task in a numa_group. This needs to
2269 * be done in a way that produces consistent results with group_weight,
2270 * otherwise workloads might not converge.
2272 static int preferred_group_nid(struct task_struct
*p
, int nid
)
2277 /* Direct connections between all NUMA nodes. */
2278 if (sched_numa_topology_type
== NUMA_DIRECT
)
2282 * On a system with glueless mesh NUMA topology, group_weight
2283 * scores nodes according to the number of NUMA hinting faults on
2284 * both the node itself, and on nearby nodes.
2286 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
2287 unsigned long score
, max_score
= 0;
2288 int node
, max_node
= nid
;
2290 dist
= sched_max_numa_distance
;
2292 for_each_online_node(node
) {
2293 score
= group_weight(p
, node
, dist
);
2294 if (score
> max_score
) {
2303 * Finding the preferred nid in a system with NUMA backplane
2304 * interconnect topology is more involved. The goal is to locate
2305 * tasks from numa_groups near each other in the system, and
2306 * untangle workloads from different sides of the system. This requires
2307 * searching down the hierarchy of node groups, recursively searching
2308 * inside the highest scoring group of nodes. The nodemask tricks
2309 * keep the complexity of the search down.
2311 nodes
= node_online_map
;
2312 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
2313 unsigned long max_faults
= 0;
2314 nodemask_t max_group
= NODE_MASK_NONE
;
2317 /* Are there nodes at this distance from each other? */
2318 if (!find_numa_distance(dist
))
2321 for_each_node_mask(a
, nodes
) {
2322 unsigned long faults
= 0;
2323 nodemask_t this_group
;
2324 nodes_clear(this_group
);
2326 /* Sum group's NUMA faults; includes a==b case. */
2327 for_each_node_mask(b
, nodes
) {
2328 if (node_distance(a
, b
) < dist
) {
2329 faults
+= group_faults(p
, b
);
2330 node_set(b
, this_group
);
2331 node_clear(b
, nodes
);
2335 /* Remember the top group. */
2336 if (faults
> max_faults
) {
2337 max_faults
= faults
;
2338 max_group
= this_group
;
2340 * subtle: at the smallest distance there is
2341 * just one node left in each "group", the
2342 * winner is the preferred nid.
2347 /* Next round, evaluate the nodes within max_group. */
2355 static void task_numa_placement(struct task_struct
*p
)
2357 int seq
, nid
, max_nid
= NUMA_NO_NODE
;
2358 unsigned long max_faults
= 0;
2359 unsigned long fault_types
[2] = { 0, 0 };
2360 unsigned long total_faults
;
2361 u64 runtime
, period
;
2362 spinlock_t
*group_lock
= NULL
;
2363 struct numa_group
*ng
;
2366 * The p->mm->numa_scan_seq field gets updated without
2367 * exclusive access. Use READ_ONCE() here to ensure
2368 * that the field is read in a single access:
2370 seq
= READ_ONCE(p
->mm
->numa_scan_seq
);
2371 if (p
->numa_scan_seq
== seq
)
2373 p
->numa_scan_seq
= seq
;
2374 p
->numa_scan_period_max
= task_scan_max(p
);
2376 total_faults
= p
->numa_faults_locality
[0] +
2377 p
->numa_faults_locality
[1];
2378 runtime
= numa_get_avg_runtime(p
, &period
);
2380 /* If the task is part of a group prevent parallel updates to group stats */
2381 ng
= deref_curr_numa_group(p
);
2383 group_lock
= &ng
->lock
;
2384 spin_lock_irq(group_lock
);
2387 /* Find the node with the highest number of faults */
2388 for_each_online_node(nid
) {
2389 /* Keep track of the offsets in numa_faults array */
2390 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
2391 unsigned long faults
= 0, group_faults
= 0;
2394 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
2395 long diff
, f_diff
, f_weight
;
2397 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
2398 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
2399 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
2400 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
2402 /* Decay existing window, copy faults since last scan */
2403 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
2404 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
2405 p
->numa_faults
[membuf_idx
] = 0;
2408 * Normalize the faults_from, so all tasks in a group
2409 * count according to CPU use, instead of by the raw
2410 * number of faults. Tasks with little runtime have
2411 * little over-all impact on throughput, and thus their
2412 * faults are less important.
2414 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
2415 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
2417 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
2418 p
->numa_faults
[cpubuf_idx
] = 0;
2420 p
->numa_faults
[mem_idx
] += diff
;
2421 p
->numa_faults
[cpu_idx
] += f_diff
;
2422 faults
+= p
->numa_faults
[mem_idx
];
2423 p
->total_numa_faults
+= diff
;
2426 * safe because we can only change our own group
2428 * mem_idx represents the offset for a given
2429 * nid and priv in a specific region because it
2430 * is at the beginning of the numa_faults array.
2432 ng
->faults
[mem_idx
] += diff
;
2433 ng
->faults_cpu
[mem_idx
] += f_diff
;
2434 ng
->total_faults
+= diff
;
2435 group_faults
+= ng
->faults
[mem_idx
];
2440 if (faults
> max_faults
) {
2441 max_faults
= faults
;
2444 } else if (group_faults
> max_faults
) {
2445 max_faults
= group_faults
;
2451 numa_group_count_active_nodes(ng
);
2452 spin_unlock_irq(group_lock
);
2453 max_nid
= preferred_group_nid(p
, max_nid
);
2457 /* Set the new preferred node */
2458 if (max_nid
!= p
->numa_preferred_nid
)
2459 sched_setnuma(p
, max_nid
);
2462 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
2465 static inline int get_numa_group(struct numa_group
*grp
)
2467 return refcount_inc_not_zero(&grp
->refcount
);
2470 static inline void put_numa_group(struct numa_group
*grp
)
2472 if (refcount_dec_and_test(&grp
->refcount
))
2473 kfree_rcu(grp
, rcu
);
2476 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
2479 struct numa_group
*grp
, *my_grp
;
2480 struct task_struct
*tsk
;
2482 int cpu
= cpupid_to_cpu(cpupid
);
2485 if (unlikely(!deref_curr_numa_group(p
))) {
2486 unsigned int size
= sizeof(struct numa_group
) +
2487 4*nr_node_ids
*sizeof(unsigned long);
2489 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
2493 refcount_set(&grp
->refcount
, 1);
2494 grp
->active_nodes
= 1;
2495 grp
->max_faults_cpu
= 0;
2496 spin_lock_init(&grp
->lock
);
2498 /* Second half of the array tracks nids where faults happen */
2499 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
2502 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2503 grp
->faults
[i
] = p
->numa_faults
[i
];
2505 grp
->total_faults
= p
->total_numa_faults
;
2508 rcu_assign_pointer(p
->numa_group
, grp
);
2512 tsk
= READ_ONCE(cpu_rq(cpu
)->curr
);
2514 if (!cpupid_match_pid(tsk
, cpupid
))
2517 grp
= rcu_dereference(tsk
->numa_group
);
2521 my_grp
= deref_curr_numa_group(p
);
2526 * Only join the other group if its bigger; if we're the bigger group,
2527 * the other task will join us.
2529 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
2533 * Tie-break on the grp address.
2535 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
2538 /* Always join threads in the same process. */
2539 if (tsk
->mm
== current
->mm
)
2542 /* Simple filter to avoid false positives due to PID collisions */
2543 if (flags
& TNF_SHARED
)
2546 /* Update priv based on whether false sharing was detected */
2549 if (join
&& !get_numa_group(grp
))
2557 BUG_ON(irqs_disabled());
2558 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
2560 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
2561 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
2562 grp
->faults
[i
] += p
->numa_faults
[i
];
2564 my_grp
->total_faults
-= p
->total_numa_faults
;
2565 grp
->total_faults
+= p
->total_numa_faults
;
2570 spin_unlock(&my_grp
->lock
);
2571 spin_unlock_irq(&grp
->lock
);
2573 rcu_assign_pointer(p
->numa_group
, grp
);
2575 put_numa_group(my_grp
);
2584 * Get rid of NUMA staticstics associated with a task (either current or dead).
2585 * If @final is set, the task is dead and has reached refcount zero, so we can
2586 * safely free all relevant data structures. Otherwise, there might be
2587 * concurrent reads from places like load balancing and procfs, and we should
2588 * reset the data back to default state without freeing ->numa_faults.
2590 void task_numa_free(struct task_struct
*p
, bool final
)
2592 /* safe: p either is current or is being freed by current */
2593 struct numa_group
*grp
= rcu_dereference_raw(p
->numa_group
);
2594 unsigned long *numa_faults
= p
->numa_faults
;
2595 unsigned long flags
;
2602 spin_lock_irqsave(&grp
->lock
, flags
);
2603 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2604 grp
->faults
[i
] -= p
->numa_faults
[i
];
2605 grp
->total_faults
-= p
->total_numa_faults
;
2608 spin_unlock_irqrestore(&grp
->lock
, flags
);
2609 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2610 put_numa_group(grp
);
2614 p
->numa_faults
= NULL
;
2617 p
->total_numa_faults
= 0;
2618 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2624 * Got a PROT_NONE fault for a page on @node.
2626 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2628 struct task_struct
*p
= current
;
2629 bool migrated
= flags
& TNF_MIGRATED
;
2630 int cpu_node
= task_node(current
);
2631 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2632 struct numa_group
*ng
;
2635 if (!static_branch_likely(&sched_numa_balancing
))
2638 /* for example, ksmd faulting in a user's mm */
2642 /* Allocate buffer to track faults on a per-node basis */
2643 if (unlikely(!p
->numa_faults
)) {
2644 int size
= sizeof(*p
->numa_faults
) *
2645 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2647 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2648 if (!p
->numa_faults
)
2651 p
->total_numa_faults
= 0;
2652 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2656 * First accesses are treated as private, otherwise consider accesses
2657 * to be private if the accessing pid has not changed
2659 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2662 priv
= cpupid_match_pid(p
, last_cpupid
);
2663 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2664 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2668 * If a workload spans multiple NUMA nodes, a shared fault that
2669 * occurs wholly within the set of nodes that the workload is
2670 * actively using should be counted as local. This allows the
2671 * scan rate to slow down when a workload has settled down.
2673 ng
= deref_curr_numa_group(p
);
2674 if (!priv
&& !local
&& ng
&& ng
->active_nodes
> 1 &&
2675 numa_is_active_node(cpu_node
, ng
) &&
2676 numa_is_active_node(mem_node
, ng
))
2680 * Retry to migrate task to preferred node periodically, in case it
2681 * previously failed, or the scheduler moved us.
2683 if (time_after(jiffies
, p
->numa_migrate_retry
)) {
2684 task_numa_placement(p
);
2685 numa_migrate_preferred(p
);
2689 p
->numa_pages_migrated
+= pages
;
2690 if (flags
& TNF_MIGRATE_FAIL
)
2691 p
->numa_faults_locality
[2] += pages
;
2693 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2694 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2695 p
->numa_faults_locality
[local
] += pages
;
2698 static void reset_ptenuma_scan(struct task_struct
*p
)
2701 * We only did a read acquisition of the mmap sem, so
2702 * p->mm->numa_scan_seq is written to without exclusive access
2703 * and the update is not guaranteed to be atomic. That's not
2704 * much of an issue though, since this is just used for
2705 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2706 * expensive, to avoid any form of compiler optimizations:
2708 WRITE_ONCE(p
->mm
->numa_scan_seq
, READ_ONCE(p
->mm
->numa_scan_seq
) + 1);
2709 p
->mm
->numa_scan_offset
= 0;
2713 * The expensive part of numa migration is done from task_work context.
2714 * Triggered from task_tick_numa().
2716 static void task_numa_work(struct callback_head
*work
)
2718 unsigned long migrate
, next_scan
, now
= jiffies
;
2719 struct task_struct
*p
= current
;
2720 struct mm_struct
*mm
= p
->mm
;
2721 u64 runtime
= p
->se
.sum_exec_runtime
;
2722 struct vm_area_struct
*vma
;
2723 unsigned long start
, end
;
2724 unsigned long nr_pte_updates
= 0;
2725 long pages
, virtpages
;
2727 SCHED_WARN_ON(p
!= container_of(work
, struct task_struct
, numa_work
));
2731 * Who cares about NUMA placement when they're dying.
2733 * NOTE: make sure not to dereference p->mm before this check,
2734 * exit_task_work() happens _after_ exit_mm() so we could be called
2735 * without p->mm even though we still had it when we enqueued this
2738 if (p
->flags
& PF_EXITING
)
2741 if (!mm
->numa_next_scan
) {
2742 mm
->numa_next_scan
= now
+
2743 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2747 * Enforce maximal scan/migration frequency..
2749 migrate
= mm
->numa_next_scan
;
2750 if (time_before(now
, migrate
))
2753 if (p
->numa_scan_period
== 0) {
2754 p
->numa_scan_period_max
= task_scan_max(p
);
2755 p
->numa_scan_period
= task_scan_start(p
);
2758 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2759 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2763 * Delay this task enough that another task of this mm will likely win
2764 * the next time around.
2766 p
->node_stamp
+= 2 * TICK_NSEC
;
2768 start
= mm
->numa_scan_offset
;
2769 pages
= sysctl_numa_balancing_scan_size
;
2770 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2771 virtpages
= pages
* 8; /* Scan up to this much virtual space */
2776 if (!mmap_read_trylock(mm
))
2778 vma
= find_vma(mm
, start
);
2780 reset_ptenuma_scan(p
);
2784 for (; vma
; vma
= vma
->vm_next
) {
2785 if (!vma_migratable(vma
) || !vma_policy_mof(vma
) ||
2786 is_vm_hugetlb_page(vma
) || (vma
->vm_flags
& VM_MIXEDMAP
)) {
2791 * Shared library pages mapped by multiple processes are not
2792 * migrated as it is expected they are cache replicated. Avoid
2793 * hinting faults in read-only file-backed mappings or the vdso
2794 * as migrating the pages will be of marginal benefit.
2797 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2801 * Skip inaccessible VMAs to avoid any confusion between
2802 * PROT_NONE and NUMA hinting ptes
2804 if (!vma_is_accessible(vma
))
2808 start
= max(start
, vma
->vm_start
);
2809 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2810 end
= min(end
, vma
->vm_end
);
2811 nr_pte_updates
= change_prot_numa(vma
, start
, end
);
2814 * Try to scan sysctl_numa_balancing_size worth of
2815 * hpages that have at least one present PTE that
2816 * is not already pte-numa. If the VMA contains
2817 * areas that are unused or already full of prot_numa
2818 * PTEs, scan up to virtpages, to skip through those
2822 pages
-= (end
- start
) >> PAGE_SHIFT
;
2823 virtpages
-= (end
- start
) >> PAGE_SHIFT
;
2826 if (pages
<= 0 || virtpages
<= 0)
2830 } while (end
!= vma
->vm_end
);
2835 * It is possible to reach the end of the VMA list but the last few
2836 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2837 * would find the !migratable VMA on the next scan but not reset the
2838 * scanner to the start so check it now.
2841 mm
->numa_scan_offset
= start
;
2843 reset_ptenuma_scan(p
);
2844 mmap_read_unlock(mm
);
2847 * Make sure tasks use at least 32x as much time to run other code
2848 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2849 * Usually update_task_scan_period slows down scanning enough; on an
2850 * overloaded system we need to limit overhead on a per task basis.
2852 if (unlikely(p
->se
.sum_exec_runtime
!= runtime
)) {
2853 u64 diff
= p
->se
.sum_exec_runtime
- runtime
;
2854 p
->node_stamp
+= 32 * diff
;
2858 void init_numa_balancing(unsigned long clone_flags
, struct task_struct
*p
)
2861 struct mm_struct
*mm
= p
->mm
;
2864 mm_users
= atomic_read(&mm
->mm_users
);
2865 if (mm_users
== 1) {
2866 mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2867 mm
->numa_scan_seq
= 0;
2871 p
->numa_scan_seq
= mm
? mm
->numa_scan_seq
: 0;
2872 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2873 /* Protect against double add, see task_tick_numa and task_numa_work */
2874 p
->numa_work
.next
= &p
->numa_work
;
2875 p
->numa_faults
= NULL
;
2876 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2877 p
->last_task_numa_placement
= 0;
2878 p
->last_sum_exec_runtime
= 0;
2880 init_task_work(&p
->numa_work
, task_numa_work
);
2882 /* New address space, reset the preferred nid */
2883 if (!(clone_flags
& CLONE_VM
)) {
2884 p
->numa_preferred_nid
= NUMA_NO_NODE
;
2889 * New thread, keep existing numa_preferred_nid which should be copied
2890 * already by arch_dup_task_struct but stagger when scans start.
2895 delay
= min_t(unsigned int, task_scan_max(current
),
2896 current
->numa_scan_period
* mm_users
* NSEC_PER_MSEC
);
2897 delay
+= 2 * TICK_NSEC
;
2898 p
->node_stamp
= delay
;
2903 * Drive the periodic memory faults..
2905 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2907 struct callback_head
*work
= &curr
->numa_work
;
2911 * We don't care about NUMA placement if we don't have memory.
2913 if ((curr
->flags
& (PF_EXITING
| PF_KTHREAD
)) || work
->next
!= work
)
2917 * Using runtime rather than walltime has the dual advantage that
2918 * we (mostly) drive the selection from busy threads and that the
2919 * task needs to have done some actual work before we bother with
2922 now
= curr
->se
.sum_exec_runtime
;
2923 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2925 if (now
> curr
->node_stamp
+ period
) {
2926 if (!curr
->node_stamp
)
2927 curr
->numa_scan_period
= task_scan_start(curr
);
2928 curr
->node_stamp
+= period
;
2930 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
))
2931 task_work_add(curr
, work
, TWA_RESUME
);
2935 static void update_scan_period(struct task_struct
*p
, int new_cpu
)
2937 int src_nid
= cpu_to_node(task_cpu(p
));
2938 int dst_nid
= cpu_to_node(new_cpu
);
2940 if (!static_branch_likely(&sched_numa_balancing
))
2943 if (!p
->mm
|| !p
->numa_faults
|| (p
->flags
& PF_EXITING
))
2946 if (src_nid
== dst_nid
)
2950 * Allow resets if faults have been trapped before one scan
2951 * has completed. This is most likely due to a new task that
2952 * is pulled cross-node due to wakeups or load balancing.
2954 if (p
->numa_scan_seq
) {
2956 * Avoid scan adjustments if moving to the preferred
2957 * node or if the task was not previously running on
2958 * the preferred node.
2960 if (dst_nid
== p
->numa_preferred_nid
||
2961 (p
->numa_preferred_nid
!= NUMA_NO_NODE
&&
2962 src_nid
!= p
->numa_preferred_nid
))
2966 p
->numa_scan_period
= task_scan_start(p
);
2970 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2974 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2978 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2982 static inline void update_scan_period(struct task_struct
*p
, int new_cpu
)
2986 #endif /* CONFIG_NUMA_BALANCING */
2989 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2991 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2993 if (entity_is_task(se
)) {
2994 struct rq
*rq
= rq_of(cfs_rq
);
2996 account_numa_enqueue(rq
, task_of(se
));
2997 list_add(&se
->group_node
, &rq
->cfs_tasks
);
3000 cfs_rq
->nr_running
++;
3004 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3006 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
3008 if (entity_is_task(se
)) {
3009 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
3010 list_del_init(&se
->group_node
);
3013 cfs_rq
->nr_running
--;
3017 * Signed add and clamp on underflow.
3019 * Explicitly do a load-store to ensure the intermediate value never hits
3020 * memory. This allows lockless observations without ever seeing the negative
3023 #define add_positive(_ptr, _val) do { \
3024 typeof(_ptr) ptr = (_ptr); \
3025 typeof(_val) val = (_val); \
3026 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3030 if (val < 0 && res > var) \
3033 WRITE_ONCE(*ptr, res); \
3037 * Unsigned subtract and clamp on underflow.
3039 * Explicitly do a load-store to ensure the intermediate value never hits
3040 * memory. This allows lockless observations without ever seeing the negative
3043 #define sub_positive(_ptr, _val) do { \
3044 typeof(_ptr) ptr = (_ptr); \
3045 typeof(*ptr) val = (_val); \
3046 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3050 WRITE_ONCE(*ptr, res); \
3054 * Remove and clamp on negative, from a local variable.
3056 * A variant of sub_positive(), which does not use explicit load-store
3057 * and is thus optimized for local variable updates.
3059 #define lsub_positive(_ptr, _val) do { \
3060 typeof(_ptr) ptr = (_ptr); \
3061 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3066 enqueue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3068 cfs_rq
->avg
.load_avg
+= se
->avg
.load_avg
;
3069 cfs_rq
->avg
.load_sum
+= se_weight(se
) * se
->avg
.load_sum
;
3073 dequeue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3075 sub_positive(&cfs_rq
->avg
.load_avg
, se
->avg
.load_avg
);
3076 sub_positive(&cfs_rq
->avg
.load_sum
, se_weight(se
) * se
->avg
.load_sum
);
3080 enqueue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
3082 dequeue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
3085 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
3086 unsigned long weight
)
3089 /* commit outstanding execution time */
3090 if (cfs_rq
->curr
== se
)
3091 update_curr(cfs_rq
);
3092 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
3094 dequeue_load_avg(cfs_rq
, se
);
3096 update_load_set(&se
->load
, weight
);
3100 u32 divider
= get_pelt_divider(&se
->avg
);
3102 se
->avg
.load_avg
= div_u64(se_weight(se
) * se
->avg
.load_sum
, divider
);
3106 enqueue_load_avg(cfs_rq
, se
);
3108 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
3112 void reweight_task(struct task_struct
*p
, int prio
)
3114 struct sched_entity
*se
= &p
->se
;
3115 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3116 struct load_weight
*load
= &se
->load
;
3117 unsigned long weight
= scale_load(sched_prio_to_weight
[prio
]);
3119 reweight_entity(cfs_rq
, se
, weight
);
3120 load
->inv_weight
= sched_prio_to_wmult
[prio
];
3123 #ifdef CONFIG_FAIR_GROUP_SCHED
3126 * All this does is approximate the hierarchical proportion which includes that
3127 * global sum we all love to hate.
3129 * That is, the weight of a group entity, is the proportional share of the
3130 * group weight based on the group runqueue weights. That is:
3132 * tg->weight * grq->load.weight
3133 * ge->load.weight = ----------------------------- (1)
3134 * \Sum grq->load.weight
3136 * Now, because computing that sum is prohibitively expensive to compute (been
3137 * there, done that) we approximate it with this average stuff. The average
3138 * moves slower and therefore the approximation is cheaper and more stable.
3140 * So instead of the above, we substitute:
3142 * grq->load.weight -> grq->avg.load_avg (2)
3144 * which yields the following:
3146 * tg->weight * grq->avg.load_avg
3147 * ge->load.weight = ------------------------------ (3)
3150 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3152 * That is shares_avg, and it is right (given the approximation (2)).
3154 * The problem with it is that because the average is slow -- it was designed
3155 * to be exactly that of course -- this leads to transients in boundary
3156 * conditions. In specific, the case where the group was idle and we start the
3157 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3158 * yielding bad latency etc..
3160 * Now, in that special case (1) reduces to:
3162 * tg->weight * grq->load.weight
3163 * ge->load.weight = ----------------------------- = tg->weight (4)
3166 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3168 * So what we do is modify our approximation (3) to approach (4) in the (near)
3173 * tg->weight * grq->load.weight
3174 * --------------------------------------------------- (5)
3175 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3177 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3178 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3181 * tg->weight * grq->load.weight
3182 * ge->load.weight = ----------------------------- (6)
3187 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3188 * max(grq->load.weight, grq->avg.load_avg)
3190 * And that is shares_weight and is icky. In the (near) UP case it approaches
3191 * (4) while in the normal case it approaches (3). It consistently
3192 * overestimates the ge->load.weight and therefore:
3194 * \Sum ge->load.weight >= tg->weight
3198 static long calc_group_shares(struct cfs_rq
*cfs_rq
)
3200 long tg_weight
, tg_shares
, load
, shares
;
3201 struct task_group
*tg
= cfs_rq
->tg
;
3203 tg_shares
= READ_ONCE(tg
->shares
);
3205 load
= max(scale_load_down(cfs_rq
->load
.weight
), cfs_rq
->avg
.load_avg
);
3207 tg_weight
= atomic_long_read(&tg
->load_avg
);
3209 /* Ensure tg_weight >= load */
3210 tg_weight
-= cfs_rq
->tg_load_avg_contrib
;
3213 shares
= (tg_shares
* load
);
3215 shares
/= tg_weight
;
3218 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3219 * of a group with small tg->shares value. It is a floor value which is
3220 * assigned as a minimum load.weight to the sched_entity representing
3221 * the group on a CPU.
3223 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3224 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3225 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3226 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3229 return clamp_t(long, shares
, MIN_SHARES
, tg_shares
);
3231 #endif /* CONFIG_SMP */
3233 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
3236 * Recomputes the group entity based on the current state of its group
3239 static void update_cfs_group(struct sched_entity
*se
)
3241 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3247 if (throttled_hierarchy(gcfs_rq
))
3251 shares
= READ_ONCE(gcfs_rq
->tg
->shares
);
3253 if (likely(se
->load
.weight
== shares
))
3256 shares
= calc_group_shares(gcfs_rq
);
3259 reweight_entity(cfs_rq_of(se
), se
, shares
);
3262 #else /* CONFIG_FAIR_GROUP_SCHED */
3263 static inline void update_cfs_group(struct sched_entity
*se
)
3266 #endif /* CONFIG_FAIR_GROUP_SCHED */
3268 static inline void cfs_rq_util_change(struct cfs_rq
*cfs_rq
, int flags
)
3270 struct rq
*rq
= rq_of(cfs_rq
);
3272 if (&rq
->cfs
== cfs_rq
) {
3274 * There are a few boundary cases this might miss but it should
3275 * get called often enough that that should (hopefully) not be
3278 * It will not get called when we go idle, because the idle
3279 * thread is a different class (!fair), nor will the utilization
3280 * number include things like RT tasks.
3282 * As is, the util number is not freq-invariant (we'd have to
3283 * implement arch_scale_freq_capacity() for that).
3287 cpufreq_update_util(rq
, flags
);
3292 #ifdef CONFIG_FAIR_GROUP_SCHED
3294 * update_tg_load_avg - update the tg's load avg
3295 * @cfs_rq: the cfs_rq whose avg changed
3297 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3298 * However, because tg->load_avg is a global value there are performance
3301 * In order to avoid having to look at the other cfs_rq's, we use a
3302 * differential update where we store the last value we propagated. This in
3303 * turn allows skipping updates if the differential is 'small'.
3305 * Updating tg's load_avg is necessary before update_cfs_share().
3307 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
)
3309 long delta
= cfs_rq
->avg
.load_avg
- cfs_rq
->tg_load_avg_contrib
;
3312 * No need to update load_avg for root_task_group as it is not used.
3314 if (cfs_rq
->tg
== &root_task_group
)
3317 if (abs(delta
) > cfs_rq
->tg_load_avg_contrib
/ 64) {
3318 atomic_long_add(delta
, &cfs_rq
->tg
->load_avg
);
3319 cfs_rq
->tg_load_avg_contrib
= cfs_rq
->avg
.load_avg
;
3324 * Called within set_task_rq() right before setting a task's CPU. The
3325 * caller only guarantees p->pi_lock is held; no other assumptions,
3326 * including the state of rq->lock, should be made.
3328 void set_task_rq_fair(struct sched_entity
*se
,
3329 struct cfs_rq
*prev
, struct cfs_rq
*next
)
3331 u64 p_last_update_time
;
3332 u64 n_last_update_time
;
3334 if (!sched_feat(ATTACH_AGE_LOAD
))
3338 * We are supposed to update the task to "current" time, then its up to
3339 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3340 * getting what current time is, so simply throw away the out-of-date
3341 * time. This will result in the wakee task is less decayed, but giving
3342 * the wakee more load sounds not bad.
3344 if (!(se
->avg
.last_update_time
&& prev
))
3347 #ifndef CONFIG_64BIT
3349 u64 p_last_update_time_copy
;
3350 u64 n_last_update_time_copy
;
3353 p_last_update_time_copy
= prev
->load_last_update_time_copy
;
3354 n_last_update_time_copy
= next
->load_last_update_time_copy
;
3358 p_last_update_time
= prev
->avg
.last_update_time
;
3359 n_last_update_time
= next
->avg
.last_update_time
;
3361 } while (p_last_update_time
!= p_last_update_time_copy
||
3362 n_last_update_time
!= n_last_update_time_copy
);
3365 p_last_update_time
= prev
->avg
.last_update_time
;
3366 n_last_update_time
= next
->avg
.last_update_time
;
3368 __update_load_avg_blocked_se(p_last_update_time
, se
);
3369 se
->avg
.last_update_time
= n_last_update_time
;
3374 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3375 * propagate its contribution. The key to this propagation is the invariant
3376 * that for each group:
3378 * ge->avg == grq->avg (1)
3380 * _IFF_ we look at the pure running and runnable sums. Because they
3381 * represent the very same entity, just at different points in the hierarchy.
3383 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3384 * and simply copies the running/runnable sum over (but still wrong, because
3385 * the group entity and group rq do not have their PELT windows aligned).
3387 * However, update_tg_cfs_load() is more complex. So we have:
3389 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3391 * And since, like util, the runnable part should be directly transferable,
3392 * the following would _appear_ to be the straight forward approach:
3394 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3396 * And per (1) we have:
3398 * ge->avg.runnable_avg == grq->avg.runnable_avg
3402 * ge->load.weight * grq->avg.load_avg
3403 * ge->avg.load_avg = ----------------------------------- (4)
3406 * Except that is wrong!
3408 * Because while for entities historical weight is not important and we
3409 * really only care about our future and therefore can consider a pure
3410 * runnable sum, runqueues can NOT do this.
3412 * We specifically want runqueues to have a load_avg that includes
3413 * historical weights. Those represent the blocked load, the load we expect
3414 * to (shortly) return to us. This only works by keeping the weights as
3415 * integral part of the sum. We therefore cannot decompose as per (3).
3417 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3418 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3419 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3420 * runnable section of these tasks overlap (or not). If they were to perfectly
3421 * align the rq as a whole would be runnable 2/3 of the time. If however we
3422 * always have at least 1 runnable task, the rq as a whole is always runnable.
3424 * So we'll have to approximate.. :/
3426 * Given the constraint:
3428 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3430 * We can construct a rule that adds runnable to a rq by assuming minimal
3433 * On removal, we'll assume each task is equally runnable; which yields:
3435 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3437 * XXX: only do this for the part of runnable > running ?
3442 update_tg_cfs_util(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3444 long delta
= gcfs_rq
->avg
.util_avg
- se
->avg
.util_avg
;
3447 /* Nothing to update */
3452 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3453 * See ___update_load_avg() for details.
3455 divider
= get_pelt_divider(&cfs_rq
->avg
);
3457 /* Set new sched_entity's utilization */
3458 se
->avg
.util_avg
= gcfs_rq
->avg
.util_avg
;
3459 se
->avg
.util_sum
= se
->avg
.util_avg
* divider
;
3461 /* Update parent cfs_rq utilization */
3462 add_positive(&cfs_rq
->avg
.util_avg
, delta
);
3463 cfs_rq
->avg
.util_sum
= cfs_rq
->avg
.util_avg
* divider
;
3467 update_tg_cfs_runnable(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3469 long delta
= gcfs_rq
->avg
.runnable_avg
- se
->avg
.runnable_avg
;
3472 /* Nothing to update */
3477 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3478 * See ___update_load_avg() for details.
3480 divider
= get_pelt_divider(&cfs_rq
->avg
);
3482 /* Set new sched_entity's runnable */
3483 se
->avg
.runnable_avg
= gcfs_rq
->avg
.runnable_avg
;
3484 se
->avg
.runnable_sum
= se
->avg
.runnable_avg
* divider
;
3486 /* Update parent cfs_rq runnable */
3487 add_positive(&cfs_rq
->avg
.runnable_avg
, delta
);
3488 cfs_rq
->avg
.runnable_sum
= cfs_rq
->avg
.runnable_avg
* divider
;
3492 update_tg_cfs_load(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3494 long delta_avg
, running_sum
, runnable_sum
= gcfs_rq
->prop_runnable_sum
;
3495 unsigned long load_avg
;
3503 gcfs_rq
->prop_runnable_sum
= 0;
3506 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3507 * See ___update_load_avg() for details.
3509 divider
= get_pelt_divider(&cfs_rq
->avg
);
3511 if (runnable_sum
>= 0) {
3513 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3514 * the CPU is saturated running == runnable.
3516 runnable_sum
+= se
->avg
.load_sum
;
3517 runnable_sum
= min_t(long, runnable_sum
, divider
);
3520 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3521 * assuming all tasks are equally runnable.
3523 if (scale_load_down(gcfs_rq
->load
.weight
)) {
3524 load_sum
= div_s64(gcfs_rq
->avg
.load_sum
,
3525 scale_load_down(gcfs_rq
->load
.weight
));
3528 /* But make sure to not inflate se's runnable */
3529 runnable_sum
= min(se
->avg
.load_sum
, load_sum
);
3533 * runnable_sum can't be lower than running_sum
3534 * Rescale running sum to be in the same range as runnable sum
3535 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3536 * runnable_sum is in [0 : LOAD_AVG_MAX]
3538 running_sum
= se
->avg
.util_sum
>> SCHED_CAPACITY_SHIFT
;
3539 runnable_sum
= max(runnable_sum
, running_sum
);
3541 load_sum
= (s64
)se_weight(se
) * runnable_sum
;
3542 load_avg
= div_s64(load_sum
, divider
);
3544 delta_sum
= load_sum
- (s64
)se_weight(se
) * se
->avg
.load_sum
;
3545 delta_avg
= load_avg
- se
->avg
.load_avg
;
3547 se
->avg
.load_sum
= runnable_sum
;
3548 se
->avg
.load_avg
= load_avg
;
3549 add_positive(&cfs_rq
->avg
.load_avg
, delta_avg
);
3550 add_positive(&cfs_rq
->avg
.load_sum
, delta_sum
);
3553 static inline void add_tg_cfs_propagate(struct cfs_rq
*cfs_rq
, long runnable_sum
)
3555 cfs_rq
->propagate
= 1;
3556 cfs_rq
->prop_runnable_sum
+= runnable_sum
;
3559 /* Update task and its cfs_rq load average */
3560 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3562 struct cfs_rq
*cfs_rq
, *gcfs_rq
;
3564 if (entity_is_task(se
))
3567 gcfs_rq
= group_cfs_rq(se
);
3568 if (!gcfs_rq
->propagate
)
3571 gcfs_rq
->propagate
= 0;
3573 cfs_rq
= cfs_rq_of(se
);
3575 add_tg_cfs_propagate(cfs_rq
, gcfs_rq
->prop_runnable_sum
);
3577 update_tg_cfs_util(cfs_rq
, se
, gcfs_rq
);
3578 update_tg_cfs_runnable(cfs_rq
, se
, gcfs_rq
);
3579 update_tg_cfs_load(cfs_rq
, se
, gcfs_rq
);
3581 trace_pelt_cfs_tp(cfs_rq
);
3582 trace_pelt_se_tp(se
);
3588 * Check if we need to update the load and the utilization of a blocked
3591 static inline bool skip_blocked_update(struct sched_entity
*se
)
3593 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3596 * If sched_entity still have not zero load or utilization, we have to
3599 if (se
->avg
.load_avg
|| se
->avg
.util_avg
)
3603 * If there is a pending propagation, we have to update the load and
3604 * the utilization of the sched_entity:
3606 if (gcfs_rq
->propagate
)
3610 * Otherwise, the load and the utilization of the sched_entity is
3611 * already zero and there is no pending propagation, so it will be a
3612 * waste of time to try to decay it:
3617 #else /* CONFIG_FAIR_GROUP_SCHED */
3619 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
) {}
3621 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3626 static inline void add_tg_cfs_propagate(struct cfs_rq
*cfs_rq
, long runnable_sum
) {}
3628 #endif /* CONFIG_FAIR_GROUP_SCHED */
3631 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3632 * @now: current time, as per cfs_rq_clock_pelt()
3633 * @cfs_rq: cfs_rq to update
3635 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3636 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3637 * post_init_entity_util_avg().
3639 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3641 * Returns true if the load decayed or we removed load.
3643 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3644 * call update_tg_load_avg() when this function returns true.
3647 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
)
3649 unsigned long removed_load
= 0, removed_util
= 0, removed_runnable
= 0;
3650 struct sched_avg
*sa
= &cfs_rq
->avg
;
3653 if (cfs_rq
->removed
.nr
) {
3655 u32 divider
= get_pelt_divider(&cfs_rq
->avg
);
3657 raw_spin_lock(&cfs_rq
->removed
.lock
);
3658 swap(cfs_rq
->removed
.util_avg
, removed_util
);
3659 swap(cfs_rq
->removed
.load_avg
, removed_load
);
3660 swap(cfs_rq
->removed
.runnable_avg
, removed_runnable
);
3661 cfs_rq
->removed
.nr
= 0;
3662 raw_spin_unlock(&cfs_rq
->removed
.lock
);
3665 sub_positive(&sa
->load_avg
, r
);
3666 sub_positive(&sa
->load_sum
, r
* divider
);
3669 sub_positive(&sa
->util_avg
, r
);
3670 sub_positive(&sa
->util_sum
, r
* divider
);
3672 r
= removed_runnable
;
3673 sub_positive(&sa
->runnable_avg
, r
);
3674 sub_positive(&sa
->runnable_sum
, r
* divider
);
3677 * removed_runnable is the unweighted version of removed_load so we
3678 * can use it to estimate removed_load_sum.
3680 add_tg_cfs_propagate(cfs_rq
,
3681 -(long)(removed_runnable
* divider
) >> SCHED_CAPACITY_SHIFT
);
3686 decayed
|= __update_load_avg_cfs_rq(now
, cfs_rq
);
3688 #ifndef CONFIG_64BIT
3690 cfs_rq
->load_last_update_time_copy
= sa
->last_update_time
;
3697 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3698 * @cfs_rq: cfs_rq to attach to
3699 * @se: sched_entity to attach
3701 * Must call update_cfs_rq_load_avg() before this, since we rely on
3702 * cfs_rq->avg.last_update_time being current.
3704 static void attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3707 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3708 * See ___update_load_avg() for details.
3710 u32 divider
= get_pelt_divider(&cfs_rq
->avg
);
3713 * When we attach the @se to the @cfs_rq, we must align the decay
3714 * window because without that, really weird and wonderful things can
3719 se
->avg
.last_update_time
= cfs_rq
->avg
.last_update_time
;
3720 se
->avg
.period_contrib
= cfs_rq
->avg
.period_contrib
;
3723 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3724 * period_contrib. This isn't strictly correct, but since we're
3725 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3728 se
->avg
.util_sum
= se
->avg
.util_avg
* divider
;
3730 se
->avg
.runnable_sum
= se
->avg
.runnable_avg
* divider
;
3732 se
->avg
.load_sum
= divider
;
3733 if (se_weight(se
)) {
3735 div_u64(se
->avg
.load_avg
* se
->avg
.load_sum
, se_weight(se
));
3738 enqueue_load_avg(cfs_rq
, se
);
3739 cfs_rq
->avg
.util_avg
+= se
->avg
.util_avg
;
3740 cfs_rq
->avg
.util_sum
+= se
->avg
.util_sum
;
3741 cfs_rq
->avg
.runnable_avg
+= se
->avg
.runnable_avg
;
3742 cfs_rq
->avg
.runnable_sum
+= se
->avg
.runnable_sum
;
3744 add_tg_cfs_propagate(cfs_rq
, se
->avg
.load_sum
);
3746 cfs_rq_util_change(cfs_rq
, 0);
3748 trace_pelt_cfs_tp(cfs_rq
);
3752 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3753 * @cfs_rq: cfs_rq to detach from
3754 * @se: sched_entity to detach
3756 * Must call update_cfs_rq_load_avg() before this, since we rely on
3757 * cfs_rq->avg.last_update_time being current.
3759 static void detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3761 dequeue_load_avg(cfs_rq
, se
);
3762 sub_positive(&cfs_rq
->avg
.util_avg
, se
->avg
.util_avg
);
3763 sub_positive(&cfs_rq
->avg
.util_sum
, se
->avg
.util_sum
);
3764 sub_positive(&cfs_rq
->avg
.runnable_avg
, se
->avg
.runnable_avg
);
3765 sub_positive(&cfs_rq
->avg
.runnable_sum
, se
->avg
.runnable_sum
);
3767 add_tg_cfs_propagate(cfs_rq
, -se
->avg
.load_sum
);
3769 cfs_rq_util_change(cfs_rq
, 0);
3771 trace_pelt_cfs_tp(cfs_rq
);
3775 * Optional action to be done while updating the load average
3777 #define UPDATE_TG 0x1
3778 #define SKIP_AGE_LOAD 0x2
3779 #define DO_ATTACH 0x4
3781 /* Update task and its cfs_rq load average */
3782 static inline void update_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3784 u64 now
= cfs_rq_clock_pelt(cfs_rq
);
3788 * Track task load average for carrying it to new CPU after migrated, and
3789 * track group sched_entity load average for task_h_load calc in migration
3791 if (se
->avg
.last_update_time
&& !(flags
& SKIP_AGE_LOAD
))
3792 __update_load_avg_se(now
, cfs_rq
, se
);
3794 decayed
= update_cfs_rq_load_avg(now
, cfs_rq
);
3795 decayed
|= propagate_entity_load_avg(se
);
3797 if (!se
->avg
.last_update_time
&& (flags
& DO_ATTACH
)) {
3800 * DO_ATTACH means we're here from enqueue_entity().
3801 * !last_update_time means we've passed through
3802 * migrate_task_rq_fair() indicating we migrated.
3804 * IOW we're enqueueing a task on a new CPU.
3806 attach_entity_load_avg(cfs_rq
, se
);
3807 update_tg_load_avg(cfs_rq
);
3809 } else if (decayed
) {
3810 cfs_rq_util_change(cfs_rq
, 0);
3812 if (flags
& UPDATE_TG
)
3813 update_tg_load_avg(cfs_rq
);
3817 #ifndef CONFIG_64BIT
3818 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3820 u64 last_update_time_copy
;
3821 u64 last_update_time
;
3824 last_update_time_copy
= cfs_rq
->load_last_update_time_copy
;
3826 last_update_time
= cfs_rq
->avg
.last_update_time
;
3827 } while (last_update_time
!= last_update_time_copy
);
3829 return last_update_time
;
3832 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3834 return cfs_rq
->avg
.last_update_time
;
3839 * Synchronize entity load avg of dequeued entity without locking
3842 static void sync_entity_load_avg(struct sched_entity
*se
)
3844 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3845 u64 last_update_time
;
3847 last_update_time
= cfs_rq_last_update_time(cfs_rq
);
3848 __update_load_avg_blocked_se(last_update_time
, se
);
3852 * Task first catches up with cfs_rq, and then subtract
3853 * itself from the cfs_rq (task must be off the queue now).
3855 static void remove_entity_load_avg(struct sched_entity
*se
)
3857 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3858 unsigned long flags
;
3861 * tasks cannot exit without having gone through wake_up_new_task() ->
3862 * post_init_entity_util_avg() which will have added things to the
3863 * cfs_rq, so we can remove unconditionally.
3866 sync_entity_load_avg(se
);
3868 raw_spin_lock_irqsave(&cfs_rq
->removed
.lock
, flags
);
3869 ++cfs_rq
->removed
.nr
;
3870 cfs_rq
->removed
.util_avg
+= se
->avg
.util_avg
;
3871 cfs_rq
->removed
.load_avg
+= se
->avg
.load_avg
;
3872 cfs_rq
->removed
.runnable_avg
+= se
->avg
.runnable_avg
;
3873 raw_spin_unlock_irqrestore(&cfs_rq
->removed
.lock
, flags
);
3876 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq
*cfs_rq
)
3878 return cfs_rq
->avg
.runnable_avg
;
3881 static inline unsigned long cfs_rq_load_avg(struct cfs_rq
*cfs_rq
)
3883 return cfs_rq
->avg
.load_avg
;
3886 static int newidle_balance(struct rq
*this_rq
, struct rq_flags
*rf
);
3888 static inline unsigned long task_util(struct task_struct
*p
)
3890 return READ_ONCE(p
->se
.avg
.util_avg
);
3893 static inline unsigned long _task_util_est(struct task_struct
*p
)
3895 struct util_est ue
= READ_ONCE(p
->se
.avg
.util_est
);
3897 return (max(ue
.ewma
, ue
.enqueued
) | UTIL_AVG_UNCHANGED
);
3900 static inline unsigned long task_util_est(struct task_struct
*p
)
3902 return max(task_util(p
), _task_util_est(p
));
3905 #ifdef CONFIG_UCLAMP_TASK
3906 static inline unsigned long uclamp_task_util(struct task_struct
*p
)
3908 return clamp(task_util_est(p
),
3909 uclamp_eff_value(p
, UCLAMP_MIN
),
3910 uclamp_eff_value(p
, UCLAMP_MAX
));
3913 static inline unsigned long uclamp_task_util(struct task_struct
*p
)
3915 return task_util_est(p
);
3919 static inline void util_est_enqueue(struct cfs_rq
*cfs_rq
,
3920 struct task_struct
*p
)
3922 unsigned int enqueued
;
3924 if (!sched_feat(UTIL_EST
))
3927 /* Update root cfs_rq's estimated utilization */
3928 enqueued
= cfs_rq
->avg
.util_est
.enqueued
;
3929 enqueued
+= _task_util_est(p
);
3930 WRITE_ONCE(cfs_rq
->avg
.util_est
.enqueued
, enqueued
);
3932 trace_sched_util_est_cfs_tp(cfs_rq
);
3936 * Check if a (signed) value is within a specified (unsigned) margin,
3937 * based on the observation that:
3939 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3941 * NOTE: this only works when value + maring < INT_MAX.
3943 static inline bool within_margin(int value
, int margin
)
3945 return ((unsigned int)(value
+ margin
- 1) < (2 * margin
- 1));
3949 util_est_dequeue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
, bool task_sleep
)
3951 long last_ewma_diff
;
3955 if (!sched_feat(UTIL_EST
))
3958 /* Update root cfs_rq's estimated utilization */
3959 ue
.enqueued
= cfs_rq
->avg
.util_est
.enqueued
;
3960 ue
.enqueued
-= min_t(unsigned int, ue
.enqueued
, _task_util_est(p
));
3961 WRITE_ONCE(cfs_rq
->avg
.util_est
.enqueued
, ue
.enqueued
);
3963 trace_sched_util_est_cfs_tp(cfs_rq
);
3966 * Skip update of task's estimated utilization when the task has not
3967 * yet completed an activation, e.g. being migrated.
3973 * If the PELT values haven't changed since enqueue time,
3974 * skip the util_est update.
3976 ue
= p
->se
.avg
.util_est
;
3977 if (ue
.enqueued
& UTIL_AVG_UNCHANGED
)
3981 * Reset EWMA on utilization increases, the moving average is used only
3982 * to smooth utilization decreases.
3984 ue
.enqueued
= (task_util(p
) | UTIL_AVG_UNCHANGED
);
3985 if (sched_feat(UTIL_EST_FASTUP
)) {
3986 if (ue
.ewma
< ue
.enqueued
) {
3987 ue
.ewma
= ue
.enqueued
;
3993 * Skip update of task's estimated utilization when its EWMA is
3994 * already ~1% close to its last activation value.
3996 last_ewma_diff
= ue
.enqueued
- ue
.ewma
;
3997 if (within_margin(last_ewma_diff
, (SCHED_CAPACITY_SCALE
/ 100)))
4001 * To avoid overestimation of actual task utilization, skip updates if
4002 * we cannot grant there is idle time in this CPU.
4004 cpu
= cpu_of(rq_of(cfs_rq
));
4005 if (task_util(p
) > capacity_orig_of(cpu
))
4009 * Update Task's estimated utilization
4011 * When *p completes an activation we can consolidate another sample
4012 * of the task size. This is done by storing the current PELT value
4013 * as ue.enqueued and by using this value to update the Exponential
4014 * Weighted Moving Average (EWMA):
4016 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4017 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4018 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4019 * = w * ( last_ewma_diff ) + ewma(t-1)
4020 * = w * (last_ewma_diff + ewma(t-1) / w)
4022 * Where 'w' is the weight of new samples, which is configured to be
4023 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4025 ue
.ewma
<<= UTIL_EST_WEIGHT_SHIFT
;
4026 ue
.ewma
+= last_ewma_diff
;
4027 ue
.ewma
>>= UTIL_EST_WEIGHT_SHIFT
;
4029 WRITE_ONCE(p
->se
.avg
.util_est
, ue
);
4031 trace_sched_util_est_se_tp(&p
->se
);
4034 static inline int task_fits_capacity(struct task_struct
*p
, long capacity
)
4036 return fits_capacity(uclamp_task_util(p
), capacity
);
4039 static inline void update_misfit_status(struct task_struct
*p
, struct rq
*rq
)
4041 if (!static_branch_unlikely(&sched_asym_cpucapacity
))
4045 rq
->misfit_task_load
= 0;
4049 if (task_fits_capacity(p
, capacity_of(cpu_of(rq
)))) {
4050 rq
->misfit_task_load
= 0;
4055 * Make sure that misfit_task_load will not be null even if
4056 * task_h_load() returns 0.
4058 rq
->misfit_task_load
= max_t(unsigned long, task_h_load(p
), 1);
4061 #else /* CONFIG_SMP */
4063 #define UPDATE_TG 0x0
4064 #define SKIP_AGE_LOAD 0x0
4065 #define DO_ATTACH 0x0
4067 static inline void update_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int not_used1
)
4069 cfs_rq_util_change(cfs_rq
, 0);
4072 static inline void remove_entity_load_avg(struct sched_entity
*se
) {}
4075 attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
4077 detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
4079 static inline int newidle_balance(struct rq
*rq
, struct rq_flags
*rf
)
4085 util_est_enqueue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
) {}
4088 util_est_dequeue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
,
4090 static inline void update_misfit_status(struct task_struct
*p
, struct rq
*rq
) {}
4092 #endif /* CONFIG_SMP */
4094 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4096 #ifdef CONFIG_SCHED_DEBUG
4097 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
4102 if (d
> 3*sysctl_sched_latency
)
4103 schedstat_inc(cfs_rq
->nr_spread_over
);
4108 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
4110 u64 vruntime
= cfs_rq
->min_vruntime
;
4113 * The 'current' period is already promised to the current tasks,
4114 * however the extra weight of the new task will slow them down a
4115 * little, place the new task so that it fits in the slot that
4116 * stays open at the end.
4118 if (initial
&& sched_feat(START_DEBIT
))
4119 vruntime
+= sched_vslice(cfs_rq
, se
);
4121 /* sleeps up to a single latency don't count. */
4123 unsigned long thresh
= sysctl_sched_latency
;
4126 * Halve their sleep time's effect, to allow
4127 * for a gentler effect of sleepers:
4129 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
4135 /* ensure we never gain time by being placed backwards. */
4136 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
4139 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
4141 static inline void check_schedstat_required(void)
4143 #ifdef CONFIG_SCHEDSTATS
4144 if (schedstat_enabled())
4147 /* Force schedstat enabled if a dependent tracepoint is active */
4148 if (trace_sched_stat_wait_enabled() ||
4149 trace_sched_stat_sleep_enabled() ||
4150 trace_sched_stat_iowait_enabled() ||
4151 trace_sched_stat_blocked_enabled() ||
4152 trace_sched_stat_runtime_enabled()) {
4153 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4154 "stat_blocked and stat_runtime require the "
4155 "kernel parameter schedstats=enable or "
4156 "kernel.sched_schedstats=1\n");
4161 static inline bool cfs_bandwidth_used(void);
4168 * update_min_vruntime()
4169 * vruntime -= min_vruntime
4173 * update_min_vruntime()
4174 * vruntime += min_vruntime
4176 * this way the vruntime transition between RQs is done when both
4177 * min_vruntime are up-to-date.
4181 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4182 * vruntime -= min_vruntime
4186 * update_min_vruntime()
4187 * vruntime += min_vruntime
4189 * this way we don't have the most up-to-date min_vruntime on the originating
4190 * CPU and an up-to-date min_vruntime on the destination CPU.
4194 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
4196 bool renorm
= !(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_MIGRATED
);
4197 bool curr
= cfs_rq
->curr
== se
;
4200 * If we're the current task, we must renormalise before calling
4204 se
->vruntime
+= cfs_rq
->min_vruntime
;
4206 update_curr(cfs_rq
);
4209 * Otherwise, renormalise after, such that we're placed at the current
4210 * moment in time, instead of some random moment in the past. Being
4211 * placed in the past could significantly boost this task to the
4212 * fairness detriment of existing tasks.
4214 if (renorm
&& !curr
)
4215 se
->vruntime
+= cfs_rq
->min_vruntime
;
4218 * When enqueuing a sched_entity, we must:
4219 * - Update loads to have both entity and cfs_rq synced with now.
4220 * - Add its load to cfs_rq->runnable_avg
4221 * - For group_entity, update its weight to reflect the new share of
4223 * - Add its new weight to cfs_rq->load.weight
4225 update_load_avg(cfs_rq
, se
, UPDATE_TG
| DO_ATTACH
);
4226 se_update_runnable(se
);
4227 update_cfs_group(se
);
4228 account_entity_enqueue(cfs_rq
, se
);
4230 if (flags
& ENQUEUE_WAKEUP
)
4231 place_entity(cfs_rq
, se
, 0);
4233 check_schedstat_required();
4234 update_stats_enqueue(cfs_rq
, se
, flags
);
4235 check_spread(cfs_rq
, se
);
4237 __enqueue_entity(cfs_rq
, se
);
4241 * When bandwidth control is enabled, cfs might have been removed
4242 * because of a parent been throttled but cfs->nr_running > 1. Try to
4243 * add it unconditionnally.
4245 if (cfs_rq
->nr_running
== 1 || cfs_bandwidth_used())
4246 list_add_leaf_cfs_rq(cfs_rq
);
4248 if (cfs_rq
->nr_running
== 1)
4249 check_enqueue_throttle(cfs_rq
);
4252 static void __clear_buddies_last(struct sched_entity
*se
)
4254 for_each_sched_entity(se
) {
4255 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4256 if (cfs_rq
->last
!= se
)
4259 cfs_rq
->last
= NULL
;
4263 static void __clear_buddies_next(struct sched_entity
*se
)
4265 for_each_sched_entity(se
) {
4266 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4267 if (cfs_rq
->next
!= se
)
4270 cfs_rq
->next
= NULL
;
4274 static void __clear_buddies_skip(struct sched_entity
*se
)
4276 for_each_sched_entity(se
) {
4277 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4278 if (cfs_rq
->skip
!= se
)
4281 cfs_rq
->skip
= NULL
;
4285 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4287 if (cfs_rq
->last
== se
)
4288 __clear_buddies_last(se
);
4290 if (cfs_rq
->next
== se
)
4291 __clear_buddies_next(se
);
4293 if (cfs_rq
->skip
== se
)
4294 __clear_buddies_skip(se
);
4297 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
4300 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
4303 * Update run-time statistics of the 'current'.
4305 update_curr(cfs_rq
);
4308 * When dequeuing a sched_entity, we must:
4309 * - Update loads to have both entity and cfs_rq synced with now.
4310 * - Subtract its load from the cfs_rq->runnable_avg.
4311 * - Subtract its previous weight from cfs_rq->load.weight.
4312 * - For group entity, update its weight to reflect the new share
4313 * of its group cfs_rq.
4315 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4316 se_update_runnable(se
);
4318 update_stats_dequeue(cfs_rq
, se
, flags
);
4320 clear_buddies(cfs_rq
, se
);
4322 if (se
!= cfs_rq
->curr
)
4323 __dequeue_entity(cfs_rq
, se
);
4325 account_entity_dequeue(cfs_rq
, se
);
4328 * Normalize after update_curr(); which will also have moved
4329 * min_vruntime if @se is the one holding it back. But before doing
4330 * update_min_vruntime() again, which will discount @se's position and
4331 * can move min_vruntime forward still more.
4333 if (!(flags
& DEQUEUE_SLEEP
))
4334 se
->vruntime
-= cfs_rq
->min_vruntime
;
4336 /* return excess runtime on last dequeue */
4337 return_cfs_rq_runtime(cfs_rq
);
4339 update_cfs_group(se
);
4342 * Now advance min_vruntime if @se was the entity holding it back,
4343 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4344 * put back on, and if we advance min_vruntime, we'll be placed back
4345 * further than we started -- ie. we'll be penalized.
4347 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) != DEQUEUE_SAVE
)
4348 update_min_vruntime(cfs_rq
);
4352 * Preempt the current task with a newly woken task if needed:
4355 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
4357 unsigned long ideal_runtime
, delta_exec
;
4358 struct sched_entity
*se
;
4361 ideal_runtime
= sched_slice(cfs_rq
, curr
);
4362 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
4363 if (delta_exec
> ideal_runtime
) {
4364 resched_curr(rq_of(cfs_rq
));
4366 * The current task ran long enough, ensure it doesn't get
4367 * re-elected due to buddy favours.
4369 clear_buddies(cfs_rq
, curr
);
4374 * Ensure that a task that missed wakeup preemption by a
4375 * narrow margin doesn't have to wait for a full slice.
4376 * This also mitigates buddy induced latencies under load.
4378 if (delta_exec
< sysctl_sched_min_granularity
)
4381 se
= __pick_first_entity(cfs_rq
);
4382 delta
= curr
->vruntime
- se
->vruntime
;
4387 if (delta
> ideal_runtime
)
4388 resched_curr(rq_of(cfs_rq
));
4392 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4394 /* 'current' is not kept within the tree. */
4397 * Any task has to be enqueued before it get to execute on
4398 * a CPU. So account for the time it spent waiting on the
4401 update_stats_wait_end(cfs_rq
, se
);
4402 __dequeue_entity(cfs_rq
, se
);
4403 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4406 update_stats_curr_start(cfs_rq
, se
);
4410 * Track our maximum slice length, if the CPU's load is at
4411 * least twice that of our own weight (i.e. dont track it
4412 * when there are only lesser-weight tasks around):
4414 if (schedstat_enabled() &&
4415 rq_of(cfs_rq
)->cfs
.load
.weight
>= 2*se
->load
.weight
) {
4416 schedstat_set(se
->statistics
.slice_max
,
4417 max((u64
)schedstat_val(se
->statistics
.slice_max
),
4418 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
));
4421 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
4425 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
4428 * Pick the next process, keeping these things in mind, in this order:
4429 * 1) keep things fair between processes/task groups
4430 * 2) pick the "next" process, since someone really wants that to run
4431 * 3) pick the "last" process, for cache locality
4432 * 4) do not run the "skip" process, if something else is available
4434 static struct sched_entity
*
4435 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
4437 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
4438 struct sched_entity
*se
;
4441 * If curr is set we have to see if its left of the leftmost entity
4442 * still in the tree, provided there was anything in the tree at all.
4444 if (!left
|| (curr
&& entity_before(curr
, left
)))
4447 se
= left
; /* ideally we run the leftmost entity */
4450 * Avoid running the skip buddy, if running something else can
4451 * be done without getting too unfair.
4453 if (cfs_rq
->skip
== se
) {
4454 struct sched_entity
*second
;
4457 second
= __pick_first_entity(cfs_rq
);
4459 second
= __pick_next_entity(se
);
4460 if (!second
|| (curr
&& entity_before(curr
, second
)))
4464 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
4468 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1) {
4470 * Someone really wants this to run. If it's not unfair, run it.
4473 } else if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1) {
4475 * Prefer last buddy, try to return the CPU to a preempted task.
4480 clear_buddies(cfs_rq
, se
);
4485 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
4487 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
4490 * If still on the runqueue then deactivate_task()
4491 * was not called and update_curr() has to be done:
4494 update_curr(cfs_rq
);
4496 /* throttle cfs_rqs exceeding runtime */
4497 check_cfs_rq_runtime(cfs_rq
);
4499 check_spread(cfs_rq
, prev
);
4502 update_stats_wait_start(cfs_rq
, prev
);
4503 /* Put 'current' back into the tree. */
4504 __enqueue_entity(cfs_rq
, prev
);
4505 /* in !on_rq case, update occurred at dequeue */
4506 update_load_avg(cfs_rq
, prev
, 0);
4508 cfs_rq
->curr
= NULL
;
4512 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
4515 * Update run-time statistics of the 'current'.
4517 update_curr(cfs_rq
);
4520 * Ensure that runnable average is periodically updated.
4522 update_load_avg(cfs_rq
, curr
, UPDATE_TG
);
4523 update_cfs_group(curr
);
4525 #ifdef CONFIG_SCHED_HRTICK
4527 * queued ticks are scheduled to match the slice, so don't bother
4528 * validating it and just reschedule.
4531 resched_curr(rq_of(cfs_rq
));
4535 * don't let the period tick interfere with the hrtick preemption
4537 if (!sched_feat(DOUBLE_TICK
) &&
4538 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
4542 if (cfs_rq
->nr_running
> 1)
4543 check_preempt_tick(cfs_rq
, curr
);
4547 /**************************************************
4548 * CFS bandwidth control machinery
4551 #ifdef CONFIG_CFS_BANDWIDTH
4553 #ifdef CONFIG_JUMP_LABEL
4554 static struct static_key __cfs_bandwidth_used
;
4556 static inline bool cfs_bandwidth_used(void)
4558 return static_key_false(&__cfs_bandwidth_used
);
4561 void cfs_bandwidth_usage_inc(void)
4563 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used
);
4566 void cfs_bandwidth_usage_dec(void)
4568 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used
);
4570 #else /* CONFIG_JUMP_LABEL */
4571 static bool cfs_bandwidth_used(void)
4576 void cfs_bandwidth_usage_inc(void) {}
4577 void cfs_bandwidth_usage_dec(void) {}
4578 #endif /* CONFIG_JUMP_LABEL */
4581 * default period for cfs group bandwidth.
4582 * default: 0.1s, units: nanoseconds
4584 static inline u64
default_cfs_period(void)
4586 return 100000000ULL;
4589 static inline u64
sched_cfs_bandwidth_slice(void)
4591 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
4595 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4596 * directly instead of rq->clock to avoid adding additional synchronization
4599 * requires cfs_b->lock
4601 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
4603 if (cfs_b
->quota
!= RUNTIME_INF
)
4604 cfs_b
->runtime
= cfs_b
->quota
;
4607 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4609 return &tg
->cfs_bandwidth
;
4612 /* returns 0 on failure to allocate runtime */
4613 static int __assign_cfs_rq_runtime(struct cfs_bandwidth
*cfs_b
,
4614 struct cfs_rq
*cfs_rq
, u64 target_runtime
)
4616 u64 min_amount
, amount
= 0;
4618 lockdep_assert_held(&cfs_b
->lock
);
4620 /* note: this is a positive sum as runtime_remaining <= 0 */
4621 min_amount
= target_runtime
- cfs_rq
->runtime_remaining
;
4623 if (cfs_b
->quota
== RUNTIME_INF
)
4624 amount
= min_amount
;
4626 start_cfs_bandwidth(cfs_b
);
4628 if (cfs_b
->runtime
> 0) {
4629 amount
= min(cfs_b
->runtime
, min_amount
);
4630 cfs_b
->runtime
-= amount
;
4635 cfs_rq
->runtime_remaining
+= amount
;
4637 return cfs_rq
->runtime_remaining
> 0;
4640 /* returns 0 on failure to allocate runtime */
4641 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4643 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4646 raw_spin_lock(&cfs_b
->lock
);
4647 ret
= __assign_cfs_rq_runtime(cfs_b
, cfs_rq
, sched_cfs_bandwidth_slice());
4648 raw_spin_unlock(&cfs_b
->lock
);
4653 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4655 /* dock delta_exec before expiring quota (as it could span periods) */
4656 cfs_rq
->runtime_remaining
-= delta_exec
;
4658 if (likely(cfs_rq
->runtime_remaining
> 0))
4661 if (cfs_rq
->throttled
)
4664 * if we're unable to extend our runtime we resched so that the active
4665 * hierarchy can be throttled
4667 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
4668 resched_curr(rq_of(cfs_rq
));
4671 static __always_inline
4672 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4674 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
4677 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
4680 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4682 return cfs_bandwidth_used() && cfs_rq
->throttled
;
4685 /* check whether cfs_rq, or any parent, is throttled */
4686 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4688 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
4692 * Ensure that neither of the group entities corresponding to src_cpu or
4693 * dest_cpu are members of a throttled hierarchy when performing group
4694 * load-balance operations.
4696 static inline int throttled_lb_pair(struct task_group
*tg
,
4697 int src_cpu
, int dest_cpu
)
4699 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
4701 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
4702 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
4704 return throttled_hierarchy(src_cfs_rq
) ||
4705 throttled_hierarchy(dest_cfs_rq
);
4708 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
4710 struct rq
*rq
= data
;
4711 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4713 cfs_rq
->throttle_count
--;
4714 if (!cfs_rq
->throttle_count
) {
4715 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
4716 cfs_rq
->throttled_clock_task
;
4718 /* Add cfs_rq with already running entity in the list */
4719 if (cfs_rq
->nr_running
>= 1)
4720 list_add_leaf_cfs_rq(cfs_rq
);
4726 static int tg_throttle_down(struct task_group
*tg
, void *data
)
4728 struct rq
*rq
= data
;
4729 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4731 /* group is entering throttled state, stop time */
4732 if (!cfs_rq
->throttle_count
) {
4733 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
4734 list_del_leaf_cfs_rq(cfs_rq
);
4736 cfs_rq
->throttle_count
++;
4741 static bool throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
4743 struct rq
*rq
= rq_of(cfs_rq
);
4744 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4745 struct sched_entity
*se
;
4746 long task_delta
, idle_task_delta
, dequeue
= 1;
4748 raw_spin_lock(&cfs_b
->lock
);
4749 /* This will start the period timer if necessary */
4750 if (__assign_cfs_rq_runtime(cfs_b
, cfs_rq
, 1)) {
4752 * We have raced with bandwidth becoming available, and if we
4753 * actually throttled the timer might not unthrottle us for an
4754 * entire period. We additionally needed to make sure that any
4755 * subsequent check_cfs_rq_runtime calls agree not to throttle
4756 * us, as we may commit to do cfs put_prev+pick_next, so we ask
4757 * for 1ns of runtime rather than just check cfs_b.
4761 list_add_tail_rcu(&cfs_rq
->throttled_list
,
4762 &cfs_b
->throttled_cfs_rq
);
4764 raw_spin_unlock(&cfs_b
->lock
);
4767 return false; /* Throttle no longer required. */
4769 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
4771 /* freeze hierarchy runnable averages while throttled */
4773 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
4776 task_delta
= cfs_rq
->h_nr_running
;
4777 idle_task_delta
= cfs_rq
->idle_h_nr_running
;
4778 for_each_sched_entity(se
) {
4779 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4780 /* throttled entity or throttle-on-deactivate */
4785 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
4787 update_load_avg(qcfs_rq
, se
, 0);
4788 se_update_runnable(se
);
4791 qcfs_rq
->h_nr_running
-= task_delta
;
4792 qcfs_rq
->idle_h_nr_running
-= idle_task_delta
;
4794 if (qcfs_rq
->load
.weight
)
4799 sub_nr_running(rq
, task_delta
);
4802 * Note: distribution will already see us throttled via the
4803 * throttled-list. rq->lock protects completion.
4805 cfs_rq
->throttled
= 1;
4806 cfs_rq
->throttled_clock
= rq_clock(rq
);
4810 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
4812 struct rq
*rq
= rq_of(cfs_rq
);
4813 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4814 struct sched_entity
*se
;
4815 long task_delta
, idle_task_delta
;
4817 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
4819 cfs_rq
->throttled
= 0;
4821 update_rq_clock(rq
);
4823 raw_spin_lock(&cfs_b
->lock
);
4824 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
4825 list_del_rcu(&cfs_rq
->throttled_list
);
4826 raw_spin_unlock(&cfs_b
->lock
);
4828 /* update hierarchical throttle state */
4829 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
4831 if (!cfs_rq
->load
.weight
)
4834 task_delta
= cfs_rq
->h_nr_running
;
4835 idle_task_delta
= cfs_rq
->idle_h_nr_running
;
4836 for_each_sched_entity(se
) {
4839 cfs_rq
= cfs_rq_of(se
);
4840 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
4842 cfs_rq
->h_nr_running
+= task_delta
;
4843 cfs_rq
->idle_h_nr_running
+= idle_task_delta
;
4845 /* end evaluation on encountering a throttled cfs_rq */
4846 if (cfs_rq_throttled(cfs_rq
))
4847 goto unthrottle_throttle
;
4850 for_each_sched_entity(se
) {
4851 cfs_rq
= cfs_rq_of(se
);
4853 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4854 se_update_runnable(se
);
4856 cfs_rq
->h_nr_running
+= task_delta
;
4857 cfs_rq
->idle_h_nr_running
+= idle_task_delta
;
4860 /* end evaluation on encountering a throttled cfs_rq */
4861 if (cfs_rq_throttled(cfs_rq
))
4862 goto unthrottle_throttle
;
4865 * One parent has been throttled and cfs_rq removed from the
4866 * list. Add it back to not break the leaf list.
4868 if (throttled_hierarchy(cfs_rq
))
4869 list_add_leaf_cfs_rq(cfs_rq
);
4872 /* At this point se is NULL and we are at root level*/
4873 add_nr_running(rq
, task_delta
);
4875 unthrottle_throttle
:
4877 * The cfs_rq_throttled() breaks in the above iteration can result in
4878 * incomplete leaf list maintenance, resulting in triggering the
4881 for_each_sched_entity(se
) {
4882 cfs_rq
= cfs_rq_of(se
);
4884 if (list_add_leaf_cfs_rq(cfs_rq
))
4888 assert_list_leaf_cfs_rq(rq
);
4890 /* Determine whether we need to wake up potentially idle CPU: */
4891 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
4895 static void distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
)
4897 struct cfs_rq
*cfs_rq
;
4898 u64 runtime
, remaining
= 1;
4901 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
4903 struct rq
*rq
= rq_of(cfs_rq
);
4906 rq_lock_irqsave(rq
, &rf
);
4907 if (!cfs_rq_throttled(cfs_rq
))
4910 /* By the above check, this should never be true */
4911 SCHED_WARN_ON(cfs_rq
->runtime_remaining
> 0);
4913 raw_spin_lock(&cfs_b
->lock
);
4914 runtime
= -cfs_rq
->runtime_remaining
+ 1;
4915 if (runtime
> cfs_b
->runtime
)
4916 runtime
= cfs_b
->runtime
;
4917 cfs_b
->runtime
-= runtime
;
4918 remaining
= cfs_b
->runtime
;
4919 raw_spin_unlock(&cfs_b
->lock
);
4921 cfs_rq
->runtime_remaining
+= runtime
;
4923 /* we check whether we're throttled above */
4924 if (cfs_rq
->runtime_remaining
> 0)
4925 unthrottle_cfs_rq(cfs_rq
);
4928 rq_unlock_irqrestore(rq
, &rf
);
4937 * Responsible for refilling a task_group's bandwidth and unthrottling its
4938 * cfs_rqs as appropriate. If there has been no activity within the last
4939 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4940 * used to track this state.
4942 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
, unsigned long flags
)
4946 /* no need to continue the timer with no bandwidth constraint */
4947 if (cfs_b
->quota
== RUNTIME_INF
)
4948 goto out_deactivate
;
4950 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4951 cfs_b
->nr_periods
+= overrun
;
4954 * idle depends on !throttled (for the case of a large deficit), and if
4955 * we're going inactive then everything else can be deferred
4957 if (cfs_b
->idle
&& !throttled
)
4958 goto out_deactivate
;
4960 __refill_cfs_bandwidth_runtime(cfs_b
);
4963 /* mark as potentially idle for the upcoming period */
4968 /* account preceding periods in which throttling occurred */
4969 cfs_b
->nr_throttled
+= overrun
;
4972 * This check is repeated as we release cfs_b->lock while we unthrottle.
4974 while (throttled
&& cfs_b
->runtime
> 0) {
4975 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
4976 /* we can't nest cfs_b->lock while distributing bandwidth */
4977 distribute_cfs_runtime(cfs_b
);
4978 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
4980 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4984 * While we are ensured activity in the period following an
4985 * unthrottle, this also covers the case in which the new bandwidth is
4986 * insufficient to cover the existing bandwidth deficit. (Forcing the
4987 * timer to remain active while there are any throttled entities.)
4997 /* a cfs_rq won't donate quota below this amount */
4998 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
4999 /* minimum remaining period time to redistribute slack quota */
5000 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
5001 /* how long we wait to gather additional slack before distributing */
5002 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
5005 * Are we near the end of the current quota period?
5007 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5008 * hrtimer base being cleared by hrtimer_start. In the case of
5009 * migrate_hrtimers, base is never cleared, so we are fine.
5011 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
5013 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
5016 /* if the call-back is running a quota refresh is already occurring */
5017 if (hrtimer_callback_running(refresh_timer
))
5020 /* is a quota refresh about to occur? */
5021 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
5022 if (remaining
< min_expire
)
5028 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
5030 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
5032 /* if there's a quota refresh soon don't bother with slack */
5033 if (runtime_refresh_within(cfs_b
, min_left
))
5036 /* don't push forwards an existing deferred unthrottle */
5037 if (cfs_b
->slack_started
)
5039 cfs_b
->slack_started
= true;
5041 hrtimer_start(&cfs_b
->slack_timer
,
5042 ns_to_ktime(cfs_bandwidth_slack_period
),
5046 /* we know any runtime found here is valid as update_curr() precedes return */
5047 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5049 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
5050 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
5052 if (slack_runtime
<= 0)
5055 raw_spin_lock(&cfs_b
->lock
);
5056 if (cfs_b
->quota
!= RUNTIME_INF
) {
5057 cfs_b
->runtime
+= slack_runtime
;
5059 /* we are under rq->lock, defer unthrottling using a timer */
5060 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
5061 !list_empty(&cfs_b
->throttled_cfs_rq
))
5062 start_cfs_slack_bandwidth(cfs_b
);
5064 raw_spin_unlock(&cfs_b
->lock
);
5066 /* even if it's not valid for return we don't want to try again */
5067 cfs_rq
->runtime_remaining
-= slack_runtime
;
5070 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5072 if (!cfs_bandwidth_used())
5075 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
5078 __return_cfs_rq_runtime(cfs_rq
);
5082 * This is done with a timer (instead of inline with bandwidth return) since
5083 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5085 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
5087 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
5088 unsigned long flags
;
5090 /* confirm we're still not at a refresh boundary */
5091 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5092 cfs_b
->slack_started
= false;
5094 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
5095 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5099 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
5100 runtime
= cfs_b
->runtime
;
5102 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5107 distribute_cfs_runtime(cfs_b
);
5109 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5110 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5114 * When a group wakes up we want to make sure that its quota is not already
5115 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5116 * runtime as update_curr() throttling can not not trigger until it's on-rq.
5118 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
5120 if (!cfs_bandwidth_used())
5123 /* an active group must be handled by the update_curr()->put() path */
5124 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
5127 /* ensure the group is not already throttled */
5128 if (cfs_rq_throttled(cfs_rq
))
5131 /* update runtime allocation */
5132 account_cfs_rq_runtime(cfs_rq
, 0);
5133 if (cfs_rq
->runtime_remaining
<= 0)
5134 throttle_cfs_rq(cfs_rq
);
5137 static void sync_throttle(struct task_group
*tg
, int cpu
)
5139 struct cfs_rq
*pcfs_rq
, *cfs_rq
;
5141 if (!cfs_bandwidth_used())
5147 cfs_rq
= tg
->cfs_rq
[cpu
];
5148 pcfs_rq
= tg
->parent
->cfs_rq
[cpu
];
5150 cfs_rq
->throttle_count
= pcfs_rq
->throttle_count
;
5151 cfs_rq
->throttled_clock_task
= rq_clock_task(cpu_rq(cpu
));
5154 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5155 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5157 if (!cfs_bandwidth_used())
5160 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
5164 * it's possible for a throttled entity to be forced into a running
5165 * state (e.g. set_curr_task), in this case we're finished.
5167 if (cfs_rq_throttled(cfs_rq
))
5170 return throttle_cfs_rq(cfs_rq
);
5173 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
5175 struct cfs_bandwidth
*cfs_b
=
5176 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
5178 do_sched_cfs_slack_timer(cfs_b
);
5180 return HRTIMER_NORESTART
;
5183 extern const u64 max_cfs_quota_period
;
5185 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
5187 struct cfs_bandwidth
*cfs_b
=
5188 container_of(timer
, struct cfs_bandwidth
, period_timer
);
5189 unsigned long flags
;
5194 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
5196 overrun
= hrtimer_forward_now(timer
, cfs_b
->period
);
5200 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
, flags
);
5203 u64
new, old
= ktime_to_ns(cfs_b
->period
);
5206 * Grow period by a factor of 2 to avoid losing precision.
5207 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5211 if (new < max_cfs_quota_period
) {
5212 cfs_b
->period
= ns_to_ktime(new);
5215 pr_warn_ratelimited(
5216 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5218 div_u64(new, NSEC_PER_USEC
),
5219 div_u64(cfs_b
->quota
, NSEC_PER_USEC
));
5221 pr_warn_ratelimited(
5222 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5224 div_u64(old
, NSEC_PER_USEC
),
5225 div_u64(cfs_b
->quota
, NSEC_PER_USEC
));
5228 /* reset count so we don't come right back in here */
5233 cfs_b
->period_active
= 0;
5234 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
5236 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
5239 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5241 raw_spin_lock_init(&cfs_b
->lock
);
5243 cfs_b
->quota
= RUNTIME_INF
;
5244 cfs_b
->period
= ns_to_ktime(default_cfs_period());
5246 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
5247 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_PINNED
);
5248 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
5249 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5250 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
5251 cfs_b
->slack_started
= false;
5254 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5256 cfs_rq
->runtime_enabled
= 0;
5257 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
5260 void start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5262 lockdep_assert_held(&cfs_b
->lock
);
5264 if (cfs_b
->period_active
)
5267 cfs_b
->period_active
= 1;
5268 hrtimer_forward_now(&cfs_b
->period_timer
, cfs_b
->period
);
5269 hrtimer_start_expires(&cfs_b
->period_timer
, HRTIMER_MODE_ABS_PINNED
);
5272 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5274 /* init_cfs_bandwidth() was not called */
5275 if (!cfs_b
->throttled_cfs_rq
.next
)
5278 hrtimer_cancel(&cfs_b
->period_timer
);
5279 hrtimer_cancel(&cfs_b
->slack_timer
);
5283 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5285 * The race is harmless, since modifying bandwidth settings of unhooked group
5286 * bits doesn't do much.
5289 /* cpu online calback */
5290 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
5292 struct task_group
*tg
;
5294 lockdep_assert_held(&rq
->lock
);
5297 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
5298 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
5299 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
5301 raw_spin_lock(&cfs_b
->lock
);
5302 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
5303 raw_spin_unlock(&cfs_b
->lock
);
5308 /* cpu offline callback */
5309 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
5311 struct task_group
*tg
;
5313 lockdep_assert_held(&rq
->lock
);
5316 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
5317 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
5319 if (!cfs_rq
->runtime_enabled
)
5323 * clock_task is not advancing so we just need to make sure
5324 * there's some valid quota amount
5326 cfs_rq
->runtime_remaining
= 1;
5328 * Offline rq is schedulable till CPU is completely disabled
5329 * in take_cpu_down(), so we prevent new cfs throttling here.
5331 cfs_rq
->runtime_enabled
= 0;
5333 if (cfs_rq_throttled(cfs_rq
))
5334 unthrottle_cfs_rq(cfs_rq
);
5339 #else /* CONFIG_CFS_BANDWIDTH */
5341 static inline bool cfs_bandwidth_used(void)
5346 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
5347 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
5348 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
5349 static inline void sync_throttle(struct task_group
*tg
, int cpu
) {}
5350 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
5352 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
5357 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
5362 static inline int throttled_lb_pair(struct task_group
*tg
,
5363 int src_cpu
, int dest_cpu
)
5368 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
5370 #ifdef CONFIG_FAIR_GROUP_SCHED
5371 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
5374 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
5378 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
5379 static inline void update_runtime_enabled(struct rq
*rq
) {}
5380 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
5382 #endif /* CONFIG_CFS_BANDWIDTH */
5384 /**************************************************
5385 * CFS operations on tasks:
5388 #ifdef CONFIG_SCHED_HRTICK
5389 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
5391 struct sched_entity
*se
= &p
->se
;
5392 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5394 SCHED_WARN_ON(task_rq(p
) != rq
);
5396 if (rq
->cfs
.h_nr_running
> 1) {
5397 u64 slice
= sched_slice(cfs_rq
, se
);
5398 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
5399 s64 delta
= slice
- ran
;
5406 hrtick_start(rq
, delta
);
5411 * called from enqueue/dequeue and updates the hrtick when the
5412 * current task is from our class and nr_running is low enough
5415 static void hrtick_update(struct rq
*rq
)
5417 struct task_struct
*curr
= rq
->curr
;
5419 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
5422 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
5423 hrtick_start_fair(rq
, curr
);
5425 #else /* !CONFIG_SCHED_HRTICK */
5427 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
5431 static inline void hrtick_update(struct rq
*rq
)
5437 static inline unsigned long cpu_util(int cpu
);
5439 static inline bool cpu_overutilized(int cpu
)
5441 return !fits_capacity(cpu_util(cpu
), capacity_of(cpu
));
5444 static inline void update_overutilized_status(struct rq
*rq
)
5446 if (!READ_ONCE(rq
->rd
->overutilized
) && cpu_overutilized(rq
->cpu
)) {
5447 WRITE_ONCE(rq
->rd
->overutilized
, SG_OVERUTILIZED
);
5448 trace_sched_overutilized_tp(rq
->rd
, SG_OVERUTILIZED
);
5452 static inline void update_overutilized_status(struct rq
*rq
) { }
5455 /* Runqueue only has SCHED_IDLE tasks enqueued */
5456 static int sched_idle_rq(struct rq
*rq
)
5458 return unlikely(rq
->nr_running
== rq
->cfs
.idle_h_nr_running
&&
5463 static int sched_idle_cpu(int cpu
)
5465 return sched_idle_rq(cpu_rq(cpu
));
5470 * The enqueue_task method is called before nr_running is
5471 * increased. Here we update the fair scheduling stats and
5472 * then put the task into the rbtree:
5475 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
5477 struct cfs_rq
*cfs_rq
;
5478 struct sched_entity
*se
= &p
->se
;
5479 int idle_h_nr_running
= task_has_idle_policy(p
);
5482 * The code below (indirectly) updates schedutil which looks at
5483 * the cfs_rq utilization to select a frequency.
5484 * Let's add the task's estimated utilization to the cfs_rq's
5485 * estimated utilization, before we update schedutil.
5487 util_est_enqueue(&rq
->cfs
, p
);
5490 * If in_iowait is set, the code below may not trigger any cpufreq
5491 * utilization updates, so do it here explicitly with the IOWAIT flag
5495 cpufreq_update_util(rq
, SCHED_CPUFREQ_IOWAIT
);
5497 for_each_sched_entity(se
) {
5500 cfs_rq
= cfs_rq_of(se
);
5501 enqueue_entity(cfs_rq
, se
, flags
);
5503 cfs_rq
->h_nr_running
++;
5504 cfs_rq
->idle_h_nr_running
+= idle_h_nr_running
;
5506 /* end evaluation on encountering a throttled cfs_rq */
5507 if (cfs_rq_throttled(cfs_rq
))
5508 goto enqueue_throttle
;
5510 flags
= ENQUEUE_WAKEUP
;
5513 for_each_sched_entity(se
) {
5514 cfs_rq
= cfs_rq_of(se
);
5516 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
5517 se_update_runnable(se
);
5518 update_cfs_group(se
);
5520 cfs_rq
->h_nr_running
++;
5521 cfs_rq
->idle_h_nr_running
+= idle_h_nr_running
;
5523 /* end evaluation on encountering a throttled cfs_rq */
5524 if (cfs_rq_throttled(cfs_rq
))
5525 goto enqueue_throttle
;
5528 * One parent has been throttled and cfs_rq removed from the
5529 * list. Add it back to not break the leaf list.
5531 if (throttled_hierarchy(cfs_rq
))
5532 list_add_leaf_cfs_rq(cfs_rq
);
5535 /* At this point se is NULL and we are at root level*/
5536 add_nr_running(rq
, 1);
5539 * Since new tasks are assigned an initial util_avg equal to
5540 * half of the spare capacity of their CPU, tiny tasks have the
5541 * ability to cross the overutilized threshold, which will
5542 * result in the load balancer ruining all the task placement
5543 * done by EAS. As a way to mitigate that effect, do not account
5544 * for the first enqueue operation of new tasks during the
5545 * overutilized flag detection.
5547 * A better way of solving this problem would be to wait for
5548 * the PELT signals of tasks to converge before taking them
5549 * into account, but that is not straightforward to implement,
5550 * and the following generally works well enough in practice.
5552 if (flags
& ENQUEUE_WAKEUP
)
5553 update_overutilized_status(rq
);
5556 if (cfs_bandwidth_used()) {
5558 * When bandwidth control is enabled; the cfs_rq_throttled()
5559 * breaks in the above iteration can result in incomplete
5560 * leaf list maintenance, resulting in triggering the assertion
5563 for_each_sched_entity(se
) {
5564 cfs_rq
= cfs_rq_of(se
);
5566 if (list_add_leaf_cfs_rq(cfs_rq
))
5571 assert_list_leaf_cfs_rq(rq
);
5576 static void set_next_buddy(struct sched_entity
*se
);
5579 * The dequeue_task method is called before nr_running is
5580 * decreased. We remove the task from the rbtree and
5581 * update the fair scheduling stats:
5583 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
5585 struct cfs_rq
*cfs_rq
;
5586 struct sched_entity
*se
= &p
->se
;
5587 int task_sleep
= flags
& DEQUEUE_SLEEP
;
5588 int idle_h_nr_running
= task_has_idle_policy(p
);
5589 bool was_sched_idle
= sched_idle_rq(rq
);
5591 for_each_sched_entity(se
) {
5592 cfs_rq
= cfs_rq_of(se
);
5593 dequeue_entity(cfs_rq
, se
, flags
);
5595 cfs_rq
->h_nr_running
--;
5596 cfs_rq
->idle_h_nr_running
-= idle_h_nr_running
;
5598 /* end evaluation on encountering a throttled cfs_rq */
5599 if (cfs_rq_throttled(cfs_rq
))
5600 goto dequeue_throttle
;
5602 /* Don't dequeue parent if it has other entities besides us */
5603 if (cfs_rq
->load
.weight
) {
5604 /* Avoid re-evaluating load for this entity: */
5605 se
= parent_entity(se
);
5607 * Bias pick_next to pick a task from this cfs_rq, as
5608 * p is sleeping when it is within its sched_slice.
5610 if (task_sleep
&& se
&& !throttled_hierarchy(cfs_rq
))
5614 flags
|= DEQUEUE_SLEEP
;
5617 for_each_sched_entity(se
) {
5618 cfs_rq
= cfs_rq_of(se
);
5620 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
5621 se_update_runnable(se
);
5622 update_cfs_group(se
);
5624 cfs_rq
->h_nr_running
--;
5625 cfs_rq
->idle_h_nr_running
-= idle_h_nr_running
;
5627 /* end evaluation on encountering a throttled cfs_rq */
5628 if (cfs_rq_throttled(cfs_rq
))
5629 goto dequeue_throttle
;
5633 /* At this point se is NULL and we are at root level*/
5634 sub_nr_running(rq
, 1);
5636 /* balance early to pull high priority tasks */
5637 if (unlikely(!was_sched_idle
&& sched_idle_rq(rq
)))
5638 rq
->next_balance
= jiffies
;
5641 util_est_dequeue(&rq
->cfs
, p
, task_sleep
);
5647 /* Working cpumask for: load_balance, load_balance_newidle. */
5648 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5649 DEFINE_PER_CPU(cpumask_var_t
, select_idle_mask
);
5651 #ifdef CONFIG_NO_HZ_COMMON
5654 cpumask_var_t idle_cpus_mask
;
5656 int has_blocked
; /* Idle CPUS has blocked load */
5657 unsigned long next_balance
; /* in jiffy units */
5658 unsigned long next_blocked
; /* Next update of blocked load in jiffies */
5659 } nohz ____cacheline_aligned
;
5661 #endif /* CONFIG_NO_HZ_COMMON */
5663 static unsigned long cpu_load(struct rq
*rq
)
5665 return cfs_rq_load_avg(&rq
->cfs
);
5669 * cpu_load_without - compute CPU load without any contributions from *p
5670 * @cpu: the CPU which load is requested
5671 * @p: the task which load should be discounted
5673 * The load of a CPU is defined by the load of tasks currently enqueued on that
5674 * CPU as well as tasks which are currently sleeping after an execution on that
5677 * This method returns the load of the specified CPU by discounting the load of
5678 * the specified task, whenever the task is currently contributing to the CPU
5681 static unsigned long cpu_load_without(struct rq
*rq
, struct task_struct
*p
)
5683 struct cfs_rq
*cfs_rq
;
5686 /* Task has no contribution or is new */
5687 if (cpu_of(rq
) != task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
5688 return cpu_load(rq
);
5691 load
= READ_ONCE(cfs_rq
->avg
.load_avg
);
5693 /* Discount task's util from CPU's util */
5694 lsub_positive(&load
, task_h_load(p
));
5699 static unsigned long cpu_runnable(struct rq
*rq
)
5701 return cfs_rq_runnable_avg(&rq
->cfs
);
5704 static unsigned long cpu_runnable_without(struct rq
*rq
, struct task_struct
*p
)
5706 struct cfs_rq
*cfs_rq
;
5707 unsigned int runnable
;
5709 /* Task has no contribution or is new */
5710 if (cpu_of(rq
) != task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
5711 return cpu_runnable(rq
);
5714 runnable
= READ_ONCE(cfs_rq
->avg
.runnable_avg
);
5716 /* Discount task's runnable from CPU's runnable */
5717 lsub_positive(&runnable
, p
->se
.avg
.runnable_avg
);
5722 static unsigned long capacity_of(int cpu
)
5724 return cpu_rq(cpu
)->cpu_capacity
;
5727 static void record_wakee(struct task_struct
*p
)
5730 * Only decay a single time; tasks that have less then 1 wakeup per
5731 * jiffy will not have built up many flips.
5733 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
5734 current
->wakee_flips
>>= 1;
5735 current
->wakee_flip_decay_ts
= jiffies
;
5738 if (current
->last_wakee
!= p
) {
5739 current
->last_wakee
= p
;
5740 current
->wakee_flips
++;
5745 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5747 * A waker of many should wake a different task than the one last awakened
5748 * at a frequency roughly N times higher than one of its wakees.
5750 * In order to determine whether we should let the load spread vs consolidating
5751 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5752 * partner, and a factor of lls_size higher frequency in the other.
5754 * With both conditions met, we can be relatively sure that the relationship is
5755 * non-monogamous, with partner count exceeding socket size.
5757 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5758 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5761 static int wake_wide(struct task_struct
*p
)
5763 unsigned int master
= current
->wakee_flips
;
5764 unsigned int slave
= p
->wakee_flips
;
5765 int factor
= __this_cpu_read(sd_llc_size
);
5768 swap(master
, slave
);
5769 if (slave
< factor
|| master
< slave
* factor
)
5775 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5776 * soonest. For the purpose of speed we only consider the waking and previous
5779 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5780 * cache-affine and is (or will be) idle.
5782 * wake_affine_weight() - considers the weight to reflect the average
5783 * scheduling latency of the CPUs. This seems to work
5784 * for the overloaded case.
5787 wake_affine_idle(int this_cpu
, int prev_cpu
, int sync
)
5790 * If this_cpu is idle, it implies the wakeup is from interrupt
5791 * context. Only allow the move if cache is shared. Otherwise an
5792 * interrupt intensive workload could force all tasks onto one
5793 * node depending on the IO topology or IRQ affinity settings.
5795 * If the prev_cpu is idle and cache affine then avoid a migration.
5796 * There is no guarantee that the cache hot data from an interrupt
5797 * is more important than cache hot data on the prev_cpu and from
5798 * a cpufreq perspective, it's better to have higher utilisation
5801 if (available_idle_cpu(this_cpu
) && cpus_share_cache(this_cpu
, prev_cpu
))
5802 return available_idle_cpu(prev_cpu
) ? prev_cpu
: this_cpu
;
5804 if (sync
&& cpu_rq(this_cpu
)->nr_running
== 1)
5807 return nr_cpumask_bits
;
5811 wake_affine_weight(struct sched_domain
*sd
, struct task_struct
*p
,
5812 int this_cpu
, int prev_cpu
, int sync
)
5814 s64 this_eff_load
, prev_eff_load
;
5815 unsigned long task_load
;
5817 this_eff_load
= cpu_load(cpu_rq(this_cpu
));
5820 unsigned long current_load
= task_h_load(current
);
5822 if (current_load
> this_eff_load
)
5825 this_eff_load
-= current_load
;
5828 task_load
= task_h_load(p
);
5830 this_eff_load
+= task_load
;
5831 if (sched_feat(WA_BIAS
))
5832 this_eff_load
*= 100;
5833 this_eff_load
*= capacity_of(prev_cpu
);
5835 prev_eff_load
= cpu_load(cpu_rq(prev_cpu
));
5836 prev_eff_load
-= task_load
;
5837 if (sched_feat(WA_BIAS
))
5838 prev_eff_load
*= 100 + (sd
->imbalance_pct
- 100) / 2;
5839 prev_eff_load
*= capacity_of(this_cpu
);
5842 * If sync, adjust the weight of prev_eff_load such that if
5843 * prev_eff == this_eff that select_idle_sibling() will consider
5844 * stacking the wakee on top of the waker if no other CPU is
5850 return this_eff_load
< prev_eff_load
? this_cpu
: nr_cpumask_bits
;
5853 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
,
5854 int this_cpu
, int prev_cpu
, int sync
)
5856 int target
= nr_cpumask_bits
;
5858 if (sched_feat(WA_IDLE
))
5859 target
= wake_affine_idle(this_cpu
, prev_cpu
, sync
);
5861 if (sched_feat(WA_WEIGHT
) && target
== nr_cpumask_bits
)
5862 target
= wake_affine_weight(sd
, p
, this_cpu
, prev_cpu
, sync
);
5864 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine_attempts
);
5865 if (target
== nr_cpumask_bits
)
5868 schedstat_inc(sd
->ttwu_move_affine
);
5869 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine
);
5873 static struct sched_group
*
5874 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
);
5877 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5880 find_idlest_group_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
5882 unsigned long load
, min_load
= ULONG_MAX
;
5883 unsigned int min_exit_latency
= UINT_MAX
;
5884 u64 latest_idle_timestamp
= 0;
5885 int least_loaded_cpu
= this_cpu
;
5886 int shallowest_idle_cpu
= -1;
5889 /* Check if we have any choice: */
5890 if (group
->group_weight
== 1)
5891 return cpumask_first(sched_group_span(group
));
5893 /* Traverse only the allowed CPUs */
5894 for_each_cpu_and(i
, sched_group_span(group
), p
->cpus_ptr
) {
5895 if (sched_idle_cpu(i
))
5898 if (available_idle_cpu(i
)) {
5899 struct rq
*rq
= cpu_rq(i
);
5900 struct cpuidle_state
*idle
= idle_get_state(rq
);
5901 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
5903 * We give priority to a CPU whose idle state
5904 * has the smallest exit latency irrespective
5905 * of any idle timestamp.
5907 min_exit_latency
= idle
->exit_latency
;
5908 latest_idle_timestamp
= rq
->idle_stamp
;
5909 shallowest_idle_cpu
= i
;
5910 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
5911 rq
->idle_stamp
> latest_idle_timestamp
) {
5913 * If equal or no active idle state, then
5914 * the most recently idled CPU might have
5917 latest_idle_timestamp
= rq
->idle_stamp
;
5918 shallowest_idle_cpu
= i
;
5920 } else if (shallowest_idle_cpu
== -1) {
5921 load
= cpu_load(cpu_rq(i
));
5922 if (load
< min_load
) {
5924 least_loaded_cpu
= i
;
5929 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
5932 static inline int find_idlest_cpu(struct sched_domain
*sd
, struct task_struct
*p
,
5933 int cpu
, int prev_cpu
, int sd_flag
)
5937 if (!cpumask_intersects(sched_domain_span(sd
), p
->cpus_ptr
))
5941 * We need task's util for cpu_util_without, sync it up to
5942 * prev_cpu's last_update_time.
5944 if (!(sd_flag
& SD_BALANCE_FORK
))
5945 sync_entity_load_avg(&p
->se
);
5948 struct sched_group
*group
;
5949 struct sched_domain
*tmp
;
5952 if (!(sd
->flags
& sd_flag
)) {
5957 group
= find_idlest_group(sd
, p
, cpu
);
5963 new_cpu
= find_idlest_group_cpu(group
, p
, cpu
);
5964 if (new_cpu
== cpu
) {
5965 /* Now try balancing at a lower domain level of 'cpu': */
5970 /* Now try balancing at a lower domain level of 'new_cpu': */
5972 weight
= sd
->span_weight
;
5974 for_each_domain(cpu
, tmp
) {
5975 if (weight
<= tmp
->span_weight
)
5977 if (tmp
->flags
& sd_flag
)
5985 #ifdef CONFIG_SCHED_SMT
5986 DEFINE_STATIC_KEY_FALSE(sched_smt_present
);
5987 EXPORT_SYMBOL_GPL(sched_smt_present
);
5989 static inline void set_idle_cores(int cpu
, int val
)
5991 struct sched_domain_shared
*sds
;
5993 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
5995 WRITE_ONCE(sds
->has_idle_cores
, val
);
5998 static inline bool test_idle_cores(int cpu
, bool def
)
6000 struct sched_domain_shared
*sds
;
6002 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
6004 return READ_ONCE(sds
->has_idle_cores
);
6010 * Scans the local SMT mask to see if the entire core is idle, and records this
6011 * information in sd_llc_shared->has_idle_cores.
6013 * Since SMT siblings share all cache levels, inspecting this limited remote
6014 * state should be fairly cheap.
6016 void __update_idle_core(struct rq
*rq
)
6018 int core
= cpu_of(rq
);
6022 if (test_idle_cores(core
, true))
6025 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
6029 if (!available_idle_cpu(cpu
))
6033 set_idle_cores(core
, 1);
6039 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6040 * there are no idle cores left in the system; tracked through
6041 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6043 static int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6045 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6048 if (!static_branch_likely(&sched_smt_present
))
6051 if (!test_idle_cores(target
, false))
6054 cpumask_and(cpus
, sched_domain_span(sd
), p
->cpus_ptr
);
6056 for_each_cpu_wrap(core
, cpus
, target
) {
6059 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
6060 if (!available_idle_cpu(cpu
)) {
6065 cpumask_andnot(cpus
, cpus
, cpu_smt_mask(core
));
6072 * Failed to find an idle core; stop looking for one.
6074 set_idle_cores(target
, 0);
6080 * Scan the local SMT mask for idle CPUs.
6082 static int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6086 if (!static_branch_likely(&sched_smt_present
))
6089 for_each_cpu(cpu
, cpu_smt_mask(target
)) {
6090 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
) ||
6091 !cpumask_test_cpu(cpu
, sched_domain_span(sd
)))
6093 if (available_idle_cpu(cpu
) || sched_idle_cpu(cpu
))
6100 #else /* CONFIG_SCHED_SMT */
6102 static inline int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6107 static inline int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6112 #endif /* CONFIG_SCHED_SMT */
6115 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6116 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6117 * average idle time for this rq (as found in rq->avg_idle).
6119 static int select_idle_cpu(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6121 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6122 struct sched_domain
*this_sd
;
6123 u64 avg_cost
, avg_idle
;
6125 int this = smp_processor_id();
6126 int cpu
, nr
= INT_MAX
;
6128 this_sd
= rcu_dereference(*this_cpu_ptr(&sd_llc
));
6133 * Due to large variance we need a large fuzz factor; hackbench in
6134 * particularly is sensitive here.
6136 avg_idle
= this_rq()->avg_idle
/ 512;
6137 avg_cost
= this_sd
->avg_scan_cost
+ 1;
6139 if (sched_feat(SIS_AVG_CPU
) && avg_idle
< avg_cost
)
6142 if (sched_feat(SIS_PROP
)) {
6143 u64 span_avg
= sd
->span_weight
* avg_idle
;
6144 if (span_avg
> 4*avg_cost
)
6145 nr
= div_u64(span_avg
, avg_cost
);
6150 time
= cpu_clock(this);
6152 cpumask_and(cpus
, sched_domain_span(sd
), p
->cpus_ptr
);
6154 for_each_cpu_wrap(cpu
, cpus
, target
) {
6157 if (available_idle_cpu(cpu
) || sched_idle_cpu(cpu
))
6161 time
= cpu_clock(this) - time
;
6162 update_avg(&this_sd
->avg_scan_cost
, time
);
6168 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6169 * the task fits. If no CPU is big enough, but there are idle ones, try to
6170 * maximize capacity.
6173 select_idle_capacity(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6175 unsigned long best_cap
= 0;
6176 int cpu
, best_cpu
= -1;
6177 struct cpumask
*cpus
;
6179 sync_entity_load_avg(&p
->se
);
6181 cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6182 cpumask_and(cpus
, sched_domain_span(sd
), p
->cpus_ptr
);
6184 for_each_cpu_wrap(cpu
, cpus
, target
) {
6185 unsigned long cpu_cap
= capacity_of(cpu
);
6187 if (!available_idle_cpu(cpu
) && !sched_idle_cpu(cpu
))
6189 if (task_fits_capacity(p
, cpu_cap
))
6192 if (cpu_cap
> best_cap
) {
6202 * Try and locate an idle core/thread in the LLC cache domain.
6204 static int select_idle_sibling(struct task_struct
*p
, int prev
, int target
)
6206 struct sched_domain
*sd
;
6207 int i
, recent_used_cpu
;
6210 * For asymmetric CPU capacity systems, our domain of interest is
6211 * sd_asym_cpucapacity rather than sd_llc.
6213 if (static_branch_unlikely(&sched_asym_cpucapacity
)) {
6214 sd
= rcu_dereference(per_cpu(sd_asym_cpucapacity
, target
));
6216 * On an asymmetric CPU capacity system where an exclusive
6217 * cpuset defines a symmetric island (i.e. one unique
6218 * capacity_orig value through the cpuset), the key will be set
6219 * but the CPUs within that cpuset will not have a domain with
6220 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6226 i
= select_idle_capacity(p
, sd
, target
);
6227 return ((unsigned)i
< nr_cpumask_bits
) ? i
: target
;
6231 if (available_idle_cpu(target
) || sched_idle_cpu(target
))
6235 * If the previous CPU is cache affine and idle, don't be stupid:
6237 if (prev
!= target
&& cpus_share_cache(prev
, target
) &&
6238 (available_idle_cpu(prev
) || sched_idle_cpu(prev
)))
6242 * Allow a per-cpu kthread to stack with the wakee if the
6243 * kworker thread and the tasks previous CPUs are the same.
6244 * The assumption is that the wakee queued work for the
6245 * per-cpu kthread that is now complete and the wakeup is
6246 * essentially a sync wakeup. An obvious example of this
6247 * pattern is IO completions.
6249 if (is_per_cpu_kthread(current
) &&
6250 prev
== smp_processor_id() &&
6251 this_rq()->nr_running
<= 1) {
6255 /* Check a recently used CPU as a potential idle candidate: */
6256 recent_used_cpu
= p
->recent_used_cpu
;
6257 if (recent_used_cpu
!= prev
&&
6258 recent_used_cpu
!= target
&&
6259 cpus_share_cache(recent_used_cpu
, target
) &&
6260 (available_idle_cpu(recent_used_cpu
) || sched_idle_cpu(recent_used_cpu
)) &&
6261 cpumask_test_cpu(p
->recent_used_cpu
, p
->cpus_ptr
)) {
6263 * Replace recent_used_cpu with prev as it is a potential
6264 * candidate for the next wake:
6266 p
->recent_used_cpu
= prev
;
6267 return recent_used_cpu
;
6270 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
6274 i
= select_idle_core(p
, sd
, target
);
6275 if ((unsigned)i
< nr_cpumask_bits
)
6278 i
= select_idle_cpu(p
, sd
, target
);
6279 if ((unsigned)i
< nr_cpumask_bits
)
6282 i
= select_idle_smt(p
, sd
, target
);
6283 if ((unsigned)i
< nr_cpumask_bits
)
6290 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6291 * @cpu: the CPU to get the utilization of
6293 * The unit of the return value must be the one of capacity so we can compare
6294 * the utilization with the capacity of the CPU that is available for CFS task
6295 * (ie cpu_capacity).
6297 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6298 * recent utilization of currently non-runnable tasks on a CPU. It represents
6299 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6300 * capacity_orig is the cpu_capacity available at the highest frequency
6301 * (arch_scale_freq_capacity()).
6302 * The utilization of a CPU converges towards a sum equal to or less than the
6303 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6304 * the running time on this CPU scaled by capacity_curr.
6306 * The estimated utilization of a CPU is defined to be the maximum between its
6307 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6308 * currently RUNNABLE on that CPU.
6309 * This allows to properly represent the expected utilization of a CPU which
6310 * has just got a big task running since a long sleep period. At the same time
6311 * however it preserves the benefits of the "blocked utilization" in
6312 * describing the potential for other tasks waking up on the same CPU.
6314 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6315 * higher than capacity_orig because of unfortunate rounding in
6316 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6317 * the average stabilizes with the new running time. We need to check that the
6318 * utilization stays within the range of [0..capacity_orig] and cap it if
6319 * necessary. Without utilization capping, a group could be seen as overloaded
6320 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6321 * available capacity. We allow utilization to overshoot capacity_curr (but not
6322 * capacity_orig) as it useful for predicting the capacity required after task
6323 * migrations (scheduler-driven DVFS).
6325 * Return: the (estimated) utilization for the specified CPU
6327 static inline unsigned long cpu_util(int cpu
)
6329 struct cfs_rq
*cfs_rq
;
6332 cfs_rq
= &cpu_rq(cpu
)->cfs
;
6333 util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6335 if (sched_feat(UTIL_EST
))
6336 util
= max(util
, READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
));
6338 return min_t(unsigned long, util
, capacity_orig_of(cpu
));
6342 * cpu_util_without: compute cpu utilization without any contributions from *p
6343 * @cpu: the CPU which utilization is requested
6344 * @p: the task which utilization should be discounted
6346 * The utilization of a CPU is defined by the utilization of tasks currently
6347 * enqueued on that CPU as well as tasks which are currently sleeping after an
6348 * execution on that CPU.
6350 * This method returns the utilization of the specified CPU by discounting the
6351 * utilization of the specified task, whenever the task is currently
6352 * contributing to the CPU utilization.
6354 static unsigned long cpu_util_without(int cpu
, struct task_struct
*p
)
6356 struct cfs_rq
*cfs_rq
;
6359 /* Task has no contribution or is new */
6360 if (cpu
!= task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
6361 return cpu_util(cpu
);
6363 cfs_rq
= &cpu_rq(cpu
)->cfs
;
6364 util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6366 /* Discount task's util from CPU's util */
6367 lsub_positive(&util
, task_util(p
));
6372 * a) if *p is the only task sleeping on this CPU, then:
6373 * cpu_util (== task_util) > util_est (== 0)
6374 * and thus we return:
6375 * cpu_util_without = (cpu_util - task_util) = 0
6377 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6379 * cpu_util >= task_util
6380 * cpu_util > util_est (== 0)
6381 * and thus we discount *p's blocked utilization to return:
6382 * cpu_util_without = (cpu_util - task_util) >= 0
6384 * c) if other tasks are RUNNABLE on that CPU and
6385 * util_est > cpu_util
6386 * then we use util_est since it returns a more restrictive
6387 * estimation of the spare capacity on that CPU, by just
6388 * considering the expected utilization of tasks already
6389 * runnable on that CPU.
6391 * Cases a) and b) are covered by the above code, while case c) is
6392 * covered by the following code when estimated utilization is
6395 if (sched_feat(UTIL_EST
)) {
6396 unsigned int estimated
=
6397 READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
);
6400 * Despite the following checks we still have a small window
6401 * for a possible race, when an execl's select_task_rq_fair()
6402 * races with LB's detach_task():
6405 * p->on_rq = TASK_ON_RQ_MIGRATING;
6406 * ---------------------------------- A
6407 * deactivate_task() \
6408 * dequeue_task() + RaceTime
6409 * util_est_dequeue() /
6410 * ---------------------------------- B
6412 * The additional check on "current == p" it's required to
6413 * properly fix the execl regression and it helps in further
6414 * reducing the chances for the above race.
6416 if (unlikely(task_on_rq_queued(p
) || current
== p
))
6417 lsub_positive(&estimated
, _task_util_est(p
));
6419 util
= max(util
, estimated
);
6423 * Utilization (estimated) can exceed the CPU capacity, thus let's
6424 * clamp to the maximum CPU capacity to ensure consistency with
6425 * the cpu_util call.
6427 return min_t(unsigned long, util
, capacity_orig_of(cpu
));
6431 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6434 static unsigned long cpu_util_next(int cpu
, struct task_struct
*p
, int dst_cpu
)
6436 struct cfs_rq
*cfs_rq
= &cpu_rq(cpu
)->cfs
;
6437 unsigned long util_est
, util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6440 * If @p migrates from @cpu to another, remove its contribution. Or,
6441 * if @p migrates from another CPU to @cpu, add its contribution. In
6442 * the other cases, @cpu is not impacted by the migration, so the
6443 * util_avg should already be correct.
6445 if (task_cpu(p
) == cpu
&& dst_cpu
!= cpu
)
6446 sub_positive(&util
, task_util(p
));
6447 else if (task_cpu(p
) != cpu
&& dst_cpu
== cpu
)
6448 util
+= task_util(p
);
6450 if (sched_feat(UTIL_EST
)) {
6451 util_est
= READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
);
6454 * During wake-up, the task isn't enqueued yet and doesn't
6455 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6456 * so just add it (if needed) to "simulate" what will be
6457 * cpu_util() after the task has been enqueued.
6460 util_est
+= _task_util_est(p
);
6462 util
= max(util
, util_est
);
6465 return min(util
, capacity_orig_of(cpu
));
6469 * compute_energy(): Estimates the energy that @pd would consume if @p was
6470 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6471 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6472 * to compute what would be the energy if we decided to actually migrate that
6476 compute_energy(struct task_struct
*p
, int dst_cpu
, struct perf_domain
*pd
)
6478 struct cpumask
*pd_mask
= perf_domain_span(pd
);
6479 unsigned long cpu_cap
= arch_scale_cpu_capacity(cpumask_first(pd_mask
));
6480 unsigned long max_util
= 0, sum_util
= 0;
6484 * The capacity state of CPUs of the current rd can be driven by CPUs
6485 * of another rd if they belong to the same pd. So, account for the
6486 * utilization of these CPUs too by masking pd with cpu_online_mask
6487 * instead of the rd span.
6489 * If an entire pd is outside of the current rd, it will not appear in
6490 * its pd list and will not be accounted by compute_energy().
6492 for_each_cpu_and(cpu
, pd_mask
, cpu_online_mask
) {
6493 unsigned long cpu_util
, util_cfs
= cpu_util_next(cpu
, p
, dst_cpu
);
6494 struct task_struct
*tsk
= cpu
== dst_cpu
? p
: NULL
;
6497 * Busy time computation: utilization clamping is not
6498 * required since the ratio (sum_util / cpu_capacity)
6499 * is already enough to scale the EM reported power
6500 * consumption at the (eventually clamped) cpu_capacity.
6502 sum_util
+= schedutil_cpu_util(cpu
, util_cfs
, cpu_cap
,
6506 * Performance domain frequency: utilization clamping
6507 * must be considered since it affects the selection
6508 * of the performance domain frequency.
6509 * NOTE: in case RT tasks are running, by default the
6510 * FREQUENCY_UTIL's utilization can be max OPP.
6512 cpu_util
= schedutil_cpu_util(cpu
, util_cfs
, cpu_cap
,
6513 FREQUENCY_UTIL
, tsk
);
6514 max_util
= max(max_util
, cpu_util
);
6517 return em_cpu_energy(pd
->em_pd
, max_util
, sum_util
);
6521 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6522 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6523 * spare capacity in each performance domain and uses it as a potential
6524 * candidate to execute the task. Then, it uses the Energy Model to figure
6525 * out which of the CPU candidates is the most energy-efficient.
6527 * The rationale for this heuristic is as follows. In a performance domain,
6528 * all the most energy efficient CPU candidates (according to the Energy
6529 * Model) are those for which we'll request a low frequency. When there are
6530 * several CPUs for which the frequency request will be the same, we don't
6531 * have enough data to break the tie between them, because the Energy Model
6532 * only includes active power costs. With this model, if we assume that
6533 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6534 * the maximum spare capacity in a performance domain is guaranteed to be among
6535 * the best candidates of the performance domain.
6537 * In practice, it could be preferable from an energy standpoint to pack
6538 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6539 * but that could also hurt our chances to go cluster idle, and we have no
6540 * ways to tell with the current Energy Model if this is actually a good
6541 * idea or not. So, find_energy_efficient_cpu() basically favors
6542 * cluster-packing, and spreading inside a cluster. That should at least be
6543 * a good thing for latency, and this is consistent with the idea that most
6544 * of the energy savings of EAS come from the asymmetry of the system, and
6545 * not so much from breaking the tie between identical CPUs. That's also the
6546 * reason why EAS is enabled in the topology code only for systems where
6547 * SD_ASYM_CPUCAPACITY is set.
6549 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6550 * they don't have any useful utilization data yet and it's not possible to
6551 * forecast their impact on energy consumption. Consequently, they will be
6552 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6553 * to be energy-inefficient in some use-cases. The alternative would be to
6554 * bias new tasks towards specific types of CPUs first, or to try to infer
6555 * their util_avg from the parent task, but those heuristics could hurt
6556 * other use-cases too. So, until someone finds a better way to solve this,
6557 * let's keep things simple by re-using the existing slow path.
6559 static int find_energy_efficient_cpu(struct task_struct
*p
, int prev_cpu
)
6561 unsigned long prev_delta
= ULONG_MAX
, best_delta
= ULONG_MAX
;
6562 struct root_domain
*rd
= cpu_rq(smp_processor_id())->rd
;
6563 unsigned long cpu_cap
, util
, base_energy
= 0;
6564 int cpu
, best_energy_cpu
= prev_cpu
;
6565 struct sched_domain
*sd
;
6566 struct perf_domain
*pd
;
6569 pd
= rcu_dereference(rd
->pd
);
6570 if (!pd
|| READ_ONCE(rd
->overutilized
))
6574 * Energy-aware wake-up happens on the lowest sched_domain starting
6575 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6577 sd
= rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity
));
6578 while (sd
&& !cpumask_test_cpu(prev_cpu
, sched_domain_span(sd
)))
6583 sync_entity_load_avg(&p
->se
);
6584 if (!task_util_est(p
))
6587 for (; pd
; pd
= pd
->next
) {
6588 unsigned long cur_delta
, spare_cap
, max_spare_cap
= 0;
6589 unsigned long base_energy_pd
;
6590 int max_spare_cap_cpu
= -1;
6592 /* Compute the 'base' energy of the pd, without @p */
6593 base_energy_pd
= compute_energy(p
, -1, pd
);
6594 base_energy
+= base_energy_pd
;
6596 for_each_cpu_and(cpu
, perf_domain_span(pd
), sched_domain_span(sd
)) {
6597 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
6600 util
= cpu_util_next(cpu
, p
, cpu
);
6601 cpu_cap
= capacity_of(cpu
);
6602 spare_cap
= cpu_cap
;
6603 lsub_positive(&spare_cap
, util
);
6606 * Skip CPUs that cannot satisfy the capacity request.
6607 * IOW, placing the task there would make the CPU
6608 * overutilized. Take uclamp into account to see how
6609 * much capacity we can get out of the CPU; this is
6610 * aligned with schedutil_cpu_util().
6612 util
= uclamp_rq_util_with(cpu_rq(cpu
), util
, p
);
6613 if (!fits_capacity(util
, cpu_cap
))
6616 /* Always use prev_cpu as a candidate. */
6617 if (cpu
== prev_cpu
) {
6618 prev_delta
= compute_energy(p
, prev_cpu
, pd
);
6619 prev_delta
-= base_energy_pd
;
6620 best_delta
= min(best_delta
, prev_delta
);
6624 * Find the CPU with the maximum spare capacity in
6625 * the performance domain
6627 if (spare_cap
> max_spare_cap
) {
6628 max_spare_cap
= spare_cap
;
6629 max_spare_cap_cpu
= cpu
;
6633 /* Evaluate the energy impact of using this CPU. */
6634 if (max_spare_cap_cpu
>= 0 && max_spare_cap_cpu
!= prev_cpu
) {
6635 cur_delta
= compute_energy(p
, max_spare_cap_cpu
, pd
);
6636 cur_delta
-= base_energy_pd
;
6637 if (cur_delta
< best_delta
) {
6638 best_delta
= cur_delta
;
6639 best_energy_cpu
= max_spare_cap_cpu
;
6647 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6648 * least 6% of the energy used by prev_cpu.
6650 if (prev_delta
== ULONG_MAX
)
6651 return best_energy_cpu
;
6653 if ((prev_delta
- best_delta
) > ((prev_delta
+ base_energy
) >> 4))
6654 return best_energy_cpu
;
6665 * select_task_rq_fair: Select target runqueue for the waking task in domains
6666 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6667 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6669 * Balances load by selecting the idlest CPU in the idlest group, or under
6670 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6672 * Returns the target CPU number.
6674 * preempt must be disabled.
6677 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
6679 struct sched_domain
*tmp
, *sd
= NULL
;
6680 int cpu
= smp_processor_id();
6681 int new_cpu
= prev_cpu
;
6682 int want_affine
= 0;
6683 int sync
= (wake_flags
& WF_SYNC
) && !(current
->flags
& PF_EXITING
);
6685 if (sd_flag
& SD_BALANCE_WAKE
) {
6688 if (sched_energy_enabled()) {
6689 new_cpu
= find_energy_efficient_cpu(p
, prev_cpu
);
6695 want_affine
= !wake_wide(p
) && cpumask_test_cpu(cpu
, p
->cpus_ptr
);
6699 for_each_domain(cpu
, tmp
) {
6701 * If both 'cpu' and 'prev_cpu' are part of this domain,
6702 * cpu is a valid SD_WAKE_AFFINE target.
6704 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
6705 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
6706 if (cpu
!= prev_cpu
)
6707 new_cpu
= wake_affine(tmp
, p
, cpu
, prev_cpu
, sync
);
6709 sd
= NULL
; /* Prefer wake_affine over balance flags */
6713 if (tmp
->flags
& sd_flag
)
6715 else if (!want_affine
)
6721 new_cpu
= find_idlest_cpu(sd
, p
, cpu
, prev_cpu
, sd_flag
);
6722 } else if (sd_flag
& SD_BALANCE_WAKE
) { /* XXX always ? */
6725 new_cpu
= select_idle_sibling(p
, prev_cpu
, new_cpu
);
6728 current
->recent_used_cpu
= cpu
;
6735 static void detach_entity_cfs_rq(struct sched_entity
*se
);
6738 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6739 * cfs_rq_of(p) references at time of call are still valid and identify the
6740 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6742 static void migrate_task_rq_fair(struct task_struct
*p
, int new_cpu
)
6745 * As blocked tasks retain absolute vruntime the migration needs to
6746 * deal with this by subtracting the old and adding the new
6747 * min_vruntime -- the latter is done by enqueue_entity() when placing
6748 * the task on the new runqueue.
6750 if (p
->state
== TASK_WAKING
) {
6751 struct sched_entity
*se
= &p
->se
;
6752 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6755 #ifndef CONFIG_64BIT
6756 u64 min_vruntime_copy
;
6759 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
6761 min_vruntime
= cfs_rq
->min_vruntime
;
6762 } while (min_vruntime
!= min_vruntime_copy
);
6764 min_vruntime
= cfs_rq
->min_vruntime
;
6767 se
->vruntime
-= min_vruntime
;
6770 if (p
->on_rq
== TASK_ON_RQ_MIGRATING
) {
6772 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6773 * rq->lock and can modify state directly.
6775 lockdep_assert_held(&task_rq(p
)->lock
);
6776 detach_entity_cfs_rq(&p
->se
);
6780 * We are supposed to update the task to "current" time, then
6781 * its up to date and ready to go to new CPU/cfs_rq. But we
6782 * have difficulty in getting what current time is, so simply
6783 * throw away the out-of-date time. This will result in the
6784 * wakee task is less decayed, but giving the wakee more load
6787 remove_entity_load_avg(&p
->se
);
6790 /* Tell new CPU we are migrated */
6791 p
->se
.avg
.last_update_time
= 0;
6793 /* We have migrated, no longer consider this task hot */
6794 p
->se
.exec_start
= 0;
6796 update_scan_period(p
, new_cpu
);
6799 static void task_dead_fair(struct task_struct
*p
)
6801 remove_entity_load_avg(&p
->se
);
6805 balance_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
6810 return newidle_balance(rq
, rf
) != 0;
6812 #endif /* CONFIG_SMP */
6814 static unsigned long wakeup_gran(struct sched_entity
*se
)
6816 unsigned long gran
= sysctl_sched_wakeup_granularity
;
6819 * Since its curr running now, convert the gran from real-time
6820 * to virtual-time in his units.
6822 * By using 'se' instead of 'curr' we penalize light tasks, so
6823 * they get preempted easier. That is, if 'se' < 'curr' then
6824 * the resulting gran will be larger, therefore penalizing the
6825 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6826 * be smaller, again penalizing the lighter task.
6828 * This is especially important for buddies when the leftmost
6829 * task is higher priority than the buddy.
6831 return calc_delta_fair(gran
, se
);
6835 * Should 'se' preempt 'curr'.
6849 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
6851 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
6856 gran
= wakeup_gran(se
);
6863 static void set_last_buddy(struct sched_entity
*se
)
6865 if (entity_is_task(se
) && unlikely(task_has_idle_policy(task_of(se
))))
6868 for_each_sched_entity(se
) {
6869 if (SCHED_WARN_ON(!se
->on_rq
))
6871 cfs_rq_of(se
)->last
= se
;
6875 static void set_next_buddy(struct sched_entity
*se
)
6877 if (entity_is_task(se
) && unlikely(task_has_idle_policy(task_of(se
))))
6880 for_each_sched_entity(se
) {
6881 if (SCHED_WARN_ON(!se
->on_rq
))
6883 cfs_rq_of(se
)->next
= se
;
6887 static void set_skip_buddy(struct sched_entity
*se
)
6889 for_each_sched_entity(se
)
6890 cfs_rq_of(se
)->skip
= se
;
6894 * Preempt the current task with a newly woken task if needed:
6896 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
6898 struct task_struct
*curr
= rq
->curr
;
6899 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
6900 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
6901 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
6902 int next_buddy_marked
= 0;
6904 if (unlikely(se
== pse
))
6908 * This is possible from callers such as attach_tasks(), in which we
6909 * unconditionally check_prempt_curr() after an enqueue (which may have
6910 * lead to a throttle). This both saves work and prevents false
6911 * next-buddy nomination below.
6913 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
6916 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
6917 set_next_buddy(pse
);
6918 next_buddy_marked
= 1;
6922 * We can come here with TIF_NEED_RESCHED already set from new task
6925 * Note: this also catches the edge-case of curr being in a throttled
6926 * group (e.g. via set_curr_task), since update_curr() (in the
6927 * enqueue of curr) will have resulted in resched being set. This
6928 * prevents us from potentially nominating it as a false LAST_BUDDY
6931 if (test_tsk_need_resched(curr
))
6934 /* Idle tasks are by definition preempted by non-idle tasks. */
6935 if (unlikely(task_has_idle_policy(curr
)) &&
6936 likely(!task_has_idle_policy(p
)))
6940 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6941 * is driven by the tick):
6943 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
6946 find_matching_se(&se
, &pse
);
6947 update_curr(cfs_rq_of(se
));
6949 if (wakeup_preempt_entity(se
, pse
) == 1) {
6951 * Bias pick_next to pick the sched entity that is
6952 * triggering this preemption.
6954 if (!next_buddy_marked
)
6955 set_next_buddy(pse
);
6964 * Only set the backward buddy when the current task is still
6965 * on the rq. This can happen when a wakeup gets interleaved
6966 * with schedule on the ->pre_schedule() or idle_balance()
6967 * point, either of which can * drop the rq lock.
6969 * Also, during early boot the idle thread is in the fair class,
6970 * for obvious reasons its a bad idea to schedule back to it.
6972 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
6975 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
6979 struct task_struct
*
6980 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
6982 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
6983 struct sched_entity
*se
;
6984 struct task_struct
*p
;
6988 if (!sched_fair_runnable(rq
))
6991 #ifdef CONFIG_FAIR_GROUP_SCHED
6992 if (!prev
|| prev
->sched_class
!= &fair_sched_class
)
6996 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6997 * likely that a next task is from the same cgroup as the current.
6999 * Therefore attempt to avoid putting and setting the entire cgroup
7000 * hierarchy, only change the part that actually changes.
7004 struct sched_entity
*curr
= cfs_rq
->curr
;
7007 * Since we got here without doing put_prev_entity() we also
7008 * have to consider cfs_rq->curr. If it is still a runnable
7009 * entity, update_curr() will update its vruntime, otherwise
7010 * forget we've ever seen it.
7014 update_curr(cfs_rq
);
7019 * This call to check_cfs_rq_runtime() will do the
7020 * throttle and dequeue its entity in the parent(s).
7021 * Therefore the nr_running test will indeed
7024 if (unlikely(check_cfs_rq_runtime(cfs_rq
))) {
7027 if (!cfs_rq
->nr_running
)
7034 se
= pick_next_entity(cfs_rq
, curr
);
7035 cfs_rq
= group_cfs_rq(se
);
7041 * Since we haven't yet done put_prev_entity and if the selected task
7042 * is a different task than we started out with, try and touch the
7043 * least amount of cfs_rqs.
7046 struct sched_entity
*pse
= &prev
->se
;
7048 while (!(cfs_rq
= is_same_group(se
, pse
))) {
7049 int se_depth
= se
->depth
;
7050 int pse_depth
= pse
->depth
;
7052 if (se_depth
<= pse_depth
) {
7053 put_prev_entity(cfs_rq_of(pse
), pse
);
7054 pse
= parent_entity(pse
);
7056 if (se_depth
>= pse_depth
) {
7057 set_next_entity(cfs_rq_of(se
), se
);
7058 se
= parent_entity(se
);
7062 put_prev_entity(cfs_rq
, pse
);
7063 set_next_entity(cfs_rq
, se
);
7070 put_prev_task(rq
, prev
);
7073 se
= pick_next_entity(cfs_rq
, NULL
);
7074 set_next_entity(cfs_rq
, se
);
7075 cfs_rq
= group_cfs_rq(se
);
7080 done
: __maybe_unused
;
7083 * Move the next running task to the front of
7084 * the list, so our cfs_tasks list becomes MRU
7087 list_move(&p
->se
.group_node
, &rq
->cfs_tasks
);
7090 if (hrtick_enabled(rq
))
7091 hrtick_start_fair(rq
, p
);
7093 update_misfit_status(p
, rq
);
7101 new_tasks
= newidle_balance(rq
, rf
);
7104 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7105 * possible for any higher priority task to appear. In that case we
7106 * must re-start the pick_next_entity() loop.
7115 * rq is about to be idle, check if we need to update the
7116 * lost_idle_time of clock_pelt
7118 update_idle_rq_clock_pelt(rq
);
7123 static struct task_struct
*__pick_next_task_fair(struct rq
*rq
)
7125 return pick_next_task_fair(rq
, NULL
, NULL
);
7129 * Account for a descheduled task:
7131 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
7133 struct sched_entity
*se
= &prev
->se
;
7134 struct cfs_rq
*cfs_rq
;
7136 for_each_sched_entity(se
) {
7137 cfs_rq
= cfs_rq_of(se
);
7138 put_prev_entity(cfs_rq
, se
);
7143 * sched_yield() is very simple
7145 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7147 static void yield_task_fair(struct rq
*rq
)
7149 struct task_struct
*curr
= rq
->curr
;
7150 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
7151 struct sched_entity
*se
= &curr
->se
;
7154 * Are we the only task in the tree?
7156 if (unlikely(rq
->nr_running
== 1))
7159 clear_buddies(cfs_rq
, se
);
7161 if (curr
->policy
!= SCHED_BATCH
) {
7162 update_rq_clock(rq
);
7164 * Update run-time statistics of the 'current'.
7166 update_curr(cfs_rq
);
7168 * Tell update_rq_clock() that we've just updated,
7169 * so we don't do microscopic update in schedule()
7170 * and double the fastpath cost.
7172 rq_clock_skip_update(rq
);
7178 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
)
7180 struct sched_entity
*se
= &p
->se
;
7182 /* throttled hierarchies are not runnable */
7183 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
7186 /* Tell the scheduler that we'd really like pse to run next. */
7189 yield_task_fair(rq
);
7195 /**************************************************
7196 * Fair scheduling class load-balancing methods.
7200 * The purpose of load-balancing is to achieve the same basic fairness the
7201 * per-CPU scheduler provides, namely provide a proportional amount of compute
7202 * time to each task. This is expressed in the following equation:
7204 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7206 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7207 * W_i,0 is defined as:
7209 * W_i,0 = \Sum_j w_i,j (2)
7211 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7212 * is derived from the nice value as per sched_prio_to_weight[].
7214 * The weight average is an exponential decay average of the instantaneous
7217 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7219 * C_i is the compute capacity of CPU i, typically it is the
7220 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7221 * can also include other factors [XXX].
7223 * To achieve this balance we define a measure of imbalance which follows
7224 * directly from (1):
7226 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7228 * We them move tasks around to minimize the imbalance. In the continuous
7229 * function space it is obvious this converges, in the discrete case we get
7230 * a few fun cases generally called infeasible weight scenarios.
7233 * - infeasible weights;
7234 * - local vs global optima in the discrete case. ]
7239 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7240 * for all i,j solution, we create a tree of CPUs that follows the hardware
7241 * topology where each level pairs two lower groups (or better). This results
7242 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7243 * tree to only the first of the previous level and we decrease the frequency
7244 * of load-balance at each level inv. proportional to the number of CPUs in
7250 * \Sum { --- * --- * 2^i } = O(n) (5)
7252 * `- size of each group
7253 * | | `- number of CPUs doing load-balance
7255 * `- sum over all levels
7257 * Coupled with a limit on how many tasks we can migrate every balance pass,
7258 * this makes (5) the runtime complexity of the balancer.
7260 * An important property here is that each CPU is still (indirectly) connected
7261 * to every other CPU in at most O(log n) steps:
7263 * The adjacency matrix of the resulting graph is given by:
7266 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7269 * And you'll find that:
7271 * A^(log_2 n)_i,j != 0 for all i,j (7)
7273 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7274 * The task movement gives a factor of O(m), giving a convergence complexity
7277 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7282 * In order to avoid CPUs going idle while there's still work to do, new idle
7283 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7284 * tree itself instead of relying on other CPUs to bring it work.
7286 * This adds some complexity to both (5) and (8) but it reduces the total idle
7294 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7297 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7302 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7304 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7306 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7309 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7310 * rewrite all of this once again.]
7313 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
7315 enum fbq_type
{ regular
, remote
, all
};
7318 * 'group_type' describes the group of CPUs at the moment of load balancing.
7320 * The enum is ordered by pulling priority, with the group with lowest priority
7321 * first so the group_type can simply be compared when selecting the busiest
7322 * group. See update_sd_pick_busiest().
7325 /* The group has spare capacity that can be used to run more tasks. */
7326 group_has_spare
= 0,
7328 * The group is fully used and the tasks don't compete for more CPU
7329 * cycles. Nevertheless, some tasks might wait before running.
7333 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
7334 * and must be migrated to a more powerful CPU.
7338 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7339 * and the task should be migrated to it instead of running on the
7344 * The tasks' affinity constraints previously prevented the scheduler
7345 * from balancing the load across the system.
7349 * The CPU is overloaded and can't provide expected CPU cycles to all
7355 enum migration_type
{
7362 #define LBF_ALL_PINNED 0x01
7363 #define LBF_NEED_BREAK 0x02
7364 #define LBF_DST_PINNED 0x04
7365 #define LBF_SOME_PINNED 0x08
7366 #define LBF_NOHZ_STATS 0x10
7367 #define LBF_NOHZ_AGAIN 0x20
7370 struct sched_domain
*sd
;
7378 struct cpumask
*dst_grpmask
;
7380 enum cpu_idle_type idle
;
7382 /* The set of CPUs under consideration for load-balancing */
7383 struct cpumask
*cpus
;
7388 unsigned int loop_break
;
7389 unsigned int loop_max
;
7391 enum fbq_type fbq_type
;
7392 enum migration_type migration_type
;
7393 struct list_head tasks
;
7397 * Is this task likely cache-hot:
7399 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
7403 lockdep_assert_held(&env
->src_rq
->lock
);
7405 if (p
->sched_class
!= &fair_sched_class
)
7408 if (unlikely(task_has_idle_policy(p
)))
7411 /* SMT siblings share cache */
7412 if (env
->sd
->flags
& SD_SHARE_CPUCAPACITY
)
7416 * Buddy candidates are cache hot:
7418 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
7419 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
7420 &p
->se
== cfs_rq_of(&p
->se
)->last
))
7423 if (sysctl_sched_migration_cost
== -1)
7425 if (sysctl_sched_migration_cost
== 0)
7428 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
7430 return delta
< (s64
)sysctl_sched_migration_cost
;
7433 #ifdef CONFIG_NUMA_BALANCING
7435 * Returns 1, if task migration degrades locality
7436 * Returns 0, if task migration improves locality i.e migration preferred.
7437 * Returns -1, if task migration is not affected by locality.
7439 static int migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
7441 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
7442 unsigned long src_weight
, dst_weight
;
7443 int src_nid
, dst_nid
, dist
;
7445 if (!static_branch_likely(&sched_numa_balancing
))
7448 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
7451 src_nid
= cpu_to_node(env
->src_cpu
);
7452 dst_nid
= cpu_to_node(env
->dst_cpu
);
7454 if (src_nid
== dst_nid
)
7457 /* Migrating away from the preferred node is always bad. */
7458 if (src_nid
== p
->numa_preferred_nid
) {
7459 if (env
->src_rq
->nr_running
> env
->src_rq
->nr_preferred_running
)
7465 /* Encourage migration to the preferred node. */
7466 if (dst_nid
== p
->numa_preferred_nid
)
7469 /* Leaving a core idle is often worse than degrading locality. */
7470 if (env
->idle
== CPU_IDLE
)
7473 dist
= node_distance(src_nid
, dst_nid
);
7475 src_weight
= group_weight(p
, src_nid
, dist
);
7476 dst_weight
= group_weight(p
, dst_nid
, dist
);
7478 src_weight
= task_weight(p
, src_nid
, dist
);
7479 dst_weight
= task_weight(p
, dst_nid
, dist
);
7482 return dst_weight
< src_weight
;
7486 static inline int migrate_degrades_locality(struct task_struct
*p
,
7494 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7497 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
7501 lockdep_assert_held(&env
->src_rq
->lock
);
7504 * We do not migrate tasks that are:
7505 * 1) throttled_lb_pair, or
7506 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7507 * 3) running (obviously), or
7508 * 4) are cache-hot on their current CPU.
7510 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
7513 if (!cpumask_test_cpu(env
->dst_cpu
, p
->cpus_ptr
)) {
7516 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_affine
);
7518 env
->flags
|= LBF_SOME_PINNED
;
7521 * Remember if this task can be migrated to any other CPU in
7522 * our sched_group. We may want to revisit it if we couldn't
7523 * meet load balance goals by pulling other tasks on src_cpu.
7525 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7526 * already computed one in current iteration.
7528 if (env
->idle
== CPU_NEWLY_IDLE
|| (env
->flags
& LBF_DST_PINNED
))
7531 /* Prevent to re-select dst_cpu via env's CPUs: */
7532 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
7533 if (cpumask_test_cpu(cpu
, p
->cpus_ptr
)) {
7534 env
->flags
|= LBF_DST_PINNED
;
7535 env
->new_dst_cpu
= cpu
;
7543 /* Record that we found atleast one task that could run on dst_cpu */
7544 env
->flags
&= ~LBF_ALL_PINNED
;
7546 if (task_running(env
->src_rq
, p
)) {
7547 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_running
);
7552 * Aggressive migration if:
7553 * 1) destination numa is preferred
7554 * 2) task is cache cold, or
7555 * 3) too many balance attempts have failed.
7557 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
7558 if (tsk_cache_hot
== -1)
7559 tsk_cache_hot
= task_hot(p
, env
);
7561 if (tsk_cache_hot
<= 0 ||
7562 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
7563 if (tsk_cache_hot
== 1) {
7564 schedstat_inc(env
->sd
->lb_hot_gained
[env
->idle
]);
7565 schedstat_inc(p
->se
.statistics
.nr_forced_migrations
);
7570 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_hot
);
7575 * detach_task() -- detach the task for the migration specified in env
7577 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
7579 lockdep_assert_held(&env
->src_rq
->lock
);
7581 deactivate_task(env
->src_rq
, p
, DEQUEUE_NOCLOCK
);
7582 set_task_cpu(p
, env
->dst_cpu
);
7586 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7587 * part of active balancing operations within "domain".
7589 * Returns a task if successful and NULL otherwise.
7591 static struct task_struct
*detach_one_task(struct lb_env
*env
)
7593 struct task_struct
*p
;
7595 lockdep_assert_held(&env
->src_rq
->lock
);
7597 list_for_each_entry_reverse(p
,
7598 &env
->src_rq
->cfs_tasks
, se
.group_node
) {
7599 if (!can_migrate_task(p
, env
))
7602 detach_task(p
, env
);
7605 * Right now, this is only the second place where
7606 * lb_gained[env->idle] is updated (other is detach_tasks)
7607 * so we can safely collect stats here rather than
7608 * inside detach_tasks().
7610 schedstat_inc(env
->sd
->lb_gained
[env
->idle
]);
7616 static const unsigned int sched_nr_migrate_break
= 32;
7619 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7620 * busiest_rq, as part of a balancing operation within domain "sd".
7622 * Returns number of detached tasks if successful and 0 otherwise.
7624 static int detach_tasks(struct lb_env
*env
)
7626 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
7627 unsigned long util
, load
;
7628 struct task_struct
*p
;
7631 lockdep_assert_held(&env
->src_rq
->lock
);
7633 if (env
->imbalance
<= 0)
7636 while (!list_empty(tasks
)) {
7638 * We don't want to steal all, otherwise we may be treated likewise,
7639 * which could at worst lead to a livelock crash.
7641 if (env
->idle
!= CPU_NOT_IDLE
&& env
->src_rq
->nr_running
<= 1)
7644 p
= list_last_entry(tasks
, struct task_struct
, se
.group_node
);
7647 /* We've more or less seen every task there is, call it quits */
7648 if (env
->loop
> env
->loop_max
)
7651 /* take a breather every nr_migrate tasks */
7652 if (env
->loop
> env
->loop_break
) {
7653 env
->loop_break
+= sched_nr_migrate_break
;
7654 env
->flags
|= LBF_NEED_BREAK
;
7658 if (!can_migrate_task(p
, env
))
7661 switch (env
->migration_type
) {
7664 * Depending of the number of CPUs and tasks and the
7665 * cgroup hierarchy, task_h_load() can return a null
7666 * value. Make sure that env->imbalance decreases
7667 * otherwise detach_tasks() will stop only after
7668 * detaching up to loop_max tasks.
7670 load
= max_t(unsigned long, task_h_load(p
), 1);
7672 if (sched_feat(LB_MIN
) &&
7673 load
< 16 && !env
->sd
->nr_balance_failed
)
7677 * Make sure that we don't migrate too much load.
7678 * Nevertheless, let relax the constraint if
7679 * scheduler fails to find a good waiting task to
7683 if ((load
>> env
->sd
->nr_balance_failed
) > env
->imbalance
)
7686 env
->imbalance
-= load
;
7690 util
= task_util_est(p
);
7692 if (util
> env
->imbalance
)
7695 env
->imbalance
-= util
;
7702 case migrate_misfit
:
7703 /* This is not a misfit task */
7704 if (task_fits_capacity(p
, capacity_of(env
->src_cpu
)))
7711 detach_task(p
, env
);
7712 list_add(&p
->se
.group_node
, &env
->tasks
);
7716 #ifdef CONFIG_PREEMPTION
7718 * NEWIDLE balancing is a source of latency, so preemptible
7719 * kernels will stop after the first task is detached to minimize
7720 * the critical section.
7722 if (env
->idle
== CPU_NEWLY_IDLE
)
7727 * We only want to steal up to the prescribed amount of
7730 if (env
->imbalance
<= 0)
7735 list_move(&p
->se
.group_node
, tasks
);
7739 * Right now, this is one of only two places we collect this stat
7740 * so we can safely collect detach_one_task() stats here rather
7741 * than inside detach_one_task().
7743 schedstat_add(env
->sd
->lb_gained
[env
->idle
], detached
);
7749 * attach_task() -- attach the task detached by detach_task() to its new rq.
7751 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
7753 lockdep_assert_held(&rq
->lock
);
7755 BUG_ON(task_rq(p
) != rq
);
7756 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
7757 check_preempt_curr(rq
, p
, 0);
7761 * attach_one_task() -- attaches the task returned from detach_one_task() to
7764 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
7769 update_rq_clock(rq
);
7775 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7778 static void attach_tasks(struct lb_env
*env
)
7780 struct list_head
*tasks
= &env
->tasks
;
7781 struct task_struct
*p
;
7784 rq_lock(env
->dst_rq
, &rf
);
7785 update_rq_clock(env
->dst_rq
);
7787 while (!list_empty(tasks
)) {
7788 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
7789 list_del_init(&p
->se
.group_node
);
7791 attach_task(env
->dst_rq
, p
);
7794 rq_unlock(env
->dst_rq
, &rf
);
7797 #ifdef CONFIG_NO_HZ_COMMON
7798 static inline bool cfs_rq_has_blocked(struct cfs_rq
*cfs_rq
)
7800 if (cfs_rq
->avg
.load_avg
)
7803 if (cfs_rq
->avg
.util_avg
)
7809 static inline bool others_have_blocked(struct rq
*rq
)
7811 if (READ_ONCE(rq
->avg_rt
.util_avg
))
7814 if (READ_ONCE(rq
->avg_dl
.util_avg
))
7817 if (thermal_load_avg(rq
))
7820 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7821 if (READ_ONCE(rq
->avg_irq
.util_avg
))
7828 static inline void update_blocked_load_status(struct rq
*rq
, bool has_blocked
)
7830 rq
->last_blocked_load_update_tick
= jiffies
;
7833 rq
->has_blocked_load
= 0;
7836 static inline bool cfs_rq_has_blocked(struct cfs_rq
*cfs_rq
) { return false; }
7837 static inline bool others_have_blocked(struct rq
*rq
) { return false; }
7838 static inline void update_blocked_load_status(struct rq
*rq
, bool has_blocked
) {}
7841 static bool __update_blocked_others(struct rq
*rq
, bool *done
)
7843 const struct sched_class
*curr_class
;
7844 u64 now
= rq_clock_pelt(rq
);
7845 unsigned long thermal_pressure
;
7849 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
7850 * DL and IRQ signals have been updated before updating CFS.
7852 curr_class
= rq
->curr
->sched_class
;
7854 thermal_pressure
= arch_scale_thermal_pressure(cpu_of(rq
));
7856 decayed
= update_rt_rq_load_avg(now
, rq
, curr_class
== &rt_sched_class
) |
7857 update_dl_rq_load_avg(now
, rq
, curr_class
== &dl_sched_class
) |
7858 update_thermal_load_avg(rq_clock_thermal(rq
), rq
, thermal_pressure
) |
7859 update_irq_load_avg(rq
, 0);
7861 if (others_have_blocked(rq
))
7867 #ifdef CONFIG_FAIR_GROUP_SCHED
7869 static inline bool cfs_rq_is_decayed(struct cfs_rq
*cfs_rq
)
7871 if (cfs_rq
->load
.weight
)
7874 if (cfs_rq
->avg
.load_sum
)
7877 if (cfs_rq
->avg
.util_sum
)
7880 if (cfs_rq
->avg
.runnable_sum
)
7886 static bool __update_blocked_fair(struct rq
*rq
, bool *done
)
7888 struct cfs_rq
*cfs_rq
, *pos
;
7889 bool decayed
= false;
7890 int cpu
= cpu_of(rq
);
7893 * Iterates the task_group tree in a bottom up fashion, see
7894 * list_add_leaf_cfs_rq() for details.
7896 for_each_leaf_cfs_rq_safe(rq
, cfs_rq
, pos
) {
7897 struct sched_entity
*se
;
7899 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq
), cfs_rq
)) {
7900 update_tg_load_avg(cfs_rq
);
7902 if (cfs_rq
== &rq
->cfs
)
7906 /* Propagate pending load changes to the parent, if any: */
7907 se
= cfs_rq
->tg
->se
[cpu
];
7908 if (se
&& !skip_blocked_update(se
))
7909 update_load_avg(cfs_rq_of(se
), se
, 0);
7912 * There can be a lot of idle CPU cgroups. Don't let fully
7913 * decayed cfs_rqs linger on the list.
7915 if (cfs_rq_is_decayed(cfs_rq
))
7916 list_del_leaf_cfs_rq(cfs_rq
);
7918 /* Don't need periodic decay once load/util_avg are null */
7919 if (cfs_rq_has_blocked(cfs_rq
))
7927 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7928 * This needs to be done in a top-down fashion because the load of a child
7929 * group is a fraction of its parents load.
7931 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
7933 struct rq
*rq
= rq_of(cfs_rq
);
7934 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
7935 unsigned long now
= jiffies
;
7938 if (cfs_rq
->last_h_load_update
== now
)
7941 WRITE_ONCE(cfs_rq
->h_load_next
, NULL
);
7942 for_each_sched_entity(se
) {
7943 cfs_rq
= cfs_rq_of(se
);
7944 WRITE_ONCE(cfs_rq
->h_load_next
, se
);
7945 if (cfs_rq
->last_h_load_update
== now
)
7950 cfs_rq
->h_load
= cfs_rq_load_avg(cfs_rq
);
7951 cfs_rq
->last_h_load_update
= now
;
7954 while ((se
= READ_ONCE(cfs_rq
->h_load_next
)) != NULL
) {
7955 load
= cfs_rq
->h_load
;
7956 load
= div64_ul(load
* se
->avg
.load_avg
,
7957 cfs_rq_load_avg(cfs_rq
) + 1);
7958 cfs_rq
= group_cfs_rq(se
);
7959 cfs_rq
->h_load
= load
;
7960 cfs_rq
->last_h_load_update
= now
;
7964 static unsigned long task_h_load(struct task_struct
*p
)
7966 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
7968 update_cfs_rq_h_load(cfs_rq
);
7969 return div64_ul(p
->se
.avg
.load_avg
* cfs_rq
->h_load
,
7970 cfs_rq_load_avg(cfs_rq
) + 1);
7973 static bool __update_blocked_fair(struct rq
*rq
, bool *done
)
7975 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
7978 decayed
= update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq
), cfs_rq
);
7979 if (cfs_rq_has_blocked(cfs_rq
))
7985 static unsigned long task_h_load(struct task_struct
*p
)
7987 return p
->se
.avg
.load_avg
;
7991 static void update_blocked_averages(int cpu
)
7993 bool decayed
= false, done
= true;
7994 struct rq
*rq
= cpu_rq(cpu
);
7997 rq_lock_irqsave(rq
, &rf
);
7998 update_rq_clock(rq
);
8000 decayed
|= __update_blocked_others(rq
, &done
);
8001 decayed
|= __update_blocked_fair(rq
, &done
);
8003 update_blocked_load_status(rq
, !done
);
8005 cpufreq_update_util(rq
, 0);
8006 rq_unlock_irqrestore(rq
, &rf
);
8009 /********** Helpers for find_busiest_group ************************/
8012 * sg_lb_stats - stats of a sched_group required for load_balancing
8014 struct sg_lb_stats
{
8015 unsigned long avg_load
; /*Avg load across the CPUs of the group */
8016 unsigned long group_load
; /* Total load over the CPUs of the group */
8017 unsigned long group_capacity
;
8018 unsigned long group_util
; /* Total utilization over the CPUs of the group */
8019 unsigned long group_runnable
; /* Total runnable time over the CPUs of the group */
8020 unsigned int sum_nr_running
; /* Nr of tasks running in the group */
8021 unsigned int sum_h_nr_running
; /* Nr of CFS tasks running in the group */
8022 unsigned int idle_cpus
;
8023 unsigned int group_weight
;
8024 enum group_type group_type
;
8025 unsigned int group_asym_packing
; /* Tasks should be moved to preferred CPU */
8026 unsigned long group_misfit_task_load
; /* A CPU has a task too big for its capacity */
8027 #ifdef CONFIG_NUMA_BALANCING
8028 unsigned int nr_numa_running
;
8029 unsigned int nr_preferred_running
;
8034 * sd_lb_stats - Structure to store the statistics of a sched_domain
8035 * during load balancing.
8037 struct sd_lb_stats
{
8038 struct sched_group
*busiest
; /* Busiest group in this sd */
8039 struct sched_group
*local
; /* Local group in this sd */
8040 unsigned long total_load
; /* Total load of all groups in sd */
8041 unsigned long total_capacity
; /* Total capacity of all groups in sd */
8042 unsigned long avg_load
; /* Average load across all groups in sd */
8043 unsigned int prefer_sibling
; /* tasks should go to sibling first */
8045 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
8046 struct sg_lb_stats local_stat
; /* Statistics of the local group */
8049 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
8052 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8053 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8054 * We must however set busiest_stat::group_type and
8055 * busiest_stat::idle_cpus to the worst busiest group because
8056 * update_sd_pick_busiest() reads these before assignment.
8058 *sds
= (struct sd_lb_stats
){
8062 .total_capacity
= 0UL,
8064 .idle_cpus
= UINT_MAX
,
8065 .group_type
= group_has_spare
,
8070 static unsigned long scale_rt_capacity(int cpu
)
8072 struct rq
*rq
= cpu_rq(cpu
);
8073 unsigned long max
= arch_scale_cpu_capacity(cpu
);
8074 unsigned long used
, free
;
8077 irq
= cpu_util_irq(rq
);
8079 if (unlikely(irq
>= max
))
8083 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8084 * (running and not running) with weights 0 and 1024 respectively.
8085 * avg_thermal.load_avg tracks thermal pressure and the weighted
8086 * average uses the actual delta max capacity(load).
8088 used
= READ_ONCE(rq
->avg_rt
.util_avg
);
8089 used
+= READ_ONCE(rq
->avg_dl
.util_avg
);
8090 used
+= thermal_load_avg(rq
);
8092 if (unlikely(used
>= max
))
8097 return scale_irq_capacity(free
, irq
, max
);
8100 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
8102 unsigned long capacity
= scale_rt_capacity(cpu
);
8103 struct sched_group
*sdg
= sd
->groups
;
8105 cpu_rq(cpu
)->cpu_capacity_orig
= arch_scale_cpu_capacity(cpu
);
8110 cpu_rq(cpu
)->cpu_capacity
= capacity
;
8111 trace_sched_cpu_capacity_tp(cpu_rq(cpu
));
8113 sdg
->sgc
->capacity
= capacity
;
8114 sdg
->sgc
->min_capacity
= capacity
;
8115 sdg
->sgc
->max_capacity
= capacity
;
8118 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
8120 struct sched_domain
*child
= sd
->child
;
8121 struct sched_group
*group
, *sdg
= sd
->groups
;
8122 unsigned long capacity
, min_capacity
, max_capacity
;
8123 unsigned long interval
;
8125 interval
= msecs_to_jiffies(sd
->balance_interval
);
8126 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
8127 sdg
->sgc
->next_update
= jiffies
+ interval
;
8130 update_cpu_capacity(sd
, cpu
);
8135 min_capacity
= ULONG_MAX
;
8138 if (child
->flags
& SD_OVERLAP
) {
8140 * SD_OVERLAP domains cannot assume that child groups
8141 * span the current group.
8144 for_each_cpu(cpu
, sched_group_span(sdg
)) {
8145 unsigned long cpu_cap
= capacity_of(cpu
);
8147 capacity
+= cpu_cap
;
8148 min_capacity
= min(cpu_cap
, min_capacity
);
8149 max_capacity
= max(cpu_cap
, max_capacity
);
8153 * !SD_OVERLAP domains can assume that child groups
8154 * span the current group.
8157 group
= child
->groups
;
8159 struct sched_group_capacity
*sgc
= group
->sgc
;
8161 capacity
+= sgc
->capacity
;
8162 min_capacity
= min(sgc
->min_capacity
, min_capacity
);
8163 max_capacity
= max(sgc
->max_capacity
, max_capacity
);
8164 group
= group
->next
;
8165 } while (group
!= child
->groups
);
8168 sdg
->sgc
->capacity
= capacity
;
8169 sdg
->sgc
->min_capacity
= min_capacity
;
8170 sdg
->sgc
->max_capacity
= max_capacity
;
8174 * Check whether the capacity of the rq has been noticeably reduced by side
8175 * activity. The imbalance_pct is used for the threshold.
8176 * Return true is the capacity is reduced
8179 check_cpu_capacity(struct rq
*rq
, struct sched_domain
*sd
)
8181 return ((rq
->cpu_capacity
* sd
->imbalance_pct
) <
8182 (rq
->cpu_capacity_orig
* 100));
8186 * Check whether a rq has a misfit task and if it looks like we can actually
8187 * help that task: we can migrate the task to a CPU of higher capacity, or
8188 * the task's current CPU is heavily pressured.
8190 static inline int check_misfit_status(struct rq
*rq
, struct sched_domain
*sd
)
8192 return rq
->misfit_task_load
&&
8193 (rq
->cpu_capacity_orig
< rq
->rd
->max_cpu_capacity
||
8194 check_cpu_capacity(rq
, sd
));
8198 * Group imbalance indicates (and tries to solve) the problem where balancing
8199 * groups is inadequate due to ->cpus_ptr constraints.
8201 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8202 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8205 * { 0 1 2 3 } { 4 5 6 7 }
8208 * If we were to balance group-wise we'd place two tasks in the first group and
8209 * two tasks in the second group. Clearly this is undesired as it will overload
8210 * cpu 3 and leave one of the CPUs in the second group unused.
8212 * The current solution to this issue is detecting the skew in the first group
8213 * by noticing the lower domain failed to reach balance and had difficulty
8214 * moving tasks due to affinity constraints.
8216 * When this is so detected; this group becomes a candidate for busiest; see
8217 * update_sd_pick_busiest(). And calculate_imbalance() and
8218 * find_busiest_group() avoid some of the usual balance conditions to allow it
8219 * to create an effective group imbalance.
8221 * This is a somewhat tricky proposition since the next run might not find the
8222 * group imbalance and decide the groups need to be balanced again. A most
8223 * subtle and fragile situation.
8226 static inline int sg_imbalanced(struct sched_group
*group
)
8228 return group
->sgc
->imbalance
;
8232 * group_has_capacity returns true if the group has spare capacity that could
8233 * be used by some tasks.
8234 * We consider that a group has spare capacity if the * number of task is
8235 * smaller than the number of CPUs or if the utilization is lower than the
8236 * available capacity for CFS tasks.
8237 * For the latter, we use a threshold to stabilize the state, to take into
8238 * account the variance of the tasks' load and to return true if the available
8239 * capacity in meaningful for the load balancer.
8240 * As an example, an available capacity of 1% can appear but it doesn't make
8241 * any benefit for the load balance.
8244 group_has_capacity(unsigned int imbalance_pct
, struct sg_lb_stats
*sgs
)
8246 if (sgs
->sum_nr_running
< sgs
->group_weight
)
8249 if ((sgs
->group_capacity
* imbalance_pct
) <
8250 (sgs
->group_runnable
* 100))
8253 if ((sgs
->group_capacity
* 100) >
8254 (sgs
->group_util
* imbalance_pct
))
8261 * group_is_overloaded returns true if the group has more tasks than it can
8263 * group_is_overloaded is not equals to !group_has_capacity because a group
8264 * with the exact right number of tasks, has no more spare capacity but is not
8265 * overloaded so both group_has_capacity and group_is_overloaded return
8269 group_is_overloaded(unsigned int imbalance_pct
, struct sg_lb_stats
*sgs
)
8271 if (sgs
->sum_nr_running
<= sgs
->group_weight
)
8274 if ((sgs
->group_capacity
* 100) <
8275 (sgs
->group_util
* imbalance_pct
))
8278 if ((sgs
->group_capacity
* imbalance_pct
) <
8279 (sgs
->group_runnable
* 100))
8286 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
8287 * per-CPU capacity than sched_group ref.
8290 group_smaller_min_cpu_capacity(struct sched_group
*sg
, struct sched_group
*ref
)
8292 return fits_capacity(sg
->sgc
->min_capacity
, ref
->sgc
->min_capacity
);
8296 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
8297 * per-CPU capacity_orig than sched_group ref.
8300 group_smaller_max_cpu_capacity(struct sched_group
*sg
, struct sched_group
*ref
)
8302 return fits_capacity(sg
->sgc
->max_capacity
, ref
->sgc
->max_capacity
);
8306 group_type
group_classify(unsigned int imbalance_pct
,
8307 struct sched_group
*group
,
8308 struct sg_lb_stats
*sgs
)
8310 if (group_is_overloaded(imbalance_pct
, sgs
))
8311 return group_overloaded
;
8313 if (sg_imbalanced(group
))
8314 return group_imbalanced
;
8316 if (sgs
->group_asym_packing
)
8317 return group_asym_packing
;
8319 if (sgs
->group_misfit_task_load
)
8320 return group_misfit_task
;
8322 if (!group_has_capacity(imbalance_pct
, sgs
))
8323 return group_fully_busy
;
8325 return group_has_spare
;
8328 static bool update_nohz_stats(struct rq
*rq
, bool force
)
8330 #ifdef CONFIG_NO_HZ_COMMON
8331 unsigned int cpu
= rq
->cpu
;
8333 if (!rq
->has_blocked_load
)
8336 if (!cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))
8339 if (!force
&& !time_after(jiffies
, rq
->last_blocked_load_update_tick
))
8342 update_blocked_averages(cpu
);
8344 return rq
->has_blocked_load
;
8351 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8352 * @env: The load balancing environment.
8353 * @group: sched_group whose statistics are to be updated.
8354 * @sgs: variable to hold the statistics for this group.
8355 * @sg_status: Holds flag indicating the status of the sched_group
8357 static inline void update_sg_lb_stats(struct lb_env
*env
,
8358 struct sched_group
*group
,
8359 struct sg_lb_stats
*sgs
,
8362 int i
, nr_running
, local_group
;
8364 memset(sgs
, 0, sizeof(*sgs
));
8366 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_span(group
));
8368 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
8369 struct rq
*rq
= cpu_rq(i
);
8371 if ((env
->flags
& LBF_NOHZ_STATS
) && update_nohz_stats(rq
, false))
8372 env
->flags
|= LBF_NOHZ_AGAIN
;
8374 sgs
->group_load
+= cpu_load(rq
);
8375 sgs
->group_util
+= cpu_util(i
);
8376 sgs
->group_runnable
+= cpu_runnable(rq
);
8377 sgs
->sum_h_nr_running
+= rq
->cfs
.h_nr_running
;
8379 nr_running
= rq
->nr_running
;
8380 sgs
->sum_nr_running
+= nr_running
;
8383 *sg_status
|= SG_OVERLOAD
;
8385 if (cpu_overutilized(i
))
8386 *sg_status
|= SG_OVERUTILIZED
;
8388 #ifdef CONFIG_NUMA_BALANCING
8389 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
8390 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
8393 * No need to call idle_cpu() if nr_running is not 0
8395 if (!nr_running
&& idle_cpu(i
)) {
8397 /* Idle cpu can't have misfit task */
8404 /* Check for a misfit task on the cpu */
8405 if (env
->sd
->flags
& SD_ASYM_CPUCAPACITY
&&
8406 sgs
->group_misfit_task_load
< rq
->misfit_task_load
) {
8407 sgs
->group_misfit_task_load
= rq
->misfit_task_load
;
8408 *sg_status
|= SG_OVERLOAD
;
8412 /* Check if dst CPU is idle and preferred to this group */
8413 if (env
->sd
->flags
& SD_ASYM_PACKING
&&
8414 env
->idle
!= CPU_NOT_IDLE
&&
8415 sgs
->sum_h_nr_running
&&
8416 sched_asym_prefer(env
->dst_cpu
, group
->asym_prefer_cpu
)) {
8417 sgs
->group_asym_packing
= 1;
8420 sgs
->group_capacity
= group
->sgc
->capacity
;
8422 sgs
->group_weight
= group
->group_weight
;
8424 sgs
->group_type
= group_classify(env
->sd
->imbalance_pct
, group
, sgs
);
8426 /* Computing avg_load makes sense only when group is overloaded */
8427 if (sgs
->group_type
== group_overloaded
)
8428 sgs
->avg_load
= (sgs
->group_load
* SCHED_CAPACITY_SCALE
) /
8429 sgs
->group_capacity
;
8433 * update_sd_pick_busiest - return 1 on busiest group
8434 * @env: The load balancing environment.
8435 * @sds: sched_domain statistics
8436 * @sg: sched_group candidate to be checked for being the busiest
8437 * @sgs: sched_group statistics
8439 * Determine if @sg is a busier group than the previously selected
8442 * Return: %true if @sg is a busier group than the previously selected
8443 * busiest group. %false otherwise.
8445 static bool update_sd_pick_busiest(struct lb_env
*env
,
8446 struct sd_lb_stats
*sds
,
8447 struct sched_group
*sg
,
8448 struct sg_lb_stats
*sgs
)
8450 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
8452 /* Make sure that there is at least one task to pull */
8453 if (!sgs
->sum_h_nr_running
)
8457 * Don't try to pull misfit tasks we can't help.
8458 * We can use max_capacity here as reduction in capacity on some
8459 * CPUs in the group should either be possible to resolve
8460 * internally or be covered by avg_load imbalance (eventually).
8462 if (sgs
->group_type
== group_misfit_task
&&
8463 (!group_smaller_max_cpu_capacity(sg
, sds
->local
) ||
8464 sds
->local_stat
.group_type
!= group_has_spare
))
8467 if (sgs
->group_type
> busiest
->group_type
)
8470 if (sgs
->group_type
< busiest
->group_type
)
8474 * The candidate and the current busiest group are the same type of
8475 * group. Let check which one is the busiest according to the type.
8478 switch (sgs
->group_type
) {
8479 case group_overloaded
:
8480 /* Select the overloaded group with highest avg_load. */
8481 if (sgs
->avg_load
<= busiest
->avg_load
)
8485 case group_imbalanced
:
8487 * Select the 1st imbalanced group as we don't have any way to
8488 * choose one more than another.
8492 case group_asym_packing
:
8493 /* Prefer to move from lowest priority CPU's work */
8494 if (sched_asym_prefer(sg
->asym_prefer_cpu
, sds
->busiest
->asym_prefer_cpu
))
8498 case group_misfit_task
:
8500 * If we have more than one misfit sg go with the biggest
8503 if (sgs
->group_misfit_task_load
< busiest
->group_misfit_task_load
)
8507 case group_fully_busy
:
8509 * Select the fully busy group with highest avg_load. In
8510 * theory, there is no need to pull task from such kind of
8511 * group because tasks have all compute capacity that they need
8512 * but we can still improve the overall throughput by reducing
8513 * contention when accessing shared HW resources.
8515 * XXX for now avg_load is not computed and always 0 so we
8516 * select the 1st one.
8518 if (sgs
->avg_load
<= busiest
->avg_load
)
8522 case group_has_spare
:
8524 * Select not overloaded group with lowest number of idle cpus
8525 * and highest number of running tasks. We could also compare
8526 * the spare capacity which is more stable but it can end up
8527 * that the group has less spare capacity but finally more idle
8528 * CPUs which means less opportunity to pull tasks.
8530 if (sgs
->idle_cpus
> busiest
->idle_cpus
)
8532 else if ((sgs
->idle_cpus
== busiest
->idle_cpus
) &&
8533 (sgs
->sum_nr_running
<= busiest
->sum_nr_running
))
8540 * Candidate sg has no more than one task per CPU and has higher
8541 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8542 * throughput. Maximize throughput, power/energy consequences are not
8545 if ((env
->sd
->flags
& SD_ASYM_CPUCAPACITY
) &&
8546 (sgs
->group_type
<= group_fully_busy
) &&
8547 (group_smaller_min_cpu_capacity(sds
->local
, sg
)))
8553 #ifdef CONFIG_NUMA_BALANCING
8554 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
8556 if (sgs
->sum_h_nr_running
> sgs
->nr_numa_running
)
8558 if (sgs
->sum_h_nr_running
> sgs
->nr_preferred_running
)
8563 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
8565 if (rq
->nr_running
> rq
->nr_numa_running
)
8567 if (rq
->nr_running
> rq
->nr_preferred_running
)
8572 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
8577 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
8581 #endif /* CONFIG_NUMA_BALANCING */
8587 * task_running_on_cpu - return 1 if @p is running on @cpu.
8590 static unsigned int task_running_on_cpu(int cpu
, struct task_struct
*p
)
8592 /* Task has no contribution or is new */
8593 if (cpu
!= task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
8596 if (task_on_rq_queued(p
))
8603 * idle_cpu_without - would a given CPU be idle without p ?
8604 * @cpu: the processor on which idleness is tested.
8605 * @p: task which should be ignored.
8607 * Return: 1 if the CPU would be idle. 0 otherwise.
8609 static int idle_cpu_without(int cpu
, struct task_struct
*p
)
8611 struct rq
*rq
= cpu_rq(cpu
);
8613 if (rq
->curr
!= rq
->idle
&& rq
->curr
!= p
)
8617 * rq->nr_running can't be used but an updated version without the
8618 * impact of p on cpu must be used instead. The updated nr_running
8619 * be computed and tested before calling idle_cpu_without().
8623 if (rq
->ttwu_pending
)
8631 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8632 * @sd: The sched_domain level to look for idlest group.
8633 * @group: sched_group whose statistics are to be updated.
8634 * @sgs: variable to hold the statistics for this group.
8635 * @p: The task for which we look for the idlest group/CPU.
8637 static inline void update_sg_wakeup_stats(struct sched_domain
*sd
,
8638 struct sched_group
*group
,
8639 struct sg_lb_stats
*sgs
,
8640 struct task_struct
*p
)
8644 memset(sgs
, 0, sizeof(*sgs
));
8646 for_each_cpu(i
, sched_group_span(group
)) {
8647 struct rq
*rq
= cpu_rq(i
);
8650 sgs
->group_load
+= cpu_load_without(rq
, p
);
8651 sgs
->group_util
+= cpu_util_without(i
, p
);
8652 sgs
->group_runnable
+= cpu_runnable_without(rq
, p
);
8653 local
= task_running_on_cpu(i
, p
);
8654 sgs
->sum_h_nr_running
+= rq
->cfs
.h_nr_running
- local
;
8656 nr_running
= rq
->nr_running
- local
;
8657 sgs
->sum_nr_running
+= nr_running
;
8660 * No need to call idle_cpu_without() if nr_running is not 0
8662 if (!nr_running
&& idle_cpu_without(i
, p
))
8667 /* Check if task fits in the group */
8668 if (sd
->flags
& SD_ASYM_CPUCAPACITY
&&
8669 !task_fits_capacity(p
, group
->sgc
->max_capacity
)) {
8670 sgs
->group_misfit_task_load
= 1;
8673 sgs
->group_capacity
= group
->sgc
->capacity
;
8675 sgs
->group_weight
= group
->group_weight
;
8677 sgs
->group_type
= group_classify(sd
->imbalance_pct
, group
, sgs
);
8680 * Computing avg_load makes sense only when group is fully busy or
8683 if (sgs
->group_type
== group_fully_busy
||
8684 sgs
->group_type
== group_overloaded
)
8685 sgs
->avg_load
= (sgs
->group_load
* SCHED_CAPACITY_SCALE
) /
8686 sgs
->group_capacity
;
8689 static bool update_pick_idlest(struct sched_group
*idlest
,
8690 struct sg_lb_stats
*idlest_sgs
,
8691 struct sched_group
*group
,
8692 struct sg_lb_stats
*sgs
)
8694 if (sgs
->group_type
< idlest_sgs
->group_type
)
8697 if (sgs
->group_type
> idlest_sgs
->group_type
)
8701 * The candidate and the current idlest group are the same type of
8702 * group. Let check which one is the idlest according to the type.
8705 switch (sgs
->group_type
) {
8706 case group_overloaded
:
8707 case group_fully_busy
:
8708 /* Select the group with lowest avg_load. */
8709 if (idlest_sgs
->avg_load
<= sgs
->avg_load
)
8713 case group_imbalanced
:
8714 case group_asym_packing
:
8715 /* Those types are not used in the slow wakeup path */
8718 case group_misfit_task
:
8719 /* Select group with the highest max capacity */
8720 if (idlest
->sgc
->max_capacity
>= group
->sgc
->max_capacity
)
8724 case group_has_spare
:
8725 /* Select group with most idle CPUs */
8726 if (idlest_sgs
->idle_cpus
> sgs
->idle_cpus
)
8729 /* Select group with lowest group_util */
8730 if (idlest_sgs
->idle_cpus
== sgs
->idle_cpus
&&
8731 idlest_sgs
->group_util
<= sgs
->group_util
)
8741 * find_idlest_group() finds and returns the least busy CPU group within the
8744 * Assumes p is allowed on at least one CPU in sd.
8746 static struct sched_group
*
8747 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
8749 struct sched_group
*idlest
= NULL
, *local
= NULL
, *group
= sd
->groups
;
8750 struct sg_lb_stats local_sgs
, tmp_sgs
;
8751 struct sg_lb_stats
*sgs
;
8752 unsigned long imbalance
;
8753 struct sg_lb_stats idlest_sgs
= {
8754 .avg_load
= UINT_MAX
,
8755 .group_type
= group_overloaded
,
8758 imbalance
= scale_load_down(NICE_0_LOAD
) *
8759 (sd
->imbalance_pct
-100) / 100;
8764 /* Skip over this group if it has no CPUs allowed */
8765 if (!cpumask_intersects(sched_group_span(group
),
8769 local_group
= cpumask_test_cpu(this_cpu
,
8770 sched_group_span(group
));
8779 update_sg_wakeup_stats(sd
, group
, sgs
, p
);
8781 if (!local_group
&& update_pick_idlest(idlest
, &idlest_sgs
, group
, sgs
)) {
8786 } while (group
= group
->next
, group
!= sd
->groups
);
8789 /* There is no idlest group to push tasks to */
8793 /* The local group has been skipped because of CPU affinity */
8798 * If the local group is idler than the selected idlest group
8799 * don't try and push the task.
8801 if (local_sgs
.group_type
< idlest_sgs
.group_type
)
8805 * If the local group is busier than the selected idlest group
8806 * try and push the task.
8808 if (local_sgs
.group_type
> idlest_sgs
.group_type
)
8811 switch (local_sgs
.group_type
) {
8812 case group_overloaded
:
8813 case group_fully_busy
:
8815 * When comparing groups across NUMA domains, it's possible for
8816 * the local domain to be very lightly loaded relative to the
8817 * remote domains but "imbalance" skews the comparison making
8818 * remote CPUs look much more favourable. When considering
8819 * cross-domain, add imbalance to the load on the remote node
8820 * and consider staying local.
8823 if ((sd
->flags
& SD_NUMA
) &&
8824 ((idlest_sgs
.avg_load
+ imbalance
) >= local_sgs
.avg_load
))
8828 * If the local group is less loaded than the selected
8829 * idlest group don't try and push any tasks.
8831 if (idlest_sgs
.avg_load
>= (local_sgs
.avg_load
+ imbalance
))
8834 if (100 * local_sgs
.avg_load
<= sd
->imbalance_pct
* idlest_sgs
.avg_load
)
8838 case group_imbalanced
:
8839 case group_asym_packing
:
8840 /* Those type are not used in the slow wakeup path */
8843 case group_misfit_task
:
8844 /* Select group with the highest max capacity */
8845 if (local
->sgc
->max_capacity
>= idlest
->sgc
->max_capacity
)
8849 case group_has_spare
:
8850 if (sd
->flags
& SD_NUMA
) {
8851 #ifdef CONFIG_NUMA_BALANCING
8854 * If there is spare capacity at NUMA, try to select
8855 * the preferred node
8857 if (cpu_to_node(this_cpu
) == p
->numa_preferred_nid
)
8860 idlest_cpu
= cpumask_first(sched_group_span(idlest
));
8861 if (cpu_to_node(idlest_cpu
) == p
->numa_preferred_nid
)
8865 * Otherwise, keep the task on this node to stay close
8866 * its wakeup source and improve locality. If there is
8867 * a real need of migration, periodic load balance will
8870 if (local_sgs
.idle_cpus
)
8875 * Select group with highest number of idle CPUs. We could also
8876 * compare the utilization which is more stable but it can end
8877 * up that the group has less spare capacity but finally more
8878 * idle CPUs which means more opportunity to run task.
8880 if (local_sgs
.idle_cpus
>= idlest_sgs
.idle_cpus
)
8889 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8890 * @env: The load balancing environment.
8891 * @sds: variable to hold the statistics for this sched_domain.
8894 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
8896 struct sched_domain
*child
= env
->sd
->child
;
8897 struct sched_group
*sg
= env
->sd
->groups
;
8898 struct sg_lb_stats
*local
= &sds
->local_stat
;
8899 struct sg_lb_stats tmp_sgs
;
8902 #ifdef CONFIG_NO_HZ_COMMON
8903 if (env
->idle
== CPU_NEWLY_IDLE
&& READ_ONCE(nohz
.has_blocked
))
8904 env
->flags
|= LBF_NOHZ_STATS
;
8908 struct sg_lb_stats
*sgs
= &tmp_sgs
;
8911 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_span(sg
));
8916 if (env
->idle
!= CPU_NEWLY_IDLE
||
8917 time_after_eq(jiffies
, sg
->sgc
->next_update
))
8918 update_group_capacity(env
->sd
, env
->dst_cpu
);
8921 update_sg_lb_stats(env
, sg
, sgs
, &sg_status
);
8927 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
8929 sds
->busiest_stat
= *sgs
;
8933 /* Now, start updating sd_lb_stats */
8934 sds
->total_load
+= sgs
->group_load
;
8935 sds
->total_capacity
+= sgs
->group_capacity
;
8938 } while (sg
!= env
->sd
->groups
);
8940 /* Tag domain that child domain prefers tasks go to siblings first */
8941 sds
->prefer_sibling
= child
&& child
->flags
& SD_PREFER_SIBLING
;
8943 #ifdef CONFIG_NO_HZ_COMMON
8944 if ((env
->flags
& LBF_NOHZ_AGAIN
) &&
8945 cpumask_subset(nohz
.idle_cpus_mask
, sched_domain_span(env
->sd
))) {
8947 WRITE_ONCE(nohz
.next_blocked
,
8948 jiffies
+ msecs_to_jiffies(LOAD_AVG_PERIOD
));
8952 if (env
->sd
->flags
& SD_NUMA
)
8953 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
8955 if (!env
->sd
->parent
) {
8956 struct root_domain
*rd
= env
->dst_rq
->rd
;
8958 /* update overload indicator if we are at root domain */
8959 WRITE_ONCE(rd
->overload
, sg_status
& SG_OVERLOAD
);
8961 /* Update over-utilization (tipping point, U >= 0) indicator */
8962 WRITE_ONCE(rd
->overutilized
, sg_status
& SG_OVERUTILIZED
);
8963 trace_sched_overutilized_tp(rd
, sg_status
& SG_OVERUTILIZED
);
8964 } else if (sg_status
& SG_OVERUTILIZED
) {
8965 struct root_domain
*rd
= env
->dst_rq
->rd
;
8967 WRITE_ONCE(rd
->overutilized
, SG_OVERUTILIZED
);
8968 trace_sched_overutilized_tp(rd
, SG_OVERUTILIZED
);
8972 static inline long adjust_numa_imbalance(int imbalance
, int nr_running
)
8974 unsigned int imbalance_min
;
8977 * Allow a small imbalance based on a simple pair of communicating
8978 * tasks that remain local when the source domain is almost idle.
8981 if (nr_running
<= imbalance_min
)
8988 * calculate_imbalance - Calculate the amount of imbalance present within the
8989 * groups of a given sched_domain during load balance.
8990 * @env: load balance environment
8991 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8993 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
8995 struct sg_lb_stats
*local
, *busiest
;
8997 local
= &sds
->local_stat
;
8998 busiest
= &sds
->busiest_stat
;
9000 if (busiest
->group_type
== group_misfit_task
) {
9001 /* Set imbalance to allow misfit tasks to be balanced. */
9002 env
->migration_type
= migrate_misfit
;
9007 if (busiest
->group_type
== group_asym_packing
) {
9009 * In case of asym capacity, we will try to migrate all load to
9010 * the preferred CPU.
9012 env
->migration_type
= migrate_task
;
9013 env
->imbalance
= busiest
->sum_h_nr_running
;
9017 if (busiest
->group_type
== group_imbalanced
) {
9019 * In the group_imb case we cannot rely on group-wide averages
9020 * to ensure CPU-load equilibrium, try to move any task to fix
9021 * the imbalance. The next load balance will take care of
9022 * balancing back the system.
9024 env
->migration_type
= migrate_task
;
9030 * Try to use spare capacity of local group without overloading it or
9033 if (local
->group_type
== group_has_spare
) {
9034 if (busiest
->group_type
> group_fully_busy
) {
9036 * If busiest is overloaded, try to fill spare
9037 * capacity. This might end up creating spare capacity
9038 * in busiest or busiest still being overloaded but
9039 * there is no simple way to directly compute the
9040 * amount of load to migrate in order to balance the
9043 env
->migration_type
= migrate_util
;
9044 env
->imbalance
= max(local
->group_capacity
, local
->group_util
) -
9048 * In some cases, the group's utilization is max or even
9049 * higher than capacity because of migrations but the
9050 * local CPU is (newly) idle. There is at least one
9051 * waiting task in this overloaded busiest group. Let's
9054 if (env
->idle
!= CPU_NOT_IDLE
&& env
->imbalance
== 0) {
9055 env
->migration_type
= migrate_task
;
9062 if (busiest
->group_weight
== 1 || sds
->prefer_sibling
) {
9063 unsigned int nr_diff
= busiest
->sum_nr_running
;
9065 * When prefer sibling, evenly spread running tasks on
9068 env
->migration_type
= migrate_task
;
9069 lsub_positive(&nr_diff
, local
->sum_nr_running
);
9070 env
->imbalance
= nr_diff
>> 1;
9074 * If there is no overload, we just want to even the number of
9077 env
->migration_type
= migrate_task
;
9078 env
->imbalance
= max_t(long, 0, (local
->idle_cpus
-
9079 busiest
->idle_cpus
) >> 1);
9082 /* Consider allowing a small imbalance between NUMA groups */
9083 if (env
->sd
->flags
& SD_NUMA
)
9084 env
->imbalance
= adjust_numa_imbalance(env
->imbalance
,
9085 busiest
->sum_nr_running
);
9091 * Local is fully busy but has to take more load to relieve the
9094 if (local
->group_type
< group_overloaded
) {
9096 * Local will become overloaded so the avg_load metrics are
9100 local
->avg_load
= (local
->group_load
* SCHED_CAPACITY_SCALE
) /
9101 local
->group_capacity
;
9103 sds
->avg_load
= (sds
->total_load
* SCHED_CAPACITY_SCALE
) /
9104 sds
->total_capacity
;
9106 * If the local group is more loaded than the selected
9107 * busiest group don't try to pull any tasks.
9109 if (local
->avg_load
>= busiest
->avg_load
) {
9116 * Both group are or will become overloaded and we're trying to get all
9117 * the CPUs to the average_load, so we don't want to push ourselves
9118 * above the average load, nor do we wish to reduce the max loaded CPU
9119 * below the average load. At the same time, we also don't want to
9120 * reduce the group load below the group capacity. Thus we look for
9121 * the minimum possible imbalance.
9123 env
->migration_type
= migrate_load
;
9124 env
->imbalance
= min(
9125 (busiest
->avg_load
- sds
->avg_load
) * busiest
->group_capacity
,
9126 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
9127 ) / SCHED_CAPACITY_SCALE
;
9130 /******* find_busiest_group() helpers end here *********************/
9133 * Decision matrix according to the local and busiest group type:
9135 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9136 * has_spare nr_idle balanced N/A N/A balanced balanced
9137 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9138 * misfit_task force N/A N/A N/A force force
9139 * asym_packing force force N/A N/A force force
9140 * imbalanced force force N/A N/A force force
9141 * overloaded force force N/A N/A force avg_load
9143 * N/A : Not Applicable because already filtered while updating
9145 * balanced : The system is balanced for these 2 groups.
9146 * force : Calculate the imbalance as load migration is probably needed.
9147 * avg_load : Only if imbalance is significant enough.
9148 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9149 * different in groups.
9153 * find_busiest_group - Returns the busiest group within the sched_domain
9154 * if there is an imbalance.
9156 * Also calculates the amount of runnable load which should be moved
9157 * to restore balance.
9159 * @env: The load balancing environment.
9161 * Return: - The busiest group if imbalance exists.
9163 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
9165 struct sg_lb_stats
*local
, *busiest
;
9166 struct sd_lb_stats sds
;
9168 init_sd_lb_stats(&sds
);
9171 * Compute the various statistics relevant for load balancing at
9174 update_sd_lb_stats(env
, &sds
);
9176 if (sched_energy_enabled()) {
9177 struct root_domain
*rd
= env
->dst_rq
->rd
;
9179 if (rcu_dereference(rd
->pd
) && !READ_ONCE(rd
->overutilized
))
9183 local
= &sds
.local_stat
;
9184 busiest
= &sds
.busiest_stat
;
9186 /* There is no busy sibling group to pull tasks from */
9190 /* Misfit tasks should be dealt with regardless of the avg load */
9191 if (busiest
->group_type
== group_misfit_task
)
9194 /* ASYM feature bypasses nice load balance check */
9195 if (busiest
->group_type
== group_asym_packing
)
9199 * If the busiest group is imbalanced the below checks don't
9200 * work because they assume all things are equal, which typically
9201 * isn't true due to cpus_ptr constraints and the like.
9203 if (busiest
->group_type
== group_imbalanced
)
9207 * If the local group is busier than the selected busiest group
9208 * don't try and pull any tasks.
9210 if (local
->group_type
> busiest
->group_type
)
9214 * When groups are overloaded, use the avg_load to ensure fairness
9217 if (local
->group_type
== group_overloaded
) {
9219 * If the local group is more loaded than the selected
9220 * busiest group don't try to pull any tasks.
9222 if (local
->avg_load
>= busiest
->avg_load
)
9225 /* XXX broken for overlapping NUMA groups */
9226 sds
.avg_load
= (sds
.total_load
* SCHED_CAPACITY_SCALE
) /
9230 * Don't pull any tasks if this group is already above the
9231 * domain average load.
9233 if (local
->avg_load
>= sds
.avg_load
)
9237 * If the busiest group is more loaded, use imbalance_pct to be
9240 if (100 * busiest
->avg_load
<=
9241 env
->sd
->imbalance_pct
* local
->avg_load
)
9245 /* Try to move all excess tasks to child's sibling domain */
9246 if (sds
.prefer_sibling
&& local
->group_type
== group_has_spare
&&
9247 busiest
->sum_nr_running
> local
->sum_nr_running
+ 1)
9250 if (busiest
->group_type
!= group_overloaded
) {
9251 if (env
->idle
== CPU_NOT_IDLE
)
9253 * If the busiest group is not overloaded (and as a
9254 * result the local one too) but this CPU is already
9255 * busy, let another idle CPU try to pull task.
9259 if (busiest
->group_weight
> 1 &&
9260 local
->idle_cpus
<= (busiest
->idle_cpus
+ 1))
9262 * If the busiest group is not overloaded
9263 * and there is no imbalance between this and busiest
9264 * group wrt idle CPUs, it is balanced. The imbalance
9265 * becomes significant if the diff is greater than 1
9266 * otherwise we might end up to just move the imbalance
9267 * on another group. Of course this applies only if
9268 * there is more than 1 CPU per group.
9272 if (busiest
->sum_h_nr_running
== 1)
9274 * busiest doesn't have any tasks waiting to run
9280 /* Looks like there is an imbalance. Compute it */
9281 calculate_imbalance(env
, &sds
);
9282 return env
->imbalance
? sds
.busiest
: NULL
;
9290 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9292 static struct rq
*find_busiest_queue(struct lb_env
*env
,
9293 struct sched_group
*group
)
9295 struct rq
*busiest
= NULL
, *rq
;
9296 unsigned long busiest_util
= 0, busiest_load
= 0, busiest_capacity
= 1;
9297 unsigned int busiest_nr
= 0;
9300 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
9301 unsigned long capacity
, load
, util
;
9302 unsigned int nr_running
;
9306 rt
= fbq_classify_rq(rq
);
9309 * We classify groups/runqueues into three groups:
9310 * - regular: there are !numa tasks
9311 * - remote: there are numa tasks that run on the 'wrong' node
9312 * - all: there is no distinction
9314 * In order to avoid migrating ideally placed numa tasks,
9315 * ignore those when there's better options.
9317 * If we ignore the actual busiest queue to migrate another
9318 * task, the next balance pass can still reduce the busiest
9319 * queue by moving tasks around inside the node.
9321 * If we cannot move enough load due to this classification
9322 * the next pass will adjust the group classification and
9323 * allow migration of more tasks.
9325 * Both cases only affect the total convergence complexity.
9327 if (rt
> env
->fbq_type
)
9330 capacity
= capacity_of(i
);
9331 nr_running
= rq
->cfs
.h_nr_running
;
9334 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9335 * eventually lead to active_balancing high->low capacity.
9336 * Higher per-CPU capacity is considered better than balancing
9339 if (env
->sd
->flags
& SD_ASYM_CPUCAPACITY
&&
9340 capacity_of(env
->dst_cpu
) < capacity
&&
9344 switch (env
->migration_type
) {
9347 * When comparing with load imbalance, use cpu_load()
9348 * which is not scaled with the CPU capacity.
9350 load
= cpu_load(rq
);
9352 if (nr_running
== 1 && load
> env
->imbalance
&&
9353 !check_cpu_capacity(rq
, env
->sd
))
9357 * For the load comparisons with the other CPUs,
9358 * consider the cpu_load() scaled with the CPU
9359 * capacity, so that the load can be moved away
9360 * from the CPU that is potentially running at a
9363 * Thus we're looking for max(load_i / capacity_i),
9364 * crosswise multiplication to rid ourselves of the
9365 * division works out to:
9366 * load_i * capacity_j > load_j * capacity_i;
9367 * where j is our previous maximum.
9369 if (load
* busiest_capacity
> busiest_load
* capacity
) {
9370 busiest_load
= load
;
9371 busiest_capacity
= capacity
;
9377 util
= cpu_util(cpu_of(rq
));
9380 * Don't try to pull utilization from a CPU with one
9381 * running task. Whatever its utilization, we will fail
9384 if (nr_running
<= 1)
9387 if (busiest_util
< util
) {
9388 busiest_util
= util
;
9394 if (busiest_nr
< nr_running
) {
9395 busiest_nr
= nr_running
;
9400 case migrate_misfit
:
9402 * For ASYM_CPUCAPACITY domains with misfit tasks we
9403 * simply seek the "biggest" misfit task.
9405 if (rq
->misfit_task_load
> busiest_load
) {
9406 busiest_load
= rq
->misfit_task_load
;
9419 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9420 * so long as it is large enough.
9422 #define MAX_PINNED_INTERVAL 512
9425 asym_active_balance(struct lb_env
*env
)
9428 * ASYM_PACKING needs to force migrate tasks from busy but
9429 * lower priority CPUs in order to pack all tasks in the
9430 * highest priority CPUs.
9432 return env
->idle
!= CPU_NOT_IDLE
&& (env
->sd
->flags
& SD_ASYM_PACKING
) &&
9433 sched_asym_prefer(env
->dst_cpu
, env
->src_cpu
);
9437 voluntary_active_balance(struct lb_env
*env
)
9439 struct sched_domain
*sd
= env
->sd
;
9441 if (asym_active_balance(env
))
9445 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
9446 * It's worth migrating the task if the src_cpu's capacity is reduced
9447 * because of other sched_class or IRQs if more capacity stays
9448 * available on dst_cpu.
9450 if ((env
->idle
!= CPU_NOT_IDLE
) &&
9451 (env
->src_rq
->cfs
.h_nr_running
== 1)) {
9452 if ((check_cpu_capacity(env
->src_rq
, sd
)) &&
9453 (capacity_of(env
->src_cpu
)*sd
->imbalance_pct
< capacity_of(env
->dst_cpu
)*100))
9457 if (env
->migration_type
== migrate_misfit
)
9463 static int need_active_balance(struct lb_env
*env
)
9465 struct sched_domain
*sd
= env
->sd
;
9467 if (voluntary_active_balance(env
))
9470 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
9473 static int active_load_balance_cpu_stop(void *data
);
9475 static int should_we_balance(struct lb_env
*env
)
9477 struct sched_group
*sg
= env
->sd
->groups
;
9481 * Ensure the balancing environment is consistent; can happen
9482 * when the softirq triggers 'during' hotplug.
9484 if (!cpumask_test_cpu(env
->dst_cpu
, env
->cpus
))
9488 * In the newly idle case, we will allow all the CPUs
9489 * to do the newly idle load balance.
9491 if (env
->idle
== CPU_NEWLY_IDLE
)
9494 /* Try to find first idle CPU */
9495 for_each_cpu_and(cpu
, group_balance_mask(sg
), env
->cpus
) {
9499 /* Are we the first idle CPU? */
9500 return cpu
== env
->dst_cpu
;
9503 /* Are we the first CPU of this group ? */
9504 return group_balance_cpu(sg
) == env
->dst_cpu
;
9508 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9509 * tasks if there is an imbalance.
9511 static int load_balance(int this_cpu
, struct rq
*this_rq
,
9512 struct sched_domain
*sd
, enum cpu_idle_type idle
,
9513 int *continue_balancing
)
9515 int ld_moved
, cur_ld_moved
, active_balance
= 0;
9516 struct sched_domain
*sd_parent
= sd
->parent
;
9517 struct sched_group
*group
;
9520 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
9522 struct lb_env env
= {
9524 .dst_cpu
= this_cpu
,
9526 .dst_grpmask
= sched_group_span(sd
->groups
),
9528 .loop_break
= sched_nr_migrate_break
,
9531 .tasks
= LIST_HEAD_INIT(env
.tasks
),
9534 cpumask_and(cpus
, sched_domain_span(sd
), cpu_active_mask
);
9536 schedstat_inc(sd
->lb_count
[idle
]);
9539 if (!should_we_balance(&env
)) {
9540 *continue_balancing
= 0;
9544 group
= find_busiest_group(&env
);
9546 schedstat_inc(sd
->lb_nobusyg
[idle
]);
9550 busiest
= find_busiest_queue(&env
, group
);
9552 schedstat_inc(sd
->lb_nobusyq
[idle
]);
9556 BUG_ON(busiest
== env
.dst_rq
);
9558 schedstat_add(sd
->lb_imbalance
[idle
], env
.imbalance
);
9560 env
.src_cpu
= busiest
->cpu
;
9561 env
.src_rq
= busiest
;
9564 if (busiest
->nr_running
> 1) {
9566 * Attempt to move tasks. If find_busiest_group has found
9567 * an imbalance but busiest->nr_running <= 1, the group is
9568 * still unbalanced. ld_moved simply stays zero, so it is
9569 * correctly treated as an imbalance.
9571 env
.flags
|= LBF_ALL_PINNED
;
9572 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
9575 rq_lock_irqsave(busiest
, &rf
);
9576 update_rq_clock(busiest
);
9579 * cur_ld_moved - load moved in current iteration
9580 * ld_moved - cumulative load moved across iterations
9582 cur_ld_moved
= detach_tasks(&env
);
9585 * We've detached some tasks from busiest_rq. Every
9586 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9587 * unlock busiest->lock, and we are able to be sure
9588 * that nobody can manipulate the tasks in parallel.
9589 * See task_rq_lock() family for the details.
9592 rq_unlock(busiest
, &rf
);
9596 ld_moved
+= cur_ld_moved
;
9599 local_irq_restore(rf
.flags
);
9601 if (env
.flags
& LBF_NEED_BREAK
) {
9602 env
.flags
&= ~LBF_NEED_BREAK
;
9607 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9608 * us and move them to an alternate dst_cpu in our sched_group
9609 * where they can run. The upper limit on how many times we
9610 * iterate on same src_cpu is dependent on number of CPUs in our
9613 * This changes load balance semantics a bit on who can move
9614 * load to a given_cpu. In addition to the given_cpu itself
9615 * (or a ilb_cpu acting on its behalf where given_cpu is
9616 * nohz-idle), we now have balance_cpu in a position to move
9617 * load to given_cpu. In rare situations, this may cause
9618 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9619 * _independently_ and at _same_ time to move some load to
9620 * given_cpu) causing exceess load to be moved to given_cpu.
9621 * This however should not happen so much in practice and
9622 * moreover subsequent load balance cycles should correct the
9623 * excess load moved.
9625 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
9627 /* Prevent to re-select dst_cpu via env's CPUs */
9628 __cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
9630 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
9631 env
.dst_cpu
= env
.new_dst_cpu
;
9632 env
.flags
&= ~LBF_DST_PINNED
;
9634 env
.loop_break
= sched_nr_migrate_break
;
9637 * Go back to "more_balance" rather than "redo" since we
9638 * need to continue with same src_cpu.
9644 * We failed to reach balance because of affinity.
9647 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
9649 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
9650 *group_imbalance
= 1;
9653 /* All tasks on this runqueue were pinned by CPU affinity */
9654 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
9655 __cpumask_clear_cpu(cpu_of(busiest
), cpus
);
9657 * Attempting to continue load balancing at the current
9658 * sched_domain level only makes sense if there are
9659 * active CPUs remaining as possible busiest CPUs to
9660 * pull load from which are not contained within the
9661 * destination group that is receiving any migrated
9664 if (!cpumask_subset(cpus
, env
.dst_grpmask
)) {
9666 env
.loop_break
= sched_nr_migrate_break
;
9669 goto out_all_pinned
;
9674 schedstat_inc(sd
->lb_failed
[idle
]);
9676 * Increment the failure counter only on periodic balance.
9677 * We do not want newidle balance, which can be very
9678 * frequent, pollute the failure counter causing
9679 * excessive cache_hot migrations and active balances.
9681 if (idle
!= CPU_NEWLY_IDLE
)
9682 sd
->nr_balance_failed
++;
9684 if (need_active_balance(&env
)) {
9685 unsigned long flags
;
9687 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
9690 * Don't kick the active_load_balance_cpu_stop,
9691 * if the curr task on busiest CPU can't be
9692 * moved to this_cpu:
9694 if (!cpumask_test_cpu(this_cpu
, busiest
->curr
->cpus_ptr
)) {
9695 raw_spin_unlock_irqrestore(&busiest
->lock
,
9697 env
.flags
|= LBF_ALL_PINNED
;
9698 goto out_one_pinned
;
9702 * ->active_balance synchronizes accesses to
9703 * ->active_balance_work. Once set, it's cleared
9704 * only after active load balance is finished.
9706 if (!busiest
->active_balance
) {
9707 busiest
->active_balance
= 1;
9708 busiest
->push_cpu
= this_cpu
;
9711 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
9713 if (active_balance
) {
9714 stop_one_cpu_nowait(cpu_of(busiest
),
9715 active_load_balance_cpu_stop
, busiest
,
9716 &busiest
->active_balance_work
);
9719 /* We've kicked active balancing, force task migration. */
9720 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
9723 sd
->nr_balance_failed
= 0;
9725 if (likely(!active_balance
) || voluntary_active_balance(&env
)) {
9726 /* We were unbalanced, so reset the balancing interval */
9727 sd
->balance_interval
= sd
->min_interval
;
9730 * If we've begun active balancing, start to back off. This
9731 * case may not be covered by the all_pinned logic if there
9732 * is only 1 task on the busy runqueue (because we don't call
9735 if (sd
->balance_interval
< sd
->max_interval
)
9736 sd
->balance_interval
*= 2;
9743 * We reach balance although we may have faced some affinity
9744 * constraints. Clear the imbalance flag only if other tasks got
9745 * a chance to move and fix the imbalance.
9747 if (sd_parent
&& !(env
.flags
& LBF_ALL_PINNED
)) {
9748 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
9750 if (*group_imbalance
)
9751 *group_imbalance
= 0;
9756 * We reach balance because all tasks are pinned at this level so
9757 * we can't migrate them. Let the imbalance flag set so parent level
9758 * can try to migrate them.
9760 schedstat_inc(sd
->lb_balanced
[idle
]);
9762 sd
->nr_balance_failed
= 0;
9768 * newidle_balance() disregards balance intervals, so we could
9769 * repeatedly reach this code, which would lead to balance_interval
9770 * skyrocketting in a short amount of time. Skip the balance_interval
9771 * increase logic to avoid that.
9773 if (env
.idle
== CPU_NEWLY_IDLE
)
9776 /* tune up the balancing interval */
9777 if ((env
.flags
& LBF_ALL_PINNED
&&
9778 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
9779 sd
->balance_interval
< sd
->max_interval
)
9780 sd
->balance_interval
*= 2;
9785 static inline unsigned long
9786 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
9788 unsigned long interval
= sd
->balance_interval
;
9791 interval
*= sd
->busy_factor
;
9793 /* scale ms to jiffies */
9794 interval
= msecs_to_jiffies(interval
);
9797 * Reduce likelihood of busy balancing at higher domains racing with
9798 * balancing at lower domains by preventing their balancing periods
9799 * from being multiples of each other.
9804 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
9810 update_next_balance(struct sched_domain
*sd
, unsigned long *next_balance
)
9812 unsigned long interval
, next
;
9814 /* used by idle balance, so cpu_busy = 0 */
9815 interval
= get_sd_balance_interval(sd
, 0);
9816 next
= sd
->last_balance
+ interval
;
9818 if (time_after(*next_balance
, next
))
9819 *next_balance
= next
;
9823 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9824 * running tasks off the busiest CPU onto idle CPUs. It requires at
9825 * least 1 task to be running on each physical CPU where possible, and
9826 * avoids physical / logical imbalances.
9828 static int active_load_balance_cpu_stop(void *data
)
9830 struct rq
*busiest_rq
= data
;
9831 int busiest_cpu
= cpu_of(busiest_rq
);
9832 int target_cpu
= busiest_rq
->push_cpu
;
9833 struct rq
*target_rq
= cpu_rq(target_cpu
);
9834 struct sched_domain
*sd
;
9835 struct task_struct
*p
= NULL
;
9838 rq_lock_irq(busiest_rq
, &rf
);
9840 * Between queueing the stop-work and running it is a hole in which
9841 * CPUs can become inactive. We should not move tasks from or to
9844 if (!cpu_active(busiest_cpu
) || !cpu_active(target_cpu
))
9847 /* Make sure the requested CPU hasn't gone down in the meantime: */
9848 if (unlikely(busiest_cpu
!= smp_processor_id() ||
9849 !busiest_rq
->active_balance
))
9852 /* Is there any task to move? */
9853 if (busiest_rq
->nr_running
<= 1)
9857 * This condition is "impossible", if it occurs
9858 * we need to fix it. Originally reported by
9859 * Bjorn Helgaas on a 128-CPU setup.
9861 BUG_ON(busiest_rq
== target_rq
);
9863 /* Search for an sd spanning us and the target CPU. */
9865 for_each_domain(target_cpu
, sd
) {
9866 if (cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
9871 struct lb_env env
= {
9873 .dst_cpu
= target_cpu
,
9874 .dst_rq
= target_rq
,
9875 .src_cpu
= busiest_rq
->cpu
,
9876 .src_rq
= busiest_rq
,
9879 * can_migrate_task() doesn't need to compute new_dst_cpu
9880 * for active balancing. Since we have CPU_IDLE, but no
9881 * @dst_grpmask we need to make that test go away with lying
9884 .flags
= LBF_DST_PINNED
,
9887 schedstat_inc(sd
->alb_count
);
9888 update_rq_clock(busiest_rq
);
9890 p
= detach_one_task(&env
);
9892 schedstat_inc(sd
->alb_pushed
);
9893 /* Active balancing done, reset the failure counter. */
9894 sd
->nr_balance_failed
= 0;
9896 schedstat_inc(sd
->alb_failed
);
9901 busiest_rq
->active_balance
= 0;
9902 rq_unlock(busiest_rq
, &rf
);
9905 attach_one_task(target_rq
, p
);
9912 static DEFINE_SPINLOCK(balancing
);
9915 * Scale the max load_balance interval with the number of CPUs in the system.
9916 * This trades load-balance latency on larger machines for less cross talk.
9918 void update_max_interval(void)
9920 max_load_balance_interval
= HZ
*num_online_cpus()/10;
9924 * It checks each scheduling domain to see if it is due to be balanced,
9925 * and initiates a balancing operation if so.
9927 * Balancing parameters are set up in init_sched_domains.
9929 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
9931 int continue_balancing
= 1;
9933 int busy
= idle
!= CPU_IDLE
&& !sched_idle_cpu(cpu
);
9934 unsigned long interval
;
9935 struct sched_domain
*sd
;
9936 /* Earliest time when we have to do rebalance again */
9937 unsigned long next_balance
= jiffies
+ 60*HZ
;
9938 int update_next_balance
= 0;
9939 int need_serialize
, need_decay
= 0;
9943 for_each_domain(cpu
, sd
) {
9945 * Decay the newidle max times here because this is a regular
9946 * visit to all the domains. Decay ~1% per second.
9948 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
9949 sd
->max_newidle_lb_cost
=
9950 (sd
->max_newidle_lb_cost
* 253) / 256;
9951 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
9954 max_cost
+= sd
->max_newidle_lb_cost
;
9957 * Stop the load balance at this level. There is another
9958 * CPU in our sched group which is doing load balancing more
9961 if (!continue_balancing
) {
9967 interval
= get_sd_balance_interval(sd
, busy
);
9969 need_serialize
= sd
->flags
& SD_SERIALIZE
;
9970 if (need_serialize
) {
9971 if (!spin_trylock(&balancing
))
9975 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
9976 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
9978 * The LBF_DST_PINNED logic could have changed
9979 * env->dst_cpu, so we can't know our idle
9980 * state even if we migrated tasks. Update it.
9982 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
9983 busy
= idle
!= CPU_IDLE
&& !sched_idle_cpu(cpu
);
9985 sd
->last_balance
= jiffies
;
9986 interval
= get_sd_balance_interval(sd
, busy
);
9989 spin_unlock(&balancing
);
9991 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
9992 next_balance
= sd
->last_balance
+ interval
;
9993 update_next_balance
= 1;
9998 * Ensure the rq-wide value also decays but keep it at a
9999 * reasonable floor to avoid funnies with rq->avg_idle.
10001 rq
->max_idle_balance_cost
=
10002 max((u64
)sysctl_sched_migration_cost
, max_cost
);
10007 * next_balance will be updated only when there is a need.
10008 * When the cpu is attached to null domain for ex, it will not be
10011 if (likely(update_next_balance
)) {
10012 rq
->next_balance
= next_balance
;
10014 #ifdef CONFIG_NO_HZ_COMMON
10016 * If this CPU has been elected to perform the nohz idle
10017 * balance. Other idle CPUs have already rebalanced with
10018 * nohz_idle_balance() and nohz.next_balance has been
10019 * updated accordingly. This CPU is now running the idle load
10020 * balance for itself and we need to update the
10021 * nohz.next_balance accordingly.
10023 if ((idle
== CPU_IDLE
) && time_after(nohz
.next_balance
, rq
->next_balance
))
10024 nohz
.next_balance
= rq
->next_balance
;
10029 static inline int on_null_domain(struct rq
*rq
)
10031 return unlikely(!rcu_dereference_sched(rq
->sd
));
10034 #ifdef CONFIG_NO_HZ_COMMON
10036 * idle load balancing details
10037 * - When one of the busy CPUs notice that there may be an idle rebalancing
10038 * needed, they will kick the idle load balancer, which then does idle
10039 * load balancing for all the idle CPUs.
10040 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
10044 static inline int find_new_ilb(void)
10048 for_each_cpu_and(ilb
, nohz
.idle_cpus_mask
,
10049 housekeeping_cpumask(HK_FLAG_MISC
)) {
10058 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
10059 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
10061 static void kick_ilb(unsigned int flags
)
10066 * Increase nohz.next_balance only when if full ilb is triggered but
10067 * not if we only update stats.
10069 if (flags
& NOHZ_BALANCE_KICK
)
10070 nohz
.next_balance
= jiffies
+1;
10072 ilb_cpu
= find_new_ilb();
10074 if (ilb_cpu
>= nr_cpu_ids
)
10078 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
10079 * the first flag owns it; cleared by nohz_csd_func().
10081 flags
= atomic_fetch_or(flags
, nohz_flags(ilb_cpu
));
10082 if (flags
& NOHZ_KICK_MASK
)
10086 * This way we generate an IPI on the target CPU which
10087 * is idle. And the softirq performing nohz idle load balance
10088 * will be run before returning from the IPI.
10090 smp_call_function_single_async(ilb_cpu
, &cpu_rq(ilb_cpu
)->nohz_csd
);
10094 * Current decision point for kicking the idle load balancer in the presence
10095 * of idle CPUs in the system.
10097 static void nohz_balancer_kick(struct rq
*rq
)
10099 unsigned long now
= jiffies
;
10100 struct sched_domain_shared
*sds
;
10101 struct sched_domain
*sd
;
10102 int nr_busy
, i
, cpu
= rq
->cpu
;
10103 unsigned int flags
= 0;
10105 if (unlikely(rq
->idle_balance
))
10109 * We may be recently in ticked or tickless idle mode. At the first
10110 * busy tick after returning from idle, we will update the busy stats.
10112 nohz_balance_exit_idle(rq
);
10115 * None are in tickless mode and hence no need for NOHZ idle load
10118 if (likely(!atomic_read(&nohz
.nr_cpus
)))
10121 if (READ_ONCE(nohz
.has_blocked
) &&
10122 time_after(now
, READ_ONCE(nohz
.next_blocked
)))
10123 flags
= NOHZ_STATS_KICK
;
10125 if (time_before(now
, nohz
.next_balance
))
10128 if (rq
->nr_running
>= 2) {
10129 flags
= NOHZ_KICK_MASK
;
10135 sd
= rcu_dereference(rq
->sd
);
10138 * If there's a CFS task and the current CPU has reduced
10139 * capacity; kick the ILB to see if there's a better CPU to run
10142 if (rq
->cfs
.h_nr_running
>= 1 && check_cpu_capacity(rq
, sd
)) {
10143 flags
= NOHZ_KICK_MASK
;
10148 sd
= rcu_dereference(per_cpu(sd_asym_packing
, cpu
));
10151 * When ASYM_PACKING; see if there's a more preferred CPU
10152 * currently idle; in which case, kick the ILB to move tasks
10155 for_each_cpu_and(i
, sched_domain_span(sd
), nohz
.idle_cpus_mask
) {
10156 if (sched_asym_prefer(i
, cpu
)) {
10157 flags
= NOHZ_KICK_MASK
;
10163 sd
= rcu_dereference(per_cpu(sd_asym_cpucapacity
, cpu
));
10166 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10167 * to run the misfit task on.
10169 if (check_misfit_status(rq
, sd
)) {
10170 flags
= NOHZ_KICK_MASK
;
10175 * For asymmetric systems, we do not want to nicely balance
10176 * cache use, instead we want to embrace asymmetry and only
10177 * ensure tasks have enough CPU capacity.
10179 * Skip the LLC logic because it's not relevant in that case.
10184 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
10187 * If there is an imbalance between LLC domains (IOW we could
10188 * increase the overall cache use), we need some less-loaded LLC
10189 * domain to pull some load. Likewise, we may need to spread
10190 * load within the current LLC domain (e.g. packed SMT cores but
10191 * other CPUs are idle). We can't really know from here how busy
10192 * the others are - so just get a nohz balance going if it looks
10193 * like this LLC domain has tasks we could move.
10195 nr_busy
= atomic_read(&sds
->nr_busy_cpus
);
10197 flags
= NOHZ_KICK_MASK
;
10208 static void set_cpu_sd_state_busy(int cpu
)
10210 struct sched_domain
*sd
;
10213 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
10215 if (!sd
|| !sd
->nohz_idle
)
10219 atomic_inc(&sd
->shared
->nr_busy_cpus
);
10224 void nohz_balance_exit_idle(struct rq
*rq
)
10226 SCHED_WARN_ON(rq
!= this_rq());
10228 if (likely(!rq
->nohz_tick_stopped
))
10231 rq
->nohz_tick_stopped
= 0;
10232 cpumask_clear_cpu(rq
->cpu
, nohz
.idle_cpus_mask
);
10233 atomic_dec(&nohz
.nr_cpus
);
10235 set_cpu_sd_state_busy(rq
->cpu
);
10238 static void set_cpu_sd_state_idle(int cpu
)
10240 struct sched_domain
*sd
;
10243 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
10245 if (!sd
|| sd
->nohz_idle
)
10249 atomic_dec(&sd
->shared
->nr_busy_cpus
);
10255 * This routine will record that the CPU is going idle with tick stopped.
10256 * This info will be used in performing idle load balancing in the future.
10258 void nohz_balance_enter_idle(int cpu
)
10260 struct rq
*rq
= cpu_rq(cpu
);
10262 SCHED_WARN_ON(cpu
!= smp_processor_id());
10264 /* If this CPU is going down, then nothing needs to be done: */
10265 if (!cpu_active(cpu
))
10268 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10269 if (!housekeeping_cpu(cpu
, HK_FLAG_SCHED
))
10273 * Can be set safely without rq->lock held
10274 * If a clear happens, it will have evaluated last additions because
10275 * rq->lock is held during the check and the clear
10277 rq
->has_blocked_load
= 1;
10280 * The tick is still stopped but load could have been added in the
10281 * meantime. We set the nohz.has_blocked flag to trig a check of the
10282 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10283 * of nohz.has_blocked can only happen after checking the new load
10285 if (rq
->nohz_tick_stopped
)
10288 /* If we're a completely isolated CPU, we don't play: */
10289 if (on_null_domain(rq
))
10292 rq
->nohz_tick_stopped
= 1;
10294 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
10295 atomic_inc(&nohz
.nr_cpus
);
10298 * Ensures that if nohz_idle_balance() fails to observe our
10299 * @idle_cpus_mask store, it must observe the @has_blocked
10302 smp_mb__after_atomic();
10304 set_cpu_sd_state_idle(cpu
);
10308 * Each time a cpu enter idle, we assume that it has blocked load and
10309 * enable the periodic update of the load of idle cpus
10311 WRITE_ONCE(nohz
.has_blocked
, 1);
10315 * Internal function that runs load balance for all idle cpus. The load balance
10316 * can be a simple update of blocked load or a complete load balance with
10317 * tasks movement depending of flags.
10318 * The function returns false if the loop has stopped before running
10319 * through all idle CPUs.
10321 static bool _nohz_idle_balance(struct rq
*this_rq
, unsigned int flags
,
10322 enum cpu_idle_type idle
)
10324 /* Earliest time when we have to do rebalance again */
10325 unsigned long now
= jiffies
;
10326 unsigned long next_balance
= now
+ 60*HZ
;
10327 bool has_blocked_load
= false;
10328 int update_next_balance
= 0;
10329 int this_cpu
= this_rq
->cpu
;
10334 SCHED_WARN_ON((flags
& NOHZ_KICK_MASK
) == NOHZ_BALANCE_KICK
);
10337 * We assume there will be no idle load after this update and clear
10338 * the has_blocked flag. If a cpu enters idle in the mean time, it will
10339 * set the has_blocked flag and trig another update of idle load.
10340 * Because a cpu that becomes idle, is added to idle_cpus_mask before
10341 * setting the flag, we are sure to not clear the state and not
10342 * check the load of an idle cpu.
10344 WRITE_ONCE(nohz
.has_blocked
, 0);
10347 * Ensures that if we miss the CPU, we must see the has_blocked
10348 * store from nohz_balance_enter_idle().
10352 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
10353 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
10357 * If this CPU gets work to do, stop the load balancing
10358 * work being done for other CPUs. Next load
10359 * balancing owner will pick it up.
10361 if (need_resched()) {
10362 has_blocked_load
= true;
10366 rq
= cpu_rq(balance_cpu
);
10368 has_blocked_load
|= update_nohz_stats(rq
, true);
10371 * If time for next balance is due,
10374 if (time_after_eq(jiffies
, rq
->next_balance
)) {
10375 struct rq_flags rf
;
10377 rq_lock_irqsave(rq
, &rf
);
10378 update_rq_clock(rq
);
10379 rq_unlock_irqrestore(rq
, &rf
);
10381 if (flags
& NOHZ_BALANCE_KICK
)
10382 rebalance_domains(rq
, CPU_IDLE
);
10385 if (time_after(next_balance
, rq
->next_balance
)) {
10386 next_balance
= rq
->next_balance
;
10387 update_next_balance
= 1;
10392 * next_balance will be updated only when there is a need.
10393 * When the CPU is attached to null domain for ex, it will not be
10396 if (likely(update_next_balance
))
10397 nohz
.next_balance
= next_balance
;
10399 /* Newly idle CPU doesn't need an update */
10400 if (idle
!= CPU_NEWLY_IDLE
) {
10401 update_blocked_averages(this_cpu
);
10402 has_blocked_load
|= this_rq
->has_blocked_load
;
10405 if (flags
& NOHZ_BALANCE_KICK
)
10406 rebalance_domains(this_rq
, CPU_IDLE
);
10408 WRITE_ONCE(nohz
.next_blocked
,
10409 now
+ msecs_to_jiffies(LOAD_AVG_PERIOD
));
10411 /* The full idle balance loop has been done */
10415 /* There is still blocked load, enable periodic update */
10416 if (has_blocked_load
)
10417 WRITE_ONCE(nohz
.has_blocked
, 1);
10423 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
10424 * rebalancing for all the cpus for whom scheduler ticks are stopped.
10426 static bool nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
10428 unsigned int flags
= this_rq
->nohz_idle_balance
;
10433 this_rq
->nohz_idle_balance
= 0;
10435 if (idle
!= CPU_IDLE
)
10438 _nohz_idle_balance(this_rq
, flags
, idle
);
10443 static void nohz_newidle_balance(struct rq
*this_rq
)
10445 int this_cpu
= this_rq
->cpu
;
10448 * This CPU doesn't want to be disturbed by scheduler
10451 if (!housekeeping_cpu(this_cpu
, HK_FLAG_SCHED
))
10454 /* Will wake up very soon. No time for doing anything else*/
10455 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
10458 /* Don't need to update blocked load of idle CPUs*/
10459 if (!READ_ONCE(nohz
.has_blocked
) ||
10460 time_before(jiffies
, READ_ONCE(nohz
.next_blocked
)))
10463 raw_spin_unlock(&this_rq
->lock
);
10465 * This CPU is going to be idle and blocked load of idle CPUs
10466 * need to be updated. Run the ilb locally as it is a good
10467 * candidate for ilb instead of waking up another idle CPU.
10468 * Kick an normal ilb if we failed to do the update.
10470 if (!_nohz_idle_balance(this_rq
, NOHZ_STATS_KICK
, CPU_NEWLY_IDLE
))
10471 kick_ilb(NOHZ_STATS_KICK
);
10472 raw_spin_lock(&this_rq
->lock
);
10475 #else /* !CONFIG_NO_HZ_COMMON */
10476 static inline void nohz_balancer_kick(struct rq
*rq
) { }
10478 static inline bool nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
10483 static inline void nohz_newidle_balance(struct rq
*this_rq
) { }
10484 #endif /* CONFIG_NO_HZ_COMMON */
10487 * idle_balance is called by schedule() if this_cpu is about to become
10488 * idle. Attempts to pull tasks from other CPUs.
10491 * < 0 - we released the lock and there are !fair tasks present
10492 * 0 - failed, no new tasks
10493 * > 0 - success, new (fair) tasks present
10495 static int newidle_balance(struct rq
*this_rq
, struct rq_flags
*rf
)
10497 unsigned long next_balance
= jiffies
+ HZ
;
10498 int this_cpu
= this_rq
->cpu
;
10499 struct sched_domain
*sd
;
10500 int pulled_task
= 0;
10503 update_misfit_status(NULL
, this_rq
);
10505 * We must set idle_stamp _before_ calling idle_balance(), such that we
10506 * measure the duration of idle_balance() as idle time.
10508 this_rq
->idle_stamp
= rq_clock(this_rq
);
10511 * Do not pull tasks towards !active CPUs...
10513 if (!cpu_active(this_cpu
))
10517 * This is OK, because current is on_cpu, which avoids it being picked
10518 * for load-balance and preemption/IRQs are still disabled avoiding
10519 * further scheduler activity on it and we're being very careful to
10520 * re-start the picking loop.
10522 rq_unpin_lock(this_rq
, rf
);
10524 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
10525 !READ_ONCE(this_rq
->rd
->overload
)) {
10528 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
10530 update_next_balance(sd
, &next_balance
);
10533 nohz_newidle_balance(this_rq
);
10538 raw_spin_unlock(&this_rq
->lock
);
10540 update_blocked_averages(this_cpu
);
10542 for_each_domain(this_cpu
, sd
) {
10543 int continue_balancing
= 1;
10544 u64 t0
, domain_cost
;
10546 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
10547 update_next_balance(sd
, &next_balance
);
10551 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
10552 t0
= sched_clock_cpu(this_cpu
);
10554 pulled_task
= load_balance(this_cpu
, this_rq
,
10555 sd
, CPU_NEWLY_IDLE
,
10556 &continue_balancing
);
10558 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
10559 if (domain_cost
> sd
->max_newidle_lb_cost
)
10560 sd
->max_newidle_lb_cost
= domain_cost
;
10562 curr_cost
+= domain_cost
;
10565 update_next_balance(sd
, &next_balance
);
10568 * Stop searching for tasks to pull if there are
10569 * now runnable tasks on this rq.
10571 if (pulled_task
|| this_rq
->nr_running
> 0)
10576 raw_spin_lock(&this_rq
->lock
);
10578 if (curr_cost
> this_rq
->max_idle_balance_cost
)
10579 this_rq
->max_idle_balance_cost
= curr_cost
;
10583 * While browsing the domains, we released the rq lock, a task could
10584 * have been enqueued in the meantime. Since we're not going idle,
10585 * pretend we pulled a task.
10587 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
10590 /* Move the next balance forward */
10591 if (time_after(this_rq
->next_balance
, next_balance
))
10592 this_rq
->next_balance
= next_balance
;
10594 /* Is there a task of a high priority class? */
10595 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
10599 this_rq
->idle_stamp
= 0;
10601 rq_repin_lock(this_rq
, rf
);
10603 return pulled_task
;
10607 * run_rebalance_domains is triggered when needed from the scheduler tick.
10608 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10610 static __latent_entropy
void run_rebalance_domains(struct softirq_action
*h
)
10612 struct rq
*this_rq
= this_rq();
10613 enum cpu_idle_type idle
= this_rq
->idle_balance
?
10614 CPU_IDLE
: CPU_NOT_IDLE
;
10617 * If this CPU has a pending nohz_balance_kick, then do the
10618 * balancing on behalf of the other idle CPUs whose ticks are
10619 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10620 * give the idle CPUs a chance to load balance. Else we may
10621 * load balance only within the local sched_domain hierarchy
10622 * and abort nohz_idle_balance altogether if we pull some load.
10624 if (nohz_idle_balance(this_rq
, idle
))
10627 /* normal load balance */
10628 update_blocked_averages(this_rq
->cpu
);
10629 rebalance_domains(this_rq
, idle
);
10633 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10635 void trigger_load_balance(struct rq
*rq
)
10637 /* Don't need to rebalance while attached to NULL domain */
10638 if (unlikely(on_null_domain(rq
)))
10641 if (time_after_eq(jiffies
, rq
->next_balance
))
10642 raise_softirq(SCHED_SOFTIRQ
);
10644 nohz_balancer_kick(rq
);
10647 static void rq_online_fair(struct rq
*rq
)
10651 update_runtime_enabled(rq
);
10654 static void rq_offline_fair(struct rq
*rq
)
10658 /* Ensure any throttled groups are reachable by pick_next_task */
10659 unthrottle_offline_cfs_rqs(rq
);
10662 #endif /* CONFIG_SMP */
10665 * scheduler tick hitting a task of our scheduling class.
10667 * NOTE: This function can be called remotely by the tick offload that
10668 * goes along full dynticks. Therefore no local assumption can be made
10669 * and everything must be accessed through the @rq and @curr passed in
10672 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
10674 struct cfs_rq
*cfs_rq
;
10675 struct sched_entity
*se
= &curr
->se
;
10677 for_each_sched_entity(se
) {
10678 cfs_rq
= cfs_rq_of(se
);
10679 entity_tick(cfs_rq
, se
, queued
);
10682 if (static_branch_unlikely(&sched_numa_balancing
))
10683 task_tick_numa(rq
, curr
);
10685 update_misfit_status(curr
, rq
);
10686 update_overutilized_status(task_rq(curr
));
10690 * called on fork with the child task as argument from the parent's context
10691 * - child not yet on the tasklist
10692 * - preemption disabled
10694 static void task_fork_fair(struct task_struct
*p
)
10696 struct cfs_rq
*cfs_rq
;
10697 struct sched_entity
*se
= &p
->se
, *curr
;
10698 struct rq
*rq
= this_rq();
10699 struct rq_flags rf
;
10702 update_rq_clock(rq
);
10704 cfs_rq
= task_cfs_rq(current
);
10705 curr
= cfs_rq
->curr
;
10707 update_curr(cfs_rq
);
10708 se
->vruntime
= curr
->vruntime
;
10710 place_entity(cfs_rq
, se
, 1);
10712 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
10714 * Upon rescheduling, sched_class::put_prev_task() will place
10715 * 'current' within the tree based on its new key value.
10717 swap(curr
->vruntime
, se
->vruntime
);
10721 se
->vruntime
-= cfs_rq
->min_vruntime
;
10722 rq_unlock(rq
, &rf
);
10726 * Priority of the task has changed. Check to see if we preempt
10727 * the current task.
10730 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
10732 if (!task_on_rq_queued(p
))
10735 if (rq
->cfs
.nr_running
== 1)
10739 * Reschedule if we are currently running on this runqueue and
10740 * our priority decreased, or if we are not currently running on
10741 * this runqueue and our priority is higher than the current's
10743 if (rq
->curr
== p
) {
10744 if (p
->prio
> oldprio
)
10747 check_preempt_curr(rq
, p
, 0);
10750 static inline bool vruntime_normalized(struct task_struct
*p
)
10752 struct sched_entity
*se
= &p
->se
;
10755 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10756 * the dequeue_entity(.flags=0) will already have normalized the
10763 * When !on_rq, vruntime of the task has usually NOT been normalized.
10764 * But there are some cases where it has already been normalized:
10766 * - A forked child which is waiting for being woken up by
10767 * wake_up_new_task().
10768 * - A task which has been woken up by try_to_wake_up() and
10769 * waiting for actually being woken up by sched_ttwu_pending().
10771 if (!se
->sum_exec_runtime
||
10772 (p
->state
== TASK_WAKING
&& p
->sched_remote_wakeup
))
10778 #ifdef CONFIG_FAIR_GROUP_SCHED
10780 * Propagate the changes of the sched_entity across the tg tree to make it
10781 * visible to the root
10783 static void propagate_entity_cfs_rq(struct sched_entity
*se
)
10785 struct cfs_rq
*cfs_rq
;
10787 /* Start to propagate at parent */
10790 for_each_sched_entity(se
) {
10791 cfs_rq
= cfs_rq_of(se
);
10793 if (cfs_rq_throttled(cfs_rq
))
10796 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
10800 static void propagate_entity_cfs_rq(struct sched_entity
*se
) { }
10803 static void detach_entity_cfs_rq(struct sched_entity
*se
)
10805 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10807 /* Catch up with the cfs_rq and remove our load when we leave */
10808 update_load_avg(cfs_rq
, se
, 0);
10809 detach_entity_load_avg(cfs_rq
, se
);
10810 update_tg_load_avg(cfs_rq
);
10811 propagate_entity_cfs_rq(se
);
10814 static void attach_entity_cfs_rq(struct sched_entity
*se
)
10816 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10818 #ifdef CONFIG_FAIR_GROUP_SCHED
10820 * Since the real-depth could have been changed (only FAIR
10821 * class maintain depth value), reset depth properly.
10823 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
10826 /* Synchronize entity with its cfs_rq */
10827 update_load_avg(cfs_rq
, se
, sched_feat(ATTACH_AGE_LOAD
) ? 0 : SKIP_AGE_LOAD
);
10828 attach_entity_load_avg(cfs_rq
, se
);
10829 update_tg_load_avg(cfs_rq
);
10830 propagate_entity_cfs_rq(se
);
10833 static void detach_task_cfs_rq(struct task_struct
*p
)
10835 struct sched_entity
*se
= &p
->se
;
10836 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10838 if (!vruntime_normalized(p
)) {
10840 * Fix up our vruntime so that the current sleep doesn't
10841 * cause 'unlimited' sleep bonus.
10843 place_entity(cfs_rq
, se
, 0);
10844 se
->vruntime
-= cfs_rq
->min_vruntime
;
10847 detach_entity_cfs_rq(se
);
10850 static void attach_task_cfs_rq(struct task_struct
*p
)
10852 struct sched_entity
*se
= &p
->se
;
10853 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10855 attach_entity_cfs_rq(se
);
10857 if (!vruntime_normalized(p
))
10858 se
->vruntime
+= cfs_rq
->min_vruntime
;
10861 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
10863 detach_task_cfs_rq(p
);
10866 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
10868 attach_task_cfs_rq(p
);
10870 if (task_on_rq_queued(p
)) {
10872 * We were most likely switched from sched_rt, so
10873 * kick off the schedule if running, otherwise just see
10874 * if we can still preempt the current task.
10879 check_preempt_curr(rq
, p
, 0);
10883 /* Account for a task changing its policy or group.
10885 * This routine is mostly called to set cfs_rq->curr field when a task
10886 * migrates between groups/classes.
10888 static void set_next_task_fair(struct rq
*rq
, struct task_struct
*p
, bool first
)
10890 struct sched_entity
*se
= &p
->se
;
10893 if (task_on_rq_queued(p
)) {
10895 * Move the next running task to the front of the list, so our
10896 * cfs_tasks list becomes MRU one.
10898 list_move(&se
->group_node
, &rq
->cfs_tasks
);
10902 for_each_sched_entity(se
) {
10903 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10905 set_next_entity(cfs_rq
, se
);
10906 /* ensure bandwidth has been allocated on our new cfs_rq */
10907 account_cfs_rq_runtime(cfs_rq
, 0);
10911 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
10913 cfs_rq
->tasks_timeline
= RB_ROOT_CACHED
;
10914 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
10915 #ifndef CONFIG_64BIT
10916 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
10919 raw_spin_lock_init(&cfs_rq
->removed
.lock
);
10923 #ifdef CONFIG_FAIR_GROUP_SCHED
10924 static void task_set_group_fair(struct task_struct
*p
)
10926 struct sched_entity
*se
= &p
->se
;
10928 set_task_rq(p
, task_cpu(p
));
10929 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
10932 static void task_move_group_fair(struct task_struct
*p
)
10934 detach_task_cfs_rq(p
);
10935 set_task_rq(p
, task_cpu(p
));
10938 /* Tell se's cfs_rq has been changed -- migrated */
10939 p
->se
.avg
.last_update_time
= 0;
10941 attach_task_cfs_rq(p
);
10944 static void task_change_group_fair(struct task_struct
*p
, int type
)
10947 case TASK_SET_GROUP
:
10948 task_set_group_fair(p
);
10951 case TASK_MOVE_GROUP
:
10952 task_move_group_fair(p
);
10957 void free_fair_sched_group(struct task_group
*tg
)
10961 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
10963 for_each_possible_cpu(i
) {
10965 kfree(tg
->cfs_rq
[i
]);
10974 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
10976 struct sched_entity
*se
;
10977 struct cfs_rq
*cfs_rq
;
10980 tg
->cfs_rq
= kcalloc(nr_cpu_ids
, sizeof(cfs_rq
), GFP_KERNEL
);
10983 tg
->se
= kcalloc(nr_cpu_ids
, sizeof(se
), GFP_KERNEL
);
10987 tg
->shares
= NICE_0_LOAD
;
10989 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
10991 for_each_possible_cpu(i
) {
10992 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
10993 GFP_KERNEL
, cpu_to_node(i
));
10997 se
= kzalloc_node(sizeof(struct sched_entity
),
10998 GFP_KERNEL
, cpu_to_node(i
));
11002 init_cfs_rq(cfs_rq
);
11003 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
11004 init_entity_runnable_average(se
);
11015 void online_fair_sched_group(struct task_group
*tg
)
11017 struct sched_entity
*se
;
11018 struct rq_flags rf
;
11022 for_each_possible_cpu(i
) {
11025 rq_lock_irq(rq
, &rf
);
11026 update_rq_clock(rq
);
11027 attach_entity_cfs_rq(se
);
11028 sync_throttle(tg
, i
);
11029 rq_unlock_irq(rq
, &rf
);
11033 void unregister_fair_sched_group(struct task_group
*tg
)
11035 unsigned long flags
;
11039 for_each_possible_cpu(cpu
) {
11041 remove_entity_load_avg(tg
->se
[cpu
]);
11044 * Only empty task groups can be destroyed; so we can speculatively
11045 * check on_list without danger of it being re-added.
11047 if (!tg
->cfs_rq
[cpu
]->on_list
)
11052 raw_spin_lock_irqsave(&rq
->lock
, flags
);
11053 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
11054 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
11058 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
11059 struct sched_entity
*se
, int cpu
,
11060 struct sched_entity
*parent
)
11062 struct rq
*rq
= cpu_rq(cpu
);
11066 init_cfs_rq_runtime(cfs_rq
);
11068 tg
->cfs_rq
[cpu
] = cfs_rq
;
11071 /* se could be NULL for root_task_group */
11076 se
->cfs_rq
= &rq
->cfs
;
11079 se
->cfs_rq
= parent
->my_q
;
11080 se
->depth
= parent
->depth
+ 1;
11084 /* guarantee group entities always have weight */
11085 update_load_set(&se
->load
, NICE_0_LOAD
);
11086 se
->parent
= parent
;
11089 static DEFINE_MUTEX(shares_mutex
);
11091 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
11096 * We can't change the weight of the root cgroup.
11101 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
11103 mutex_lock(&shares_mutex
);
11104 if (tg
->shares
== shares
)
11107 tg
->shares
= shares
;
11108 for_each_possible_cpu(i
) {
11109 struct rq
*rq
= cpu_rq(i
);
11110 struct sched_entity
*se
= tg
->se
[i
];
11111 struct rq_flags rf
;
11113 /* Propagate contribution to hierarchy */
11114 rq_lock_irqsave(rq
, &rf
);
11115 update_rq_clock(rq
);
11116 for_each_sched_entity(se
) {
11117 update_load_avg(cfs_rq_of(se
), se
, UPDATE_TG
);
11118 update_cfs_group(se
);
11120 rq_unlock_irqrestore(rq
, &rf
);
11124 mutex_unlock(&shares_mutex
);
11127 #else /* CONFIG_FAIR_GROUP_SCHED */
11129 void free_fair_sched_group(struct task_group
*tg
) { }
11131 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
11136 void online_fair_sched_group(struct task_group
*tg
) { }
11138 void unregister_fair_sched_group(struct task_group
*tg
) { }
11140 #endif /* CONFIG_FAIR_GROUP_SCHED */
11143 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
11145 struct sched_entity
*se
= &task
->se
;
11146 unsigned int rr_interval
= 0;
11149 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11152 if (rq
->cfs
.load
.weight
)
11153 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
11155 return rr_interval
;
11159 * All the scheduling class methods:
11161 const struct sched_class fair_sched_class
11162 __section("__fair_sched_class") = {
11163 .enqueue_task
= enqueue_task_fair
,
11164 .dequeue_task
= dequeue_task_fair
,
11165 .yield_task
= yield_task_fair
,
11166 .yield_to_task
= yield_to_task_fair
,
11168 .check_preempt_curr
= check_preempt_wakeup
,
11170 .pick_next_task
= __pick_next_task_fair
,
11171 .put_prev_task
= put_prev_task_fair
,
11172 .set_next_task
= set_next_task_fair
,
11175 .balance
= balance_fair
,
11176 .select_task_rq
= select_task_rq_fair
,
11177 .migrate_task_rq
= migrate_task_rq_fair
,
11179 .rq_online
= rq_online_fair
,
11180 .rq_offline
= rq_offline_fair
,
11182 .task_dead
= task_dead_fair
,
11183 .set_cpus_allowed
= set_cpus_allowed_common
,
11186 .task_tick
= task_tick_fair
,
11187 .task_fork
= task_fork_fair
,
11189 .prio_changed
= prio_changed_fair
,
11190 .switched_from
= switched_from_fair
,
11191 .switched_to
= switched_to_fair
,
11193 .get_rr_interval
= get_rr_interval_fair
,
11195 .update_curr
= update_curr_fair
,
11197 #ifdef CONFIG_FAIR_GROUP_SCHED
11198 .task_change_group
= task_change_group_fair
,
11201 #ifdef CONFIG_UCLAMP_TASK
11202 .uclamp_enabled
= 1,
11206 #ifdef CONFIG_SCHED_DEBUG
11207 void print_cfs_stats(struct seq_file
*m
, int cpu
)
11209 struct cfs_rq
*cfs_rq
, *pos
;
11212 for_each_leaf_cfs_rq_safe(cpu_rq(cpu
), cfs_rq
, pos
)
11213 print_cfs_rq(m
, cpu
, cfs_rq
);
11217 #ifdef CONFIG_NUMA_BALANCING
11218 void show_numa_stats(struct task_struct
*p
, struct seq_file
*m
)
11221 unsigned long tsf
= 0, tpf
= 0, gsf
= 0, gpf
= 0;
11222 struct numa_group
*ng
;
11225 ng
= rcu_dereference(p
->numa_group
);
11226 for_each_online_node(node
) {
11227 if (p
->numa_faults
) {
11228 tsf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
11229 tpf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
11232 gsf
= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)],
11233 gpf
= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
11235 print_numa_stats(m
, node
, tsf
, tpf
, gsf
, gpf
);
11239 #endif /* CONFIG_NUMA_BALANCING */
11240 #endif /* CONFIG_SCHED_DEBUG */
11242 __init
void init_sched_fair_class(void)
11245 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
11247 #ifdef CONFIG_NO_HZ_COMMON
11248 nohz
.next_balance
= jiffies
;
11249 nohz
.next_blocked
= jiffies
;
11250 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
11257 * Helper functions to facilitate extracting info from tracepoints.
11260 const struct sched_avg
*sched_trace_cfs_rq_avg(struct cfs_rq
*cfs_rq
)
11263 return cfs_rq
? &cfs_rq
->avg
: NULL
;
11268 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg
);
11270 char *sched_trace_cfs_rq_path(struct cfs_rq
*cfs_rq
, char *str
, int len
)
11274 strlcpy(str
, "(null)", len
);
11279 cfs_rq_tg_path(cfs_rq
, str
, len
);
11282 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path
);
11284 int sched_trace_cfs_rq_cpu(struct cfs_rq
*cfs_rq
)
11286 return cfs_rq
? cpu_of(rq_of(cfs_rq
)) : -1;
11288 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu
);
11290 const struct sched_avg
*sched_trace_rq_avg_rt(struct rq
*rq
)
11293 return rq
? &rq
->avg_rt
: NULL
;
11298 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt
);
11300 const struct sched_avg
*sched_trace_rq_avg_dl(struct rq
*rq
)
11303 return rq
? &rq
->avg_dl
: NULL
;
11308 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl
);
11310 const struct sched_avg
*sched_trace_rq_avg_irq(struct rq
*rq
)
11312 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
11313 return rq
? &rq
->avg_irq
: NULL
;
11318 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq
);
11320 int sched_trace_rq_cpu(struct rq
*rq
)
11322 return rq
? cpu_of(rq
) : -1;
11324 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu
);
11326 int sched_trace_rq_cpu_capacity(struct rq
*rq
)
11332 SCHED_CAPACITY_SCALE
11336 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity
);
11338 const struct cpumask
*sched_trace_rd_span(struct root_domain
*rd
)
11341 return rd
? rd
->span
: NULL
;
11346 EXPORT_SYMBOL_GPL(sched_trace_rd_span
);
11348 int sched_trace_rq_nr_running(struct rq
*rq
)
11350 return rq
? rq
->nr_running
: -1;
11352 EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running
);