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
24 #include <linux/sched/mm.h>
25 #include <linux/sched/topology.h>
27 #include <linux/latencytop.h>
28 #include <linux/cpumask.h>
29 #include <linux/cpuidle.h>
30 #include <linux/slab.h>
31 #include <linux/profile.h>
32 #include <linux/interrupt.h>
33 #include <linux/mempolicy.h>
34 #include <linux/migrate.h>
35 #include <linux/task_work.h>
36 #include <linux/sched/isolation.h>
38 #include <trace/events/sched.h>
43 * Targeted preemption latency for CPU-bound tasks:
45 * NOTE: this latency value is not the same as the concept of
46 * 'timeslice length' - timeslices in CFS are of variable length
47 * and have no persistent notion like in traditional, time-slice
48 * based scheduling concepts.
50 * (to see the precise effective timeslice length of your workload,
51 * run vmstat and monitor the context-switches (cs) field)
53 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
55 unsigned int sysctl_sched_latency
= 6000000ULL;
56 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
59 * The initial- and re-scaling of tunables is configurable
63 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
64 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
65 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
67 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
69 enum sched_tunable_scaling sysctl_sched_tunable_scaling
= SCHED_TUNABLESCALING_LOG
;
72 * Minimal preemption granularity for CPU-bound tasks:
74 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
76 unsigned int sysctl_sched_min_granularity
= 750000ULL;
77 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
80 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
82 static unsigned int sched_nr_latency
= 8;
85 * After fork, child runs first. If set to 0 (default) then
86 * parent will (try to) run first.
88 unsigned int sysctl_sched_child_runs_first __read_mostly
;
91 * SCHED_OTHER wake-up granularity.
93 * This option delays the preemption effects of decoupled workloads
94 * and reduces their over-scheduling. Synchronous workloads will still
95 * have immediate wakeup/sleep latencies.
97 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
99 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
100 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
102 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
106 * For asym packing, by default the lower numbered cpu has higher priority.
108 int __weak
arch_asym_cpu_priority(int cpu
)
114 #ifdef CONFIG_CFS_BANDWIDTH
116 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
117 * each time a cfs_rq requests quota.
119 * Note: in the case that the slice exceeds the runtime remaining (either due
120 * to consumption or the quota being specified to be smaller than the slice)
121 * we will always only issue the remaining available time.
123 * (default: 5 msec, units: microseconds)
125 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
129 * The margin used when comparing utilization with CPU capacity:
130 * util * margin < capacity * 1024
134 unsigned int capacity_margin
= 1280;
136 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
142 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
148 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
155 * Increase the granularity value when there are more CPUs,
156 * because with more CPUs the 'effective latency' as visible
157 * to users decreases. But the relationship is not linear,
158 * so pick a second-best guess by going with the log2 of the
161 * This idea comes from the SD scheduler of Con Kolivas:
163 static unsigned int get_update_sysctl_factor(void)
165 unsigned int cpus
= min_t(unsigned int, num_online_cpus(), 8);
168 switch (sysctl_sched_tunable_scaling
) {
169 case SCHED_TUNABLESCALING_NONE
:
172 case SCHED_TUNABLESCALING_LINEAR
:
175 case SCHED_TUNABLESCALING_LOG
:
177 factor
= 1 + ilog2(cpus
);
184 static void update_sysctl(void)
186 unsigned int factor
= get_update_sysctl_factor();
188 #define SET_SYSCTL(name) \
189 (sysctl_##name = (factor) * normalized_sysctl_##name)
190 SET_SYSCTL(sched_min_granularity
);
191 SET_SYSCTL(sched_latency
);
192 SET_SYSCTL(sched_wakeup_granularity
);
196 void sched_init_granularity(void)
201 #define WMULT_CONST (~0U)
202 #define WMULT_SHIFT 32
204 static void __update_inv_weight(struct load_weight
*lw
)
208 if (likely(lw
->inv_weight
))
211 w
= scale_load_down(lw
->weight
);
213 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
215 else if (unlikely(!w
))
216 lw
->inv_weight
= WMULT_CONST
;
218 lw
->inv_weight
= WMULT_CONST
/ w
;
222 * delta_exec * weight / lw.weight
224 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
226 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
227 * we're guaranteed shift stays positive because inv_weight is guaranteed to
228 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
230 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
231 * weight/lw.weight <= 1, and therefore our shift will also be positive.
233 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
235 u64 fact
= scale_load_down(weight
);
236 int shift
= WMULT_SHIFT
;
238 __update_inv_weight(lw
);
240 if (unlikely(fact
>> 32)) {
247 /* hint to use a 32x32->64 mul */
248 fact
= (u64
)(u32
)fact
* lw
->inv_weight
;
255 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
259 const struct sched_class fair_sched_class
;
261 /**************************************************************
262 * CFS operations on generic schedulable entities:
265 #ifdef CONFIG_FAIR_GROUP_SCHED
267 /* cpu runqueue to which this cfs_rq is attached */
268 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
273 /* An entity is a task if it doesn't "own" a runqueue */
274 #define entity_is_task(se) (!se->my_q)
276 static inline struct task_struct
*task_of(struct sched_entity
*se
)
278 SCHED_WARN_ON(!entity_is_task(se
));
279 return container_of(se
, struct task_struct
, se
);
282 /* Walk up scheduling entities hierarchy */
283 #define for_each_sched_entity(se) \
284 for (; se; se = se->parent)
286 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
291 /* runqueue on which this entity is (to be) queued */
292 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
297 /* runqueue "owned" by this group */
298 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
303 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
305 struct rq
*rq
= rq_of(cfs_rq
);
306 int cpu
= cpu_of(rq
);
314 * Ensure we either appear before our parent (if already
315 * enqueued) or force our parent to appear after us when it is
316 * enqueued. The fact that we always enqueue bottom-up
317 * reduces this to two cases and a special case for the root
318 * cfs_rq. Furthermore, it also means that we will always reset
319 * tmp_alone_branch either when the branch is connected
320 * to a tree or when we reach the top of the tree
322 if (cfs_rq
->tg
->parent
&&
323 cfs_rq
->tg
->parent
->cfs_rq
[cpu
]->on_list
) {
325 * If parent is already on the list, we add the child
326 * just before. Thanks to circular linked property of
327 * the list, this means to put the child at the tail
328 * of the list that starts by parent.
330 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
331 &(cfs_rq
->tg
->parent
->cfs_rq
[cpu
]->leaf_cfs_rq_list
));
333 * The branch is now connected to its tree so we can
334 * reset tmp_alone_branch to the beginning of the
337 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
341 if (!cfs_rq
->tg
->parent
) {
343 * cfs rq without parent should be put
344 * at the tail of the list.
346 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
347 &rq
->leaf_cfs_rq_list
);
349 * We have reach the top of a tree so we can reset
350 * tmp_alone_branch to the beginning of the list.
352 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
357 * The parent has not already been added so we want to
358 * make sure that it will be put after us.
359 * tmp_alone_branch points to the begin of the branch
360 * where we will add parent.
362 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, rq
->tmp_alone_branch
);
364 * update tmp_alone_branch to points to the new begin
367 rq
->tmp_alone_branch
= &cfs_rq
->leaf_cfs_rq_list
;
370 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
372 if (cfs_rq
->on_list
) {
373 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
378 static inline void assert_list_leaf_cfs_rq(struct rq
*rq
)
380 SCHED_WARN_ON(rq
->tmp_alone_branch
!= &rq
->leaf_cfs_rq_list
);
383 /* Iterate through all cfs_rq's on a runqueue in bottom-up order */
384 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
385 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
387 /* Do the two (enqueued) entities belong to the same group ? */
388 static inline struct cfs_rq
*
389 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
391 if (se
->cfs_rq
== pse
->cfs_rq
)
397 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
403 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
405 int se_depth
, pse_depth
;
408 * preemption test can be made between sibling entities who are in the
409 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
410 * both tasks until we find their ancestors who are siblings of common
414 /* First walk up until both entities are at same depth */
415 se_depth
= (*se
)->depth
;
416 pse_depth
= (*pse
)->depth
;
418 while (se_depth
> pse_depth
) {
420 *se
= parent_entity(*se
);
423 while (pse_depth
> se_depth
) {
425 *pse
= parent_entity(*pse
);
428 while (!is_same_group(*se
, *pse
)) {
429 *se
= parent_entity(*se
);
430 *pse
= parent_entity(*pse
);
434 #else /* !CONFIG_FAIR_GROUP_SCHED */
436 static inline struct task_struct
*task_of(struct sched_entity
*se
)
438 return container_of(se
, struct task_struct
, se
);
441 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
443 return container_of(cfs_rq
, struct rq
, cfs
);
446 #define entity_is_task(se) 1
448 #define for_each_sched_entity(se) \
449 for (; se; se = NULL)
451 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
453 return &task_rq(p
)->cfs
;
456 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
458 struct task_struct
*p
= task_of(se
);
459 struct rq
*rq
= task_rq(p
);
464 /* runqueue "owned" by this group */
465 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
470 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
474 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
478 static inline void assert_list_leaf_cfs_rq(struct rq
*rq
)
482 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
483 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
485 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
491 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
495 #endif /* CONFIG_FAIR_GROUP_SCHED */
497 static __always_inline
498 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
500 /**************************************************************
501 * Scheduling class tree data structure manipulation methods:
504 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
506 s64 delta
= (s64
)(vruntime
- max_vruntime
);
508 max_vruntime
= vruntime
;
513 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
515 s64 delta
= (s64
)(vruntime
- min_vruntime
);
517 min_vruntime
= vruntime
;
522 static inline int entity_before(struct sched_entity
*a
,
523 struct sched_entity
*b
)
525 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
528 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
530 struct sched_entity
*curr
= cfs_rq
->curr
;
531 struct rb_node
*leftmost
= rb_first_cached(&cfs_rq
->tasks_timeline
);
533 u64 vruntime
= cfs_rq
->min_vruntime
;
537 vruntime
= curr
->vruntime
;
542 if (leftmost
) { /* non-empty tree */
543 struct sched_entity
*se
;
544 se
= rb_entry(leftmost
, struct sched_entity
, run_node
);
547 vruntime
= se
->vruntime
;
549 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
552 /* ensure we never gain time by being placed backwards. */
553 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
556 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
561 * Enqueue an entity into the rb-tree:
563 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
565 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_root
.rb_node
;
566 struct rb_node
*parent
= NULL
;
567 struct sched_entity
*entry
;
568 bool leftmost
= true;
571 * Find the right place in the rbtree:
575 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
577 * We dont care about collisions. Nodes with
578 * the same key stay together.
580 if (entity_before(se
, entry
)) {
581 link
= &parent
->rb_left
;
583 link
= &parent
->rb_right
;
588 rb_link_node(&se
->run_node
, parent
, link
);
589 rb_insert_color_cached(&se
->run_node
,
590 &cfs_rq
->tasks_timeline
, leftmost
);
593 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
595 rb_erase_cached(&se
->run_node
, &cfs_rq
->tasks_timeline
);
598 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
600 struct rb_node
*left
= rb_first_cached(&cfs_rq
->tasks_timeline
);
605 return rb_entry(left
, struct sched_entity
, run_node
);
608 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
610 struct rb_node
*next
= rb_next(&se
->run_node
);
615 return rb_entry(next
, struct sched_entity
, run_node
);
618 #ifdef CONFIG_SCHED_DEBUG
619 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
621 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
.rb_root
);
626 return rb_entry(last
, struct sched_entity
, run_node
);
629 /**************************************************************
630 * Scheduling class statistics methods:
633 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
634 void __user
*buffer
, size_t *lenp
,
637 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
638 unsigned int factor
= get_update_sysctl_factor();
643 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
644 sysctl_sched_min_granularity
);
646 #define WRT_SYSCTL(name) \
647 (normalized_sysctl_##name = sysctl_##name / (factor))
648 WRT_SYSCTL(sched_min_granularity
);
649 WRT_SYSCTL(sched_latency
);
650 WRT_SYSCTL(sched_wakeup_granularity
);
660 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
662 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
663 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
669 * The idea is to set a period in which each task runs once.
671 * When there are too many tasks (sched_nr_latency) we have to stretch
672 * this period because otherwise the slices get too small.
674 * p = (nr <= nl) ? l : l*nr/nl
676 static u64
__sched_period(unsigned long nr_running
)
678 if (unlikely(nr_running
> sched_nr_latency
))
679 return nr_running
* sysctl_sched_min_granularity
;
681 return sysctl_sched_latency
;
685 * We calculate the wall-time slice from the period by taking a part
686 * proportional to the weight.
690 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
692 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
694 for_each_sched_entity(se
) {
695 struct load_weight
*load
;
696 struct load_weight lw
;
698 cfs_rq
= cfs_rq_of(se
);
699 load
= &cfs_rq
->load
;
701 if (unlikely(!se
->on_rq
)) {
704 update_load_add(&lw
, se
->load
.weight
);
707 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
713 * We calculate the vruntime slice of a to-be-inserted task.
717 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
719 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
724 #include "sched-pelt.h"
726 static int select_idle_sibling(struct task_struct
*p
, int prev_cpu
, int cpu
);
727 static unsigned long task_h_load(struct task_struct
*p
);
729 /* Give new sched_entity start runnable values to heavy its load in infant time */
730 void init_entity_runnable_average(struct sched_entity
*se
)
732 struct sched_avg
*sa
= &se
->avg
;
734 memset(sa
, 0, sizeof(*sa
));
737 * Tasks are intialized with full load to be seen as heavy tasks until
738 * they get a chance to stabilize to their real load level.
739 * Group entities are intialized with zero load to reflect the fact that
740 * nothing has been attached to the task group yet.
742 if (entity_is_task(se
))
743 sa
->runnable_load_avg
= sa
->load_avg
= scale_load_down(se
->load
.weight
);
745 se
->runnable_weight
= se
->load
.weight
;
747 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
750 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
751 static void attach_entity_cfs_rq(struct sched_entity
*se
);
754 * With new tasks being created, their initial util_avgs are extrapolated
755 * based on the cfs_rq's current util_avg:
757 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
759 * However, in many cases, the above util_avg does not give a desired
760 * value. Moreover, the sum of the util_avgs may be divergent, such
761 * as when the series is a harmonic series.
763 * To solve this problem, we also cap the util_avg of successive tasks to
764 * only 1/2 of the left utilization budget:
766 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
768 * where n denotes the nth task and cpu_scale the CPU capacity.
770 * For example, for a CPU with 1024 of capacity, a simplest series from
771 * the beginning would be like:
773 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
774 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
776 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
777 * if util_avg > util_avg_cap.
779 void post_init_entity_util_avg(struct sched_entity
*se
)
781 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
782 struct sched_avg
*sa
= &se
->avg
;
783 long cpu_scale
= arch_scale_cpu_capacity(NULL
, cpu_of(rq_of(cfs_rq
)));
784 long cap
= (long)(cpu_scale
- cfs_rq
->avg
.util_avg
) / 2;
787 if (cfs_rq
->avg
.util_avg
!= 0) {
788 sa
->util_avg
= cfs_rq
->avg
.util_avg
* se
->load
.weight
;
789 sa
->util_avg
/= (cfs_rq
->avg
.load_avg
+ 1);
791 if (sa
->util_avg
> cap
)
798 if (entity_is_task(se
)) {
799 struct task_struct
*p
= task_of(se
);
800 if (p
->sched_class
!= &fair_sched_class
) {
802 * For !fair tasks do:
804 update_cfs_rq_load_avg(now, cfs_rq);
805 attach_entity_load_avg(cfs_rq, se);
806 switched_from_fair(rq, p);
808 * such that the next switched_to_fair() has the
811 se
->avg
.last_update_time
= cfs_rq_clock_task(cfs_rq
);
816 attach_entity_cfs_rq(se
);
819 #else /* !CONFIG_SMP */
820 void init_entity_runnable_average(struct sched_entity
*se
)
823 void post_init_entity_util_avg(struct sched_entity
*se
)
826 static void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
)
829 #endif /* CONFIG_SMP */
832 * Update the current task's runtime statistics.
834 static void update_curr(struct cfs_rq
*cfs_rq
)
836 struct sched_entity
*curr
= cfs_rq
->curr
;
837 u64 now
= rq_clock_task(rq_of(cfs_rq
));
843 delta_exec
= now
- curr
->exec_start
;
844 if (unlikely((s64
)delta_exec
<= 0))
847 curr
->exec_start
= now
;
849 schedstat_set(curr
->statistics
.exec_max
,
850 max(delta_exec
, curr
->statistics
.exec_max
));
852 curr
->sum_exec_runtime
+= delta_exec
;
853 schedstat_add(cfs_rq
->exec_clock
, delta_exec
);
855 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
856 update_min_vruntime(cfs_rq
);
858 if (entity_is_task(curr
)) {
859 struct task_struct
*curtask
= task_of(curr
);
861 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
862 cgroup_account_cputime(curtask
, delta_exec
);
863 account_group_exec_runtime(curtask
, delta_exec
);
866 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
869 static void update_curr_fair(struct rq
*rq
)
871 update_curr(cfs_rq_of(&rq
->curr
->se
));
875 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
877 u64 wait_start
, prev_wait_start
;
879 if (!schedstat_enabled())
882 wait_start
= rq_clock(rq_of(cfs_rq
));
883 prev_wait_start
= schedstat_val(se
->statistics
.wait_start
);
885 if (entity_is_task(se
) && task_on_rq_migrating(task_of(se
)) &&
886 likely(wait_start
> prev_wait_start
))
887 wait_start
-= prev_wait_start
;
889 schedstat_set(se
->statistics
.wait_start
, wait_start
);
893 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
895 struct task_struct
*p
;
898 if (!schedstat_enabled())
901 delta
= rq_clock(rq_of(cfs_rq
)) - schedstat_val(se
->statistics
.wait_start
);
903 if (entity_is_task(se
)) {
905 if (task_on_rq_migrating(p
)) {
907 * Preserve migrating task's wait time so wait_start
908 * time stamp can be adjusted to accumulate wait time
909 * prior to migration.
911 schedstat_set(se
->statistics
.wait_start
, delta
);
914 trace_sched_stat_wait(p
, delta
);
917 schedstat_set(se
->statistics
.wait_max
,
918 max(schedstat_val(se
->statistics
.wait_max
), delta
));
919 schedstat_inc(se
->statistics
.wait_count
);
920 schedstat_add(se
->statistics
.wait_sum
, delta
);
921 schedstat_set(se
->statistics
.wait_start
, 0);
925 update_stats_enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
927 struct task_struct
*tsk
= NULL
;
928 u64 sleep_start
, block_start
;
930 if (!schedstat_enabled())
933 sleep_start
= schedstat_val(se
->statistics
.sleep_start
);
934 block_start
= schedstat_val(se
->statistics
.block_start
);
936 if (entity_is_task(se
))
940 u64 delta
= rq_clock(rq_of(cfs_rq
)) - sleep_start
;
945 if (unlikely(delta
> schedstat_val(se
->statistics
.sleep_max
)))
946 schedstat_set(se
->statistics
.sleep_max
, delta
);
948 schedstat_set(se
->statistics
.sleep_start
, 0);
949 schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
952 account_scheduler_latency(tsk
, delta
>> 10, 1);
953 trace_sched_stat_sleep(tsk
, delta
);
957 u64 delta
= rq_clock(rq_of(cfs_rq
)) - block_start
;
962 if (unlikely(delta
> schedstat_val(se
->statistics
.block_max
)))
963 schedstat_set(se
->statistics
.block_max
, delta
);
965 schedstat_set(se
->statistics
.block_start
, 0);
966 schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
969 if (tsk
->in_iowait
) {
970 schedstat_add(se
->statistics
.iowait_sum
, delta
);
971 schedstat_inc(se
->statistics
.iowait_count
);
972 trace_sched_stat_iowait(tsk
, delta
);
975 trace_sched_stat_blocked(tsk
, delta
);
978 * Blocking time is in units of nanosecs, so shift by
979 * 20 to get a milliseconds-range estimation of the
980 * amount of time that the task spent sleeping:
982 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
983 profile_hits(SLEEP_PROFILING
,
984 (void *)get_wchan(tsk
),
987 account_scheduler_latency(tsk
, delta
>> 10, 0);
993 * Task is being enqueued - update stats:
996 update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
998 if (!schedstat_enabled())
1002 * Are we enqueueing a waiting task? (for current tasks
1003 * a dequeue/enqueue event is a NOP)
1005 if (se
!= cfs_rq
->curr
)
1006 update_stats_wait_start(cfs_rq
, se
);
1008 if (flags
& ENQUEUE_WAKEUP
)
1009 update_stats_enqueue_sleeper(cfs_rq
, se
);
1013 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1016 if (!schedstat_enabled())
1020 * Mark the end of the wait period if dequeueing a
1023 if (se
!= cfs_rq
->curr
)
1024 update_stats_wait_end(cfs_rq
, se
);
1026 if ((flags
& DEQUEUE_SLEEP
) && entity_is_task(se
)) {
1027 struct task_struct
*tsk
= task_of(se
);
1029 if (tsk
->state
& TASK_INTERRUPTIBLE
)
1030 schedstat_set(se
->statistics
.sleep_start
,
1031 rq_clock(rq_of(cfs_rq
)));
1032 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
1033 schedstat_set(se
->statistics
.block_start
,
1034 rq_clock(rq_of(cfs_rq
)));
1039 * We are picking a new current task - update its stats:
1042 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1045 * We are starting a new run period:
1047 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
1050 /**************************************************
1051 * Scheduling class queueing methods:
1054 #ifdef CONFIG_NUMA_BALANCING
1056 * Approximate time to scan a full NUMA task in ms. The task scan period is
1057 * calculated based on the tasks virtual memory size and
1058 * numa_balancing_scan_size.
1060 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
1061 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
1063 /* Portion of address space to scan in MB */
1064 unsigned int sysctl_numa_balancing_scan_size
= 256;
1066 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1067 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
1072 spinlock_t lock
; /* nr_tasks, tasks */
1077 struct rcu_head rcu
;
1078 unsigned long total_faults
;
1079 unsigned long max_faults_cpu
;
1081 * Faults_cpu is used to decide whether memory should move
1082 * towards the CPU. As a consequence, these stats are weighted
1083 * more by CPU use than by memory faults.
1085 unsigned long *faults_cpu
;
1086 unsigned long faults
[0];
1089 static inline unsigned long group_faults_priv(struct numa_group
*ng
);
1090 static inline unsigned long group_faults_shared(struct numa_group
*ng
);
1092 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
1094 unsigned long rss
= 0;
1095 unsigned long nr_scan_pages
;
1098 * Calculations based on RSS as non-present and empty pages are skipped
1099 * by the PTE scanner and NUMA hinting faults should be trapped based
1102 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
1103 rss
= get_mm_rss(p
->mm
);
1105 rss
= nr_scan_pages
;
1107 rss
= round_up(rss
, nr_scan_pages
);
1108 return rss
/ nr_scan_pages
;
1111 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1112 #define MAX_SCAN_WINDOW 2560
1114 static unsigned int task_scan_min(struct task_struct
*p
)
1116 unsigned int scan_size
= READ_ONCE(sysctl_numa_balancing_scan_size
);
1117 unsigned int scan
, floor
;
1118 unsigned int windows
= 1;
1120 if (scan_size
< MAX_SCAN_WINDOW
)
1121 windows
= MAX_SCAN_WINDOW
/ scan_size
;
1122 floor
= 1000 / windows
;
1124 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
1125 return max_t(unsigned int, floor
, scan
);
1128 static unsigned int task_scan_start(struct task_struct
*p
)
1130 unsigned long smin
= task_scan_min(p
);
1131 unsigned long period
= smin
;
1133 /* Scale the maximum scan period with the amount of shared memory. */
1134 if (p
->numa_group
) {
1135 struct numa_group
*ng
= p
->numa_group
;
1136 unsigned long shared
= group_faults_shared(ng
);
1137 unsigned long private = group_faults_priv(ng
);
1139 period
*= atomic_read(&ng
->refcount
);
1140 period
*= shared
+ 1;
1141 period
/= private + shared
+ 1;
1144 return max(smin
, period
);
1147 static unsigned int task_scan_max(struct task_struct
*p
)
1149 unsigned long smin
= task_scan_min(p
);
1152 /* Watch for min being lower than max due to floor calculations */
1153 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
1155 /* Scale the maximum scan period with the amount of shared memory. */
1156 if (p
->numa_group
) {
1157 struct numa_group
*ng
= p
->numa_group
;
1158 unsigned long shared
= group_faults_shared(ng
);
1159 unsigned long private = group_faults_priv(ng
);
1160 unsigned long period
= smax
;
1162 period
*= atomic_read(&ng
->refcount
);
1163 period
*= shared
+ 1;
1164 period
/= private + shared
+ 1;
1166 smax
= max(smax
, period
);
1169 return max(smin
, smax
);
1172 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
1174 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
1175 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
1178 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
1180 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
1181 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
1184 /* Shared or private faults. */
1185 #define NR_NUMA_HINT_FAULT_TYPES 2
1187 /* Memory and CPU locality */
1188 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1190 /* Averaged statistics, and temporary buffers. */
1191 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1193 pid_t
task_numa_group_id(struct task_struct
*p
)
1195 return p
->numa_group
? p
->numa_group
->gid
: 0;
1199 * The averaged statistics, shared & private, memory & cpu,
1200 * occupy the first half of the array. The second half of the
1201 * array is for current counters, which are averaged into the
1202 * first set by task_numa_placement.
1204 static inline int task_faults_idx(enum numa_faults_stats s
, int nid
, int priv
)
1206 return NR_NUMA_HINT_FAULT_TYPES
* (s
* nr_node_ids
+ nid
) + priv
;
1209 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
1211 if (!p
->numa_faults
)
1214 return p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1215 p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1218 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
1223 return p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1224 p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1227 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
1229 return group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1230 group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1233 static inline unsigned long group_faults_priv(struct numa_group
*ng
)
1235 unsigned long faults
= 0;
1238 for_each_online_node(node
) {
1239 faults
+= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
1245 static inline unsigned long group_faults_shared(struct numa_group
*ng
)
1247 unsigned long faults
= 0;
1250 for_each_online_node(node
) {
1251 faults
+= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
1258 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1259 * considered part of a numa group's pseudo-interleaving set. Migrations
1260 * between these nodes are slowed down, to allow things to settle down.
1262 #define ACTIVE_NODE_FRACTION 3
1264 static bool numa_is_active_node(int nid
, struct numa_group
*ng
)
1266 return group_faults_cpu(ng
, nid
) * ACTIVE_NODE_FRACTION
> ng
->max_faults_cpu
;
1269 /* Handle placement on systems where not all nodes are directly connected. */
1270 static unsigned long score_nearby_nodes(struct task_struct
*p
, int nid
,
1271 int maxdist
, bool task
)
1273 unsigned long score
= 0;
1277 * All nodes are directly connected, and the same distance
1278 * from each other. No need for fancy placement algorithms.
1280 if (sched_numa_topology_type
== NUMA_DIRECT
)
1284 * This code is called for each node, introducing N^2 complexity,
1285 * which should be ok given the number of nodes rarely exceeds 8.
1287 for_each_online_node(node
) {
1288 unsigned long faults
;
1289 int dist
= node_distance(nid
, node
);
1292 * The furthest away nodes in the system are not interesting
1293 * for placement; nid was already counted.
1295 if (dist
== sched_max_numa_distance
|| node
== nid
)
1299 * On systems with a backplane NUMA topology, compare groups
1300 * of nodes, and move tasks towards the group with the most
1301 * memory accesses. When comparing two nodes at distance
1302 * "hoplimit", only nodes closer by than "hoplimit" are part
1303 * of each group. Skip other nodes.
1305 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1309 /* Add up the faults from nearby nodes. */
1311 faults
= task_faults(p
, node
);
1313 faults
= group_faults(p
, node
);
1316 * On systems with a glueless mesh NUMA topology, there are
1317 * no fixed "groups of nodes". Instead, nodes that are not
1318 * directly connected bounce traffic through intermediate
1319 * nodes; a numa_group can occupy any set of nodes.
1320 * The further away a node is, the less the faults count.
1321 * This seems to result in good task placement.
1323 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1324 faults
*= (sched_max_numa_distance
- dist
);
1325 faults
/= (sched_max_numa_distance
- LOCAL_DISTANCE
);
1335 * These return the fraction of accesses done by a particular task, or
1336 * task group, on a particular numa node. The group weight is given a
1337 * larger multiplier, in order to group tasks together that are almost
1338 * evenly spread out between numa nodes.
1340 static inline unsigned long task_weight(struct task_struct
*p
, int nid
,
1343 unsigned long faults
, total_faults
;
1345 if (!p
->numa_faults
)
1348 total_faults
= p
->total_numa_faults
;
1353 faults
= task_faults(p
, nid
);
1354 faults
+= score_nearby_nodes(p
, nid
, dist
, true);
1356 return 1000 * faults
/ total_faults
;
1359 static inline unsigned long group_weight(struct task_struct
*p
, int nid
,
1362 unsigned long faults
, total_faults
;
1367 total_faults
= p
->numa_group
->total_faults
;
1372 faults
= group_faults(p
, nid
);
1373 faults
+= score_nearby_nodes(p
, nid
, dist
, false);
1375 return 1000 * faults
/ total_faults
;
1378 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
1379 int src_nid
, int dst_cpu
)
1381 struct numa_group
*ng
= p
->numa_group
;
1382 int dst_nid
= cpu_to_node(dst_cpu
);
1383 int last_cpupid
, this_cpupid
;
1385 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
1388 * Multi-stage node selection is used in conjunction with a periodic
1389 * migration fault to build a temporal task<->page relation. By using
1390 * a two-stage filter we remove short/unlikely relations.
1392 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1393 * a task's usage of a particular page (n_p) per total usage of this
1394 * page (n_t) (in a given time-span) to a probability.
1396 * Our periodic faults will sample this probability and getting the
1397 * same result twice in a row, given these samples are fully
1398 * independent, is then given by P(n)^2, provided our sample period
1399 * is sufficiently short compared to the usage pattern.
1401 * This quadric squishes small probabilities, making it less likely we
1402 * act on an unlikely task<->page relation.
1404 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
1405 if (!cpupid_pid_unset(last_cpupid
) &&
1406 cpupid_to_nid(last_cpupid
) != dst_nid
)
1409 /* Always allow migrate on private faults */
1410 if (cpupid_match_pid(p
, last_cpupid
))
1413 /* A shared fault, but p->numa_group has not been set up yet. */
1418 * Destination node is much more heavily used than the source
1419 * node? Allow migration.
1421 if (group_faults_cpu(ng
, dst_nid
) > group_faults_cpu(ng
, src_nid
) *
1422 ACTIVE_NODE_FRACTION
)
1426 * Distribute memory according to CPU & memory use on each node,
1427 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1429 * faults_cpu(dst) 3 faults_cpu(src)
1430 * --------------- * - > ---------------
1431 * faults_mem(dst) 4 faults_mem(src)
1433 return group_faults_cpu(ng
, dst_nid
) * group_faults(p
, src_nid
) * 3 >
1434 group_faults_cpu(ng
, src_nid
) * group_faults(p
, dst_nid
) * 4;
1437 static unsigned long weighted_cpuload(struct rq
*rq
);
1438 static unsigned long source_load(int cpu
, int type
);
1439 static unsigned long target_load(int cpu
, int type
);
1440 static unsigned long capacity_of(int cpu
);
1442 /* Cached statistics for all CPUs within a node */
1444 unsigned long nr_running
;
1447 /* Total compute capacity of CPUs on a node */
1448 unsigned long compute_capacity
;
1450 /* Approximate capacity in terms of runnable tasks on a node */
1451 unsigned long task_capacity
;
1452 int has_free_capacity
;
1456 * XXX borrowed from update_sg_lb_stats
1458 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1460 int smt
, cpu
, cpus
= 0;
1461 unsigned long capacity
;
1463 memset(ns
, 0, sizeof(*ns
));
1464 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1465 struct rq
*rq
= cpu_rq(cpu
);
1467 ns
->nr_running
+= rq
->nr_running
;
1468 ns
->load
+= weighted_cpuload(rq
);
1469 ns
->compute_capacity
+= capacity_of(cpu
);
1475 * If we raced with hotplug and there are no CPUs left in our mask
1476 * the @ns structure is NULL'ed and task_numa_compare() will
1477 * not find this node attractive.
1479 * We'll either bail at !has_free_capacity, or we'll detect a huge
1480 * imbalance and bail there.
1485 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1486 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, ns
->compute_capacity
);
1487 capacity
= cpus
/ smt
; /* cores */
1489 ns
->task_capacity
= min_t(unsigned, capacity
,
1490 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
));
1491 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1494 struct task_numa_env
{
1495 struct task_struct
*p
;
1497 int src_cpu
, src_nid
;
1498 int dst_cpu
, dst_nid
;
1500 struct numa_stats src_stats
, dst_stats
;
1505 struct task_struct
*best_task
;
1510 static void task_numa_assign(struct task_numa_env
*env
,
1511 struct task_struct
*p
, long imp
)
1514 put_task_struct(env
->best_task
);
1519 env
->best_imp
= imp
;
1520 env
->best_cpu
= env
->dst_cpu
;
1523 static bool load_too_imbalanced(long src_load
, long dst_load
,
1524 struct task_numa_env
*env
)
1527 long orig_src_load
, orig_dst_load
;
1528 long src_capacity
, dst_capacity
;
1531 * The load is corrected for the CPU capacity available on each node.
1534 * ------------ vs ---------
1535 * src_capacity dst_capacity
1537 src_capacity
= env
->src_stats
.compute_capacity
;
1538 dst_capacity
= env
->dst_stats
.compute_capacity
;
1540 /* We care about the slope of the imbalance, not the direction. */
1541 if (dst_load
< src_load
)
1542 swap(dst_load
, src_load
);
1544 /* Is the difference below the threshold? */
1545 imb
= dst_load
* src_capacity
* 100 -
1546 src_load
* dst_capacity
* env
->imbalance_pct
;
1551 * The imbalance is above the allowed threshold.
1552 * Compare it with the old imbalance.
1554 orig_src_load
= env
->src_stats
.load
;
1555 orig_dst_load
= env
->dst_stats
.load
;
1557 if (orig_dst_load
< orig_src_load
)
1558 swap(orig_dst_load
, orig_src_load
);
1560 old_imb
= orig_dst_load
* src_capacity
* 100 -
1561 orig_src_load
* dst_capacity
* env
->imbalance_pct
;
1563 /* Would this change make things worse? */
1564 return (imb
> old_imb
);
1568 * This checks if the overall compute and NUMA accesses of the system would
1569 * be improved if the source tasks was migrated to the target dst_cpu taking
1570 * into account that it might be best if task running on the dst_cpu should
1571 * be exchanged with the source task
1573 static void task_numa_compare(struct task_numa_env
*env
,
1574 long taskimp
, long groupimp
)
1576 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1577 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1578 struct task_struct
*cur
;
1579 long src_load
, dst_load
;
1581 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1583 int dist
= env
->dist
;
1586 cur
= task_rcu_dereference(&dst_rq
->curr
);
1587 if (cur
&& ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
)))
1591 * Because we have preemption enabled we can get migrated around and
1592 * end try selecting ourselves (current == env->p) as a swap candidate.
1598 * "imp" is the fault differential for the source task between the
1599 * source and destination node. Calculate the total differential for
1600 * the source task and potential destination task. The more negative
1601 * the value is, the more rmeote accesses that would be expected to
1602 * be incurred if the tasks were swapped.
1605 /* Skip this swap candidate if cannot move to the source cpu */
1606 if (!cpumask_test_cpu(env
->src_cpu
, &cur
->cpus_allowed
))
1610 * If dst and source tasks are in the same NUMA group, or not
1611 * in any group then look only at task weights.
1613 if (cur
->numa_group
== env
->p
->numa_group
) {
1614 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1615 task_weight(cur
, env
->dst_nid
, dist
);
1617 * Add some hysteresis to prevent swapping the
1618 * tasks within a group over tiny differences.
1620 if (cur
->numa_group
)
1624 * Compare the group weights. If a task is all by
1625 * itself (not part of a group), use the task weight
1628 if (cur
->numa_group
)
1629 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1630 group_weight(cur
, env
->dst_nid
, dist
);
1632 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1633 task_weight(cur
, env
->dst_nid
, dist
);
1637 if (imp
<= env
->best_imp
&& moveimp
<= env
->best_imp
)
1641 /* Is there capacity at our destination? */
1642 if (env
->src_stats
.nr_running
<= env
->src_stats
.task_capacity
&&
1643 !env
->dst_stats
.has_free_capacity
)
1649 /* Balance doesn't matter much if we're running a task per cpu */
1650 if (imp
> env
->best_imp
&& src_rq
->nr_running
== 1 &&
1651 dst_rq
->nr_running
== 1)
1655 * In the overloaded case, try and keep the load balanced.
1658 load
= task_h_load(env
->p
);
1659 dst_load
= env
->dst_stats
.load
+ load
;
1660 src_load
= env
->src_stats
.load
- load
;
1662 if (moveimp
> imp
&& moveimp
> env
->best_imp
) {
1664 * If the improvement from just moving env->p direction is
1665 * better than swapping tasks around, check if a move is
1666 * possible. Store a slightly smaller score than moveimp,
1667 * so an actually idle CPU will win.
1669 if (!load_too_imbalanced(src_load
, dst_load
, env
)) {
1676 if (imp
<= env
->best_imp
)
1680 load
= task_h_load(cur
);
1685 if (load_too_imbalanced(src_load
, dst_load
, env
))
1689 * One idle CPU per node is evaluated for a task numa move.
1690 * Call select_idle_sibling to maybe find a better one.
1694 * select_idle_siblings() uses an per-cpu cpumask that
1695 * can be used from IRQ context.
1697 local_irq_disable();
1698 env
->dst_cpu
= select_idle_sibling(env
->p
, env
->src_cpu
,
1704 task_numa_assign(env
, cur
, imp
);
1709 static void task_numa_find_cpu(struct task_numa_env
*env
,
1710 long taskimp
, long groupimp
)
1714 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1715 /* Skip this CPU if the source task cannot migrate */
1716 if (!cpumask_test_cpu(cpu
, &env
->p
->cpus_allowed
))
1720 task_numa_compare(env
, taskimp
, groupimp
);
1724 /* Only move tasks to a NUMA node less busy than the current node. */
1725 static bool numa_has_capacity(struct task_numa_env
*env
)
1727 struct numa_stats
*src
= &env
->src_stats
;
1728 struct numa_stats
*dst
= &env
->dst_stats
;
1730 if (src
->has_free_capacity
&& !dst
->has_free_capacity
)
1734 * Only consider a task move if the source has a higher load
1735 * than the destination, corrected for CPU capacity on each node.
1737 * src->load dst->load
1738 * --------------------- vs ---------------------
1739 * src->compute_capacity dst->compute_capacity
1741 if (src
->load
* dst
->compute_capacity
* env
->imbalance_pct
>
1743 dst
->load
* src
->compute_capacity
* 100)
1749 static int task_numa_migrate(struct task_struct
*p
)
1751 struct task_numa_env env
= {
1754 .src_cpu
= task_cpu(p
),
1755 .src_nid
= task_node(p
),
1757 .imbalance_pct
= 112,
1763 struct sched_domain
*sd
;
1764 unsigned long taskweight
, groupweight
;
1766 long taskimp
, groupimp
;
1769 * Pick the lowest SD_NUMA domain, as that would have the smallest
1770 * imbalance and would be the first to start moving tasks about.
1772 * And we want to avoid any moving of tasks about, as that would create
1773 * random movement of tasks -- counter the numa conditions we're trying
1777 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1779 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1783 * Cpusets can break the scheduler domain tree into smaller
1784 * balance domains, some of which do not cross NUMA boundaries.
1785 * Tasks that are "trapped" in such domains cannot be migrated
1786 * elsewhere, so there is no point in (re)trying.
1788 if (unlikely(!sd
)) {
1789 p
->numa_preferred_nid
= task_node(p
);
1793 env
.dst_nid
= p
->numa_preferred_nid
;
1794 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1795 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1796 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1797 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1798 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
1799 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
1800 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1802 /* Try to find a spot on the preferred nid. */
1803 if (numa_has_capacity(&env
))
1804 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1807 * Look at other nodes in these cases:
1808 * - there is no space available on the preferred_nid
1809 * - the task is part of a numa_group that is interleaved across
1810 * multiple NUMA nodes; in order to better consolidate the group,
1811 * we need to check other locations.
1813 if (env
.best_cpu
== -1 || (p
->numa_group
&& p
->numa_group
->active_nodes
> 1)) {
1814 for_each_online_node(nid
) {
1815 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1818 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1819 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1821 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1822 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1825 /* Only consider nodes where both task and groups benefit */
1826 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
1827 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
1828 if (taskimp
< 0 && groupimp
< 0)
1833 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1834 if (numa_has_capacity(&env
))
1835 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1840 * If the task is part of a workload that spans multiple NUMA nodes,
1841 * and is migrating into one of the workload's active nodes, remember
1842 * this node as the task's preferred numa node, so the workload can
1844 * A task that migrated to a second choice node will be better off
1845 * trying for a better one later. Do not set the preferred node here.
1847 if (p
->numa_group
) {
1848 struct numa_group
*ng
= p
->numa_group
;
1850 if (env
.best_cpu
== -1)
1855 if (ng
->active_nodes
> 1 && numa_is_active_node(env
.dst_nid
, ng
))
1856 sched_setnuma(p
, env
.dst_nid
);
1859 /* No better CPU than the current one was found. */
1860 if (env
.best_cpu
== -1)
1864 * Reset the scan period if the task is being rescheduled on an
1865 * alternative node to recheck if the tasks is now properly placed.
1867 p
->numa_scan_period
= task_scan_start(p
);
1869 if (env
.best_task
== NULL
) {
1870 ret
= migrate_task_to(p
, env
.best_cpu
);
1872 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1876 ret
= migrate_swap(p
, env
.best_task
);
1878 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1879 put_task_struct(env
.best_task
);
1883 /* Attempt to migrate a task to a CPU on the preferred node. */
1884 static void numa_migrate_preferred(struct task_struct
*p
)
1886 unsigned long interval
= HZ
;
1888 /* This task has no NUMA fault statistics yet */
1889 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults
))
1892 /* Periodically retry migrating the task to the preferred node */
1893 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1894 p
->numa_migrate_retry
= jiffies
+ interval
;
1896 /* Success if task is already running on preferred CPU */
1897 if (task_node(p
) == p
->numa_preferred_nid
)
1900 /* Otherwise, try migrate to a CPU on the preferred node */
1901 task_numa_migrate(p
);
1905 * Find out how many nodes on the workload is actively running on. Do this by
1906 * tracking the nodes from which NUMA hinting faults are triggered. This can
1907 * be different from the set of nodes where the workload's memory is currently
1910 static void numa_group_count_active_nodes(struct numa_group
*numa_group
)
1912 unsigned long faults
, max_faults
= 0;
1913 int nid
, active_nodes
= 0;
1915 for_each_online_node(nid
) {
1916 faults
= group_faults_cpu(numa_group
, nid
);
1917 if (faults
> max_faults
)
1918 max_faults
= faults
;
1921 for_each_online_node(nid
) {
1922 faults
= group_faults_cpu(numa_group
, nid
);
1923 if (faults
* ACTIVE_NODE_FRACTION
> max_faults
)
1927 numa_group
->max_faults_cpu
= max_faults
;
1928 numa_group
->active_nodes
= active_nodes
;
1932 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1933 * increments. The more local the fault statistics are, the higher the scan
1934 * period will be for the next scan window. If local/(local+remote) ratio is
1935 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1936 * the scan period will decrease. Aim for 70% local accesses.
1938 #define NUMA_PERIOD_SLOTS 10
1939 #define NUMA_PERIOD_THRESHOLD 7
1942 * Increase the scan period (slow down scanning) if the majority of
1943 * our memory is already on our local node, or if the majority of
1944 * the page accesses are shared with other processes.
1945 * Otherwise, decrease the scan period.
1947 static void update_task_scan_period(struct task_struct
*p
,
1948 unsigned long shared
, unsigned long private)
1950 unsigned int period_slot
;
1951 int lr_ratio
, ps_ratio
;
1954 unsigned long remote
= p
->numa_faults_locality
[0];
1955 unsigned long local
= p
->numa_faults_locality
[1];
1958 * If there were no record hinting faults then either the task is
1959 * completely idle or all activity is areas that are not of interest
1960 * to automatic numa balancing. Related to that, if there were failed
1961 * migration then it implies we are migrating too quickly or the local
1962 * node is overloaded. In either case, scan slower
1964 if (local
+ shared
== 0 || p
->numa_faults_locality
[2]) {
1965 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1966 p
->numa_scan_period
<< 1);
1968 p
->mm
->numa_next_scan
= jiffies
+
1969 msecs_to_jiffies(p
->numa_scan_period
);
1975 * Prepare to scale scan period relative to the current period.
1976 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1977 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1978 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1980 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1981 lr_ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1982 ps_ratio
= (private * NUMA_PERIOD_SLOTS
) / (private + shared
);
1984 if (ps_ratio
>= NUMA_PERIOD_THRESHOLD
) {
1986 * Most memory accesses are local. There is no need to
1987 * do fast NUMA scanning, since memory is already local.
1989 int slot
= ps_ratio
- NUMA_PERIOD_THRESHOLD
;
1992 diff
= slot
* period_slot
;
1993 } else if (lr_ratio
>= NUMA_PERIOD_THRESHOLD
) {
1995 * Most memory accesses are shared with other tasks.
1996 * There is no point in continuing fast NUMA scanning,
1997 * since other tasks may just move the memory elsewhere.
1999 int slot
= lr_ratio
- NUMA_PERIOD_THRESHOLD
;
2002 diff
= slot
* period_slot
;
2005 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2006 * yet they are not on the local NUMA node. Speed up
2007 * NUMA scanning to get the memory moved over.
2009 int ratio
= max(lr_ratio
, ps_ratio
);
2010 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
2013 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
2014 task_scan_min(p
), task_scan_max(p
));
2015 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2019 * Get the fraction of time the task has been running since the last
2020 * NUMA placement cycle. The scheduler keeps similar statistics, but
2021 * decays those on a 32ms period, which is orders of magnitude off
2022 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2023 * stats only if the task is so new there are no NUMA statistics yet.
2025 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
2027 u64 runtime
, delta
, now
;
2028 /* Use the start of this time slice to avoid calculations. */
2029 now
= p
->se
.exec_start
;
2030 runtime
= p
->se
.sum_exec_runtime
;
2032 if (p
->last_task_numa_placement
) {
2033 delta
= runtime
- p
->last_sum_exec_runtime
;
2034 *period
= now
- p
->last_task_numa_placement
;
2036 /* Avoid time going backwards, prevent potential divide error: */
2037 if (unlikely((s64
)*period
< 0))
2040 delta
= p
->se
.avg
.load_sum
;
2041 *period
= LOAD_AVG_MAX
;
2044 p
->last_sum_exec_runtime
= runtime
;
2045 p
->last_task_numa_placement
= now
;
2051 * Determine the preferred nid for a task in a numa_group. This needs to
2052 * be done in a way that produces consistent results with group_weight,
2053 * otherwise workloads might not converge.
2055 static int preferred_group_nid(struct task_struct
*p
, int nid
)
2060 /* Direct connections between all NUMA nodes. */
2061 if (sched_numa_topology_type
== NUMA_DIRECT
)
2065 * On a system with glueless mesh NUMA topology, group_weight
2066 * scores nodes according to the number of NUMA hinting faults on
2067 * both the node itself, and on nearby nodes.
2069 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
2070 unsigned long score
, max_score
= 0;
2071 int node
, max_node
= nid
;
2073 dist
= sched_max_numa_distance
;
2075 for_each_online_node(node
) {
2076 score
= group_weight(p
, node
, dist
);
2077 if (score
> max_score
) {
2086 * Finding the preferred nid in a system with NUMA backplane
2087 * interconnect topology is more involved. The goal is to locate
2088 * tasks from numa_groups near each other in the system, and
2089 * untangle workloads from different sides of the system. This requires
2090 * searching down the hierarchy of node groups, recursively searching
2091 * inside the highest scoring group of nodes. The nodemask tricks
2092 * keep the complexity of the search down.
2094 nodes
= node_online_map
;
2095 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
2096 unsigned long max_faults
= 0;
2097 nodemask_t max_group
= NODE_MASK_NONE
;
2100 /* Are there nodes at this distance from each other? */
2101 if (!find_numa_distance(dist
))
2104 for_each_node_mask(a
, nodes
) {
2105 unsigned long faults
= 0;
2106 nodemask_t this_group
;
2107 nodes_clear(this_group
);
2109 /* Sum group's NUMA faults; includes a==b case. */
2110 for_each_node_mask(b
, nodes
) {
2111 if (node_distance(a
, b
) < dist
) {
2112 faults
+= group_faults(p
, b
);
2113 node_set(b
, this_group
);
2114 node_clear(b
, nodes
);
2118 /* Remember the top group. */
2119 if (faults
> max_faults
) {
2120 max_faults
= faults
;
2121 max_group
= this_group
;
2123 * subtle: at the smallest distance there is
2124 * just one node left in each "group", the
2125 * winner is the preferred nid.
2130 /* Next round, evaluate the nodes within max_group. */
2138 static void task_numa_placement(struct task_struct
*p
)
2140 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
2141 unsigned long max_faults
= 0, max_group_faults
= 0;
2142 unsigned long fault_types
[2] = { 0, 0 };
2143 unsigned long total_faults
;
2144 u64 runtime
, period
;
2145 spinlock_t
*group_lock
= NULL
;
2148 * The p->mm->numa_scan_seq field gets updated without
2149 * exclusive access. Use READ_ONCE() here to ensure
2150 * that the field is read in a single access:
2152 seq
= READ_ONCE(p
->mm
->numa_scan_seq
);
2153 if (p
->numa_scan_seq
== seq
)
2155 p
->numa_scan_seq
= seq
;
2156 p
->numa_scan_period_max
= task_scan_max(p
);
2158 total_faults
= p
->numa_faults_locality
[0] +
2159 p
->numa_faults_locality
[1];
2160 runtime
= numa_get_avg_runtime(p
, &period
);
2162 /* If the task is part of a group prevent parallel updates to group stats */
2163 if (p
->numa_group
) {
2164 group_lock
= &p
->numa_group
->lock
;
2165 spin_lock_irq(group_lock
);
2168 /* Find the node with the highest number of faults */
2169 for_each_online_node(nid
) {
2170 /* Keep track of the offsets in numa_faults array */
2171 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
2172 unsigned long faults
= 0, group_faults
= 0;
2175 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
2176 long diff
, f_diff
, f_weight
;
2178 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
2179 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
2180 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
2181 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
2183 /* Decay existing window, copy faults since last scan */
2184 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
2185 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
2186 p
->numa_faults
[membuf_idx
] = 0;
2189 * Normalize the faults_from, so all tasks in a group
2190 * count according to CPU use, instead of by the raw
2191 * number of faults. Tasks with little runtime have
2192 * little over-all impact on throughput, and thus their
2193 * faults are less important.
2195 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
2196 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
2198 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
2199 p
->numa_faults
[cpubuf_idx
] = 0;
2201 p
->numa_faults
[mem_idx
] += diff
;
2202 p
->numa_faults
[cpu_idx
] += f_diff
;
2203 faults
+= p
->numa_faults
[mem_idx
];
2204 p
->total_numa_faults
+= diff
;
2205 if (p
->numa_group
) {
2207 * safe because we can only change our own group
2209 * mem_idx represents the offset for a given
2210 * nid and priv in a specific region because it
2211 * is at the beginning of the numa_faults array.
2213 p
->numa_group
->faults
[mem_idx
] += diff
;
2214 p
->numa_group
->faults_cpu
[mem_idx
] += f_diff
;
2215 p
->numa_group
->total_faults
+= diff
;
2216 group_faults
+= p
->numa_group
->faults
[mem_idx
];
2220 if (faults
> max_faults
) {
2221 max_faults
= faults
;
2225 if (group_faults
> max_group_faults
) {
2226 max_group_faults
= group_faults
;
2227 max_group_nid
= nid
;
2231 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
2233 if (p
->numa_group
) {
2234 numa_group_count_active_nodes(p
->numa_group
);
2235 spin_unlock_irq(group_lock
);
2236 max_nid
= preferred_group_nid(p
, max_group_nid
);
2240 /* Set the new preferred node */
2241 if (max_nid
!= p
->numa_preferred_nid
)
2242 sched_setnuma(p
, max_nid
);
2244 if (task_node(p
) != p
->numa_preferred_nid
)
2245 numa_migrate_preferred(p
);
2249 static inline int get_numa_group(struct numa_group
*grp
)
2251 return atomic_inc_not_zero(&grp
->refcount
);
2254 static inline void put_numa_group(struct numa_group
*grp
)
2256 if (atomic_dec_and_test(&grp
->refcount
))
2257 kfree_rcu(grp
, rcu
);
2260 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
2263 struct numa_group
*grp
, *my_grp
;
2264 struct task_struct
*tsk
;
2266 int cpu
= cpupid_to_cpu(cpupid
);
2269 if (unlikely(!p
->numa_group
)) {
2270 unsigned int size
= sizeof(struct numa_group
) +
2271 4*nr_node_ids
*sizeof(unsigned long);
2273 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
2277 atomic_set(&grp
->refcount
, 1);
2278 grp
->active_nodes
= 1;
2279 grp
->max_faults_cpu
= 0;
2280 spin_lock_init(&grp
->lock
);
2282 /* Second half of the array tracks nids where faults happen */
2283 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
2286 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2287 grp
->faults
[i
] = p
->numa_faults
[i
];
2289 grp
->total_faults
= p
->total_numa_faults
;
2292 rcu_assign_pointer(p
->numa_group
, grp
);
2296 tsk
= READ_ONCE(cpu_rq(cpu
)->curr
);
2298 if (!cpupid_match_pid(tsk
, cpupid
))
2301 grp
= rcu_dereference(tsk
->numa_group
);
2305 my_grp
= p
->numa_group
;
2310 * Only join the other group if its bigger; if we're the bigger group,
2311 * the other task will join us.
2313 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
2317 * Tie-break on the grp address.
2319 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
2322 /* Always join threads in the same process. */
2323 if (tsk
->mm
== current
->mm
)
2326 /* Simple filter to avoid false positives due to PID collisions */
2327 if (flags
& TNF_SHARED
)
2330 /* Update priv based on whether false sharing was detected */
2333 if (join
&& !get_numa_group(grp
))
2341 BUG_ON(irqs_disabled());
2342 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
2344 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
2345 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
2346 grp
->faults
[i
] += p
->numa_faults
[i
];
2348 my_grp
->total_faults
-= p
->total_numa_faults
;
2349 grp
->total_faults
+= p
->total_numa_faults
;
2354 spin_unlock(&my_grp
->lock
);
2355 spin_unlock_irq(&grp
->lock
);
2357 rcu_assign_pointer(p
->numa_group
, grp
);
2359 put_numa_group(my_grp
);
2368 * Get rid of NUMA staticstics associated with a task (either current or dead).
2369 * If @final is set, the task is dead and has reached refcount zero, so we can
2370 * safely free all relevant data structures. Otherwise, there might be
2371 * concurrent reads from places like load balancing and procfs, and we should
2372 * reset the data back to default state without freeing ->numa_faults.
2374 void task_numa_free(struct task_struct
*p
, bool final
)
2376 struct numa_group
*grp
= p
->numa_group
;
2377 unsigned long *numa_faults
= p
->numa_faults
;
2378 unsigned long flags
;
2385 spin_lock_irqsave(&grp
->lock
, flags
);
2386 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2387 grp
->faults
[i
] -= p
->numa_faults
[i
];
2388 grp
->total_faults
-= p
->total_numa_faults
;
2391 spin_unlock_irqrestore(&grp
->lock
, flags
);
2392 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2393 put_numa_group(grp
);
2397 p
->numa_faults
= NULL
;
2400 p
->total_numa_faults
= 0;
2401 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2407 * Got a PROT_NONE fault for a page on @node.
2409 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2411 struct task_struct
*p
= current
;
2412 bool migrated
= flags
& TNF_MIGRATED
;
2413 int cpu_node
= task_node(current
);
2414 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2415 struct numa_group
*ng
;
2418 if (!static_branch_likely(&sched_numa_balancing
))
2421 /* for example, ksmd faulting in a user's mm */
2425 /* Allocate buffer to track faults on a per-node basis */
2426 if (unlikely(!p
->numa_faults
)) {
2427 int size
= sizeof(*p
->numa_faults
) *
2428 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2430 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2431 if (!p
->numa_faults
)
2434 p
->total_numa_faults
= 0;
2435 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2439 * First accesses are treated as private, otherwise consider accesses
2440 * to be private if the accessing pid has not changed
2442 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2445 priv
= cpupid_match_pid(p
, last_cpupid
);
2446 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2447 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2451 * If a workload spans multiple NUMA nodes, a shared fault that
2452 * occurs wholly within the set of nodes that the workload is
2453 * actively using should be counted as local. This allows the
2454 * scan rate to slow down when a workload has settled down.
2457 if (!priv
&& !local
&& ng
&& ng
->active_nodes
> 1 &&
2458 numa_is_active_node(cpu_node
, ng
) &&
2459 numa_is_active_node(mem_node
, ng
))
2462 task_numa_placement(p
);
2465 * Retry task to preferred node migration periodically, in case it
2466 * case it previously failed, or the scheduler moved us.
2468 if (time_after(jiffies
, p
->numa_migrate_retry
))
2469 numa_migrate_preferred(p
);
2472 p
->numa_pages_migrated
+= pages
;
2473 if (flags
& TNF_MIGRATE_FAIL
)
2474 p
->numa_faults_locality
[2] += pages
;
2476 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2477 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2478 p
->numa_faults_locality
[local
] += pages
;
2481 static void reset_ptenuma_scan(struct task_struct
*p
)
2484 * We only did a read acquisition of the mmap sem, so
2485 * p->mm->numa_scan_seq is written to without exclusive access
2486 * and the update is not guaranteed to be atomic. That's not
2487 * much of an issue though, since this is just used for
2488 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2489 * expensive, to avoid any form of compiler optimizations:
2491 WRITE_ONCE(p
->mm
->numa_scan_seq
, READ_ONCE(p
->mm
->numa_scan_seq
) + 1);
2492 p
->mm
->numa_scan_offset
= 0;
2496 * The expensive part of numa migration is done from task_work context.
2497 * Triggered from task_tick_numa().
2499 void task_numa_work(struct callback_head
*work
)
2501 unsigned long migrate
, next_scan
, now
= jiffies
;
2502 struct task_struct
*p
= current
;
2503 struct mm_struct
*mm
= p
->mm
;
2504 u64 runtime
= p
->se
.sum_exec_runtime
;
2505 struct vm_area_struct
*vma
;
2506 unsigned long start
, end
;
2507 unsigned long nr_pte_updates
= 0;
2508 long pages
, virtpages
;
2510 SCHED_WARN_ON(p
!= container_of(work
, struct task_struct
, numa_work
));
2512 work
->next
= work
; /* protect against double add */
2514 * Who cares about NUMA placement when they're dying.
2516 * NOTE: make sure not to dereference p->mm before this check,
2517 * exit_task_work() happens _after_ exit_mm() so we could be called
2518 * without p->mm even though we still had it when we enqueued this
2521 if (p
->flags
& PF_EXITING
)
2524 if (!mm
->numa_next_scan
) {
2525 mm
->numa_next_scan
= now
+
2526 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2530 * Enforce maximal scan/migration frequency..
2532 migrate
= mm
->numa_next_scan
;
2533 if (time_before(now
, migrate
))
2536 if (p
->numa_scan_period
== 0) {
2537 p
->numa_scan_period_max
= task_scan_max(p
);
2538 p
->numa_scan_period
= task_scan_start(p
);
2541 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2542 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2546 * Delay this task enough that another task of this mm will likely win
2547 * the next time around.
2549 p
->node_stamp
+= 2 * TICK_NSEC
;
2551 start
= mm
->numa_scan_offset
;
2552 pages
= sysctl_numa_balancing_scan_size
;
2553 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2554 virtpages
= pages
* 8; /* Scan up to this much virtual space */
2559 if (!down_read_trylock(&mm
->mmap_sem
))
2561 vma
= find_vma(mm
, start
);
2563 reset_ptenuma_scan(p
);
2567 for (; vma
; vma
= vma
->vm_next
) {
2568 if (!vma_migratable(vma
) || !vma_policy_mof(vma
) ||
2569 is_vm_hugetlb_page(vma
) || (vma
->vm_flags
& VM_MIXEDMAP
)) {
2574 * Shared library pages mapped by multiple processes are not
2575 * migrated as it is expected they are cache replicated. Avoid
2576 * hinting faults in read-only file-backed mappings or the vdso
2577 * as migrating the pages will be of marginal benefit.
2580 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2584 * Skip inaccessible VMAs to avoid any confusion between
2585 * PROT_NONE and NUMA hinting ptes
2587 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
2591 start
= max(start
, vma
->vm_start
);
2592 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2593 end
= min(end
, vma
->vm_end
);
2594 nr_pte_updates
= change_prot_numa(vma
, start
, end
);
2597 * Try to scan sysctl_numa_balancing_size worth of
2598 * hpages that have at least one present PTE that
2599 * is not already pte-numa. If the VMA contains
2600 * areas that are unused or already full of prot_numa
2601 * PTEs, scan up to virtpages, to skip through those
2605 pages
-= (end
- start
) >> PAGE_SHIFT
;
2606 virtpages
-= (end
- start
) >> PAGE_SHIFT
;
2609 if (pages
<= 0 || virtpages
<= 0)
2613 } while (end
!= vma
->vm_end
);
2618 * It is possible to reach the end of the VMA list but the last few
2619 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2620 * would find the !migratable VMA on the next scan but not reset the
2621 * scanner to the start so check it now.
2624 mm
->numa_scan_offset
= start
;
2626 reset_ptenuma_scan(p
);
2627 up_read(&mm
->mmap_sem
);
2630 * Make sure tasks use at least 32x as much time to run other code
2631 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2632 * Usually update_task_scan_period slows down scanning enough; on an
2633 * overloaded system we need to limit overhead on a per task basis.
2635 if (unlikely(p
->se
.sum_exec_runtime
!= runtime
)) {
2636 u64 diff
= p
->se
.sum_exec_runtime
- runtime
;
2637 p
->node_stamp
+= 32 * diff
;
2642 * Drive the periodic memory faults..
2644 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2646 struct callback_head
*work
= &curr
->numa_work
;
2650 * We don't care about NUMA placement if we don't have memory.
2652 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2656 * Using runtime rather than walltime has the dual advantage that
2657 * we (mostly) drive the selection from busy threads and that the
2658 * task needs to have done some actual work before we bother with
2661 now
= curr
->se
.sum_exec_runtime
;
2662 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2664 if (now
> curr
->node_stamp
+ period
) {
2665 if (!curr
->node_stamp
)
2666 curr
->numa_scan_period
= task_scan_start(curr
);
2667 curr
->node_stamp
+= period
;
2669 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2670 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2671 task_work_add(curr
, work
, true);
2677 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2681 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2685 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2689 #endif /* CONFIG_NUMA_BALANCING */
2692 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2694 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2695 if (!parent_entity(se
))
2696 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2698 if (entity_is_task(se
)) {
2699 struct rq
*rq
= rq_of(cfs_rq
);
2701 account_numa_enqueue(rq
, task_of(se
));
2702 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2705 cfs_rq
->nr_running
++;
2709 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2711 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2712 if (!parent_entity(se
))
2713 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2715 if (entity_is_task(se
)) {
2716 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2717 list_del_init(&se
->group_node
);
2720 cfs_rq
->nr_running
--;
2724 * Signed add and clamp on underflow.
2726 * Explicitly do a load-store to ensure the intermediate value never hits
2727 * memory. This allows lockless observations without ever seeing the negative
2730 #define add_positive(_ptr, _val) do { \
2731 typeof(_ptr) ptr = (_ptr); \
2732 typeof(_val) val = (_val); \
2733 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2737 if (val < 0 && res > var) \
2740 WRITE_ONCE(*ptr, res); \
2744 * Unsigned subtract and clamp on underflow.
2746 * Explicitly do a load-store to ensure the intermediate value never hits
2747 * memory. This allows lockless observations without ever seeing the negative
2750 #define sub_positive(_ptr, _val) do { \
2751 typeof(_ptr) ptr = (_ptr); \
2752 typeof(*ptr) val = (_val); \
2753 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2757 WRITE_ONCE(*ptr, res); \
2761 * Remove and clamp on negative, from a local variable.
2763 * A variant of sub_positive(), which does not use explicit load-store
2764 * and is thus optimized for local variable updates.
2766 #define lsub_positive(_ptr, _val) do { \
2767 typeof(_ptr) ptr = (_ptr); \
2768 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2773 * XXX we want to get rid of these helpers and use the full load resolution.
2775 static inline long se_weight(struct sched_entity
*se
)
2777 return scale_load_down(se
->load
.weight
);
2780 static inline long se_runnable(struct sched_entity
*se
)
2782 return scale_load_down(se
->runnable_weight
);
2786 enqueue_runnable_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2788 cfs_rq
->runnable_weight
+= se
->runnable_weight
;
2790 cfs_rq
->avg
.runnable_load_avg
+= se
->avg
.runnable_load_avg
;
2791 cfs_rq
->avg
.runnable_load_sum
+= se_runnable(se
) * se
->avg
.runnable_load_sum
;
2795 dequeue_runnable_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2797 cfs_rq
->runnable_weight
-= se
->runnable_weight
;
2799 sub_positive(&cfs_rq
->avg
.runnable_load_avg
, se
->avg
.runnable_load_avg
);
2800 sub_positive(&cfs_rq
->avg
.runnable_load_sum
,
2801 se_runnable(se
) * se
->avg
.runnable_load_sum
);
2805 enqueue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2807 cfs_rq
->avg
.load_avg
+= se
->avg
.load_avg
;
2808 cfs_rq
->avg
.load_sum
+= se_weight(se
) * se
->avg
.load_sum
;
2812 dequeue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2814 sub_positive(&cfs_rq
->avg
.load_avg
, se
->avg
.load_avg
);
2815 sub_positive(&cfs_rq
->avg
.load_sum
, se_weight(se
) * se
->avg
.load_sum
);
2819 enqueue_runnable_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
2821 dequeue_runnable_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
2823 enqueue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
2825 dequeue_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) { }
2828 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2829 unsigned long weight
, unsigned long runnable
)
2832 /* commit outstanding execution time */
2833 if (cfs_rq
->curr
== se
)
2834 update_curr(cfs_rq
);
2835 account_entity_dequeue(cfs_rq
, se
);
2836 dequeue_runnable_load_avg(cfs_rq
, se
);
2838 dequeue_load_avg(cfs_rq
, se
);
2840 se
->runnable_weight
= runnable
;
2841 update_load_set(&se
->load
, weight
);
2845 u32 divider
= LOAD_AVG_MAX
- 1024 + se
->avg
.period_contrib
;
2847 se
->avg
.load_avg
= div_u64(se_weight(se
) * se
->avg
.load_sum
, divider
);
2848 se
->avg
.runnable_load_avg
=
2849 div_u64(se_runnable(se
) * se
->avg
.runnable_load_sum
, divider
);
2853 enqueue_load_avg(cfs_rq
, se
);
2855 account_entity_enqueue(cfs_rq
, se
);
2856 enqueue_runnable_load_avg(cfs_rq
, se
);
2860 void reweight_task(struct task_struct
*p
, int prio
)
2862 struct sched_entity
*se
= &p
->se
;
2863 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2864 struct load_weight
*load
= &se
->load
;
2865 unsigned long weight
= scale_load(sched_prio_to_weight
[prio
]);
2867 reweight_entity(cfs_rq
, se
, weight
, weight
);
2868 load
->inv_weight
= sched_prio_to_wmult
[prio
];
2871 #ifdef CONFIG_FAIR_GROUP_SCHED
2874 * All this does is approximate the hierarchical proportion which includes that
2875 * global sum we all love to hate.
2877 * That is, the weight of a group entity, is the proportional share of the
2878 * group weight based on the group runqueue weights. That is:
2880 * tg->weight * grq->load.weight
2881 * ge->load.weight = ----------------------------- (1)
2882 * \Sum grq->load.weight
2884 * Now, because computing that sum is prohibitively expensive to compute (been
2885 * there, done that) we approximate it with this average stuff. The average
2886 * moves slower and therefore the approximation is cheaper and more stable.
2888 * So instead of the above, we substitute:
2890 * grq->load.weight -> grq->avg.load_avg (2)
2892 * which yields the following:
2894 * tg->weight * grq->avg.load_avg
2895 * ge->load.weight = ------------------------------ (3)
2898 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2900 * That is shares_avg, and it is right (given the approximation (2)).
2902 * The problem with it is that because the average is slow -- it was designed
2903 * to be exactly that of course -- this leads to transients in boundary
2904 * conditions. In specific, the case where the group was idle and we start the
2905 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2906 * yielding bad latency etc..
2908 * Now, in that special case (1) reduces to:
2910 * tg->weight * grq->load.weight
2911 * ge->load.weight = ----------------------------- = tg->weight (4)
2914 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2916 * So what we do is modify our approximation (3) to approach (4) in the (near)
2921 * tg->weight * grq->load.weight
2922 * --------------------------------------------------- (5)
2923 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2925 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2926 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2929 * tg->weight * grq->load.weight
2930 * ge->load.weight = ----------------------------- (6)
2935 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2936 * max(grq->load.weight, grq->avg.load_avg)
2938 * And that is shares_weight and is icky. In the (near) UP case it approaches
2939 * (4) while in the normal case it approaches (3). It consistently
2940 * overestimates the ge->load.weight and therefore:
2942 * \Sum ge->load.weight >= tg->weight
2946 static long calc_group_shares(struct cfs_rq
*cfs_rq
)
2948 long tg_weight
, tg_shares
, load
, shares
;
2949 struct task_group
*tg
= cfs_rq
->tg
;
2951 tg_shares
= READ_ONCE(tg
->shares
);
2953 load
= max(scale_load_down(cfs_rq
->load
.weight
), cfs_rq
->avg
.load_avg
);
2955 tg_weight
= atomic_long_read(&tg
->load_avg
);
2957 /* Ensure tg_weight >= load */
2958 tg_weight
-= cfs_rq
->tg_load_avg_contrib
;
2961 shares
= (tg_shares
* load
);
2963 shares
/= tg_weight
;
2966 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2967 * of a group with small tg->shares value. It is a floor value which is
2968 * assigned as a minimum load.weight to the sched_entity representing
2969 * the group on a CPU.
2971 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2972 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2973 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2974 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2977 return clamp_t(long, shares
, MIN_SHARES
, tg_shares
);
2981 * This calculates the effective runnable weight for a group entity based on
2982 * the group entity weight calculated above.
2984 * Because of the above approximation (2), our group entity weight is
2985 * an load_avg based ratio (3). This means that it includes blocked load and
2986 * does not represent the runnable weight.
2988 * Approximate the group entity's runnable weight per ratio from the group
2991 * grq->avg.runnable_load_avg
2992 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2995 * However, analogous to above, since the avg numbers are slow, this leads to
2996 * transients in the from-idle case. Instead we use:
2998 * ge->runnable_weight = ge->load.weight *
3000 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
3001 * ----------------------------------------------------- (8)
3002 * max(grq->avg.load_avg, grq->load.weight)
3004 * Where these max() serve both to use the 'instant' values to fix the slow
3005 * from-idle and avoid the /0 on to-idle, similar to (6).
3007 static long calc_group_runnable(struct cfs_rq
*cfs_rq
, long shares
)
3009 long runnable
, load_avg
;
3011 load_avg
= max(cfs_rq
->avg
.load_avg
,
3012 scale_load_down(cfs_rq
->load
.weight
));
3014 runnable
= max(cfs_rq
->avg
.runnable_load_avg
,
3015 scale_load_down(cfs_rq
->runnable_weight
));
3019 runnable
/= load_avg
;
3021 return clamp_t(long, runnable
, MIN_SHARES
, shares
);
3023 # endif /* CONFIG_SMP */
3025 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
3028 * Recomputes the group entity based on the current state of its group
3031 static void update_cfs_group(struct sched_entity
*se
)
3033 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3034 long shares
, runnable
;
3039 if (throttled_hierarchy(gcfs_rq
))
3043 runnable
= shares
= READ_ONCE(gcfs_rq
->tg
->shares
);
3045 if (likely(se
->load
.weight
== shares
))
3048 shares
= calc_group_shares(gcfs_rq
);
3049 runnable
= calc_group_runnable(gcfs_rq
, shares
);
3052 reweight_entity(cfs_rq_of(se
), se
, shares
, runnable
);
3055 #else /* CONFIG_FAIR_GROUP_SCHED */
3056 static inline void update_cfs_group(struct sched_entity
*se
)
3059 #endif /* CONFIG_FAIR_GROUP_SCHED */
3061 static inline void cfs_rq_util_change(struct cfs_rq
*cfs_rq
)
3063 struct rq
*rq
= rq_of(cfs_rq
);
3065 if (&rq
->cfs
== cfs_rq
) {
3067 * There are a few boundary cases this might miss but it should
3068 * get called often enough that that should (hopefully) not be
3069 * a real problem -- added to that it only calls on the local
3070 * CPU, so if we enqueue remotely we'll miss an update, but
3071 * the next tick/schedule should update.
3073 * It will not get called when we go idle, because the idle
3074 * thread is a different class (!fair), nor will the utilization
3075 * number include things like RT tasks.
3077 * As is, the util number is not freq-invariant (we'd have to
3078 * implement arch_scale_freq_capacity() for that).
3082 cpufreq_update_util(rq
, 0);
3089 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
3091 static u64
decay_load(u64 val
, u64 n
)
3093 unsigned int local_n
;
3095 if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
3098 /* after bounds checking we can collapse to 32-bit */
3102 * As y^PERIOD = 1/2, we can combine
3103 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
3104 * With a look-up table which covers y^n (n<PERIOD)
3106 * To achieve constant time decay_load.
3108 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
3109 val
>>= local_n
/ LOAD_AVG_PERIOD
;
3110 local_n
%= LOAD_AVG_PERIOD
;
3113 val
= mul_u64_u32_shr(val
, runnable_avg_yN_inv
[local_n
], 32);
3117 static u32
__accumulate_pelt_segments(u64 periods
, u32 d1
, u32 d3
)
3119 u32 c1
, c2
, c3
= d3
; /* y^0 == 1 */
3124 c1
= decay_load((u64
)d1
, periods
);
3128 * c2 = 1024 \Sum y^n
3132 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
3135 c2
= LOAD_AVG_MAX
- decay_load(LOAD_AVG_MAX
, periods
) - 1024;
3137 return c1
+ c2
+ c3
;
3140 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
3143 * Accumulate the three separate parts of the sum; d1 the remainder
3144 * of the last (incomplete) period, d2 the span of full periods and d3
3145 * the remainder of the (incomplete) current period.
3150 * |<->|<----------------->|<--->|
3151 * ... |---x---|------| ... |------|-----x (now)
3154 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
3157 * = u y^p + (Step 1)
3160 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
3163 static __always_inline u32
3164 accumulate_sum(u64 delta
, int cpu
, struct sched_avg
*sa
,
3165 unsigned long load
, unsigned long runnable
, int running
)
3167 unsigned long scale_freq
, scale_cpu
;
3168 u32 contrib
= (u32
)delta
; /* p == 0 -> delta < 1024 */
3171 scale_freq
= arch_scale_freq_capacity(NULL
, cpu
);
3172 scale_cpu
= arch_scale_cpu_capacity(NULL
, cpu
);
3174 delta
+= sa
->period_contrib
;
3175 periods
= delta
/ 1024; /* A period is 1024us (~1ms) */
3178 * Step 1: decay old *_sum if we crossed period boundaries.
3181 sa
->load_sum
= decay_load(sa
->load_sum
, periods
);
3182 sa
->runnable_load_sum
=
3183 decay_load(sa
->runnable_load_sum
, periods
);
3184 sa
->util_sum
= decay_load((u64
)(sa
->util_sum
), periods
);
3190 contrib
= __accumulate_pelt_segments(periods
,
3191 1024 - sa
->period_contrib
, delta
);
3193 sa
->period_contrib
= delta
;
3195 contrib
= cap_scale(contrib
, scale_freq
);
3197 sa
->load_sum
+= load
* contrib
;
3199 sa
->runnable_load_sum
+= runnable
* contrib
;
3201 sa
->util_sum
+= contrib
* scale_cpu
;
3207 * We can represent the historical contribution to runnable average as the
3208 * coefficients of a geometric series. To do this we sub-divide our runnable
3209 * history into segments of approximately 1ms (1024us); label the segment that
3210 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
3212 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
3214 * (now) (~1ms ago) (~2ms ago)
3216 * Let u_i denote the fraction of p_i that the entity was runnable.
3218 * We then designate the fractions u_i as our co-efficients, yielding the
3219 * following representation of historical load:
3220 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
3222 * We choose y based on the with of a reasonably scheduling period, fixing:
3225 * This means that the contribution to load ~32ms ago (u_32) will be weighted
3226 * approximately half as much as the contribution to load within the last ms
3229 * When a period "rolls over" and we have new u_0`, multiplying the previous
3230 * sum again by y is sufficient to update:
3231 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
3232 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
3234 static __always_inline
int
3235 ___update_load_sum(u64 now
, int cpu
, struct sched_avg
*sa
,
3236 unsigned long load
, unsigned long runnable
, int running
)
3240 delta
= now
- sa
->last_update_time
;
3242 * This should only happen when time goes backwards, which it
3243 * unfortunately does during sched clock init when we swap over to TSC.
3245 if ((s64
)delta
< 0) {
3246 sa
->last_update_time
= now
;
3251 * Use 1024ns as the unit of measurement since it's a reasonable
3252 * approximation of 1us and fast to compute.
3258 sa
->last_update_time
+= delta
<< 10;
3261 * running is a subset of runnable (weight) so running can't be set if
3262 * runnable is clear. But there are some corner cases where the current
3263 * se has been already dequeued but cfs_rq->curr still points to it.
3264 * This means that weight will be 0 but not running for a sched_entity
3265 * but also for a cfs_rq if the latter becomes idle. As an example,
3266 * this happens during idle_balance() which calls
3267 * update_blocked_averages()
3270 runnable
= running
= 0;
3273 * Now we know we crossed measurement unit boundaries. The *_avg
3274 * accrues by two steps:
3276 * Step 1: accumulate *_sum since last_update_time. If we haven't
3277 * crossed period boundaries, finish.
3279 if (!accumulate_sum(delta
, cpu
, sa
, load
, runnable
, running
))
3285 static __always_inline
void
3286 ___update_load_avg(struct sched_avg
*sa
, unsigned long load
, unsigned long runnable
)
3288 u32 divider
= LOAD_AVG_MAX
- 1024 + sa
->period_contrib
;
3291 * Step 2: update *_avg.
3293 sa
->load_avg
= div_u64(load
* sa
->load_sum
, divider
);
3294 sa
->runnable_load_avg
= div_u64(runnable
* sa
->runnable_load_sum
, divider
);
3295 sa
->util_avg
= sa
->util_sum
/ divider
;
3302 * se_runnable() == se_weight()
3304 * group: [ see update_cfs_group() ]
3305 * se_weight() = tg->weight * grq->load_avg / tg->load_avg
3306 * se_runnable() = se_weight(se) * grq->runnable_load_avg / grq->load_avg
3308 * load_sum := runnable_sum
3309 * load_avg = se_weight(se) * runnable_avg
3311 * runnable_load_sum := runnable_sum
3312 * runnable_load_avg = se_runnable(se) * runnable_avg
3314 * XXX collapse load_sum and runnable_load_sum
3318 * load_sum = \Sum se_weight(se) * se->avg.load_sum
3319 * load_avg = \Sum se->avg.load_avg
3321 * runnable_load_sum = \Sum se_runnable(se) * se->avg.runnable_load_sum
3322 * runnable_load_avg = \Sum se->avg.runable_load_avg
3326 __update_load_avg_blocked_se(u64 now
, int cpu
, struct sched_entity
*se
)
3328 if (entity_is_task(se
))
3329 se
->runnable_weight
= se
->load
.weight
;
3331 if (___update_load_sum(now
, cpu
, &se
->avg
, 0, 0, 0)) {
3332 ___update_load_avg(&se
->avg
, se_weight(se
), se_runnable(se
));
3340 __update_load_avg_se(u64 now
, int cpu
, struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3342 if (entity_is_task(se
))
3343 se
->runnable_weight
= se
->load
.weight
;
3345 if (___update_load_sum(now
, cpu
, &se
->avg
, !!se
->on_rq
, !!se
->on_rq
,
3346 cfs_rq
->curr
== se
)) {
3348 ___update_load_avg(&se
->avg
, se_weight(se
), se_runnable(se
));
3356 __update_load_avg_cfs_rq(u64 now
, int cpu
, struct cfs_rq
*cfs_rq
)
3358 if (___update_load_sum(now
, cpu
, &cfs_rq
->avg
,
3359 scale_load_down(cfs_rq
->load
.weight
),
3360 scale_load_down(cfs_rq
->runnable_weight
),
3361 cfs_rq
->curr
!= NULL
)) {
3363 ___update_load_avg(&cfs_rq
->avg
, 1, 1);
3370 #ifdef CONFIG_FAIR_GROUP_SCHED
3372 * update_tg_load_avg - update the tg's load avg
3373 * @cfs_rq: the cfs_rq whose avg changed
3374 * @force: update regardless of how small the difference
3376 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3377 * However, because tg->load_avg is a global value there are performance
3380 * In order to avoid having to look at the other cfs_rq's, we use a
3381 * differential update where we store the last value we propagated. This in
3382 * turn allows skipping updates if the differential is 'small'.
3384 * Updating tg's load_avg is necessary before update_cfs_share().
3386 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
)
3388 long delta
= cfs_rq
->avg
.load_avg
- cfs_rq
->tg_load_avg_contrib
;
3391 * No need to update load_avg for root_task_group as it is not used.
3393 if (cfs_rq
->tg
== &root_task_group
)
3396 if (force
|| abs(delta
) > cfs_rq
->tg_load_avg_contrib
/ 64) {
3397 atomic_long_add(delta
, &cfs_rq
->tg
->load_avg
);
3398 cfs_rq
->tg_load_avg_contrib
= cfs_rq
->avg
.load_avg
;
3403 * Called within set_task_rq() right before setting a task's cpu. The
3404 * caller only guarantees p->pi_lock is held; no other assumptions,
3405 * including the state of rq->lock, should be made.
3407 void set_task_rq_fair(struct sched_entity
*se
,
3408 struct cfs_rq
*prev
, struct cfs_rq
*next
)
3410 u64 p_last_update_time
;
3411 u64 n_last_update_time
;
3413 if (!sched_feat(ATTACH_AGE_LOAD
))
3417 * We are supposed to update the task to "current" time, then its up to
3418 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3419 * getting what current time is, so simply throw away the out-of-date
3420 * time. This will result in the wakee task is less decayed, but giving
3421 * the wakee more load sounds not bad.
3423 if (!(se
->avg
.last_update_time
&& prev
))
3426 #ifndef CONFIG_64BIT
3428 u64 p_last_update_time_copy
;
3429 u64 n_last_update_time_copy
;
3432 p_last_update_time_copy
= prev
->load_last_update_time_copy
;
3433 n_last_update_time_copy
= next
->load_last_update_time_copy
;
3437 p_last_update_time
= prev
->avg
.last_update_time
;
3438 n_last_update_time
= next
->avg
.last_update_time
;
3440 } while (p_last_update_time
!= p_last_update_time_copy
||
3441 n_last_update_time
!= n_last_update_time_copy
);
3444 p_last_update_time
= prev
->avg
.last_update_time
;
3445 n_last_update_time
= next
->avg
.last_update_time
;
3447 __update_load_avg_blocked_se(p_last_update_time
, cpu_of(rq_of(prev
)), se
);
3448 se
->avg
.last_update_time
= n_last_update_time
;
3453 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3454 * propagate its contribution. The key to this propagation is the invariant
3455 * that for each group:
3457 * ge->avg == grq->avg (1)
3459 * _IFF_ we look at the pure running and runnable sums. Because they
3460 * represent the very same entity, just at different points in the hierarchy.
3462 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3463 * sum over (but still wrong, because the group entity and group rq do not have
3464 * their PELT windows aligned).
3466 * However, update_tg_cfs_runnable() is more complex. So we have:
3468 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3470 * And since, like util, the runnable part should be directly transferable,
3471 * the following would _appear_ to be the straight forward approach:
3473 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3475 * And per (1) we have:
3477 * ge->avg.runnable_avg == grq->avg.runnable_avg
3481 * ge->load.weight * grq->avg.load_avg
3482 * ge->avg.load_avg = ----------------------------------- (4)
3485 * Except that is wrong!
3487 * Because while for entities historical weight is not important and we
3488 * really only care about our future and therefore can consider a pure
3489 * runnable sum, runqueues can NOT do this.
3491 * We specifically want runqueues to have a load_avg that includes
3492 * historical weights. Those represent the blocked load, the load we expect
3493 * to (shortly) return to us. This only works by keeping the weights as
3494 * integral part of the sum. We therefore cannot decompose as per (3).
3496 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3497 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3498 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3499 * runnable section of these tasks overlap (or not). If they were to perfectly
3500 * align the rq as a whole would be runnable 2/3 of the time. If however we
3501 * always have at least 1 runnable task, the rq as a whole is always runnable.
3503 * So we'll have to approximate.. :/
3505 * Given the constraint:
3507 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3509 * We can construct a rule that adds runnable to a rq by assuming minimal
3512 * On removal, we'll assume each task is equally runnable; which yields:
3514 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3516 * XXX: only do this for the part of runnable > running ?
3521 update_tg_cfs_util(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3523 long delta
= gcfs_rq
->avg
.util_avg
- se
->avg
.util_avg
;
3525 /* Nothing to update */
3530 * The relation between sum and avg is:
3532 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3534 * however, the PELT windows are not aligned between grq and gse.
3537 /* Set new sched_entity's utilization */
3538 se
->avg
.util_avg
= gcfs_rq
->avg
.util_avg
;
3539 se
->avg
.util_sum
= se
->avg
.util_avg
* LOAD_AVG_MAX
;
3541 /* Update parent cfs_rq utilization */
3542 add_positive(&cfs_rq
->avg
.util_avg
, delta
);
3543 cfs_rq
->avg
.util_sum
= cfs_rq
->avg
.util_avg
* LOAD_AVG_MAX
;
3547 update_tg_cfs_runnable(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, struct cfs_rq
*gcfs_rq
)
3549 long delta_avg
, running_sum
, runnable_sum
= gcfs_rq
->prop_runnable_sum
;
3550 unsigned long runnable_load_avg
, load_avg
;
3551 u64 runnable_load_sum
, load_sum
= 0;
3557 gcfs_rq
->prop_runnable_sum
= 0;
3559 if (runnable_sum
>= 0) {
3561 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3562 * the CPU is saturated running == runnable.
3564 runnable_sum
+= se
->avg
.load_sum
;
3565 runnable_sum
= min(runnable_sum
, (long)LOAD_AVG_MAX
);
3568 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3569 * assuming all tasks are equally runnable.
3571 if (scale_load_down(gcfs_rq
->load
.weight
)) {
3572 load_sum
= div_s64(gcfs_rq
->avg
.load_sum
,
3573 scale_load_down(gcfs_rq
->load
.weight
));
3576 /* But make sure to not inflate se's runnable */
3577 runnable_sum
= min(se
->avg
.load_sum
, load_sum
);
3581 * runnable_sum can't be lower than running_sum
3582 * As running sum is scale with cpu capacity wehreas the runnable sum
3583 * is not we rescale running_sum 1st
3585 running_sum
= se
->avg
.util_sum
/
3586 arch_scale_cpu_capacity(NULL
, cpu_of(rq_of(cfs_rq
)));
3587 runnable_sum
= max(runnable_sum
, running_sum
);
3589 load_sum
= (s64
)se_weight(se
) * runnable_sum
;
3590 load_avg
= div_s64(load_sum
, LOAD_AVG_MAX
);
3592 delta_sum
= load_sum
- (s64
)se_weight(se
) * se
->avg
.load_sum
;
3593 delta_avg
= load_avg
- se
->avg
.load_avg
;
3595 se
->avg
.load_sum
= runnable_sum
;
3596 se
->avg
.load_avg
= load_avg
;
3597 add_positive(&cfs_rq
->avg
.load_avg
, delta_avg
);
3598 add_positive(&cfs_rq
->avg
.load_sum
, delta_sum
);
3600 runnable_load_sum
= (s64
)se_runnable(se
) * runnable_sum
;
3601 runnable_load_avg
= div_s64(runnable_load_sum
, LOAD_AVG_MAX
);
3602 delta_sum
= runnable_load_sum
- se_weight(se
) * se
->avg
.runnable_load_sum
;
3603 delta_avg
= runnable_load_avg
- se
->avg
.runnable_load_avg
;
3605 se
->avg
.runnable_load_sum
= runnable_sum
;
3606 se
->avg
.runnable_load_avg
= runnable_load_avg
;
3609 add_positive(&cfs_rq
->avg
.runnable_load_avg
, delta_avg
);
3610 add_positive(&cfs_rq
->avg
.runnable_load_sum
, delta_sum
);
3614 static inline void add_tg_cfs_propagate(struct cfs_rq
*cfs_rq
, long runnable_sum
)
3616 cfs_rq
->propagate
= 1;
3617 cfs_rq
->prop_runnable_sum
+= runnable_sum
;
3620 /* Update task and its cfs_rq load average */
3621 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3623 struct cfs_rq
*cfs_rq
, *gcfs_rq
;
3625 if (entity_is_task(se
))
3628 gcfs_rq
= group_cfs_rq(se
);
3629 if (!gcfs_rq
->propagate
)
3632 gcfs_rq
->propagate
= 0;
3634 cfs_rq
= cfs_rq_of(se
);
3636 add_tg_cfs_propagate(cfs_rq
, gcfs_rq
->prop_runnable_sum
);
3638 update_tg_cfs_util(cfs_rq
, se
, gcfs_rq
);
3639 update_tg_cfs_runnable(cfs_rq
, se
, gcfs_rq
);
3645 * Check if we need to update the load and the utilization of a blocked
3648 static inline bool skip_blocked_update(struct sched_entity
*se
)
3650 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3653 * If sched_entity still have not zero load or utilization, we have to
3656 if (se
->avg
.load_avg
|| se
->avg
.util_avg
)
3660 * If there is a pending propagation, we have to update the load and
3661 * the utilization of the sched_entity:
3663 if (gcfs_rq
->propagate
)
3667 * Otherwise, the load and the utilization of the sched_entity is
3668 * already zero and there is no pending propagation, so it will be a
3669 * waste of time to try to decay it:
3674 #else /* CONFIG_FAIR_GROUP_SCHED */
3676 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
) {}
3678 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3683 static inline void add_tg_cfs_propagate(struct cfs_rq
*cfs_rq
, long runnable_sum
) {}
3685 #endif /* CONFIG_FAIR_GROUP_SCHED */
3688 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3689 * @now: current time, as per cfs_rq_clock_task()
3690 * @cfs_rq: cfs_rq to update
3692 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3693 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3694 * post_init_entity_util_avg().
3696 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3698 * Returns true if the load decayed or we removed load.
3700 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3701 * call update_tg_load_avg() when this function returns true.
3704 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
)
3706 unsigned long removed_load
= 0, removed_util
= 0, removed_runnable_sum
= 0;
3707 struct sched_avg
*sa
= &cfs_rq
->avg
;
3710 if (cfs_rq
->removed
.nr
) {
3712 u32 divider
= LOAD_AVG_MAX
- 1024 + sa
->period_contrib
;
3714 raw_spin_lock(&cfs_rq
->removed
.lock
);
3715 swap(cfs_rq
->removed
.util_avg
, removed_util
);
3716 swap(cfs_rq
->removed
.load_avg
, removed_load
);
3717 swap(cfs_rq
->removed
.runnable_sum
, removed_runnable_sum
);
3718 cfs_rq
->removed
.nr
= 0;
3719 raw_spin_unlock(&cfs_rq
->removed
.lock
);
3722 sub_positive(&sa
->load_avg
, r
);
3723 sub_positive(&sa
->load_sum
, r
* divider
);
3726 sub_positive(&sa
->util_avg
, r
);
3727 sub_positive(&sa
->util_sum
, r
* divider
);
3729 add_tg_cfs_propagate(cfs_rq
, -(long)removed_runnable_sum
);
3734 decayed
|= __update_load_avg_cfs_rq(now
, cpu_of(rq_of(cfs_rq
)), cfs_rq
);
3736 #ifndef CONFIG_64BIT
3738 cfs_rq
->load_last_update_time_copy
= sa
->last_update_time
;
3742 cfs_rq_util_change(cfs_rq
);
3748 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3749 * @cfs_rq: cfs_rq to attach to
3750 * @se: sched_entity to attach
3752 * Must call update_cfs_rq_load_avg() before this, since we rely on
3753 * cfs_rq->avg.last_update_time being current.
3755 static void attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3757 u32 divider
= LOAD_AVG_MAX
- 1024 + cfs_rq
->avg
.period_contrib
;
3760 * When we attach the @se to the @cfs_rq, we must align the decay
3761 * window because without that, really weird and wonderful things can
3766 se
->avg
.last_update_time
= cfs_rq
->avg
.last_update_time
;
3767 se
->avg
.period_contrib
= cfs_rq
->avg
.period_contrib
;
3770 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3771 * period_contrib. This isn't strictly correct, but since we're
3772 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3775 se
->avg
.util_sum
= se
->avg
.util_avg
* divider
;
3777 se
->avg
.load_sum
= divider
;
3778 if (se_weight(se
)) {
3780 div_u64(se
->avg
.load_avg
* se
->avg
.load_sum
, se_weight(se
));
3783 se
->avg
.runnable_load_sum
= se
->avg
.load_sum
;
3785 enqueue_load_avg(cfs_rq
, se
);
3786 cfs_rq
->avg
.util_avg
+= se
->avg
.util_avg
;
3787 cfs_rq
->avg
.util_sum
+= se
->avg
.util_sum
;
3789 add_tg_cfs_propagate(cfs_rq
, se
->avg
.load_sum
);
3791 cfs_rq_util_change(cfs_rq
);
3795 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3796 * @cfs_rq: cfs_rq to detach from
3797 * @se: sched_entity to detach
3799 * Must call update_cfs_rq_load_avg() before this, since we rely on
3800 * cfs_rq->avg.last_update_time being current.
3802 static void detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3804 dequeue_load_avg(cfs_rq
, se
);
3805 sub_positive(&cfs_rq
->avg
.util_avg
, se
->avg
.util_avg
);
3806 sub_positive(&cfs_rq
->avg
.util_sum
, se
->avg
.util_sum
);
3808 add_tg_cfs_propagate(cfs_rq
, -se
->avg
.load_sum
);
3810 cfs_rq_util_change(cfs_rq
);
3814 * Optional action to be done while updating the load average
3816 #define UPDATE_TG 0x1
3817 #define SKIP_AGE_LOAD 0x2
3818 #define DO_ATTACH 0x4
3820 /* Update task and its cfs_rq load average */
3821 static inline void update_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3823 u64 now
= cfs_rq_clock_task(cfs_rq
);
3824 struct rq
*rq
= rq_of(cfs_rq
);
3825 int cpu
= cpu_of(rq
);
3829 * Track task load average for carrying it to new CPU after migrated, and
3830 * track group sched_entity load average for task_h_load calc in migration
3832 if (se
->avg
.last_update_time
&& !(flags
& SKIP_AGE_LOAD
))
3833 __update_load_avg_se(now
, cpu
, cfs_rq
, se
);
3835 decayed
= update_cfs_rq_load_avg(now
, cfs_rq
);
3836 decayed
|= propagate_entity_load_avg(se
);
3838 if (!se
->avg
.last_update_time
&& (flags
& DO_ATTACH
)) {
3840 attach_entity_load_avg(cfs_rq
, se
);
3841 update_tg_load_avg(cfs_rq
, 0);
3843 } else if (decayed
&& (flags
& UPDATE_TG
))
3844 update_tg_load_avg(cfs_rq
, 0);
3847 #ifndef CONFIG_64BIT
3848 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3850 u64 last_update_time_copy
;
3851 u64 last_update_time
;
3854 last_update_time_copy
= cfs_rq
->load_last_update_time_copy
;
3856 last_update_time
= cfs_rq
->avg
.last_update_time
;
3857 } while (last_update_time
!= last_update_time_copy
);
3859 return last_update_time
;
3862 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3864 return cfs_rq
->avg
.last_update_time
;
3869 * Synchronize entity load avg of dequeued entity without locking
3872 void sync_entity_load_avg(struct sched_entity
*se
)
3874 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3875 u64 last_update_time
;
3877 last_update_time
= cfs_rq_last_update_time(cfs_rq
);
3878 __update_load_avg_blocked_se(last_update_time
, cpu_of(rq_of(cfs_rq
)), se
);
3882 * Task first catches up with cfs_rq, and then subtract
3883 * itself from the cfs_rq (task must be off the queue now).
3885 void remove_entity_load_avg(struct sched_entity
*se
)
3887 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3888 unsigned long flags
;
3891 * tasks cannot exit without having gone through wake_up_new_task() ->
3892 * post_init_entity_util_avg() which will have added things to the
3893 * cfs_rq, so we can remove unconditionally.
3895 * Similarly for groups, they will have passed through
3896 * post_init_entity_util_avg() before unregister_sched_fair_group()
3900 sync_entity_load_avg(se
);
3902 raw_spin_lock_irqsave(&cfs_rq
->removed
.lock
, flags
);
3903 ++cfs_rq
->removed
.nr
;
3904 cfs_rq
->removed
.util_avg
+= se
->avg
.util_avg
;
3905 cfs_rq
->removed
.load_avg
+= se
->avg
.load_avg
;
3906 cfs_rq
->removed
.runnable_sum
+= se
->avg
.load_sum
; /* == runnable_sum */
3907 raw_spin_unlock_irqrestore(&cfs_rq
->removed
.lock
, flags
);
3910 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq
*cfs_rq
)
3912 return cfs_rq
->avg
.runnable_load_avg
;
3915 static inline unsigned long cfs_rq_load_avg(struct cfs_rq
*cfs_rq
)
3917 return cfs_rq
->avg
.load_avg
;
3920 static int idle_balance(struct rq
*this_rq
, struct rq_flags
*rf
);
3922 #else /* CONFIG_SMP */
3925 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
)
3930 #define UPDATE_TG 0x0
3931 #define SKIP_AGE_LOAD 0x0
3932 #define DO_ATTACH 0x0
3934 static inline void update_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int not_used1
)
3936 cfs_rq_util_change(cfs_rq
);
3939 static inline void remove_entity_load_avg(struct sched_entity
*se
) {}
3942 attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3944 detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3946 static inline int idle_balance(struct rq
*rq
, struct rq_flags
*rf
)
3951 #endif /* CONFIG_SMP */
3953 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3955 #ifdef CONFIG_SCHED_DEBUG
3956 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
3961 if (d
> 3*sysctl_sched_latency
)
3962 schedstat_inc(cfs_rq
->nr_spread_over
);
3967 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
3969 u64 vruntime
= cfs_rq
->min_vruntime
;
3972 * The 'current' period is already promised to the current tasks,
3973 * however the extra weight of the new task will slow them down a
3974 * little, place the new task so that it fits in the slot that
3975 * stays open at the end.
3977 if (initial
&& sched_feat(START_DEBIT
))
3978 vruntime
+= sched_vslice(cfs_rq
, se
);
3980 /* sleeps up to a single latency don't count. */
3982 unsigned long thresh
= sysctl_sched_latency
;
3985 * Halve their sleep time's effect, to allow
3986 * for a gentler effect of sleepers:
3988 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
3994 /* ensure we never gain time by being placed backwards. */
3995 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
3998 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
4000 static inline void check_schedstat_required(void)
4002 #ifdef CONFIG_SCHEDSTATS
4003 if (schedstat_enabled())
4006 /* Force schedstat enabled if a dependent tracepoint is active */
4007 if (trace_sched_stat_wait_enabled() ||
4008 trace_sched_stat_sleep_enabled() ||
4009 trace_sched_stat_iowait_enabled() ||
4010 trace_sched_stat_blocked_enabled() ||
4011 trace_sched_stat_runtime_enabled()) {
4012 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4013 "stat_blocked and stat_runtime require the "
4014 "kernel parameter schedstats=enable or "
4015 "kernel.sched_schedstats=1\n");
4026 * update_min_vruntime()
4027 * vruntime -= min_vruntime
4031 * update_min_vruntime()
4032 * vruntime += min_vruntime
4034 * this way the vruntime transition between RQs is done when both
4035 * min_vruntime are up-to-date.
4039 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4040 * vruntime -= min_vruntime
4044 * update_min_vruntime()
4045 * vruntime += min_vruntime
4047 * this way we don't have the most up-to-date min_vruntime on the originating
4048 * CPU and an up-to-date min_vruntime on the destination CPU.
4052 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
4054 bool renorm
= !(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_MIGRATED
);
4055 bool curr
= cfs_rq
->curr
== se
;
4058 * If we're the current task, we must renormalise before calling
4062 se
->vruntime
+= cfs_rq
->min_vruntime
;
4064 update_curr(cfs_rq
);
4067 * Otherwise, renormalise after, such that we're placed at the current
4068 * moment in time, instead of some random moment in the past. Being
4069 * placed in the past could significantly boost this task to the
4070 * fairness detriment of existing tasks.
4072 if (renorm
&& !curr
)
4073 se
->vruntime
+= cfs_rq
->min_vruntime
;
4076 * When enqueuing a sched_entity, we must:
4077 * - Update loads to have both entity and cfs_rq synced with now.
4078 * - Add its load to cfs_rq->runnable_avg
4079 * - For group_entity, update its weight to reflect the new share of
4081 * - Add its new weight to cfs_rq->load.weight
4083 update_load_avg(cfs_rq
, se
, UPDATE_TG
| DO_ATTACH
);
4084 update_cfs_group(se
);
4085 enqueue_runnable_load_avg(cfs_rq
, se
);
4086 account_entity_enqueue(cfs_rq
, se
);
4088 if (flags
& ENQUEUE_WAKEUP
)
4089 place_entity(cfs_rq
, se
, 0);
4091 check_schedstat_required();
4092 update_stats_enqueue(cfs_rq
, se
, flags
);
4093 check_spread(cfs_rq
, se
);
4095 __enqueue_entity(cfs_rq
, se
);
4098 if (cfs_rq
->nr_running
== 1) {
4099 list_add_leaf_cfs_rq(cfs_rq
);
4100 check_enqueue_throttle(cfs_rq
);
4104 static void __clear_buddies_last(struct sched_entity
*se
)
4106 for_each_sched_entity(se
) {
4107 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4108 if (cfs_rq
->last
!= se
)
4111 cfs_rq
->last
= NULL
;
4115 static void __clear_buddies_next(struct sched_entity
*se
)
4117 for_each_sched_entity(se
) {
4118 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4119 if (cfs_rq
->next
!= se
)
4122 cfs_rq
->next
= NULL
;
4126 static void __clear_buddies_skip(struct sched_entity
*se
)
4128 for_each_sched_entity(se
) {
4129 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4130 if (cfs_rq
->skip
!= se
)
4133 cfs_rq
->skip
= NULL
;
4137 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4139 if (cfs_rq
->last
== se
)
4140 __clear_buddies_last(se
);
4142 if (cfs_rq
->next
== se
)
4143 __clear_buddies_next(se
);
4145 if (cfs_rq
->skip
== se
)
4146 __clear_buddies_skip(se
);
4149 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
4152 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
4155 * Update run-time statistics of the 'current'.
4157 update_curr(cfs_rq
);
4160 * When dequeuing a sched_entity, we must:
4161 * - Update loads to have both entity and cfs_rq synced with now.
4162 * - Substract its load from the cfs_rq->runnable_avg.
4163 * - Substract its previous weight from cfs_rq->load.weight.
4164 * - For group entity, update its weight to reflect the new share
4165 * of its group cfs_rq.
4167 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4168 dequeue_runnable_load_avg(cfs_rq
, se
);
4170 update_stats_dequeue(cfs_rq
, se
, flags
);
4172 clear_buddies(cfs_rq
, se
);
4174 if (se
!= cfs_rq
->curr
)
4175 __dequeue_entity(cfs_rq
, se
);
4177 account_entity_dequeue(cfs_rq
, se
);
4180 * Normalize after update_curr(); which will also have moved
4181 * min_vruntime if @se is the one holding it back. But before doing
4182 * update_min_vruntime() again, which will discount @se's position and
4183 * can move min_vruntime forward still more.
4185 if (!(flags
& DEQUEUE_SLEEP
))
4186 se
->vruntime
-= cfs_rq
->min_vruntime
;
4188 /* return excess runtime on last dequeue */
4189 return_cfs_rq_runtime(cfs_rq
);
4191 update_cfs_group(se
);
4194 * Now advance min_vruntime if @se was the entity holding it back,
4195 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4196 * put back on, and if we advance min_vruntime, we'll be placed back
4197 * further than we started -- ie. we'll be penalized.
4199 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) != DEQUEUE_SAVE
)
4200 update_min_vruntime(cfs_rq
);
4204 * Preempt the current task with a newly woken task if needed:
4207 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
4209 unsigned long ideal_runtime
, delta_exec
;
4210 struct sched_entity
*se
;
4213 ideal_runtime
= sched_slice(cfs_rq
, curr
);
4214 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
4215 if (delta_exec
> ideal_runtime
) {
4216 resched_curr(rq_of(cfs_rq
));
4218 * The current task ran long enough, ensure it doesn't get
4219 * re-elected due to buddy favours.
4221 clear_buddies(cfs_rq
, curr
);
4226 * Ensure that a task that missed wakeup preemption by a
4227 * narrow margin doesn't have to wait for a full slice.
4228 * This also mitigates buddy induced latencies under load.
4230 if (delta_exec
< sysctl_sched_min_granularity
)
4233 se
= __pick_first_entity(cfs_rq
);
4234 delta
= curr
->vruntime
- se
->vruntime
;
4239 if (delta
> ideal_runtime
)
4240 resched_curr(rq_of(cfs_rq
));
4244 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
4246 /* 'current' is not kept within the tree. */
4249 * Any task has to be enqueued before it get to execute on
4250 * a CPU. So account for the time it spent waiting on the
4253 update_stats_wait_end(cfs_rq
, se
);
4254 __dequeue_entity(cfs_rq
, se
);
4255 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
4258 update_stats_curr_start(cfs_rq
, se
);
4262 * Track our maximum slice length, if the CPU's load is at
4263 * least twice that of our own weight (i.e. dont track it
4264 * when there are only lesser-weight tasks around):
4266 if (schedstat_enabled() && rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
4267 schedstat_set(se
->statistics
.slice_max
,
4268 max((u64
)schedstat_val(se
->statistics
.slice_max
),
4269 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
));
4272 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
4276 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
4279 * Pick the next process, keeping these things in mind, in this order:
4280 * 1) keep things fair between processes/task groups
4281 * 2) pick the "next" process, since someone really wants that to run
4282 * 3) pick the "last" process, for cache locality
4283 * 4) do not run the "skip" process, if something else is available
4285 static struct sched_entity
*
4286 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
4288 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
4289 struct sched_entity
*se
;
4292 * If curr is set we have to see if its left of the leftmost entity
4293 * still in the tree, provided there was anything in the tree at all.
4295 if (!left
|| (curr
&& entity_before(curr
, left
)))
4298 se
= left
; /* ideally we run the leftmost entity */
4301 * Avoid running the skip buddy, if running something else can
4302 * be done without getting too unfair.
4304 if (cfs_rq
->skip
== se
) {
4305 struct sched_entity
*second
;
4308 second
= __pick_first_entity(cfs_rq
);
4310 second
= __pick_next_entity(se
);
4311 if (!second
|| (curr
&& entity_before(curr
, second
)))
4315 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
4320 * Prefer last buddy, try to return the CPU to a preempted task.
4322 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
4326 * Someone really wants this to run. If it's not unfair, run it.
4328 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
4331 clear_buddies(cfs_rq
, se
);
4336 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
4338 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
4341 * If still on the runqueue then deactivate_task()
4342 * was not called and update_curr() has to be done:
4345 update_curr(cfs_rq
);
4347 /* throttle cfs_rqs exceeding runtime */
4348 check_cfs_rq_runtime(cfs_rq
);
4350 check_spread(cfs_rq
, prev
);
4353 update_stats_wait_start(cfs_rq
, prev
);
4354 /* Put 'current' back into the tree. */
4355 __enqueue_entity(cfs_rq
, prev
);
4356 /* in !on_rq case, update occurred at dequeue */
4357 update_load_avg(cfs_rq
, prev
, 0);
4359 cfs_rq
->curr
= NULL
;
4363 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
4366 * Update run-time statistics of the 'current'.
4368 update_curr(cfs_rq
);
4371 * Ensure that runnable average is periodically updated.
4373 update_load_avg(cfs_rq
, curr
, UPDATE_TG
);
4374 update_cfs_group(curr
);
4376 #ifdef CONFIG_SCHED_HRTICK
4378 * queued ticks are scheduled to match the slice, so don't bother
4379 * validating it and just reschedule.
4382 resched_curr(rq_of(cfs_rq
));
4386 * don't let the period tick interfere with the hrtick preemption
4388 if (!sched_feat(DOUBLE_TICK
) &&
4389 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
4393 if (cfs_rq
->nr_running
> 1)
4394 check_preempt_tick(cfs_rq
, curr
);
4398 /**************************************************
4399 * CFS bandwidth control machinery
4402 #ifdef CONFIG_CFS_BANDWIDTH
4404 #ifdef HAVE_JUMP_LABEL
4405 static struct static_key __cfs_bandwidth_used
;
4407 static inline bool cfs_bandwidth_used(void)
4409 return static_key_false(&__cfs_bandwidth_used
);
4412 void cfs_bandwidth_usage_inc(void)
4414 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used
);
4417 void cfs_bandwidth_usage_dec(void)
4419 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used
);
4421 #else /* HAVE_JUMP_LABEL */
4422 static bool cfs_bandwidth_used(void)
4427 void cfs_bandwidth_usage_inc(void) {}
4428 void cfs_bandwidth_usage_dec(void) {}
4429 #endif /* HAVE_JUMP_LABEL */
4432 * default period for cfs group bandwidth.
4433 * default: 0.1s, units: nanoseconds
4435 static inline u64
default_cfs_period(void)
4437 return 100000000ULL;
4440 static inline u64
sched_cfs_bandwidth_slice(void)
4442 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
4446 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4447 * directly instead of rq->clock to avoid adding additional synchronization
4450 * requires cfs_b->lock
4452 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
4454 if (cfs_b
->quota
!= RUNTIME_INF
)
4455 cfs_b
->runtime
= cfs_b
->quota
;
4458 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4460 return &tg
->cfs_bandwidth
;
4463 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4464 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
4466 if (unlikely(cfs_rq
->throttle_count
))
4467 return cfs_rq
->throttled_clock_task
- cfs_rq
->throttled_clock_task_time
;
4469 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
4472 /* returns 0 on failure to allocate runtime */
4473 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4475 struct task_group
*tg
= cfs_rq
->tg
;
4476 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
4477 u64 amount
= 0, min_amount
;
4479 /* note: this is a positive sum as runtime_remaining <= 0 */
4480 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
4482 raw_spin_lock(&cfs_b
->lock
);
4483 if (cfs_b
->quota
== RUNTIME_INF
)
4484 amount
= min_amount
;
4486 start_cfs_bandwidth(cfs_b
);
4488 if (cfs_b
->runtime
> 0) {
4489 amount
= min(cfs_b
->runtime
, min_amount
);
4490 cfs_b
->runtime
-= amount
;
4494 raw_spin_unlock(&cfs_b
->lock
);
4496 cfs_rq
->runtime_remaining
+= amount
;
4498 return cfs_rq
->runtime_remaining
> 0;
4501 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4503 /* dock delta_exec before expiring quota (as it could span periods) */
4504 cfs_rq
->runtime_remaining
-= delta_exec
;
4506 if (likely(cfs_rq
->runtime_remaining
> 0))
4509 if (cfs_rq
->throttled
)
4512 * if we're unable to extend our runtime we resched so that the active
4513 * hierarchy can be throttled
4515 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
4516 resched_curr(rq_of(cfs_rq
));
4519 static __always_inline
4520 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4522 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
4525 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
4528 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4530 return cfs_bandwidth_used() && cfs_rq
->throttled
;
4533 /* check whether cfs_rq, or any parent, is throttled */
4534 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4536 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
4540 * Ensure that neither of the group entities corresponding to src_cpu or
4541 * dest_cpu are members of a throttled hierarchy when performing group
4542 * load-balance operations.
4544 static inline int throttled_lb_pair(struct task_group
*tg
,
4545 int src_cpu
, int dest_cpu
)
4547 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
4549 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
4550 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
4552 return throttled_hierarchy(src_cfs_rq
) ||
4553 throttled_hierarchy(dest_cfs_rq
);
4556 /* updated child weight may affect parent so we have to do this bottom up */
4557 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
4559 struct rq
*rq
= data
;
4560 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4562 cfs_rq
->throttle_count
--;
4563 if (!cfs_rq
->throttle_count
) {
4564 /* adjust cfs_rq_clock_task() */
4565 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
4566 cfs_rq
->throttled_clock_task
;
4572 static int tg_throttle_down(struct task_group
*tg
, void *data
)
4574 struct rq
*rq
= data
;
4575 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4577 /* group is entering throttled state, stop time */
4578 if (!cfs_rq
->throttle_count
)
4579 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
4580 cfs_rq
->throttle_count
++;
4585 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
4587 struct rq
*rq
= rq_of(cfs_rq
);
4588 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4589 struct sched_entity
*se
;
4590 long task_delta
, dequeue
= 1;
4593 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
4595 /* freeze hierarchy runnable averages while throttled */
4597 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
4600 task_delta
= cfs_rq
->h_nr_running
;
4601 for_each_sched_entity(se
) {
4602 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4603 /* throttled entity or throttle-on-deactivate */
4608 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
4609 qcfs_rq
->h_nr_running
-= task_delta
;
4611 if (qcfs_rq
->load
.weight
)
4616 sub_nr_running(rq
, task_delta
);
4618 cfs_rq
->throttled
= 1;
4619 cfs_rq
->throttled_clock
= rq_clock(rq
);
4620 raw_spin_lock(&cfs_b
->lock
);
4621 empty
= list_empty(&cfs_b
->throttled_cfs_rq
);
4624 * Add to the _head_ of the list, so that an already-started
4625 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4626 * not running add to the tail so that later runqueues don't get starved.
4628 if (cfs_b
->distribute_running
)
4629 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
4631 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
4634 * If we're the first throttled task, make sure the bandwidth
4638 start_cfs_bandwidth(cfs_b
);
4640 raw_spin_unlock(&cfs_b
->lock
);
4643 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
4645 struct rq
*rq
= rq_of(cfs_rq
);
4646 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4647 struct sched_entity
*se
;
4651 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
4653 cfs_rq
->throttled
= 0;
4655 update_rq_clock(rq
);
4657 raw_spin_lock(&cfs_b
->lock
);
4658 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
4659 list_del_rcu(&cfs_rq
->throttled_list
);
4660 raw_spin_unlock(&cfs_b
->lock
);
4662 /* update hierarchical throttle state */
4663 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
4665 if (!cfs_rq
->load
.weight
)
4668 task_delta
= cfs_rq
->h_nr_running
;
4669 for_each_sched_entity(se
) {
4673 cfs_rq
= cfs_rq_of(se
);
4675 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
4676 cfs_rq
->h_nr_running
+= task_delta
;
4678 if (cfs_rq_throttled(cfs_rq
))
4683 add_nr_running(rq
, task_delta
);
4685 /* determine whether we need to wake up potentially idle cpu */
4686 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
4690 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
, u64 remaining
)
4692 struct cfs_rq
*cfs_rq
;
4694 u64 starting_runtime
= remaining
;
4697 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
4699 struct rq
*rq
= rq_of(cfs_rq
);
4703 if (!cfs_rq_throttled(cfs_rq
))
4706 /* By the above check, this should never be true */
4707 SCHED_WARN_ON(cfs_rq
->runtime_remaining
> 0);
4709 runtime
= -cfs_rq
->runtime_remaining
+ 1;
4710 if (runtime
> remaining
)
4711 runtime
= remaining
;
4712 remaining
-= runtime
;
4714 cfs_rq
->runtime_remaining
+= runtime
;
4716 /* we check whether we're throttled above */
4717 if (cfs_rq
->runtime_remaining
> 0)
4718 unthrottle_cfs_rq(cfs_rq
);
4728 return starting_runtime
- remaining
;
4732 * Responsible for refilling a task_group's bandwidth and unthrottling its
4733 * cfs_rqs as appropriate. If there has been no activity within the last
4734 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4735 * used to track this state.
4737 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
4742 /* no need to continue the timer with no bandwidth constraint */
4743 if (cfs_b
->quota
== RUNTIME_INF
)
4744 goto out_deactivate
;
4746 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4747 cfs_b
->nr_periods
+= overrun
;
4750 * idle depends on !throttled (for the case of a large deficit), and if
4751 * we're going inactive then everything else can be deferred
4753 if (cfs_b
->idle
&& !throttled
)
4754 goto out_deactivate
;
4756 __refill_cfs_bandwidth_runtime(cfs_b
);
4759 /* mark as potentially idle for the upcoming period */
4764 /* account preceding periods in which throttling occurred */
4765 cfs_b
->nr_throttled
+= overrun
;
4768 * This check is repeated as we are holding onto the new bandwidth while
4769 * we unthrottle. This can potentially race with an unthrottled group
4770 * trying to acquire new bandwidth from the global pool. This can result
4771 * in us over-using our runtime if it is all used during this loop, but
4772 * only by limited amounts in that extreme case.
4774 while (throttled
&& cfs_b
->runtime
> 0 && !cfs_b
->distribute_running
) {
4775 runtime
= cfs_b
->runtime
;
4776 cfs_b
->distribute_running
= 1;
4777 raw_spin_unlock(&cfs_b
->lock
);
4778 /* we can't nest cfs_b->lock while distributing bandwidth */
4779 runtime
= distribute_cfs_runtime(cfs_b
, runtime
);
4780 raw_spin_lock(&cfs_b
->lock
);
4782 cfs_b
->distribute_running
= 0;
4783 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4785 lsub_positive(&cfs_b
->runtime
, runtime
);
4789 * While we are ensured activity in the period following an
4790 * unthrottle, this also covers the case in which the new bandwidth is
4791 * insufficient to cover the existing bandwidth deficit. (Forcing the
4792 * timer to remain active while there are any throttled entities.)
4802 /* a cfs_rq won't donate quota below this amount */
4803 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
4804 /* minimum remaining period time to redistribute slack quota */
4805 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
4806 /* how long we wait to gather additional slack before distributing */
4807 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
4810 * Are we near the end of the current quota period?
4812 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4813 * hrtimer base being cleared by hrtimer_start. In the case of
4814 * migrate_hrtimers, base is never cleared, so we are fine.
4816 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
4818 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
4821 /* if the call-back is running a quota refresh is already occurring */
4822 if (hrtimer_callback_running(refresh_timer
))
4825 /* is a quota refresh about to occur? */
4826 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
4827 if (remaining
< min_expire
)
4833 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
4835 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
4837 /* if there's a quota refresh soon don't bother with slack */
4838 if (runtime_refresh_within(cfs_b
, min_left
))
4841 hrtimer_start(&cfs_b
->slack_timer
,
4842 ns_to_ktime(cfs_bandwidth_slack_period
),
4846 /* we know any runtime found here is valid as update_curr() precedes return */
4847 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4849 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4850 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
4852 if (slack_runtime
<= 0)
4855 raw_spin_lock(&cfs_b
->lock
);
4856 if (cfs_b
->quota
!= RUNTIME_INF
) {
4857 cfs_b
->runtime
+= slack_runtime
;
4859 /* we are under rq->lock, defer unthrottling using a timer */
4860 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
4861 !list_empty(&cfs_b
->throttled_cfs_rq
))
4862 start_cfs_slack_bandwidth(cfs_b
);
4864 raw_spin_unlock(&cfs_b
->lock
);
4866 /* even if it's not valid for return we don't want to try again */
4867 cfs_rq
->runtime_remaining
-= slack_runtime
;
4870 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4872 if (!cfs_bandwidth_used())
4875 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
4878 __return_cfs_rq_runtime(cfs_rq
);
4882 * This is done with a timer (instead of inline with bandwidth return) since
4883 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4885 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
4887 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
4889 /* confirm we're still not at a refresh boundary */
4890 raw_spin_lock(&cfs_b
->lock
);
4891 if (cfs_b
->distribute_running
) {
4892 raw_spin_unlock(&cfs_b
->lock
);
4896 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
4897 raw_spin_unlock(&cfs_b
->lock
);
4901 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
4902 runtime
= cfs_b
->runtime
;
4905 cfs_b
->distribute_running
= 1;
4907 raw_spin_unlock(&cfs_b
->lock
);
4912 runtime
= distribute_cfs_runtime(cfs_b
, runtime
);
4914 raw_spin_lock(&cfs_b
->lock
);
4915 lsub_positive(&cfs_b
->runtime
, runtime
);
4916 cfs_b
->distribute_running
= 0;
4917 raw_spin_unlock(&cfs_b
->lock
);
4921 * When a group wakes up we want to make sure that its quota is not already
4922 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4923 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4925 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
4927 if (!cfs_bandwidth_used())
4930 /* an active group must be handled by the update_curr()->put() path */
4931 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
4934 /* ensure the group is not already throttled */
4935 if (cfs_rq_throttled(cfs_rq
))
4938 /* update runtime allocation */
4939 account_cfs_rq_runtime(cfs_rq
, 0);
4940 if (cfs_rq
->runtime_remaining
<= 0)
4941 throttle_cfs_rq(cfs_rq
);
4944 static void sync_throttle(struct task_group
*tg
, int cpu
)
4946 struct cfs_rq
*pcfs_rq
, *cfs_rq
;
4948 if (!cfs_bandwidth_used())
4954 cfs_rq
= tg
->cfs_rq
[cpu
];
4955 pcfs_rq
= tg
->parent
->cfs_rq
[cpu
];
4957 cfs_rq
->throttle_count
= pcfs_rq
->throttle_count
;
4958 cfs_rq
->throttled_clock_task
= rq_clock_task(cpu_rq(cpu
));
4961 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4962 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4964 if (!cfs_bandwidth_used())
4967 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
4971 * it's possible for a throttled entity to be forced into a running
4972 * state (e.g. set_curr_task), in this case we're finished.
4974 if (cfs_rq_throttled(cfs_rq
))
4977 throttle_cfs_rq(cfs_rq
);
4981 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
4983 struct cfs_bandwidth
*cfs_b
=
4984 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
4986 do_sched_cfs_slack_timer(cfs_b
);
4988 return HRTIMER_NORESTART
;
4991 extern const u64 max_cfs_quota_period
;
4993 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
4995 struct cfs_bandwidth
*cfs_b
=
4996 container_of(timer
, struct cfs_bandwidth
, period_timer
);
5001 raw_spin_lock(&cfs_b
->lock
);
5003 overrun
= hrtimer_forward_now(timer
, cfs_b
->period
);
5008 u64
new, old
= ktime_to_ns(cfs_b
->period
);
5011 * Grow period by a factor of 2 to avoid losing precision.
5012 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5016 if (new < max_cfs_quota_period
) {
5017 cfs_b
->period
= ns_to_ktime(new);
5020 pr_warn_ratelimited(
5021 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5023 div_u64(new, NSEC_PER_USEC
),
5024 div_u64(cfs_b
->quota
, NSEC_PER_USEC
));
5026 pr_warn_ratelimited(
5027 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5029 div_u64(old
, NSEC_PER_USEC
),
5030 div_u64(cfs_b
->quota
, NSEC_PER_USEC
));
5033 /* reset count so we don't come right back in here */
5037 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
5040 cfs_b
->period_active
= 0;
5041 raw_spin_unlock(&cfs_b
->lock
);
5043 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
5046 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5048 raw_spin_lock_init(&cfs_b
->lock
);
5050 cfs_b
->quota
= RUNTIME_INF
;
5051 cfs_b
->period
= ns_to_ktime(default_cfs_period());
5053 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
5054 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_PINNED
);
5055 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
5056 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
5057 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
5058 cfs_b
->distribute_running
= 0;
5061 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
5063 cfs_rq
->runtime_enabled
= 0;
5064 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
5067 void start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5069 lockdep_assert_held(&cfs_b
->lock
);
5071 if (!cfs_b
->period_active
) {
5072 cfs_b
->period_active
= 1;
5073 hrtimer_forward_now(&cfs_b
->period_timer
, cfs_b
->period
);
5074 hrtimer_start_expires(&cfs_b
->period_timer
, HRTIMER_MODE_ABS_PINNED
);
5078 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
5080 /* init_cfs_bandwidth() was not called */
5081 if (!cfs_b
->throttled_cfs_rq
.next
)
5084 hrtimer_cancel(&cfs_b
->period_timer
);
5085 hrtimer_cancel(&cfs_b
->slack_timer
);
5089 * Both these cpu hotplug callbacks race against unregister_fair_sched_group()
5091 * The race is harmless, since modifying bandwidth settings of unhooked group
5092 * bits doesn't do much.
5095 /* cpu online calback */
5096 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
5098 struct task_group
*tg
;
5100 lockdep_assert_held(&rq
->lock
);
5103 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
5104 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
5105 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
5107 raw_spin_lock(&cfs_b
->lock
);
5108 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
5109 raw_spin_unlock(&cfs_b
->lock
);
5114 /* cpu offline callback */
5115 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
5117 struct task_group
*tg
;
5119 lockdep_assert_held(&rq
->lock
);
5122 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
5123 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
5125 if (!cfs_rq
->runtime_enabled
)
5129 * clock_task is not advancing so we just need to make sure
5130 * there's some valid quota amount
5132 cfs_rq
->runtime_remaining
= 1;
5134 * Offline rq is schedulable till cpu is completely disabled
5135 * in take_cpu_down(), so we prevent new cfs throttling here.
5137 cfs_rq
->runtime_enabled
= 0;
5139 if (cfs_rq_throttled(cfs_rq
))
5140 unthrottle_cfs_rq(cfs_rq
);
5145 #else /* CONFIG_CFS_BANDWIDTH */
5146 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
5148 return rq_clock_task(rq_of(cfs_rq
));
5151 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
5152 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
5153 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
5154 static inline void sync_throttle(struct task_group
*tg
, int cpu
) {}
5155 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
5157 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
5162 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
5167 static inline int throttled_lb_pair(struct task_group
*tg
,
5168 int src_cpu
, int dest_cpu
)
5173 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
5175 #ifdef CONFIG_FAIR_GROUP_SCHED
5176 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
5179 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
5183 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
5184 static inline void update_runtime_enabled(struct rq
*rq
) {}
5185 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
5187 #endif /* CONFIG_CFS_BANDWIDTH */
5189 /**************************************************
5190 * CFS operations on tasks:
5193 #ifdef CONFIG_SCHED_HRTICK
5194 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
5196 struct sched_entity
*se
= &p
->se
;
5197 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5199 SCHED_WARN_ON(task_rq(p
) != rq
);
5201 if (rq
->cfs
.h_nr_running
> 1) {
5202 u64 slice
= sched_slice(cfs_rq
, se
);
5203 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
5204 s64 delta
= slice
- ran
;
5211 hrtick_start(rq
, delta
);
5216 * called from enqueue/dequeue and updates the hrtick when the
5217 * current task is from our class and nr_running is low enough
5220 static void hrtick_update(struct rq
*rq
)
5222 struct task_struct
*curr
= rq
->curr
;
5224 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
5227 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
5228 hrtick_start_fair(rq
, curr
);
5230 #else /* !CONFIG_SCHED_HRTICK */
5232 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
5236 static inline void hrtick_update(struct rq
*rq
)
5242 * The enqueue_task method is called before nr_running is
5243 * increased. Here we update the fair scheduling stats and
5244 * then put the task into the rbtree:
5247 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
5249 struct cfs_rq
*cfs_rq
;
5250 struct sched_entity
*se
= &p
->se
;
5253 * If in_iowait is set, the code below may not trigger any cpufreq
5254 * utilization updates, so do it here explicitly with the IOWAIT flag
5258 cpufreq_update_util(rq
, SCHED_CPUFREQ_IOWAIT
);
5260 for_each_sched_entity(se
) {
5263 cfs_rq
= cfs_rq_of(se
);
5264 enqueue_entity(cfs_rq
, se
, flags
);
5267 * end evaluation on encountering a throttled cfs_rq
5269 * note: in the case of encountering a throttled cfs_rq we will
5270 * post the final h_nr_running increment below.
5272 if (cfs_rq_throttled(cfs_rq
))
5274 cfs_rq
->h_nr_running
++;
5276 flags
= ENQUEUE_WAKEUP
;
5279 for_each_sched_entity(se
) {
5280 cfs_rq
= cfs_rq_of(se
);
5281 cfs_rq
->h_nr_running
++;
5283 if (cfs_rq_throttled(cfs_rq
))
5286 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
5287 update_cfs_group(se
);
5291 add_nr_running(rq
, 1);
5293 assert_list_leaf_cfs_rq(rq
);
5298 static void set_next_buddy(struct sched_entity
*se
);
5301 * The dequeue_task method is called before nr_running is
5302 * decreased. We remove the task from the rbtree and
5303 * update the fair scheduling stats:
5305 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
5307 struct cfs_rq
*cfs_rq
;
5308 struct sched_entity
*se
= &p
->se
;
5309 int task_sleep
= flags
& DEQUEUE_SLEEP
;
5311 for_each_sched_entity(se
) {
5312 cfs_rq
= cfs_rq_of(se
);
5313 dequeue_entity(cfs_rq
, se
, flags
);
5316 * end evaluation on encountering a throttled cfs_rq
5318 * note: in the case of encountering a throttled cfs_rq we will
5319 * post the final h_nr_running decrement below.
5321 if (cfs_rq_throttled(cfs_rq
))
5323 cfs_rq
->h_nr_running
--;
5325 /* Don't dequeue parent if it has other entities besides us */
5326 if (cfs_rq
->load
.weight
) {
5327 /* Avoid re-evaluating load for this entity: */
5328 se
= parent_entity(se
);
5330 * Bias pick_next to pick a task from this cfs_rq, as
5331 * p is sleeping when it is within its sched_slice.
5333 if (task_sleep
&& se
&& !throttled_hierarchy(cfs_rq
))
5337 flags
|= DEQUEUE_SLEEP
;
5340 for_each_sched_entity(se
) {
5341 cfs_rq
= cfs_rq_of(se
);
5342 cfs_rq
->h_nr_running
--;
5344 if (cfs_rq_throttled(cfs_rq
))
5347 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
5348 update_cfs_group(se
);
5352 sub_nr_running(rq
, 1);
5359 /* Working cpumask for: load_balance, load_balance_newidle. */
5360 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5361 DEFINE_PER_CPU(cpumask_var_t
, select_idle_mask
);
5363 #ifdef CONFIG_NO_HZ_COMMON
5365 * per rq 'load' arrray crap; XXX kill this.
5369 * The exact cpuload calculated at every tick would be:
5371 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5373 * If a cpu misses updates for n ticks (as it was idle) and update gets
5374 * called on the n+1-th tick when cpu may be busy, then we have:
5376 * load_n = (1 - 1/2^i)^n * load_0
5377 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5379 * decay_load_missed() below does efficient calculation of
5381 * load' = (1 - 1/2^i)^n * load
5383 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5384 * This allows us to precompute the above in said factors, thereby allowing the
5385 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5386 * fixed_power_int())
5388 * The calculation is approximated on a 128 point scale.
5390 #define DEGRADE_SHIFT 7
5392 static const u8 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
5393 static const u8 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
5394 { 0, 0, 0, 0, 0, 0, 0, 0 },
5395 { 64, 32, 8, 0, 0, 0, 0, 0 },
5396 { 96, 72, 40, 12, 1, 0, 0, 0 },
5397 { 112, 98, 75, 43, 15, 1, 0, 0 },
5398 { 120, 112, 98, 76, 45, 16, 2, 0 }
5402 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5403 * would be when CPU is idle and so we just decay the old load without
5404 * adding any new load.
5406 static unsigned long
5407 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
5411 if (!missed_updates
)
5414 if (missed_updates
>= degrade_zero_ticks
[idx
])
5418 return load
>> missed_updates
;
5420 while (missed_updates
) {
5421 if (missed_updates
% 2)
5422 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
5424 missed_updates
>>= 1;
5429 #endif /* CONFIG_NO_HZ_COMMON */
5432 * __cpu_load_update - update the rq->cpu_load[] statistics
5433 * @this_rq: The rq to update statistics for
5434 * @this_load: The current load
5435 * @pending_updates: The number of missed updates
5437 * Update rq->cpu_load[] statistics. This function is usually called every
5438 * scheduler tick (TICK_NSEC).
5440 * This function computes a decaying average:
5442 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5444 * Because of NOHZ it might not get called on every tick which gives need for
5445 * the @pending_updates argument.
5447 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5448 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5449 * = A * (A * load[i]_n-2 + B) + B
5450 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5451 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5452 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5453 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5454 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5456 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5457 * any change in load would have resulted in the tick being turned back on.
5459 * For regular NOHZ, this reduces to:
5461 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5463 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5466 static void cpu_load_update(struct rq
*this_rq
, unsigned long this_load
,
5467 unsigned long pending_updates
)
5469 unsigned long __maybe_unused tickless_load
= this_rq
->cpu_load
[0];
5472 this_rq
->nr_load_updates
++;
5474 /* Update our load: */
5475 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
5476 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
5477 unsigned long old_load
, new_load
;
5479 /* scale is effectively 1 << i now, and >> i divides by scale */
5481 old_load
= this_rq
->cpu_load
[i
];
5482 #ifdef CONFIG_NO_HZ_COMMON
5483 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
5484 if (tickless_load
) {
5485 old_load
-= decay_load_missed(tickless_load
, pending_updates
- 1, i
);
5487 * old_load can never be a negative value because a
5488 * decayed tickless_load cannot be greater than the
5489 * original tickless_load.
5491 old_load
+= tickless_load
;
5494 new_load
= this_load
;
5496 * Round up the averaging division if load is increasing. This
5497 * prevents us from getting stuck on 9 if the load is 10, for
5500 if (new_load
> old_load
)
5501 new_load
+= scale
- 1;
5503 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
5506 sched_avg_update(this_rq
);
5509 /* Used instead of source_load when we know the type == 0 */
5510 static unsigned long weighted_cpuload(struct rq
*rq
)
5512 return cfs_rq_runnable_load_avg(&rq
->cfs
);
5515 #ifdef CONFIG_NO_HZ_COMMON
5517 * There is no sane way to deal with nohz on smp when using jiffies because the
5518 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5519 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5521 * Therefore we need to avoid the delta approach from the regular tick when
5522 * possible since that would seriously skew the load calculation. This is why we
5523 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5524 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5525 * loop exit, nohz_idle_balance, nohz full exit...)
5527 * This means we might still be one tick off for nohz periods.
5530 static void cpu_load_update_nohz(struct rq
*this_rq
,
5531 unsigned long curr_jiffies
,
5534 unsigned long pending_updates
;
5536 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
5537 if (pending_updates
) {
5538 this_rq
->last_load_update_tick
= curr_jiffies
;
5540 * In the regular NOHZ case, we were idle, this means load 0.
5541 * In the NOHZ_FULL case, we were non-idle, we should consider
5542 * its weighted load.
5544 cpu_load_update(this_rq
, load
, pending_updates
);
5549 * Called from nohz_idle_balance() to update the load ratings before doing the
5552 static void cpu_load_update_idle(struct rq
*this_rq
)
5555 * bail if there's load or we're actually up-to-date.
5557 if (weighted_cpuload(this_rq
))
5560 cpu_load_update_nohz(this_rq
, READ_ONCE(jiffies
), 0);
5564 * Record CPU load on nohz entry so we know the tickless load to account
5565 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5566 * than other cpu_load[idx] but it should be fine as cpu_load readers
5567 * shouldn't rely into synchronized cpu_load[*] updates.
5569 void cpu_load_update_nohz_start(void)
5571 struct rq
*this_rq
= this_rq();
5574 * This is all lockless but should be fine. If weighted_cpuload changes
5575 * concurrently we'll exit nohz. And cpu_load write can race with
5576 * cpu_load_update_idle() but both updater would be writing the same.
5578 this_rq
->cpu_load
[0] = weighted_cpuload(this_rq
);
5582 * Account the tickless load in the end of a nohz frame.
5584 void cpu_load_update_nohz_stop(void)
5586 unsigned long curr_jiffies
= READ_ONCE(jiffies
);
5587 struct rq
*this_rq
= this_rq();
5591 if (curr_jiffies
== this_rq
->last_load_update_tick
)
5594 load
= weighted_cpuload(this_rq
);
5595 rq_lock(this_rq
, &rf
);
5596 update_rq_clock(this_rq
);
5597 cpu_load_update_nohz(this_rq
, curr_jiffies
, load
);
5598 rq_unlock(this_rq
, &rf
);
5600 #else /* !CONFIG_NO_HZ_COMMON */
5601 static inline void cpu_load_update_nohz(struct rq
*this_rq
,
5602 unsigned long curr_jiffies
,
5603 unsigned long load
) { }
5604 #endif /* CONFIG_NO_HZ_COMMON */
5606 static void cpu_load_update_periodic(struct rq
*this_rq
, unsigned long load
)
5608 #ifdef CONFIG_NO_HZ_COMMON
5609 /* See the mess around cpu_load_update_nohz(). */
5610 this_rq
->last_load_update_tick
= READ_ONCE(jiffies
);
5612 cpu_load_update(this_rq
, load
, 1);
5616 * Called from scheduler_tick()
5618 void cpu_load_update_active(struct rq
*this_rq
)
5620 unsigned long load
= weighted_cpuload(this_rq
);
5622 if (tick_nohz_tick_stopped())
5623 cpu_load_update_nohz(this_rq
, READ_ONCE(jiffies
), load
);
5625 cpu_load_update_periodic(this_rq
, load
);
5629 * Return a low guess at the load of a migration-source cpu weighted
5630 * according to the scheduling class and "nice" value.
5632 * We want to under-estimate the load of migration sources, to
5633 * balance conservatively.
5635 static unsigned long source_load(int cpu
, int type
)
5637 struct rq
*rq
= cpu_rq(cpu
);
5638 unsigned long total
= weighted_cpuload(rq
);
5640 if (type
== 0 || !sched_feat(LB_BIAS
))
5643 return min(rq
->cpu_load
[type
-1], total
);
5647 * Return a high guess at the load of a migration-target cpu weighted
5648 * according to the scheduling class and "nice" value.
5650 static unsigned long target_load(int cpu
, int type
)
5652 struct rq
*rq
= cpu_rq(cpu
);
5653 unsigned long total
= weighted_cpuload(rq
);
5655 if (type
== 0 || !sched_feat(LB_BIAS
))
5658 return max(rq
->cpu_load
[type
-1], total
);
5661 static unsigned long capacity_of(int cpu
)
5663 return cpu_rq(cpu
)->cpu_capacity
;
5666 static unsigned long capacity_orig_of(int cpu
)
5668 return cpu_rq(cpu
)->cpu_capacity_orig
;
5671 static unsigned long cpu_avg_load_per_task(int cpu
)
5673 struct rq
*rq
= cpu_rq(cpu
);
5674 unsigned long nr_running
= READ_ONCE(rq
->cfs
.h_nr_running
);
5675 unsigned long load_avg
= weighted_cpuload(rq
);
5678 return load_avg
/ nr_running
;
5683 static void record_wakee(struct task_struct
*p
)
5686 * Only decay a single time; tasks that have less then 1 wakeup per
5687 * jiffy will not have built up many flips.
5689 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
5690 current
->wakee_flips
>>= 1;
5691 current
->wakee_flip_decay_ts
= jiffies
;
5694 if (current
->last_wakee
!= p
) {
5695 current
->last_wakee
= p
;
5696 current
->wakee_flips
++;
5701 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5703 * A waker of many should wake a different task than the one last awakened
5704 * at a frequency roughly N times higher than one of its wakees.
5706 * In order to determine whether we should let the load spread vs consolidating
5707 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5708 * partner, and a factor of lls_size higher frequency in the other.
5710 * With both conditions met, we can be relatively sure that the relationship is
5711 * non-monogamous, with partner count exceeding socket size.
5713 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5714 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5717 static int wake_wide(struct task_struct
*p
)
5719 unsigned int master
= current
->wakee_flips
;
5720 unsigned int slave
= p
->wakee_flips
;
5721 int factor
= this_cpu_read(sd_llc_size
);
5724 swap(master
, slave
);
5725 if (slave
< factor
|| master
< slave
* factor
)
5731 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5732 * soonest. For the purpose of speed we only consider the waking and previous
5735 * wake_affine_idle() - only considers 'now', it check if the waking CPU is (or
5738 * wake_affine_weight() - considers the weight to reflect the average
5739 * scheduling latency of the CPUs. This seems to work
5740 * for the overloaded case.
5744 wake_affine_idle(struct sched_domain
*sd
, struct task_struct
*p
,
5745 int this_cpu
, int prev_cpu
, int sync
)
5747 if (idle_cpu(this_cpu
))
5750 if (sync
&& cpu_rq(this_cpu
)->nr_running
== 1)
5757 wake_affine_weight(struct sched_domain
*sd
, struct task_struct
*p
,
5758 int this_cpu
, int prev_cpu
, int sync
)
5760 s64 this_eff_load
, prev_eff_load
;
5761 unsigned long task_load
;
5763 this_eff_load
= target_load(this_cpu
, sd
->wake_idx
);
5764 prev_eff_load
= source_load(prev_cpu
, sd
->wake_idx
);
5767 unsigned long current_load
= task_h_load(current
);
5769 if (current_load
> this_eff_load
)
5772 this_eff_load
-= current_load
;
5775 task_load
= task_h_load(p
);
5777 this_eff_load
+= task_load
;
5778 if (sched_feat(WA_BIAS
))
5779 this_eff_load
*= 100;
5780 this_eff_load
*= capacity_of(prev_cpu
);
5782 prev_eff_load
-= task_load
;
5783 if (sched_feat(WA_BIAS
))
5784 prev_eff_load
*= 100 + (sd
->imbalance_pct
- 100) / 2;
5785 prev_eff_load
*= capacity_of(this_cpu
);
5787 return this_eff_load
<= prev_eff_load
;
5790 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
,
5791 int prev_cpu
, int sync
)
5793 int this_cpu
= smp_processor_id();
5794 bool affine
= false;
5796 if (sched_feat(WA_IDLE
) && !affine
)
5797 affine
= wake_affine_idle(sd
, p
, this_cpu
, prev_cpu
, sync
);
5799 if (sched_feat(WA_WEIGHT
) && !affine
)
5800 affine
= wake_affine_weight(sd
, p
, this_cpu
, prev_cpu
, sync
);
5802 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine_attempts
);
5804 schedstat_inc(sd
->ttwu_move_affine
);
5805 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine
);
5811 static inline int task_util(struct task_struct
*p
);
5812 static int cpu_util_wake(int cpu
, struct task_struct
*p
);
5814 static unsigned long capacity_spare_wake(int cpu
, struct task_struct
*p
)
5816 return capacity_orig_of(cpu
) - cpu_util_wake(cpu
, p
);
5820 * find_idlest_group finds and returns the least busy CPU group within the
5823 * Assumes p is allowed on at least one CPU in sd.
5825 static struct sched_group
*
5826 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
5827 int this_cpu
, int sd_flag
)
5829 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
5830 struct sched_group
*most_spare_sg
= NULL
;
5831 unsigned long min_runnable_load
= ULONG_MAX
;
5832 unsigned long this_runnable_load
= ULONG_MAX
;
5833 unsigned long min_avg_load
= ULONG_MAX
, this_avg_load
= ULONG_MAX
;
5834 unsigned long most_spare
= 0, this_spare
= 0;
5835 int load_idx
= sd
->forkexec_idx
;
5836 int imbalance_scale
= 100 + (sd
->imbalance_pct
-100)/2;
5837 unsigned long imbalance
= scale_load_down(NICE_0_LOAD
) *
5838 (sd
->imbalance_pct
-100) / 100;
5840 if (sd_flag
& SD_BALANCE_WAKE
)
5841 load_idx
= sd
->wake_idx
;
5844 unsigned long load
, avg_load
, runnable_load
;
5845 unsigned long spare_cap
, max_spare_cap
;
5849 /* Skip over this group if it has no CPUs allowed */
5850 if (!cpumask_intersects(sched_group_span(group
),
5854 local_group
= cpumask_test_cpu(this_cpu
,
5855 sched_group_span(group
));
5858 * Tally up the load of all CPUs in the group and find
5859 * the group containing the CPU with most spare capacity.
5865 for_each_cpu(i
, sched_group_span(group
)) {
5866 /* Bias balancing toward cpus of our domain */
5868 load
= source_load(i
, load_idx
);
5870 load
= target_load(i
, load_idx
);
5872 runnable_load
+= load
;
5874 avg_load
+= cfs_rq_load_avg(&cpu_rq(i
)->cfs
);
5876 spare_cap
= capacity_spare_wake(i
, p
);
5878 if (spare_cap
> max_spare_cap
)
5879 max_spare_cap
= spare_cap
;
5882 /* Adjust by relative CPU capacity of the group */
5883 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) /
5884 group
->sgc
->capacity
;
5885 runnable_load
= (runnable_load
* SCHED_CAPACITY_SCALE
) /
5886 group
->sgc
->capacity
;
5889 this_runnable_load
= runnable_load
;
5890 this_avg_load
= avg_load
;
5891 this_spare
= max_spare_cap
;
5893 if (min_runnable_load
> (runnable_load
+ imbalance
)) {
5895 * The runnable load is significantly smaller
5896 * so we can pick this new cpu
5898 min_runnable_load
= runnable_load
;
5899 min_avg_load
= avg_load
;
5901 } else if ((runnable_load
< (min_runnable_load
+ imbalance
)) &&
5902 (100*min_avg_load
> imbalance_scale
*avg_load
)) {
5904 * The runnable loads are close so take the
5905 * blocked load into account through avg_load.
5907 min_avg_load
= avg_load
;
5911 if (most_spare
< max_spare_cap
) {
5912 most_spare
= max_spare_cap
;
5913 most_spare_sg
= group
;
5916 } while (group
= group
->next
, group
!= sd
->groups
);
5919 * The cross-over point between using spare capacity or least load
5920 * is too conservative for high utilization tasks on partially
5921 * utilized systems if we require spare_capacity > task_util(p),
5922 * so we allow for some task stuffing by using
5923 * spare_capacity > task_util(p)/2.
5925 * Spare capacity can't be used for fork because the utilization has
5926 * not been set yet, we must first select a rq to compute the initial
5929 if (sd_flag
& SD_BALANCE_FORK
)
5932 if (this_spare
> task_util(p
) / 2 &&
5933 imbalance_scale
*this_spare
> 100*most_spare
)
5936 if (most_spare
> task_util(p
) / 2)
5937 return most_spare_sg
;
5943 if (min_runnable_load
> (this_runnable_load
+ imbalance
))
5946 if ((this_runnable_load
< (min_runnable_load
+ imbalance
)) &&
5947 (100*this_avg_load
< imbalance_scale
*min_avg_load
))
5954 * find_idlest_group_cpu - find the idlest cpu among the cpus in group.
5957 find_idlest_group_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
5959 unsigned long load
, min_load
= ULONG_MAX
;
5960 unsigned int min_exit_latency
= UINT_MAX
;
5961 u64 latest_idle_timestamp
= 0;
5962 int least_loaded_cpu
= this_cpu
;
5963 int shallowest_idle_cpu
= -1;
5966 /* Check if we have any choice: */
5967 if (group
->group_weight
== 1)
5968 return cpumask_first(sched_group_span(group
));
5970 /* Traverse only the allowed CPUs */
5971 for_each_cpu_and(i
, sched_group_span(group
), &p
->cpus_allowed
) {
5973 struct rq
*rq
= cpu_rq(i
);
5974 struct cpuidle_state
*idle
= idle_get_state(rq
);
5975 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
5977 * We give priority to a CPU whose idle state
5978 * has the smallest exit latency irrespective
5979 * of any idle timestamp.
5981 min_exit_latency
= idle
->exit_latency
;
5982 latest_idle_timestamp
= rq
->idle_stamp
;
5983 shallowest_idle_cpu
= i
;
5984 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
5985 rq
->idle_stamp
> latest_idle_timestamp
) {
5987 * If equal or no active idle state, then
5988 * the most recently idled CPU might have
5991 latest_idle_timestamp
= rq
->idle_stamp
;
5992 shallowest_idle_cpu
= i
;
5994 } else if (shallowest_idle_cpu
== -1) {
5995 load
= weighted_cpuload(cpu_rq(i
));
5996 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
5998 least_loaded_cpu
= i
;
6003 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
6006 static inline int find_idlest_cpu(struct sched_domain
*sd
, struct task_struct
*p
,
6007 int cpu
, int prev_cpu
, int sd_flag
)
6011 if (!cpumask_intersects(sched_domain_span(sd
), &p
->cpus_allowed
))
6015 struct sched_group
*group
;
6016 struct sched_domain
*tmp
;
6019 if (!(sd
->flags
& sd_flag
)) {
6024 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
6030 new_cpu
= find_idlest_group_cpu(group
, p
, cpu
);
6031 if (new_cpu
== cpu
) {
6032 /* Now try balancing at a lower domain level of cpu */
6037 /* Now try balancing at a lower domain level of new_cpu */
6039 weight
= sd
->span_weight
;
6041 for_each_domain(cpu
, tmp
) {
6042 if (weight
<= tmp
->span_weight
)
6044 if (tmp
->flags
& sd_flag
)
6047 /* while loop will break here if sd == NULL */
6053 #ifdef CONFIG_SCHED_SMT
6054 DEFINE_STATIC_KEY_FALSE(sched_smt_present
);
6055 EXPORT_SYMBOL_GPL(sched_smt_present
);
6057 static inline void set_idle_cores(int cpu
, int val
)
6059 struct sched_domain_shared
*sds
;
6061 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
6063 WRITE_ONCE(sds
->has_idle_cores
, val
);
6066 static inline bool test_idle_cores(int cpu
, bool def
)
6068 struct sched_domain_shared
*sds
;
6070 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
6072 return READ_ONCE(sds
->has_idle_cores
);
6078 * Scans the local SMT mask to see if the entire core is idle, and records this
6079 * information in sd_llc_shared->has_idle_cores.
6081 * Since SMT siblings share all cache levels, inspecting this limited remote
6082 * state should be fairly cheap.
6084 void __update_idle_core(struct rq
*rq
)
6086 int core
= cpu_of(rq
);
6090 if (test_idle_cores(core
, true))
6093 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
6101 set_idle_cores(core
, 1);
6107 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6108 * there are no idle cores left in the system; tracked through
6109 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6111 static int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6113 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
6116 if (!static_branch_likely(&sched_smt_present
))
6119 if (!test_idle_cores(target
, false))
6122 cpumask_and(cpus
, sched_domain_span(sd
), &p
->cpus_allowed
);
6124 for_each_cpu_wrap(core
, cpus
, target
) {
6127 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
6128 cpumask_clear_cpu(cpu
, cpus
);
6138 * Failed to find an idle core; stop looking for one.
6140 set_idle_cores(target
, 0);
6146 * Scan the local SMT mask for idle CPUs.
6148 static int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6152 if (!static_branch_likely(&sched_smt_present
))
6155 for_each_cpu(cpu
, cpu_smt_mask(target
)) {
6156 if (!cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
6165 #else /* CONFIG_SCHED_SMT */
6167 static inline int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6172 static inline int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6177 #endif /* CONFIG_SCHED_SMT */
6180 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6181 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6182 * average idle time for this rq (as found in rq->avg_idle).
6184 static int select_idle_cpu(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
6186 struct sched_domain
*this_sd
;
6187 u64 avg_cost
, avg_idle
;
6190 int cpu
, nr
= INT_MAX
;
6192 this_sd
= rcu_dereference(*this_cpu_ptr(&sd_llc
));
6197 * Due to large variance we need a large fuzz factor; hackbench in
6198 * particularly is sensitive here.
6200 avg_idle
= this_rq()->avg_idle
/ 512;
6201 avg_cost
= this_sd
->avg_scan_cost
+ 1;
6203 if (sched_feat(SIS_AVG_CPU
) && avg_idle
< avg_cost
)
6206 if (sched_feat(SIS_PROP
)) {
6207 u64 span_avg
= sd
->span_weight
* avg_idle
;
6208 if (span_avg
> 4*avg_cost
)
6209 nr
= div_u64(span_avg
, avg_cost
);
6214 time
= local_clock();
6216 for_each_cpu_wrap(cpu
, sched_domain_span(sd
), target
) {
6219 if (!cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
6225 time
= local_clock() - time
;
6226 cost
= this_sd
->avg_scan_cost
;
6227 delta
= (s64
)(time
- cost
) / 8;
6228 this_sd
->avg_scan_cost
+= delta
;
6234 * Try and locate an idle core/thread in the LLC cache domain.
6236 static int select_idle_sibling(struct task_struct
*p
, int prev
, int target
)
6238 struct sched_domain
*sd
;
6241 if (idle_cpu(target
))
6245 * If the previous cpu is cache affine and idle, don't be stupid.
6247 if (prev
!= target
&& cpus_share_cache(prev
, target
) && idle_cpu(prev
))
6250 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
6254 i
= select_idle_core(p
, sd
, target
);
6255 if ((unsigned)i
< nr_cpumask_bits
)
6258 i
= select_idle_cpu(p
, sd
, target
);
6259 if ((unsigned)i
< nr_cpumask_bits
)
6262 i
= select_idle_smt(p
, sd
, target
);
6263 if ((unsigned)i
< nr_cpumask_bits
)
6270 * cpu_util returns the amount of capacity of a CPU that is used by CFS
6271 * tasks. The unit of the return value must be the one of capacity so we can
6272 * compare the utilization with the capacity of the CPU that is available for
6273 * CFS task (ie cpu_capacity).
6275 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6276 * recent utilization of currently non-runnable tasks on a CPU. It represents
6277 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6278 * capacity_orig is the cpu_capacity available at the highest frequency
6279 * (arch_scale_freq_capacity()).
6280 * The utilization of a CPU converges towards a sum equal to or less than the
6281 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6282 * the running time on this CPU scaled by capacity_curr.
6284 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6285 * higher than capacity_orig because of unfortunate rounding in
6286 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6287 * the average stabilizes with the new running time. We need to check that the
6288 * utilization stays within the range of [0..capacity_orig] and cap it if
6289 * necessary. Without utilization capping, a group could be seen as overloaded
6290 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6291 * available capacity. We allow utilization to overshoot capacity_curr (but not
6292 * capacity_orig) as it useful for predicting the capacity required after task
6293 * migrations (scheduler-driven DVFS).
6295 static int cpu_util(int cpu
)
6297 unsigned long util
= cpu_rq(cpu
)->cfs
.avg
.util_avg
;
6298 unsigned long capacity
= capacity_orig_of(cpu
);
6300 return (util
>= capacity
) ? capacity
: util
;
6303 static inline int task_util(struct task_struct
*p
)
6305 return p
->se
.avg
.util_avg
;
6309 * cpu_util_wake: Compute cpu utilization with any contributions from
6310 * the waking task p removed.
6312 static int cpu_util_wake(int cpu
, struct task_struct
*p
)
6314 unsigned long util
, capacity
;
6316 /* Task has no contribution or is new */
6317 if (cpu
!= task_cpu(p
) || !p
->se
.avg
.last_update_time
)
6318 return cpu_util(cpu
);
6320 capacity
= capacity_orig_of(cpu
);
6321 util
= max_t(long, cpu_rq(cpu
)->cfs
.avg
.util_avg
- task_util(p
), 0);
6323 return (util
>= capacity
) ? capacity
: util
;
6327 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6328 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6330 * In that case WAKE_AFFINE doesn't make sense and we'll let
6331 * BALANCE_WAKE sort things out.
6333 static int wake_cap(struct task_struct
*p
, int cpu
, int prev_cpu
)
6335 long min_cap
, max_cap
;
6337 min_cap
= min(capacity_orig_of(prev_cpu
), capacity_orig_of(cpu
));
6338 max_cap
= cpu_rq(cpu
)->rd
->max_cpu_capacity
;
6340 /* Minimum capacity is close to max, no need to abort wake_affine */
6341 if (max_cap
- min_cap
< max_cap
>> 3)
6344 /* Bring task utilization in sync with prev_cpu */
6345 sync_entity_load_avg(&p
->se
);
6347 return min_cap
* 1024 < task_util(p
) * capacity_margin
;
6351 * select_task_rq_fair: Select target runqueue for the waking task in domains
6352 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6353 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6355 * Balances load by selecting the idlest cpu in the idlest group, or under
6356 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6358 * Returns the target cpu number.
6360 * preempt must be disabled.
6363 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
6365 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
6366 int cpu
= smp_processor_id();
6367 int new_cpu
= prev_cpu
;
6368 int want_affine
= 0;
6369 int sync
= wake_flags
& WF_SYNC
;
6371 if (sd_flag
& SD_BALANCE_WAKE
) {
6373 want_affine
= !wake_wide(p
) && !wake_cap(p
, cpu
, prev_cpu
)
6374 && cpumask_test_cpu(cpu
, &p
->cpus_allowed
);
6378 for_each_domain(cpu
, tmp
) {
6379 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
6383 * If both cpu and prev_cpu are part of this domain,
6384 * cpu is a valid SD_WAKE_AFFINE target.
6386 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
6387 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
6392 if (tmp
->flags
& sd_flag
)
6394 else if (!want_affine
)
6399 sd
= NULL
; /* Prefer wake_affine over balance flags */
6400 if (cpu
== prev_cpu
)
6403 if (wake_affine(affine_sd
, p
, prev_cpu
, sync
))
6407 if (sd
&& !(sd_flag
& SD_BALANCE_FORK
)) {
6409 * We're going to need the task's util for capacity_spare_wake
6410 * in find_idlest_group. Sync it up to prev_cpu's
6413 sync_entity_load_avg(&p
->se
);
6418 if (sd_flag
& SD_BALANCE_WAKE
) /* XXX always ? */
6419 new_cpu
= select_idle_sibling(p
, prev_cpu
, new_cpu
);
6422 new_cpu
= find_idlest_cpu(sd
, p
, cpu
, prev_cpu
, sd_flag
);
6429 static void detach_entity_cfs_rq(struct sched_entity
*se
);
6432 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6433 * cfs_rq_of(p) references at time of call are still valid and identify the
6434 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6436 static void migrate_task_rq_fair(struct task_struct
*p
)
6439 * As blocked tasks retain absolute vruntime the migration needs to
6440 * deal with this by subtracting the old and adding the new
6441 * min_vruntime -- the latter is done by enqueue_entity() when placing
6442 * the task on the new runqueue.
6444 if (p
->state
== TASK_WAKING
) {
6445 struct sched_entity
*se
= &p
->se
;
6446 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6449 #ifndef CONFIG_64BIT
6450 u64 min_vruntime_copy
;
6453 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
6455 min_vruntime
= cfs_rq
->min_vruntime
;
6456 } while (min_vruntime
!= min_vruntime_copy
);
6458 min_vruntime
= cfs_rq
->min_vruntime
;
6461 se
->vruntime
-= min_vruntime
;
6464 if (p
->on_rq
== TASK_ON_RQ_MIGRATING
) {
6466 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6467 * rq->lock and can modify state directly.
6469 lockdep_assert_held(&task_rq(p
)->lock
);
6470 detach_entity_cfs_rq(&p
->se
);
6474 * We are supposed to update the task to "current" time, then
6475 * its up to date and ready to go to new CPU/cfs_rq. But we
6476 * have difficulty in getting what current time is, so simply
6477 * throw away the out-of-date time. This will result in the
6478 * wakee task is less decayed, but giving the wakee more load
6481 remove_entity_load_avg(&p
->se
);
6484 /* Tell new CPU we are migrated */
6485 p
->se
.avg
.last_update_time
= 0;
6487 /* We have migrated, no longer consider this task hot */
6488 p
->se
.exec_start
= 0;
6491 static void task_dead_fair(struct task_struct
*p
)
6493 remove_entity_load_avg(&p
->se
);
6495 #endif /* CONFIG_SMP */
6497 static unsigned long
6498 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
6500 unsigned long gran
= sysctl_sched_wakeup_granularity
;
6503 * Since its curr running now, convert the gran from real-time
6504 * to virtual-time in his units.
6506 * By using 'se' instead of 'curr' we penalize light tasks, so
6507 * they get preempted easier. That is, if 'se' < 'curr' then
6508 * the resulting gran will be larger, therefore penalizing the
6509 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6510 * be smaller, again penalizing the lighter task.
6512 * This is especially important for buddies when the leftmost
6513 * task is higher priority than the buddy.
6515 return calc_delta_fair(gran
, se
);
6519 * Should 'se' preempt 'curr'.
6533 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
6535 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
6540 gran
= wakeup_gran(curr
, se
);
6547 static void set_last_buddy(struct sched_entity
*se
)
6549 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
6552 for_each_sched_entity(se
) {
6553 if (SCHED_WARN_ON(!se
->on_rq
))
6555 cfs_rq_of(se
)->last
= se
;
6559 static void set_next_buddy(struct sched_entity
*se
)
6561 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
6564 for_each_sched_entity(se
) {
6565 if (SCHED_WARN_ON(!se
->on_rq
))
6567 cfs_rq_of(se
)->next
= se
;
6571 static void set_skip_buddy(struct sched_entity
*se
)
6573 for_each_sched_entity(se
)
6574 cfs_rq_of(se
)->skip
= se
;
6578 * Preempt the current task with a newly woken task if needed:
6580 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
6582 struct task_struct
*curr
= rq
->curr
;
6583 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
6584 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
6585 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
6586 int next_buddy_marked
= 0;
6588 if (unlikely(se
== pse
))
6592 * This is possible from callers such as attach_tasks(), in which we
6593 * unconditionally check_prempt_curr() after an enqueue (which may have
6594 * lead to a throttle). This both saves work and prevents false
6595 * next-buddy nomination below.
6597 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
6600 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
6601 set_next_buddy(pse
);
6602 next_buddy_marked
= 1;
6606 * We can come here with TIF_NEED_RESCHED already set from new task
6609 * Note: this also catches the edge-case of curr being in a throttled
6610 * group (e.g. via set_curr_task), since update_curr() (in the
6611 * enqueue of curr) will have resulted in resched being set. This
6612 * prevents us from potentially nominating it as a false LAST_BUDDY
6615 if (test_tsk_need_resched(curr
))
6618 /* Idle tasks are by definition preempted by non-idle tasks. */
6619 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
6620 likely(p
->policy
!= SCHED_IDLE
))
6624 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6625 * is driven by the tick):
6627 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
6630 find_matching_se(&se
, &pse
);
6631 update_curr(cfs_rq_of(se
));
6633 if (wakeup_preempt_entity(se
, pse
) == 1) {
6635 * Bias pick_next to pick the sched entity that is
6636 * triggering this preemption.
6638 if (!next_buddy_marked
)
6639 set_next_buddy(pse
);
6648 * Only set the backward buddy when the current task is still
6649 * on the rq. This can happen when a wakeup gets interleaved
6650 * with schedule on the ->pre_schedule() or idle_balance()
6651 * point, either of which can * drop the rq lock.
6653 * Also, during early boot the idle thread is in the fair class,
6654 * for obvious reasons its a bad idea to schedule back to it.
6656 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
6659 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
6663 static struct task_struct
*
6664 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
6666 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
6667 struct sched_entity
*se
;
6668 struct task_struct
*p
;
6672 if (!cfs_rq
->nr_running
)
6675 #ifdef CONFIG_FAIR_GROUP_SCHED
6676 if (prev
->sched_class
!= &fair_sched_class
)
6680 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6681 * likely that a next task is from the same cgroup as the current.
6683 * Therefore attempt to avoid putting and setting the entire cgroup
6684 * hierarchy, only change the part that actually changes.
6688 struct sched_entity
*curr
= cfs_rq
->curr
;
6691 * Since we got here without doing put_prev_entity() we also
6692 * have to consider cfs_rq->curr. If it is still a runnable
6693 * entity, update_curr() will update its vruntime, otherwise
6694 * forget we've ever seen it.
6698 update_curr(cfs_rq
);
6703 * This call to check_cfs_rq_runtime() will do the
6704 * throttle and dequeue its entity in the parent(s).
6705 * Therefore the nr_running test will indeed
6708 if (unlikely(check_cfs_rq_runtime(cfs_rq
))) {
6711 if (!cfs_rq
->nr_running
)
6718 se
= pick_next_entity(cfs_rq
, curr
);
6719 cfs_rq
= group_cfs_rq(se
);
6725 * Since we haven't yet done put_prev_entity and if the selected task
6726 * is a different task than we started out with, try and touch the
6727 * least amount of cfs_rqs.
6730 struct sched_entity
*pse
= &prev
->se
;
6732 while (!(cfs_rq
= is_same_group(se
, pse
))) {
6733 int se_depth
= se
->depth
;
6734 int pse_depth
= pse
->depth
;
6736 if (se_depth
<= pse_depth
) {
6737 put_prev_entity(cfs_rq_of(pse
), pse
);
6738 pse
= parent_entity(pse
);
6740 if (se_depth
>= pse_depth
) {
6741 set_next_entity(cfs_rq_of(se
), se
);
6742 se
= parent_entity(se
);
6746 put_prev_entity(cfs_rq
, pse
);
6747 set_next_entity(cfs_rq
, se
);
6754 put_prev_task(rq
, prev
);
6757 se
= pick_next_entity(cfs_rq
, NULL
);
6758 set_next_entity(cfs_rq
, se
);
6759 cfs_rq
= group_cfs_rq(se
);
6764 done
: __maybe_unused
6767 * Move the next running task to the front of
6768 * the list, so our cfs_tasks list becomes MRU
6771 list_move(&p
->se
.group_node
, &rq
->cfs_tasks
);
6774 if (hrtick_enabled(rq
))
6775 hrtick_start_fair(rq
, p
);
6780 new_tasks
= idle_balance(rq
, rf
);
6783 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6784 * possible for any higher priority task to appear. In that case we
6785 * must re-start the pick_next_entity() loop.
6797 * Account for a descheduled task:
6799 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
6801 struct sched_entity
*se
= &prev
->se
;
6802 struct cfs_rq
*cfs_rq
;
6804 for_each_sched_entity(se
) {
6805 cfs_rq
= cfs_rq_of(se
);
6806 put_prev_entity(cfs_rq
, se
);
6811 * sched_yield() is very simple
6813 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6815 static void yield_task_fair(struct rq
*rq
)
6817 struct task_struct
*curr
= rq
->curr
;
6818 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
6819 struct sched_entity
*se
= &curr
->se
;
6822 * Are we the only task in the tree?
6824 if (unlikely(rq
->nr_running
== 1))
6827 clear_buddies(cfs_rq
, se
);
6829 if (curr
->policy
!= SCHED_BATCH
) {
6830 update_rq_clock(rq
);
6832 * Update run-time statistics of the 'current'.
6834 update_curr(cfs_rq
);
6836 * Tell update_rq_clock() that we've just updated,
6837 * so we don't do microscopic update in schedule()
6838 * and double the fastpath cost.
6840 rq_clock_skip_update(rq
, true);
6846 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
6848 struct sched_entity
*se
= &p
->se
;
6850 /* throttled hierarchies are not runnable */
6851 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
6854 /* Tell the scheduler that we'd really like pse to run next. */
6857 yield_task_fair(rq
);
6863 /**************************************************
6864 * Fair scheduling class load-balancing methods.
6868 * The purpose of load-balancing is to achieve the same basic fairness the
6869 * per-cpu scheduler provides, namely provide a proportional amount of compute
6870 * time to each task. This is expressed in the following equation:
6872 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6874 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6875 * W_i,0 is defined as:
6877 * W_i,0 = \Sum_j w_i,j (2)
6879 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6880 * is derived from the nice value as per sched_prio_to_weight[].
6882 * The weight average is an exponential decay average of the instantaneous
6885 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6887 * C_i is the compute capacity of cpu i, typically it is the
6888 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6889 * can also include other factors [XXX].
6891 * To achieve this balance we define a measure of imbalance which follows
6892 * directly from (1):
6894 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6896 * We them move tasks around to minimize the imbalance. In the continuous
6897 * function space it is obvious this converges, in the discrete case we get
6898 * a few fun cases generally called infeasible weight scenarios.
6901 * - infeasible weights;
6902 * - local vs global optima in the discrete case. ]
6907 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6908 * for all i,j solution, we create a tree of cpus that follows the hardware
6909 * topology where each level pairs two lower groups (or better). This results
6910 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6911 * tree to only the first of the previous level and we decrease the frequency
6912 * of load-balance at each level inv. proportional to the number of cpus in
6918 * \Sum { --- * --- * 2^i } = O(n) (5)
6920 * `- size of each group
6921 * | | `- number of cpus doing load-balance
6923 * `- sum over all levels
6925 * Coupled with a limit on how many tasks we can migrate every balance pass,
6926 * this makes (5) the runtime complexity of the balancer.
6928 * An important property here is that each CPU is still (indirectly) connected
6929 * to every other cpu in at most O(log n) steps:
6931 * The adjacency matrix of the resulting graph is given by:
6934 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6937 * And you'll find that:
6939 * A^(log_2 n)_i,j != 0 for all i,j (7)
6941 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6942 * The task movement gives a factor of O(m), giving a convergence complexity
6945 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6950 * In order to avoid CPUs going idle while there's still work to do, new idle
6951 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6952 * tree itself instead of relying on other CPUs to bring it work.
6954 * This adds some complexity to both (5) and (8) but it reduces the total idle
6962 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6965 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6970 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6972 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6974 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6977 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6978 * rewrite all of this once again.]
6981 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
6983 enum fbq_type
{ regular
, remote
, all
};
6985 #define LBF_ALL_PINNED 0x01
6986 #define LBF_NEED_BREAK 0x02
6987 #define LBF_DST_PINNED 0x04
6988 #define LBF_SOME_PINNED 0x08
6991 struct sched_domain
*sd
;
6999 struct cpumask
*dst_grpmask
;
7001 enum cpu_idle_type idle
;
7003 /* The set of CPUs under consideration for load-balancing */
7004 struct cpumask
*cpus
;
7009 unsigned int loop_break
;
7010 unsigned int loop_max
;
7012 enum fbq_type fbq_type
;
7013 struct list_head tasks
;
7017 * Is this task likely cache-hot:
7019 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
7023 lockdep_assert_held(&env
->src_rq
->lock
);
7025 if (p
->sched_class
!= &fair_sched_class
)
7028 if (unlikely(p
->policy
== SCHED_IDLE
))
7032 * Buddy candidates are cache hot:
7034 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
7035 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
7036 &p
->se
== cfs_rq_of(&p
->se
)->last
))
7039 if (sysctl_sched_migration_cost
== -1)
7041 if (sysctl_sched_migration_cost
== 0)
7044 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
7046 return delta
< (s64
)sysctl_sched_migration_cost
;
7049 #ifdef CONFIG_NUMA_BALANCING
7051 * Returns 1, if task migration degrades locality
7052 * Returns 0, if task migration improves locality i.e migration preferred.
7053 * Returns -1, if task migration is not affected by locality.
7055 static int migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
7057 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
7058 unsigned long src_faults
, dst_faults
;
7059 int src_nid
, dst_nid
;
7061 if (!static_branch_likely(&sched_numa_balancing
))
7064 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
7067 src_nid
= cpu_to_node(env
->src_cpu
);
7068 dst_nid
= cpu_to_node(env
->dst_cpu
);
7070 if (src_nid
== dst_nid
)
7073 /* Migrating away from the preferred node is always bad. */
7074 if (src_nid
== p
->numa_preferred_nid
) {
7075 if (env
->src_rq
->nr_running
> env
->src_rq
->nr_preferred_running
)
7081 /* Encourage migration to the preferred node. */
7082 if (dst_nid
== p
->numa_preferred_nid
)
7085 /* Leaving a core idle is often worse than degrading locality. */
7086 if (env
->idle
!= CPU_NOT_IDLE
)
7090 src_faults
= group_faults(p
, src_nid
);
7091 dst_faults
= group_faults(p
, dst_nid
);
7093 src_faults
= task_faults(p
, src_nid
);
7094 dst_faults
= task_faults(p
, dst_nid
);
7097 return dst_faults
< src_faults
;
7101 static inline int migrate_degrades_locality(struct task_struct
*p
,
7109 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7112 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
7116 lockdep_assert_held(&env
->src_rq
->lock
);
7119 * We do not migrate tasks that are:
7120 * 1) throttled_lb_pair, or
7121 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7122 * 3) running (obviously), or
7123 * 4) are cache-hot on their current CPU.
7125 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
7128 if (!cpumask_test_cpu(env
->dst_cpu
, &p
->cpus_allowed
)) {
7131 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_affine
);
7133 env
->flags
|= LBF_SOME_PINNED
;
7136 * Remember if this task can be migrated to any other cpu in
7137 * our sched_group. We may want to revisit it if we couldn't
7138 * meet load balance goals by pulling other tasks on src_cpu.
7140 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7141 * already computed one in current iteration.
7143 if (env
->idle
== CPU_NEWLY_IDLE
|| (env
->flags
& LBF_DST_PINNED
))
7146 /* Prevent to re-select dst_cpu via env's cpus */
7147 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
7148 if (cpumask_test_cpu(cpu
, &p
->cpus_allowed
)) {
7149 env
->flags
|= LBF_DST_PINNED
;
7150 env
->new_dst_cpu
= cpu
;
7158 /* Record that we found atleast one task that could run on dst_cpu */
7159 env
->flags
&= ~LBF_ALL_PINNED
;
7161 if (task_running(env
->src_rq
, p
)) {
7162 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_running
);
7167 * Aggressive migration if:
7168 * 1) destination numa is preferred
7169 * 2) task is cache cold, or
7170 * 3) too many balance attempts have failed.
7172 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
7173 if (tsk_cache_hot
== -1)
7174 tsk_cache_hot
= task_hot(p
, env
);
7176 if (tsk_cache_hot
<= 0 ||
7177 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
7178 if (tsk_cache_hot
== 1) {
7179 schedstat_inc(env
->sd
->lb_hot_gained
[env
->idle
]);
7180 schedstat_inc(p
->se
.statistics
.nr_forced_migrations
);
7185 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_hot
);
7190 * detach_task() -- detach the task for the migration specified in env
7192 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
7194 lockdep_assert_held(&env
->src_rq
->lock
);
7196 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
7197 deactivate_task(env
->src_rq
, p
, DEQUEUE_NOCLOCK
);
7198 set_task_cpu(p
, env
->dst_cpu
);
7202 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7203 * part of active balancing operations within "domain".
7205 * Returns a task if successful and NULL otherwise.
7207 static struct task_struct
*detach_one_task(struct lb_env
*env
)
7209 struct task_struct
*p
;
7211 lockdep_assert_held(&env
->src_rq
->lock
);
7213 list_for_each_entry_reverse(p
,
7214 &env
->src_rq
->cfs_tasks
, se
.group_node
) {
7215 if (!can_migrate_task(p
, env
))
7218 detach_task(p
, env
);
7221 * Right now, this is only the second place where
7222 * lb_gained[env->idle] is updated (other is detach_tasks)
7223 * so we can safely collect stats here rather than
7224 * inside detach_tasks().
7226 schedstat_inc(env
->sd
->lb_gained
[env
->idle
]);
7232 static const unsigned int sched_nr_migrate_break
= 32;
7235 * detach_tasks() -- tries to detach up to imbalance weighted load from
7236 * busiest_rq, as part of a balancing operation within domain "sd".
7238 * Returns number of detached tasks if successful and 0 otherwise.
7240 static int detach_tasks(struct lb_env
*env
)
7242 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
7243 struct task_struct
*p
;
7247 lockdep_assert_held(&env
->src_rq
->lock
);
7249 if (env
->imbalance
<= 0)
7252 while (!list_empty(tasks
)) {
7254 * We don't want to steal all, otherwise we may be treated likewise,
7255 * which could at worst lead to a livelock crash.
7257 if (env
->idle
!= CPU_NOT_IDLE
&& env
->src_rq
->nr_running
<= 1)
7260 p
= list_last_entry(tasks
, struct task_struct
, se
.group_node
);
7263 /* We've more or less seen every task there is, call it quits */
7264 if (env
->loop
> env
->loop_max
)
7267 /* take a breather every nr_migrate tasks */
7268 if (env
->loop
> env
->loop_break
) {
7269 env
->loop_break
+= sched_nr_migrate_break
;
7270 env
->flags
|= LBF_NEED_BREAK
;
7274 if (!can_migrate_task(p
, env
))
7277 load
= task_h_load(p
);
7279 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
7282 if ((load
/ 2) > env
->imbalance
)
7285 detach_task(p
, env
);
7286 list_add(&p
->se
.group_node
, &env
->tasks
);
7289 env
->imbalance
-= load
;
7291 #ifdef CONFIG_PREEMPT
7293 * NEWIDLE balancing is a source of latency, so preemptible
7294 * kernels will stop after the first task is detached to minimize
7295 * the critical section.
7297 if (env
->idle
== CPU_NEWLY_IDLE
)
7302 * We only want to steal up to the prescribed amount of
7305 if (env
->imbalance
<= 0)
7310 list_move(&p
->se
.group_node
, tasks
);
7314 * Right now, this is one of only two places we collect this stat
7315 * so we can safely collect detach_one_task() stats here rather
7316 * than inside detach_one_task().
7318 schedstat_add(env
->sd
->lb_gained
[env
->idle
], detached
);
7324 * attach_task() -- attach the task detached by detach_task() to its new rq.
7326 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
7328 lockdep_assert_held(&rq
->lock
);
7330 BUG_ON(task_rq(p
) != rq
);
7331 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
7332 p
->on_rq
= TASK_ON_RQ_QUEUED
;
7333 check_preempt_curr(rq
, p
, 0);
7337 * attach_one_task() -- attaches the task returned from detach_one_task() to
7340 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
7345 update_rq_clock(rq
);
7351 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7354 static void attach_tasks(struct lb_env
*env
)
7356 struct list_head
*tasks
= &env
->tasks
;
7357 struct task_struct
*p
;
7360 rq_lock(env
->dst_rq
, &rf
);
7361 update_rq_clock(env
->dst_rq
);
7363 while (!list_empty(tasks
)) {
7364 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
7365 list_del_init(&p
->se
.group_node
);
7367 attach_task(env
->dst_rq
, p
);
7370 rq_unlock(env
->dst_rq
, &rf
);
7373 #ifdef CONFIG_FAIR_GROUP_SCHED
7375 static void update_blocked_averages(int cpu
)
7377 struct rq
*rq
= cpu_rq(cpu
);
7378 struct cfs_rq
*cfs_rq
;
7381 rq_lock_irqsave(rq
, &rf
);
7382 update_rq_clock(rq
);
7385 * Iterates the task_group tree in a bottom up fashion, see
7386 * list_add_leaf_cfs_rq() for details.
7388 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
7389 struct sched_entity
*se
;
7391 /* throttled entities do not contribute to load */
7392 if (throttled_hierarchy(cfs_rq
))
7395 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq
), cfs_rq
))
7396 update_tg_load_avg(cfs_rq
, 0);
7398 /* Propagate pending load changes to the parent, if any: */
7399 se
= cfs_rq
->tg
->se
[cpu
];
7400 if (se
&& !skip_blocked_update(se
))
7401 update_load_avg(cfs_rq_of(se
), se
, 0);
7403 rq_unlock_irqrestore(rq
, &rf
);
7407 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7408 * This needs to be done in a top-down fashion because the load of a child
7409 * group is a fraction of its parents load.
7411 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
7413 struct rq
*rq
= rq_of(cfs_rq
);
7414 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
7415 unsigned long now
= jiffies
;
7418 if (cfs_rq
->last_h_load_update
== now
)
7421 WRITE_ONCE(cfs_rq
->h_load_next
, NULL
);
7422 for_each_sched_entity(se
) {
7423 cfs_rq
= cfs_rq_of(se
);
7424 WRITE_ONCE(cfs_rq
->h_load_next
, se
);
7425 if (cfs_rq
->last_h_load_update
== now
)
7430 cfs_rq
->h_load
= cfs_rq_load_avg(cfs_rq
);
7431 cfs_rq
->last_h_load_update
= now
;
7434 while ((se
= READ_ONCE(cfs_rq
->h_load_next
)) != NULL
) {
7435 load
= cfs_rq
->h_load
;
7436 load
= div64_ul(load
* se
->avg
.load_avg
,
7437 cfs_rq_load_avg(cfs_rq
) + 1);
7438 cfs_rq
= group_cfs_rq(se
);
7439 cfs_rq
->h_load
= load
;
7440 cfs_rq
->last_h_load_update
= now
;
7444 static unsigned long task_h_load(struct task_struct
*p
)
7446 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
7448 update_cfs_rq_h_load(cfs_rq
);
7449 return div64_ul(p
->se
.avg
.load_avg
* cfs_rq
->h_load
,
7450 cfs_rq_load_avg(cfs_rq
) + 1);
7453 static inline void update_blocked_averages(int cpu
)
7455 struct rq
*rq
= cpu_rq(cpu
);
7456 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
7459 rq_lock_irqsave(rq
, &rf
);
7460 update_rq_clock(rq
);
7461 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq
), cfs_rq
);
7462 rq_unlock_irqrestore(rq
, &rf
);
7465 static unsigned long task_h_load(struct task_struct
*p
)
7467 return p
->se
.avg
.load_avg
;
7471 /********** Helpers for find_busiest_group ************************/
7480 * sg_lb_stats - stats of a sched_group required for load_balancing
7482 struct sg_lb_stats
{
7483 unsigned long avg_load
; /*Avg load across the CPUs of the group */
7484 unsigned long group_load
; /* Total load over the CPUs of the group */
7485 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
7486 unsigned long load_per_task
;
7487 unsigned long group_capacity
;
7488 unsigned long group_util
; /* Total utilization of the group */
7489 unsigned int sum_nr_running
; /* Nr tasks running in the group */
7490 unsigned int idle_cpus
;
7491 unsigned int group_weight
;
7492 enum group_type group_type
;
7493 int group_no_capacity
;
7494 #ifdef CONFIG_NUMA_BALANCING
7495 unsigned int nr_numa_running
;
7496 unsigned int nr_preferred_running
;
7501 * sd_lb_stats - Structure to store the statistics of a sched_domain
7502 * during load balancing.
7504 struct sd_lb_stats
{
7505 struct sched_group
*busiest
; /* Busiest group in this sd */
7506 struct sched_group
*local
; /* Local group in this sd */
7507 unsigned long total_running
;
7508 unsigned long total_load
; /* Total load of all groups in sd */
7509 unsigned long total_capacity
; /* Total capacity of all groups in sd */
7510 unsigned long avg_load
; /* Average load across all groups in sd */
7512 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
7513 struct sg_lb_stats local_stat
; /* Statistics of the local group */
7516 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
7519 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7520 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7521 * We must however clear busiest_stat::avg_load because
7522 * update_sd_pick_busiest() reads this before assignment.
7524 *sds
= (struct sd_lb_stats
){
7527 .total_running
= 0UL,
7529 .total_capacity
= 0UL,
7532 .sum_nr_running
= 0,
7533 .group_type
= group_other
,
7539 * get_sd_load_idx - Obtain the load index for a given sched domain.
7540 * @sd: The sched_domain whose load_idx is to be obtained.
7541 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7543 * Return: The load index.
7545 static inline int get_sd_load_idx(struct sched_domain
*sd
,
7546 enum cpu_idle_type idle
)
7552 load_idx
= sd
->busy_idx
;
7555 case CPU_NEWLY_IDLE
:
7556 load_idx
= sd
->newidle_idx
;
7559 load_idx
= sd
->idle_idx
;
7566 static unsigned long scale_rt_capacity(int cpu
)
7568 struct rq
*rq
= cpu_rq(cpu
);
7569 u64 total
, used
, age_stamp
, avg
;
7573 * Since we're reading these variables without serialization make sure
7574 * we read them once before doing sanity checks on them.
7576 age_stamp
= READ_ONCE(rq
->age_stamp
);
7577 avg
= READ_ONCE(rq
->rt_avg
);
7578 delta
= __rq_clock_broken(rq
) - age_stamp
;
7580 if (unlikely(delta
< 0))
7583 total
= sched_avg_period() + delta
;
7585 used
= div_u64(avg
, total
);
7587 if (likely(used
< SCHED_CAPACITY_SCALE
))
7588 return SCHED_CAPACITY_SCALE
- used
;
7593 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
7595 unsigned long capacity
= arch_scale_cpu_capacity(sd
, cpu
);
7596 struct sched_group
*sdg
= sd
->groups
;
7598 cpu_rq(cpu
)->cpu_capacity_orig
= capacity
;
7600 capacity
*= scale_rt_capacity(cpu
);
7601 capacity
>>= SCHED_CAPACITY_SHIFT
;
7606 cpu_rq(cpu
)->cpu_capacity
= capacity
;
7607 sdg
->sgc
->capacity
= capacity
;
7608 sdg
->sgc
->min_capacity
= capacity
;
7611 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
7613 struct sched_domain
*child
= sd
->child
;
7614 struct sched_group
*group
, *sdg
= sd
->groups
;
7615 unsigned long capacity
, min_capacity
;
7616 unsigned long interval
;
7618 interval
= msecs_to_jiffies(sd
->balance_interval
);
7619 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
7620 sdg
->sgc
->next_update
= jiffies
+ interval
;
7623 update_cpu_capacity(sd
, cpu
);
7628 min_capacity
= ULONG_MAX
;
7630 if (child
->flags
& SD_OVERLAP
) {
7632 * SD_OVERLAP domains cannot assume that child groups
7633 * span the current group.
7636 for_each_cpu(cpu
, sched_group_span(sdg
)) {
7637 struct sched_group_capacity
*sgc
;
7638 struct rq
*rq
= cpu_rq(cpu
);
7641 * build_sched_domains() -> init_sched_groups_capacity()
7642 * gets here before we've attached the domains to the
7645 * Use capacity_of(), which is set irrespective of domains
7646 * in update_cpu_capacity().
7648 * This avoids capacity from being 0 and
7649 * causing divide-by-zero issues on boot.
7651 if (unlikely(!rq
->sd
)) {
7652 capacity
+= capacity_of(cpu
);
7654 sgc
= rq
->sd
->groups
->sgc
;
7655 capacity
+= sgc
->capacity
;
7658 min_capacity
= min(capacity
, min_capacity
);
7662 * !SD_OVERLAP domains can assume that child groups
7663 * span the current group.
7666 group
= child
->groups
;
7668 struct sched_group_capacity
*sgc
= group
->sgc
;
7670 capacity
+= sgc
->capacity
;
7671 min_capacity
= min(sgc
->min_capacity
, min_capacity
);
7672 group
= group
->next
;
7673 } while (group
!= child
->groups
);
7676 sdg
->sgc
->capacity
= capacity
;
7677 sdg
->sgc
->min_capacity
= min_capacity
;
7681 * Check whether the capacity of the rq has been noticeably reduced by side
7682 * activity. The imbalance_pct is used for the threshold.
7683 * Return true is the capacity is reduced
7686 check_cpu_capacity(struct rq
*rq
, struct sched_domain
*sd
)
7688 return ((rq
->cpu_capacity
* sd
->imbalance_pct
) <
7689 (rq
->cpu_capacity_orig
* 100));
7693 * Group imbalance indicates (and tries to solve) the problem where balancing
7694 * groups is inadequate due to ->cpus_allowed constraints.
7696 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7697 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7700 * { 0 1 2 3 } { 4 5 6 7 }
7703 * If we were to balance group-wise we'd place two tasks in the first group and
7704 * two tasks in the second group. Clearly this is undesired as it will overload
7705 * cpu 3 and leave one of the cpus in the second group unused.
7707 * The current solution to this issue is detecting the skew in the first group
7708 * by noticing the lower domain failed to reach balance and had difficulty
7709 * moving tasks due to affinity constraints.
7711 * When this is so detected; this group becomes a candidate for busiest; see
7712 * update_sd_pick_busiest(). And calculate_imbalance() and
7713 * find_busiest_group() avoid some of the usual balance conditions to allow it
7714 * to create an effective group imbalance.
7716 * This is a somewhat tricky proposition since the next run might not find the
7717 * group imbalance and decide the groups need to be balanced again. A most
7718 * subtle and fragile situation.
7721 static inline int sg_imbalanced(struct sched_group
*group
)
7723 return group
->sgc
->imbalance
;
7727 * group_has_capacity returns true if the group has spare capacity that could
7728 * be used by some tasks.
7729 * We consider that a group has spare capacity if the * number of task is
7730 * smaller than the number of CPUs or if the utilization is lower than the
7731 * available capacity for CFS tasks.
7732 * For the latter, we use a threshold to stabilize the state, to take into
7733 * account the variance of the tasks' load and to return true if the available
7734 * capacity in meaningful for the load balancer.
7735 * As an example, an available capacity of 1% can appear but it doesn't make
7736 * any benefit for the load balance.
7739 group_has_capacity(struct lb_env
*env
, struct sg_lb_stats
*sgs
)
7741 if (sgs
->sum_nr_running
< sgs
->group_weight
)
7744 if ((sgs
->group_capacity
* 100) >
7745 (sgs
->group_util
* env
->sd
->imbalance_pct
))
7752 * group_is_overloaded returns true if the group has more tasks than it can
7754 * group_is_overloaded is not equals to !group_has_capacity because a group
7755 * with the exact right number of tasks, has no more spare capacity but is not
7756 * overloaded so both group_has_capacity and group_is_overloaded return
7760 group_is_overloaded(struct lb_env
*env
, struct sg_lb_stats
*sgs
)
7762 if (sgs
->sum_nr_running
<= sgs
->group_weight
)
7765 if ((sgs
->group_capacity
* 100) <
7766 (sgs
->group_util
* env
->sd
->imbalance_pct
))
7773 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7774 * per-CPU capacity than sched_group ref.
7777 group_smaller_cpu_capacity(struct sched_group
*sg
, struct sched_group
*ref
)
7779 return sg
->sgc
->min_capacity
* capacity_margin
<
7780 ref
->sgc
->min_capacity
* 1024;
7784 group_type
group_classify(struct sched_group
*group
,
7785 struct sg_lb_stats
*sgs
)
7787 if (sgs
->group_no_capacity
)
7788 return group_overloaded
;
7790 if (sg_imbalanced(group
))
7791 return group_imbalanced
;
7797 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7798 * @env: The load balancing environment.
7799 * @group: sched_group whose statistics are to be updated.
7800 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7801 * @local_group: Does group contain this_cpu.
7802 * @sgs: variable to hold the statistics for this group.
7803 * @overload: Indicate more than one runnable task for any CPU.
7805 static inline void update_sg_lb_stats(struct lb_env
*env
,
7806 struct sched_group
*group
, int load_idx
,
7807 int local_group
, struct sg_lb_stats
*sgs
,
7813 memset(sgs
, 0, sizeof(*sgs
));
7815 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
7816 struct rq
*rq
= cpu_rq(i
);
7818 /* Bias balancing toward cpus of our domain */
7820 load
= target_load(i
, load_idx
);
7822 load
= source_load(i
, load_idx
);
7824 sgs
->group_load
+= load
;
7825 sgs
->group_util
+= cpu_util(i
);
7826 sgs
->sum_nr_running
+= rq
->cfs
.h_nr_running
;
7828 nr_running
= rq
->nr_running
;
7832 #ifdef CONFIG_NUMA_BALANCING
7833 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
7834 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
7836 sgs
->sum_weighted_load
+= weighted_cpuload(rq
);
7838 * No need to call idle_cpu() if nr_running is not 0
7840 if (!nr_running
&& idle_cpu(i
))
7844 /* Adjust by relative CPU capacity of the group */
7845 sgs
->group_capacity
= group
->sgc
->capacity
;
7846 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
7848 if (sgs
->sum_nr_running
)
7849 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
7851 sgs
->group_weight
= group
->group_weight
;
7853 sgs
->group_no_capacity
= group_is_overloaded(env
, sgs
);
7854 sgs
->group_type
= group_classify(group
, sgs
);
7858 * update_sd_pick_busiest - return 1 on busiest group
7859 * @env: The load balancing environment.
7860 * @sds: sched_domain statistics
7861 * @sg: sched_group candidate to be checked for being the busiest
7862 * @sgs: sched_group statistics
7864 * Determine if @sg is a busier group than the previously selected
7867 * Return: %true if @sg is a busier group than the previously selected
7868 * busiest group. %false otherwise.
7870 static bool update_sd_pick_busiest(struct lb_env
*env
,
7871 struct sd_lb_stats
*sds
,
7872 struct sched_group
*sg
,
7873 struct sg_lb_stats
*sgs
)
7875 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
7877 if (sgs
->group_type
> busiest
->group_type
)
7880 if (sgs
->group_type
< busiest
->group_type
)
7883 if (sgs
->avg_load
<= busiest
->avg_load
)
7886 if (!(env
->sd
->flags
& SD_ASYM_CPUCAPACITY
))
7890 * Candidate sg has no more than one task per CPU and
7891 * has higher per-CPU capacity. Migrating tasks to less
7892 * capable CPUs may harm throughput. Maximize throughput,
7893 * power/energy consequences are not considered.
7895 if (sgs
->sum_nr_running
<= sgs
->group_weight
&&
7896 group_smaller_cpu_capacity(sds
->local
, sg
))
7900 /* This is the busiest node in its class. */
7901 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
7904 /* No ASYM_PACKING if target cpu is already busy */
7905 if (env
->idle
== CPU_NOT_IDLE
)
7908 * ASYM_PACKING needs to move all the work to the highest
7909 * prority CPUs in the group, therefore mark all groups
7910 * of lower priority than ourself as busy.
7912 if (sgs
->sum_nr_running
&&
7913 sched_asym_prefer(env
->dst_cpu
, sg
->asym_prefer_cpu
)) {
7917 /* Prefer to move from lowest priority cpu's work */
7918 if (sched_asym_prefer(sds
->busiest
->asym_prefer_cpu
,
7919 sg
->asym_prefer_cpu
))
7926 #ifdef CONFIG_NUMA_BALANCING
7927 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
7929 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
7931 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
7936 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
7938 if (rq
->nr_running
> rq
->nr_numa_running
)
7940 if (rq
->nr_running
> rq
->nr_preferred_running
)
7945 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
7950 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
7954 #endif /* CONFIG_NUMA_BALANCING */
7957 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7958 * @env: The load balancing environment.
7959 * @sds: variable to hold the statistics for this sched_domain.
7961 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
7963 struct sched_domain
*child
= env
->sd
->child
;
7964 struct sched_group
*sg
= env
->sd
->groups
;
7965 struct sg_lb_stats
*local
= &sds
->local_stat
;
7966 struct sg_lb_stats tmp_sgs
;
7967 int load_idx
, prefer_sibling
= 0;
7968 bool overload
= false;
7970 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
7973 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
7976 struct sg_lb_stats
*sgs
= &tmp_sgs
;
7979 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_span(sg
));
7984 if (env
->idle
!= CPU_NEWLY_IDLE
||
7985 time_after_eq(jiffies
, sg
->sgc
->next_update
))
7986 update_group_capacity(env
->sd
, env
->dst_cpu
);
7989 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
,
7996 * In case the child domain prefers tasks go to siblings
7997 * first, lower the sg capacity so that we'll try
7998 * and move all the excess tasks away. We lower the capacity
7999 * of a group only if the local group has the capacity to fit
8000 * these excess tasks. The extra check prevents the case where
8001 * you always pull from the heaviest group when it is already
8002 * under-utilized (possible with a large weight task outweighs
8003 * the tasks on the system).
8005 if (prefer_sibling
&& sds
->local
&&
8006 group_has_capacity(env
, local
) &&
8007 (sgs
->sum_nr_running
> local
->sum_nr_running
+ 1)) {
8008 sgs
->group_no_capacity
= 1;
8009 sgs
->group_type
= group_classify(sg
, sgs
);
8012 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
8014 sds
->busiest_stat
= *sgs
;
8018 /* Now, start updating sd_lb_stats */
8019 sds
->total_running
+= sgs
->sum_nr_running
;
8020 sds
->total_load
+= sgs
->group_load
;
8021 sds
->total_capacity
+= sgs
->group_capacity
;
8024 } while (sg
!= env
->sd
->groups
);
8026 if (env
->sd
->flags
& SD_NUMA
)
8027 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
8029 if (!env
->sd
->parent
) {
8030 /* update overload indicator if we are at root domain */
8031 if (env
->dst_rq
->rd
->overload
!= overload
)
8032 env
->dst_rq
->rd
->overload
= overload
;
8037 * check_asym_packing - Check to see if the group is packed into the
8040 * This is primarily intended to used at the sibling level. Some
8041 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8042 * case of POWER7, it can move to lower SMT modes only when higher
8043 * threads are idle. When in lower SMT modes, the threads will
8044 * perform better since they share less core resources. Hence when we
8045 * have idle threads, we want them to be the higher ones.
8047 * This packing function is run on idle threads. It checks to see if
8048 * the busiest CPU in this domain (core in the P7 case) has a higher
8049 * CPU number than the packing function is being run on. Here we are
8050 * assuming lower CPU number will be equivalent to lower a SMT thread
8053 * Return: 1 when packing is required and a task should be moved to
8054 * this CPU. The amount of the imbalance is returned in env->imbalance.
8056 * @env: The load balancing environment.
8057 * @sds: Statistics of the sched_domain which is to be packed
8059 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
8063 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
8066 if (env
->idle
== CPU_NOT_IDLE
)
8072 busiest_cpu
= sds
->busiest
->asym_prefer_cpu
;
8073 if (sched_asym_prefer(busiest_cpu
, env
->dst_cpu
))
8076 env
->imbalance
= DIV_ROUND_CLOSEST(
8077 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
8078 SCHED_CAPACITY_SCALE
);
8084 * fix_small_imbalance - Calculate the minor imbalance that exists
8085 * amongst the groups of a sched_domain, during
8087 * @env: The load balancing environment.
8088 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8091 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
8093 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
8094 unsigned int imbn
= 2;
8095 unsigned long scaled_busy_load_per_task
;
8096 struct sg_lb_stats
*local
, *busiest
;
8098 local
= &sds
->local_stat
;
8099 busiest
= &sds
->busiest_stat
;
8101 if (!local
->sum_nr_running
)
8102 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
8103 else if (busiest
->load_per_task
> local
->load_per_task
)
8106 scaled_busy_load_per_task
=
8107 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
8108 busiest
->group_capacity
;
8110 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
8111 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
8112 env
->imbalance
= busiest
->load_per_task
;
8117 * OK, we don't have enough imbalance to justify moving tasks,
8118 * however we may be able to increase total CPU capacity used by
8122 capa_now
+= busiest
->group_capacity
*
8123 min(busiest
->load_per_task
, busiest
->avg_load
);
8124 capa_now
+= local
->group_capacity
*
8125 min(local
->load_per_task
, local
->avg_load
);
8126 capa_now
/= SCHED_CAPACITY_SCALE
;
8128 /* Amount of load we'd subtract */
8129 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
8130 capa_move
+= busiest
->group_capacity
*
8131 min(busiest
->load_per_task
,
8132 busiest
->avg_load
- scaled_busy_load_per_task
);
8135 /* Amount of load we'd add */
8136 if (busiest
->avg_load
* busiest
->group_capacity
<
8137 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
8138 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
8139 local
->group_capacity
;
8141 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
8142 local
->group_capacity
;
8144 capa_move
+= local
->group_capacity
*
8145 min(local
->load_per_task
, local
->avg_load
+ tmp
);
8146 capa_move
/= SCHED_CAPACITY_SCALE
;
8148 /* Move if we gain throughput */
8149 if (capa_move
> capa_now
)
8150 env
->imbalance
= busiest
->load_per_task
;
8154 * calculate_imbalance - Calculate the amount of imbalance present within the
8155 * groups of a given sched_domain during load balance.
8156 * @env: load balance environment
8157 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8159 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
8161 unsigned long max_pull
, load_above_capacity
= ~0UL;
8162 struct sg_lb_stats
*local
, *busiest
;
8164 local
= &sds
->local_stat
;
8165 busiest
= &sds
->busiest_stat
;
8167 if (busiest
->group_type
== group_imbalanced
) {
8169 * In the group_imb case we cannot rely on group-wide averages
8170 * to ensure cpu-load equilibrium, look at wider averages. XXX
8172 busiest
->load_per_task
=
8173 min(busiest
->load_per_task
, sds
->avg_load
);
8177 * Avg load of busiest sg can be less and avg load of local sg can
8178 * be greater than avg load across all sgs of sd because avg load
8179 * factors in sg capacity and sgs with smaller group_type are
8180 * skipped when updating the busiest sg:
8182 if (busiest
->avg_load
<= sds
->avg_load
||
8183 local
->avg_load
>= sds
->avg_load
) {
8185 return fix_small_imbalance(env
, sds
);
8189 * If there aren't any idle cpus, avoid creating some.
8191 if (busiest
->group_type
== group_overloaded
&&
8192 local
->group_type
== group_overloaded
) {
8193 load_above_capacity
= busiest
->sum_nr_running
* SCHED_CAPACITY_SCALE
;
8194 if (load_above_capacity
> busiest
->group_capacity
) {
8195 load_above_capacity
-= busiest
->group_capacity
;
8196 load_above_capacity
*= scale_load_down(NICE_0_LOAD
);
8197 load_above_capacity
/= busiest
->group_capacity
;
8199 load_above_capacity
= ~0UL;
8203 * We're trying to get all the cpus to the average_load, so we don't
8204 * want to push ourselves above the average load, nor do we wish to
8205 * reduce the max loaded cpu below the average load. At the same time,
8206 * we also don't want to reduce the group load below the group
8207 * capacity. Thus we look for the minimum possible imbalance.
8209 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
8211 /* How much load to actually move to equalise the imbalance */
8212 env
->imbalance
= min(
8213 max_pull
* busiest
->group_capacity
,
8214 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
8215 ) / SCHED_CAPACITY_SCALE
;
8218 * if *imbalance is less than the average load per runnable task
8219 * there is no guarantee that any tasks will be moved so we'll have
8220 * a think about bumping its value to force at least one task to be
8223 if (env
->imbalance
< busiest
->load_per_task
)
8224 return fix_small_imbalance(env
, sds
);
8227 /******* find_busiest_group() helpers end here *********************/
8230 * find_busiest_group - Returns the busiest group within the sched_domain
8231 * if there is an imbalance.
8233 * Also calculates the amount of weighted load which should be moved
8234 * to restore balance.
8236 * @env: The load balancing environment.
8238 * Return: - The busiest group if imbalance exists.
8240 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
8242 struct sg_lb_stats
*local
, *busiest
;
8243 struct sd_lb_stats sds
;
8245 init_sd_lb_stats(&sds
);
8248 * Compute the various statistics relavent for load balancing at
8251 update_sd_lb_stats(env
, &sds
);
8252 local
= &sds
.local_stat
;
8253 busiest
= &sds
.busiest_stat
;
8255 /* ASYM feature bypasses nice load balance check */
8256 if (check_asym_packing(env
, &sds
))
8259 /* There is no busy sibling group to pull tasks from */
8260 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
8263 /* XXX broken for overlapping NUMA groups */
8264 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
8265 / sds
.total_capacity
;
8268 * If the busiest group is imbalanced the below checks don't
8269 * work because they assume all things are equal, which typically
8270 * isn't true due to cpus_allowed constraints and the like.
8272 if (busiest
->group_type
== group_imbalanced
)
8276 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8277 * capacities from resulting in underutilization due to avg_load.
8279 if (env
->idle
!= CPU_NOT_IDLE
&& group_has_capacity(env
, local
) &&
8280 busiest
->group_no_capacity
)
8284 * If the local group is busier than the selected busiest group
8285 * don't try and pull any tasks.
8287 if (local
->avg_load
>= busiest
->avg_load
)
8291 * Don't pull any tasks if this group is already above the domain
8294 if (local
->avg_load
>= sds
.avg_load
)
8297 if (env
->idle
== CPU_IDLE
) {
8299 * This cpu is idle. If the busiest group is not overloaded
8300 * and there is no imbalance between this and busiest group
8301 * wrt idle cpus, it is balanced. The imbalance becomes
8302 * significant if the diff is greater than 1 otherwise we
8303 * might end up to just move the imbalance on another group
8305 if ((busiest
->group_type
!= group_overloaded
) &&
8306 (local
->idle_cpus
<= (busiest
->idle_cpus
+ 1)))
8310 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8311 * imbalance_pct to be conservative.
8313 if (100 * busiest
->avg_load
<=
8314 env
->sd
->imbalance_pct
* local
->avg_load
)
8319 /* Looks like there is an imbalance. Compute it */
8320 calculate_imbalance(env
, &sds
);
8329 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8331 static struct rq
*find_busiest_queue(struct lb_env
*env
,
8332 struct sched_group
*group
)
8334 struct rq
*busiest
= NULL
, *rq
;
8335 unsigned long busiest_load
= 0, busiest_capacity
= 1;
8338 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
8339 unsigned long capacity
, wl
;
8343 rt
= fbq_classify_rq(rq
);
8346 * We classify groups/runqueues into three groups:
8347 * - regular: there are !numa tasks
8348 * - remote: there are numa tasks that run on the 'wrong' node
8349 * - all: there is no distinction
8351 * In order to avoid migrating ideally placed numa tasks,
8352 * ignore those when there's better options.
8354 * If we ignore the actual busiest queue to migrate another
8355 * task, the next balance pass can still reduce the busiest
8356 * queue by moving tasks around inside the node.
8358 * If we cannot move enough load due to this classification
8359 * the next pass will adjust the group classification and
8360 * allow migration of more tasks.
8362 * Both cases only affect the total convergence complexity.
8364 if (rt
> env
->fbq_type
)
8367 capacity
= capacity_of(i
);
8369 wl
= weighted_cpuload(rq
);
8372 * When comparing with imbalance, use weighted_cpuload()
8373 * which is not scaled with the cpu capacity.
8376 if (rq
->nr_running
== 1 && wl
> env
->imbalance
&&
8377 !check_cpu_capacity(rq
, env
->sd
))
8381 * For the load comparisons with the other cpu's, consider
8382 * the weighted_cpuload() scaled with the cpu capacity, so
8383 * that the load can be moved away from the cpu that is
8384 * potentially running at a lower capacity.
8386 * Thus we're looking for max(wl_i / capacity_i), crosswise
8387 * multiplication to rid ourselves of the division works out
8388 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8389 * our previous maximum.
8391 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
8393 busiest_capacity
= capacity
;
8402 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8403 * so long as it is large enough.
8405 #define MAX_PINNED_INTERVAL 512
8407 static int need_active_balance(struct lb_env
*env
)
8409 struct sched_domain
*sd
= env
->sd
;
8411 if (env
->idle
== CPU_NEWLY_IDLE
) {
8414 * ASYM_PACKING needs to force migrate tasks from busy but
8415 * lower priority CPUs in order to pack all tasks in the
8416 * highest priority CPUs.
8418 if ((sd
->flags
& SD_ASYM_PACKING
) &&
8419 sched_asym_prefer(env
->dst_cpu
, env
->src_cpu
))
8424 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8425 * It's worth migrating the task if the src_cpu's capacity is reduced
8426 * because of other sched_class or IRQs if more capacity stays
8427 * available on dst_cpu.
8429 if ((env
->idle
!= CPU_NOT_IDLE
) &&
8430 (env
->src_rq
->cfs
.h_nr_running
== 1)) {
8431 if ((check_cpu_capacity(env
->src_rq
, sd
)) &&
8432 (capacity_of(env
->src_cpu
)*sd
->imbalance_pct
< capacity_of(env
->dst_cpu
)*100))
8436 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
8439 static int active_load_balance_cpu_stop(void *data
);
8441 static int should_we_balance(struct lb_env
*env
)
8443 struct sched_group
*sg
= env
->sd
->groups
;
8444 int cpu
, balance_cpu
= -1;
8447 * Ensure the balancing environment is consistent; can happen
8448 * when the softirq triggers 'during' hotplug.
8450 if (!cpumask_test_cpu(env
->dst_cpu
, env
->cpus
))
8454 * In the newly idle case, we will allow all the cpu's
8455 * to do the newly idle load balance.
8457 if (env
->idle
== CPU_NEWLY_IDLE
)
8460 /* Try to find first idle cpu */
8461 for_each_cpu_and(cpu
, group_balance_mask(sg
), env
->cpus
) {
8469 if (balance_cpu
== -1)
8470 balance_cpu
= group_balance_cpu(sg
);
8473 * First idle cpu or the first cpu(busiest) in this sched group
8474 * is eligible for doing load balancing at this and above domains.
8476 return balance_cpu
== env
->dst_cpu
;
8480 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8481 * tasks if there is an imbalance.
8483 static int load_balance(int this_cpu
, struct rq
*this_rq
,
8484 struct sched_domain
*sd
, enum cpu_idle_type idle
,
8485 int *continue_balancing
)
8487 int ld_moved
, cur_ld_moved
, active_balance
= 0;
8488 struct sched_domain
*sd_parent
= sd
->parent
;
8489 struct sched_group
*group
;
8492 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
8494 struct lb_env env
= {
8496 .dst_cpu
= this_cpu
,
8498 .dst_grpmask
= sched_group_span(sd
->groups
),
8500 .loop_break
= sched_nr_migrate_break
,
8503 .tasks
= LIST_HEAD_INIT(env
.tasks
),
8506 cpumask_and(cpus
, sched_domain_span(sd
), cpu_active_mask
);
8508 schedstat_inc(sd
->lb_count
[idle
]);
8511 if (!should_we_balance(&env
)) {
8512 *continue_balancing
= 0;
8516 group
= find_busiest_group(&env
);
8518 schedstat_inc(sd
->lb_nobusyg
[idle
]);
8522 busiest
= find_busiest_queue(&env
, group
);
8524 schedstat_inc(sd
->lb_nobusyq
[idle
]);
8528 BUG_ON(busiest
== env
.dst_rq
);
8530 schedstat_add(sd
->lb_imbalance
[idle
], env
.imbalance
);
8532 env
.src_cpu
= busiest
->cpu
;
8533 env
.src_rq
= busiest
;
8536 if (busiest
->nr_running
> 1) {
8538 * Attempt to move tasks. If find_busiest_group has found
8539 * an imbalance but busiest->nr_running <= 1, the group is
8540 * still unbalanced. ld_moved simply stays zero, so it is
8541 * correctly treated as an imbalance.
8543 env
.flags
|= LBF_ALL_PINNED
;
8544 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
8547 rq_lock_irqsave(busiest
, &rf
);
8548 update_rq_clock(busiest
);
8551 * cur_ld_moved - load moved in current iteration
8552 * ld_moved - cumulative load moved across iterations
8554 cur_ld_moved
= detach_tasks(&env
);
8557 * We've detached some tasks from busiest_rq. Every
8558 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8559 * unlock busiest->lock, and we are able to be sure
8560 * that nobody can manipulate the tasks in parallel.
8561 * See task_rq_lock() family for the details.
8564 rq_unlock(busiest
, &rf
);
8568 ld_moved
+= cur_ld_moved
;
8571 local_irq_restore(rf
.flags
);
8573 if (env
.flags
& LBF_NEED_BREAK
) {
8574 env
.flags
&= ~LBF_NEED_BREAK
;
8579 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8580 * us and move them to an alternate dst_cpu in our sched_group
8581 * where they can run. The upper limit on how many times we
8582 * iterate on same src_cpu is dependent on number of cpus in our
8585 * This changes load balance semantics a bit on who can move
8586 * load to a given_cpu. In addition to the given_cpu itself
8587 * (or a ilb_cpu acting on its behalf where given_cpu is
8588 * nohz-idle), we now have balance_cpu in a position to move
8589 * load to given_cpu. In rare situations, this may cause
8590 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8591 * _independently_ and at _same_ time to move some load to
8592 * given_cpu) causing exceess load to be moved to given_cpu.
8593 * This however should not happen so much in practice and
8594 * moreover subsequent load balance cycles should correct the
8595 * excess load moved.
8597 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
8599 /* Prevent to re-select dst_cpu via env's cpus */
8600 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
8602 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
8603 env
.dst_cpu
= env
.new_dst_cpu
;
8604 env
.flags
&= ~LBF_DST_PINNED
;
8606 env
.loop_break
= sched_nr_migrate_break
;
8609 * Go back to "more_balance" rather than "redo" since we
8610 * need to continue with same src_cpu.
8616 * We failed to reach balance because of affinity.
8619 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
8621 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
8622 *group_imbalance
= 1;
8625 /* All tasks on this runqueue were pinned by CPU affinity */
8626 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
8627 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
8629 * Attempting to continue load balancing at the current
8630 * sched_domain level only makes sense if there are
8631 * active CPUs remaining as possible busiest CPUs to
8632 * pull load from which are not contained within the
8633 * destination group that is receiving any migrated
8636 if (!cpumask_subset(cpus
, env
.dst_grpmask
)) {
8638 env
.loop_break
= sched_nr_migrate_break
;
8641 goto out_all_pinned
;
8646 schedstat_inc(sd
->lb_failed
[idle
]);
8648 * Increment the failure counter only on periodic balance.
8649 * We do not want newidle balance, which can be very
8650 * frequent, pollute the failure counter causing
8651 * excessive cache_hot migrations and active balances.
8653 if (idle
!= CPU_NEWLY_IDLE
)
8654 sd
->nr_balance_failed
++;
8656 if (need_active_balance(&env
)) {
8657 unsigned long flags
;
8659 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
8661 /* don't kick the active_load_balance_cpu_stop,
8662 * if the curr task on busiest cpu can't be
8665 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
8666 raw_spin_unlock_irqrestore(&busiest
->lock
,
8668 env
.flags
|= LBF_ALL_PINNED
;
8669 goto out_one_pinned
;
8673 * ->active_balance synchronizes accesses to
8674 * ->active_balance_work. Once set, it's cleared
8675 * only after active load balance is finished.
8677 if (!busiest
->active_balance
) {
8678 busiest
->active_balance
= 1;
8679 busiest
->push_cpu
= this_cpu
;
8682 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
8684 if (active_balance
) {
8685 stop_one_cpu_nowait(cpu_of(busiest
),
8686 active_load_balance_cpu_stop
, busiest
,
8687 &busiest
->active_balance_work
);
8690 /* We've kicked active balancing, force task migration. */
8691 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
8694 sd
->nr_balance_failed
= 0;
8696 if (likely(!active_balance
)) {
8697 /* We were unbalanced, so reset the balancing interval */
8698 sd
->balance_interval
= sd
->min_interval
;
8701 * If we've begun active balancing, start to back off. This
8702 * case may not be covered by the all_pinned logic if there
8703 * is only 1 task on the busy runqueue (because we don't call
8706 if (sd
->balance_interval
< sd
->max_interval
)
8707 sd
->balance_interval
*= 2;
8714 * We reach balance although we may have faced some affinity
8715 * constraints. Clear the imbalance flag only if other tasks got
8716 * a chance to move and fix the imbalance.
8718 if (sd_parent
&& !(env
.flags
& LBF_ALL_PINNED
)) {
8719 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
8721 if (*group_imbalance
)
8722 *group_imbalance
= 0;
8727 * We reach balance because all tasks are pinned at this level so
8728 * we can't migrate them. Let the imbalance flag set so parent level
8729 * can try to migrate them.
8731 schedstat_inc(sd
->lb_balanced
[idle
]);
8733 sd
->nr_balance_failed
= 0;
8739 * idle_balance() disregards balance intervals, so we could repeatedly
8740 * reach this code, which would lead to balance_interval skyrocketting
8741 * in a short amount of time. Skip the balance_interval increase logic
8744 if (env
.idle
== CPU_NEWLY_IDLE
)
8747 /* tune up the balancing interval */
8748 if (((env
.flags
& LBF_ALL_PINNED
) &&
8749 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
8750 (sd
->balance_interval
< sd
->max_interval
))
8751 sd
->balance_interval
*= 2;
8756 static inline unsigned long
8757 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
8759 unsigned long interval
= sd
->balance_interval
;
8762 interval
*= sd
->busy_factor
;
8764 /* scale ms to jiffies */
8765 interval
= msecs_to_jiffies(interval
);
8766 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
8772 update_next_balance(struct sched_domain
*sd
, unsigned long *next_balance
)
8774 unsigned long interval
, next
;
8776 /* used by idle balance, so cpu_busy = 0 */
8777 interval
= get_sd_balance_interval(sd
, 0);
8778 next
= sd
->last_balance
+ interval
;
8780 if (time_after(*next_balance
, next
))
8781 *next_balance
= next
;
8785 * idle_balance is called by schedule() if this_cpu is about to become
8786 * idle. Attempts to pull tasks from other CPUs.
8788 static int idle_balance(struct rq
*this_rq
, struct rq_flags
*rf
)
8790 unsigned long next_balance
= jiffies
+ HZ
;
8791 int this_cpu
= this_rq
->cpu
;
8792 struct sched_domain
*sd
;
8793 int pulled_task
= 0;
8797 * We must set idle_stamp _before_ calling idle_balance(), such that we
8798 * measure the duration of idle_balance() as idle time.
8800 this_rq
->idle_stamp
= rq_clock(this_rq
);
8803 * Do not pull tasks towards !active CPUs...
8805 if (!cpu_active(this_cpu
))
8809 * This is OK, because current is on_cpu, which avoids it being picked
8810 * for load-balance and preemption/IRQs are still disabled avoiding
8811 * further scheduler activity on it and we're being very careful to
8812 * re-start the picking loop.
8814 rq_unpin_lock(this_rq
, rf
);
8816 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
8817 !this_rq
->rd
->overload
) {
8819 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
8821 update_next_balance(sd
, &next_balance
);
8827 raw_spin_unlock(&this_rq
->lock
);
8829 update_blocked_averages(this_cpu
);
8831 for_each_domain(this_cpu
, sd
) {
8832 int continue_balancing
= 1;
8833 u64 t0
, domain_cost
;
8835 if (!(sd
->flags
& SD_LOAD_BALANCE
))
8838 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
8839 update_next_balance(sd
, &next_balance
);
8843 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
8844 t0
= sched_clock_cpu(this_cpu
);
8846 pulled_task
= load_balance(this_cpu
, this_rq
,
8848 &continue_balancing
);
8850 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
8851 if (domain_cost
> sd
->max_newidle_lb_cost
)
8852 sd
->max_newidle_lb_cost
= domain_cost
;
8854 curr_cost
+= domain_cost
;
8857 update_next_balance(sd
, &next_balance
);
8860 * Stop searching for tasks to pull if there are
8861 * now runnable tasks on this rq.
8863 if (pulled_task
|| this_rq
->nr_running
> 0)
8868 raw_spin_lock(&this_rq
->lock
);
8870 if (curr_cost
> this_rq
->max_idle_balance_cost
)
8871 this_rq
->max_idle_balance_cost
= curr_cost
;
8874 * While browsing the domains, we released the rq lock, a task could
8875 * have been enqueued in the meantime. Since we're not going idle,
8876 * pretend we pulled a task.
8878 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
8882 /* Move the next balance forward */
8883 if (time_after(this_rq
->next_balance
, next_balance
))
8884 this_rq
->next_balance
= next_balance
;
8886 /* Is there a task of a high priority class? */
8887 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
8891 this_rq
->idle_stamp
= 0;
8893 rq_repin_lock(this_rq
, rf
);
8899 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8900 * running tasks off the busiest CPU onto idle CPUs. It requires at
8901 * least 1 task to be running on each physical CPU where possible, and
8902 * avoids physical / logical imbalances.
8904 static int active_load_balance_cpu_stop(void *data
)
8906 struct rq
*busiest_rq
= data
;
8907 int busiest_cpu
= cpu_of(busiest_rq
);
8908 int target_cpu
= busiest_rq
->push_cpu
;
8909 struct rq
*target_rq
= cpu_rq(target_cpu
);
8910 struct sched_domain
*sd
;
8911 struct task_struct
*p
= NULL
;
8914 rq_lock_irq(busiest_rq
, &rf
);
8916 * Between queueing the stop-work and running it is a hole in which
8917 * CPUs can become inactive. We should not move tasks from or to
8920 if (!cpu_active(busiest_cpu
) || !cpu_active(target_cpu
))
8923 /* make sure the requested cpu hasn't gone down in the meantime */
8924 if (unlikely(busiest_cpu
!= smp_processor_id() ||
8925 !busiest_rq
->active_balance
))
8928 /* Is there any task to move? */
8929 if (busiest_rq
->nr_running
<= 1)
8933 * This condition is "impossible", if it occurs
8934 * we need to fix it. Originally reported by
8935 * Bjorn Helgaas on a 128-cpu setup.
8937 BUG_ON(busiest_rq
== target_rq
);
8939 /* Search for an sd spanning us and the target CPU. */
8941 for_each_domain(target_cpu
, sd
) {
8942 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
8943 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
8948 struct lb_env env
= {
8950 .dst_cpu
= target_cpu
,
8951 .dst_rq
= target_rq
,
8952 .src_cpu
= busiest_rq
->cpu
,
8953 .src_rq
= busiest_rq
,
8956 * can_migrate_task() doesn't need to compute new_dst_cpu
8957 * for active balancing. Since we have CPU_IDLE, but no
8958 * @dst_grpmask we need to make that test go away with lying
8961 .flags
= LBF_DST_PINNED
,
8964 schedstat_inc(sd
->alb_count
);
8965 update_rq_clock(busiest_rq
);
8967 p
= detach_one_task(&env
);
8969 schedstat_inc(sd
->alb_pushed
);
8970 /* Active balancing done, reset the failure counter. */
8971 sd
->nr_balance_failed
= 0;
8973 schedstat_inc(sd
->alb_failed
);
8978 busiest_rq
->active_balance
= 0;
8979 rq_unlock(busiest_rq
, &rf
);
8982 attach_one_task(target_rq
, p
);
8989 static inline int on_null_domain(struct rq
*rq
)
8991 return unlikely(!rcu_dereference_sched(rq
->sd
));
8994 #ifdef CONFIG_NO_HZ_COMMON
8996 * idle load balancing details
8997 * - When one of the busy CPUs notice that there may be an idle rebalancing
8998 * needed, they will kick the idle load balancer, which then does idle
8999 * load balancing for all the idle CPUs.
9000 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9004 cpumask_var_t idle_cpus_mask
;
9006 unsigned long next_balance
; /* in jiffy units */
9007 } nohz ____cacheline_aligned
;
9009 static inline int find_new_ilb(void)
9013 for_each_cpu_and(ilb
, nohz
.idle_cpus_mask
,
9014 housekeeping_cpumask(HK_FLAG_MISC
)) {
9023 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9024 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
9026 static void nohz_balancer_kick(void)
9030 nohz
.next_balance
++;
9032 ilb_cpu
= find_new_ilb();
9034 if (ilb_cpu
>= nr_cpu_ids
)
9037 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
9040 * Use smp_send_reschedule() instead of resched_cpu().
9041 * This way we generate a sched IPI on the target cpu which
9042 * is idle. And the softirq performing nohz idle load balance
9043 * will be run before returning from the IPI.
9045 smp_send_reschedule(ilb_cpu
);
9049 void nohz_balance_exit_idle(unsigned int cpu
)
9051 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
9053 * Completely isolated CPUs don't ever set, so we must test.
9055 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
9056 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
9057 atomic_dec(&nohz
.nr_cpus
);
9059 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
9063 static inline void set_cpu_sd_state_busy(void)
9065 struct sched_domain
*sd
;
9066 int cpu
= smp_processor_id();
9069 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
9071 if (!sd
|| !sd
->nohz_idle
)
9075 atomic_inc(&sd
->shared
->nr_busy_cpus
);
9080 void set_cpu_sd_state_idle(void)
9082 struct sched_domain
*sd
;
9083 int cpu
= smp_processor_id();
9086 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
9088 if (!sd
|| sd
->nohz_idle
)
9092 atomic_dec(&sd
->shared
->nr_busy_cpus
);
9098 * This routine will record that the cpu is going idle with tick stopped.
9099 * This info will be used in performing idle load balancing in the future.
9101 void nohz_balance_enter_idle(int cpu
)
9104 * If this cpu is going down, then nothing needs to be done.
9106 if (!cpu_active(cpu
))
9109 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9110 if (!housekeeping_cpu(cpu
, HK_FLAG_SCHED
))
9113 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
9117 * If we're a completely isolated CPU, we don't play.
9119 if (on_null_domain(cpu_rq(cpu
)))
9122 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
9123 atomic_inc(&nohz
.nr_cpus
);
9124 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
9128 static DEFINE_SPINLOCK(balancing
);
9131 * Scale the max load_balance interval with the number of CPUs in the system.
9132 * This trades load-balance latency on larger machines for less cross talk.
9134 void update_max_interval(void)
9136 max_load_balance_interval
= HZ
*num_online_cpus()/10;
9140 * It checks each scheduling domain to see if it is due to be balanced,
9141 * and initiates a balancing operation if so.
9143 * Balancing parameters are set up in init_sched_domains.
9145 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
9147 int continue_balancing
= 1;
9149 unsigned long interval
;
9150 struct sched_domain
*sd
;
9151 /* Earliest time when we have to do rebalance again */
9152 unsigned long next_balance
= jiffies
+ 60*HZ
;
9153 int update_next_balance
= 0;
9154 int need_serialize
, need_decay
= 0;
9157 update_blocked_averages(cpu
);
9160 for_each_domain(cpu
, sd
) {
9162 * Decay the newidle max times here because this is a regular
9163 * visit to all the domains. Decay ~1% per second.
9165 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
9166 sd
->max_newidle_lb_cost
=
9167 (sd
->max_newidle_lb_cost
* 253) / 256;
9168 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
9171 max_cost
+= sd
->max_newidle_lb_cost
;
9173 if (!(sd
->flags
& SD_LOAD_BALANCE
))
9177 * Stop the load balance at this level. There is another
9178 * CPU in our sched group which is doing load balancing more
9181 if (!continue_balancing
) {
9187 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
9189 need_serialize
= sd
->flags
& SD_SERIALIZE
;
9190 if (need_serialize
) {
9191 if (!spin_trylock(&balancing
))
9195 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
9196 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
9198 * The LBF_DST_PINNED logic could have changed
9199 * env->dst_cpu, so we can't know our idle
9200 * state even if we migrated tasks. Update it.
9202 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
9204 sd
->last_balance
= jiffies
;
9205 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
9208 spin_unlock(&balancing
);
9210 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
9211 next_balance
= sd
->last_balance
+ interval
;
9212 update_next_balance
= 1;
9217 * Ensure the rq-wide value also decays but keep it at a
9218 * reasonable floor to avoid funnies with rq->avg_idle.
9220 rq
->max_idle_balance_cost
=
9221 max((u64
)sysctl_sched_migration_cost
, max_cost
);
9226 * next_balance will be updated only when there is a need.
9227 * When the cpu is attached to null domain for ex, it will not be
9230 if (likely(update_next_balance
)) {
9231 rq
->next_balance
= next_balance
;
9233 #ifdef CONFIG_NO_HZ_COMMON
9235 * If this CPU has been elected to perform the nohz idle
9236 * balance. Other idle CPUs have already rebalanced with
9237 * nohz_idle_balance() and nohz.next_balance has been
9238 * updated accordingly. This CPU is now running the idle load
9239 * balance for itself and we need to update the
9240 * nohz.next_balance accordingly.
9242 if ((idle
== CPU_IDLE
) && time_after(nohz
.next_balance
, rq
->next_balance
))
9243 nohz
.next_balance
= rq
->next_balance
;
9248 #ifdef CONFIG_NO_HZ_COMMON
9250 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9251 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9253 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
9255 int this_cpu
= this_rq
->cpu
;
9258 /* Earliest time when we have to do rebalance again */
9259 unsigned long next_balance
= jiffies
+ 60*HZ
;
9260 int update_next_balance
= 0;
9262 if (idle
!= CPU_IDLE
||
9263 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
9266 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
9267 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
9271 * If this cpu gets work to do, stop the load balancing
9272 * work being done for other cpus. Next load
9273 * balancing owner will pick it up.
9278 rq
= cpu_rq(balance_cpu
);
9281 * If time for next balance is due,
9284 if (time_after_eq(jiffies
, rq
->next_balance
)) {
9287 rq_lock_irq(rq
, &rf
);
9288 update_rq_clock(rq
);
9289 cpu_load_update_idle(rq
);
9290 rq_unlock_irq(rq
, &rf
);
9292 rebalance_domains(rq
, CPU_IDLE
);
9295 if (time_after(next_balance
, rq
->next_balance
)) {
9296 next_balance
= rq
->next_balance
;
9297 update_next_balance
= 1;
9302 * next_balance will be updated only when there is a need.
9303 * When the CPU is attached to null domain for ex, it will not be
9306 if (likely(update_next_balance
))
9307 nohz
.next_balance
= next_balance
;
9309 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
9313 * Current heuristic for kicking the idle load balancer in the presence
9314 * of an idle cpu in the system.
9315 * - This rq has more than one task.
9316 * - This rq has at least one CFS task and the capacity of the CPU is
9317 * significantly reduced because of RT tasks or IRQs.
9318 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9319 * multiple busy cpu.
9320 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9321 * domain span are idle.
9323 static inline bool nohz_kick_needed(struct rq
*rq
)
9325 unsigned long now
= jiffies
;
9326 struct sched_domain_shared
*sds
;
9327 struct sched_domain
*sd
;
9328 int nr_busy
, i
, cpu
= rq
->cpu
;
9331 if (unlikely(rq
->idle_balance
))
9335 * We may be recently in ticked or tickless idle mode. At the first
9336 * busy tick after returning from idle, we will update the busy stats.
9338 set_cpu_sd_state_busy();
9339 nohz_balance_exit_idle(cpu
);
9342 * None are in tickless mode and hence no need for NOHZ idle load
9345 if (likely(!atomic_read(&nohz
.nr_cpus
)))
9348 if (time_before(now
, nohz
.next_balance
))
9351 if (rq
->nr_running
>= 2)
9355 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
9358 * XXX: write a coherent comment on why we do this.
9359 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9361 nr_busy
= atomic_read(&sds
->nr_busy_cpus
);
9369 sd
= rcu_dereference(rq
->sd
);
9371 if ((rq
->cfs
.h_nr_running
>= 1) &&
9372 check_cpu_capacity(rq
, sd
)) {
9378 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
9380 for_each_cpu(i
, sched_domain_span(sd
)) {
9382 !cpumask_test_cpu(i
, nohz
.idle_cpus_mask
))
9385 if (sched_asym_prefer(i
, cpu
)) {
9396 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
9400 * run_rebalance_domains is triggered when needed from the scheduler tick.
9401 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9403 static __latent_entropy
void run_rebalance_domains(struct softirq_action
*h
)
9405 struct rq
*this_rq
= this_rq();
9406 enum cpu_idle_type idle
= this_rq
->idle_balance
?
9407 CPU_IDLE
: CPU_NOT_IDLE
;
9410 * If this cpu has a pending nohz_balance_kick, then do the
9411 * balancing on behalf of the other idle cpus whose ticks are
9412 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9413 * give the idle cpus a chance to load balance. Else we may
9414 * load balance only within the local sched_domain hierarchy
9415 * and abort nohz_idle_balance altogether if we pull some load.
9417 nohz_idle_balance(this_rq
, idle
);
9418 rebalance_domains(this_rq
, idle
);
9422 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9424 void trigger_load_balance(struct rq
*rq
)
9426 /* Don't need to rebalance while attached to NULL domain */
9427 if (unlikely(on_null_domain(rq
)))
9430 if (time_after_eq(jiffies
, rq
->next_balance
))
9431 raise_softirq(SCHED_SOFTIRQ
);
9432 #ifdef CONFIG_NO_HZ_COMMON
9433 if (nohz_kick_needed(rq
))
9434 nohz_balancer_kick();
9438 static void rq_online_fair(struct rq
*rq
)
9442 update_runtime_enabled(rq
);
9445 static void rq_offline_fair(struct rq
*rq
)
9449 /* Ensure any throttled groups are reachable by pick_next_task */
9450 unthrottle_offline_cfs_rqs(rq
);
9453 #endif /* CONFIG_SMP */
9456 * scheduler tick hitting a task of our scheduling class:
9458 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
9460 struct cfs_rq
*cfs_rq
;
9461 struct sched_entity
*se
= &curr
->se
;
9463 for_each_sched_entity(se
) {
9464 cfs_rq
= cfs_rq_of(se
);
9465 entity_tick(cfs_rq
, se
, queued
);
9468 if (static_branch_unlikely(&sched_numa_balancing
))
9469 task_tick_numa(rq
, curr
);
9473 * called on fork with the child task as argument from the parent's context
9474 * - child not yet on the tasklist
9475 * - preemption disabled
9477 static void task_fork_fair(struct task_struct
*p
)
9479 struct cfs_rq
*cfs_rq
;
9480 struct sched_entity
*se
= &p
->se
, *curr
;
9481 struct rq
*rq
= this_rq();
9485 update_rq_clock(rq
);
9487 cfs_rq
= task_cfs_rq(current
);
9488 curr
= cfs_rq
->curr
;
9490 update_curr(cfs_rq
);
9491 se
->vruntime
= curr
->vruntime
;
9493 place_entity(cfs_rq
, se
, 1);
9495 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
9497 * Upon rescheduling, sched_class::put_prev_task() will place
9498 * 'current' within the tree based on its new key value.
9500 swap(curr
->vruntime
, se
->vruntime
);
9504 se
->vruntime
-= cfs_rq
->min_vruntime
;
9509 * Priority of the task has changed. Check to see if we preempt
9513 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
9515 if (!task_on_rq_queued(p
))
9519 * Reschedule if we are currently running on this runqueue and
9520 * our priority decreased, or if we are not currently running on
9521 * this runqueue and our priority is higher than the current's
9523 if (rq
->curr
== p
) {
9524 if (p
->prio
> oldprio
)
9527 check_preempt_curr(rq
, p
, 0);
9530 static inline bool vruntime_normalized(struct task_struct
*p
)
9532 struct sched_entity
*se
= &p
->se
;
9535 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9536 * the dequeue_entity(.flags=0) will already have normalized the
9543 * When !on_rq, vruntime of the task has usually NOT been normalized.
9544 * But there are some cases where it has already been normalized:
9546 * - A forked child which is waiting for being woken up by
9547 * wake_up_new_task().
9548 * - A task which has been woken up by try_to_wake_up() and
9549 * waiting for actually being woken up by sched_ttwu_pending().
9551 if (!se
->sum_exec_runtime
||
9552 (p
->state
== TASK_WAKING
&& p
->sched_remote_wakeup
))
9558 #ifdef CONFIG_FAIR_GROUP_SCHED
9560 * Propagate the changes of the sched_entity across the tg tree to make it
9561 * visible to the root
9563 static void propagate_entity_cfs_rq(struct sched_entity
*se
)
9565 struct cfs_rq
*cfs_rq
;
9567 /* Start to propagate at parent */
9570 for_each_sched_entity(se
) {
9571 cfs_rq
= cfs_rq_of(se
);
9573 if (cfs_rq_throttled(cfs_rq
))
9576 update_load_avg(cfs_rq
, se
, UPDATE_TG
);
9580 static void propagate_entity_cfs_rq(struct sched_entity
*se
) { }
9583 static void detach_entity_cfs_rq(struct sched_entity
*se
)
9585 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9587 /* Catch up with the cfs_rq and remove our load when we leave */
9588 update_load_avg(cfs_rq
, se
, 0);
9589 detach_entity_load_avg(cfs_rq
, se
);
9590 update_tg_load_avg(cfs_rq
, false);
9591 propagate_entity_cfs_rq(se
);
9594 static void attach_entity_cfs_rq(struct sched_entity
*se
)
9596 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9598 #ifdef CONFIG_FAIR_GROUP_SCHED
9600 * Since the real-depth could have been changed (only FAIR
9601 * class maintain depth value), reset depth properly.
9603 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
9606 /* Synchronize entity with its cfs_rq */
9607 update_load_avg(cfs_rq
, se
, sched_feat(ATTACH_AGE_LOAD
) ? 0 : SKIP_AGE_LOAD
);
9608 attach_entity_load_avg(cfs_rq
, se
);
9609 update_tg_load_avg(cfs_rq
, false);
9610 propagate_entity_cfs_rq(se
);
9613 static void detach_task_cfs_rq(struct task_struct
*p
)
9615 struct sched_entity
*se
= &p
->se
;
9616 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9618 if (!vruntime_normalized(p
)) {
9620 * Fix up our vruntime so that the current sleep doesn't
9621 * cause 'unlimited' sleep bonus.
9623 place_entity(cfs_rq
, se
, 0);
9624 se
->vruntime
-= cfs_rq
->min_vruntime
;
9627 detach_entity_cfs_rq(se
);
9630 static void attach_task_cfs_rq(struct task_struct
*p
)
9632 struct sched_entity
*se
= &p
->se
;
9633 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9635 attach_entity_cfs_rq(se
);
9637 if (!vruntime_normalized(p
))
9638 se
->vruntime
+= cfs_rq
->min_vruntime
;
9641 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
9643 detach_task_cfs_rq(p
);
9646 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
9648 attach_task_cfs_rq(p
);
9650 if (task_on_rq_queued(p
)) {
9652 * We were most likely switched from sched_rt, so
9653 * kick off the schedule if running, otherwise just see
9654 * if we can still preempt the current task.
9659 check_preempt_curr(rq
, p
, 0);
9663 /* Account for a task changing its policy or group.
9665 * This routine is mostly called to set cfs_rq->curr field when a task
9666 * migrates between groups/classes.
9668 static void set_curr_task_fair(struct rq
*rq
)
9670 struct sched_entity
*se
= &rq
->curr
->se
;
9672 for_each_sched_entity(se
) {
9673 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9675 set_next_entity(cfs_rq
, se
);
9676 /* ensure bandwidth has been allocated on our new cfs_rq */
9677 account_cfs_rq_runtime(cfs_rq
, 0);
9681 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
9683 cfs_rq
->tasks_timeline
= RB_ROOT_CACHED
;
9684 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9685 #ifndef CONFIG_64BIT
9686 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
9689 raw_spin_lock_init(&cfs_rq
->removed
.lock
);
9693 #ifdef CONFIG_FAIR_GROUP_SCHED
9694 static void task_set_group_fair(struct task_struct
*p
)
9696 struct sched_entity
*se
= &p
->se
;
9698 set_task_rq(p
, task_cpu(p
));
9699 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
9702 static void task_move_group_fair(struct task_struct
*p
)
9704 detach_task_cfs_rq(p
);
9705 set_task_rq(p
, task_cpu(p
));
9708 /* Tell se's cfs_rq has been changed -- migrated */
9709 p
->se
.avg
.last_update_time
= 0;
9711 attach_task_cfs_rq(p
);
9714 static void task_change_group_fair(struct task_struct
*p
, int type
)
9717 case TASK_SET_GROUP
:
9718 task_set_group_fair(p
);
9721 case TASK_MOVE_GROUP
:
9722 task_move_group_fair(p
);
9727 void free_fair_sched_group(struct task_group
*tg
)
9731 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
9733 for_each_possible_cpu(i
) {
9735 kfree(tg
->cfs_rq
[i
]);
9744 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9746 struct sched_entity
*se
;
9747 struct cfs_rq
*cfs_rq
;
9750 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9753 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9757 tg
->shares
= NICE_0_LOAD
;
9759 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
9761 for_each_possible_cpu(i
) {
9762 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9763 GFP_KERNEL
, cpu_to_node(i
));
9767 se
= kzalloc_node(sizeof(struct sched_entity
),
9768 GFP_KERNEL
, cpu_to_node(i
));
9772 init_cfs_rq(cfs_rq
);
9773 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
9774 init_entity_runnable_average(se
);
9785 void online_fair_sched_group(struct task_group
*tg
)
9787 struct sched_entity
*se
;
9792 for_each_possible_cpu(i
) {
9795 rq_lock_irq(rq
, &rf
);
9796 update_rq_clock(rq
);
9797 attach_entity_cfs_rq(se
);
9798 sync_throttle(tg
, i
);
9799 rq_unlock_irq(rq
, &rf
);
9803 void unregister_fair_sched_group(struct task_group
*tg
)
9805 unsigned long flags
;
9809 for_each_possible_cpu(cpu
) {
9811 remove_entity_load_avg(tg
->se
[cpu
]);
9814 * Only empty task groups can be destroyed; so we can speculatively
9815 * check on_list without danger of it being re-added.
9817 if (!tg
->cfs_rq
[cpu
]->on_list
)
9822 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9823 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
9824 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9828 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9829 struct sched_entity
*se
, int cpu
,
9830 struct sched_entity
*parent
)
9832 struct rq
*rq
= cpu_rq(cpu
);
9836 init_cfs_rq_runtime(cfs_rq
);
9838 tg
->cfs_rq
[cpu
] = cfs_rq
;
9841 /* se could be NULL for root_task_group */
9846 se
->cfs_rq
= &rq
->cfs
;
9849 se
->cfs_rq
= parent
->my_q
;
9850 se
->depth
= parent
->depth
+ 1;
9854 /* guarantee group entities always have weight */
9855 update_load_set(&se
->load
, NICE_0_LOAD
);
9856 se
->parent
= parent
;
9859 static DEFINE_MUTEX(shares_mutex
);
9861 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9866 * We can't change the weight of the root cgroup.
9871 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
9873 mutex_lock(&shares_mutex
);
9874 if (tg
->shares
== shares
)
9877 tg
->shares
= shares
;
9878 for_each_possible_cpu(i
) {
9879 struct rq
*rq
= cpu_rq(i
);
9880 struct sched_entity
*se
= tg
->se
[i
];
9883 /* Propagate contribution to hierarchy */
9884 rq_lock_irqsave(rq
, &rf
);
9885 update_rq_clock(rq
);
9886 for_each_sched_entity(se
) {
9887 update_load_avg(cfs_rq_of(se
), se
, UPDATE_TG
);
9888 update_cfs_group(se
);
9890 rq_unlock_irqrestore(rq
, &rf
);
9894 mutex_unlock(&shares_mutex
);
9897 #else /* CONFIG_FAIR_GROUP_SCHED */
9899 void free_fair_sched_group(struct task_group
*tg
) { }
9901 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9906 void online_fair_sched_group(struct task_group
*tg
) { }
9908 void unregister_fair_sched_group(struct task_group
*tg
) { }
9910 #endif /* CONFIG_FAIR_GROUP_SCHED */
9913 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
9915 struct sched_entity
*se
= &task
->se
;
9916 unsigned int rr_interval
= 0;
9919 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9922 if (rq
->cfs
.load
.weight
)
9923 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
9929 * All the scheduling class methods:
9931 const struct sched_class fair_sched_class
= {
9932 .next
= &idle_sched_class
,
9933 .enqueue_task
= enqueue_task_fair
,
9934 .dequeue_task
= dequeue_task_fair
,
9935 .yield_task
= yield_task_fair
,
9936 .yield_to_task
= yield_to_task_fair
,
9938 .check_preempt_curr
= check_preempt_wakeup
,
9940 .pick_next_task
= pick_next_task_fair
,
9941 .put_prev_task
= put_prev_task_fair
,
9944 .select_task_rq
= select_task_rq_fair
,
9945 .migrate_task_rq
= migrate_task_rq_fair
,
9947 .rq_online
= rq_online_fair
,
9948 .rq_offline
= rq_offline_fair
,
9950 .task_dead
= task_dead_fair
,
9951 .set_cpus_allowed
= set_cpus_allowed_common
,
9954 .set_curr_task
= set_curr_task_fair
,
9955 .task_tick
= task_tick_fair
,
9956 .task_fork
= task_fork_fair
,
9958 .prio_changed
= prio_changed_fair
,
9959 .switched_from
= switched_from_fair
,
9960 .switched_to
= switched_to_fair
,
9962 .get_rr_interval
= get_rr_interval_fair
,
9964 .update_curr
= update_curr_fair
,
9966 #ifdef CONFIG_FAIR_GROUP_SCHED
9967 .task_change_group
= task_change_group_fair
,
9971 #ifdef CONFIG_SCHED_DEBUG
9972 void print_cfs_stats(struct seq_file
*m
, int cpu
)
9974 struct cfs_rq
*cfs_rq
;
9977 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
9978 print_cfs_rq(m
, cpu
, cfs_rq
);
9982 #ifdef CONFIG_NUMA_BALANCING
9983 void show_numa_stats(struct task_struct
*p
, struct seq_file
*m
)
9986 unsigned long tsf
= 0, tpf
= 0, gsf
= 0, gpf
= 0;
9988 for_each_online_node(node
) {
9989 if (p
->numa_faults
) {
9990 tsf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
9991 tpf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
9993 if (p
->numa_group
) {
9994 gsf
= p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)],
9995 gpf
= p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
9997 print_numa_stats(m
, node
, tsf
, tpf
, gsf
, gpf
);
10000 #endif /* CONFIG_NUMA_BALANCING */
10001 #endif /* CONFIG_SCHED_DEBUG */
10003 __init
void init_sched_fair_class(void)
10006 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
10008 #ifdef CONFIG_NO_HZ_COMMON
10009 nohz
.next_balance
= jiffies
;
10010 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);