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
25 #include <trace/events/sched.h>
28 * Targeted preemption latency for CPU-bound tasks:
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
40 unsigned int sysctl_sched_latency
= 6000000ULL;
41 static unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
44 * The initial- and re-scaling of tunables is configurable
48 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
49 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
50 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
52 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
54 enum sched_tunable_scaling sysctl_sched_tunable_scaling
= SCHED_TUNABLESCALING_LOG
;
57 * Minimal preemption granularity for CPU-bound tasks:
59 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
61 unsigned int sysctl_sched_min_granularity
= 750000ULL;
62 static unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
65 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
67 static unsigned int sched_nr_latency
= 8;
70 * After fork, child runs first. If set to 0 (default) then
71 * parent will (try to) run first.
73 unsigned int sysctl_sched_child_runs_first __read_mostly
;
76 * SCHED_OTHER wake-up granularity.
78 * This option delays the preemption effects of decoupled workloads
79 * and reduces their over-scheduling. Synchronous workloads will still
80 * have immediate wakeup/sleep latencies.
82 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
84 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
85 static unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
87 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
91 * For asym packing, by default the lower numbered CPU has higher priority.
93 int __weak
arch_asym_cpu_priority(int cpu
)
99 * The margin used when comparing utilization with CPU capacity:
100 * util * margin < capacity * 1024
104 static unsigned int capacity_margin
= 1280;
107 #ifdef CONFIG_CFS_BANDWIDTH
109 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
110 * each time a cfs_rq requests quota.
112 * Note: in the case that the slice exceeds the runtime remaining (either due
113 * to consumption or the quota being specified to be smaller than the slice)
114 * we will always only issue the remaining available time.
116 * (default: 5 msec, units: microseconds)
118 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
121 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
127 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
133 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
140 * Increase the granularity value when there are more CPUs,
141 * because with more CPUs the 'effective latency' as visible
142 * to users decreases. But the relationship is not linear,
143 * so pick a second-best guess by going with the log2 of the
146 * This idea comes from the SD scheduler of Con Kolivas:
148 static unsigned int get_update_sysctl_factor(void)
150 unsigned int cpus
= min_t(unsigned int, num_online_cpus(), 8);
153 switch (sysctl_sched_tunable_scaling
) {
154 case SCHED_TUNABLESCALING_NONE
:
157 case SCHED_TUNABLESCALING_LINEAR
:
160 case SCHED_TUNABLESCALING_LOG
:
162 factor
= 1 + ilog2(cpus
);
169 static void update_sysctl(void)
171 unsigned int factor
= get_update_sysctl_factor();
173 #define SET_SYSCTL(name) \
174 (sysctl_##name = (factor) * normalized_sysctl_##name)
175 SET_SYSCTL(sched_min_granularity
);
176 SET_SYSCTL(sched_latency
);
177 SET_SYSCTL(sched_wakeup_granularity
);
181 void sched_init_granularity(void)
186 #define WMULT_CONST (~0U)
187 #define WMULT_SHIFT 32
189 static void __update_inv_weight(struct load_weight
*lw
)
193 if (likely(lw
->inv_weight
))
196 w
= scale_load_down(lw
->weight
);
198 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
200 else if (unlikely(!w
))
201 lw
->inv_weight
= WMULT_CONST
;
203 lw
->inv_weight
= WMULT_CONST
/ w
;
207 * delta_exec * weight / lw.weight
209 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
211 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
212 * we're guaranteed shift stays positive because inv_weight is guaranteed to
213 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
215 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
216 * weight/lw.weight <= 1, and therefore our shift will also be positive.
218 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
220 u64 fact
= scale_load_down(weight
);
221 int shift
= WMULT_SHIFT
;
223 __update_inv_weight(lw
);
225 if (unlikely(fact
>> 32)) {
232 /* hint to use a 32x32->64 mul */
233 fact
= (u64
)(u32
)fact
* lw
->inv_weight
;
240 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
244 const struct sched_class fair_sched_class
;
246 /**************************************************************
247 * CFS operations on generic schedulable entities:
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 static inline struct task_struct
*task_of(struct sched_entity
*se
)
253 SCHED_WARN_ON(!entity_is_task(se
));
254 return container_of(se
, struct task_struct
, se
);
257 /* Walk up scheduling entities hierarchy */
258 #define for_each_sched_entity(se) \
259 for (; se; se = se->parent)
261 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
266 /* runqueue on which this entity is (to be) queued */
267 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
272 /* runqueue "owned" by this group */
273 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
278 static inline bool list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
280 struct rq
*rq
= rq_of(cfs_rq
);
281 int cpu
= cpu_of(rq
);
284 return rq
->tmp_alone_branch
== &rq
->leaf_cfs_rq_list
;
289 * Ensure we either appear before our parent (if already
290 * enqueued) or force our parent to appear after us when it is
291 * enqueued. The fact that we always enqueue bottom-up
292 * reduces this to two cases and a special case for the root
293 * cfs_rq. Furthermore, it also means that we will always reset
294 * tmp_alone_branch either when the branch is connected
295 * to a tree or when we reach the top of the tree
297 if (cfs_rq
->tg
->parent
&&
298 cfs_rq
->tg
->parent
->cfs_rq
[cpu
]->on_list
) {
300 * If parent is already on the list, we add the child
301 * just before. Thanks to circular linked property of
302 * the list, this means to put the child at the tail
303 * of the list that starts by parent.
305 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
306 &(cfs_rq
->tg
->parent
->cfs_rq
[cpu
]->leaf_cfs_rq_list
));
308 * The branch is now connected to its tree so we can
309 * reset tmp_alone_branch to the beginning of the
312 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
316 if (!cfs_rq
->tg
->parent
) {
318 * cfs rq without parent should be put
319 * at the tail of the list.
321 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
322 &rq
->leaf_cfs_rq_list
);
324 * We have reach the top of a tree so we can reset
325 * tmp_alone_branch to the beginning of the list.
327 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
332 * The parent has not already been added so we want to
333 * make sure that it will be put after us.
334 * tmp_alone_branch points to the begin of the branch
335 * where we will add parent.
337 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, rq
->tmp_alone_branch
);
339 * update tmp_alone_branch to points to the new begin
342 rq
->tmp_alone_branch
= &cfs_rq
->leaf_cfs_rq_list
;
346 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
348 if (cfs_rq
->on_list
) {
349 struct rq
*rq
= rq_of(cfs_rq
);
352 * With cfs_rq being unthrottled/throttled during an enqueue,
353 * it can happen the tmp_alone_branch points the a leaf that
354 * we finally want to del. In this case, tmp_alone_branch moves
355 * to the prev element but it will point to rq->leaf_cfs_rq_list
356 * at the end of the enqueue.
358 if (rq
->tmp_alone_branch
== &cfs_rq
->leaf_cfs_rq_list
)
359 rq
->tmp_alone_branch
= cfs_rq
->leaf_cfs_rq_list
.prev
;
361 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
366 static inline void assert_list_leaf_cfs_rq(struct rq
*rq
)
368 SCHED_WARN_ON(rq
->tmp_alone_branch
!= &rq
->leaf_cfs_rq_list
);
371 /* Iterate thr' all leaf cfs_rq's on a runqueue */
372 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
373 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
376 /* Do the two (enqueued) entities belong to the same group ? */
377 static inline struct cfs_rq
*
378 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
380 if (se
->cfs_rq
== pse
->cfs_rq
)
386 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
392 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
394 int se_depth
, pse_depth
;
397 * preemption test can be made between sibling entities who are in the
398 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
399 * both tasks until we find their ancestors who are siblings of common
403 /* First walk up until both entities are at same depth */
404 se_depth
= (*se
)->depth
;
405 pse_depth
= (*pse
)->depth
;
407 while (se_depth
> pse_depth
) {
409 *se
= parent_entity(*se
);
412 while (pse_depth
> se_depth
) {
414 *pse
= parent_entity(*pse
);
417 while (!is_same_group(*se
, *pse
)) {
418 *se
= parent_entity(*se
);
419 *pse
= parent_entity(*pse
);
423 #else /* !CONFIG_FAIR_GROUP_SCHED */
425 static inline struct task_struct
*task_of(struct sched_entity
*se
)
427 return container_of(se
, struct task_struct
, se
);
430 #define for_each_sched_entity(se) \
431 for (; se; se = NULL)
433 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
435 return &task_rq(p
)->cfs
;
438 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
440 struct task_struct
*p
= task_of(se
);
441 struct rq
*rq
= task_rq(p
);
446 /* runqueue "owned" by this group */
447 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
452 static inline bool list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
457 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
461 static inline void assert_list_leaf_cfs_rq(struct rq
*rq
)
465 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
466 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
468 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
474 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
478 #endif /* CONFIG_FAIR_GROUP_SCHED */
480 static __always_inline
481 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
483 /**************************************************************
484 * Scheduling class tree data structure manipulation methods:
487 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
489 s64 delta
= (s64
)(vruntime
- max_vruntime
);
491 max_vruntime
= vruntime
;
496 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
498 s64 delta
= (s64
)(vruntime
- min_vruntime
);
500 min_vruntime
= vruntime
;
505 static inline int entity_before(struct sched_entity
*a
,
506 struct sched_entity
*b
)
508 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
511 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
513 struct sched_entity
*curr
= cfs_rq
->curr
;
514 struct rb_node
*leftmost
= rb_first_cached(&cfs_rq
->tasks_timeline
);
516 u64 vruntime
= cfs_rq
->min_vruntime
;
520 vruntime
= curr
->vruntime
;
525 if (leftmost
) { /* non-empty tree */
526 struct sched_entity
*se
;
527 se
= rb_entry(leftmost
, struct sched_entity
, run_node
);
530 vruntime
= se
->vruntime
;
532 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
535 /* ensure we never gain time by being placed backwards. */
536 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
539 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
544 * Enqueue an entity into the rb-tree:
546 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
548 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_root
.rb_node
;
549 struct rb_node
*parent
= NULL
;
550 struct sched_entity
*entry
;
551 bool leftmost
= true;
554 * Find the right place in the rbtree:
558 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
560 * We dont care about collisions. Nodes with
561 * the same key stay together.
563 if (entity_before(se
, entry
)) {
564 link
= &parent
->rb_left
;
566 link
= &parent
->rb_right
;
571 rb_link_node(&se
->run_node
, parent
, link
);
572 rb_insert_color_cached(&se
->run_node
,
573 &cfs_rq
->tasks_timeline
, leftmost
);
576 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
578 rb_erase_cached(&se
->run_node
, &cfs_rq
->tasks_timeline
);
581 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
583 struct rb_node
*left
= rb_first_cached(&cfs_rq
->tasks_timeline
);
588 return rb_entry(left
, struct sched_entity
, run_node
);
591 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
593 struct rb_node
*next
= rb_next(&se
->run_node
);
598 return rb_entry(next
, struct sched_entity
, run_node
);
601 #ifdef CONFIG_SCHED_DEBUG
602 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
604 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
.rb_root
);
609 return rb_entry(last
, struct sched_entity
, run_node
);
612 /**************************************************************
613 * Scheduling class statistics methods:
616 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
617 void __user
*buffer
, size_t *lenp
,
620 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
621 unsigned int factor
= get_update_sysctl_factor();
626 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
627 sysctl_sched_min_granularity
);
629 #define WRT_SYSCTL(name) \
630 (normalized_sysctl_##name = sysctl_##name / (factor))
631 WRT_SYSCTL(sched_min_granularity
);
632 WRT_SYSCTL(sched_latency
);
633 WRT_SYSCTL(sched_wakeup_granularity
);
643 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
645 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
646 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
652 * The idea is to set a period in which each task runs once.
654 * When there are too many tasks (sched_nr_latency) we have to stretch
655 * this period because otherwise the slices get too small.
657 * p = (nr <= nl) ? l : l*nr/nl
659 static u64
__sched_period(unsigned long nr_running
)
661 if (unlikely(nr_running
> sched_nr_latency
))
662 return nr_running
* sysctl_sched_min_granularity
;
664 return sysctl_sched_latency
;
668 * We calculate the wall-time slice from the period by taking a part
669 * proportional to the weight.
673 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
675 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
677 for_each_sched_entity(se
) {
678 struct load_weight
*load
;
679 struct load_weight lw
;
681 cfs_rq
= cfs_rq_of(se
);
682 load
= &cfs_rq
->load
;
684 if (unlikely(!se
->on_rq
)) {
687 update_load_add(&lw
, se
->load
.weight
);
690 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
696 * We calculate the vruntime slice of a to-be-inserted task.
700 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
702 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
708 static int select_idle_sibling(struct task_struct
*p
, int prev_cpu
, int cpu
);
709 static unsigned long task_h_load(struct task_struct
*p
);
710 static unsigned long capacity_of(int cpu
);
712 /* Give new sched_entity start runnable values to heavy its load in infant time */
713 void init_entity_runnable_average(struct sched_entity
*se
)
715 struct sched_avg
*sa
= &se
->avg
;
717 memset(sa
, 0, sizeof(*sa
));
720 * Tasks are initialized with full load to be seen as heavy tasks until
721 * they get a chance to stabilize to their real load level.
722 * Group entities are initialized with zero load to reflect the fact that
723 * nothing has been attached to the task group yet.
725 if (entity_is_task(se
))
726 sa
->runnable_load_avg
= sa
->load_avg
= scale_load_down(se
->load
.weight
);
728 se
->runnable_weight
= se
->load
.weight
;
730 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
733 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
734 static void attach_entity_cfs_rq(struct sched_entity
*se
);
737 * With new tasks being created, their initial util_avgs are extrapolated
738 * based on the cfs_rq's current util_avg:
740 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
742 * However, in many cases, the above util_avg does not give a desired
743 * value. Moreover, the sum of the util_avgs may be divergent, such
744 * as when the series is a harmonic series.
746 * To solve this problem, we also cap the util_avg of successive tasks to
747 * only 1/2 of the left utilization budget:
749 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
751 * where n denotes the nth task and cpu_scale the CPU capacity.
753 * For example, for a CPU with 1024 of capacity, a simplest series from
754 * the beginning would be like:
756 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
757 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
759 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
760 * if util_avg > util_avg_cap.
762 void post_init_entity_util_avg(struct task_struct
*p
)
764 struct sched_entity
*se
= &p
->se
;
765 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
766 struct sched_avg
*sa
= &se
->avg
;
767 long cpu_scale
= arch_scale_cpu_capacity(NULL
, cpu_of(rq_of(cfs_rq
)));
768 long cap
= (long)(cpu_scale
- cfs_rq
->avg
.util_avg
) / 2;
771 if (cfs_rq
->avg
.util_avg
!= 0) {
772 sa
->util_avg
= cfs_rq
->avg
.util_avg
* se
->load
.weight
;
773 sa
->util_avg
/= (cfs_rq
->avg
.load_avg
+ 1);
775 if (sa
->util_avg
> cap
)
782 if (p
->sched_class
!= &fair_sched_class
) {
784 * For !fair tasks do:
786 update_cfs_rq_load_avg(now, cfs_rq);
787 attach_entity_load_avg(cfs_rq, se, 0);
788 switched_from_fair(rq, p);
790 * such that the next switched_to_fair() has the
793 se
->avg
.last_update_time
= cfs_rq_clock_pelt(cfs_rq
);
797 attach_entity_cfs_rq(se
);
800 #else /* !CONFIG_SMP */
801 void init_entity_runnable_average(struct sched_entity
*se
)
804 void post_init_entity_util_avg(struct task_struct
*p
)
807 static void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
)
810 #endif /* CONFIG_SMP */
813 * Update the current task's runtime statistics.
815 static void update_curr(struct cfs_rq
*cfs_rq
)
817 struct sched_entity
*curr
= cfs_rq
->curr
;
818 u64 now
= rq_clock_task(rq_of(cfs_rq
));
824 delta_exec
= now
- curr
->exec_start
;
825 if (unlikely((s64
)delta_exec
<= 0))
828 curr
->exec_start
= now
;
830 schedstat_set(curr
->statistics
.exec_max
,
831 max(delta_exec
, curr
->statistics
.exec_max
));
833 curr
->sum_exec_runtime
+= delta_exec
;
834 schedstat_add(cfs_rq
->exec_clock
, delta_exec
);
836 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
837 update_min_vruntime(cfs_rq
);
839 if (entity_is_task(curr
)) {
840 struct task_struct
*curtask
= task_of(curr
);
842 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
843 cgroup_account_cputime(curtask
, delta_exec
);
844 account_group_exec_runtime(curtask
, delta_exec
);
847 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
850 static void update_curr_fair(struct rq
*rq
)
852 update_curr(cfs_rq_of(&rq
->curr
->se
));
856 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
858 u64 wait_start
, prev_wait_start
;
860 if (!schedstat_enabled())
863 wait_start
= rq_clock(rq_of(cfs_rq
));
864 prev_wait_start
= schedstat_val(se
->statistics
.wait_start
);
866 if (entity_is_task(se
) && task_on_rq_migrating(task_of(se
)) &&
867 likely(wait_start
> prev_wait_start
))
868 wait_start
-= prev_wait_start
;
870 __schedstat_set(se
->statistics
.wait_start
, wait_start
);
874 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
876 struct task_struct
*p
;
879 if (!schedstat_enabled())
882 delta
= rq_clock(rq_of(cfs_rq
)) - schedstat_val(se
->statistics
.wait_start
);
884 if (entity_is_task(se
)) {
886 if (task_on_rq_migrating(p
)) {
888 * Preserve migrating task's wait time so wait_start
889 * time stamp can be adjusted to accumulate wait time
890 * prior to migration.
892 __schedstat_set(se
->statistics
.wait_start
, delta
);
895 trace_sched_stat_wait(p
, delta
);
898 __schedstat_set(se
->statistics
.wait_max
,
899 max(schedstat_val(se
->statistics
.wait_max
), delta
));
900 __schedstat_inc(se
->statistics
.wait_count
);
901 __schedstat_add(se
->statistics
.wait_sum
, delta
);
902 __schedstat_set(se
->statistics
.wait_start
, 0);
906 update_stats_enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
908 struct task_struct
*tsk
= NULL
;
909 u64 sleep_start
, block_start
;
911 if (!schedstat_enabled())
914 sleep_start
= schedstat_val(se
->statistics
.sleep_start
);
915 block_start
= schedstat_val(se
->statistics
.block_start
);
917 if (entity_is_task(se
))
921 u64 delta
= rq_clock(rq_of(cfs_rq
)) - sleep_start
;
926 if (unlikely(delta
> schedstat_val(se
->statistics
.sleep_max
)))
927 __schedstat_set(se
->statistics
.sleep_max
, delta
);
929 __schedstat_set(se
->statistics
.sleep_start
, 0);
930 __schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
933 account_scheduler_latency(tsk
, delta
>> 10, 1);
934 trace_sched_stat_sleep(tsk
, delta
);
938 u64 delta
= rq_clock(rq_of(cfs_rq
)) - block_start
;
943 if (unlikely(delta
> schedstat_val(se
->statistics
.block_max
)))
944 __schedstat_set(se
->statistics
.block_max
, delta
);
946 __schedstat_set(se
->statistics
.block_start
, 0);
947 __schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
950 if (tsk
->in_iowait
) {
951 __schedstat_add(se
->statistics
.iowait_sum
, delta
);
952 __schedstat_inc(se
->statistics
.iowait_count
);
953 trace_sched_stat_iowait(tsk
, delta
);
956 trace_sched_stat_blocked(tsk
, delta
);
959 * Blocking time is in units of nanosecs, so shift by
960 * 20 to get a milliseconds-range estimation of the
961 * amount of time that the task spent sleeping:
963 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
964 profile_hits(SLEEP_PROFILING
,
965 (void *)get_wchan(tsk
),
968 account_scheduler_latency(tsk
, delta
>> 10, 0);
974 * Task is being enqueued - update stats:
977 update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
979 if (!schedstat_enabled())
983 * Are we enqueueing a waiting task? (for current tasks
984 * a dequeue/enqueue event is a NOP)
986 if (se
!= cfs_rq
->curr
)
987 update_stats_wait_start(cfs_rq
, se
);
989 if (flags
& ENQUEUE_WAKEUP
)
990 update_stats_enqueue_sleeper(cfs_rq
, se
);
994 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
997 if (!schedstat_enabled())
1001 * Mark the end of the wait period if dequeueing a
1004 if (se
!= cfs_rq
->curr
)
1005 update_stats_wait_end(cfs_rq
, se
);
1007 if ((flags
& DEQUEUE_SLEEP
) && entity_is_task(se
)) {
1008 struct task_struct
*tsk
= task_of(se
);
1010 if (tsk
->state
& TASK_INTERRUPTIBLE
)
1011 __schedstat_set(se
->statistics
.sleep_start
,
1012 rq_clock(rq_of(cfs_rq
)));
1013 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
1014 __schedstat_set(se
->statistics
.block_start
,
1015 rq_clock(rq_of(cfs_rq
)));
1020 * We are picking a new current task - update its stats:
1023 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1026 * We are starting a new run period:
1028 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
1031 /**************************************************
1032 * Scheduling class queueing methods:
1035 #ifdef CONFIG_NUMA_BALANCING
1037 * Approximate time to scan a full NUMA task in ms. The task scan period is
1038 * calculated based on the tasks virtual memory size and
1039 * numa_balancing_scan_size.
1041 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
1042 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
1044 /* Portion of address space to scan in MB */
1045 unsigned int sysctl_numa_balancing_scan_size
= 256;
1047 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1048 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
1051 refcount_t refcount
;
1053 spinlock_t lock
; /* nr_tasks, tasks */
1058 struct rcu_head rcu
;
1059 unsigned long total_faults
;
1060 unsigned long max_faults_cpu
;
1062 * Faults_cpu is used to decide whether memory should move
1063 * towards the CPU. As a consequence, these stats are weighted
1064 * more by CPU use than by memory faults.
1066 unsigned long *faults_cpu
;
1067 unsigned long faults
[0];
1070 static inline unsigned long group_faults_priv(struct numa_group
*ng
);
1071 static inline unsigned long group_faults_shared(struct numa_group
*ng
);
1073 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
1075 unsigned long rss
= 0;
1076 unsigned long nr_scan_pages
;
1079 * Calculations based on RSS as non-present and empty pages are skipped
1080 * by the PTE scanner and NUMA hinting faults should be trapped based
1083 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
1084 rss
= get_mm_rss(p
->mm
);
1086 rss
= nr_scan_pages
;
1088 rss
= round_up(rss
, nr_scan_pages
);
1089 return rss
/ nr_scan_pages
;
1092 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1093 #define MAX_SCAN_WINDOW 2560
1095 static unsigned int task_scan_min(struct task_struct
*p
)
1097 unsigned int scan_size
= READ_ONCE(sysctl_numa_balancing_scan_size
);
1098 unsigned int scan
, floor
;
1099 unsigned int windows
= 1;
1101 if (scan_size
< MAX_SCAN_WINDOW
)
1102 windows
= MAX_SCAN_WINDOW
/ scan_size
;
1103 floor
= 1000 / windows
;
1105 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
1106 return max_t(unsigned int, floor
, scan
);
1109 static unsigned int task_scan_start(struct task_struct
*p
)
1111 unsigned long smin
= task_scan_min(p
);
1112 unsigned long period
= smin
;
1114 /* Scale the maximum scan period with the amount of shared memory. */
1115 if (p
->numa_group
) {
1116 struct numa_group
*ng
= p
->numa_group
;
1117 unsigned long shared
= group_faults_shared(ng
);
1118 unsigned long private = group_faults_priv(ng
);
1120 period
*= refcount_read(&ng
->refcount
);
1121 period
*= shared
+ 1;
1122 period
/= private + shared
+ 1;
1125 return max(smin
, period
);
1128 static unsigned int task_scan_max(struct task_struct
*p
)
1130 unsigned long smin
= task_scan_min(p
);
1133 /* Watch for min being lower than max due to floor calculations */
1134 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
1136 /* Scale the maximum scan period with the amount of shared memory. */
1137 if (p
->numa_group
) {
1138 struct numa_group
*ng
= p
->numa_group
;
1139 unsigned long shared
= group_faults_shared(ng
);
1140 unsigned long private = group_faults_priv(ng
);
1141 unsigned long period
= smax
;
1143 period
*= refcount_read(&ng
->refcount
);
1144 period
*= shared
+ 1;
1145 period
/= private + shared
+ 1;
1147 smax
= max(smax
, period
);
1150 return max(smin
, smax
);
1153 void init_numa_balancing(unsigned long clone_flags
, struct task_struct
*p
)
1156 struct mm_struct
*mm
= p
->mm
;
1159 mm_users
= atomic_read(&mm
->mm_users
);
1160 if (mm_users
== 1) {
1161 mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1162 mm
->numa_scan_seq
= 0;
1166 p
->numa_scan_seq
= mm
? mm
->numa_scan_seq
: 0;
1167 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1168 p
->numa_work
.next
= &p
->numa_work
;
1169 p
->numa_faults
= NULL
;
1170 p
->numa_group
= NULL
;
1171 p
->last_task_numa_placement
= 0;
1172 p
->last_sum_exec_runtime
= 0;
1174 /* New address space, reset the preferred nid */
1175 if (!(clone_flags
& CLONE_VM
)) {
1176 p
->numa_preferred_nid
= NUMA_NO_NODE
;
1181 * New thread, keep existing numa_preferred_nid which should be copied
1182 * already by arch_dup_task_struct but stagger when scans start.
1187 delay
= min_t(unsigned int, task_scan_max(current
),
1188 current
->numa_scan_period
* mm_users
* NSEC_PER_MSEC
);
1189 delay
+= 2 * TICK_NSEC
;
1190 p
->node_stamp
= delay
;
1194 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
1196 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= NUMA_NO_NODE
);
1197 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
1200 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
1202 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= NUMA_NO_NODE
);
1203 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
1206 /* Shared or private faults. */
1207 #define NR_NUMA_HINT_FAULT_TYPES 2
1209 /* Memory and CPU locality */
1210 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1212 /* Averaged statistics, and temporary buffers. */
1213 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1215 pid_t
task_numa_group_id(struct task_struct
*p
)
1217 return p
->numa_group
? p
->numa_group
->gid
: 0;
1221 * The averaged statistics, shared & private, memory & CPU,
1222 * occupy the first half of the array. The second half of the
1223 * array is for current counters, which are averaged into the
1224 * first set by task_numa_placement.
1226 static inline int task_faults_idx(enum numa_faults_stats s
, int nid
, int priv
)
1228 return NR_NUMA_HINT_FAULT_TYPES
* (s
* nr_node_ids
+ nid
) + priv
;
1231 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
1233 if (!p
->numa_faults
)
1236 return p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1237 p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1240 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
1245 return p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1246 p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1249 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
1251 return group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1252 group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1255 static inline unsigned long group_faults_priv(struct numa_group
*ng
)
1257 unsigned long faults
= 0;
1260 for_each_online_node(node
) {
1261 faults
+= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
1267 static inline unsigned long group_faults_shared(struct numa_group
*ng
)
1269 unsigned long faults
= 0;
1272 for_each_online_node(node
) {
1273 faults
+= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
1280 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1281 * considered part of a numa group's pseudo-interleaving set. Migrations
1282 * between these nodes are slowed down, to allow things to settle down.
1284 #define ACTIVE_NODE_FRACTION 3
1286 static bool numa_is_active_node(int nid
, struct numa_group
*ng
)
1288 return group_faults_cpu(ng
, nid
) * ACTIVE_NODE_FRACTION
> ng
->max_faults_cpu
;
1291 /* Handle placement on systems where not all nodes are directly connected. */
1292 static unsigned long score_nearby_nodes(struct task_struct
*p
, int nid
,
1293 int maxdist
, bool task
)
1295 unsigned long score
= 0;
1299 * All nodes are directly connected, and the same distance
1300 * from each other. No need for fancy placement algorithms.
1302 if (sched_numa_topology_type
== NUMA_DIRECT
)
1306 * This code is called for each node, introducing N^2 complexity,
1307 * which should be ok given the number of nodes rarely exceeds 8.
1309 for_each_online_node(node
) {
1310 unsigned long faults
;
1311 int dist
= node_distance(nid
, node
);
1314 * The furthest away nodes in the system are not interesting
1315 * for placement; nid was already counted.
1317 if (dist
== sched_max_numa_distance
|| node
== nid
)
1321 * On systems with a backplane NUMA topology, compare groups
1322 * of nodes, and move tasks towards the group with the most
1323 * memory accesses. When comparing two nodes at distance
1324 * "hoplimit", only nodes closer by than "hoplimit" are part
1325 * of each group. Skip other nodes.
1327 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1331 /* Add up the faults from nearby nodes. */
1333 faults
= task_faults(p
, node
);
1335 faults
= group_faults(p
, node
);
1338 * On systems with a glueless mesh NUMA topology, there are
1339 * no fixed "groups of nodes". Instead, nodes that are not
1340 * directly connected bounce traffic through intermediate
1341 * nodes; a numa_group can occupy any set of nodes.
1342 * The further away a node is, the less the faults count.
1343 * This seems to result in good task placement.
1345 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1346 faults
*= (sched_max_numa_distance
- dist
);
1347 faults
/= (sched_max_numa_distance
- LOCAL_DISTANCE
);
1357 * These return the fraction of accesses done by a particular task, or
1358 * task group, on a particular numa node. The group weight is given a
1359 * larger multiplier, in order to group tasks together that are almost
1360 * evenly spread out between numa nodes.
1362 static inline unsigned long task_weight(struct task_struct
*p
, int nid
,
1365 unsigned long faults
, total_faults
;
1367 if (!p
->numa_faults
)
1370 total_faults
= p
->total_numa_faults
;
1375 faults
= task_faults(p
, nid
);
1376 faults
+= score_nearby_nodes(p
, nid
, dist
, true);
1378 return 1000 * faults
/ total_faults
;
1381 static inline unsigned long group_weight(struct task_struct
*p
, int nid
,
1384 unsigned long faults
, total_faults
;
1389 total_faults
= p
->numa_group
->total_faults
;
1394 faults
= group_faults(p
, nid
);
1395 faults
+= score_nearby_nodes(p
, nid
, dist
, false);
1397 return 1000 * faults
/ total_faults
;
1400 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
1401 int src_nid
, int dst_cpu
)
1403 struct numa_group
*ng
= p
->numa_group
;
1404 int dst_nid
= cpu_to_node(dst_cpu
);
1405 int last_cpupid
, this_cpupid
;
1407 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
1408 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
1411 * Allow first faults or private faults to migrate immediately early in
1412 * the lifetime of a task. The magic number 4 is based on waiting for
1413 * two full passes of the "multi-stage node selection" test that is
1416 if ((p
->numa_preferred_nid
== NUMA_NO_NODE
|| p
->numa_scan_seq
<= 4) &&
1417 (cpupid_pid_unset(last_cpupid
) || cpupid_match_pid(p
, last_cpupid
)))
1421 * Multi-stage node selection is used in conjunction with a periodic
1422 * migration fault to build a temporal task<->page relation. By using
1423 * a two-stage filter we remove short/unlikely relations.
1425 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1426 * a task's usage of a particular page (n_p) per total usage of this
1427 * page (n_t) (in a given time-span) to a probability.
1429 * Our periodic faults will sample this probability and getting the
1430 * same result twice in a row, given these samples are fully
1431 * independent, is then given by P(n)^2, provided our sample period
1432 * is sufficiently short compared to the usage pattern.
1434 * This quadric squishes small probabilities, making it less likely we
1435 * act on an unlikely task<->page relation.
1437 if (!cpupid_pid_unset(last_cpupid
) &&
1438 cpupid_to_nid(last_cpupid
) != dst_nid
)
1441 /* Always allow migrate on private faults */
1442 if (cpupid_match_pid(p
, last_cpupid
))
1445 /* A shared fault, but p->numa_group has not been set up yet. */
1450 * Destination node is much more heavily used than the source
1451 * node? Allow migration.
1453 if (group_faults_cpu(ng
, dst_nid
) > group_faults_cpu(ng
, src_nid
) *
1454 ACTIVE_NODE_FRACTION
)
1458 * Distribute memory according to CPU & memory use on each node,
1459 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1461 * faults_cpu(dst) 3 faults_cpu(src)
1462 * --------------- * - > ---------------
1463 * faults_mem(dst) 4 faults_mem(src)
1465 return group_faults_cpu(ng
, dst_nid
) * group_faults(p
, src_nid
) * 3 >
1466 group_faults_cpu(ng
, src_nid
) * group_faults(p
, dst_nid
) * 4;
1469 static unsigned long weighted_cpuload(struct rq
*rq
);
1470 static unsigned long source_load(int cpu
, int type
);
1471 static unsigned long target_load(int cpu
, int type
);
1473 /* Cached statistics for all CPUs within a node */
1477 /* Total compute capacity of CPUs on a node */
1478 unsigned long compute_capacity
;
1482 * XXX borrowed from update_sg_lb_stats
1484 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1488 memset(ns
, 0, sizeof(*ns
));
1489 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1490 struct rq
*rq
= cpu_rq(cpu
);
1492 ns
->load
+= weighted_cpuload(rq
);
1493 ns
->compute_capacity
+= capacity_of(cpu
);
1498 struct task_numa_env
{
1499 struct task_struct
*p
;
1501 int src_cpu
, src_nid
;
1502 int dst_cpu
, dst_nid
;
1504 struct numa_stats src_stats
, dst_stats
;
1509 struct task_struct
*best_task
;
1514 static void task_numa_assign(struct task_numa_env
*env
,
1515 struct task_struct
*p
, long imp
)
1517 struct rq
*rq
= cpu_rq(env
->dst_cpu
);
1519 /* Bail out if run-queue part of active NUMA balance. */
1520 if (xchg(&rq
->numa_migrate_on
, 1))
1524 * Clear previous best_cpu/rq numa-migrate flag, since task now
1525 * found a better CPU to move/swap.
1527 if (env
->best_cpu
!= -1) {
1528 rq
= cpu_rq(env
->best_cpu
);
1529 WRITE_ONCE(rq
->numa_migrate_on
, 0);
1533 put_task_struct(env
->best_task
);
1538 env
->best_imp
= imp
;
1539 env
->best_cpu
= env
->dst_cpu
;
1542 static bool load_too_imbalanced(long src_load
, long dst_load
,
1543 struct task_numa_env
*env
)
1546 long orig_src_load
, orig_dst_load
;
1547 long src_capacity
, dst_capacity
;
1550 * The load is corrected for the CPU capacity available on each node.
1553 * ------------ vs ---------
1554 * src_capacity dst_capacity
1556 src_capacity
= env
->src_stats
.compute_capacity
;
1557 dst_capacity
= env
->dst_stats
.compute_capacity
;
1559 imb
= abs(dst_load
* src_capacity
- src_load
* dst_capacity
);
1561 orig_src_load
= env
->src_stats
.load
;
1562 orig_dst_load
= env
->dst_stats
.load
;
1564 old_imb
= abs(orig_dst_load
* src_capacity
- orig_src_load
* dst_capacity
);
1566 /* Would this change make things worse? */
1567 return (imb
> old_imb
);
1571 * Maximum NUMA importance can be 1998 (2*999);
1572 * SMALLIMP @ 30 would be close to 1998/64.
1573 * Used to deter task migration.
1578 * This checks if the overall compute and NUMA accesses of the system would
1579 * be improved if the source tasks was migrated to the target dst_cpu taking
1580 * into account that it might be best if task running on the dst_cpu should
1581 * be exchanged with the source task
1583 static void task_numa_compare(struct task_numa_env
*env
,
1584 long taskimp
, long groupimp
, bool maymove
)
1586 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1587 struct task_struct
*cur
;
1588 long src_load
, dst_load
;
1590 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1592 int dist
= env
->dist
;
1594 if (READ_ONCE(dst_rq
->numa_migrate_on
))
1598 cur
= task_rcu_dereference(&dst_rq
->curr
);
1599 if (cur
&& ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
)))
1603 * Because we have preemption enabled we can get migrated around and
1604 * end try selecting ourselves (current == env->p) as a swap candidate.
1610 if (maymove
&& moveimp
>= env
->best_imp
)
1617 * "imp" is the fault differential for the source task between the
1618 * source and destination node. Calculate the total differential for
1619 * the source task and potential destination task. The more negative
1620 * the value is, the more remote accesses that would be expected to
1621 * be incurred if the tasks were swapped.
1623 /* Skip this swap candidate if cannot move to the source cpu */
1624 if (!cpumask_test_cpu(env
->src_cpu
, &cur
->cpus_allowed
))
1628 * If dst and source tasks are in the same NUMA group, or not
1629 * in any group then look only at task weights.
1631 if (cur
->numa_group
== env
->p
->numa_group
) {
1632 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1633 task_weight(cur
, env
->dst_nid
, dist
);
1635 * Add some hysteresis to prevent swapping the
1636 * tasks within a group over tiny differences.
1638 if (cur
->numa_group
)
1642 * Compare the group weights. If a task is all by itself
1643 * (not part of a group), use the task weight instead.
1645 if (cur
->numa_group
&& env
->p
->numa_group
)
1646 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1647 group_weight(cur
, env
->dst_nid
, dist
);
1649 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1650 task_weight(cur
, env
->dst_nid
, dist
);
1653 if (maymove
&& moveimp
> imp
&& moveimp
> env
->best_imp
) {
1660 * If the NUMA importance is less than SMALLIMP,
1661 * task migration might only result in ping pong
1662 * of tasks and also hurt performance due to cache
1665 if (imp
< SMALLIMP
|| imp
<= env
->best_imp
+ SMALLIMP
/ 2)
1669 * In the overloaded case, try and keep the load balanced.
1671 load
= task_h_load(env
->p
) - task_h_load(cur
);
1675 dst_load
= env
->dst_stats
.load
+ load
;
1676 src_load
= env
->src_stats
.load
- load
;
1678 if (load_too_imbalanced(src_load
, dst_load
, env
))
1683 * One idle CPU per node is evaluated for a task numa move.
1684 * Call select_idle_sibling to maybe find a better one.
1688 * select_idle_siblings() uses an per-CPU cpumask that
1689 * can be used from IRQ context.
1691 local_irq_disable();
1692 env
->dst_cpu
= select_idle_sibling(env
->p
, env
->src_cpu
,
1697 task_numa_assign(env
, cur
, imp
);
1702 static void task_numa_find_cpu(struct task_numa_env
*env
,
1703 long taskimp
, long groupimp
)
1705 long src_load
, dst_load
, load
;
1706 bool maymove
= false;
1709 load
= task_h_load(env
->p
);
1710 dst_load
= env
->dst_stats
.load
+ load
;
1711 src_load
= env
->src_stats
.load
- load
;
1714 * If the improvement from just moving env->p direction is better
1715 * than swapping tasks around, check if a move is possible.
1717 maymove
= !load_too_imbalanced(src_load
, dst_load
, env
);
1719 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1720 /* Skip this CPU if the source task cannot migrate */
1721 if (!cpumask_test_cpu(cpu
, &env
->p
->cpus_allowed
))
1725 task_numa_compare(env
, taskimp
, groupimp
, maymove
);
1729 static int task_numa_migrate(struct task_struct
*p
)
1731 struct task_numa_env env
= {
1734 .src_cpu
= task_cpu(p
),
1735 .src_nid
= task_node(p
),
1737 .imbalance_pct
= 112,
1743 struct sched_domain
*sd
;
1745 unsigned long taskweight
, groupweight
;
1747 long taskimp
, groupimp
;
1750 * Pick the lowest SD_NUMA domain, as that would have the smallest
1751 * imbalance and would be the first to start moving tasks about.
1753 * And we want to avoid any moving of tasks about, as that would create
1754 * random movement of tasks -- counter the numa conditions we're trying
1758 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1760 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1764 * Cpusets can break the scheduler domain tree into smaller
1765 * balance domains, some of which do not cross NUMA boundaries.
1766 * Tasks that are "trapped" in such domains cannot be migrated
1767 * elsewhere, so there is no point in (re)trying.
1769 if (unlikely(!sd
)) {
1770 sched_setnuma(p
, task_node(p
));
1774 env
.dst_nid
= p
->numa_preferred_nid
;
1775 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1776 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1777 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1778 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1779 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
1780 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
1781 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1783 /* Try to find a spot on the preferred nid. */
1784 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1787 * Look at other nodes in these cases:
1788 * - there is no space available on the preferred_nid
1789 * - the task is part of a numa_group that is interleaved across
1790 * multiple NUMA nodes; in order to better consolidate the group,
1791 * we need to check other locations.
1793 if (env
.best_cpu
== -1 || (p
->numa_group
&& p
->numa_group
->active_nodes
> 1)) {
1794 for_each_online_node(nid
) {
1795 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1798 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1799 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1801 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1802 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1805 /* Only consider nodes where both task and groups benefit */
1806 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
1807 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
1808 if (taskimp
< 0 && groupimp
< 0)
1813 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1814 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1819 * If the task is part of a workload that spans multiple NUMA nodes,
1820 * and is migrating into one of the workload's active nodes, remember
1821 * this node as the task's preferred numa node, so the workload can
1823 * A task that migrated to a second choice node will be better off
1824 * trying for a better one later. Do not set the preferred node here.
1826 if (p
->numa_group
) {
1827 if (env
.best_cpu
== -1)
1830 nid
= cpu_to_node(env
.best_cpu
);
1832 if (nid
!= p
->numa_preferred_nid
)
1833 sched_setnuma(p
, nid
);
1836 /* No better CPU than the current one was found. */
1837 if (env
.best_cpu
== -1)
1840 best_rq
= cpu_rq(env
.best_cpu
);
1841 if (env
.best_task
== NULL
) {
1842 ret
= migrate_task_to(p
, env
.best_cpu
);
1843 WRITE_ONCE(best_rq
->numa_migrate_on
, 0);
1845 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1849 ret
= migrate_swap(p
, env
.best_task
, env
.best_cpu
, env
.src_cpu
);
1850 WRITE_ONCE(best_rq
->numa_migrate_on
, 0);
1853 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1854 put_task_struct(env
.best_task
);
1858 /* Attempt to migrate a task to a CPU on the preferred node. */
1859 static void numa_migrate_preferred(struct task_struct
*p
)
1861 unsigned long interval
= HZ
;
1863 /* This task has no NUMA fault statistics yet */
1864 if (unlikely(p
->numa_preferred_nid
== NUMA_NO_NODE
|| !p
->numa_faults
))
1867 /* Periodically retry migrating the task to the preferred node */
1868 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1869 p
->numa_migrate_retry
= jiffies
+ interval
;
1871 /* Success if task is already running on preferred CPU */
1872 if (task_node(p
) == p
->numa_preferred_nid
)
1875 /* Otherwise, try migrate to a CPU on the preferred node */
1876 task_numa_migrate(p
);
1880 * Find out how many nodes on the workload is actively running on. Do this by
1881 * tracking the nodes from which NUMA hinting faults are triggered. This can
1882 * be different from the set of nodes where the workload's memory is currently
1885 static void numa_group_count_active_nodes(struct numa_group
*numa_group
)
1887 unsigned long faults
, max_faults
= 0;
1888 int nid
, active_nodes
= 0;
1890 for_each_online_node(nid
) {
1891 faults
= group_faults_cpu(numa_group
, nid
);
1892 if (faults
> max_faults
)
1893 max_faults
= faults
;
1896 for_each_online_node(nid
) {
1897 faults
= group_faults_cpu(numa_group
, nid
);
1898 if (faults
* ACTIVE_NODE_FRACTION
> max_faults
)
1902 numa_group
->max_faults_cpu
= max_faults
;
1903 numa_group
->active_nodes
= active_nodes
;
1907 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1908 * increments. The more local the fault statistics are, the higher the scan
1909 * period will be for the next scan window. If local/(local+remote) ratio is
1910 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1911 * the scan period will decrease. Aim for 70% local accesses.
1913 #define NUMA_PERIOD_SLOTS 10
1914 #define NUMA_PERIOD_THRESHOLD 7
1917 * Increase the scan period (slow down scanning) if the majority of
1918 * our memory is already on our local node, or if the majority of
1919 * the page accesses are shared with other processes.
1920 * Otherwise, decrease the scan period.
1922 static void update_task_scan_period(struct task_struct
*p
,
1923 unsigned long shared
, unsigned long private)
1925 unsigned int period_slot
;
1926 int lr_ratio
, ps_ratio
;
1929 unsigned long remote
= p
->numa_faults_locality
[0];
1930 unsigned long local
= p
->numa_faults_locality
[1];
1933 * If there were no record hinting faults then either the task is
1934 * completely idle or all activity is areas that are not of interest
1935 * to automatic numa balancing. Related to that, if there were failed
1936 * migration then it implies we are migrating too quickly or the local
1937 * node is overloaded. In either case, scan slower
1939 if (local
+ shared
== 0 || p
->numa_faults_locality
[2]) {
1940 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1941 p
->numa_scan_period
<< 1);
1943 p
->mm
->numa_next_scan
= jiffies
+
1944 msecs_to_jiffies(p
->numa_scan_period
);
1950 * Prepare to scale scan period relative to the current period.
1951 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1952 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1953 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1955 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1956 lr_ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1957 ps_ratio
= (private * NUMA_PERIOD_SLOTS
) / (private + shared
);
1959 if (ps_ratio
>= NUMA_PERIOD_THRESHOLD
) {
1961 * Most memory accesses are local. There is no need to
1962 * do fast NUMA scanning, since memory is already local.
1964 int slot
= ps_ratio
- NUMA_PERIOD_THRESHOLD
;
1967 diff
= slot
* period_slot
;
1968 } else if (lr_ratio
>= NUMA_PERIOD_THRESHOLD
) {
1970 * Most memory accesses are shared with other tasks.
1971 * There is no point in continuing fast NUMA scanning,
1972 * since other tasks may just move the memory elsewhere.
1974 int slot
= lr_ratio
- NUMA_PERIOD_THRESHOLD
;
1977 diff
= slot
* period_slot
;
1980 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1981 * yet they are not on the local NUMA node. Speed up
1982 * NUMA scanning to get the memory moved over.
1984 int ratio
= max(lr_ratio
, ps_ratio
);
1985 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1988 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1989 task_scan_min(p
), task_scan_max(p
));
1990 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1994 * Get the fraction of time the task has been running since the last
1995 * NUMA placement cycle. The scheduler keeps similar statistics, but
1996 * decays those on a 32ms period, which is orders of magnitude off
1997 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1998 * stats only if the task is so new there are no NUMA statistics yet.
2000 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
2002 u64 runtime
, delta
, now
;
2003 /* Use the start of this time slice to avoid calculations. */
2004 now
= p
->se
.exec_start
;
2005 runtime
= p
->se
.sum_exec_runtime
;
2007 if (p
->last_task_numa_placement
) {
2008 delta
= runtime
- p
->last_sum_exec_runtime
;
2009 *period
= now
- p
->last_task_numa_placement
;
2011 delta
= p
->se
.avg
.load_sum
;
2012 *period
= LOAD_AVG_MAX
;
2015 p
->last_sum_exec_runtime
= runtime
;
2016 p
->last_task_numa_placement
= now
;
2022 * Determine the preferred nid for a task in a numa_group. This needs to
2023 * be done in a way that produces consistent results with group_weight,
2024 * otherwise workloads might not converge.
2026 static int preferred_group_nid(struct task_struct
*p
, int nid
)
2031 /* Direct connections between all NUMA nodes. */
2032 if (sched_numa_topology_type
== NUMA_DIRECT
)
2036 * On a system with glueless mesh NUMA topology, group_weight
2037 * scores nodes according to the number of NUMA hinting faults on
2038 * both the node itself, and on nearby nodes.
2040 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
2041 unsigned long score
, max_score
= 0;
2042 int node
, max_node
= nid
;
2044 dist
= sched_max_numa_distance
;
2046 for_each_online_node(node
) {
2047 score
= group_weight(p
, node
, dist
);
2048 if (score
> max_score
) {
2057 * Finding the preferred nid in a system with NUMA backplane
2058 * interconnect topology is more involved. The goal is to locate
2059 * tasks from numa_groups near each other in the system, and
2060 * untangle workloads from different sides of the system. This requires
2061 * searching down the hierarchy of node groups, recursively searching
2062 * inside the highest scoring group of nodes. The nodemask tricks
2063 * keep the complexity of the search down.
2065 nodes
= node_online_map
;
2066 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
2067 unsigned long max_faults
= 0;
2068 nodemask_t max_group
= NODE_MASK_NONE
;
2071 /* Are there nodes at this distance from each other? */
2072 if (!find_numa_distance(dist
))
2075 for_each_node_mask(a
, nodes
) {
2076 unsigned long faults
= 0;
2077 nodemask_t this_group
;
2078 nodes_clear(this_group
);
2080 /* Sum group's NUMA faults; includes a==b case. */
2081 for_each_node_mask(b
, nodes
) {
2082 if (node_distance(a
, b
) < dist
) {
2083 faults
+= group_faults(p
, b
);
2084 node_set(b
, this_group
);
2085 node_clear(b
, nodes
);
2089 /* Remember the top group. */
2090 if (faults
> max_faults
) {
2091 max_faults
= faults
;
2092 max_group
= this_group
;
2094 * subtle: at the smallest distance there is
2095 * just one node left in each "group", the
2096 * winner is the preferred nid.
2101 /* Next round, evaluate the nodes within max_group. */
2109 static void task_numa_placement(struct task_struct
*p
)
2111 int seq
, nid
, max_nid
= NUMA_NO_NODE
;
2112 unsigned long max_faults
= 0;
2113 unsigned long fault_types
[2] = { 0, 0 };
2114 unsigned long total_faults
;
2115 u64 runtime
, period
;
2116 spinlock_t
*group_lock
= NULL
;
2119 * The p->mm->numa_scan_seq field gets updated without
2120 * exclusive access. Use READ_ONCE() here to ensure
2121 * that the field is read in a single access:
2123 seq
= READ_ONCE(p
->mm
->numa_scan_seq
);
2124 if (p
->numa_scan_seq
== seq
)
2126 p
->numa_scan_seq
= seq
;
2127 p
->numa_scan_period_max
= task_scan_max(p
);
2129 total_faults
= p
->numa_faults_locality
[0] +
2130 p
->numa_faults_locality
[1];
2131 runtime
= numa_get_avg_runtime(p
, &period
);
2133 /* If the task is part of a group prevent parallel updates to group stats */
2134 if (p
->numa_group
) {
2135 group_lock
= &p
->numa_group
->lock
;
2136 spin_lock_irq(group_lock
);
2139 /* Find the node with the highest number of faults */
2140 for_each_online_node(nid
) {
2141 /* Keep track of the offsets in numa_faults array */
2142 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
2143 unsigned long faults
= 0, group_faults
= 0;
2146 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
2147 long diff
, f_diff
, f_weight
;
2149 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
2150 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
2151 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
2152 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
2154 /* Decay existing window, copy faults since last scan */
2155 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
2156 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
2157 p
->numa_faults
[membuf_idx
] = 0;
2160 * Normalize the faults_from, so all tasks in a group
2161 * count according to CPU use, instead of by the raw
2162 * number of faults. Tasks with little runtime have
2163 * little over-all impact on throughput, and thus their
2164 * faults are less important.
2166 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
2167 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
2169 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
2170 p
->numa_faults
[cpubuf_idx
] = 0;
2172 p
->numa_faults
[mem_idx
] += diff
;
2173 p
->numa_faults
[cpu_idx
] += f_diff
;
2174 faults
+= p
->numa_faults
[mem_idx
];
2175 p
->total_numa_faults
+= diff
;
2176 if (p
->numa_group
) {
2178 * safe because we can only change our own group
2180 * mem_idx represents the offset for a given
2181 * nid and priv in a specific region because it
2182 * is at the beginning of the numa_faults array.
2184 p
->numa_group
->faults
[mem_idx
] += diff
;
2185 p
->numa_group
->faults_cpu
[mem_idx
] += f_diff
;
2186 p
->numa_group
->total_faults
+= diff
;
2187 group_faults
+= p
->numa_group
->faults
[mem_idx
];
2191 if (!p
->numa_group
) {
2192 if (faults
> max_faults
) {
2193 max_faults
= faults
;
2196 } else if (group_faults
> max_faults
) {
2197 max_faults
= group_faults
;
2202 if (p
->numa_group
) {
2203 numa_group_count_active_nodes(p
->numa_group
);
2204 spin_unlock_irq(group_lock
);
2205 max_nid
= preferred_group_nid(p
, max_nid
);
2209 /* Set the new preferred node */
2210 if (max_nid
!= p
->numa_preferred_nid
)
2211 sched_setnuma(p
, max_nid
);
2214 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
2217 static inline int get_numa_group(struct numa_group
*grp
)
2219 return refcount_inc_not_zero(&grp
->refcount
);
2222 static inline void put_numa_group(struct numa_group
*grp
)
2224 if (refcount_dec_and_test(&grp
->refcount
))
2225 kfree_rcu(grp
, rcu
);
2228 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
2231 struct numa_group
*grp
, *my_grp
;
2232 struct task_struct
*tsk
;
2234 int cpu
= cpupid_to_cpu(cpupid
);
2237 if (unlikely(!p
->numa_group
)) {
2238 unsigned int size
= sizeof(struct numa_group
) +
2239 4*nr_node_ids
*sizeof(unsigned long);
2241 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
2245 refcount_set(&grp
->refcount
, 1);
2246 grp
->active_nodes
= 1;
2247 grp
->max_faults_cpu
= 0;
2248 spin_lock_init(&grp
->lock
);
2250 /* Second half of the array tracks nids where faults happen */
2251 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
2254 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2255 grp
->faults
[i
] = p
->numa_faults
[i
];
2257 grp
->total_faults
= p
->total_numa_faults
;
2260 rcu_assign_pointer(p
->numa_group
, grp
);
2264 tsk
= READ_ONCE(cpu_rq(cpu
)->curr
);
2266 if (!cpupid_match_pid(tsk
, cpupid
))
2269 grp
= rcu_dereference(tsk
->numa_group
);
2273 my_grp
= p
->numa_group
;
2278 * Only join the other group if its bigger; if we're the bigger group,
2279 * the other task will join us.
2281 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
2285 * Tie-break on the grp address.
2287 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
2290 /* Always join threads in the same process. */
2291 if (tsk
->mm
== current
->mm
)
2294 /* Simple filter to avoid false positives due to PID collisions */
2295 if (flags
& TNF_SHARED
)
2298 /* Update priv based on whether false sharing was detected */
2301 if (join
&& !get_numa_group(grp
))
2309 BUG_ON(irqs_disabled());
2310 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
2312 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
2313 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
2314 grp
->faults
[i
] += p
->numa_faults
[i
];
2316 my_grp
->total_faults
-= p
->total_numa_faults
;
2317 grp
->total_faults
+= p
->total_numa_faults
;
2322 spin_unlock(&my_grp
->lock
);
2323 spin_unlock_irq(&grp
->lock
);
2325 rcu_assign_pointer(p
->numa_group
, grp
);
2327 put_numa_group(my_grp
);
2335 void task_numa_free(struct task_struct
*p
)
2337 struct numa_group
*grp
= p
->numa_group
;
2338 void *numa_faults
= p
->numa_faults
;
2339 unsigned long flags
;
2343 spin_lock_irqsave(&grp
->lock
, flags
);
2344 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2345 grp
->faults
[i
] -= p
->numa_faults
[i
];
2346 grp
->total_faults
-= p
->total_numa_faults
;
2349 spin_unlock_irqrestore(&grp
->lock
, flags
);
2350 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2351 put_numa_group(grp
);
2354 p
->numa_faults
= NULL
;
2359 * Got a PROT_NONE fault for a page on @node.
2361 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2363 struct task_struct
*p
= current
;
2364 bool migrated
= flags
& TNF_MIGRATED
;
2365 int cpu_node
= task_node(current
);
2366 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2367 struct numa_group
*ng
;
2370 if (!static_branch_likely(&sched_numa_balancing
))
2373 /* for example, ksmd faulting in a user's mm */
2377 /* Allocate buffer to track faults on a per-node basis */
2378 if (unlikely(!p
->numa_faults
)) {
2379 int size
= sizeof(*p
->numa_faults
) *
2380 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2382 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2383 if (!p
->numa_faults
)
2386 p
->total_numa_faults
= 0;
2387 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2391 * First accesses are treated as private, otherwise consider accesses
2392 * to be private if the accessing pid has not changed
2394 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2397 priv
= cpupid_match_pid(p
, last_cpupid
);
2398 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2399 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2403 * If a workload spans multiple NUMA nodes, a shared fault that
2404 * occurs wholly within the set of nodes that the workload is
2405 * actively using should be counted as local. This allows the
2406 * scan rate to slow down when a workload has settled down.
2409 if (!priv
&& !local
&& ng
&& ng
->active_nodes
> 1 &&
2410 numa_is_active_node(cpu_node
, ng
) &&
2411 numa_is_active_node(mem_node
, ng
))
2415 * Retry to migrate task to preferred node periodically, in case it
2416 * previously failed, or the scheduler moved us.
2418 if (time_after(jiffies
, p
->numa_migrate_retry
)) {
2419 task_numa_placement(p
);
2420 numa_migrate_preferred(p
);
2424 p
->numa_pages_migrated
+= pages
;
2425 if (flags
& TNF_MIGRATE_FAIL
)
2426 p
->numa_faults_locality
[2] += pages
;
2428 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2429 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2430 p
->numa_faults_locality
[local
] += pages
;
2433 static void reset_ptenuma_scan(struct task_struct
*p
)
2436 * We only did a read acquisition of the mmap sem, so
2437 * p->mm->numa_scan_seq is written to without exclusive access
2438 * and the update is not guaranteed to be atomic. That's not
2439 * much of an issue though, since this is just used for
2440 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2441 * expensive, to avoid any form of compiler optimizations:
2443 WRITE_ONCE(p
->mm
->numa_scan_seq
, READ_ONCE(p
->mm
->numa_scan_seq
) + 1);
2444 p
->mm
->numa_scan_offset
= 0;
2448 * The expensive part of numa migration is done from task_work context.
2449 * Triggered from task_tick_numa().
2451 void task_numa_work(struct callback_head
*work
)
2453 unsigned long migrate
, next_scan
, now
= jiffies
;
2454 struct task_struct
*p
= current
;
2455 struct mm_struct
*mm
= p
->mm
;
2456 u64 runtime
= p
->se
.sum_exec_runtime
;
2457 struct vm_area_struct
*vma
;
2458 unsigned long start
, end
;
2459 unsigned long nr_pte_updates
= 0;
2460 long pages
, virtpages
;
2462 SCHED_WARN_ON(p
!= container_of(work
, struct task_struct
, numa_work
));
2464 work
->next
= work
; /* protect against double add */
2466 * Who cares about NUMA placement when they're dying.
2468 * NOTE: make sure not to dereference p->mm before this check,
2469 * exit_task_work() happens _after_ exit_mm() so we could be called
2470 * without p->mm even though we still had it when we enqueued this
2473 if (p
->flags
& PF_EXITING
)
2476 if (!mm
->numa_next_scan
) {
2477 mm
->numa_next_scan
= now
+
2478 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2482 * Enforce maximal scan/migration frequency..
2484 migrate
= mm
->numa_next_scan
;
2485 if (time_before(now
, migrate
))
2488 if (p
->numa_scan_period
== 0) {
2489 p
->numa_scan_period_max
= task_scan_max(p
);
2490 p
->numa_scan_period
= task_scan_start(p
);
2493 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2494 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2498 * Delay this task enough that another task of this mm will likely win
2499 * the next time around.
2501 p
->node_stamp
+= 2 * TICK_NSEC
;
2503 start
= mm
->numa_scan_offset
;
2504 pages
= sysctl_numa_balancing_scan_size
;
2505 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2506 virtpages
= pages
* 8; /* Scan up to this much virtual space */
2511 if (!down_read_trylock(&mm
->mmap_sem
))
2513 vma
= find_vma(mm
, start
);
2515 reset_ptenuma_scan(p
);
2519 for (; vma
; vma
= vma
->vm_next
) {
2520 if (!vma_migratable(vma
) || !vma_policy_mof(vma
) ||
2521 is_vm_hugetlb_page(vma
) || (vma
->vm_flags
& VM_MIXEDMAP
)) {
2526 * Shared library pages mapped by multiple processes are not
2527 * migrated as it is expected they are cache replicated. Avoid
2528 * hinting faults in read-only file-backed mappings or the vdso
2529 * as migrating the pages will be of marginal benefit.
2532 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2536 * Skip inaccessible VMAs to avoid any confusion between
2537 * PROT_NONE and NUMA hinting ptes
2539 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
2543 start
= max(start
, vma
->vm_start
);
2544 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2545 end
= min(end
, vma
->vm_end
);
2546 nr_pte_updates
= change_prot_numa(vma
, start
, end
);
2549 * Try to scan sysctl_numa_balancing_size worth of
2550 * hpages that have at least one present PTE that
2551 * is not already pte-numa. If the VMA contains
2552 * areas that are unused or already full of prot_numa
2553 * PTEs, scan up to virtpages, to skip through those
2557 pages
-= (end
- start
) >> PAGE_SHIFT
;
2558 virtpages
-= (end
- start
) >> PAGE_SHIFT
;
2561 if (pages
<= 0 || virtpages
<= 0)
2565 } while (end
!= vma
->vm_end
);
2570 * It is possible to reach the end of the VMA list but the last few
2571 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2572 * would find the !migratable VMA on the next scan but not reset the
2573 * scanner to the start so check it now.
2576 mm
->numa_scan_offset
= start
;
2578 reset_ptenuma_scan(p
);
2579 up_read(&mm
->mmap_sem
);
2582 * Make sure tasks use at least 32x as much time to run other code
2583 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2584 * Usually update_task_scan_period slows down scanning enough; on an
2585 * overloaded system we need to limit overhead on a per task basis.
2587 if (unlikely(p
->se
.sum_exec_runtime
!= runtime
)) {
2588 u64 diff
= p
->se
.sum_exec_runtime
- runtime
;
2589 p
->node_stamp
+= 32 * diff
;
2594 * Drive the periodic memory faults..
2596 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2598 struct callback_head
*work
= &curr
->numa_work
;
2602 * We don't care about NUMA placement if we don't have memory.
2604 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2608 * Using runtime rather than walltime has the dual advantage that
2609 * we (mostly) drive the selection from busy threads and that the
2610 * task needs to have done some actual work before we bother with
2613 now
= curr
->se
.sum_exec_runtime
;
2614 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2616 if (now
> curr
->node_stamp
+ period
) {
2617 if (!curr
->node_stamp
)
2618 curr
->numa_scan_period
= task_scan_start(curr
);
2619 curr
->node_stamp
+= period
;
2621 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2622 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2623 task_work_add(curr
, work
, true);
2628 static void update_scan_period(struct task_struct
*p
, int new_cpu
)
2630 int src_nid
= cpu_to_node(task_cpu(p
));
2631 int dst_nid
= cpu_to_node(new_cpu
);
2633 if (!static_branch_likely(&sched_numa_balancing
))
2636 if (!p
->mm
|| !p
->numa_faults
|| (p
->flags
& PF_EXITING
))
2639 if (src_nid
== dst_nid
)
2643 * Allow resets if faults have been trapped before one scan
2644 * has completed. This is most likely due to a new task that
2645 * is pulled cross-node due to wakeups or load balancing.
2647 if (p
->numa_scan_seq
) {
2649 * Avoid scan adjustments if moving to the preferred
2650 * node or if the task was not previously running on
2651 * the preferred node.
2653 if (dst_nid
== p
->numa_preferred_nid
||
2654 (p
->numa_preferred_nid
!= NUMA_NO_NODE
&&
2655 src_nid
!= p
->numa_preferred_nid
))
2659 p
->numa_scan_period
= task_scan_start(p
);
2663 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2667 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2671 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2675 static inline void update_scan_period(struct task_struct
*p
, int new_cpu
)
2679 #endif /* CONFIG_NUMA_BALANCING */
2682 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2684 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2685 if (!parent_entity(se
))
2686 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2688 if (entity_is_task(se
)) {
2689 struct rq
*rq
= rq_of(cfs_rq
);
2691 account_numa_enqueue(rq
, task_of(se
));
2692 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2695 cfs_rq
->nr_running
++;
2699 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2701 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2702 if (!parent_entity(se
))
2703 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2705 if (entity_is_task(se
)) {
2706 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2707 list_del_init(&se
->group_node
);
2710 cfs_rq
->nr_running
--;
2714 * Signed add and clamp on underflow.
2716 * Explicitly do a load-store to ensure the intermediate value never hits
2717 * memory. This allows lockless observations without ever seeing the negative
2720 #define add_positive(_ptr, _val) do { \
2721 typeof(_ptr) ptr = (_ptr); \
2722 typeof(_val) val = (_val); \
2723 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2727 if (val < 0 && res > var) \
2730 WRITE_ONCE(*ptr, res); \
2734 * Unsigned subtract and clamp on underflow.
2736 * Explicitly do a load-store to ensure the intermediate value never hits
2737 * memory. This allows lockless observations without ever seeing the negative
2740 #define sub_positive(_ptr, _val) do { \
2741 typeof(_ptr) ptr = (_ptr); \
2742 typeof(*ptr) val = (_val); \
2743 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2747 WRITE_ONCE(*ptr, res); \
2751 * Remove and clamp on negative, from a local variable.
2753 * A variant of sub_positive(), which does not use explicit load-store
2754 * and is thus optimized for local variable updates.
2756 #define lsub_positive(_ptr, _val) do { \
2757 typeof(_ptr) ptr = (_ptr); \
2758 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2763 enqueue_runnable_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2765 cfs_rq
->runnable_weight
+= se
->runnable_weight
;
2767 cfs_rq
->avg
.runnable_load_avg
+= se
->avg
.runnable_load_avg
;
2768 cfs_rq
->avg
.runnable_load_sum
+= se_runnable(se
) * se
->avg
.runnable_load_sum
;
2772 dequeue_runnable_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2774 cfs_rq
->runnable_weight
-= se
->runnable_weight
;
2776 sub_positive(&cfs_rq
->avg
.runnable_load_avg
, se
->avg
.runnable_load_avg
);
2777 sub_positive(&cfs_rq
->avg
.runnable_load_sum
,
2778 se_runnable(se
) * se
->avg
.runnable_load_sum
);
2782 enqueue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2784 cfs_rq
->avg
.load_avg
+= se
->avg
.load_avg
;
2785 cfs_rq
->avg
.load_sum
+= se_weight(se
) * se
->avg
.load_sum
;
2789 dequeue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2791 sub_positive(&cfs_rq
->avg
.load_avg
, se
->avg
.load_avg
);
2792 sub_positive(&cfs_rq
->avg
.load_sum
, se_weight(se
) * se
->avg
.load_sum
);
2796 enqueue_runnable_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
2798 dequeue_runnable_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
2800 enqueue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
2802 dequeue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
2805 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2806 unsigned long weight
, unsigned long runnable
)
2809 /* commit outstanding execution time */
2810 if (cfs_rq
->curr
== se
)
2811 update_curr(cfs_rq
);
2812 account_entity_dequeue(cfs_rq
, se
);
2813 dequeue_runnable_load_avg(cfs_rq
, se
);
2815 dequeue_load_avg(cfs_rq
, se
);
2817 se
->runnable_weight
= runnable
;
2818 update_load_set(&se
->load
, weight
);
2822 u32 divider
= LOAD_AVG_MAX
- 1024 + se
->avg
.period_contrib
;
2824 se
->avg
.load_avg
= div_u64(se_weight(se
) * se
->avg
.load_sum
, divider
);
2825 se
->avg
.runnable_load_avg
=
2826 div_u64(se_runnable(se
) * se
->avg
.runnable_load_sum
, divider
);
2830 enqueue_load_avg(cfs_rq
, se
);
2832 account_entity_enqueue(cfs_rq
, se
);
2833 enqueue_runnable_load_avg(cfs_rq
, se
);
2837 void reweight_task(struct task_struct
*p
, int prio
)
2839 struct sched_entity
*se
= &p
->se
;
2840 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2841 struct load_weight
*load
= &se
->load
;
2842 unsigned long weight
= scale_load(sched_prio_to_weight
[prio
]);
2844 reweight_entity(cfs_rq
, se
, weight
, weight
);
2845 load
->inv_weight
= sched_prio_to_wmult
[prio
];
2848 #ifdef CONFIG_FAIR_GROUP_SCHED
2851 * All this does is approximate the hierarchical proportion which includes that
2852 * global sum we all love to hate.
2854 * That is, the weight of a group entity, is the proportional share of the
2855 * group weight based on the group runqueue weights. That is:
2857 * tg->weight * grq->load.weight
2858 * ge->load.weight = ----------------------------- (1)
2859 * \Sum grq->load.weight
2861 * Now, because computing that sum is prohibitively expensive to compute (been
2862 * there, done that) we approximate it with this average stuff. The average
2863 * moves slower and therefore the approximation is cheaper and more stable.
2865 * So instead of the above, we substitute:
2867 * grq->load.weight -> grq->avg.load_avg (2)
2869 * which yields the following:
2871 * tg->weight * grq->avg.load_avg
2872 * ge->load.weight = ------------------------------ (3)
2875 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2877 * That is shares_avg, and it is right (given the approximation (2)).
2879 * The problem with it is that because the average is slow -- it was designed
2880 * to be exactly that of course -- this leads to transients in boundary
2881 * conditions. In specific, the case where the group was idle and we start the
2882 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2883 * yielding bad latency etc..
2885 * Now, in that special case (1) reduces to:
2887 * tg->weight * grq->load.weight
2888 * ge->load.weight = ----------------------------- = tg->weight (4)
2891 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2893 * So what we do is modify our approximation (3) to approach (4) in the (near)
2898 * tg->weight * grq->load.weight
2899 * --------------------------------------------------- (5)
2900 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2902 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2903 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2906 * tg->weight * grq->load.weight
2907 * ge->load.weight = ----------------------------- (6)
2912 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2913 * max(grq->load.weight, grq->avg.load_avg)
2915 * And that is shares_weight and is icky. In the (near) UP case it approaches
2916 * (4) while in the normal case it approaches (3). It consistently
2917 * overestimates the ge->load.weight and therefore:
2919 * \Sum ge->load.weight >= tg->weight
2923 static long calc_group_shares(struct cfs_rq
*cfs_rq
)
2925 long tg_weight
, tg_shares
, load
, shares
;
2926 struct task_group
*tg
= cfs_rq
->tg
;
2928 tg_shares
= READ_ONCE(tg
->shares
);
2930 load
= max(scale_load_down(cfs_rq
->load
.weight
), cfs_rq
->avg
.load_avg
);
2932 tg_weight
= atomic_long_read(&tg
->load_avg
);
2934 /* Ensure tg_weight >= load */
2935 tg_weight
-= cfs_rq
->tg_load_avg_contrib
;
2938 shares
= (tg_shares
* load
);
2940 shares
/= tg_weight
;
2943 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2944 * of a group with small tg->shares value. It is a floor value which is
2945 * assigned as a minimum load.weight to the sched_entity representing
2946 * the group on a CPU.
2948 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2949 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2950 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2951 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2954 return clamp_t(long, shares
, MIN_SHARES
, tg_shares
);
2958 * This calculates the effective runnable weight for a group entity based on
2959 * the group entity weight calculated above.
2961 * Because of the above approximation (2), our group entity weight is
2962 * an load_avg based ratio (3). This means that it includes blocked load and
2963 * does not represent the runnable weight.
2965 * Approximate the group entity's runnable weight per ratio from the group
2968 * grq->avg.runnable_load_avg
2969 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2972 * However, analogous to above, since the avg numbers are slow, this leads to
2973 * transients in the from-idle case. Instead we use:
2975 * ge->runnable_weight = ge->load.weight *
2977 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2978 * ----------------------------------------------------- (8)
2979 * max(grq->avg.load_avg, grq->load.weight)
2981 * Where these max() serve both to use the 'instant' values to fix the slow
2982 * from-idle and avoid the /0 on to-idle, similar to (6).
2984 static long calc_group_runnable(struct cfs_rq
*cfs_rq
, long shares
)
2986 long runnable
, load_avg
;
2988 load_avg
= max(cfs_rq
->avg
.load_avg
,
2989 scale_load_down(cfs_rq
->load
.weight
));
2991 runnable
= max(cfs_rq
->avg
.runnable_load_avg
,
2992 scale_load_down(cfs_rq
->runnable_weight
));
2996 runnable
/= load_avg
;
2998 return clamp_t(long, runnable
, MIN_SHARES
, shares
);
3000 #endif /* CONFIG_SMP */
3002 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
3005 * Recomputes the group entity based on the current state of its group
3008 static void update_cfs_group(struct sched_entity
*se
)
3010 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3011 long shares
, runnable
;
3016 if (throttled_hierarchy(gcfs_rq
))
3020 runnable
= shares
= READ_ONCE(gcfs_rq
->tg
->shares
);
3022 if (likely(se
->load
.weight
== shares
))
3025 shares
= calc_group_shares(gcfs_rq
);
3026 runnable
= calc_group_runnable(gcfs_rq
, shares
);
3029 reweight_entity(cfs_rq_of(se
), se
, shares
, runnable
);
3032 #else /* CONFIG_FAIR_GROUP_SCHED */
3033 static inline void update_cfs_group(struct sched_entity
*se
)
3036 #endif /* CONFIG_FAIR_GROUP_SCHED */
3038 static inline void cfs_rq_util_change(struct cfs_rq
*cfs_rq
, int flags
)
3040 struct rq
*rq
= rq_of(cfs_rq
);
3042 if (&rq
->cfs
== cfs_rq
|| (flags
& SCHED_CPUFREQ_MIGRATION
)) {
3044 * There are a few boundary cases this might miss but it should
3045 * get called often enough that that should (hopefully) not be
3048 * It will not get called when we go idle, because the idle
3049 * thread is a different class (!fair), nor will the utilization
3050 * number include things like RT tasks.
3052 * As is, the util number is not freq-invariant (we'd have to
3053 * implement arch_scale_freq_capacity() for that).
3057 cpufreq_update_util(rq
, flags
);
3062 #ifdef CONFIG_FAIR_GROUP_SCHED
3064 * update_tg_load_avg - update the tg's load avg
3065 * @cfs_rq: the cfs_rq whose avg changed
3066 * @force: update regardless of how small the difference
3068 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3069 * However, because tg->load_avg is a global value there are performance
3072 * In order to avoid having to look at the other cfs_rq's, we use a
3073 * differential update where we store the last value we propagated. This in
3074 * turn allows skipping updates if the differential is 'small'.
3076 * Updating tg's load_avg is necessary before update_cfs_share().
3078 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
)
3080 long delta
= cfs_rq
->avg
.load_avg
- cfs_rq
->tg_load_avg_contrib
;
3083 * No need to update load_avg for root_task_group as it is not used.
3085 if (cfs_rq
->tg
== &root_task_group
)
3088 if (force
|| abs(delta
) > cfs_rq
->tg_load_avg_contrib
/ 64) {
3089 atomic_long_add(delta
, &cfs_rq
->tg
->load_avg
);
3090 cfs_rq
->tg_load_avg_contrib
= cfs_rq
->avg
.load_avg
;
3095 * Called within set_task_rq() right before setting a task's CPU. The
3096 * caller only guarantees p->pi_lock is held; no other assumptions,
3097 * including the state of rq->lock, should be made.
3099 void set_task_rq_fair(struct sched_entity
*se
,
3100 struct cfs_rq
*prev
, struct cfs_rq
*next
)
3102 u64 p_last_update_time
;
3103 u64 n_last_update_time
;
3105 if (!sched_feat(ATTACH_AGE_LOAD
))
3109 * We are supposed to update the task to "current" time, then its up to
3110 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3111 * getting what current time is, so simply throw away the out-of-date
3112 * time. This will result in the wakee task is less decayed, but giving
3113 * the wakee more load sounds not bad.
3115 if (!(se
->avg
.last_update_time
&& prev
))
3118 #ifndef CONFIG_64BIT
3120 u64 p_last_update_time_copy
;
3121 u64 n_last_update_time_copy
;
3124 p_last_update_time_copy
= prev
->load_last_update_time_copy
;
3125 n_last_update_time_copy
= next
->load_last_update_time_copy
;
3129 p_last_update_time
= prev
->avg
.last_update_time
;
3130 n_last_update_time
= next
->avg
.last_update_time
;
3132 } while (p_last_update_time
!= p_last_update_time_copy
||
3133 n_last_update_time
!= n_last_update_time_copy
);
3136 p_last_update_time
= prev
->avg
.last_update_time
;
3137 n_last_update_time
= next
->avg
.last_update_time
;
3139 __update_load_avg_blocked_se(p_last_update_time
, se
);
3140 se
->avg
.last_update_time
= n_last_update_time
;
3145 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3146 * propagate its contribution. The key to this propagation is the invariant
3147 * that for each group:
3149 * ge->avg == grq->avg (1)
3151 * _IFF_ we look at the pure running and runnable sums. Because they
3152 * represent the very same entity, just at different points in the hierarchy.
3154 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3155 * sum over (but still wrong, because the group entity and group rq do not have
3156 * their PELT windows aligned).
3158 * However, update_tg_cfs_runnable() is more complex. So we have:
3160 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3162 * And since, like util, the runnable part should be directly transferable,
3163 * the following would _appear_ to be the straight forward approach:
3165 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3167 * And per (1) we have:
3169 * ge->avg.runnable_avg == grq->avg.runnable_avg
3173 * ge->load.weight * grq->avg.load_avg
3174 * ge->avg.load_avg = ----------------------------------- (4)
3177 * Except that is wrong!
3179 * Because while for entities historical weight is not important and we
3180 * really only care about our future and therefore can consider a pure
3181 * runnable sum, runqueues can NOT do this.
3183 * We specifically want runqueues to have a load_avg that includes
3184 * historical weights. Those represent the blocked load, the load we expect
3185 * to (shortly) return to us. This only works by keeping the weights as
3186 * integral part of the sum. We therefore cannot decompose as per (3).
3188 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3189 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3190 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3191 * runnable section of these tasks overlap (or not). If they were to perfectly
3192 * align the rq as a whole would be runnable 2/3 of the time. If however we
3193 * always have at least 1 runnable task, the rq as a whole is always runnable.
3195 * So we'll have to approximate.. :/
3197 * Given the constraint:
3199 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3201 * We can construct a rule that adds runnable to a rq by assuming minimal
3204 * On removal, we'll assume each task is equally runnable; which yields:
3206 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3208 * XXX: only do this for the part of runnable > running ?
3213 update_tg_cfs_util(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3215 long delta
= gcfs_rq
->avg
.util_avg
- se
->avg
.util_avg
;
3217 /* Nothing to update */
3222 * The relation between sum and avg is:
3224 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3226 * however, the PELT windows are not aligned between grq and gse.
3229 /* Set new sched_entity's utilization */
3230 se
->avg
.util_avg
= gcfs_rq
->avg
.util_avg
;
3231 se
->avg
.util_sum
= se
->avg
.util_avg
* LOAD_AVG_MAX
;
3233 /* Update parent cfs_rq utilization */
3234 add_positive(&cfs_rq
->avg
.util_avg
, delta
);
3235 cfs_rq
->avg
.util_sum
= cfs_rq
->avg
.util_avg
* LOAD_AVG_MAX
;
3239 update_tg_cfs_runnable(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3241 long delta_avg
, running_sum
, runnable_sum
= gcfs_rq
->prop_runnable_sum
;
3242 unsigned long runnable_load_avg
, load_avg
;
3243 u64 runnable_load_sum
, load_sum
= 0;
3249 gcfs_rq
->prop_runnable_sum
= 0;
3251 if (runnable_sum
>= 0) {
3253 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3254 * the CPU is saturated running == runnable.
3256 runnable_sum
+= se
->avg
.load_sum
;
3257 runnable_sum
= min(runnable_sum
, (long)LOAD_AVG_MAX
);
3260 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3261 * assuming all tasks are equally runnable.
3263 if (scale_load_down(gcfs_rq
->load
.weight
)) {
3264 load_sum
= div_s64(gcfs_rq
->avg
.load_sum
,
3265 scale_load_down(gcfs_rq
->load
.weight
));
3268 /* But make sure to not inflate se's runnable */
3269 runnable_sum
= min(se
->avg
.load_sum
, load_sum
);
3273 * runnable_sum can't be lower than running_sum
3274 * Rescale running sum to be in the same range as runnable sum
3275 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3276 * runnable_sum is in [0 : LOAD_AVG_MAX]
3278 running_sum
= se
->avg
.util_sum
>> SCHED_CAPACITY_SHIFT
;
3279 runnable_sum
= max(runnable_sum
, running_sum
);
3281 load_sum
= (s64
)se_weight(se
) * runnable_sum
;
3282 load_avg
= div_s64(load_sum
, LOAD_AVG_MAX
);
3284 delta_sum
= load_sum
- (s64
)se_weight(se
) * se
->avg
.load_sum
;
3285 delta_avg
= load_avg
- se
->avg
.load_avg
;
3287 se
->avg
.load_sum
= runnable_sum
;
3288 se
->avg
.load_avg
= load_avg
;
3289 add_positive(&cfs_rq
->avg
.load_avg
, delta_avg
);
3290 add_positive(&cfs_rq
->avg
.load_sum
, delta_sum
);
3292 runnable_load_sum
= (s64
)se_runnable(se
) * runnable_sum
;
3293 runnable_load_avg
= div_s64(runnable_load_sum
, LOAD_AVG_MAX
);
3294 delta_sum
= runnable_load_sum
- se_weight(se
) * se
->avg
.runnable_load_sum
;
3295 delta_avg
= runnable_load_avg
- se
->avg
.runnable_load_avg
;
3297 se
->avg
.runnable_load_sum
= runnable_sum
;
3298 se
->avg
.runnable_load_avg
= runnable_load_avg
;
3301 add_positive(&cfs_rq
->avg
.runnable_load_avg
, delta_avg
);
3302 add_positive(&cfs_rq
->avg
.runnable_load_sum
, delta_sum
);
3306 static inline void add_tg_cfs_propagate(struct cfs_rq
*cfs_rq
, long runnable_sum
)
3308 cfs_rq
->propagate
= 1;
3309 cfs_rq
->prop_runnable_sum
+= runnable_sum
;
3312 /* Update task and its cfs_rq load average */
3313 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3315 struct cfs_rq
*cfs_rq
, *gcfs_rq
;
3317 if (entity_is_task(se
))
3320 gcfs_rq
= group_cfs_rq(se
);
3321 if (!gcfs_rq
->propagate
)
3324 gcfs_rq
->propagate
= 0;
3326 cfs_rq
= cfs_rq_of(se
);
3328 add_tg_cfs_propagate(cfs_rq
, gcfs_rq
->prop_runnable_sum
);
3330 update_tg_cfs_util(cfs_rq
, se
, gcfs_rq
);
3331 update_tg_cfs_runnable(cfs_rq
, se
, gcfs_rq
);
3337 * Check if we need to update the load and the utilization of a blocked
3340 static inline bool skip_blocked_update(struct sched_entity
*se
)
3342 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3345 * If sched_entity still have not zero load or utilization, we have to
3348 if (se
->avg
.load_avg
|| se
->avg
.util_avg
)
3352 * If there is a pending propagation, we have to update the load and
3353 * the utilization of the sched_entity:
3355 if (gcfs_rq
->propagate
)
3359 * Otherwise, the load and the utilization of the sched_entity is
3360 * already zero and there is no pending propagation, so it will be a
3361 * waste of time to try to decay it:
3366 #else /* CONFIG_FAIR_GROUP_SCHED */
3368 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
) {}
3370 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3375 static inline void add_tg_cfs_propagate(struct cfs_rq
*cfs_rq
, long runnable_sum
) {}
3377 #endif /* CONFIG_FAIR_GROUP_SCHED */
3380 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3381 * @now: current time, as per cfs_rq_clock_pelt()
3382 * @cfs_rq: cfs_rq to update
3384 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3385 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3386 * post_init_entity_util_avg().
3388 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3390 * Returns true if the load decayed or we removed load.
3392 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3393 * call update_tg_load_avg() when this function returns true.
3396 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
)
3398 unsigned long removed_load
= 0, removed_util
= 0, removed_runnable_sum
= 0;
3399 struct sched_avg
*sa
= &cfs_rq
->avg
;
3402 if (cfs_rq
->removed
.nr
) {
3404 u32 divider
= LOAD_AVG_MAX
- 1024 + sa
->period_contrib
;
3406 raw_spin_lock(&cfs_rq
->removed
.lock
);
3407 swap(cfs_rq
->removed
.util_avg
, removed_util
);
3408 swap(cfs_rq
->removed
.load_avg
, removed_load
);
3409 swap(cfs_rq
->removed
.runnable_sum
, removed_runnable_sum
);
3410 cfs_rq
->removed
.nr
= 0;
3411 raw_spin_unlock(&cfs_rq
->removed
.lock
);
3414 sub_positive(&sa
->load_avg
, r
);
3415 sub_positive(&sa
->load_sum
, r
* divider
);
3418 sub_positive(&sa
->util_avg
, r
);
3419 sub_positive(&sa
->util_sum
, r
* divider
);
3421 add_tg_cfs_propagate(cfs_rq
, -(long)removed_runnable_sum
);
3426 decayed
|= __update_load_avg_cfs_rq(now
, cfs_rq
);
3428 #ifndef CONFIG_64BIT
3430 cfs_rq
->load_last_update_time_copy
= sa
->last_update_time
;
3434 cfs_rq_util_change(cfs_rq
, 0);
3440 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3441 * @cfs_rq: cfs_rq to attach to
3442 * @se: sched_entity to attach
3443 * @flags: migration hints
3445 * Must call update_cfs_rq_load_avg() before this, since we rely on
3446 * cfs_rq->avg.last_update_time being current.
3448 static void attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3450 u32 divider
= LOAD_AVG_MAX
- 1024 + cfs_rq
->avg
.period_contrib
;
3453 * When we attach the @se to the @cfs_rq, we must align the decay
3454 * window because without that, really weird and wonderful things can
3459 se
->avg
.last_update_time
= cfs_rq
->avg
.last_update_time
;
3460 se
->avg
.period_contrib
= cfs_rq
->avg
.period_contrib
;
3463 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3464 * period_contrib. This isn't strictly correct, but since we're
3465 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3468 se
->avg
.util_sum
= se
->avg
.util_avg
* divider
;
3470 se
->avg
.load_sum
= divider
;
3471 if (se_weight(se
)) {
3473 div_u64(se
->avg
.load_avg
* se
->avg
.load_sum
, se_weight(se
));
3476 se
->avg
.runnable_load_sum
= se
->avg
.load_sum
;
3478 enqueue_load_avg(cfs_rq
, se
);
3479 cfs_rq
->avg
.util_avg
+= se
->avg
.util_avg
;
3480 cfs_rq
->avg
.util_sum
+= se
->avg
.util_sum
;
3482 add_tg_cfs_propagate(cfs_rq
, se
->avg
.load_sum
);
3484 cfs_rq_util_change(cfs_rq
, flags
);
3488 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3489 * @cfs_rq: cfs_rq to detach from
3490 * @se: sched_entity to detach
3492 * Must call update_cfs_rq_load_avg() before this, since we rely on
3493 * cfs_rq->avg.last_update_time being current.
3495 static void detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3497 dequeue_load_avg(cfs_rq
, se
);
3498 sub_positive(&cfs_rq
->avg
.util_avg
, se
->avg
.util_avg
);
3499 sub_positive(&cfs_rq
->avg
.util_sum
, se
->avg
.util_sum
);
3501 add_tg_cfs_propagate(cfs_rq
, -se
->avg
.load_sum
);
3503 cfs_rq_util_change(cfs_rq
, 0);
3507 * Optional action to be done while updating the load average
3509 #define UPDATE_TG 0x1
3510 #define SKIP_AGE_LOAD 0x2
3511 #define DO_ATTACH 0x4
3513 /* Update task and its cfs_rq load average */
3514 static inline void update_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3516 u64 now
= cfs_rq_clock_pelt(cfs_rq
);
3520 * Track task load average for carrying it to new CPU after migrated, and
3521 * track group sched_entity load average for task_h_load calc in migration
3523 if (se
->avg
.last_update_time
&& !(flags
& SKIP_AGE_LOAD
))
3524 __update_load_avg_se(now
, cfs_rq
, se
);
3526 decayed
= update_cfs_rq_load_avg(now
, cfs_rq
);
3527 decayed
|= propagate_entity_load_avg(se
);
3529 if (!se
->avg
.last_update_time
&& (flags
& DO_ATTACH
)) {
3532 * DO_ATTACH means we're here from enqueue_entity().
3533 * !last_update_time means we've passed through
3534 * migrate_task_rq_fair() indicating we migrated.
3536 * IOW we're enqueueing a task on a new CPU.
3538 attach_entity_load_avg(cfs_rq
, se
, SCHED_CPUFREQ_MIGRATION
);
3539 update_tg_load_avg(cfs_rq
, 0);
3541 } else if (decayed
&& (flags
& UPDATE_TG
))
3542 update_tg_load_avg(cfs_rq
, 0);
3545 #ifndef CONFIG_64BIT
3546 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3548 u64 last_update_time_copy
;
3549 u64 last_update_time
;
3552 last_update_time_copy
= cfs_rq
->load_last_update_time_copy
;
3554 last_update_time
= cfs_rq
->avg
.last_update_time
;
3555 } while (last_update_time
!= last_update_time_copy
);
3557 return last_update_time
;
3560 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3562 return cfs_rq
->avg
.last_update_time
;
3567 * Synchronize entity load avg of dequeued entity without locking
3570 void sync_entity_load_avg(struct sched_entity
*se
)
3572 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3573 u64 last_update_time
;
3575 last_update_time
= cfs_rq_last_update_time(cfs_rq
);
3576 __update_load_avg_blocked_se(last_update_time
, se
);
3580 * Task first catches up with cfs_rq, and then subtract
3581 * itself from the cfs_rq (task must be off the queue now).
3583 void remove_entity_load_avg(struct sched_entity
*se
)
3585 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3586 unsigned long flags
;
3589 * tasks cannot exit without having gone through wake_up_new_task() ->
3590 * post_init_entity_util_avg() which will have added things to the
3591 * cfs_rq, so we can remove unconditionally.
3594 sync_entity_load_avg(se
);
3596 raw_spin_lock_irqsave(&cfs_rq
->removed
.lock
, flags
);
3597 ++cfs_rq
->removed
.nr
;
3598 cfs_rq
->removed
.util_avg
+= se
->avg
.util_avg
;
3599 cfs_rq
->removed
.load_avg
+= se
->avg
.load_avg
;
3600 cfs_rq
->removed
.runnable_sum
+= se
->avg
.load_sum
; /* == runnable_sum */
3601 raw_spin_unlock_irqrestore(&cfs_rq
->removed
.lock
, flags
);
3604 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq
*cfs_rq
)
3606 return cfs_rq
->avg
.runnable_load_avg
;
3609 static inline unsigned long cfs_rq_load_avg(struct cfs_rq
*cfs_rq
)
3611 return cfs_rq
->avg
.load_avg
;
3614 static int idle_balance(struct rq
*this_rq
, struct rq_flags
*rf
);
3616 static inline unsigned long task_util(struct task_struct
*p
)
3618 return READ_ONCE(p
->se
.avg
.util_avg
);
3621 static inline unsigned long _task_util_est(struct task_struct
*p
)
3623 struct util_est ue
= READ_ONCE(p
->se
.avg
.util_est
);
3625 return (max(ue
.ewma
, ue
.enqueued
) | UTIL_AVG_UNCHANGED
);
3628 static inline unsigned long task_util_est(struct task_struct
*p
)
3630 return max(task_util(p
), _task_util_est(p
));
3633 static inline void util_est_enqueue(struct cfs_rq
*cfs_rq
,
3634 struct task_struct
*p
)
3636 unsigned int enqueued
;
3638 if (!sched_feat(UTIL_EST
))
3641 /* Update root cfs_rq's estimated utilization */
3642 enqueued
= cfs_rq
->avg
.util_est
.enqueued
;
3643 enqueued
+= _task_util_est(p
);
3644 WRITE_ONCE(cfs_rq
->avg
.util_est
.enqueued
, enqueued
);
3648 * Check if a (signed) value is within a specified (unsigned) margin,
3649 * based on the observation that:
3651 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3653 * NOTE: this only works when value + maring < INT_MAX.
3655 static inline bool within_margin(int value
, int margin
)
3657 return ((unsigned int)(value
+ margin
- 1) < (2 * margin
- 1));
3661 util_est_dequeue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
, bool task_sleep
)
3663 long last_ewma_diff
;
3667 if (!sched_feat(UTIL_EST
))
3670 /* Update root cfs_rq's estimated utilization */
3671 ue
.enqueued
= cfs_rq
->avg
.util_est
.enqueued
;
3672 ue
.enqueued
-= min_t(unsigned int, ue
.enqueued
, _task_util_est(p
));
3673 WRITE_ONCE(cfs_rq
->avg
.util_est
.enqueued
, ue
.enqueued
);
3676 * Skip update of task's estimated utilization when the task has not
3677 * yet completed an activation, e.g. being migrated.
3683 * If the PELT values haven't changed since enqueue time,
3684 * skip the util_est update.
3686 ue
= p
->se
.avg
.util_est
;
3687 if (ue
.enqueued
& UTIL_AVG_UNCHANGED
)
3691 * Skip update of task's estimated utilization when its EWMA is
3692 * already ~1% close to its last activation value.
3694 ue
.enqueued
= (task_util(p
) | UTIL_AVG_UNCHANGED
);
3695 last_ewma_diff
= ue
.enqueued
- ue
.ewma
;
3696 if (within_margin(last_ewma_diff
, (SCHED_CAPACITY_SCALE
/ 100)))
3700 * To avoid overestimation of actual task utilization, skip updates if
3701 * we cannot grant there is idle time in this CPU.
3703 cpu
= cpu_of(rq_of(cfs_rq
));
3704 if (task_util(p
) > capacity_orig_of(cpu
))
3708 * Update Task's estimated utilization
3710 * When *p completes an activation we can consolidate another sample
3711 * of the task size. This is done by storing the current PELT value
3712 * as ue.enqueued and by using this value to update the Exponential
3713 * Weighted Moving Average (EWMA):
3715 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3716 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3717 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3718 * = w * ( last_ewma_diff ) + ewma(t-1)
3719 * = w * (last_ewma_diff + ewma(t-1) / w)
3721 * Where 'w' is the weight of new samples, which is configured to be
3722 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3724 ue
.ewma
<<= UTIL_EST_WEIGHT_SHIFT
;
3725 ue
.ewma
+= last_ewma_diff
;
3726 ue
.ewma
>>= UTIL_EST_WEIGHT_SHIFT
;
3727 WRITE_ONCE(p
->se
.avg
.util_est
, ue
);
3730 static inline int task_fits_capacity(struct task_struct
*p
, long capacity
)
3732 return capacity
* 1024 > task_util_est(p
) * capacity_margin
;
3735 static inline void update_misfit_status(struct task_struct
*p
, struct rq
*rq
)
3737 if (!static_branch_unlikely(&sched_asym_cpucapacity
))
3741 rq
->misfit_task_load
= 0;
3745 if (task_fits_capacity(p
, capacity_of(cpu_of(rq
)))) {
3746 rq
->misfit_task_load
= 0;
3750 rq
->misfit_task_load
= task_h_load(p
);
3753 #else /* CONFIG_SMP */
3755 #define UPDATE_TG 0x0
3756 #define SKIP_AGE_LOAD 0x0
3757 #define DO_ATTACH 0x0
3759 static inline void update_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int not_used1
)
3761 cfs_rq_util_change(cfs_rq
, 0);
3764 static inline void remove_entity_load_avg(struct sched_entity
*se
) {}
3767 attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
) {}
3769 detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3771 static inline int idle_balance(struct rq
*rq
, struct rq_flags
*rf
)
3777 util_est_enqueue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
) {}
3780 util_est_dequeue(struct cfs_rq
*cfs_rq
, struct task_struct
*p
,
3782 static inline void update_misfit_status(struct task_struct
*p
, struct rq
*rq
) {}
3784 #endif /* CONFIG_SMP */
3786 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3788 #ifdef CONFIG_SCHED_DEBUG
3789 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
3794 if (d
> 3*sysctl_sched_latency
)
3795 schedstat_inc(cfs_rq
->nr_spread_over
);
3800 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
3802 u64 vruntime
= cfs_rq
->min_vruntime
;
3805 * The 'current' period is already promised to the current tasks,
3806 * however the extra weight of the new task will slow them down a
3807 * little, place the new task so that it fits in the slot that
3808 * stays open at the end.
3810 if (initial
&& sched_feat(START_DEBIT
))
3811 vruntime
+= sched_vslice(cfs_rq
, se
);
3813 /* sleeps up to a single latency don't count. */
3815 unsigned long thresh
= sysctl_sched_latency
;
3818 * Halve their sleep time's effect, to allow
3819 * for a gentler effect of sleepers:
3821 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
3827 /* ensure we never gain time by being placed backwards. */
3828 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
3831 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
3833 static inline void check_schedstat_required(void)
3835 #ifdef CONFIG_SCHEDSTATS
3836 if (schedstat_enabled())
3839 /* Force schedstat enabled if a dependent tracepoint is active */
3840 if (trace_sched_stat_wait_enabled() ||
3841 trace_sched_stat_sleep_enabled() ||
3842 trace_sched_stat_iowait_enabled() ||
3843 trace_sched_stat_blocked_enabled() ||
3844 trace_sched_stat_runtime_enabled()) {
3845 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3846 "stat_blocked and stat_runtime require the "
3847 "kernel parameter schedstats=enable or "
3848 "kernel.sched_schedstats=1\n");
3859 * update_min_vruntime()
3860 * vruntime -= min_vruntime
3864 * update_min_vruntime()
3865 * vruntime += min_vruntime
3867 * this way the vruntime transition between RQs is done when both
3868 * min_vruntime are up-to-date.
3872 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3873 * vruntime -= min_vruntime
3877 * update_min_vruntime()
3878 * vruntime += min_vruntime
3880 * this way we don't have the most up-to-date min_vruntime on the originating
3881 * CPU and an up-to-date min_vruntime on the destination CPU.
3885 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3887 bool renorm
= !(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_MIGRATED
);
3888 bool curr
= cfs_rq
->curr
== se
;
3891 * If we're the current task, we must renormalise before calling
3895 se
->vruntime
+= cfs_rq
->min_vruntime
;
3897 update_curr(cfs_rq
);
3900 * Otherwise, renormalise after, such that we're placed at the current
3901 * moment in time, instead of some random moment in the past. Being
3902 * placed in the past could significantly boost this task to the
3903 * fairness detriment of existing tasks.
3905 if (renorm
&& !curr
)
3906 se
->vruntime
+= cfs_rq
->min_vruntime
;
3909 * When enqueuing a sched_entity, we must:
3910 * - Update loads to have both entity and cfs_rq synced with now.
3911 * - Add its load to cfs_rq->runnable_avg
3912 * - For group_entity, update its weight to reflect the new share of
3914 * - Add its new weight to cfs_rq->load.weight
3916 update_load_avg(cfs_rq
, se
, UPDATE_TG
| DO_ATTACH
);
3917 update_cfs_group(se
);
3918 enqueue_runnable_load_avg(cfs_rq
, se
);
3919 account_entity_enqueue(cfs_rq
, se
);
3921 if (flags
& ENQUEUE_WAKEUP
)
3922 place_entity(cfs_rq
, se
, 0);
3924 check_schedstat_required();
3925 update_stats_enqueue(cfs_rq
, se
, flags
);
3926 check_spread(cfs_rq
, se
);
3928 __enqueue_entity(cfs_rq
, se
);
3931 if (cfs_rq
->nr_running
== 1) {
3932 list_add_leaf_cfs_rq(cfs_rq
);
3933 check_enqueue_throttle(cfs_rq
);
3937 static void __clear_buddies_last(struct sched_entity
*se
)
3939 for_each_sched_entity(se
) {
3940 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3941 if (cfs_rq
->last
!= se
)
3944 cfs_rq
->last
= NULL
;
3948 static void __clear_buddies_next(struct sched_entity
*se
)
3950 for_each_sched_entity(se
) {
3951 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3952 if (cfs_rq
->next
!= se
)
3955 cfs_rq
->next
= NULL
;
3959 static void __clear_buddies_skip(struct sched_entity
*se
)
3961 for_each_sched_entity(se
) {
3962 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3963 if (cfs_rq
->skip
!= se
)
3966 cfs_rq
->skip
= NULL
;
3970 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3972 if (cfs_rq
->last
== se
)
3973 __clear_buddies_last(se
);
3975 if (cfs_rq
->next
== se
)
3976 __clear_buddies_next(se
);
3978 if (cfs_rq
->skip
== se
)
3979 __clear_buddies_skip(se
);
3982 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3985 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3988 * Update run-time statistics of the 'current'.
3990 update_curr(cfs_rq
);
3993 * When dequeuing a sched_entity, we must:
3994 * - Update loads to have both entity and cfs_rq synced with now.
3995 * - Subtract its load from the cfs_rq->runnable_avg.
3996 * - Subtract its previous weight from cfs_rq->load.weight.
3997 * - For group entity, update its weight to reflect the new share
3998 * of its group cfs_rq.
4000 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4001 dequeue_runnable_load_avg(cfs_rq
, se
);
4003 update_stats_dequeue(cfs_rq
, se
, flags
);
4005 clear_buddies(cfs_rq
, se
);
4007 if (se
!= cfs_rq
->curr
)
4008 __dequeue_entity(cfs_rq
, se
);
4010 account_entity_dequeue(cfs_rq
, se
);
4013 * Normalize after update_curr(); which will also have moved
4014 * min_vruntime if @se is the one holding it back. But before doing
4015 * update_min_vruntime() again, which will discount @se's position and
4016 * can move min_vruntime forward still more.
4018 if (!(flags
& DEQUEUE_SLEEP
))
4019 se
->vruntime
-= cfs_rq
->min_vruntime
;
4021 /* return excess runtime on last dequeue */
4022 return_cfs_rq_runtime(cfs_rq
);
4024 update_cfs_group(se
);
4027 * Now advance min_vruntime if @se was the entity holding it back,
4028 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4029 * put back on, and if we advance min_vruntime, we'll be placed back
4030 * further than we started -- ie. we'll be penalized.
4032 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) != DEQUEUE_SAVE
)
4033 update_min_vruntime(cfs_rq
);
4037 * Preempt the current task with a newly woken task if needed:
4040 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
4042 unsigned long ideal_runtime
, delta_exec
;
4043 struct sched_entity
*se
;
4046 ideal_runtime
= sched_slice(cfs_rq
, curr
);
4047 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
4048 if (delta_exec
> ideal_runtime
) {
4049 resched_curr(rq_of(cfs_rq
));
4051 * The current task ran long enough, ensure it doesn't get
4052 * re-elected due to buddy favours.
4054 clear_buddies(cfs_rq
, curr
);
4059 * Ensure that a task that missed wakeup preemption by a
4060 * narrow margin doesn't have to wait for a full slice.
4061 * This also mitigates buddy induced latencies under load.
4063 if (delta_exec
< sysctl_sched_min_granularity
)
4066 se
= __pick_first_entity(cfs_rq
);
4067 delta
= curr
->vruntime
- se
->vruntime
;
4072 if (delta
> ideal_runtime
)
4073 resched_curr(rq_of(cfs_rq
));
4077 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4079 /* 'current' is not kept within the tree. */
4082 * Any task has to be enqueued before it get to execute on
4083 * a CPU. So account for the time it spent waiting on the
4086 update_stats_wait_end(cfs_rq
, se
);
4087 __dequeue_entity(cfs_rq
, se
);
4088 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4091 update_stats_curr_start(cfs_rq
, se
);
4095 * Track our maximum slice length, if the CPU's load is at
4096 * least twice that of our own weight (i.e. dont track it
4097 * when there are only lesser-weight tasks around):
4099 if (schedstat_enabled() && rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
4100 schedstat_set(se
->statistics
.slice_max
,
4101 max((u64
)schedstat_val(se
->statistics
.slice_max
),
4102 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
));
4105 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
4109 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
4112 * Pick the next process, keeping these things in mind, in this order:
4113 * 1) keep things fair between processes/task groups
4114 * 2) pick the "next" process, since someone really wants that to run
4115 * 3) pick the "last" process, for cache locality
4116 * 4) do not run the "skip" process, if something else is available
4118 static struct sched_entity
*
4119 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
4121 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
4122 struct sched_entity
*se
;
4125 * If curr is set we have to see if its left of the leftmost entity
4126 * still in the tree, provided there was anything in the tree at all.
4128 if (!left
|| (curr
&& entity_before(curr
, left
)))
4131 se
= left
; /* ideally we run the leftmost entity */
4134 * Avoid running the skip buddy, if running something else can
4135 * be done without getting too unfair.
4137 if (cfs_rq
->skip
== se
) {
4138 struct sched_entity
*second
;
4141 second
= __pick_first_entity(cfs_rq
);
4143 second
= __pick_next_entity(se
);
4144 if (!second
|| (curr
&& entity_before(curr
, second
)))
4148 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
4153 * Prefer last buddy, try to return the CPU to a preempted task.
4155 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
4159 * Someone really wants this to run. If it's not unfair, run it.
4161 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
4164 clear_buddies(cfs_rq
, se
);
4169 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
4171 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
4174 * If still on the runqueue then deactivate_task()
4175 * was not called and update_curr() has to be done:
4178 update_curr(cfs_rq
);
4180 /* throttle cfs_rqs exceeding runtime */
4181 check_cfs_rq_runtime(cfs_rq
);
4183 check_spread(cfs_rq
, prev
);
4186 update_stats_wait_start(cfs_rq
, prev
);
4187 /* Put 'current' back into the tree. */
4188 __enqueue_entity(cfs_rq
, prev
);
4189 /* in !on_rq case, update occurred at dequeue */
4190 update_load_avg(cfs_rq
, prev
, 0);
4192 cfs_rq
->curr
= NULL
;
4196 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
4199 * Update run-time statistics of the 'current'.
4201 update_curr(cfs_rq
);
4204 * Ensure that runnable average is periodically updated.
4206 update_load_avg(cfs_rq
, curr
, UPDATE_TG
);
4207 update_cfs_group(curr
);
4209 #ifdef CONFIG_SCHED_HRTICK
4211 * queued ticks are scheduled to match the slice, so don't bother
4212 * validating it and just reschedule.
4215 resched_curr(rq_of(cfs_rq
));
4219 * don't let the period tick interfere with the hrtick preemption
4221 if (!sched_feat(DOUBLE_TICK
) &&
4222 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
4226 if (cfs_rq
->nr_running
> 1)
4227 check_preempt_tick(cfs_rq
, curr
);
4231 /**************************************************
4232 * CFS bandwidth control machinery
4235 #ifdef CONFIG_CFS_BANDWIDTH
4237 #ifdef CONFIG_JUMP_LABEL
4238 static struct static_key __cfs_bandwidth_used
;
4240 static inline bool cfs_bandwidth_used(void)
4242 return static_key_false(&__cfs_bandwidth_used
);
4245 void cfs_bandwidth_usage_inc(void)
4247 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used
);
4250 void cfs_bandwidth_usage_dec(void)
4252 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used
);
4254 #else /* CONFIG_JUMP_LABEL */
4255 static bool cfs_bandwidth_used(void)
4260 void cfs_bandwidth_usage_inc(void) {}
4261 void cfs_bandwidth_usage_dec(void) {}
4262 #endif /* CONFIG_JUMP_LABEL */
4265 * default period for cfs group bandwidth.
4266 * default: 0.1s, units: nanoseconds
4268 static inline u64
default_cfs_period(void)
4270 return 100000000ULL;
4273 static inline u64
sched_cfs_bandwidth_slice(void)
4275 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
4279 * Replenish runtime according to assigned quota and update expiration time.
4280 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4281 * additional synchronization around rq->lock.
4283 * requires cfs_b->lock
4285 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
4289 if (cfs_b
->quota
== RUNTIME_INF
)
4292 now
= sched_clock_cpu(smp_processor_id());
4293 cfs_b
->runtime
= cfs_b
->quota
;
4294 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
4295 cfs_b
->expires_seq
++;
4298 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4300 return &tg
->cfs_bandwidth
;
4303 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4304 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
4306 if (unlikely(cfs_rq
->throttle_count
))
4307 return cfs_rq
->throttled_clock_task
- cfs_rq
->throttled_clock_task_time
;
4309 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
4312 /* returns 0 on failure to allocate runtime */
4313 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4315 struct task_group
*tg
= cfs_rq
->tg
;
4316 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
4317 u64 amount
= 0, min_amount
, expires
;
4320 /* note: this is a positive sum as runtime_remaining <= 0 */
4321 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
4323 raw_spin_lock(&cfs_b
->lock
);
4324 if (cfs_b
->quota
== RUNTIME_INF
)
4325 amount
= min_amount
;
4327 start_cfs_bandwidth(cfs_b
);
4329 if (cfs_b
->runtime
> 0) {
4330 amount
= min(cfs_b
->runtime
, min_amount
);
4331 cfs_b
->runtime
-= amount
;
4335 expires_seq
= cfs_b
->expires_seq
;
4336 expires
= cfs_b
->runtime_expires
;
4337 raw_spin_unlock(&cfs_b
->lock
);
4339 cfs_rq
->runtime_remaining
+= amount
;
4341 * we may have advanced our local expiration to account for allowed
4342 * spread between our sched_clock and the one on which runtime was
4345 if (cfs_rq
->expires_seq
!= expires_seq
) {
4346 cfs_rq
->expires_seq
= expires_seq
;
4347 cfs_rq
->runtime_expires
= expires
;
4350 return cfs_rq
->runtime_remaining
> 0;
4354 * Note: This depends on the synchronization provided by sched_clock and the
4355 * fact that rq->clock snapshots this value.
4357 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4359 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4361 /* if the deadline is ahead of our clock, nothing to do */
4362 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
4365 if (cfs_rq
->runtime_remaining
< 0)
4369 * If the local deadline has passed we have to consider the
4370 * possibility that our sched_clock is 'fast' and the global deadline
4371 * has not truly expired.
4373 * Fortunately we can check determine whether this the case by checking
4374 * whether the global deadline(cfs_b->expires_seq) has advanced.
4376 if (cfs_rq
->expires_seq
== cfs_b
->expires_seq
) {
4377 /* extend local deadline, drift is bounded above by 2 ticks */
4378 cfs_rq
->runtime_expires
+= TICK_NSEC
;
4380 /* global deadline is ahead, expiration has passed */
4381 cfs_rq
->runtime_remaining
= 0;
4385 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4387 /* dock delta_exec before expiring quota (as it could span periods) */
4388 cfs_rq
->runtime_remaining
-= delta_exec
;
4389 expire_cfs_rq_runtime(cfs_rq
);
4391 if (likely(cfs_rq
->runtime_remaining
> 0))
4395 * if we're unable to extend our runtime we resched so that the active
4396 * hierarchy can be throttled
4398 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
4399 resched_curr(rq_of(cfs_rq
));
4402 static __always_inline
4403 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4405 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
4408 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
4411 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4413 return cfs_bandwidth_used() && cfs_rq
->throttled
;
4416 /* check whether cfs_rq, or any parent, is throttled */
4417 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4419 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
4423 * Ensure that neither of the group entities corresponding to src_cpu or
4424 * dest_cpu are members of a throttled hierarchy when performing group
4425 * load-balance operations.
4427 static inline int throttled_lb_pair(struct task_group
*tg
,
4428 int src_cpu
, int dest_cpu
)
4430 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
4432 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
4433 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
4435 return throttled_hierarchy(src_cfs_rq
) ||
4436 throttled_hierarchy(dest_cfs_rq
);
4439 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
4441 struct rq
*rq
= data
;
4442 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4444 cfs_rq
->throttle_count
--;
4445 if (!cfs_rq
->throttle_count
) {
4446 /* adjust cfs_rq_clock_task() */
4447 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
4448 cfs_rq
->throttled_clock_task
;
4450 /* Add cfs_rq with already running entity in the list */
4451 if (cfs_rq
->nr_running
>= 1)
4452 list_add_leaf_cfs_rq(cfs_rq
);
4458 static int tg_throttle_down(struct task_group
*tg
, void *data
)
4460 struct rq
*rq
= data
;
4461 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4463 /* group is entering throttled state, stop time */
4464 if (!cfs_rq
->throttle_count
) {
4465 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
4466 list_del_leaf_cfs_rq(cfs_rq
);
4468 cfs_rq
->throttle_count
++;
4473 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
4475 struct rq
*rq
= rq_of(cfs_rq
);
4476 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4477 struct sched_entity
*se
;
4478 long task_delta
, dequeue
= 1;
4481 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
4483 /* freeze hierarchy runnable averages while throttled */
4485 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
4488 task_delta
= cfs_rq
->h_nr_running
;
4489 for_each_sched_entity(se
) {
4490 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4491 /* throttled entity or throttle-on-deactivate */
4496 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
4497 qcfs_rq
->h_nr_running
-= task_delta
;
4499 if (qcfs_rq
->load
.weight
)
4504 sub_nr_running(rq
, task_delta
);
4506 cfs_rq
->throttled
= 1;
4507 cfs_rq
->throttled_clock
= rq_clock(rq
);
4508 raw_spin_lock(&cfs_b
->lock
);
4509 empty
= list_empty(&cfs_b
->throttled_cfs_rq
);
4512 * Add to the _head_ of the list, so that an already-started
4513 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4514 * not running add to the tail so that later runqueues don't get starved.
4516 if (cfs_b
->distribute_running
)
4517 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
4519 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
4522 * If we're the first throttled task, make sure the bandwidth
4526 start_cfs_bandwidth(cfs_b
);
4528 raw_spin_unlock(&cfs_b
->lock
);
4531 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
4533 struct rq
*rq
= rq_of(cfs_rq
);
4534 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4535 struct sched_entity
*se
;
4539 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
4541 cfs_rq
->throttled
= 0;
4543 update_rq_clock(rq
);
4545 raw_spin_lock(&cfs_b
->lock
);
4546 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
4547 list_del_rcu(&cfs_rq
->throttled_list
);
4548 raw_spin_unlock(&cfs_b
->lock
);
4550 /* update hierarchical throttle state */
4551 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
4553 if (!cfs_rq
->load
.weight
)
4556 task_delta
= cfs_rq
->h_nr_running
;
4557 for_each_sched_entity(se
) {
4561 cfs_rq
= cfs_rq_of(se
);
4563 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
4564 cfs_rq
->h_nr_running
+= task_delta
;
4566 if (cfs_rq_throttled(cfs_rq
))
4570 assert_list_leaf_cfs_rq(rq
);
4573 add_nr_running(rq
, task_delta
);
4575 /* Determine whether we need to wake up potentially idle CPU: */
4576 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
4580 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
4581 u64 remaining
, u64 expires
)
4583 struct cfs_rq
*cfs_rq
;
4585 u64 starting_runtime
= remaining
;
4588 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
4590 struct rq
*rq
= rq_of(cfs_rq
);
4593 rq_lock_irqsave(rq
, &rf
);
4594 if (!cfs_rq_throttled(cfs_rq
))
4597 runtime
= -cfs_rq
->runtime_remaining
+ 1;
4598 if (runtime
> remaining
)
4599 runtime
= remaining
;
4600 remaining
-= runtime
;
4602 cfs_rq
->runtime_remaining
+= runtime
;
4603 cfs_rq
->runtime_expires
= expires
;
4605 /* we check whether we're throttled above */
4606 if (cfs_rq
->runtime_remaining
> 0)
4607 unthrottle_cfs_rq(cfs_rq
);
4610 rq_unlock_irqrestore(rq
, &rf
);
4617 return starting_runtime
- remaining
;
4621 * Responsible for refilling a task_group's bandwidth and unthrottling its
4622 * cfs_rqs as appropriate. If there has been no activity within the last
4623 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4624 * used to track this state.
4626 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
, unsigned long flags
)
4628 u64 runtime
, runtime_expires
;
4631 /* no need to continue the timer with no bandwidth constraint */
4632 if (cfs_b
->quota
== RUNTIME_INF
)
4633 goto out_deactivate
;
4635 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4636 cfs_b
->nr_periods
+= overrun
;
4639 * idle depends on !throttled (for the case of a large deficit), and if
4640 * we're going inactive then everything else can be deferred
4642 if (cfs_b
->idle
&& !throttled
)
4643 goto out_deactivate
;
4645 __refill_cfs_bandwidth_runtime(cfs_b
);
4648 /* mark as potentially idle for the upcoming period */
4653 /* account preceding periods in which throttling occurred */
4654 cfs_b
->nr_throttled
+= overrun
;
4656 runtime_expires
= cfs_b
->runtime_expires
;
4659 * This check is repeated as we are holding onto the new bandwidth while
4660 * we unthrottle. This can potentially race with an unthrottled group
4661 * trying to acquire new bandwidth from the global pool. This can result
4662 * in us over-using our runtime if it is all used during this loop, but
4663 * only by limited amounts in that extreme case.
4665 while (throttled
&& cfs_b
->runtime
> 0 && !cfs_b
->distribute_running
) {
4666 runtime
= cfs_b
->runtime
;
4667 cfs_b
->distribute_running
= 1;
4668 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
4669 /* we can't nest cfs_b->lock while distributing bandwidth */
4670 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
4672 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
4674 cfs_b
->distribute_running
= 0;
4675 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4677 lsub_positive(&cfs_b
->runtime
, runtime
);
4681 * While we are ensured activity in the period following an
4682 * unthrottle, this also covers the case in which the new bandwidth is
4683 * insufficient to cover the existing bandwidth deficit. (Forcing the
4684 * timer to remain active while there are any throttled entities.)
4694 /* a cfs_rq won't donate quota below this amount */
4695 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
4696 /* minimum remaining period time to redistribute slack quota */
4697 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
4698 /* how long we wait to gather additional slack before distributing */
4699 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
4702 * Are we near the end of the current quota period?
4704 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4705 * hrtimer base being cleared by hrtimer_start. In the case of
4706 * migrate_hrtimers, base is never cleared, so we are fine.
4708 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
4710 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
4713 /* if the call-back is running a quota refresh is already occurring */
4714 if (hrtimer_callback_running(refresh_timer
))
4717 /* is a quota refresh about to occur? */
4718 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
4719 if (remaining
< min_expire
)
4725 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
4727 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
4729 /* if there's a quota refresh soon don't bother with slack */
4730 if (runtime_refresh_within(cfs_b
, min_left
))
4733 hrtimer_start(&cfs_b
->slack_timer
,
4734 ns_to_ktime(cfs_bandwidth_slack_period
),
4738 /* we know any runtime found here is valid as update_curr() precedes return */
4739 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4741 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4742 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
4744 if (slack_runtime
<= 0)
4747 raw_spin_lock(&cfs_b
->lock
);
4748 if (cfs_b
->quota
!= RUNTIME_INF
&&
4749 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
4750 cfs_b
->runtime
+= slack_runtime
;
4752 /* we are under rq->lock, defer unthrottling using a timer */
4753 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
4754 !list_empty(&cfs_b
->throttled_cfs_rq
))
4755 start_cfs_slack_bandwidth(cfs_b
);
4757 raw_spin_unlock(&cfs_b
->lock
);
4759 /* even if it's not valid for return we don't want to try again */
4760 cfs_rq
->runtime_remaining
-= slack_runtime
;
4763 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4765 if (!cfs_bandwidth_used())
4768 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
4771 __return_cfs_rq_runtime(cfs_rq
);
4775 * This is done with a timer (instead of inline with bandwidth return) since
4776 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4778 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
4780 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
4781 unsigned long flags
;
4784 /* confirm we're still not at a refresh boundary */
4785 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
4786 if (cfs_b
->distribute_running
) {
4787 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
4791 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
4792 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
4796 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
4797 runtime
= cfs_b
->runtime
;
4799 expires
= cfs_b
->runtime_expires
;
4801 cfs_b
->distribute_running
= 1;
4803 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
4808 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
4810 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
4811 if (expires
== cfs_b
->runtime_expires
)
4812 lsub_positive(&cfs_b
->runtime
, runtime
);
4813 cfs_b
->distribute_running
= 0;
4814 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
4818 * When a group wakes up we want to make sure that its quota is not already
4819 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4820 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4822 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
4824 if (!cfs_bandwidth_used())
4827 /* an active group must be handled by the update_curr()->put() path */
4828 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
4831 /* ensure the group is not already throttled */
4832 if (cfs_rq_throttled(cfs_rq
))
4835 /* update runtime allocation */
4836 account_cfs_rq_runtime(cfs_rq
, 0);
4837 if (cfs_rq
->runtime_remaining
<= 0)
4838 throttle_cfs_rq(cfs_rq
);
4841 static void sync_throttle(struct task_group
*tg
, int cpu
)
4843 struct cfs_rq
*pcfs_rq
, *cfs_rq
;
4845 if (!cfs_bandwidth_used())
4851 cfs_rq
= tg
->cfs_rq
[cpu
];
4852 pcfs_rq
= tg
->parent
->cfs_rq
[cpu
];
4854 cfs_rq
->throttle_count
= pcfs_rq
->throttle_count
;
4855 cfs_rq
->throttled_clock_task
= rq_clock_task(cpu_rq(cpu
));
4858 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4859 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4861 if (!cfs_bandwidth_used())
4864 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
4868 * it's possible for a throttled entity to be forced into a running
4869 * state (e.g. set_curr_task), in this case we're finished.
4871 if (cfs_rq_throttled(cfs_rq
))
4874 throttle_cfs_rq(cfs_rq
);
4878 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
4880 struct cfs_bandwidth
*cfs_b
=
4881 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
4883 do_sched_cfs_slack_timer(cfs_b
);
4885 return HRTIMER_NORESTART
;
4888 extern const u64 max_cfs_quota_period
;
4890 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
4892 struct cfs_bandwidth
*cfs_b
=
4893 container_of(timer
, struct cfs_bandwidth
, period_timer
);
4894 unsigned long flags
;
4899 raw_spin_lock_irqsave(&cfs_b
->lock
, flags
);
4901 overrun
= hrtimer_forward_now(timer
, cfs_b
->period
);
4906 u64
new, old
= ktime_to_ns(cfs_b
->period
);
4908 new = (old
* 147) / 128; /* ~115% */
4909 new = min(new, max_cfs_quota_period
);
4911 cfs_b
->period
= ns_to_ktime(new);
4913 /* since max is 1s, this is limited to 1e9^2, which fits in u64 */
4914 cfs_b
->quota
*= new;
4915 cfs_b
->quota
= div64_u64(cfs_b
->quota
, old
);
4917 pr_warn_ratelimited(
4918 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us %lld, cfs_quota_us = %lld)\n",
4920 div_u64(new, NSEC_PER_USEC
),
4921 div_u64(cfs_b
->quota
, NSEC_PER_USEC
));
4923 /* reset count so we don't come right back in here */
4927 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
, flags
);
4930 cfs_b
->period_active
= 0;
4931 raw_spin_unlock_irqrestore(&cfs_b
->lock
, flags
);
4933 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
4936 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4938 raw_spin_lock_init(&cfs_b
->lock
);
4940 cfs_b
->quota
= RUNTIME_INF
;
4941 cfs_b
->period
= ns_to_ktime(default_cfs_period());
4943 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
4944 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_PINNED
);
4945 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
4946 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4947 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
4948 cfs_b
->distribute_running
= 0;
4951 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4953 cfs_rq
->runtime_enabled
= 0;
4954 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
4957 void start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4961 lockdep_assert_held(&cfs_b
->lock
);
4963 if (cfs_b
->period_active
)
4966 cfs_b
->period_active
= 1;
4967 overrun
= hrtimer_forward_now(&cfs_b
->period_timer
, cfs_b
->period
);
4968 cfs_b
->runtime_expires
+= (overrun
+ 1) * ktime_to_ns(cfs_b
->period
);
4969 cfs_b
->expires_seq
++;
4970 hrtimer_start_expires(&cfs_b
->period_timer
, HRTIMER_MODE_ABS_PINNED
);
4973 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4975 /* init_cfs_bandwidth() was not called */
4976 if (!cfs_b
->throttled_cfs_rq
.next
)
4979 hrtimer_cancel(&cfs_b
->period_timer
);
4980 hrtimer_cancel(&cfs_b
->slack_timer
);
4984 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
4986 * The race is harmless, since modifying bandwidth settings of unhooked group
4987 * bits doesn't do much.
4990 /* cpu online calback */
4991 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
4993 struct task_group
*tg
;
4995 lockdep_assert_held(&rq
->lock
);
4998 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
4999 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
5000 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
5002 raw_spin_lock(&cfs_b
->lock
);
5003 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
5004 raw_spin_unlock(&cfs_b
->lock
);
5009 /* cpu offline callback */
5010 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
5012 struct task_group
*tg
;
5014 lockdep_assert_held(&rq
->lock
);
5017 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
5018 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
5020 if (!cfs_rq
->runtime_enabled
)
5024 * clock_task is not advancing so we just need to make sure
5025 * there's some valid quota amount
5027 cfs_rq
->runtime_remaining
= 1;
5029 * Offline rq is schedulable till CPU is completely disabled
5030 * in take_cpu_down(), so we prevent new cfs throttling here.
5032 cfs_rq
->runtime_enabled
= 0;
5034 if (cfs_rq_throttled(cfs_rq
))
5035 unthrottle_cfs_rq(cfs_rq
);
5040 #else /* CONFIG_CFS_BANDWIDTH */
5042 static inline bool cfs_bandwidth_used(void)
5047 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
5049 return rq_clock_task(rq_of(cfs_rq
));
5052 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
5053 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
5054 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
5055 static inline void sync_throttle(struct task_group
*tg
, int cpu
) {}
5056 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
5058 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
5063 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
5068 static inline int throttled_lb_pair(struct task_group
*tg
,
5069 int src_cpu
, int dest_cpu
)
5074 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
5076 #ifdef CONFIG_FAIR_GROUP_SCHED
5077 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
5080 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
5084 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
5085 static inline void update_runtime_enabled(struct rq
*rq
) {}
5086 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
5088 #endif /* CONFIG_CFS_BANDWIDTH */
5090 /**************************************************
5091 * CFS operations on tasks:
5094 #ifdef CONFIG_SCHED_HRTICK
5095 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
5097 struct sched_entity
*se
= &p
->se
;
5098 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5100 SCHED_WARN_ON(task_rq(p
) != rq
);
5102 if (rq
->cfs
.h_nr_running
> 1) {
5103 u64 slice
= sched_slice(cfs_rq
, se
);
5104 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
5105 s64 delta
= slice
- ran
;
5112 hrtick_start(rq
, delta
);
5117 * called from enqueue/dequeue and updates the hrtick when the
5118 * current task is from our class and nr_running is low enough
5121 static void hrtick_update(struct rq
*rq
)
5123 struct task_struct
*curr
= rq
->curr
;
5125 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
5128 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
5129 hrtick_start_fair(rq
, curr
);
5131 #else /* !CONFIG_SCHED_HRTICK */
5133 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
5137 static inline void hrtick_update(struct rq
*rq
)
5143 static inline unsigned long cpu_util(int cpu
);
5144 static unsigned long capacity_of(int cpu
);
5146 static inline bool cpu_overutilized(int cpu
)
5148 return (capacity_of(cpu
) * 1024) < (cpu_util(cpu
) * capacity_margin
);
5151 static inline void update_overutilized_status(struct rq
*rq
)
5153 if (!READ_ONCE(rq
->rd
->overutilized
) && cpu_overutilized(rq
->cpu
))
5154 WRITE_ONCE(rq
->rd
->overutilized
, SG_OVERUTILIZED
);
5157 static inline void update_overutilized_status(struct rq
*rq
) { }
5161 * The enqueue_task method is called before nr_running is
5162 * increased. Here we update the fair scheduling stats and
5163 * then put the task into the rbtree:
5166 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
5168 struct cfs_rq
*cfs_rq
;
5169 struct sched_entity
*se
= &p
->se
;
5172 * The code below (indirectly) updates schedutil which looks at
5173 * the cfs_rq utilization to select a frequency.
5174 * Let's add the task's estimated utilization to the cfs_rq's
5175 * estimated utilization, before we update schedutil.
5177 util_est_enqueue(&rq
->cfs
, p
);
5180 * If in_iowait is set, the code below may not trigger any cpufreq
5181 * utilization updates, so do it here explicitly with the IOWAIT flag
5185 cpufreq_update_util(rq
, SCHED_CPUFREQ_IOWAIT
);
5187 for_each_sched_entity(se
) {
5190 cfs_rq
= cfs_rq_of(se
);
5191 enqueue_entity(cfs_rq
, se
, flags
);
5194 * end evaluation on encountering a throttled cfs_rq
5196 * note: in the case of encountering a throttled cfs_rq we will
5197 * post the final h_nr_running increment below.
5199 if (cfs_rq_throttled(cfs_rq
))
5201 cfs_rq
->h_nr_running
++;
5203 flags
= ENQUEUE_WAKEUP
;
5206 for_each_sched_entity(se
) {
5207 cfs_rq
= cfs_rq_of(se
);
5208 cfs_rq
->h_nr_running
++;
5210 if (cfs_rq_throttled(cfs_rq
))
5213 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
5214 update_cfs_group(se
);
5218 add_nr_running(rq
, 1);
5220 * Since new tasks are assigned an initial util_avg equal to
5221 * half of the spare capacity of their CPU, tiny tasks have the
5222 * ability to cross the overutilized threshold, which will
5223 * result in the load balancer ruining all the task placement
5224 * done by EAS. As a way to mitigate that effect, do not account
5225 * for the first enqueue operation of new tasks during the
5226 * overutilized flag detection.
5228 * A better way of solving this problem would be to wait for
5229 * the PELT signals of tasks to converge before taking them
5230 * into account, but that is not straightforward to implement,
5231 * and the following generally works well enough in practice.
5233 if (flags
& ENQUEUE_WAKEUP
)
5234 update_overutilized_status(rq
);
5238 if (cfs_bandwidth_used()) {
5240 * When bandwidth control is enabled; the cfs_rq_throttled()
5241 * breaks in the above iteration can result in incomplete
5242 * leaf list maintenance, resulting in triggering the assertion
5245 for_each_sched_entity(se
) {
5246 cfs_rq
= cfs_rq_of(se
);
5248 if (list_add_leaf_cfs_rq(cfs_rq
))
5253 assert_list_leaf_cfs_rq(rq
);
5258 static void set_next_buddy(struct sched_entity
*se
);
5261 * The dequeue_task method is called before nr_running is
5262 * decreased. We remove the task from the rbtree and
5263 * update the fair scheduling stats:
5265 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
5267 struct cfs_rq
*cfs_rq
;
5268 struct sched_entity
*se
= &p
->se
;
5269 int task_sleep
= flags
& DEQUEUE_SLEEP
;
5271 for_each_sched_entity(se
) {
5272 cfs_rq
= cfs_rq_of(se
);
5273 dequeue_entity(cfs_rq
, se
, flags
);
5276 * end evaluation on encountering a throttled cfs_rq
5278 * note: in the case of encountering a throttled cfs_rq we will
5279 * post the final h_nr_running decrement below.
5281 if (cfs_rq_throttled(cfs_rq
))
5283 cfs_rq
->h_nr_running
--;
5285 /* Don't dequeue parent if it has other entities besides us */
5286 if (cfs_rq
->load
.weight
) {
5287 /* Avoid re-evaluating load for this entity: */
5288 se
= parent_entity(se
);
5290 * Bias pick_next to pick a task from this cfs_rq, as
5291 * p is sleeping when it is within its sched_slice.
5293 if (task_sleep
&& se
&& !throttled_hierarchy(cfs_rq
))
5297 flags
|= DEQUEUE_SLEEP
;
5300 for_each_sched_entity(se
) {
5301 cfs_rq
= cfs_rq_of(se
);
5302 cfs_rq
->h_nr_running
--;
5304 if (cfs_rq_throttled(cfs_rq
))
5307 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
5308 update_cfs_group(se
);
5312 sub_nr_running(rq
, 1);
5314 util_est_dequeue(&rq
->cfs
, p
, task_sleep
);
5320 /* Working cpumask for: load_balance, load_balance_newidle. */
5321 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5322 DEFINE_PER_CPU(cpumask_var_t
, select_idle_mask
);
5324 #ifdef CONFIG_NO_HZ_COMMON
5326 * per rq 'load' arrray crap; XXX kill this.
5330 * The exact cpuload calculated at every tick would be:
5332 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5334 * If a CPU misses updates for n ticks (as it was idle) and update gets
5335 * called on the n+1-th tick when CPU may be busy, then we have:
5337 * load_n = (1 - 1/2^i)^n * load_0
5338 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5340 * decay_load_missed() below does efficient calculation of
5342 * load' = (1 - 1/2^i)^n * load
5344 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5345 * This allows us to precompute the above in said factors, thereby allowing the
5346 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5347 * fixed_power_int())
5349 * The calculation is approximated on a 128 point scale.
5351 #define DEGRADE_SHIFT 7
5353 static const u8 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
5354 static const u8 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
5355 { 0, 0, 0, 0, 0, 0, 0, 0 },
5356 { 64, 32, 8, 0, 0, 0, 0, 0 },
5357 { 96, 72, 40, 12, 1, 0, 0, 0 },
5358 { 112, 98, 75, 43, 15, 1, 0, 0 },
5359 { 120, 112, 98, 76, 45, 16, 2, 0 }
5363 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5364 * would be when CPU is idle and so we just decay the old load without
5365 * adding any new load.
5367 static unsigned long
5368 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
5372 if (!missed_updates
)
5375 if (missed_updates
>= degrade_zero_ticks
[idx
])
5379 return load
>> missed_updates
;
5381 while (missed_updates
) {
5382 if (missed_updates
% 2)
5383 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
5385 missed_updates
>>= 1;
5392 cpumask_var_t idle_cpus_mask
;
5394 int has_blocked
; /* Idle CPUS has blocked load */
5395 unsigned long next_balance
; /* in jiffy units */
5396 unsigned long next_blocked
; /* Next update of blocked load in jiffies */
5397 } nohz ____cacheline_aligned
;
5399 #endif /* CONFIG_NO_HZ_COMMON */
5402 * __cpu_load_update - update the rq->cpu_load[] statistics
5403 * @this_rq: The rq to update statistics for
5404 * @this_load: The current load
5405 * @pending_updates: The number of missed updates
5407 * Update rq->cpu_load[] statistics. This function is usually called every
5408 * scheduler tick (TICK_NSEC).
5410 * This function computes a decaying average:
5412 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5414 * Because of NOHZ it might not get called on every tick which gives need for
5415 * the @pending_updates argument.
5417 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5418 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5419 * = A * (A * load[i]_n-2 + B) + B
5420 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5421 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5422 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5423 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5424 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5426 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5427 * any change in load would have resulted in the tick being turned back on.
5429 * For regular NOHZ, this reduces to:
5431 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5433 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5436 static void cpu_load_update(struct rq
*this_rq
, unsigned long this_load
,
5437 unsigned long pending_updates
)
5439 unsigned long __maybe_unused tickless_load
= this_rq
->cpu_load
[0];
5442 this_rq
->nr_load_updates
++;
5444 /* Update our load: */
5445 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
5446 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
5447 unsigned long old_load
, new_load
;
5449 /* scale is effectively 1 << i now, and >> i divides by scale */
5451 old_load
= this_rq
->cpu_load
[i
];
5452 #ifdef CONFIG_NO_HZ_COMMON
5453 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
5454 if (tickless_load
) {
5455 old_load
-= decay_load_missed(tickless_load
, pending_updates
- 1, i
);
5457 * old_load can never be a negative value because a
5458 * decayed tickless_load cannot be greater than the
5459 * original tickless_load.
5461 old_load
+= tickless_load
;
5464 new_load
= this_load
;
5466 * Round up the averaging division if load is increasing. This
5467 * prevents us from getting stuck on 9 if the load is 10, for
5470 if (new_load
> old_load
)
5471 new_load
+= scale
- 1;
5473 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
5477 /* Used instead of source_load when we know the type == 0 */
5478 static unsigned long weighted_cpuload(struct rq
*rq
)
5480 return cfs_rq_runnable_load_avg(&rq
->cfs
);
5483 #ifdef CONFIG_NO_HZ_COMMON
5485 * There is no sane way to deal with nohz on smp when using jiffies because the
5486 * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5487 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5489 * Therefore we need to avoid the delta approach from the regular tick when
5490 * possible since that would seriously skew the load calculation. This is why we
5491 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5492 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5493 * loop exit, nohz_idle_balance, nohz full exit...)
5495 * This means we might still be one tick off for nohz periods.
5498 static void cpu_load_update_nohz(struct rq
*this_rq
,
5499 unsigned long curr_jiffies
,
5502 unsigned long pending_updates
;
5504 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
5505 if (pending_updates
) {
5506 this_rq
->last_load_update_tick
= curr_jiffies
;
5508 * In the regular NOHZ case, we were idle, this means load 0.
5509 * In the NOHZ_FULL case, we were non-idle, we should consider
5510 * its weighted load.
5512 cpu_load_update(this_rq
, load
, pending_updates
);
5517 * Called from nohz_idle_balance() to update the load ratings before doing the
5520 static void cpu_load_update_idle(struct rq
*this_rq
)
5523 * bail if there's load or we're actually up-to-date.
5525 if (weighted_cpuload(this_rq
))
5528 cpu_load_update_nohz(this_rq
, READ_ONCE(jiffies
), 0);
5532 * Record CPU load on nohz entry so we know the tickless load to account
5533 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5534 * than other cpu_load[idx] but it should be fine as cpu_load readers
5535 * shouldn't rely into synchronized cpu_load[*] updates.
5537 void cpu_load_update_nohz_start(void)
5539 struct rq
*this_rq
= this_rq();
5542 * This is all lockless but should be fine. If weighted_cpuload changes
5543 * concurrently we'll exit nohz. And cpu_load write can race with
5544 * cpu_load_update_idle() but both updater would be writing the same.
5546 this_rq
->cpu_load
[0] = weighted_cpuload(this_rq
);
5550 * Account the tickless load in the end of a nohz frame.
5552 void cpu_load_update_nohz_stop(void)
5554 unsigned long curr_jiffies
= READ_ONCE(jiffies
);
5555 struct rq
*this_rq
= this_rq();
5559 if (curr_jiffies
== this_rq
->last_load_update_tick
)
5562 load
= weighted_cpuload(this_rq
);
5563 rq_lock(this_rq
, &rf
);
5564 update_rq_clock(this_rq
);
5565 cpu_load_update_nohz(this_rq
, curr_jiffies
, load
);
5566 rq_unlock(this_rq
, &rf
);
5568 #else /* !CONFIG_NO_HZ_COMMON */
5569 static inline void cpu_load_update_nohz(struct rq
*this_rq
,
5570 unsigned long curr_jiffies
,
5571 unsigned long load
) { }
5572 #endif /* CONFIG_NO_HZ_COMMON */
5574 static void cpu_load_update_periodic(struct rq
*this_rq
, unsigned long load
)
5576 #ifdef CONFIG_NO_HZ_COMMON
5577 /* See the mess around cpu_load_update_nohz(). */
5578 this_rq
->last_load_update_tick
= READ_ONCE(jiffies
);
5580 cpu_load_update(this_rq
, load
, 1);
5584 * Called from scheduler_tick()
5586 void cpu_load_update_active(struct rq
*this_rq
)
5588 unsigned long load
= weighted_cpuload(this_rq
);
5590 if (tick_nohz_tick_stopped())
5591 cpu_load_update_nohz(this_rq
, READ_ONCE(jiffies
), load
);
5593 cpu_load_update_periodic(this_rq
, load
);
5597 * Return a low guess at the load of a migration-source CPU weighted
5598 * according to the scheduling class and "nice" value.
5600 * We want to under-estimate the load of migration sources, to
5601 * balance conservatively.
5603 static unsigned long source_load(int cpu
, int type
)
5605 struct rq
*rq
= cpu_rq(cpu
);
5606 unsigned long total
= weighted_cpuload(rq
);
5608 if (type
== 0 || !sched_feat(LB_BIAS
))
5611 return min(rq
->cpu_load
[type
-1], total
);
5615 * Return a high guess at the load of a migration-target CPU weighted
5616 * according to the scheduling class and "nice" value.
5618 static unsigned long target_load(int cpu
, int type
)
5620 struct rq
*rq
= cpu_rq(cpu
);
5621 unsigned long total
= weighted_cpuload(rq
);
5623 if (type
== 0 || !sched_feat(LB_BIAS
))
5626 return max(rq
->cpu_load
[type
-1], total
);
5629 static unsigned long capacity_of(int cpu
)
5631 return cpu_rq(cpu
)->cpu_capacity
;
5634 static unsigned long cpu_avg_load_per_task(int cpu
)
5636 struct rq
*rq
= cpu_rq(cpu
);
5637 unsigned long nr_running
= READ_ONCE(rq
->cfs
.h_nr_running
);
5638 unsigned long load_avg
= weighted_cpuload(rq
);
5641 return load_avg
/ nr_running
;
5646 static void record_wakee(struct task_struct
*p
)
5649 * Only decay a single time; tasks that have less then 1 wakeup per
5650 * jiffy will not have built up many flips.
5652 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
5653 current
->wakee_flips
>>= 1;
5654 current
->wakee_flip_decay_ts
= jiffies
;
5657 if (current
->last_wakee
!= p
) {
5658 current
->last_wakee
= p
;
5659 current
->wakee_flips
++;
5664 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5666 * A waker of many should wake a different task than the one last awakened
5667 * at a frequency roughly N times higher than one of its wakees.
5669 * In order to determine whether we should let the load spread vs consolidating
5670 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5671 * partner, and a factor of lls_size higher frequency in the other.
5673 * With both conditions met, we can be relatively sure that the relationship is
5674 * non-monogamous, with partner count exceeding socket size.
5676 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5677 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5680 static int wake_wide(struct task_struct
*p
)
5682 unsigned int master
= current
->wakee_flips
;
5683 unsigned int slave
= p
->wakee_flips
;
5684 int factor
= this_cpu_read(sd_llc_size
);
5687 swap(master
, slave
);
5688 if (slave
< factor
|| master
< slave
* factor
)
5694 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5695 * soonest. For the purpose of speed we only consider the waking and previous
5698 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5699 * cache-affine and is (or will be) idle.
5701 * wake_affine_weight() - considers the weight to reflect the average
5702 * scheduling latency of the CPUs. This seems to work
5703 * for the overloaded case.
5706 wake_affine_idle(int this_cpu
, int prev_cpu
, int sync
)
5709 * If this_cpu is idle, it implies the wakeup is from interrupt
5710 * context. Only allow the move if cache is shared. Otherwise an
5711 * interrupt intensive workload could force all tasks onto one
5712 * node depending on the IO topology or IRQ affinity settings.
5714 * If the prev_cpu is idle and cache affine then avoid a migration.
5715 * There is no guarantee that the cache hot data from an interrupt
5716 * is more important than cache hot data on the prev_cpu and from
5717 * a cpufreq perspective, it's better to have higher utilisation
5720 if (available_idle_cpu(this_cpu
) && cpus_share_cache(this_cpu
, prev_cpu
))
5721 return available_idle_cpu(prev_cpu
) ? prev_cpu
: this_cpu
;
5723 if (sync
&& cpu_rq(this_cpu
)->nr_running
== 1)
5726 return nr_cpumask_bits
;
5730 wake_affine_weight(struct sched_domain
*sd
, struct task_struct
*p
,
5731 int this_cpu
, int prev_cpu
, int sync
)
5733 s64 this_eff_load
, prev_eff_load
;
5734 unsigned long task_load
;
5736 this_eff_load
= target_load(this_cpu
, sd
->wake_idx
);
5739 unsigned long current_load
= task_h_load(current
);
5741 if (current_load
> this_eff_load
)
5744 this_eff_load
-= current_load
;
5747 task_load
= task_h_load(p
);
5749 this_eff_load
+= task_load
;
5750 if (sched_feat(WA_BIAS
))
5751 this_eff_load
*= 100;
5752 this_eff_load
*= capacity_of(prev_cpu
);
5754 prev_eff_load
= source_load(prev_cpu
, sd
->wake_idx
);
5755 prev_eff_load
-= task_load
;
5756 if (sched_feat(WA_BIAS
))
5757 prev_eff_load
*= 100 + (sd
->imbalance_pct
- 100) / 2;
5758 prev_eff_load
*= capacity_of(this_cpu
);
5761 * If sync, adjust the weight of prev_eff_load such that if
5762 * prev_eff == this_eff that select_idle_sibling() will consider
5763 * stacking the wakee on top of the waker if no other CPU is
5769 return this_eff_load
< prev_eff_load
? this_cpu
: nr_cpumask_bits
;
5772 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
,
5773 int this_cpu
, int prev_cpu
, int sync
)
5775 int target
= nr_cpumask_bits
;
5777 if (sched_feat(WA_IDLE
))
5778 target
= wake_affine_idle(this_cpu
, prev_cpu
, sync
);
5780 if (sched_feat(WA_WEIGHT
) && target
== nr_cpumask_bits
)
5781 target
= wake_affine_weight(sd
, p
, this_cpu
, prev_cpu
, sync
);
5783 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine_attempts
);
5784 if (target
== nr_cpumask_bits
)
5787 schedstat_inc(sd
->ttwu_move_affine
);
5788 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine
);
5792 static unsigned long cpu_util_without(int cpu
, struct task_struct
*p
);
5794 static unsigned long capacity_spare_without(int cpu
, struct task_struct
*p
)
5796 return max_t(long, capacity_of(cpu
) - cpu_util_without(cpu
, p
), 0);
5800 * find_idlest_group finds and returns the least busy CPU group within the
5803 * Assumes p is allowed on at least one CPU in sd.
5805 static struct sched_group
*
5806 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
5807 int this_cpu
, int sd_flag
)
5809 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
5810 struct sched_group
*most_spare_sg
= NULL
;
5811 unsigned long min_runnable_load
= ULONG_MAX
;
5812 unsigned long this_runnable_load
= ULONG_MAX
;
5813 unsigned long min_avg_load
= ULONG_MAX
, this_avg_load
= ULONG_MAX
;
5814 unsigned long most_spare
= 0, this_spare
= 0;
5815 int load_idx
= sd
->forkexec_idx
;
5816 int imbalance_scale
= 100 + (sd
->imbalance_pct
-100)/2;
5817 unsigned long imbalance
= scale_load_down(NICE_0_LOAD
) *
5818 (sd
->imbalance_pct
-100) / 100;
5820 if (sd_flag
& SD_BALANCE_WAKE
)
5821 load_idx
= sd
->wake_idx
;
5824 unsigned long load
, avg_load
, runnable_load
;
5825 unsigned long spare_cap
, max_spare_cap
;
5829 /* Skip over this group if it has no CPUs allowed */
5830 if (!cpumask_intersects(sched_group_span(group
),
5834 local_group
= cpumask_test_cpu(this_cpu
,
5835 sched_group_span(group
));
5838 * Tally up the load of all CPUs in the group and find
5839 * the group containing the CPU with most spare capacity.
5845 for_each_cpu(i
, sched_group_span(group
)) {
5846 /* Bias balancing toward CPUs of our domain */
5848 load
= source_load(i
, load_idx
);
5850 load
= target_load(i
, load_idx
);
5852 runnable_load
+= load
;
5854 avg_load
+= cfs_rq_load_avg(&cpu_rq(i
)->cfs
);
5856 spare_cap
= capacity_spare_without(i
, p
);
5858 if (spare_cap
> max_spare_cap
)
5859 max_spare_cap
= spare_cap
;
5862 /* Adjust by relative CPU capacity of the group */
5863 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) /
5864 group
->sgc
->capacity
;
5865 runnable_load
= (runnable_load
* SCHED_CAPACITY_SCALE
) /
5866 group
->sgc
->capacity
;
5869 this_runnable_load
= runnable_load
;
5870 this_avg_load
= avg_load
;
5871 this_spare
= max_spare_cap
;
5873 if (min_runnable_load
> (runnable_load
+ imbalance
)) {
5875 * The runnable load is significantly smaller
5876 * so we can pick this new CPU:
5878 min_runnable_load
= runnable_load
;
5879 min_avg_load
= avg_load
;
5881 } else if ((runnable_load
< (min_runnable_load
+ imbalance
)) &&
5882 (100*min_avg_load
> imbalance_scale
*avg_load
)) {
5884 * The runnable loads are close so take the
5885 * blocked load into account through avg_load:
5887 min_avg_load
= avg_load
;
5891 if (most_spare
< max_spare_cap
) {
5892 most_spare
= max_spare_cap
;
5893 most_spare_sg
= group
;
5896 } while (group
= group
->next
, group
!= sd
->groups
);
5899 * The cross-over point between using spare capacity or least load
5900 * is too conservative for high utilization tasks on partially
5901 * utilized systems if we require spare_capacity > task_util(p),
5902 * so we allow for some task stuffing by using
5903 * spare_capacity > task_util(p)/2.
5905 * Spare capacity can't be used for fork because the utilization has
5906 * not been set yet, we must first select a rq to compute the initial
5909 if (sd_flag
& SD_BALANCE_FORK
)
5912 if (this_spare
> task_util(p
) / 2 &&
5913 imbalance_scale
*this_spare
> 100*most_spare
)
5916 if (most_spare
> task_util(p
) / 2)
5917 return most_spare_sg
;
5924 * When comparing groups across NUMA domains, it's possible for the
5925 * local domain to be very lightly loaded relative to the remote
5926 * domains but "imbalance" skews the comparison making remote CPUs
5927 * look much more favourable. When considering cross-domain, add
5928 * imbalance to the runnable load on the remote node and consider
5931 if ((sd
->flags
& SD_NUMA
) &&
5932 min_runnable_load
+ imbalance
>= this_runnable_load
)
5935 if (min_runnable_load
> (this_runnable_load
+ imbalance
))
5938 if ((this_runnable_load
< (min_runnable_load
+ imbalance
)) &&
5939 (100*this_avg_load
< imbalance_scale
*min_avg_load
))
5946 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5949 find_idlest_group_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
5951 unsigned long load
, min_load
= ULONG_MAX
;
5952 unsigned int min_exit_latency
= UINT_MAX
;
5953 u64 latest_idle_timestamp
= 0;
5954 int least_loaded_cpu
= this_cpu
;
5955 int shallowest_idle_cpu
= -1;
5958 /* Check if we have any choice: */
5959 if (group
->group_weight
== 1)
5960 return cpumask_first(sched_group_span(group
));
5962 /* Traverse only the allowed CPUs */
5963 for_each_cpu_and(i
, sched_group_span(group
), &p
->cpus_allowed
) {
5964 if (available_idle_cpu(i
)) {
5965 struct rq
*rq
= cpu_rq(i
);
5966 struct cpuidle_state
*idle
= idle_get_state(rq
);
5967 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
5969 * We give priority to a CPU whose idle state
5970 * has the smallest exit latency irrespective
5971 * of any idle timestamp.
5973 min_exit_latency
= idle
->exit_latency
;
5974 latest_idle_timestamp
= rq
->idle_stamp
;
5975 shallowest_idle_cpu
= i
;
5976 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
5977 rq
->idle_stamp
> latest_idle_timestamp
) {
5979 * If equal or no active idle state, then
5980 * the most recently idled CPU might have
5983 latest_idle_timestamp
= rq
->idle_stamp
;
5984 shallowest_idle_cpu
= i
;
5986 } else if (shallowest_idle_cpu
== -1) {
5987 load
= weighted_cpuload(cpu_rq(i
));
5988 if (load
< min_load
) {
5990 least_loaded_cpu
= i
;
5995 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
5998 static inline int find_idlest_cpu(struct sched_domain
*sd
, struct task_struct
*p
,
5999 int cpu
, int prev_cpu
, int sd_flag
)
6003 if (!cpumask_intersects(sched_domain_span(sd
), &p
->cpus_allowed
))
6007 * We need task's util for capacity_spare_without, sync it up to
6008 * prev_cpu's last_update_time.
6010 if (!(sd_flag
& SD_BALANCE_FORK
))
6011 sync_entity_load_avg(&p
->se
);
6014 struct sched_group
*group
;
6015 struct sched_domain
*tmp
;
6018 if (!(sd
->flags
& sd_flag
)) {
6023 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
6029 new_cpu
= find_idlest_group_cpu(group
, p
, cpu
);
6030 if (new_cpu
== cpu
) {
6031 /* Now try balancing at a lower domain level of 'cpu': */
6036 /* Now try balancing at a lower domain level of 'new_cpu': */
6038 weight
= sd
->span_weight
;
6040 for_each_domain(cpu
, tmp
) {
6041 if (weight
<= tmp
->span_weight
)
6043 if (tmp
->flags
& sd_flag
)
6051 #ifdef CONFIG_SCHED_SMT
6052 DEFINE_STATIC_KEY_FALSE(sched_smt_present
);
6053 EXPORT_SYMBOL_GPL(sched_smt_present
);
6055 static inline void set_idle_cores(int cpu
, int val
)
6057 struct sched_domain_shared
*sds
;
6059 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
6061 WRITE_ONCE(sds
->has_idle_cores
, val
);
6064 static inline bool test_idle_cores(int cpu
, bool def
)
6066 struct sched_domain_shared
*sds
;
6068 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
6070 return READ_ONCE(sds
->has_idle_cores
);
6076 * Scans the local SMT mask to see if the entire core is idle, and records this
6077 * information in sd_llc_shared->has_idle_cores.
6079 * Since SMT siblings share all cache levels, inspecting this limited remote
6080 * state should be fairly cheap.
6082 void __update_idle_core(struct rq
*rq
)
6084 int core
= cpu_of(rq
);
6088 if (test_idle_cores(core
, true))
6091 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
6095 if (!available_idle_cpu(cpu
))
6099 set_idle_cores(core
, 1);
6105 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6106 * there are no idle cores left in the system; tracked through
6107 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6109 static int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6111 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6114 if (!static_branch_likely(&sched_smt_present
))
6117 if (!test_idle_cores(target
, false))
6120 cpumask_and(cpus
, sched_domain_span(sd
), &p
->cpus_allowed
);
6122 for_each_cpu_wrap(core
, cpus
, target
) {
6125 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
6126 __cpumask_clear_cpu(cpu
, cpus
);
6127 if (!available_idle_cpu(cpu
))
6136 * Failed to find an idle core; stop looking for one.
6138 set_idle_cores(target
, 0);
6144 * Scan the local SMT mask for idle CPUs.
6146 static int select_idle_smt(struct task_struct
*p
, int target
)
6150 if (!static_branch_likely(&sched_smt_present
))
6153 for_each_cpu(cpu
, cpu_smt_mask(target
)) {
6154 if (!cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
6156 if (available_idle_cpu(cpu
))
6163 #else /* CONFIG_SCHED_SMT */
6165 static inline int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6170 static inline int select_idle_smt(struct task_struct
*p
, int target
)
6175 #endif /* CONFIG_SCHED_SMT */
6178 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6179 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6180 * average idle time for this rq (as found in rq->avg_idle).
6182 static int select_idle_cpu(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6184 struct sched_domain
*this_sd
;
6185 u64 avg_cost
, avg_idle
;
6188 int cpu
, nr
= INT_MAX
;
6190 this_sd
= rcu_dereference(*this_cpu_ptr(&sd_llc
));
6195 * Due to large variance we need a large fuzz factor; hackbench in
6196 * particularly is sensitive here.
6198 avg_idle
= this_rq()->avg_idle
/ 512;
6199 avg_cost
= this_sd
->avg_scan_cost
+ 1;
6201 if (sched_feat(SIS_AVG_CPU
) && avg_idle
< avg_cost
)
6204 if (sched_feat(SIS_PROP
)) {
6205 u64 span_avg
= sd
->span_weight
* avg_idle
;
6206 if (span_avg
> 4*avg_cost
)
6207 nr
= div_u64(span_avg
, avg_cost
);
6212 time
= local_clock();
6214 for_each_cpu_wrap(cpu
, sched_domain_span(sd
), target
) {
6217 if (!cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
6219 if (available_idle_cpu(cpu
))
6223 time
= local_clock() - time
;
6224 cost
= this_sd
->avg_scan_cost
;
6225 delta
= (s64
)(time
- cost
) / 8;
6226 this_sd
->avg_scan_cost
+= delta
;
6232 * Try and locate an idle core/thread in the LLC cache domain.
6234 static int select_idle_sibling(struct task_struct
*p
, int prev
, int target
)
6236 struct sched_domain
*sd
;
6237 int i
, recent_used_cpu
;
6239 if (available_idle_cpu(target
))
6243 * If the previous CPU is cache affine and idle, don't be stupid:
6245 if (prev
!= target
&& cpus_share_cache(prev
, target
) && available_idle_cpu(prev
))
6248 /* Check a recently used CPU as a potential idle candidate: */
6249 recent_used_cpu
= p
->recent_used_cpu
;
6250 if (recent_used_cpu
!= prev
&&
6251 recent_used_cpu
!= target
&&
6252 cpus_share_cache(recent_used_cpu
, target
) &&
6253 available_idle_cpu(recent_used_cpu
) &&
6254 cpumask_test_cpu(p
->recent_used_cpu
, &p
->cpus_allowed
)) {
6256 * Replace recent_used_cpu with prev as it is a potential
6257 * candidate for the next wake:
6259 p
->recent_used_cpu
= prev
;
6260 return recent_used_cpu
;
6263 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
6267 i
= select_idle_core(p
, sd
, target
);
6268 if ((unsigned)i
< nr_cpumask_bits
)
6271 i
= select_idle_cpu(p
, sd
, target
);
6272 if ((unsigned)i
< nr_cpumask_bits
)
6275 i
= select_idle_smt(p
, target
);
6276 if ((unsigned)i
< nr_cpumask_bits
)
6283 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6284 * @cpu: the CPU to get the utilization of
6286 * The unit of the return value must be the one of capacity so we can compare
6287 * the utilization with the capacity of the CPU that is available for CFS task
6288 * (ie cpu_capacity).
6290 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6291 * recent utilization of currently non-runnable tasks on a CPU. It represents
6292 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6293 * capacity_orig is the cpu_capacity available at the highest frequency
6294 * (arch_scale_freq_capacity()).
6295 * The utilization of a CPU converges towards a sum equal to or less than the
6296 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6297 * the running time on this CPU scaled by capacity_curr.
6299 * The estimated utilization of a CPU is defined to be the maximum between its
6300 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6301 * currently RUNNABLE on that CPU.
6302 * This allows to properly represent the expected utilization of a CPU which
6303 * has just got a big task running since a long sleep period. At the same time
6304 * however it preserves the benefits of the "blocked utilization" in
6305 * describing the potential for other tasks waking up on the same CPU.
6307 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6308 * higher than capacity_orig because of unfortunate rounding in
6309 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6310 * the average stabilizes with the new running time. We need to check that the
6311 * utilization stays within the range of [0..capacity_orig] and cap it if
6312 * necessary. Without utilization capping, a group could be seen as overloaded
6313 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6314 * available capacity. We allow utilization to overshoot capacity_curr (but not
6315 * capacity_orig) as it useful for predicting the capacity required after task
6316 * migrations (scheduler-driven DVFS).
6318 * Return: the (estimated) utilization for the specified CPU
6320 static inline unsigned long cpu_util(int cpu
)
6322 struct cfs_rq
*cfs_rq
;
6325 cfs_rq
= &cpu_rq(cpu
)->cfs
;
6326 util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6328 if (sched_feat(UTIL_EST
))
6329 util
= max(util
, READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
));
6331 return min_t(unsigned long, util
, capacity_orig_of(cpu
));
6335 * cpu_util_without: compute cpu utilization without any contributions from *p
6336 * @cpu: the CPU which utilization is requested
6337 * @p: the task which utilization should be discounted
6339 * The utilization of a CPU is defined by the utilization of tasks currently
6340 * enqueued on that CPU as well as tasks which are currently sleeping after an
6341 * execution on that CPU.
6343 * This method returns the utilization of the specified CPU by discounting the
6344 * utilization of the specified task, whenever the task is currently
6345 * contributing to the CPU utilization.
6347 static unsigned long cpu_util_without(int cpu
, struct task_struct
*p
)
6349 struct cfs_rq
*cfs_rq
;
6352 /* Task has no contribution or is new */
6353 if (cpu
!= task_cpu(p
) || !READ_ONCE(p
->se
.avg
.last_update_time
))
6354 return cpu_util(cpu
);
6356 cfs_rq
= &cpu_rq(cpu
)->cfs
;
6357 util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6359 /* Discount task's util from CPU's util */
6360 lsub_positive(&util
, task_util(p
));
6365 * a) if *p is the only task sleeping on this CPU, then:
6366 * cpu_util (== task_util) > util_est (== 0)
6367 * and thus we return:
6368 * cpu_util_without = (cpu_util - task_util) = 0
6370 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6372 * cpu_util >= task_util
6373 * cpu_util > util_est (== 0)
6374 * and thus we discount *p's blocked utilization to return:
6375 * cpu_util_without = (cpu_util - task_util) >= 0
6377 * c) if other tasks are RUNNABLE on that CPU and
6378 * util_est > cpu_util
6379 * then we use util_est since it returns a more restrictive
6380 * estimation of the spare capacity on that CPU, by just
6381 * considering the expected utilization of tasks already
6382 * runnable on that CPU.
6384 * Cases a) and b) are covered by the above code, while case c) is
6385 * covered by the following code when estimated utilization is
6388 if (sched_feat(UTIL_EST
)) {
6389 unsigned int estimated
=
6390 READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
);
6393 * Despite the following checks we still have a small window
6394 * for a possible race, when an execl's select_task_rq_fair()
6395 * races with LB's detach_task():
6398 * p->on_rq = TASK_ON_RQ_MIGRATING;
6399 * ---------------------------------- A
6400 * deactivate_task() \
6401 * dequeue_task() + RaceTime
6402 * util_est_dequeue() /
6403 * ---------------------------------- B
6405 * The additional check on "current == p" it's required to
6406 * properly fix the execl regression and it helps in further
6407 * reducing the chances for the above race.
6409 if (unlikely(task_on_rq_queued(p
) || current
== p
))
6410 lsub_positive(&estimated
, _task_util_est(p
));
6412 util
= max(util
, estimated
);
6416 * Utilization (estimated) can exceed the CPU capacity, thus let's
6417 * clamp to the maximum CPU capacity to ensure consistency with
6418 * the cpu_util call.
6420 return min_t(unsigned long, util
, capacity_orig_of(cpu
));
6424 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6425 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6427 * In that case WAKE_AFFINE doesn't make sense and we'll let
6428 * BALANCE_WAKE sort things out.
6430 static int wake_cap(struct task_struct
*p
, int cpu
, int prev_cpu
)
6432 long min_cap
, max_cap
;
6434 if (!static_branch_unlikely(&sched_asym_cpucapacity
))
6437 min_cap
= min(capacity_orig_of(prev_cpu
), capacity_orig_of(cpu
));
6438 max_cap
= cpu_rq(cpu
)->rd
->max_cpu_capacity
;
6440 /* Minimum capacity is close to max, no need to abort wake_affine */
6441 if (max_cap
- min_cap
< max_cap
>> 3)
6444 /* Bring task utilization in sync with prev_cpu */
6445 sync_entity_load_avg(&p
->se
);
6447 return !task_fits_capacity(p
, min_cap
);
6451 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6454 static unsigned long cpu_util_next(int cpu
, struct task_struct
*p
, int dst_cpu
)
6456 struct cfs_rq
*cfs_rq
= &cpu_rq(cpu
)->cfs
;
6457 unsigned long util_est
, util
= READ_ONCE(cfs_rq
->avg
.util_avg
);
6460 * If @p migrates from @cpu to another, remove its contribution. Or,
6461 * if @p migrates from another CPU to @cpu, add its contribution. In
6462 * the other cases, @cpu is not impacted by the migration, so the
6463 * util_avg should already be correct.
6465 if (task_cpu(p
) == cpu
&& dst_cpu
!= cpu
)
6466 sub_positive(&util
, task_util(p
));
6467 else if (task_cpu(p
) != cpu
&& dst_cpu
== cpu
)
6468 util
+= task_util(p
);
6470 if (sched_feat(UTIL_EST
)) {
6471 util_est
= READ_ONCE(cfs_rq
->avg
.util_est
.enqueued
);
6474 * During wake-up, the task isn't enqueued yet and doesn't
6475 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6476 * so just add it (if needed) to "simulate" what will be
6477 * cpu_util() after the task has been enqueued.
6480 util_est
+= _task_util_est(p
);
6482 util
= max(util
, util_est
);
6485 return min(util
, capacity_orig_of(cpu
));
6489 * compute_energy(): Estimates the energy that would be consumed if @p was
6490 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6491 * landscape of the * CPUs after the task migration, and uses the Energy Model
6492 * to compute what would be the energy if we decided to actually migrate that
6496 compute_energy(struct task_struct
*p
, int dst_cpu
, struct perf_domain
*pd
)
6498 long util
, max_util
, sum_util
, energy
= 0;
6501 for (; pd
; pd
= pd
->next
) {
6502 max_util
= sum_util
= 0;
6504 * The capacity state of CPUs of the current rd can be driven by
6505 * CPUs of another rd if they belong to the same performance
6506 * domain. So, account for the utilization of these CPUs too
6507 * by masking pd with cpu_online_mask instead of the rd span.
6509 * If an entire performance domain is outside of the current rd,
6510 * it will not appear in its pd list and will not be accounted
6511 * by compute_energy().
6513 for_each_cpu_and(cpu
, perf_domain_span(pd
), cpu_online_mask
) {
6514 util
= cpu_util_next(cpu
, p
, dst_cpu
);
6515 util
= schedutil_energy_util(cpu
, util
);
6516 max_util
= max(util
, max_util
);
6520 energy
+= em_pd_energy(pd
->em_pd
, max_util
, sum_util
);
6527 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6528 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6529 * spare capacity in each performance domain and uses it as a potential
6530 * candidate to execute the task. Then, it uses the Energy Model to figure
6531 * out which of the CPU candidates is the most energy-efficient.
6533 * The rationale for this heuristic is as follows. In a performance domain,
6534 * all the most energy efficient CPU candidates (according to the Energy
6535 * Model) are those for which we'll request a low frequency. When there are
6536 * several CPUs for which the frequency request will be the same, we don't
6537 * have enough data to break the tie between them, because the Energy Model
6538 * only includes active power costs. With this model, if we assume that
6539 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6540 * the maximum spare capacity in a performance domain is guaranteed to be among
6541 * the best candidates of the performance domain.
6543 * In practice, it could be preferable from an energy standpoint to pack
6544 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6545 * but that could also hurt our chances to go cluster idle, and we have no
6546 * ways to tell with the current Energy Model if this is actually a good
6547 * idea or not. So, find_energy_efficient_cpu() basically favors
6548 * cluster-packing, and spreading inside a cluster. That should at least be
6549 * a good thing for latency, and this is consistent with the idea that most
6550 * of the energy savings of EAS come from the asymmetry of the system, and
6551 * not so much from breaking the tie between identical CPUs. That's also the
6552 * reason why EAS is enabled in the topology code only for systems where
6553 * SD_ASYM_CPUCAPACITY is set.
6555 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6556 * they don't have any useful utilization data yet and it's not possible to
6557 * forecast their impact on energy consumption. Consequently, they will be
6558 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6559 * to be energy-inefficient in some use-cases. The alternative would be to
6560 * bias new tasks towards specific types of CPUs first, or to try to infer
6561 * their util_avg from the parent task, but those heuristics could hurt
6562 * other use-cases too. So, until someone finds a better way to solve this,
6563 * let's keep things simple by re-using the existing slow path.
6566 static int find_energy_efficient_cpu(struct task_struct
*p
, int prev_cpu
)
6568 unsigned long prev_energy
= ULONG_MAX
, best_energy
= ULONG_MAX
;
6569 struct root_domain
*rd
= cpu_rq(smp_processor_id())->rd
;
6570 int cpu
, best_energy_cpu
= prev_cpu
;
6571 struct perf_domain
*head
, *pd
;
6572 unsigned long cpu_cap
, util
;
6573 struct sched_domain
*sd
;
6576 pd
= rcu_dereference(rd
->pd
);
6577 if (!pd
|| READ_ONCE(rd
->overutilized
))
6582 * Energy-aware wake-up happens on the lowest sched_domain starting
6583 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6585 sd
= rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity
));
6586 while (sd
&& !cpumask_test_cpu(prev_cpu
, sched_domain_span(sd
)))
6591 sync_entity_load_avg(&p
->se
);
6592 if (!task_util_est(p
))
6595 for (; pd
; pd
= pd
->next
) {
6596 unsigned long cur_energy
, spare_cap
, max_spare_cap
= 0;
6597 int max_spare_cap_cpu
= -1;
6599 for_each_cpu_and(cpu
, perf_domain_span(pd
), sched_domain_span(sd
)) {
6600 if (!cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
6603 /* Skip CPUs that will be overutilized. */
6604 util
= cpu_util_next(cpu
, p
, cpu
);
6605 cpu_cap
= capacity_of(cpu
);
6606 if (cpu_cap
* 1024 < util
* capacity_margin
)
6609 /* Always use prev_cpu as a candidate. */
6610 if (cpu
== prev_cpu
) {
6611 prev_energy
= compute_energy(p
, prev_cpu
, head
);
6612 best_energy
= min(best_energy
, prev_energy
);
6617 * Find the CPU with the maximum spare capacity in
6618 * the performance domain
6620 spare_cap
= cpu_cap
- util
;
6621 if (spare_cap
> max_spare_cap
) {
6622 max_spare_cap
= spare_cap
;
6623 max_spare_cap_cpu
= cpu
;
6627 /* Evaluate the energy impact of using this CPU. */
6628 if (max_spare_cap_cpu
>= 0) {
6629 cur_energy
= compute_energy(p
, max_spare_cap_cpu
, head
);
6630 if (cur_energy
< best_energy
) {
6631 best_energy
= cur_energy
;
6632 best_energy_cpu
= max_spare_cap_cpu
;
6640 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6641 * least 6% of the energy used by prev_cpu.
6643 if (prev_energy
== ULONG_MAX
)
6644 return best_energy_cpu
;
6646 if ((prev_energy
- best_energy
) > (prev_energy
>> 4))
6647 return best_energy_cpu
;
6658 * select_task_rq_fair: Select target runqueue for the waking task in domains
6659 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6660 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6662 * Balances load by selecting the idlest CPU in the idlest group, or under
6663 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6665 * Returns the target CPU number.
6667 * preempt must be disabled.
6670 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
6672 struct sched_domain
*tmp
, *sd
= NULL
;
6673 int cpu
= smp_processor_id();
6674 int new_cpu
= prev_cpu
;
6675 int want_affine
= 0;
6676 int sync
= (wake_flags
& WF_SYNC
) && !(current
->flags
& PF_EXITING
);
6678 if (sd_flag
& SD_BALANCE_WAKE
) {
6681 if (sched_energy_enabled()) {
6682 new_cpu
= find_energy_efficient_cpu(p
, prev_cpu
);
6688 want_affine
= !wake_wide(p
) && !wake_cap(p
, cpu
, prev_cpu
) &&
6689 cpumask_test_cpu(cpu
, &p
->cpus_allowed
);
6693 for_each_domain(cpu
, tmp
) {
6694 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
6698 * If both 'cpu' and 'prev_cpu' are part of this domain,
6699 * cpu is a valid SD_WAKE_AFFINE target.
6701 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
6702 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
6703 if (cpu
!= prev_cpu
)
6704 new_cpu
= wake_affine(tmp
, p
, cpu
, prev_cpu
, sync
);
6706 sd
= NULL
; /* Prefer wake_affine over balance flags */
6710 if (tmp
->flags
& sd_flag
)
6712 else if (!want_affine
)
6718 new_cpu
= find_idlest_cpu(sd
, p
, cpu
, prev_cpu
, sd_flag
);
6719 } else if (sd_flag
& SD_BALANCE_WAKE
) { /* XXX always ? */
6722 new_cpu
= select_idle_sibling(p
, prev_cpu
, new_cpu
);
6725 current
->recent_used_cpu
= cpu
;
6732 static void detach_entity_cfs_rq(struct sched_entity
*se
);
6735 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6736 * cfs_rq_of(p) references at time of call are still valid and identify the
6737 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6739 static void migrate_task_rq_fair(struct task_struct
*p
, int new_cpu
)
6742 * As blocked tasks retain absolute vruntime the migration needs to
6743 * deal with this by subtracting the old and adding the new
6744 * min_vruntime -- the latter is done by enqueue_entity() when placing
6745 * the task on the new runqueue.
6747 if (p
->state
== TASK_WAKING
) {
6748 struct sched_entity
*se
= &p
->se
;
6749 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6752 #ifndef CONFIG_64BIT
6753 u64 min_vruntime_copy
;
6756 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
6758 min_vruntime
= cfs_rq
->min_vruntime
;
6759 } while (min_vruntime
!= min_vruntime_copy
);
6761 min_vruntime
= cfs_rq
->min_vruntime
;
6764 se
->vruntime
-= min_vruntime
;
6767 if (p
->on_rq
== TASK_ON_RQ_MIGRATING
) {
6769 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6770 * rq->lock and can modify state directly.
6772 lockdep_assert_held(&task_rq(p
)->lock
);
6773 detach_entity_cfs_rq(&p
->se
);
6777 * We are supposed to update the task to "current" time, then
6778 * its up to date and ready to go to new CPU/cfs_rq. But we
6779 * have difficulty in getting what current time is, so simply
6780 * throw away the out-of-date time. This will result in the
6781 * wakee task is less decayed, but giving the wakee more load
6784 remove_entity_load_avg(&p
->se
);
6787 /* Tell new CPU we are migrated */
6788 p
->se
.avg
.last_update_time
= 0;
6790 /* We have migrated, no longer consider this task hot */
6791 p
->se
.exec_start
= 0;
6793 update_scan_period(p
, new_cpu
);
6796 static void task_dead_fair(struct task_struct
*p
)
6798 remove_entity_load_avg(&p
->se
);
6800 #endif /* CONFIG_SMP */
6802 static unsigned long wakeup_gran(struct sched_entity
*se
)
6804 unsigned long gran
= sysctl_sched_wakeup_granularity
;
6807 * Since its curr running now, convert the gran from real-time
6808 * to virtual-time in his units.
6810 * By using 'se' instead of 'curr' we penalize light tasks, so
6811 * they get preempted easier. That is, if 'se' < 'curr' then
6812 * the resulting gran will be larger, therefore penalizing the
6813 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6814 * be smaller, again penalizing the lighter task.
6816 * This is especially important for buddies when the leftmost
6817 * task is higher priority than the buddy.
6819 return calc_delta_fair(gran
, se
);
6823 * Should 'se' preempt 'curr'.
6837 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
6839 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
6844 gran
= wakeup_gran(se
);
6851 static void set_last_buddy(struct sched_entity
*se
)
6853 if (entity_is_task(se
) && unlikely(task_has_idle_policy(task_of(se
))))
6856 for_each_sched_entity(se
) {
6857 if (SCHED_WARN_ON(!se
->on_rq
))
6859 cfs_rq_of(se
)->last
= se
;
6863 static void set_next_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
)->next
= se
;
6875 static void set_skip_buddy(struct sched_entity
*se
)
6877 for_each_sched_entity(se
)
6878 cfs_rq_of(se
)->skip
= se
;
6882 * Preempt the current task with a newly woken task if needed:
6884 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
6886 struct task_struct
*curr
= rq
->curr
;
6887 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
6888 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
6889 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
6890 int next_buddy_marked
= 0;
6892 if (unlikely(se
== pse
))
6896 * This is possible from callers such as attach_tasks(), in which we
6897 * unconditionally check_prempt_curr() after an enqueue (which may have
6898 * lead to a throttle). This both saves work and prevents false
6899 * next-buddy nomination below.
6901 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
6904 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
6905 set_next_buddy(pse
);
6906 next_buddy_marked
= 1;
6910 * We can come here with TIF_NEED_RESCHED already set from new task
6913 * Note: this also catches the edge-case of curr being in a throttled
6914 * group (e.g. via set_curr_task), since update_curr() (in the
6915 * enqueue of curr) will have resulted in resched being set. This
6916 * prevents us from potentially nominating it as a false LAST_BUDDY
6919 if (test_tsk_need_resched(curr
))
6922 /* Idle tasks are by definition preempted by non-idle tasks. */
6923 if (unlikely(task_has_idle_policy(curr
)) &&
6924 likely(!task_has_idle_policy(p
)))
6928 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6929 * is driven by the tick):
6931 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
6934 find_matching_se(&se
, &pse
);
6935 update_curr(cfs_rq_of(se
));
6937 if (wakeup_preempt_entity(se
, pse
) == 1) {
6939 * Bias pick_next to pick the sched entity that is
6940 * triggering this preemption.
6942 if (!next_buddy_marked
)
6943 set_next_buddy(pse
);
6952 * Only set the backward buddy when the current task is still
6953 * on the rq. This can happen when a wakeup gets interleaved
6954 * with schedule on the ->pre_schedule() or idle_balance()
6955 * point, either of which can * drop the rq lock.
6957 * Also, during early boot the idle thread is in the fair class,
6958 * for obvious reasons its a bad idea to schedule back to it.
6960 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
6963 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
6967 static struct task_struct
*
6968 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
6970 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
6971 struct sched_entity
*se
;
6972 struct task_struct
*p
;
6976 if (!cfs_rq
->nr_running
)
6979 #ifdef CONFIG_FAIR_GROUP_SCHED
6980 if (prev
->sched_class
!= &fair_sched_class
)
6984 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6985 * likely that a next task is from the same cgroup as the current.
6987 * Therefore attempt to avoid putting and setting the entire cgroup
6988 * hierarchy, only change the part that actually changes.
6992 struct sched_entity
*curr
= cfs_rq
->curr
;
6995 * Since we got here without doing put_prev_entity() we also
6996 * have to consider cfs_rq->curr. If it is still a runnable
6997 * entity, update_curr() will update its vruntime, otherwise
6998 * forget we've ever seen it.
7002 update_curr(cfs_rq
);
7007 * This call to check_cfs_rq_runtime() will do the
7008 * throttle and dequeue its entity in the parent(s).
7009 * Therefore the nr_running test will indeed
7012 if (unlikely(check_cfs_rq_runtime(cfs_rq
))) {
7015 if (!cfs_rq
->nr_running
)
7022 se
= pick_next_entity(cfs_rq
, curr
);
7023 cfs_rq
= group_cfs_rq(se
);
7029 * Since we haven't yet done put_prev_entity and if the selected task
7030 * is a different task than we started out with, try and touch the
7031 * least amount of cfs_rqs.
7034 struct sched_entity
*pse
= &prev
->se
;
7036 while (!(cfs_rq
= is_same_group(se
, pse
))) {
7037 int se_depth
= se
->depth
;
7038 int pse_depth
= pse
->depth
;
7040 if (se_depth
<= pse_depth
) {
7041 put_prev_entity(cfs_rq_of(pse
), pse
);
7042 pse
= parent_entity(pse
);
7044 if (se_depth
>= pse_depth
) {
7045 set_next_entity(cfs_rq_of(se
), se
);
7046 se
= parent_entity(se
);
7050 put_prev_entity(cfs_rq
, pse
);
7051 set_next_entity(cfs_rq
, se
);
7058 put_prev_task(rq
, prev
);
7061 se
= pick_next_entity(cfs_rq
, NULL
);
7062 set_next_entity(cfs_rq
, se
);
7063 cfs_rq
= group_cfs_rq(se
);
7068 done
: __maybe_unused
;
7071 * Move the next running task to the front of
7072 * the list, so our cfs_tasks list becomes MRU
7075 list_move(&p
->se
.group_node
, &rq
->cfs_tasks
);
7078 if (hrtick_enabled(rq
))
7079 hrtick_start_fair(rq
, p
);
7081 update_misfit_status(p
, rq
);
7086 update_misfit_status(NULL
, rq
);
7087 new_tasks
= idle_balance(rq
, rf
);
7090 * Because idle_balance() releases (and re-acquires) rq->lock, it is
7091 * possible for any higher priority task to appear. In that case we
7092 * must re-start the pick_next_entity() loop.
7101 * rq is about to be idle, check if we need to update the
7102 * lost_idle_time of clock_pelt
7104 update_idle_rq_clock_pelt(rq
);
7110 * Account for a descheduled task:
7112 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
7114 struct sched_entity
*se
= &prev
->se
;
7115 struct cfs_rq
*cfs_rq
;
7117 for_each_sched_entity(se
) {
7118 cfs_rq
= cfs_rq_of(se
);
7119 put_prev_entity(cfs_rq
, se
);
7124 * sched_yield() is very simple
7126 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7128 static void yield_task_fair(struct rq
*rq
)
7130 struct task_struct
*curr
= rq
->curr
;
7131 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
7132 struct sched_entity
*se
= &curr
->se
;
7135 * Are we the only task in the tree?
7137 if (unlikely(rq
->nr_running
== 1))
7140 clear_buddies(cfs_rq
, se
);
7142 if (curr
->policy
!= SCHED_BATCH
) {
7143 update_rq_clock(rq
);
7145 * Update run-time statistics of the 'current'.
7147 update_curr(cfs_rq
);
7149 * Tell update_rq_clock() that we've just updated,
7150 * so we don't do microscopic update in schedule()
7151 * and double the fastpath cost.
7153 rq_clock_skip_update(rq
);
7159 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
7161 struct sched_entity
*se
= &p
->se
;
7163 /* throttled hierarchies are not runnable */
7164 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
7167 /* Tell the scheduler that we'd really like pse to run next. */
7170 yield_task_fair(rq
);
7176 /**************************************************
7177 * Fair scheduling class load-balancing methods.
7181 * The purpose of load-balancing is to achieve the same basic fairness the
7182 * per-CPU scheduler provides, namely provide a proportional amount of compute
7183 * time to each task. This is expressed in the following equation:
7185 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7187 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7188 * W_i,0 is defined as:
7190 * W_i,0 = \Sum_j w_i,j (2)
7192 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7193 * is derived from the nice value as per sched_prio_to_weight[].
7195 * The weight average is an exponential decay average of the instantaneous
7198 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7200 * C_i is the compute capacity of CPU i, typically it is the
7201 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7202 * can also include other factors [XXX].
7204 * To achieve this balance we define a measure of imbalance which follows
7205 * directly from (1):
7207 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7209 * We them move tasks around to minimize the imbalance. In the continuous
7210 * function space it is obvious this converges, in the discrete case we get
7211 * a few fun cases generally called infeasible weight scenarios.
7214 * - infeasible weights;
7215 * - local vs global optima in the discrete case. ]
7220 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7221 * for all i,j solution, we create a tree of CPUs that follows the hardware
7222 * topology where each level pairs two lower groups (or better). This results
7223 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7224 * tree to only the first of the previous level and we decrease the frequency
7225 * of load-balance at each level inv. proportional to the number of CPUs in
7231 * \Sum { --- * --- * 2^i } = O(n) (5)
7233 * `- size of each group
7234 * | | `- number of CPUs doing load-balance
7236 * `- sum over all levels
7238 * Coupled with a limit on how many tasks we can migrate every balance pass,
7239 * this makes (5) the runtime complexity of the balancer.
7241 * An important property here is that each CPU is still (indirectly) connected
7242 * to every other CPU in at most O(log n) steps:
7244 * The adjacency matrix of the resulting graph is given by:
7247 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7250 * And you'll find that:
7252 * A^(log_2 n)_i,j != 0 for all i,j (7)
7254 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7255 * The task movement gives a factor of O(m), giving a convergence complexity
7258 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7263 * In order to avoid CPUs going idle while there's still work to do, new idle
7264 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7265 * tree itself instead of relying on other CPUs to bring it work.
7267 * This adds some complexity to both (5) and (8) but it reduces the total idle
7275 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7278 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7283 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7285 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7287 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7290 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7291 * rewrite all of this once again.]
7294 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
7296 enum fbq_type
{ regular
, remote
, all
};
7305 #define LBF_ALL_PINNED 0x01
7306 #define LBF_NEED_BREAK 0x02
7307 #define LBF_DST_PINNED 0x04
7308 #define LBF_SOME_PINNED 0x08
7309 #define LBF_NOHZ_STATS 0x10
7310 #define LBF_NOHZ_AGAIN 0x20
7313 struct sched_domain
*sd
;
7321 struct cpumask
*dst_grpmask
;
7323 enum cpu_idle_type idle
;
7325 /* The set of CPUs under consideration for load-balancing */
7326 struct cpumask
*cpus
;
7331 unsigned int loop_break
;
7332 unsigned int loop_max
;
7334 enum fbq_type fbq_type
;
7335 enum group_type src_grp_type
;
7336 struct list_head tasks
;
7340 * Is this task likely cache-hot:
7342 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
7346 lockdep_assert_held(&env
->src_rq
->lock
);
7348 if (p
->sched_class
!= &fair_sched_class
)
7351 if (unlikely(task_has_idle_policy(p
)))
7355 * Buddy candidates are cache hot:
7357 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
7358 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
7359 &p
->se
== cfs_rq_of(&p
->se
)->last
))
7362 if (sysctl_sched_migration_cost
== -1)
7364 if (sysctl_sched_migration_cost
== 0)
7367 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
7369 return delta
< (s64
)sysctl_sched_migration_cost
;
7372 #ifdef CONFIG_NUMA_BALANCING
7374 * Returns 1, if task migration degrades locality
7375 * Returns 0, if task migration improves locality i.e migration preferred.
7376 * Returns -1, if task migration is not affected by locality.
7378 static int migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
7380 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
7381 unsigned long src_weight
, dst_weight
;
7382 int src_nid
, dst_nid
, dist
;
7384 if (!static_branch_likely(&sched_numa_balancing
))
7387 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
7390 src_nid
= cpu_to_node(env
->src_cpu
);
7391 dst_nid
= cpu_to_node(env
->dst_cpu
);
7393 if (src_nid
== dst_nid
)
7396 /* Migrating away from the preferred node is always bad. */
7397 if (src_nid
== p
->numa_preferred_nid
) {
7398 if (env
->src_rq
->nr_running
> env
->src_rq
->nr_preferred_running
)
7404 /* Encourage migration to the preferred node. */
7405 if (dst_nid
== p
->numa_preferred_nid
)
7408 /* Leaving a core idle is often worse than degrading locality. */
7409 if (env
->idle
== CPU_IDLE
)
7412 dist
= node_distance(src_nid
, dst_nid
);
7414 src_weight
= group_weight(p
, src_nid
, dist
);
7415 dst_weight
= group_weight(p
, dst_nid
, dist
);
7417 src_weight
= task_weight(p
, src_nid
, dist
);
7418 dst_weight
= task_weight(p
, dst_nid
, dist
);
7421 return dst_weight
< src_weight
;
7425 static inline int migrate_degrades_locality(struct task_struct
*p
,
7433 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7436 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
7440 lockdep_assert_held(&env
->src_rq
->lock
);
7443 * We do not migrate tasks that are:
7444 * 1) throttled_lb_pair, or
7445 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7446 * 3) running (obviously), or
7447 * 4) are cache-hot on their current CPU.
7449 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
7452 if (!cpumask_test_cpu(env
->dst_cpu
, &p
->cpus_allowed
)) {
7455 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_affine
);
7457 env
->flags
|= LBF_SOME_PINNED
;
7460 * Remember if this task can be migrated to any other CPU in
7461 * our sched_group. We may want to revisit it if we couldn't
7462 * meet load balance goals by pulling other tasks on src_cpu.
7464 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7465 * already computed one in current iteration.
7467 if (env
->idle
== CPU_NEWLY_IDLE
|| (env
->flags
& LBF_DST_PINNED
))
7470 /* Prevent to re-select dst_cpu via env's CPUs: */
7471 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
7472 if (cpumask_test_cpu(cpu
, &p
->cpus_allowed
)) {
7473 env
->flags
|= LBF_DST_PINNED
;
7474 env
->new_dst_cpu
= cpu
;
7482 /* Record that we found atleast one task that could run on dst_cpu */
7483 env
->flags
&= ~LBF_ALL_PINNED
;
7485 if (task_running(env
->src_rq
, p
)) {
7486 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_running
);
7491 * Aggressive migration if:
7492 * 1) destination numa is preferred
7493 * 2) task is cache cold, or
7494 * 3) too many balance attempts have failed.
7496 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
7497 if (tsk_cache_hot
== -1)
7498 tsk_cache_hot
= task_hot(p
, env
);
7500 if (tsk_cache_hot
<= 0 ||
7501 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
7502 if (tsk_cache_hot
== 1) {
7503 schedstat_inc(env
->sd
->lb_hot_gained
[env
->idle
]);
7504 schedstat_inc(p
->se
.statistics
.nr_forced_migrations
);
7509 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_hot
);
7514 * detach_task() -- detach the task for the migration specified in env
7516 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
7518 lockdep_assert_held(&env
->src_rq
->lock
);
7520 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
7521 deactivate_task(env
->src_rq
, p
, DEQUEUE_NOCLOCK
);
7522 set_task_cpu(p
, env
->dst_cpu
);
7526 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7527 * part of active balancing operations within "domain".
7529 * Returns a task if successful and NULL otherwise.
7531 static struct task_struct
*detach_one_task(struct lb_env
*env
)
7533 struct task_struct
*p
;
7535 lockdep_assert_held(&env
->src_rq
->lock
);
7537 list_for_each_entry_reverse(p
,
7538 &env
->src_rq
->cfs_tasks
, se
.group_node
) {
7539 if (!can_migrate_task(p
, env
))
7542 detach_task(p
, env
);
7545 * Right now, this is only the second place where
7546 * lb_gained[env->idle] is updated (other is detach_tasks)
7547 * so we can safely collect stats here rather than
7548 * inside detach_tasks().
7550 schedstat_inc(env
->sd
->lb_gained
[env
->idle
]);
7556 static const unsigned int sched_nr_migrate_break
= 32;
7559 * detach_tasks() -- tries to detach up to imbalance weighted load from
7560 * busiest_rq, as part of a balancing operation within domain "sd".
7562 * Returns number of detached tasks if successful and 0 otherwise.
7564 static int detach_tasks(struct lb_env
*env
)
7566 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
7567 struct task_struct
*p
;
7571 lockdep_assert_held(&env
->src_rq
->lock
);
7573 if (env
->imbalance
<= 0)
7576 while (!list_empty(tasks
)) {
7578 * We don't want to steal all, otherwise we may be treated likewise,
7579 * which could at worst lead to a livelock crash.
7581 if (env
->idle
!= CPU_NOT_IDLE
&& env
->src_rq
->nr_running
<= 1)
7584 p
= list_last_entry(tasks
, struct task_struct
, se
.group_node
);
7587 /* We've more or less seen every task there is, call it quits */
7588 if (env
->loop
> env
->loop_max
)
7591 /* take a breather every nr_migrate tasks */
7592 if (env
->loop
> env
->loop_break
) {
7593 env
->loop_break
+= sched_nr_migrate_break
;
7594 env
->flags
|= LBF_NEED_BREAK
;
7598 if (!can_migrate_task(p
, env
))
7601 load
= task_h_load(p
);
7603 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
7606 if ((load
/ 2) > env
->imbalance
)
7609 detach_task(p
, env
);
7610 list_add(&p
->se
.group_node
, &env
->tasks
);
7613 env
->imbalance
-= load
;
7615 #ifdef CONFIG_PREEMPT
7617 * NEWIDLE balancing is a source of latency, so preemptible
7618 * kernels will stop after the first task is detached to minimize
7619 * the critical section.
7621 if (env
->idle
== CPU_NEWLY_IDLE
)
7626 * We only want to steal up to the prescribed amount of
7629 if (env
->imbalance
<= 0)
7634 list_move(&p
->se
.group_node
, tasks
);
7638 * Right now, this is one of only two places we collect this stat
7639 * so we can safely collect detach_one_task() stats here rather
7640 * than inside detach_one_task().
7642 schedstat_add(env
->sd
->lb_gained
[env
->idle
], detached
);
7648 * attach_task() -- attach the task detached by detach_task() to its new rq.
7650 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
7652 lockdep_assert_held(&rq
->lock
);
7654 BUG_ON(task_rq(p
) != rq
);
7655 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
7656 p
->on_rq
= TASK_ON_RQ_QUEUED
;
7657 check_preempt_curr(rq
, p
, 0);
7661 * attach_one_task() -- attaches the task returned from detach_one_task() to
7664 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
7669 update_rq_clock(rq
);
7675 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7678 static void attach_tasks(struct lb_env
*env
)
7680 struct list_head
*tasks
= &env
->tasks
;
7681 struct task_struct
*p
;
7684 rq_lock(env
->dst_rq
, &rf
);
7685 update_rq_clock(env
->dst_rq
);
7687 while (!list_empty(tasks
)) {
7688 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
7689 list_del_init(&p
->se
.group_node
);
7691 attach_task(env
->dst_rq
, p
);
7694 rq_unlock(env
->dst_rq
, &rf
);
7697 static inline bool cfs_rq_has_blocked(struct cfs_rq
*cfs_rq
)
7699 if (cfs_rq
->avg
.load_avg
)
7702 if (cfs_rq
->avg
.util_avg
)
7708 static inline bool others_have_blocked(struct rq
*rq
)
7710 if (READ_ONCE(rq
->avg_rt
.util_avg
))
7713 if (READ_ONCE(rq
->avg_dl
.util_avg
))
7716 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7717 if (READ_ONCE(rq
->avg_irq
.util_avg
))
7724 #ifdef CONFIG_FAIR_GROUP_SCHED
7726 static inline bool cfs_rq_is_decayed(struct cfs_rq
*cfs_rq
)
7728 if (cfs_rq
->load
.weight
)
7731 if (cfs_rq
->avg
.load_sum
)
7734 if (cfs_rq
->avg
.util_sum
)
7737 if (cfs_rq
->avg
.runnable_load_sum
)
7743 static void update_blocked_averages(int cpu
)
7745 struct rq
*rq
= cpu_rq(cpu
);
7746 struct cfs_rq
*cfs_rq
, *pos
;
7747 const struct sched_class
*curr_class
;
7751 rq_lock_irqsave(rq
, &rf
);
7752 update_rq_clock(rq
);
7755 * Iterates the task_group tree in a bottom up fashion, see
7756 * list_add_leaf_cfs_rq() for details.
7758 for_each_leaf_cfs_rq_safe(rq
, cfs_rq
, pos
) {
7759 struct sched_entity
*se
;
7761 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq
), cfs_rq
))
7762 update_tg_load_avg(cfs_rq
, 0);
7764 /* Propagate pending load changes to the parent, if any: */
7765 se
= cfs_rq
->tg
->se
[cpu
];
7766 if (se
&& !skip_blocked_update(se
))
7767 update_load_avg(cfs_rq_of(se
), se
, 0);
7770 * There can be a lot of idle CPU cgroups. Don't let fully
7771 * decayed cfs_rqs linger on the list.
7773 if (cfs_rq_is_decayed(cfs_rq
))
7774 list_del_leaf_cfs_rq(cfs_rq
);
7776 /* Don't need periodic decay once load/util_avg are null */
7777 if (cfs_rq_has_blocked(cfs_rq
))
7781 curr_class
= rq
->curr
->sched_class
;
7782 update_rt_rq_load_avg(rq_clock_pelt(rq
), rq
, curr_class
== &rt_sched_class
);
7783 update_dl_rq_load_avg(rq_clock_pelt(rq
), rq
, curr_class
== &dl_sched_class
);
7784 update_irq_load_avg(rq
, 0);
7785 /* Don't need periodic decay once load/util_avg are null */
7786 if (others_have_blocked(rq
))
7789 #ifdef CONFIG_NO_HZ_COMMON
7790 rq
->last_blocked_load_update_tick
= jiffies
;
7792 rq
->has_blocked_load
= 0;
7794 rq_unlock_irqrestore(rq
, &rf
);
7798 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7799 * This needs to be done in a top-down fashion because the load of a child
7800 * group is a fraction of its parents load.
7802 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
7804 struct rq
*rq
= rq_of(cfs_rq
);
7805 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
7806 unsigned long now
= jiffies
;
7809 if (cfs_rq
->last_h_load_update
== now
)
7812 WRITE_ONCE(cfs_rq
->h_load_next
, NULL
);
7813 for_each_sched_entity(se
) {
7814 cfs_rq
= cfs_rq_of(se
);
7815 WRITE_ONCE(cfs_rq
->h_load_next
, se
);
7816 if (cfs_rq
->last_h_load_update
== now
)
7821 cfs_rq
->h_load
= cfs_rq_load_avg(cfs_rq
);
7822 cfs_rq
->last_h_load_update
= now
;
7825 while ((se
= READ_ONCE(cfs_rq
->h_load_next
)) != NULL
) {
7826 load
= cfs_rq
->h_load
;
7827 load
= div64_ul(load
* se
->avg
.load_avg
,
7828 cfs_rq_load_avg(cfs_rq
) + 1);
7829 cfs_rq
= group_cfs_rq(se
);
7830 cfs_rq
->h_load
= load
;
7831 cfs_rq
->last_h_load_update
= now
;
7835 static unsigned long task_h_load(struct task_struct
*p
)
7837 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
7839 update_cfs_rq_h_load(cfs_rq
);
7840 return div64_ul(p
->se
.avg
.load_avg
* cfs_rq
->h_load
,
7841 cfs_rq_load_avg(cfs_rq
) + 1);
7844 static inline void update_blocked_averages(int cpu
)
7846 struct rq
*rq
= cpu_rq(cpu
);
7847 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
7848 const struct sched_class
*curr_class
;
7851 rq_lock_irqsave(rq
, &rf
);
7852 update_rq_clock(rq
);
7853 update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq
), cfs_rq
);
7855 curr_class
= rq
->curr
->sched_class
;
7856 update_rt_rq_load_avg(rq_clock_pelt(rq
), rq
, curr_class
== &rt_sched_class
);
7857 update_dl_rq_load_avg(rq_clock_pelt(rq
), rq
, curr_class
== &dl_sched_class
);
7858 update_irq_load_avg(rq
, 0);
7859 #ifdef CONFIG_NO_HZ_COMMON
7860 rq
->last_blocked_load_update_tick
= jiffies
;
7861 if (!cfs_rq_has_blocked(cfs_rq
) && !others_have_blocked(rq
))
7862 rq
->has_blocked_load
= 0;
7864 rq_unlock_irqrestore(rq
, &rf
);
7867 static unsigned long task_h_load(struct task_struct
*p
)
7869 return p
->se
.avg
.load_avg
;
7873 /********** Helpers for find_busiest_group ************************/
7876 * sg_lb_stats - stats of a sched_group required for load_balancing
7878 struct sg_lb_stats
{
7879 unsigned long avg_load
; /*Avg load across the CPUs of the group */
7880 unsigned long group_load
; /* Total load over the CPUs of the group */
7881 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
7882 unsigned long load_per_task
;
7883 unsigned long group_capacity
;
7884 unsigned long group_util
; /* Total utilization of the group */
7885 unsigned int sum_nr_running
; /* Nr tasks running in the group */
7886 unsigned int idle_cpus
;
7887 unsigned int group_weight
;
7888 enum group_type group_type
;
7889 int group_no_capacity
;
7890 unsigned long group_misfit_task_load
; /* A CPU has a task too big for its capacity */
7891 #ifdef CONFIG_NUMA_BALANCING
7892 unsigned int nr_numa_running
;
7893 unsigned int nr_preferred_running
;
7898 * sd_lb_stats - Structure to store the statistics of a sched_domain
7899 * during load balancing.
7901 struct sd_lb_stats
{
7902 struct sched_group
*busiest
; /* Busiest group in this sd */
7903 struct sched_group
*local
; /* Local group in this sd */
7904 unsigned long total_running
;
7905 unsigned long total_load
; /* Total load of all groups in sd */
7906 unsigned long total_capacity
; /* Total capacity of all groups in sd */
7907 unsigned long avg_load
; /* Average load across all groups in sd */
7909 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
7910 struct sg_lb_stats local_stat
; /* Statistics of the local group */
7913 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
7916 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7917 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7918 * We must however clear busiest_stat::avg_load because
7919 * update_sd_pick_busiest() reads this before assignment.
7921 *sds
= (struct sd_lb_stats
){
7924 .total_running
= 0UL,
7926 .total_capacity
= 0UL,
7929 .sum_nr_running
= 0,
7930 .group_type
= group_other
,
7936 * get_sd_load_idx - Obtain the load index for a given sched domain.
7937 * @sd: The sched_domain whose load_idx is to be obtained.
7938 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7940 * Return: The load index.
7942 static inline int get_sd_load_idx(struct sched_domain
*sd
,
7943 enum cpu_idle_type idle
)
7949 load_idx
= sd
->busy_idx
;
7952 case CPU_NEWLY_IDLE
:
7953 load_idx
= sd
->newidle_idx
;
7956 load_idx
= sd
->idle_idx
;
7963 static unsigned long scale_rt_capacity(struct sched_domain
*sd
, int cpu
)
7965 struct rq
*rq
= cpu_rq(cpu
);
7966 unsigned long max
= arch_scale_cpu_capacity(sd
, cpu
);
7967 unsigned long used
, free
;
7970 irq
= cpu_util_irq(rq
);
7972 if (unlikely(irq
>= max
))
7975 used
= READ_ONCE(rq
->avg_rt
.util_avg
);
7976 used
+= READ_ONCE(rq
->avg_dl
.util_avg
);
7978 if (unlikely(used
>= max
))
7983 return scale_irq_capacity(free
, irq
, max
);
7986 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
7988 unsigned long capacity
= scale_rt_capacity(sd
, cpu
);
7989 struct sched_group
*sdg
= sd
->groups
;
7991 cpu_rq(cpu
)->cpu_capacity_orig
= arch_scale_cpu_capacity(sd
, cpu
);
7996 cpu_rq(cpu
)->cpu_capacity
= capacity
;
7997 sdg
->sgc
->capacity
= capacity
;
7998 sdg
->sgc
->min_capacity
= capacity
;
7999 sdg
->sgc
->max_capacity
= capacity
;
8002 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
8004 struct sched_domain
*child
= sd
->child
;
8005 struct sched_group
*group
, *sdg
= sd
->groups
;
8006 unsigned long capacity
, min_capacity
, max_capacity
;
8007 unsigned long interval
;
8009 interval
= msecs_to_jiffies(sd
->balance_interval
);
8010 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
8011 sdg
->sgc
->next_update
= jiffies
+ interval
;
8014 update_cpu_capacity(sd
, cpu
);
8019 min_capacity
= ULONG_MAX
;
8022 if (child
->flags
& SD_OVERLAP
) {
8024 * SD_OVERLAP domains cannot assume that child groups
8025 * span the current group.
8028 for_each_cpu(cpu
, sched_group_span(sdg
)) {
8029 struct sched_group_capacity
*sgc
;
8030 struct rq
*rq
= cpu_rq(cpu
);
8033 * build_sched_domains() -> init_sched_groups_capacity()
8034 * gets here before we've attached the domains to the
8037 * Use capacity_of(), which is set irrespective of domains
8038 * in update_cpu_capacity().
8040 * This avoids capacity from being 0 and
8041 * causing divide-by-zero issues on boot.
8043 if (unlikely(!rq
->sd
)) {
8044 capacity
+= capacity_of(cpu
);
8046 sgc
= rq
->sd
->groups
->sgc
;
8047 capacity
+= sgc
->capacity
;
8050 min_capacity
= min(capacity
, min_capacity
);
8051 max_capacity
= max(capacity
, max_capacity
);
8055 * !SD_OVERLAP domains can assume that child groups
8056 * span the current group.
8059 group
= child
->groups
;
8061 struct sched_group_capacity
*sgc
= group
->sgc
;
8063 capacity
+= sgc
->capacity
;
8064 min_capacity
= min(sgc
->min_capacity
, min_capacity
);
8065 max_capacity
= max(sgc
->max_capacity
, max_capacity
);
8066 group
= group
->next
;
8067 } while (group
!= child
->groups
);
8070 sdg
->sgc
->capacity
= capacity
;
8071 sdg
->sgc
->min_capacity
= min_capacity
;
8072 sdg
->sgc
->max_capacity
= max_capacity
;
8076 * Check whether the capacity of the rq has been noticeably reduced by side
8077 * activity. The imbalance_pct is used for the threshold.
8078 * Return true is the capacity is reduced
8081 check_cpu_capacity(struct rq
*rq
, struct sched_domain
*sd
)
8083 return ((rq
->cpu_capacity
* sd
->imbalance_pct
) <
8084 (rq
->cpu_capacity_orig
* 100));
8088 * Check whether a rq has a misfit task and if it looks like we can actually
8089 * help that task: we can migrate the task to a CPU of higher capacity, or
8090 * the task's current CPU is heavily pressured.
8092 static inline int check_misfit_status(struct rq
*rq
, struct sched_domain
*sd
)
8094 return rq
->misfit_task_load
&&
8095 (rq
->cpu_capacity_orig
< rq
->rd
->max_cpu_capacity
||
8096 check_cpu_capacity(rq
, sd
));
8100 * Group imbalance indicates (and tries to solve) the problem where balancing
8101 * groups is inadequate due to ->cpus_allowed constraints.
8103 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8104 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8107 * { 0 1 2 3 } { 4 5 6 7 }
8110 * If we were to balance group-wise we'd place two tasks in the first group and
8111 * two tasks in the second group. Clearly this is undesired as it will overload
8112 * cpu 3 and leave one of the CPUs in the second group unused.
8114 * The current solution to this issue is detecting the skew in the first group
8115 * by noticing the lower domain failed to reach balance and had difficulty
8116 * moving tasks due to affinity constraints.
8118 * When this is so detected; this group becomes a candidate for busiest; see
8119 * update_sd_pick_busiest(). And calculate_imbalance() and
8120 * find_busiest_group() avoid some of the usual balance conditions to allow it
8121 * to create an effective group imbalance.
8123 * This is a somewhat tricky proposition since the next run might not find the
8124 * group imbalance and decide the groups need to be balanced again. A most
8125 * subtle and fragile situation.
8128 static inline int sg_imbalanced(struct sched_group
*group
)
8130 return group
->sgc
->imbalance
;
8134 * group_has_capacity returns true if the group has spare capacity that could
8135 * be used by some tasks.
8136 * We consider that a group has spare capacity if the * number of task is
8137 * smaller than the number of CPUs or if the utilization is lower than the
8138 * available capacity for CFS tasks.
8139 * For the latter, we use a threshold to stabilize the state, to take into
8140 * account the variance of the tasks' load and to return true if the available
8141 * capacity in meaningful for the load balancer.
8142 * As an example, an available capacity of 1% can appear but it doesn't make
8143 * any benefit for the load balance.
8146 group_has_capacity(struct lb_env
*env
, struct sg_lb_stats
*sgs
)
8148 if (sgs
->sum_nr_running
< sgs
->group_weight
)
8151 if ((sgs
->group_capacity
* 100) >
8152 (sgs
->group_util
* env
->sd
->imbalance_pct
))
8159 * group_is_overloaded returns true if the group has more tasks than it can
8161 * group_is_overloaded is not equals to !group_has_capacity because a group
8162 * with the exact right number of tasks, has no more spare capacity but is not
8163 * overloaded so both group_has_capacity and group_is_overloaded return
8167 group_is_overloaded(struct lb_env
*env
, struct sg_lb_stats
*sgs
)
8169 if (sgs
->sum_nr_running
<= sgs
->group_weight
)
8172 if ((sgs
->group_capacity
* 100) <
8173 (sgs
->group_util
* env
->sd
->imbalance_pct
))
8180 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
8181 * per-CPU capacity than sched_group ref.
8184 group_smaller_min_cpu_capacity(struct sched_group
*sg
, struct sched_group
*ref
)
8186 return sg
->sgc
->min_capacity
* capacity_margin
<
8187 ref
->sgc
->min_capacity
* 1024;
8191 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
8192 * per-CPU capacity_orig than sched_group ref.
8195 group_smaller_max_cpu_capacity(struct sched_group
*sg
, struct sched_group
*ref
)
8197 return sg
->sgc
->max_capacity
* capacity_margin
<
8198 ref
->sgc
->max_capacity
* 1024;
8202 group_type
group_classify(struct sched_group
*group
,
8203 struct sg_lb_stats
*sgs
)
8205 if (sgs
->group_no_capacity
)
8206 return group_overloaded
;
8208 if (sg_imbalanced(group
))
8209 return group_imbalanced
;
8211 if (sgs
->group_misfit_task_load
)
8212 return group_misfit_task
;
8217 static bool update_nohz_stats(struct rq
*rq
, bool force
)
8219 #ifdef CONFIG_NO_HZ_COMMON
8220 unsigned int cpu
= rq
->cpu
;
8222 if (!rq
->has_blocked_load
)
8225 if (!cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))
8228 if (!force
&& !time_after(jiffies
, rq
->last_blocked_load_update_tick
))
8231 update_blocked_averages(cpu
);
8233 return rq
->has_blocked_load
;
8240 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8241 * @env: The load balancing environment.
8242 * @group: sched_group whose statistics are to be updated.
8243 * @sgs: variable to hold the statistics for this group.
8244 * @sg_status: Holds flag indicating the status of the sched_group
8246 static inline void update_sg_lb_stats(struct lb_env
*env
,
8247 struct sched_group
*group
,
8248 struct sg_lb_stats
*sgs
,
8251 int local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_span(group
));
8252 int load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
8256 memset(sgs
, 0, sizeof(*sgs
));
8258 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
8259 struct rq
*rq
= cpu_rq(i
);
8261 if ((env
->flags
& LBF_NOHZ_STATS
) && update_nohz_stats(rq
, false))
8262 env
->flags
|= LBF_NOHZ_AGAIN
;
8264 /* Bias balancing toward CPUs of our domain: */
8266 load
= target_load(i
, load_idx
);
8268 load
= source_load(i
, load_idx
);
8270 sgs
->group_load
+= load
;
8271 sgs
->group_util
+= cpu_util(i
);
8272 sgs
->sum_nr_running
+= rq
->cfs
.h_nr_running
;
8274 nr_running
= rq
->nr_running
;
8276 *sg_status
|= SG_OVERLOAD
;
8278 if (cpu_overutilized(i
))
8279 *sg_status
|= SG_OVERUTILIZED
;
8281 #ifdef CONFIG_NUMA_BALANCING
8282 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
8283 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
8285 sgs
->sum_weighted_load
+= weighted_cpuload(rq
);
8287 * No need to call idle_cpu() if nr_running is not 0
8289 if (!nr_running
&& idle_cpu(i
))
8292 if (env
->sd
->flags
& SD_ASYM_CPUCAPACITY
&&
8293 sgs
->group_misfit_task_load
< rq
->misfit_task_load
) {
8294 sgs
->group_misfit_task_load
= rq
->misfit_task_load
;
8295 *sg_status
|= SG_OVERLOAD
;
8299 /* Adjust by relative CPU capacity of the group */
8300 sgs
->group_capacity
= group
->sgc
->capacity
;
8301 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
8303 if (sgs
->sum_nr_running
)
8304 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
8306 sgs
->group_weight
= group
->group_weight
;
8308 sgs
->group_no_capacity
= group_is_overloaded(env
, sgs
);
8309 sgs
->group_type
= group_classify(group
, sgs
);
8313 * update_sd_pick_busiest - return 1 on busiest group
8314 * @env: The load balancing environment.
8315 * @sds: sched_domain statistics
8316 * @sg: sched_group candidate to be checked for being the busiest
8317 * @sgs: sched_group statistics
8319 * Determine if @sg is a busier group than the previously selected
8322 * Return: %true if @sg is a busier group than the previously selected
8323 * busiest group. %false otherwise.
8325 static bool update_sd_pick_busiest(struct lb_env
*env
,
8326 struct sd_lb_stats
*sds
,
8327 struct sched_group
*sg
,
8328 struct sg_lb_stats
*sgs
)
8330 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
8333 * Don't try to pull misfit tasks we can't help.
8334 * We can use max_capacity here as reduction in capacity on some
8335 * CPUs in the group should either be possible to resolve
8336 * internally or be covered by avg_load imbalance (eventually).
8338 if (sgs
->group_type
== group_misfit_task
&&
8339 (!group_smaller_max_cpu_capacity(sg
, sds
->local
) ||
8340 !group_has_capacity(env
, &sds
->local_stat
)))
8343 if (sgs
->group_type
> busiest
->group_type
)
8346 if (sgs
->group_type
< busiest
->group_type
)
8349 if (sgs
->avg_load
<= busiest
->avg_load
)
8352 if (!(env
->sd
->flags
& SD_ASYM_CPUCAPACITY
))
8356 * Candidate sg has no more than one task per CPU and
8357 * has higher per-CPU capacity. Migrating tasks to less
8358 * capable CPUs may harm throughput. Maximize throughput,
8359 * power/energy consequences are not considered.
8361 if (sgs
->sum_nr_running
<= sgs
->group_weight
&&
8362 group_smaller_min_cpu_capacity(sds
->local
, sg
))
8366 * If we have more than one misfit sg go with the biggest misfit.
8368 if (sgs
->group_type
== group_misfit_task
&&
8369 sgs
->group_misfit_task_load
< busiest
->group_misfit_task_load
)
8373 /* This is the busiest node in its class. */
8374 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
8377 /* No ASYM_PACKING if target CPU is already busy */
8378 if (env
->idle
== CPU_NOT_IDLE
)
8381 * ASYM_PACKING needs to move all the work to the highest
8382 * prority CPUs in the group, therefore mark all groups
8383 * of lower priority than ourself as busy.
8385 if (sgs
->sum_nr_running
&&
8386 sched_asym_prefer(env
->dst_cpu
, sg
->asym_prefer_cpu
)) {
8390 /* Prefer to move from lowest priority CPU's work */
8391 if (sched_asym_prefer(sds
->busiest
->asym_prefer_cpu
,
8392 sg
->asym_prefer_cpu
))
8399 #ifdef CONFIG_NUMA_BALANCING
8400 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
8402 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
8404 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
8409 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
8411 if (rq
->nr_running
> rq
->nr_numa_running
)
8413 if (rq
->nr_running
> rq
->nr_preferred_running
)
8418 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
8423 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
8427 #endif /* CONFIG_NUMA_BALANCING */
8430 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8431 * @env: The load balancing environment.
8432 * @sds: variable to hold the statistics for this sched_domain.
8434 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
8436 struct sched_domain
*child
= env
->sd
->child
;
8437 struct sched_group
*sg
= env
->sd
->groups
;
8438 struct sg_lb_stats
*local
= &sds
->local_stat
;
8439 struct sg_lb_stats tmp_sgs
;
8440 bool prefer_sibling
= child
&& child
->flags
& SD_PREFER_SIBLING
;
8443 #ifdef CONFIG_NO_HZ_COMMON
8444 if (env
->idle
== CPU_NEWLY_IDLE
&& READ_ONCE(nohz
.has_blocked
))
8445 env
->flags
|= LBF_NOHZ_STATS
;
8449 struct sg_lb_stats
*sgs
= &tmp_sgs
;
8452 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_span(sg
));
8457 if (env
->idle
!= CPU_NEWLY_IDLE
||
8458 time_after_eq(jiffies
, sg
->sgc
->next_update
))
8459 update_group_capacity(env
->sd
, env
->dst_cpu
);
8462 update_sg_lb_stats(env
, sg
, sgs
, &sg_status
);
8468 * In case the child domain prefers tasks go to siblings
8469 * first, lower the sg capacity so that we'll try
8470 * and move all the excess tasks away. We lower the capacity
8471 * of a group only if the local group has the capacity to fit
8472 * these excess tasks. The extra check prevents the case where
8473 * you always pull from the heaviest group when it is already
8474 * under-utilized (possible with a large weight task outweighs
8475 * the tasks on the system).
8477 if (prefer_sibling
&& sds
->local
&&
8478 group_has_capacity(env
, local
) &&
8479 (sgs
->sum_nr_running
> local
->sum_nr_running
+ 1)) {
8480 sgs
->group_no_capacity
= 1;
8481 sgs
->group_type
= group_classify(sg
, sgs
);
8484 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
8486 sds
->busiest_stat
= *sgs
;
8490 /* Now, start updating sd_lb_stats */
8491 sds
->total_running
+= sgs
->sum_nr_running
;
8492 sds
->total_load
+= sgs
->group_load
;
8493 sds
->total_capacity
+= sgs
->group_capacity
;
8496 } while (sg
!= env
->sd
->groups
);
8498 #ifdef CONFIG_NO_HZ_COMMON
8499 if ((env
->flags
& LBF_NOHZ_AGAIN
) &&
8500 cpumask_subset(nohz
.idle_cpus_mask
, sched_domain_span(env
->sd
))) {
8502 WRITE_ONCE(nohz
.next_blocked
,
8503 jiffies
+ msecs_to_jiffies(LOAD_AVG_PERIOD
));
8507 if (env
->sd
->flags
& SD_NUMA
)
8508 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
8510 if (!env
->sd
->parent
) {
8511 struct root_domain
*rd
= env
->dst_rq
->rd
;
8513 /* update overload indicator if we are at root domain */
8514 WRITE_ONCE(rd
->overload
, sg_status
& SG_OVERLOAD
);
8516 /* Update over-utilization (tipping point, U >= 0) indicator */
8517 WRITE_ONCE(rd
->overutilized
, sg_status
& SG_OVERUTILIZED
);
8518 } else if (sg_status
& SG_OVERUTILIZED
) {
8519 WRITE_ONCE(env
->dst_rq
->rd
->overutilized
, SG_OVERUTILIZED
);
8524 * check_asym_packing - Check to see if the group is packed into the
8527 * This is primarily intended to used at the sibling level. Some
8528 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8529 * case of POWER7, it can move to lower SMT modes only when higher
8530 * threads are idle. When in lower SMT modes, the threads will
8531 * perform better since they share less core resources. Hence when we
8532 * have idle threads, we want them to be the higher ones.
8534 * This packing function is run on idle threads. It checks to see if
8535 * the busiest CPU in this domain (core in the P7 case) has a higher
8536 * CPU number than the packing function is being run on. Here we are
8537 * assuming lower CPU number will be equivalent to lower a SMT thread
8540 * Return: 1 when packing is required and a task should be moved to
8541 * this CPU. The amount of the imbalance is returned in env->imbalance.
8543 * @env: The load balancing environment.
8544 * @sds: Statistics of the sched_domain which is to be packed
8546 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
8550 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
8553 if (env
->idle
== CPU_NOT_IDLE
)
8559 busiest_cpu
= sds
->busiest
->asym_prefer_cpu
;
8560 if (sched_asym_prefer(busiest_cpu
, env
->dst_cpu
))
8563 env
->imbalance
= sds
->busiest_stat
.group_load
;
8569 * fix_small_imbalance - Calculate the minor imbalance that exists
8570 * amongst the groups of a sched_domain, during
8572 * @env: The load balancing environment.
8573 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8576 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
8578 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
8579 unsigned int imbn
= 2;
8580 unsigned long scaled_busy_load_per_task
;
8581 struct sg_lb_stats
*local
, *busiest
;
8583 local
= &sds
->local_stat
;
8584 busiest
= &sds
->busiest_stat
;
8586 if (!local
->sum_nr_running
)
8587 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
8588 else if (busiest
->load_per_task
> local
->load_per_task
)
8591 scaled_busy_load_per_task
=
8592 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
8593 busiest
->group_capacity
;
8595 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
8596 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
8597 env
->imbalance
= busiest
->load_per_task
;
8602 * OK, we don't have enough imbalance to justify moving tasks,
8603 * however we may be able to increase total CPU capacity used by
8607 capa_now
+= busiest
->group_capacity
*
8608 min(busiest
->load_per_task
, busiest
->avg_load
);
8609 capa_now
+= local
->group_capacity
*
8610 min(local
->load_per_task
, local
->avg_load
);
8611 capa_now
/= SCHED_CAPACITY_SCALE
;
8613 /* Amount of load we'd subtract */
8614 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
8615 capa_move
+= busiest
->group_capacity
*
8616 min(busiest
->load_per_task
,
8617 busiest
->avg_load
- scaled_busy_load_per_task
);
8620 /* Amount of load we'd add */
8621 if (busiest
->avg_load
* busiest
->group_capacity
<
8622 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
8623 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
8624 local
->group_capacity
;
8626 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
8627 local
->group_capacity
;
8629 capa_move
+= local
->group_capacity
*
8630 min(local
->load_per_task
, local
->avg_load
+ tmp
);
8631 capa_move
/= SCHED_CAPACITY_SCALE
;
8633 /* Move if we gain throughput */
8634 if (capa_move
> capa_now
)
8635 env
->imbalance
= busiest
->load_per_task
;
8639 * calculate_imbalance - Calculate the amount of imbalance present within the
8640 * groups of a given sched_domain during load balance.
8641 * @env: load balance environment
8642 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8644 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
8646 unsigned long max_pull
, load_above_capacity
= ~0UL;
8647 struct sg_lb_stats
*local
, *busiest
;
8649 local
= &sds
->local_stat
;
8650 busiest
= &sds
->busiest_stat
;
8652 if (busiest
->group_type
== group_imbalanced
) {
8654 * In the group_imb case we cannot rely on group-wide averages
8655 * to ensure CPU-load equilibrium, look at wider averages. XXX
8657 busiest
->load_per_task
=
8658 min(busiest
->load_per_task
, sds
->avg_load
);
8662 * Avg load of busiest sg can be less and avg load of local sg can
8663 * be greater than avg load across all sgs of sd because avg load
8664 * factors in sg capacity and sgs with smaller group_type are
8665 * skipped when updating the busiest sg:
8667 if (busiest
->group_type
!= group_misfit_task
&&
8668 (busiest
->avg_load
<= sds
->avg_load
||
8669 local
->avg_load
>= sds
->avg_load
)) {
8671 return fix_small_imbalance(env
, sds
);
8675 * If there aren't any idle CPUs, avoid creating some.
8677 if (busiest
->group_type
== group_overloaded
&&
8678 local
->group_type
== group_overloaded
) {
8679 load_above_capacity
= busiest
->sum_nr_running
* SCHED_CAPACITY_SCALE
;
8680 if (load_above_capacity
> busiest
->group_capacity
) {
8681 load_above_capacity
-= busiest
->group_capacity
;
8682 load_above_capacity
*= scale_load_down(NICE_0_LOAD
);
8683 load_above_capacity
/= busiest
->group_capacity
;
8685 load_above_capacity
= ~0UL;
8689 * We're trying to get all the CPUs to the average_load, so we don't
8690 * want to push ourselves above the average load, nor do we wish to
8691 * reduce the max loaded CPU below the average load. At the same time,
8692 * we also don't want to reduce the group load below the group
8693 * capacity. Thus we look for the minimum possible imbalance.
8695 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
8697 /* How much load to actually move to equalise the imbalance */
8698 env
->imbalance
= min(
8699 max_pull
* busiest
->group_capacity
,
8700 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
8701 ) / SCHED_CAPACITY_SCALE
;
8703 /* Boost imbalance to allow misfit task to be balanced. */
8704 if (busiest
->group_type
== group_misfit_task
) {
8705 env
->imbalance
= max_t(long, env
->imbalance
,
8706 busiest
->group_misfit_task_load
);
8710 * if *imbalance is less than the average load per runnable task
8711 * there is no guarantee that any tasks will be moved so we'll have
8712 * a think about bumping its value to force at least one task to be
8715 if (env
->imbalance
< busiest
->load_per_task
)
8716 return fix_small_imbalance(env
, sds
);
8719 /******* find_busiest_group() helpers end here *********************/
8722 * find_busiest_group - Returns the busiest group within the sched_domain
8723 * if there is an imbalance.
8725 * Also calculates the amount of weighted load which should be moved
8726 * to restore balance.
8728 * @env: The load balancing environment.
8730 * Return: - The busiest group if imbalance exists.
8732 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
8734 struct sg_lb_stats
*local
, *busiest
;
8735 struct sd_lb_stats sds
;
8737 init_sd_lb_stats(&sds
);
8740 * Compute the various statistics relavent for load balancing at
8743 update_sd_lb_stats(env
, &sds
);
8745 if (sched_energy_enabled()) {
8746 struct root_domain
*rd
= env
->dst_rq
->rd
;
8748 if (rcu_dereference(rd
->pd
) && !READ_ONCE(rd
->overutilized
))
8752 local
= &sds
.local_stat
;
8753 busiest
= &sds
.busiest_stat
;
8755 /* ASYM feature bypasses nice load balance check */
8756 if (check_asym_packing(env
, &sds
))
8759 /* There is no busy sibling group to pull tasks from */
8760 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
8763 /* XXX broken for overlapping NUMA groups */
8764 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
8765 / sds
.total_capacity
;
8768 * If the busiest group is imbalanced the below checks don't
8769 * work because they assume all things are equal, which typically
8770 * isn't true due to cpus_allowed constraints and the like.
8772 if (busiest
->group_type
== group_imbalanced
)
8776 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8777 * capacities from resulting in underutilization due to avg_load.
8779 if (env
->idle
!= CPU_NOT_IDLE
&& group_has_capacity(env
, local
) &&
8780 busiest
->group_no_capacity
)
8783 /* Misfit tasks should be dealt with regardless of the avg load */
8784 if (busiest
->group_type
== group_misfit_task
)
8788 * If the local group is busier than the selected busiest group
8789 * don't try and pull any tasks.
8791 if (local
->avg_load
>= busiest
->avg_load
)
8795 * Don't pull any tasks if this group is already above the domain
8798 if (local
->avg_load
>= sds
.avg_load
)
8801 if (env
->idle
== CPU_IDLE
) {
8803 * This CPU is idle. If the busiest group is not overloaded
8804 * and there is no imbalance between this and busiest group
8805 * wrt idle CPUs, it is balanced. The imbalance becomes
8806 * significant if the diff is greater than 1 otherwise we
8807 * might end up to just move the imbalance on another group
8809 if ((busiest
->group_type
!= group_overloaded
) &&
8810 (local
->idle_cpus
<= (busiest
->idle_cpus
+ 1)))
8814 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8815 * imbalance_pct to be conservative.
8817 if (100 * busiest
->avg_load
<=
8818 env
->sd
->imbalance_pct
* local
->avg_load
)
8823 /* Looks like there is an imbalance. Compute it */
8824 env
->src_grp_type
= busiest
->group_type
;
8825 calculate_imbalance(env
, &sds
);
8826 return env
->imbalance
? sds
.busiest
: NULL
;
8834 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8836 static struct rq
*find_busiest_queue(struct lb_env
*env
,
8837 struct sched_group
*group
)
8839 struct rq
*busiest
= NULL
, *rq
;
8840 unsigned long busiest_load
= 0, busiest_capacity
= 1;
8843 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
8844 unsigned long capacity
, wl
;
8848 rt
= fbq_classify_rq(rq
);
8851 * We classify groups/runqueues into three groups:
8852 * - regular: there are !numa tasks
8853 * - remote: there are numa tasks that run on the 'wrong' node
8854 * - all: there is no distinction
8856 * In order to avoid migrating ideally placed numa tasks,
8857 * ignore those when there's better options.
8859 * If we ignore the actual busiest queue to migrate another
8860 * task, the next balance pass can still reduce the busiest
8861 * queue by moving tasks around inside the node.
8863 * If we cannot move enough load due to this classification
8864 * the next pass will adjust the group classification and
8865 * allow migration of more tasks.
8867 * Both cases only affect the total convergence complexity.
8869 if (rt
> env
->fbq_type
)
8873 * For ASYM_CPUCAPACITY domains with misfit tasks we simply
8874 * seek the "biggest" misfit task.
8876 if (env
->src_grp_type
== group_misfit_task
) {
8877 if (rq
->misfit_task_load
> busiest_load
) {
8878 busiest_load
= rq
->misfit_task_load
;
8885 capacity
= capacity_of(i
);
8888 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8889 * eventually lead to active_balancing high->low capacity.
8890 * Higher per-CPU capacity is considered better than balancing
8893 if (env
->sd
->flags
& SD_ASYM_CPUCAPACITY
&&
8894 capacity_of(env
->dst_cpu
) < capacity
&&
8895 rq
->nr_running
== 1)
8898 wl
= weighted_cpuload(rq
);
8901 * When comparing with imbalance, use weighted_cpuload()
8902 * which is not scaled with the CPU capacity.
8905 if (rq
->nr_running
== 1 && wl
> env
->imbalance
&&
8906 !check_cpu_capacity(rq
, env
->sd
))
8910 * For the load comparisons with the other CPU's, consider
8911 * the weighted_cpuload() scaled with the CPU capacity, so
8912 * that the load can be moved away from the CPU that is
8913 * potentially running at a lower capacity.
8915 * Thus we're looking for max(wl_i / capacity_i), crosswise
8916 * multiplication to rid ourselves of the division works out
8917 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8918 * our previous maximum.
8920 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
8922 busiest_capacity
= capacity
;
8931 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8932 * so long as it is large enough.
8934 #define MAX_PINNED_INTERVAL 512
8937 asym_active_balance(struct lb_env
*env
)
8940 * ASYM_PACKING needs to force migrate tasks from busy but
8941 * lower priority CPUs in order to pack all tasks in the
8942 * highest priority CPUs.
8944 return env
->idle
!= CPU_NOT_IDLE
&& (env
->sd
->flags
& SD_ASYM_PACKING
) &&
8945 sched_asym_prefer(env
->dst_cpu
, env
->src_cpu
);
8949 voluntary_active_balance(struct lb_env
*env
)
8951 struct sched_domain
*sd
= env
->sd
;
8953 if (asym_active_balance(env
))
8957 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8958 * It's worth migrating the task if the src_cpu's capacity is reduced
8959 * because of other sched_class or IRQs if more capacity stays
8960 * available on dst_cpu.
8962 if ((env
->idle
!= CPU_NOT_IDLE
) &&
8963 (env
->src_rq
->cfs
.h_nr_running
== 1)) {
8964 if ((check_cpu_capacity(env
->src_rq
, sd
)) &&
8965 (capacity_of(env
->src_cpu
)*sd
->imbalance_pct
< capacity_of(env
->dst_cpu
)*100))
8969 if (env
->src_grp_type
== group_misfit_task
)
8975 static int need_active_balance(struct lb_env
*env
)
8977 struct sched_domain
*sd
= env
->sd
;
8979 if (voluntary_active_balance(env
))
8982 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
8985 static int active_load_balance_cpu_stop(void *data
);
8987 static int should_we_balance(struct lb_env
*env
)
8989 struct sched_group
*sg
= env
->sd
->groups
;
8990 int cpu
, balance_cpu
= -1;
8993 * Ensure the balancing environment is consistent; can happen
8994 * when the softirq triggers 'during' hotplug.
8996 if (!cpumask_test_cpu(env
->dst_cpu
, env
->cpus
))
9000 * In the newly idle case, we will allow all the CPUs
9001 * to do the newly idle load balance.
9003 if (env
->idle
== CPU_NEWLY_IDLE
)
9006 /* Try to find first idle CPU */
9007 for_each_cpu_and(cpu
, group_balance_mask(sg
), env
->cpus
) {
9015 if (balance_cpu
== -1)
9016 balance_cpu
= group_balance_cpu(sg
);
9019 * First idle CPU or the first CPU(busiest) in this sched group
9020 * is eligible for doing load balancing at this and above domains.
9022 return balance_cpu
== env
->dst_cpu
;
9026 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9027 * tasks if there is an imbalance.
9029 static int load_balance(int this_cpu
, struct rq
*this_rq
,
9030 struct sched_domain
*sd
, enum cpu_idle_type idle
,
9031 int *continue_balancing
)
9033 int ld_moved
, cur_ld_moved
, active_balance
= 0;
9034 struct sched_domain
*sd_parent
= sd
->parent
;
9035 struct sched_group
*group
;
9038 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
9040 struct lb_env env
= {
9042 .dst_cpu
= this_cpu
,
9044 .dst_grpmask
= sched_group_span(sd
->groups
),
9046 .loop_break
= sched_nr_migrate_break
,
9049 .tasks
= LIST_HEAD_INIT(env
.tasks
),
9052 cpumask_and(cpus
, sched_domain_span(sd
), cpu_active_mask
);
9054 schedstat_inc(sd
->lb_count
[idle
]);
9057 if (!should_we_balance(&env
)) {
9058 *continue_balancing
= 0;
9062 group
= find_busiest_group(&env
);
9064 schedstat_inc(sd
->lb_nobusyg
[idle
]);
9068 busiest
= find_busiest_queue(&env
, group
);
9070 schedstat_inc(sd
->lb_nobusyq
[idle
]);
9074 BUG_ON(busiest
== env
.dst_rq
);
9076 schedstat_add(sd
->lb_imbalance
[idle
], env
.imbalance
);
9078 env
.src_cpu
= busiest
->cpu
;
9079 env
.src_rq
= busiest
;
9082 if (busiest
->nr_running
> 1) {
9084 * Attempt to move tasks. If find_busiest_group has found
9085 * an imbalance but busiest->nr_running <= 1, the group is
9086 * still unbalanced. ld_moved simply stays zero, so it is
9087 * correctly treated as an imbalance.
9089 env
.flags
|= LBF_ALL_PINNED
;
9090 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
9093 rq_lock_irqsave(busiest
, &rf
);
9094 update_rq_clock(busiest
);
9097 * cur_ld_moved - load moved in current iteration
9098 * ld_moved - cumulative load moved across iterations
9100 cur_ld_moved
= detach_tasks(&env
);
9103 * We've detached some tasks from busiest_rq. Every
9104 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9105 * unlock busiest->lock, and we are able to be sure
9106 * that nobody can manipulate the tasks in parallel.
9107 * See task_rq_lock() family for the details.
9110 rq_unlock(busiest
, &rf
);
9114 ld_moved
+= cur_ld_moved
;
9117 local_irq_restore(rf
.flags
);
9119 if (env
.flags
& LBF_NEED_BREAK
) {
9120 env
.flags
&= ~LBF_NEED_BREAK
;
9125 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9126 * us and move them to an alternate dst_cpu in our sched_group
9127 * where they can run. The upper limit on how many times we
9128 * iterate on same src_cpu is dependent on number of CPUs in our
9131 * This changes load balance semantics a bit on who can move
9132 * load to a given_cpu. In addition to the given_cpu itself
9133 * (or a ilb_cpu acting on its behalf where given_cpu is
9134 * nohz-idle), we now have balance_cpu in a position to move
9135 * load to given_cpu. In rare situations, this may cause
9136 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9137 * _independently_ and at _same_ time to move some load to
9138 * given_cpu) causing exceess load to be moved to given_cpu.
9139 * This however should not happen so much in practice and
9140 * moreover subsequent load balance cycles should correct the
9141 * excess load moved.
9143 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
9145 /* Prevent to re-select dst_cpu via env's CPUs */
9146 __cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
9148 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
9149 env
.dst_cpu
= env
.new_dst_cpu
;
9150 env
.flags
&= ~LBF_DST_PINNED
;
9152 env
.loop_break
= sched_nr_migrate_break
;
9155 * Go back to "more_balance" rather than "redo" since we
9156 * need to continue with same src_cpu.
9162 * We failed to reach balance because of affinity.
9165 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
9167 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
9168 *group_imbalance
= 1;
9171 /* All tasks on this runqueue were pinned by CPU affinity */
9172 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
9173 __cpumask_clear_cpu(cpu_of(busiest
), cpus
);
9175 * Attempting to continue load balancing at the current
9176 * sched_domain level only makes sense if there are
9177 * active CPUs remaining as possible busiest CPUs to
9178 * pull load from which are not contained within the
9179 * destination group that is receiving any migrated
9182 if (!cpumask_subset(cpus
, env
.dst_grpmask
)) {
9184 env
.loop_break
= sched_nr_migrate_break
;
9187 goto out_all_pinned
;
9192 schedstat_inc(sd
->lb_failed
[idle
]);
9194 * Increment the failure counter only on periodic balance.
9195 * We do not want newidle balance, which can be very
9196 * frequent, pollute the failure counter causing
9197 * excessive cache_hot migrations and active balances.
9199 if (idle
!= CPU_NEWLY_IDLE
)
9200 sd
->nr_balance_failed
++;
9202 if (need_active_balance(&env
)) {
9203 unsigned long flags
;
9205 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
9208 * Don't kick the active_load_balance_cpu_stop,
9209 * if the curr task on busiest CPU can't be
9210 * moved to this_cpu:
9212 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
9213 raw_spin_unlock_irqrestore(&busiest
->lock
,
9215 env
.flags
|= LBF_ALL_PINNED
;
9216 goto out_one_pinned
;
9220 * ->active_balance synchronizes accesses to
9221 * ->active_balance_work. Once set, it's cleared
9222 * only after active load balance is finished.
9224 if (!busiest
->active_balance
) {
9225 busiest
->active_balance
= 1;
9226 busiest
->push_cpu
= this_cpu
;
9229 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
9231 if (active_balance
) {
9232 stop_one_cpu_nowait(cpu_of(busiest
),
9233 active_load_balance_cpu_stop
, busiest
,
9234 &busiest
->active_balance_work
);
9237 /* We've kicked active balancing, force task migration. */
9238 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
9241 sd
->nr_balance_failed
= 0;
9243 if (likely(!active_balance
) || voluntary_active_balance(&env
)) {
9244 /* We were unbalanced, so reset the balancing interval */
9245 sd
->balance_interval
= sd
->min_interval
;
9248 * If we've begun active balancing, start to back off. This
9249 * case may not be covered by the all_pinned logic if there
9250 * is only 1 task on the busy runqueue (because we don't call
9253 if (sd
->balance_interval
< sd
->max_interval
)
9254 sd
->balance_interval
*= 2;
9261 * We reach balance although we may have faced some affinity
9262 * constraints. Clear the imbalance flag if it was set.
9265 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
9267 if (*group_imbalance
)
9268 *group_imbalance
= 0;
9273 * We reach balance because all tasks are pinned at this level so
9274 * we can't migrate them. Let the imbalance flag set so parent level
9275 * can try to migrate them.
9277 schedstat_inc(sd
->lb_balanced
[idle
]);
9279 sd
->nr_balance_failed
= 0;
9285 * idle_balance() disregards balance intervals, so we could repeatedly
9286 * reach this code, which would lead to balance_interval skyrocketting
9287 * in a short amount of time. Skip the balance_interval increase logic
9290 if (env
.idle
== CPU_NEWLY_IDLE
)
9293 /* tune up the balancing interval */
9294 if ((env
.flags
& LBF_ALL_PINNED
&&
9295 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
9296 sd
->balance_interval
< sd
->max_interval
)
9297 sd
->balance_interval
*= 2;
9302 static inline unsigned long
9303 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
9305 unsigned long interval
= sd
->balance_interval
;
9308 interval
*= sd
->busy_factor
;
9310 /* scale ms to jiffies */
9311 interval
= msecs_to_jiffies(interval
);
9312 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
9318 update_next_balance(struct sched_domain
*sd
, unsigned long *next_balance
)
9320 unsigned long interval
, next
;
9322 /* used by idle balance, so cpu_busy = 0 */
9323 interval
= get_sd_balance_interval(sd
, 0);
9324 next
= sd
->last_balance
+ interval
;
9326 if (time_after(*next_balance
, next
))
9327 *next_balance
= next
;
9331 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9332 * running tasks off the busiest CPU onto idle CPUs. It requires at
9333 * least 1 task to be running on each physical CPU where possible, and
9334 * avoids physical / logical imbalances.
9336 static int active_load_balance_cpu_stop(void *data
)
9338 struct rq
*busiest_rq
= data
;
9339 int busiest_cpu
= cpu_of(busiest_rq
);
9340 int target_cpu
= busiest_rq
->push_cpu
;
9341 struct rq
*target_rq
= cpu_rq(target_cpu
);
9342 struct sched_domain
*sd
;
9343 struct task_struct
*p
= NULL
;
9346 rq_lock_irq(busiest_rq
, &rf
);
9348 * Between queueing the stop-work and running it is a hole in which
9349 * CPUs can become inactive. We should not move tasks from or to
9352 if (!cpu_active(busiest_cpu
) || !cpu_active(target_cpu
))
9355 /* Make sure the requested CPU hasn't gone down in the meantime: */
9356 if (unlikely(busiest_cpu
!= smp_processor_id() ||
9357 !busiest_rq
->active_balance
))
9360 /* Is there any task to move? */
9361 if (busiest_rq
->nr_running
<= 1)
9365 * This condition is "impossible", if it occurs
9366 * we need to fix it. Originally reported by
9367 * Bjorn Helgaas on a 128-CPU setup.
9369 BUG_ON(busiest_rq
== target_rq
);
9371 /* Search for an sd spanning us and the target CPU. */
9373 for_each_domain(target_cpu
, sd
) {
9374 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
9375 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
9380 struct lb_env env
= {
9382 .dst_cpu
= target_cpu
,
9383 .dst_rq
= target_rq
,
9384 .src_cpu
= busiest_rq
->cpu
,
9385 .src_rq
= busiest_rq
,
9388 * can_migrate_task() doesn't need to compute new_dst_cpu
9389 * for active balancing. Since we have CPU_IDLE, but no
9390 * @dst_grpmask we need to make that test go away with lying
9393 .flags
= LBF_DST_PINNED
,
9396 schedstat_inc(sd
->alb_count
);
9397 update_rq_clock(busiest_rq
);
9399 p
= detach_one_task(&env
);
9401 schedstat_inc(sd
->alb_pushed
);
9402 /* Active balancing done, reset the failure counter. */
9403 sd
->nr_balance_failed
= 0;
9405 schedstat_inc(sd
->alb_failed
);
9410 busiest_rq
->active_balance
= 0;
9411 rq_unlock(busiest_rq
, &rf
);
9414 attach_one_task(target_rq
, p
);
9421 static DEFINE_SPINLOCK(balancing
);
9424 * Scale the max load_balance interval with the number of CPUs in the system.
9425 * This trades load-balance latency on larger machines for less cross talk.
9427 void update_max_interval(void)
9429 max_load_balance_interval
= HZ
*num_online_cpus()/10;
9433 * It checks each scheduling domain to see if it is due to be balanced,
9434 * and initiates a balancing operation if so.
9436 * Balancing parameters are set up in init_sched_domains.
9438 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
9440 int continue_balancing
= 1;
9442 unsigned long interval
;
9443 struct sched_domain
*sd
;
9444 /* Earliest time when we have to do rebalance again */
9445 unsigned long next_balance
= jiffies
+ 60*HZ
;
9446 int update_next_balance
= 0;
9447 int need_serialize
, need_decay
= 0;
9451 for_each_domain(cpu
, sd
) {
9453 * Decay the newidle max times here because this is a regular
9454 * visit to all the domains. Decay ~1% per second.
9456 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
9457 sd
->max_newidle_lb_cost
=
9458 (sd
->max_newidle_lb_cost
* 253) / 256;
9459 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
9462 max_cost
+= sd
->max_newidle_lb_cost
;
9464 if (!(sd
->flags
& SD_LOAD_BALANCE
))
9468 * Stop the load balance at this level. There is another
9469 * CPU in our sched group which is doing load balancing more
9472 if (!continue_balancing
) {
9478 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
9480 need_serialize
= sd
->flags
& SD_SERIALIZE
;
9481 if (need_serialize
) {
9482 if (!spin_trylock(&balancing
))
9486 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
9487 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
9489 * The LBF_DST_PINNED logic could have changed
9490 * env->dst_cpu, so we can't know our idle
9491 * state even if we migrated tasks. Update it.
9493 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
9495 sd
->last_balance
= jiffies
;
9496 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
9499 spin_unlock(&balancing
);
9501 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
9502 next_balance
= sd
->last_balance
+ interval
;
9503 update_next_balance
= 1;
9508 * Ensure the rq-wide value also decays but keep it at a
9509 * reasonable floor to avoid funnies with rq->avg_idle.
9511 rq
->max_idle_balance_cost
=
9512 max((u64
)sysctl_sched_migration_cost
, max_cost
);
9517 * next_balance will be updated only when there is a need.
9518 * When the cpu is attached to null domain for ex, it will not be
9521 if (likely(update_next_balance
)) {
9522 rq
->next_balance
= next_balance
;
9524 #ifdef CONFIG_NO_HZ_COMMON
9526 * If this CPU has been elected to perform the nohz idle
9527 * balance. Other idle CPUs have already rebalanced with
9528 * nohz_idle_balance() and nohz.next_balance has been
9529 * updated accordingly. This CPU is now running the idle load
9530 * balance for itself and we need to update the
9531 * nohz.next_balance accordingly.
9533 if ((idle
== CPU_IDLE
) && time_after(nohz
.next_balance
, rq
->next_balance
))
9534 nohz
.next_balance
= rq
->next_balance
;
9539 static inline int on_null_domain(struct rq
*rq
)
9541 return unlikely(!rcu_dereference_sched(rq
->sd
));
9544 #ifdef CONFIG_NO_HZ_COMMON
9546 * idle load balancing details
9547 * - When one of the busy CPUs notice that there may be an idle rebalancing
9548 * needed, they will kick the idle load balancer, which then does idle
9549 * load balancing for all the idle CPUs.
9552 static inline int find_new_ilb(void)
9554 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
9556 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
9563 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9564 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9565 * CPU (if there is one).
9567 static void kick_ilb(unsigned int flags
)
9571 nohz
.next_balance
++;
9573 ilb_cpu
= find_new_ilb();
9575 if (ilb_cpu
>= nr_cpu_ids
)
9578 flags
= atomic_fetch_or(flags
, nohz_flags(ilb_cpu
));
9579 if (flags
& NOHZ_KICK_MASK
)
9583 * Use smp_send_reschedule() instead of resched_cpu().
9584 * This way we generate a sched IPI on the target CPU which
9585 * is idle. And the softirq performing nohz idle load balance
9586 * will be run before returning from the IPI.
9588 smp_send_reschedule(ilb_cpu
);
9592 * Current decision point for kicking the idle load balancer in the presence
9593 * of idle CPUs in the system.
9595 static void nohz_balancer_kick(struct rq
*rq
)
9597 unsigned long now
= jiffies
;
9598 struct sched_domain_shared
*sds
;
9599 struct sched_domain
*sd
;
9600 int nr_busy
, i
, cpu
= rq
->cpu
;
9601 unsigned int flags
= 0;
9603 if (unlikely(rq
->idle_balance
))
9607 * We may be recently in ticked or tickless idle mode. At the first
9608 * busy tick after returning from idle, we will update the busy stats.
9610 nohz_balance_exit_idle(rq
);
9613 * None are in tickless mode and hence no need for NOHZ idle load
9616 if (likely(!atomic_read(&nohz
.nr_cpus
)))
9619 if (READ_ONCE(nohz
.has_blocked
) &&
9620 time_after(now
, READ_ONCE(nohz
.next_blocked
)))
9621 flags
= NOHZ_STATS_KICK
;
9623 if (time_before(now
, nohz
.next_balance
))
9626 if (rq
->nr_running
>= 2) {
9627 flags
= NOHZ_KICK_MASK
;
9633 sd
= rcu_dereference(rq
->sd
);
9636 * If there's a CFS task and the current CPU has reduced
9637 * capacity; kick the ILB to see if there's a better CPU to run
9640 if (rq
->cfs
.h_nr_running
>= 1 && check_cpu_capacity(rq
, sd
)) {
9641 flags
= NOHZ_KICK_MASK
;
9646 sd
= rcu_dereference(per_cpu(sd_asym_packing
, cpu
));
9649 * When ASYM_PACKING; see if there's a more preferred CPU
9650 * currently idle; in which case, kick the ILB to move tasks
9653 for_each_cpu_and(i
, sched_domain_span(sd
), nohz
.idle_cpus_mask
) {
9654 if (sched_asym_prefer(i
, cpu
)) {
9655 flags
= NOHZ_KICK_MASK
;
9661 sd
= rcu_dereference(per_cpu(sd_asym_cpucapacity
, cpu
));
9664 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
9665 * to run the misfit task on.
9667 if (check_misfit_status(rq
, sd
)) {
9668 flags
= NOHZ_KICK_MASK
;
9673 * For asymmetric systems, we do not want to nicely balance
9674 * cache use, instead we want to embrace asymmetry and only
9675 * ensure tasks have enough CPU capacity.
9677 * Skip the LLC logic because it's not relevant in that case.
9682 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
9685 * If there is an imbalance between LLC domains (IOW we could
9686 * increase the overall cache use), we need some less-loaded LLC
9687 * domain to pull some load. Likewise, we may need to spread
9688 * load within the current LLC domain (e.g. packed SMT cores but
9689 * other CPUs are idle). We can't really know from here how busy
9690 * the others are - so just get a nohz balance going if it looks
9691 * like this LLC domain has tasks we could move.
9693 nr_busy
= atomic_read(&sds
->nr_busy_cpus
);
9695 flags
= NOHZ_KICK_MASK
;
9706 static void set_cpu_sd_state_busy(int cpu
)
9708 struct sched_domain
*sd
;
9711 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
9713 if (!sd
|| !sd
->nohz_idle
)
9717 atomic_inc(&sd
->shared
->nr_busy_cpus
);
9722 void nohz_balance_exit_idle(struct rq
*rq
)
9724 SCHED_WARN_ON(rq
!= this_rq());
9726 if (likely(!rq
->nohz_tick_stopped
))
9729 rq
->nohz_tick_stopped
= 0;
9730 cpumask_clear_cpu(rq
->cpu
, nohz
.idle_cpus_mask
);
9731 atomic_dec(&nohz
.nr_cpus
);
9733 set_cpu_sd_state_busy(rq
->cpu
);
9736 static void set_cpu_sd_state_idle(int cpu
)
9738 struct sched_domain
*sd
;
9741 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
9743 if (!sd
|| sd
->nohz_idle
)
9747 atomic_dec(&sd
->shared
->nr_busy_cpus
);
9753 * This routine will record that the CPU is going idle with tick stopped.
9754 * This info will be used in performing idle load balancing in the future.
9756 void nohz_balance_enter_idle(int cpu
)
9758 struct rq
*rq
= cpu_rq(cpu
);
9760 SCHED_WARN_ON(cpu
!= smp_processor_id());
9762 /* If this CPU is going down, then nothing needs to be done: */
9763 if (!cpu_active(cpu
))
9766 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9767 if (!housekeeping_cpu(cpu
, HK_FLAG_SCHED
))
9771 * Can be set safely without rq->lock held
9772 * If a clear happens, it will have evaluated last additions because
9773 * rq->lock is held during the check and the clear
9775 rq
->has_blocked_load
= 1;
9778 * The tick is still stopped but load could have been added in the
9779 * meantime. We set the nohz.has_blocked flag to trig a check of the
9780 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9781 * of nohz.has_blocked can only happen after checking the new load
9783 if (rq
->nohz_tick_stopped
)
9786 /* If we're a completely isolated CPU, we don't play: */
9787 if (on_null_domain(rq
))
9790 rq
->nohz_tick_stopped
= 1;
9792 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
9793 atomic_inc(&nohz
.nr_cpus
);
9796 * Ensures that if nohz_idle_balance() fails to observe our
9797 * @idle_cpus_mask store, it must observe the @has_blocked
9800 smp_mb__after_atomic();
9802 set_cpu_sd_state_idle(cpu
);
9806 * Each time a cpu enter idle, we assume that it has blocked load and
9807 * enable the periodic update of the load of idle cpus
9809 WRITE_ONCE(nohz
.has_blocked
, 1);
9813 * Internal function that runs load balance for all idle cpus. The load balance
9814 * can be a simple update of blocked load or a complete load balance with
9815 * tasks movement depending of flags.
9816 * The function returns false if the loop has stopped before running
9817 * through all idle CPUs.
9819 static bool _nohz_idle_balance(struct rq
*this_rq
, unsigned int flags
,
9820 enum cpu_idle_type idle
)
9822 /* Earliest time when we have to do rebalance again */
9823 unsigned long now
= jiffies
;
9824 unsigned long next_balance
= now
+ 60*HZ
;
9825 bool has_blocked_load
= false;
9826 int update_next_balance
= 0;
9827 int this_cpu
= this_rq
->cpu
;
9832 SCHED_WARN_ON((flags
& NOHZ_KICK_MASK
) == NOHZ_BALANCE_KICK
);
9835 * We assume there will be no idle load after this update and clear
9836 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9837 * set the has_blocked flag and trig another update of idle load.
9838 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9839 * setting the flag, we are sure to not clear the state and not
9840 * check the load of an idle cpu.
9842 WRITE_ONCE(nohz
.has_blocked
, 0);
9845 * Ensures that if we miss the CPU, we must see the has_blocked
9846 * store from nohz_balance_enter_idle().
9850 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
9851 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
9855 * If this CPU gets work to do, stop the load balancing
9856 * work being done for other CPUs. Next load
9857 * balancing owner will pick it up.
9859 if (need_resched()) {
9860 has_blocked_load
= true;
9864 rq
= cpu_rq(balance_cpu
);
9866 has_blocked_load
|= update_nohz_stats(rq
, true);
9869 * If time for next balance is due,
9872 if (time_after_eq(jiffies
, rq
->next_balance
)) {
9875 rq_lock_irqsave(rq
, &rf
);
9876 update_rq_clock(rq
);
9877 cpu_load_update_idle(rq
);
9878 rq_unlock_irqrestore(rq
, &rf
);
9880 if (flags
& NOHZ_BALANCE_KICK
)
9881 rebalance_domains(rq
, CPU_IDLE
);
9884 if (time_after(next_balance
, rq
->next_balance
)) {
9885 next_balance
= rq
->next_balance
;
9886 update_next_balance
= 1;
9890 /* Newly idle CPU doesn't need an update */
9891 if (idle
!= CPU_NEWLY_IDLE
) {
9892 update_blocked_averages(this_cpu
);
9893 has_blocked_load
|= this_rq
->has_blocked_load
;
9896 if (flags
& NOHZ_BALANCE_KICK
)
9897 rebalance_domains(this_rq
, CPU_IDLE
);
9899 WRITE_ONCE(nohz
.next_blocked
,
9900 now
+ msecs_to_jiffies(LOAD_AVG_PERIOD
));
9902 /* The full idle balance loop has been done */
9906 /* There is still blocked load, enable periodic update */
9907 if (has_blocked_load
)
9908 WRITE_ONCE(nohz
.has_blocked
, 1);
9911 * next_balance will be updated only when there is a need.
9912 * When the CPU is attached to null domain for ex, it will not be
9915 if (likely(update_next_balance
))
9916 nohz
.next_balance
= next_balance
;
9922 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9923 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9925 static bool nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
9927 int this_cpu
= this_rq
->cpu
;
9930 if (!(atomic_read(nohz_flags(this_cpu
)) & NOHZ_KICK_MASK
))
9933 if (idle
!= CPU_IDLE
) {
9934 atomic_andnot(NOHZ_KICK_MASK
, nohz_flags(this_cpu
));
9938 /* could be _relaxed() */
9939 flags
= atomic_fetch_andnot(NOHZ_KICK_MASK
, nohz_flags(this_cpu
));
9940 if (!(flags
& NOHZ_KICK_MASK
))
9943 _nohz_idle_balance(this_rq
, flags
, idle
);
9948 static void nohz_newidle_balance(struct rq
*this_rq
)
9950 int this_cpu
= this_rq
->cpu
;
9953 * This CPU doesn't want to be disturbed by scheduler
9956 if (!housekeeping_cpu(this_cpu
, HK_FLAG_SCHED
))
9959 /* Will wake up very soon. No time for doing anything else*/
9960 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
9963 /* Don't need to update blocked load of idle CPUs*/
9964 if (!READ_ONCE(nohz
.has_blocked
) ||
9965 time_before(jiffies
, READ_ONCE(nohz
.next_blocked
)))
9968 raw_spin_unlock(&this_rq
->lock
);
9970 * This CPU is going to be idle and blocked load of idle CPUs
9971 * need to be updated. Run the ilb locally as it is a good
9972 * candidate for ilb instead of waking up another idle CPU.
9973 * Kick an normal ilb if we failed to do the update.
9975 if (!_nohz_idle_balance(this_rq
, NOHZ_STATS_KICK
, CPU_NEWLY_IDLE
))
9976 kick_ilb(NOHZ_STATS_KICK
);
9977 raw_spin_lock(&this_rq
->lock
);
9980 #else /* !CONFIG_NO_HZ_COMMON */
9981 static inline void nohz_balancer_kick(struct rq
*rq
) { }
9983 static inline bool nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
9988 static inline void nohz_newidle_balance(struct rq
*this_rq
) { }
9989 #endif /* CONFIG_NO_HZ_COMMON */
9992 * idle_balance is called by schedule() if this_cpu is about to become
9993 * idle. Attempts to pull tasks from other CPUs.
9995 static int idle_balance(struct rq
*this_rq
, struct rq_flags
*rf
)
9997 unsigned long next_balance
= jiffies
+ HZ
;
9998 int this_cpu
= this_rq
->cpu
;
9999 struct sched_domain
*sd
;
10000 int pulled_task
= 0;
10004 * We must set idle_stamp _before_ calling idle_balance(), such that we
10005 * measure the duration of idle_balance() as idle time.
10007 this_rq
->idle_stamp
= rq_clock(this_rq
);
10010 * Do not pull tasks towards !active CPUs...
10012 if (!cpu_active(this_cpu
))
10016 * This is OK, because current is on_cpu, which avoids it being picked
10017 * for load-balance and preemption/IRQs are still disabled avoiding
10018 * further scheduler activity on it and we're being very careful to
10019 * re-start the picking loop.
10021 rq_unpin_lock(this_rq
, rf
);
10023 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
10024 !READ_ONCE(this_rq
->rd
->overload
)) {
10027 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
10029 update_next_balance(sd
, &next_balance
);
10032 nohz_newidle_balance(this_rq
);
10037 raw_spin_unlock(&this_rq
->lock
);
10039 update_blocked_averages(this_cpu
);
10041 for_each_domain(this_cpu
, sd
) {
10042 int continue_balancing
= 1;
10043 u64 t0
, domain_cost
;
10045 if (!(sd
->flags
& SD_LOAD_BALANCE
))
10048 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
10049 update_next_balance(sd
, &next_balance
);
10053 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
10054 t0
= sched_clock_cpu(this_cpu
);
10056 pulled_task
= load_balance(this_cpu
, this_rq
,
10057 sd
, CPU_NEWLY_IDLE
,
10058 &continue_balancing
);
10060 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
10061 if (domain_cost
> sd
->max_newidle_lb_cost
)
10062 sd
->max_newidle_lb_cost
= domain_cost
;
10064 curr_cost
+= domain_cost
;
10067 update_next_balance(sd
, &next_balance
);
10070 * Stop searching for tasks to pull if there are
10071 * now runnable tasks on this rq.
10073 if (pulled_task
|| this_rq
->nr_running
> 0)
10078 raw_spin_lock(&this_rq
->lock
);
10080 if (curr_cost
> this_rq
->max_idle_balance_cost
)
10081 this_rq
->max_idle_balance_cost
= curr_cost
;
10085 * While browsing the domains, we released the rq lock, a task could
10086 * have been enqueued in the meantime. Since we're not going idle,
10087 * pretend we pulled a task.
10089 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
10092 /* Move the next balance forward */
10093 if (time_after(this_rq
->next_balance
, next_balance
))
10094 this_rq
->next_balance
= next_balance
;
10096 /* Is there a task of a high priority class? */
10097 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
10101 this_rq
->idle_stamp
= 0;
10103 rq_repin_lock(this_rq
, rf
);
10105 return pulled_task
;
10109 * run_rebalance_domains is triggered when needed from the scheduler tick.
10110 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10112 static __latent_entropy
void run_rebalance_domains(struct softirq_action
*h
)
10114 struct rq
*this_rq
= this_rq();
10115 enum cpu_idle_type idle
= this_rq
->idle_balance
?
10116 CPU_IDLE
: CPU_NOT_IDLE
;
10119 * If this CPU has a pending nohz_balance_kick, then do the
10120 * balancing on behalf of the other idle CPUs whose ticks are
10121 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10122 * give the idle CPUs a chance to load balance. Else we may
10123 * load balance only within the local sched_domain hierarchy
10124 * and abort nohz_idle_balance altogether if we pull some load.
10126 if (nohz_idle_balance(this_rq
, idle
))
10129 /* normal load balance */
10130 update_blocked_averages(this_rq
->cpu
);
10131 rebalance_domains(this_rq
, idle
);
10135 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10137 void trigger_load_balance(struct rq
*rq
)
10139 /* Don't need to rebalance while attached to NULL domain */
10140 if (unlikely(on_null_domain(rq
)))
10143 if (time_after_eq(jiffies
, rq
->next_balance
))
10144 raise_softirq(SCHED_SOFTIRQ
);
10146 nohz_balancer_kick(rq
);
10149 static void rq_online_fair(struct rq
*rq
)
10153 update_runtime_enabled(rq
);
10156 static void rq_offline_fair(struct rq
*rq
)
10160 /* Ensure any throttled groups are reachable by pick_next_task */
10161 unthrottle_offline_cfs_rqs(rq
);
10164 #endif /* CONFIG_SMP */
10167 * scheduler tick hitting a task of our scheduling class.
10169 * NOTE: This function can be called remotely by the tick offload that
10170 * goes along full dynticks. Therefore no local assumption can be made
10171 * and everything must be accessed through the @rq and @curr passed in
10174 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
10176 struct cfs_rq
*cfs_rq
;
10177 struct sched_entity
*se
= &curr
->se
;
10179 for_each_sched_entity(se
) {
10180 cfs_rq
= cfs_rq_of(se
);
10181 entity_tick(cfs_rq
, se
, queued
);
10184 if (static_branch_unlikely(&sched_numa_balancing
))
10185 task_tick_numa(rq
, curr
);
10187 update_misfit_status(curr
, rq
);
10188 update_overutilized_status(task_rq(curr
));
10192 * called on fork with the child task as argument from the parent's context
10193 * - child not yet on the tasklist
10194 * - preemption disabled
10196 static void task_fork_fair(struct task_struct
*p
)
10198 struct cfs_rq
*cfs_rq
;
10199 struct sched_entity
*se
= &p
->se
, *curr
;
10200 struct rq
*rq
= this_rq();
10201 struct rq_flags rf
;
10204 update_rq_clock(rq
);
10206 cfs_rq
= task_cfs_rq(current
);
10207 curr
= cfs_rq
->curr
;
10209 update_curr(cfs_rq
);
10210 se
->vruntime
= curr
->vruntime
;
10212 place_entity(cfs_rq
, se
, 1);
10214 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
10216 * Upon rescheduling, sched_class::put_prev_task() will place
10217 * 'current' within the tree based on its new key value.
10219 swap(curr
->vruntime
, se
->vruntime
);
10223 se
->vruntime
-= cfs_rq
->min_vruntime
;
10224 rq_unlock(rq
, &rf
);
10228 * Priority of the task has changed. Check to see if we preempt
10229 * the current task.
10232 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
10234 if (!task_on_rq_queued(p
))
10238 * Reschedule if we are currently running on this runqueue and
10239 * our priority decreased, or if we are not currently running on
10240 * this runqueue and our priority is higher than the current's
10242 if (rq
->curr
== p
) {
10243 if (p
->prio
> oldprio
)
10246 check_preempt_curr(rq
, p
, 0);
10249 static inline bool vruntime_normalized(struct task_struct
*p
)
10251 struct sched_entity
*se
= &p
->se
;
10254 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10255 * the dequeue_entity(.flags=0) will already have normalized the
10262 * When !on_rq, vruntime of the task has usually NOT been normalized.
10263 * But there are some cases where it has already been normalized:
10265 * - A forked child which is waiting for being woken up by
10266 * wake_up_new_task().
10267 * - A task which has been woken up by try_to_wake_up() and
10268 * waiting for actually being woken up by sched_ttwu_pending().
10270 if (!se
->sum_exec_runtime
||
10271 (p
->state
== TASK_WAKING
&& p
->sched_remote_wakeup
))
10277 #ifdef CONFIG_FAIR_GROUP_SCHED
10279 * Propagate the changes of the sched_entity across the tg tree to make it
10280 * visible to the root
10282 static void propagate_entity_cfs_rq(struct sched_entity
*se
)
10284 struct cfs_rq
*cfs_rq
;
10286 /* Start to propagate at parent */
10289 for_each_sched_entity(se
) {
10290 cfs_rq
= cfs_rq_of(se
);
10292 if (cfs_rq_throttled(cfs_rq
))
10295 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
10299 static void propagate_entity_cfs_rq(struct sched_entity
*se
) { }
10302 static void detach_entity_cfs_rq(struct sched_entity
*se
)
10304 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10306 /* Catch up with the cfs_rq and remove our load when we leave */
10307 update_load_avg(cfs_rq
, se
, 0);
10308 detach_entity_load_avg(cfs_rq
, se
);
10309 update_tg_load_avg(cfs_rq
, false);
10310 propagate_entity_cfs_rq(se
);
10313 static void attach_entity_cfs_rq(struct sched_entity
*se
)
10315 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10317 #ifdef CONFIG_FAIR_GROUP_SCHED
10319 * Since the real-depth could have been changed (only FAIR
10320 * class maintain depth value), reset depth properly.
10322 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
10325 /* Synchronize entity with its cfs_rq */
10326 update_load_avg(cfs_rq
, se
, sched_feat(ATTACH_AGE_LOAD
) ? 0 : SKIP_AGE_LOAD
);
10327 attach_entity_load_avg(cfs_rq
, se
, 0);
10328 update_tg_load_avg(cfs_rq
, false);
10329 propagate_entity_cfs_rq(se
);
10332 static void detach_task_cfs_rq(struct task_struct
*p
)
10334 struct sched_entity
*se
= &p
->se
;
10335 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10337 if (!vruntime_normalized(p
)) {
10339 * Fix up our vruntime so that the current sleep doesn't
10340 * cause 'unlimited' sleep bonus.
10342 place_entity(cfs_rq
, se
, 0);
10343 se
->vruntime
-= cfs_rq
->min_vruntime
;
10346 detach_entity_cfs_rq(se
);
10349 static void attach_task_cfs_rq(struct task_struct
*p
)
10351 struct sched_entity
*se
= &p
->se
;
10352 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10354 attach_entity_cfs_rq(se
);
10356 if (!vruntime_normalized(p
))
10357 se
->vruntime
+= cfs_rq
->min_vruntime
;
10360 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
10362 detach_task_cfs_rq(p
);
10365 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
10367 attach_task_cfs_rq(p
);
10369 if (task_on_rq_queued(p
)) {
10371 * We were most likely switched from sched_rt, so
10372 * kick off the schedule if running, otherwise just see
10373 * if we can still preempt the current task.
10378 check_preempt_curr(rq
, p
, 0);
10382 /* Account for a task changing its policy or group.
10384 * This routine is mostly called to set cfs_rq->curr field when a task
10385 * migrates between groups/classes.
10387 static void set_curr_task_fair(struct rq
*rq
)
10389 struct sched_entity
*se
= &rq
->curr
->se
;
10391 for_each_sched_entity(se
) {
10392 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
10394 set_next_entity(cfs_rq
, se
);
10395 /* ensure bandwidth has been allocated on our new cfs_rq */
10396 account_cfs_rq_runtime(cfs_rq
, 0);
10400 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
10402 cfs_rq
->tasks_timeline
= RB_ROOT_CACHED
;
10403 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
10404 #ifndef CONFIG_64BIT
10405 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
10408 raw_spin_lock_init(&cfs_rq
->removed
.lock
);
10412 #ifdef CONFIG_FAIR_GROUP_SCHED
10413 static void task_set_group_fair(struct task_struct
*p
)
10415 struct sched_entity
*se
= &p
->se
;
10417 set_task_rq(p
, task_cpu(p
));
10418 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
10421 static void task_move_group_fair(struct task_struct
*p
)
10423 detach_task_cfs_rq(p
);
10424 set_task_rq(p
, task_cpu(p
));
10427 /* Tell se's cfs_rq has been changed -- migrated */
10428 p
->se
.avg
.last_update_time
= 0;
10430 attach_task_cfs_rq(p
);
10433 static void task_change_group_fair(struct task_struct
*p
, int type
)
10436 case TASK_SET_GROUP
:
10437 task_set_group_fair(p
);
10440 case TASK_MOVE_GROUP
:
10441 task_move_group_fair(p
);
10446 void free_fair_sched_group(struct task_group
*tg
)
10450 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
10452 for_each_possible_cpu(i
) {
10454 kfree(tg
->cfs_rq
[i
]);
10463 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
10465 struct sched_entity
*se
;
10466 struct cfs_rq
*cfs_rq
;
10469 tg
->cfs_rq
= kcalloc(nr_cpu_ids
, sizeof(cfs_rq
), GFP_KERNEL
);
10472 tg
->se
= kcalloc(nr_cpu_ids
, sizeof(se
), GFP_KERNEL
);
10476 tg
->shares
= NICE_0_LOAD
;
10478 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
10480 for_each_possible_cpu(i
) {
10481 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
10482 GFP_KERNEL
, cpu_to_node(i
));
10486 se
= kzalloc_node(sizeof(struct sched_entity
),
10487 GFP_KERNEL
, cpu_to_node(i
));
10491 init_cfs_rq(cfs_rq
);
10492 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
10493 init_entity_runnable_average(se
);
10504 void online_fair_sched_group(struct task_group
*tg
)
10506 struct sched_entity
*se
;
10510 for_each_possible_cpu(i
) {
10514 raw_spin_lock_irq(&rq
->lock
);
10515 update_rq_clock(rq
);
10516 attach_entity_cfs_rq(se
);
10517 sync_throttle(tg
, i
);
10518 raw_spin_unlock_irq(&rq
->lock
);
10522 void unregister_fair_sched_group(struct task_group
*tg
)
10524 unsigned long flags
;
10528 for_each_possible_cpu(cpu
) {
10530 remove_entity_load_avg(tg
->se
[cpu
]);
10533 * Only empty task groups can be destroyed; so we can speculatively
10534 * check on_list without danger of it being re-added.
10536 if (!tg
->cfs_rq
[cpu
]->on_list
)
10541 raw_spin_lock_irqsave(&rq
->lock
, flags
);
10542 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
10543 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
10547 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
10548 struct sched_entity
*se
, int cpu
,
10549 struct sched_entity
*parent
)
10551 struct rq
*rq
= cpu_rq(cpu
);
10555 init_cfs_rq_runtime(cfs_rq
);
10557 tg
->cfs_rq
[cpu
] = cfs_rq
;
10560 /* se could be NULL for root_task_group */
10565 se
->cfs_rq
= &rq
->cfs
;
10568 se
->cfs_rq
= parent
->my_q
;
10569 se
->depth
= parent
->depth
+ 1;
10573 /* guarantee group entities always have weight */
10574 update_load_set(&se
->load
, NICE_0_LOAD
);
10575 se
->parent
= parent
;
10578 static DEFINE_MUTEX(shares_mutex
);
10580 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
10585 * We can't change the weight of the root cgroup.
10590 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
10592 mutex_lock(&shares_mutex
);
10593 if (tg
->shares
== shares
)
10596 tg
->shares
= shares
;
10597 for_each_possible_cpu(i
) {
10598 struct rq
*rq
= cpu_rq(i
);
10599 struct sched_entity
*se
= tg
->se
[i
];
10600 struct rq_flags rf
;
10602 /* Propagate contribution to hierarchy */
10603 rq_lock_irqsave(rq
, &rf
);
10604 update_rq_clock(rq
);
10605 for_each_sched_entity(se
) {
10606 update_load_avg(cfs_rq_of(se
), se
, UPDATE_TG
);
10607 update_cfs_group(se
);
10609 rq_unlock_irqrestore(rq
, &rf
);
10613 mutex_unlock(&shares_mutex
);
10616 #else /* CONFIG_FAIR_GROUP_SCHED */
10618 void free_fair_sched_group(struct task_group
*tg
) { }
10620 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
10625 void online_fair_sched_group(struct task_group
*tg
) { }
10627 void unregister_fair_sched_group(struct task_group
*tg
) { }
10629 #endif /* CONFIG_FAIR_GROUP_SCHED */
10632 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
10634 struct sched_entity
*se
= &task
->se
;
10635 unsigned int rr_interval
= 0;
10638 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10641 if (rq
->cfs
.load
.weight
)
10642 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
10644 return rr_interval
;
10648 * All the scheduling class methods:
10650 const struct sched_class fair_sched_class
= {
10651 .next
= &idle_sched_class
,
10652 .enqueue_task
= enqueue_task_fair
,
10653 .dequeue_task
= dequeue_task_fair
,
10654 .yield_task
= yield_task_fair
,
10655 .yield_to_task
= yield_to_task_fair
,
10657 .check_preempt_curr
= check_preempt_wakeup
,
10659 .pick_next_task
= pick_next_task_fair
,
10660 .put_prev_task
= put_prev_task_fair
,
10663 .select_task_rq
= select_task_rq_fair
,
10664 .migrate_task_rq
= migrate_task_rq_fair
,
10666 .rq_online
= rq_online_fair
,
10667 .rq_offline
= rq_offline_fair
,
10669 .task_dead
= task_dead_fair
,
10670 .set_cpus_allowed
= set_cpus_allowed_common
,
10673 .set_curr_task
= set_curr_task_fair
,
10674 .task_tick
= task_tick_fair
,
10675 .task_fork
= task_fork_fair
,
10677 .prio_changed
= prio_changed_fair
,
10678 .switched_from
= switched_from_fair
,
10679 .switched_to
= switched_to_fair
,
10681 .get_rr_interval
= get_rr_interval_fair
,
10683 .update_curr
= update_curr_fair
,
10685 #ifdef CONFIG_FAIR_GROUP_SCHED
10686 .task_change_group
= task_change_group_fair
,
10690 #ifdef CONFIG_SCHED_DEBUG
10691 void print_cfs_stats(struct seq_file
*m
, int cpu
)
10693 struct cfs_rq
*cfs_rq
, *pos
;
10696 for_each_leaf_cfs_rq_safe(cpu_rq(cpu
), cfs_rq
, pos
)
10697 print_cfs_rq(m
, cpu
, cfs_rq
);
10701 #ifdef CONFIG_NUMA_BALANCING
10702 void show_numa_stats(struct task_struct
*p
, struct seq_file
*m
)
10705 unsigned long tsf
= 0, tpf
= 0, gsf
= 0, gpf
= 0;
10707 for_each_online_node(node
) {
10708 if (p
->numa_faults
) {
10709 tsf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
10710 tpf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
10712 if (p
->numa_group
) {
10713 gsf
= p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)],
10714 gpf
= p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
10716 print_numa_stats(m
, node
, tsf
, tpf
, gsf
, gpf
);
10719 #endif /* CONFIG_NUMA_BALANCING */
10720 #endif /* CONFIG_SCHED_DEBUG */
10722 __init
void init_sched_fair_class(void)
10725 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
10727 #ifdef CONFIG_NO_HZ_COMMON
10728 nohz
.next_balance
= jiffies
;
10729 nohz
.next_blocked
= jiffies
;
10730 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);