2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 #include <linux/sched/mm.h>
24 #include <linux/sched/topology.h>
26 #include <linux/latencytop.h>
27 #include <linux/cpumask.h>
28 #include <linux/cpuidle.h>
29 #include <linux/slab.h>
30 #include <linux/profile.h>
31 #include <linux/interrupt.h>
32 #include <linux/mempolicy.h>
33 #include <linux/migrate.h>
34 #include <linux/task_work.h>
36 #include <trace/events/sched.h>
41 * Targeted preemption latency for CPU-bound tasks:
43 * NOTE: this latency value is not the same as the concept of
44 * 'timeslice length' - timeslices in CFS are of variable length
45 * and have no persistent notion like in traditional, time-slice
46 * based scheduling concepts.
48 * (to see the precise effective timeslice length of your workload,
49 * run vmstat and monitor the context-switches (cs) field)
51 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
53 unsigned int sysctl_sched_latency
= 6000000ULL;
54 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
57 * The initial- and re-scaling of tunables is configurable
61 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
62 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
63 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
65 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
67 enum sched_tunable_scaling sysctl_sched_tunable_scaling
= SCHED_TUNABLESCALING_LOG
;
70 * Minimal preemption granularity for CPU-bound tasks:
72 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
74 unsigned int sysctl_sched_min_granularity
= 750000ULL;
75 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
78 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
80 static unsigned int sched_nr_latency
= 8;
83 * After fork, child runs first. If set to 0 (default) then
84 * parent will (try to) run first.
86 unsigned int sysctl_sched_child_runs_first __read_mostly
;
89 * SCHED_OTHER wake-up granularity.
91 * This option delays the preemption effects of decoupled workloads
92 * and reduces their over-scheduling. Synchronous workloads will still
93 * have immediate wakeup/sleep latencies.
95 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
97 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
98 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
100 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
104 * For asym packing, by default the lower numbered cpu has higher priority.
106 int __weak
arch_asym_cpu_priority(int cpu
)
112 #ifdef CONFIG_CFS_BANDWIDTH
114 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
115 * each time a cfs_rq requests quota.
117 * Note: in the case that the slice exceeds the runtime remaining (either due
118 * to consumption or the quota being specified to be smaller than the slice)
119 * we will always only issue the remaining available time.
121 * (default: 5 msec, units: microseconds)
123 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
127 * The margin used when comparing utilization with CPU capacity:
128 * util * margin < capacity * 1024
132 unsigned int capacity_margin
= 1280;
134 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
140 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
146 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
153 * Increase the granularity value when there are more CPUs,
154 * because with more CPUs the 'effective latency' as visible
155 * to users decreases. But the relationship is not linear,
156 * so pick a second-best guess by going with the log2 of the
159 * This idea comes from the SD scheduler of Con Kolivas:
161 static unsigned int get_update_sysctl_factor(void)
163 unsigned int cpus
= min_t(unsigned int, num_online_cpus(), 8);
166 switch (sysctl_sched_tunable_scaling
) {
167 case SCHED_TUNABLESCALING_NONE
:
170 case SCHED_TUNABLESCALING_LINEAR
:
173 case SCHED_TUNABLESCALING_LOG
:
175 factor
= 1 + ilog2(cpus
);
182 static void update_sysctl(void)
184 unsigned int factor
= get_update_sysctl_factor();
186 #define SET_SYSCTL(name) \
187 (sysctl_##name = (factor) * normalized_sysctl_##name)
188 SET_SYSCTL(sched_min_granularity
);
189 SET_SYSCTL(sched_latency
);
190 SET_SYSCTL(sched_wakeup_granularity
);
194 void sched_init_granularity(void)
199 #define WMULT_CONST (~0U)
200 #define WMULT_SHIFT 32
202 static void __update_inv_weight(struct load_weight
*lw
)
206 if (likely(lw
->inv_weight
))
209 w
= scale_load_down(lw
->weight
);
211 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
213 else if (unlikely(!w
))
214 lw
->inv_weight
= WMULT_CONST
;
216 lw
->inv_weight
= WMULT_CONST
/ w
;
220 * delta_exec * weight / lw.weight
222 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
224 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
225 * we're guaranteed shift stays positive because inv_weight is guaranteed to
226 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
228 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
229 * weight/lw.weight <= 1, and therefore our shift will also be positive.
231 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
233 u64 fact
= scale_load_down(weight
);
234 int shift
= WMULT_SHIFT
;
236 __update_inv_weight(lw
);
238 if (unlikely(fact
>> 32)) {
245 /* hint to use a 32x32->64 mul */
246 fact
= (u64
)(u32
)fact
* lw
->inv_weight
;
253 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
257 const struct sched_class fair_sched_class
;
259 /**************************************************************
260 * CFS operations on generic schedulable entities:
263 #ifdef CONFIG_FAIR_GROUP_SCHED
265 /* cpu runqueue to which this cfs_rq is attached */
266 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
271 /* An entity is a task if it doesn't "own" a runqueue */
272 #define entity_is_task(se) (!se->my_q)
274 static inline struct task_struct
*task_of(struct sched_entity
*se
)
276 SCHED_WARN_ON(!entity_is_task(se
));
277 return container_of(se
, struct task_struct
, se
);
280 /* Walk up scheduling entities hierarchy */
281 #define for_each_sched_entity(se) \
282 for (; se; se = se->parent)
284 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
289 /* runqueue on which this entity is (to be) queued */
290 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
295 /* runqueue "owned" by this group */
296 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
301 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
303 if (!cfs_rq
->on_list
) {
304 struct rq
*rq
= rq_of(cfs_rq
);
305 int cpu
= cpu_of(rq
);
307 * Ensure we either appear before our parent (if already
308 * enqueued) or force our parent to appear after us when it is
309 * enqueued. The fact that we always enqueue bottom-up
310 * reduces this to two cases and a special case for the root
311 * cfs_rq. Furthermore, it also means that we will always reset
312 * tmp_alone_branch either when the branch is connected
313 * to a tree or when we reach the beg of the tree
315 if (cfs_rq
->tg
->parent
&&
316 cfs_rq
->tg
->parent
->cfs_rq
[cpu
]->on_list
) {
318 * If parent is already on the list, we add the child
319 * just before. Thanks to circular linked property of
320 * the list, this means to put the child at the tail
321 * of the list that starts by parent.
323 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
324 &(cfs_rq
->tg
->parent
->cfs_rq
[cpu
]->leaf_cfs_rq_list
));
326 * The branch is now connected to its tree so we can
327 * reset tmp_alone_branch to the beginning of the
330 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
331 } else if (!cfs_rq
->tg
->parent
) {
333 * cfs rq without parent should be put
334 * at the tail of the list.
336 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
337 &rq
->leaf_cfs_rq_list
);
339 * We have reach the beg of a tree so we can reset
340 * tmp_alone_branch to the beginning of the list.
342 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
345 * The parent has not already been added so we want to
346 * make sure that it will be put after us.
347 * tmp_alone_branch points to the beg of the branch
348 * where we will add parent.
350 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
351 rq
->tmp_alone_branch
);
353 * update tmp_alone_branch to points to the new beg
356 rq
->tmp_alone_branch
= &cfs_rq
->leaf_cfs_rq_list
;
363 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
365 if (cfs_rq
->on_list
) {
366 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
371 /* Iterate thr' all leaf cfs_rq's on a runqueue */
372 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
373 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
376 /* Do the two (enqueued) entities belong to the same group ? */
377 static inline struct cfs_rq
*
378 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
380 if (se
->cfs_rq
== pse
->cfs_rq
)
386 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
392 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
394 int se_depth
, pse_depth
;
397 * preemption test can be made between sibling entities who are in the
398 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
399 * both tasks until we find their ancestors who are siblings of common
403 /* First walk up until both entities are at same depth */
404 se_depth
= (*se
)->depth
;
405 pse_depth
= (*pse
)->depth
;
407 while (se_depth
> pse_depth
) {
409 *se
= parent_entity(*se
);
412 while (pse_depth
> se_depth
) {
414 *pse
= parent_entity(*pse
);
417 while (!is_same_group(*se
, *pse
)) {
418 *se
= parent_entity(*se
);
419 *pse
= parent_entity(*pse
);
423 #else /* !CONFIG_FAIR_GROUP_SCHED */
425 static inline struct task_struct
*task_of(struct sched_entity
*se
)
427 return container_of(se
, struct task_struct
, se
);
430 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
432 return container_of(cfs_rq
, struct rq
, cfs
);
435 #define entity_is_task(se) 1
437 #define for_each_sched_entity(se) \
438 for (; se; se = NULL)
440 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
442 return &task_rq(p
)->cfs
;
445 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
447 struct task_struct
*p
= task_of(se
);
448 struct rq
*rq
= task_rq(p
);
453 /* runqueue "owned" by this group */
454 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
459 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
463 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
467 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
468 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
470 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
476 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
480 #endif /* CONFIG_FAIR_GROUP_SCHED */
482 static __always_inline
483 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
485 /**************************************************************
486 * Scheduling class tree data structure manipulation methods:
489 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
491 s64 delta
= (s64
)(vruntime
- max_vruntime
);
493 max_vruntime
= vruntime
;
498 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
500 s64 delta
= (s64
)(vruntime
- min_vruntime
);
502 min_vruntime
= vruntime
;
507 static inline int entity_before(struct sched_entity
*a
,
508 struct sched_entity
*b
)
510 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
513 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
515 struct sched_entity
*curr
= cfs_rq
->curr
;
517 u64 vruntime
= cfs_rq
->min_vruntime
;
521 vruntime
= curr
->vruntime
;
526 if (cfs_rq
->rb_leftmost
) {
527 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
532 vruntime
= se
->vruntime
;
534 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
537 /* ensure we never gain time by being placed backwards. */
538 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
541 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
546 * Enqueue an entity into the rb-tree:
548 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
550 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
551 struct rb_node
*parent
= NULL
;
552 struct sched_entity
*entry
;
556 * Find the right place in the rbtree:
560 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
562 * We dont care about collisions. Nodes with
563 * the same key stay together.
565 if (entity_before(se
, entry
)) {
566 link
= &parent
->rb_left
;
568 link
= &parent
->rb_right
;
574 * Maintain a cache of leftmost tree entries (it is frequently
578 cfs_rq
->rb_leftmost
= &se
->run_node
;
580 rb_link_node(&se
->run_node
, parent
, link
);
581 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
584 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
586 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
587 struct rb_node
*next_node
;
589 next_node
= rb_next(&se
->run_node
);
590 cfs_rq
->rb_leftmost
= next_node
;
593 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
596 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
598 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
603 return rb_entry(left
, struct sched_entity
, run_node
);
606 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
608 struct rb_node
*next
= rb_next(&se
->run_node
);
613 return rb_entry(next
, struct sched_entity
, run_node
);
616 #ifdef CONFIG_SCHED_DEBUG
617 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
619 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
624 return rb_entry(last
, struct sched_entity
, run_node
);
627 /**************************************************************
628 * Scheduling class statistics methods:
631 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
632 void __user
*buffer
, size_t *lenp
,
635 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
636 unsigned int factor
= get_update_sysctl_factor();
641 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
642 sysctl_sched_min_granularity
);
644 #define WRT_SYSCTL(name) \
645 (normalized_sysctl_##name = sysctl_##name / (factor))
646 WRT_SYSCTL(sched_min_granularity
);
647 WRT_SYSCTL(sched_latency
);
648 WRT_SYSCTL(sched_wakeup_granularity
);
658 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
660 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
661 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
667 * The idea is to set a period in which each task runs once.
669 * When there are too many tasks (sched_nr_latency) we have to stretch
670 * this period because otherwise the slices get too small.
672 * p = (nr <= nl) ? l : l*nr/nl
674 static u64
__sched_period(unsigned long nr_running
)
676 if (unlikely(nr_running
> sched_nr_latency
))
677 return nr_running
* sysctl_sched_min_granularity
;
679 return sysctl_sched_latency
;
683 * We calculate the wall-time slice from the period by taking a part
684 * proportional to the weight.
688 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
690 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
692 for_each_sched_entity(se
) {
693 struct load_weight
*load
;
694 struct load_weight lw
;
696 cfs_rq
= cfs_rq_of(se
);
697 load
= &cfs_rq
->load
;
699 if (unlikely(!se
->on_rq
)) {
702 update_load_add(&lw
, se
->load
.weight
);
705 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
711 * We calculate the vruntime slice of a to-be-inserted task.
715 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
717 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
722 #include "sched-pelt.h"
724 static int select_idle_sibling(struct task_struct
*p
, int prev_cpu
, int cpu
);
725 static unsigned long task_h_load(struct task_struct
*p
);
727 /* Give new sched_entity start runnable values to heavy its load in infant time */
728 void init_entity_runnable_average(struct sched_entity
*se
)
730 struct sched_avg
*sa
= &se
->avg
;
732 sa
->last_update_time
= 0;
734 * sched_avg's period_contrib should be strictly less then 1024, so
735 * we give it 1023 to make sure it is almost a period (1024us), and
736 * will definitely be update (after enqueue).
738 sa
->period_contrib
= 1023;
740 * Tasks are intialized with full load to be seen as heavy tasks until
741 * they get a chance to stabilize to their real load level.
742 * Group entities are intialized with zero load to reflect the fact that
743 * nothing has been attached to the task group yet.
745 if (entity_is_task(se
))
746 sa
->load_avg
= scale_load_down(se
->load
.weight
);
747 sa
->load_sum
= sa
->load_avg
* LOAD_AVG_MAX
;
749 * At this point, util_avg won't be used in select_task_rq_fair anyway
753 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
756 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
757 static void attach_entity_cfs_rq(struct sched_entity
*se
);
760 * With new tasks being created, their initial util_avgs are extrapolated
761 * based on the cfs_rq's current util_avg:
763 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
765 * However, in many cases, the above util_avg does not give a desired
766 * value. Moreover, the sum of the util_avgs may be divergent, such
767 * as when the series is a harmonic series.
769 * To solve this problem, we also cap the util_avg of successive tasks to
770 * only 1/2 of the left utilization budget:
772 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
774 * where n denotes the nth task.
776 * For example, a simplest series from the beginning would be like:
778 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
779 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
781 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
782 * if util_avg > util_avg_cap.
784 void post_init_entity_util_avg(struct sched_entity
*se
)
786 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
787 struct sched_avg
*sa
= &se
->avg
;
788 long cap
= (long)(SCHED_CAPACITY_SCALE
- cfs_rq
->avg
.util_avg
) / 2;
791 if (cfs_rq
->avg
.util_avg
!= 0) {
792 sa
->util_avg
= cfs_rq
->avg
.util_avg
* se
->load
.weight
;
793 sa
->util_avg
/= (cfs_rq
->avg
.load_avg
+ 1);
795 if (sa
->util_avg
> cap
)
800 sa
->util_sum
= sa
->util_avg
* LOAD_AVG_MAX
;
803 if (entity_is_task(se
)) {
804 struct task_struct
*p
= task_of(se
);
805 if (p
->sched_class
!= &fair_sched_class
) {
807 * For !fair tasks do:
809 update_cfs_rq_load_avg(now, cfs_rq, false);
810 attach_entity_load_avg(cfs_rq, se);
811 switched_from_fair(rq, p);
813 * such that the next switched_to_fair() has the
816 se
->avg
.last_update_time
= cfs_rq_clock_task(cfs_rq
);
821 attach_entity_cfs_rq(se
);
824 #else /* !CONFIG_SMP */
825 void init_entity_runnable_average(struct sched_entity
*se
)
828 void post_init_entity_util_avg(struct sched_entity
*se
)
831 static void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
)
834 #endif /* CONFIG_SMP */
837 * Update the current task's runtime statistics.
839 static void update_curr(struct cfs_rq
*cfs_rq
)
841 struct sched_entity
*curr
= cfs_rq
->curr
;
842 u64 now
= rq_clock_task(rq_of(cfs_rq
));
848 delta_exec
= now
- curr
->exec_start
;
849 if (unlikely((s64
)delta_exec
<= 0))
852 curr
->exec_start
= now
;
854 schedstat_set(curr
->statistics
.exec_max
,
855 max(delta_exec
, curr
->statistics
.exec_max
));
857 curr
->sum_exec_runtime
+= delta_exec
;
858 schedstat_add(cfs_rq
->exec_clock
, delta_exec
);
860 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
861 update_min_vruntime(cfs_rq
);
863 if (entity_is_task(curr
)) {
864 struct task_struct
*curtask
= task_of(curr
);
866 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
867 cpuacct_charge(curtask
, delta_exec
);
868 account_group_exec_runtime(curtask
, delta_exec
);
871 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
874 static void update_curr_fair(struct rq
*rq
)
876 update_curr(cfs_rq_of(&rq
->curr
->se
));
880 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
882 u64 wait_start
, prev_wait_start
;
884 if (!schedstat_enabled())
887 wait_start
= rq_clock(rq_of(cfs_rq
));
888 prev_wait_start
= schedstat_val(se
->statistics
.wait_start
);
890 if (entity_is_task(se
) && task_on_rq_migrating(task_of(se
)) &&
891 likely(wait_start
> prev_wait_start
))
892 wait_start
-= prev_wait_start
;
894 schedstat_set(se
->statistics
.wait_start
, wait_start
);
898 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
900 struct task_struct
*p
;
903 if (!schedstat_enabled())
906 delta
= rq_clock(rq_of(cfs_rq
)) - schedstat_val(se
->statistics
.wait_start
);
908 if (entity_is_task(se
)) {
910 if (task_on_rq_migrating(p
)) {
912 * Preserve migrating task's wait time so wait_start
913 * time stamp can be adjusted to accumulate wait time
914 * prior to migration.
916 schedstat_set(se
->statistics
.wait_start
, delta
);
919 trace_sched_stat_wait(p
, delta
);
922 schedstat_set(se
->statistics
.wait_max
,
923 max(schedstat_val(se
->statistics
.wait_max
), delta
));
924 schedstat_inc(se
->statistics
.wait_count
);
925 schedstat_add(se
->statistics
.wait_sum
, delta
);
926 schedstat_set(se
->statistics
.wait_start
, 0);
930 update_stats_enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
932 struct task_struct
*tsk
= NULL
;
933 u64 sleep_start
, block_start
;
935 if (!schedstat_enabled())
938 sleep_start
= schedstat_val(se
->statistics
.sleep_start
);
939 block_start
= schedstat_val(se
->statistics
.block_start
);
941 if (entity_is_task(se
))
945 u64 delta
= rq_clock(rq_of(cfs_rq
)) - sleep_start
;
950 if (unlikely(delta
> schedstat_val(se
->statistics
.sleep_max
)))
951 schedstat_set(se
->statistics
.sleep_max
, delta
);
953 schedstat_set(se
->statistics
.sleep_start
, 0);
954 schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
957 account_scheduler_latency(tsk
, delta
>> 10, 1);
958 trace_sched_stat_sleep(tsk
, delta
);
962 u64 delta
= rq_clock(rq_of(cfs_rq
)) - block_start
;
967 if (unlikely(delta
> schedstat_val(se
->statistics
.block_max
)))
968 schedstat_set(se
->statistics
.block_max
, delta
);
970 schedstat_set(se
->statistics
.block_start
, 0);
971 schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
974 if (tsk
->in_iowait
) {
975 schedstat_add(se
->statistics
.iowait_sum
, delta
);
976 schedstat_inc(se
->statistics
.iowait_count
);
977 trace_sched_stat_iowait(tsk
, delta
);
980 trace_sched_stat_blocked(tsk
, delta
);
983 * Blocking time is in units of nanosecs, so shift by
984 * 20 to get a milliseconds-range estimation of the
985 * amount of time that the task spent sleeping:
987 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
988 profile_hits(SLEEP_PROFILING
,
989 (void *)get_wchan(tsk
),
992 account_scheduler_latency(tsk
, delta
>> 10, 0);
998 * Task is being enqueued - update stats:
1001 update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1003 if (!schedstat_enabled())
1007 * Are we enqueueing a waiting task? (for current tasks
1008 * a dequeue/enqueue event is a NOP)
1010 if (se
!= cfs_rq
->curr
)
1011 update_stats_wait_start(cfs_rq
, se
);
1013 if (flags
& ENQUEUE_WAKEUP
)
1014 update_stats_enqueue_sleeper(cfs_rq
, se
);
1018 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1021 if (!schedstat_enabled())
1025 * Mark the end of the wait period if dequeueing a
1028 if (se
!= cfs_rq
->curr
)
1029 update_stats_wait_end(cfs_rq
, se
);
1031 if ((flags
& DEQUEUE_SLEEP
) && entity_is_task(se
)) {
1032 struct task_struct
*tsk
= task_of(se
);
1034 if (tsk
->state
& TASK_INTERRUPTIBLE
)
1035 schedstat_set(se
->statistics
.sleep_start
,
1036 rq_clock(rq_of(cfs_rq
)));
1037 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
1038 schedstat_set(se
->statistics
.block_start
,
1039 rq_clock(rq_of(cfs_rq
)));
1044 * We are picking a new current task - update its stats:
1047 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1050 * We are starting a new run period:
1052 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
1055 /**************************************************
1056 * Scheduling class queueing methods:
1059 #ifdef CONFIG_NUMA_BALANCING
1061 * Approximate time to scan a full NUMA task in ms. The task scan period is
1062 * calculated based on the tasks virtual memory size and
1063 * numa_balancing_scan_size.
1065 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
1066 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
1068 /* Portion of address space to scan in MB */
1069 unsigned int sysctl_numa_balancing_scan_size
= 256;
1071 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1072 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
1074 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
1076 unsigned long rss
= 0;
1077 unsigned long nr_scan_pages
;
1080 * Calculations based on RSS as non-present and empty pages are skipped
1081 * by the PTE scanner and NUMA hinting faults should be trapped based
1084 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
1085 rss
= get_mm_rss(p
->mm
);
1087 rss
= nr_scan_pages
;
1089 rss
= round_up(rss
, nr_scan_pages
);
1090 return rss
/ nr_scan_pages
;
1093 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1094 #define MAX_SCAN_WINDOW 2560
1096 static unsigned int task_scan_min(struct task_struct
*p
)
1098 unsigned int scan_size
= READ_ONCE(sysctl_numa_balancing_scan_size
);
1099 unsigned int scan
, floor
;
1100 unsigned int windows
= 1;
1102 if (scan_size
< MAX_SCAN_WINDOW
)
1103 windows
= MAX_SCAN_WINDOW
/ scan_size
;
1104 floor
= 1000 / windows
;
1106 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
1107 return max_t(unsigned int, floor
, scan
);
1110 static unsigned int task_scan_max(struct task_struct
*p
)
1112 unsigned int smin
= task_scan_min(p
);
1115 /* Watch for min being lower than max due to floor calculations */
1116 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
1117 return max(smin
, smax
);
1120 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
1122 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
1123 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
1126 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
1128 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
1129 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
1135 spinlock_t lock
; /* nr_tasks, tasks */
1140 struct rcu_head rcu
;
1141 unsigned long total_faults
;
1142 unsigned long max_faults_cpu
;
1144 * Faults_cpu is used to decide whether memory should move
1145 * towards the CPU. As a consequence, these stats are weighted
1146 * more by CPU use than by memory faults.
1148 unsigned long *faults_cpu
;
1149 unsigned long faults
[0];
1152 /* Shared or private faults. */
1153 #define NR_NUMA_HINT_FAULT_TYPES 2
1155 /* Memory and CPU locality */
1156 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1158 /* Averaged statistics, and temporary buffers. */
1159 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1161 pid_t
task_numa_group_id(struct task_struct
*p
)
1163 return p
->numa_group
? p
->numa_group
->gid
: 0;
1167 * The averaged statistics, shared & private, memory & cpu,
1168 * occupy the first half of the array. The second half of the
1169 * array is for current counters, which are averaged into the
1170 * first set by task_numa_placement.
1172 static inline int task_faults_idx(enum numa_faults_stats s
, int nid
, int priv
)
1174 return NR_NUMA_HINT_FAULT_TYPES
* (s
* nr_node_ids
+ nid
) + priv
;
1177 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
1179 if (!p
->numa_faults
)
1182 return p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1183 p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1186 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
1191 return p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1192 p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1195 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
1197 return group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1198 group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1202 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1203 * considered part of a numa group's pseudo-interleaving set. Migrations
1204 * between these nodes are slowed down, to allow things to settle down.
1206 #define ACTIVE_NODE_FRACTION 3
1208 static bool numa_is_active_node(int nid
, struct numa_group
*ng
)
1210 return group_faults_cpu(ng
, nid
) * ACTIVE_NODE_FRACTION
> ng
->max_faults_cpu
;
1213 /* Handle placement on systems where not all nodes are directly connected. */
1214 static unsigned long score_nearby_nodes(struct task_struct
*p
, int nid
,
1215 int maxdist
, bool task
)
1217 unsigned long score
= 0;
1221 * All nodes are directly connected, and the same distance
1222 * from each other. No need for fancy placement algorithms.
1224 if (sched_numa_topology_type
== NUMA_DIRECT
)
1228 * This code is called for each node, introducing N^2 complexity,
1229 * which should be ok given the number of nodes rarely exceeds 8.
1231 for_each_online_node(node
) {
1232 unsigned long faults
;
1233 int dist
= node_distance(nid
, node
);
1236 * The furthest away nodes in the system are not interesting
1237 * for placement; nid was already counted.
1239 if (dist
== sched_max_numa_distance
|| node
== nid
)
1243 * On systems with a backplane NUMA topology, compare groups
1244 * of nodes, and move tasks towards the group with the most
1245 * memory accesses. When comparing two nodes at distance
1246 * "hoplimit", only nodes closer by than "hoplimit" are part
1247 * of each group. Skip other nodes.
1249 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1253 /* Add up the faults from nearby nodes. */
1255 faults
= task_faults(p
, node
);
1257 faults
= group_faults(p
, node
);
1260 * On systems with a glueless mesh NUMA topology, there are
1261 * no fixed "groups of nodes". Instead, nodes that are not
1262 * directly connected bounce traffic through intermediate
1263 * nodes; a numa_group can occupy any set of nodes.
1264 * The further away a node is, the less the faults count.
1265 * This seems to result in good task placement.
1267 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1268 faults
*= (sched_max_numa_distance
- dist
);
1269 faults
/= (sched_max_numa_distance
- LOCAL_DISTANCE
);
1279 * These return the fraction of accesses done by a particular task, or
1280 * task group, on a particular numa node. The group weight is given a
1281 * larger multiplier, in order to group tasks together that are almost
1282 * evenly spread out between numa nodes.
1284 static inline unsigned long task_weight(struct task_struct
*p
, int nid
,
1287 unsigned long faults
, total_faults
;
1289 if (!p
->numa_faults
)
1292 total_faults
= p
->total_numa_faults
;
1297 faults
= task_faults(p
, nid
);
1298 faults
+= score_nearby_nodes(p
, nid
, dist
, true);
1300 return 1000 * faults
/ total_faults
;
1303 static inline unsigned long group_weight(struct task_struct
*p
, int nid
,
1306 unsigned long faults
, total_faults
;
1311 total_faults
= p
->numa_group
->total_faults
;
1316 faults
= group_faults(p
, nid
);
1317 faults
+= score_nearby_nodes(p
, nid
, dist
, false);
1319 return 1000 * faults
/ total_faults
;
1322 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
1323 int src_nid
, int dst_cpu
)
1325 struct numa_group
*ng
= p
->numa_group
;
1326 int dst_nid
= cpu_to_node(dst_cpu
);
1327 int last_cpupid
, this_cpupid
;
1329 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
1332 * Multi-stage node selection is used in conjunction with a periodic
1333 * migration fault to build a temporal task<->page relation. By using
1334 * a two-stage filter we remove short/unlikely relations.
1336 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1337 * a task's usage of a particular page (n_p) per total usage of this
1338 * page (n_t) (in a given time-span) to a probability.
1340 * Our periodic faults will sample this probability and getting the
1341 * same result twice in a row, given these samples are fully
1342 * independent, is then given by P(n)^2, provided our sample period
1343 * is sufficiently short compared to the usage pattern.
1345 * This quadric squishes small probabilities, making it less likely we
1346 * act on an unlikely task<->page relation.
1348 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
1349 if (!cpupid_pid_unset(last_cpupid
) &&
1350 cpupid_to_nid(last_cpupid
) != dst_nid
)
1353 /* Always allow migrate on private faults */
1354 if (cpupid_match_pid(p
, last_cpupid
))
1357 /* A shared fault, but p->numa_group has not been set up yet. */
1362 * Destination node is much more heavily used than the source
1363 * node? Allow migration.
1365 if (group_faults_cpu(ng
, dst_nid
) > group_faults_cpu(ng
, src_nid
) *
1366 ACTIVE_NODE_FRACTION
)
1370 * Distribute memory according to CPU & memory use on each node,
1371 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1373 * faults_cpu(dst) 3 faults_cpu(src)
1374 * --------------- * - > ---------------
1375 * faults_mem(dst) 4 faults_mem(src)
1377 return group_faults_cpu(ng
, dst_nid
) * group_faults(p
, src_nid
) * 3 >
1378 group_faults_cpu(ng
, src_nid
) * group_faults(p
, dst_nid
) * 4;
1381 static unsigned long weighted_cpuload(const int cpu
);
1382 static unsigned long source_load(int cpu
, int type
);
1383 static unsigned long target_load(int cpu
, int type
);
1384 static unsigned long capacity_of(int cpu
);
1385 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
1387 /* Cached statistics for all CPUs within a node */
1389 unsigned long nr_running
;
1392 /* Total compute capacity of CPUs on a node */
1393 unsigned long compute_capacity
;
1395 /* Approximate capacity in terms of runnable tasks on a node */
1396 unsigned long task_capacity
;
1397 int has_free_capacity
;
1401 * XXX borrowed from update_sg_lb_stats
1403 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1405 int smt
, cpu
, cpus
= 0;
1406 unsigned long capacity
;
1408 memset(ns
, 0, sizeof(*ns
));
1409 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1410 struct rq
*rq
= cpu_rq(cpu
);
1412 ns
->nr_running
+= rq
->nr_running
;
1413 ns
->load
+= weighted_cpuload(cpu
);
1414 ns
->compute_capacity
+= capacity_of(cpu
);
1420 * If we raced with hotplug and there are no CPUs left in our mask
1421 * the @ns structure is NULL'ed and task_numa_compare() will
1422 * not find this node attractive.
1424 * We'll either bail at !has_free_capacity, or we'll detect a huge
1425 * imbalance and bail there.
1430 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1431 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, ns
->compute_capacity
);
1432 capacity
= cpus
/ smt
; /* cores */
1434 ns
->task_capacity
= min_t(unsigned, capacity
,
1435 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
));
1436 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1439 struct task_numa_env
{
1440 struct task_struct
*p
;
1442 int src_cpu
, src_nid
;
1443 int dst_cpu
, dst_nid
;
1445 struct numa_stats src_stats
, dst_stats
;
1450 struct task_struct
*best_task
;
1455 static void task_numa_assign(struct task_numa_env
*env
,
1456 struct task_struct
*p
, long imp
)
1459 put_task_struct(env
->best_task
);
1464 env
->best_imp
= imp
;
1465 env
->best_cpu
= env
->dst_cpu
;
1468 static bool load_too_imbalanced(long src_load
, long dst_load
,
1469 struct task_numa_env
*env
)
1472 long orig_src_load
, orig_dst_load
;
1473 long src_capacity
, dst_capacity
;
1476 * The load is corrected for the CPU capacity available on each node.
1479 * ------------ vs ---------
1480 * src_capacity dst_capacity
1482 src_capacity
= env
->src_stats
.compute_capacity
;
1483 dst_capacity
= env
->dst_stats
.compute_capacity
;
1485 /* We care about the slope of the imbalance, not the direction. */
1486 if (dst_load
< src_load
)
1487 swap(dst_load
, src_load
);
1489 /* Is the difference below the threshold? */
1490 imb
= dst_load
* src_capacity
* 100 -
1491 src_load
* dst_capacity
* env
->imbalance_pct
;
1496 * The imbalance is above the allowed threshold.
1497 * Compare it with the old imbalance.
1499 orig_src_load
= env
->src_stats
.load
;
1500 orig_dst_load
= env
->dst_stats
.load
;
1502 if (orig_dst_load
< orig_src_load
)
1503 swap(orig_dst_load
, orig_src_load
);
1505 old_imb
= orig_dst_load
* src_capacity
* 100 -
1506 orig_src_load
* dst_capacity
* env
->imbalance_pct
;
1508 /* Would this change make things worse? */
1509 return (imb
> old_imb
);
1513 * This checks if the overall compute and NUMA accesses of the system would
1514 * be improved if the source tasks was migrated to the target dst_cpu taking
1515 * into account that it might be best if task running on the dst_cpu should
1516 * be exchanged with the source task
1518 static void task_numa_compare(struct task_numa_env
*env
,
1519 long taskimp
, long groupimp
)
1521 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1522 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1523 struct task_struct
*cur
;
1524 long src_load
, dst_load
;
1526 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1528 int dist
= env
->dist
;
1531 cur
= task_rcu_dereference(&dst_rq
->curr
);
1532 if (cur
&& ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
)))
1536 * Because we have preemption enabled we can get migrated around and
1537 * end try selecting ourselves (current == env->p) as a swap candidate.
1543 * "imp" is the fault differential for the source task between the
1544 * source and destination node. Calculate the total differential for
1545 * the source task and potential destination task. The more negative
1546 * the value is, the more rmeote accesses that would be expected to
1547 * be incurred if the tasks were swapped.
1550 /* Skip this swap candidate if cannot move to the source cpu */
1551 if (!cpumask_test_cpu(env
->src_cpu
, &cur
->cpus_allowed
))
1555 * If dst and source tasks are in the same NUMA group, or not
1556 * in any group then look only at task weights.
1558 if (cur
->numa_group
== env
->p
->numa_group
) {
1559 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1560 task_weight(cur
, env
->dst_nid
, dist
);
1562 * Add some hysteresis to prevent swapping the
1563 * tasks within a group over tiny differences.
1565 if (cur
->numa_group
)
1569 * Compare the group weights. If a task is all by
1570 * itself (not part of a group), use the task weight
1573 if (cur
->numa_group
)
1574 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1575 group_weight(cur
, env
->dst_nid
, dist
);
1577 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1578 task_weight(cur
, env
->dst_nid
, dist
);
1582 if (imp
<= env
->best_imp
&& moveimp
<= env
->best_imp
)
1586 /* Is there capacity at our destination? */
1587 if (env
->src_stats
.nr_running
<= env
->src_stats
.task_capacity
&&
1588 !env
->dst_stats
.has_free_capacity
)
1594 /* Balance doesn't matter much if we're running a task per cpu */
1595 if (imp
> env
->best_imp
&& src_rq
->nr_running
== 1 &&
1596 dst_rq
->nr_running
== 1)
1600 * In the overloaded case, try and keep the load balanced.
1603 load
= task_h_load(env
->p
);
1604 dst_load
= env
->dst_stats
.load
+ load
;
1605 src_load
= env
->src_stats
.load
- load
;
1607 if (moveimp
> imp
&& moveimp
> env
->best_imp
) {
1609 * If the improvement from just moving env->p direction is
1610 * better than swapping tasks around, check if a move is
1611 * possible. Store a slightly smaller score than moveimp,
1612 * so an actually idle CPU will win.
1614 if (!load_too_imbalanced(src_load
, dst_load
, env
)) {
1621 if (imp
<= env
->best_imp
)
1625 load
= task_h_load(cur
);
1630 if (load_too_imbalanced(src_load
, dst_load
, env
))
1634 * One idle CPU per node is evaluated for a task numa move.
1635 * Call select_idle_sibling to maybe find a better one.
1639 * select_idle_siblings() uses an per-cpu cpumask that
1640 * can be used from IRQ context.
1642 local_irq_disable();
1643 env
->dst_cpu
= select_idle_sibling(env
->p
, env
->src_cpu
,
1649 task_numa_assign(env
, cur
, imp
);
1654 static void task_numa_find_cpu(struct task_numa_env
*env
,
1655 long taskimp
, long groupimp
)
1659 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1660 /* Skip this CPU if the source task cannot migrate */
1661 if (!cpumask_test_cpu(cpu
, &env
->p
->cpus_allowed
))
1665 task_numa_compare(env
, taskimp
, groupimp
);
1669 /* Only move tasks to a NUMA node less busy than the current node. */
1670 static bool numa_has_capacity(struct task_numa_env
*env
)
1672 struct numa_stats
*src
= &env
->src_stats
;
1673 struct numa_stats
*dst
= &env
->dst_stats
;
1675 if (src
->has_free_capacity
&& !dst
->has_free_capacity
)
1679 * Only consider a task move if the source has a higher load
1680 * than the destination, corrected for CPU capacity on each node.
1682 * src->load dst->load
1683 * --------------------- vs ---------------------
1684 * src->compute_capacity dst->compute_capacity
1686 if (src
->load
* dst
->compute_capacity
* env
->imbalance_pct
>
1688 dst
->load
* src
->compute_capacity
* 100)
1694 static int task_numa_migrate(struct task_struct
*p
)
1696 struct task_numa_env env
= {
1699 .src_cpu
= task_cpu(p
),
1700 .src_nid
= task_node(p
),
1702 .imbalance_pct
= 112,
1708 struct sched_domain
*sd
;
1709 unsigned long taskweight
, groupweight
;
1711 long taskimp
, groupimp
;
1714 * Pick the lowest SD_NUMA domain, as that would have the smallest
1715 * imbalance and would be the first to start moving tasks about.
1717 * And we want to avoid any moving of tasks about, as that would create
1718 * random movement of tasks -- counter the numa conditions we're trying
1722 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1724 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1728 * Cpusets can break the scheduler domain tree into smaller
1729 * balance domains, some of which do not cross NUMA boundaries.
1730 * Tasks that are "trapped" in such domains cannot be migrated
1731 * elsewhere, so there is no point in (re)trying.
1733 if (unlikely(!sd
)) {
1734 p
->numa_preferred_nid
= task_node(p
);
1738 env
.dst_nid
= p
->numa_preferred_nid
;
1739 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1740 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1741 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1742 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1743 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
1744 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
1745 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1747 /* Try to find a spot on the preferred nid. */
1748 if (numa_has_capacity(&env
))
1749 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1752 * Look at other nodes in these cases:
1753 * - there is no space available on the preferred_nid
1754 * - the task is part of a numa_group that is interleaved across
1755 * multiple NUMA nodes; in order to better consolidate the group,
1756 * we need to check other locations.
1758 if (env
.best_cpu
== -1 || (p
->numa_group
&& p
->numa_group
->active_nodes
> 1)) {
1759 for_each_online_node(nid
) {
1760 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1763 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1764 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1766 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1767 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1770 /* Only consider nodes where both task and groups benefit */
1771 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
1772 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
1773 if (taskimp
< 0 && groupimp
< 0)
1778 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1779 if (numa_has_capacity(&env
))
1780 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1785 * If the task is part of a workload that spans multiple NUMA nodes,
1786 * and is migrating into one of the workload's active nodes, remember
1787 * this node as the task's preferred numa node, so the workload can
1789 * A task that migrated to a second choice node will be better off
1790 * trying for a better one later. Do not set the preferred node here.
1792 if (p
->numa_group
) {
1793 struct numa_group
*ng
= p
->numa_group
;
1795 if (env
.best_cpu
== -1)
1800 if (ng
->active_nodes
> 1 && numa_is_active_node(env
.dst_nid
, ng
))
1801 sched_setnuma(p
, env
.dst_nid
);
1804 /* No better CPU than the current one was found. */
1805 if (env
.best_cpu
== -1)
1809 * Reset the scan period if the task is being rescheduled on an
1810 * alternative node to recheck if the tasks is now properly placed.
1812 p
->numa_scan_period
= task_scan_min(p
);
1814 if (env
.best_task
== NULL
) {
1815 ret
= migrate_task_to(p
, env
.best_cpu
);
1817 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1821 ret
= migrate_swap(p
, env
.best_task
);
1823 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1824 put_task_struct(env
.best_task
);
1828 /* Attempt to migrate a task to a CPU on the preferred node. */
1829 static void numa_migrate_preferred(struct task_struct
*p
)
1831 unsigned long interval
= HZ
;
1833 /* This task has no NUMA fault statistics yet */
1834 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults
))
1837 /* Periodically retry migrating the task to the preferred node */
1838 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1839 p
->numa_migrate_retry
= jiffies
+ interval
;
1841 /* Success if task is already running on preferred CPU */
1842 if (task_node(p
) == p
->numa_preferred_nid
)
1845 /* Otherwise, try migrate to a CPU on the preferred node */
1846 task_numa_migrate(p
);
1850 * Find out how many nodes on the workload is actively running on. Do this by
1851 * tracking the nodes from which NUMA hinting faults are triggered. This can
1852 * be different from the set of nodes where the workload's memory is currently
1855 static void numa_group_count_active_nodes(struct numa_group
*numa_group
)
1857 unsigned long faults
, max_faults
= 0;
1858 int nid
, active_nodes
= 0;
1860 for_each_online_node(nid
) {
1861 faults
= group_faults_cpu(numa_group
, nid
);
1862 if (faults
> max_faults
)
1863 max_faults
= faults
;
1866 for_each_online_node(nid
) {
1867 faults
= group_faults_cpu(numa_group
, nid
);
1868 if (faults
* ACTIVE_NODE_FRACTION
> max_faults
)
1872 numa_group
->max_faults_cpu
= max_faults
;
1873 numa_group
->active_nodes
= active_nodes
;
1877 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1878 * increments. The more local the fault statistics are, the higher the scan
1879 * period will be for the next scan window. If local/(local+remote) ratio is
1880 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1881 * the scan period will decrease. Aim for 70% local accesses.
1883 #define NUMA_PERIOD_SLOTS 10
1884 #define NUMA_PERIOD_THRESHOLD 7
1887 * Increase the scan period (slow down scanning) if the majority of
1888 * our memory is already on our local node, or if the majority of
1889 * the page accesses are shared with other processes.
1890 * Otherwise, decrease the scan period.
1892 static void update_task_scan_period(struct task_struct
*p
,
1893 unsigned long shared
, unsigned long private)
1895 unsigned int period_slot
;
1899 unsigned long remote
= p
->numa_faults_locality
[0];
1900 unsigned long local
= p
->numa_faults_locality
[1];
1903 * If there were no record hinting faults then either the task is
1904 * completely idle or all activity is areas that are not of interest
1905 * to automatic numa balancing. Related to that, if there were failed
1906 * migration then it implies we are migrating too quickly or the local
1907 * node is overloaded. In either case, scan slower
1909 if (local
+ shared
== 0 || p
->numa_faults_locality
[2]) {
1910 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1911 p
->numa_scan_period
<< 1);
1913 p
->mm
->numa_next_scan
= jiffies
+
1914 msecs_to_jiffies(p
->numa_scan_period
);
1920 * Prepare to scale scan period relative to the current period.
1921 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1922 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1923 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1925 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1926 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1927 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1928 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1931 diff
= slot
* period_slot
;
1933 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1936 * Scale scan rate increases based on sharing. There is an
1937 * inverse relationship between the degree of sharing and
1938 * the adjustment made to the scanning period. Broadly
1939 * speaking the intent is that there is little point
1940 * scanning faster if shared accesses dominate as it may
1941 * simply bounce migrations uselessly
1943 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
+ 1));
1944 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1947 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1948 task_scan_min(p
), task_scan_max(p
));
1949 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1953 * Get the fraction of time the task has been running since the last
1954 * NUMA placement cycle. The scheduler keeps similar statistics, but
1955 * decays those on a 32ms period, which is orders of magnitude off
1956 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1957 * stats only if the task is so new there are no NUMA statistics yet.
1959 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1961 u64 runtime
, delta
, now
;
1962 /* Use the start of this time slice to avoid calculations. */
1963 now
= p
->se
.exec_start
;
1964 runtime
= p
->se
.sum_exec_runtime
;
1966 if (p
->last_task_numa_placement
) {
1967 delta
= runtime
- p
->last_sum_exec_runtime
;
1968 *period
= now
- p
->last_task_numa_placement
;
1970 delta
= p
->se
.avg
.load_sum
/ p
->se
.load
.weight
;
1971 *period
= LOAD_AVG_MAX
;
1974 p
->last_sum_exec_runtime
= runtime
;
1975 p
->last_task_numa_placement
= now
;
1981 * Determine the preferred nid for a task in a numa_group. This needs to
1982 * be done in a way that produces consistent results with group_weight,
1983 * otherwise workloads might not converge.
1985 static int preferred_group_nid(struct task_struct
*p
, int nid
)
1990 /* Direct connections between all NUMA nodes. */
1991 if (sched_numa_topology_type
== NUMA_DIRECT
)
1995 * On a system with glueless mesh NUMA topology, group_weight
1996 * scores nodes according to the number of NUMA hinting faults on
1997 * both the node itself, and on nearby nodes.
1999 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
2000 unsigned long score
, max_score
= 0;
2001 int node
, max_node
= nid
;
2003 dist
= sched_max_numa_distance
;
2005 for_each_online_node(node
) {
2006 score
= group_weight(p
, node
, dist
);
2007 if (score
> max_score
) {
2016 * Finding the preferred nid in a system with NUMA backplane
2017 * interconnect topology is more involved. The goal is to locate
2018 * tasks from numa_groups near each other in the system, and
2019 * untangle workloads from different sides of the system. This requires
2020 * searching down the hierarchy of node groups, recursively searching
2021 * inside the highest scoring group of nodes. The nodemask tricks
2022 * keep the complexity of the search down.
2024 nodes
= node_online_map
;
2025 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
2026 unsigned long max_faults
= 0;
2027 nodemask_t max_group
= NODE_MASK_NONE
;
2030 /* Are there nodes at this distance from each other? */
2031 if (!find_numa_distance(dist
))
2034 for_each_node_mask(a
, nodes
) {
2035 unsigned long faults
= 0;
2036 nodemask_t this_group
;
2037 nodes_clear(this_group
);
2039 /* Sum group's NUMA faults; includes a==b case. */
2040 for_each_node_mask(b
, nodes
) {
2041 if (node_distance(a
, b
) < dist
) {
2042 faults
+= group_faults(p
, b
);
2043 node_set(b
, this_group
);
2044 node_clear(b
, nodes
);
2048 /* Remember the top group. */
2049 if (faults
> max_faults
) {
2050 max_faults
= faults
;
2051 max_group
= this_group
;
2053 * subtle: at the smallest distance there is
2054 * just one node left in each "group", the
2055 * winner is the preferred nid.
2060 /* Next round, evaluate the nodes within max_group. */
2068 static void task_numa_placement(struct task_struct
*p
)
2070 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
2071 unsigned long max_faults
= 0, max_group_faults
= 0;
2072 unsigned long fault_types
[2] = { 0, 0 };
2073 unsigned long total_faults
;
2074 u64 runtime
, period
;
2075 spinlock_t
*group_lock
= NULL
;
2078 * The p->mm->numa_scan_seq field gets updated without
2079 * exclusive access. Use READ_ONCE() here to ensure
2080 * that the field is read in a single access:
2082 seq
= READ_ONCE(p
->mm
->numa_scan_seq
);
2083 if (p
->numa_scan_seq
== seq
)
2085 p
->numa_scan_seq
= seq
;
2086 p
->numa_scan_period_max
= task_scan_max(p
);
2088 total_faults
= p
->numa_faults_locality
[0] +
2089 p
->numa_faults_locality
[1];
2090 runtime
= numa_get_avg_runtime(p
, &period
);
2092 /* If the task is part of a group prevent parallel updates to group stats */
2093 if (p
->numa_group
) {
2094 group_lock
= &p
->numa_group
->lock
;
2095 spin_lock_irq(group_lock
);
2098 /* Find the node with the highest number of faults */
2099 for_each_online_node(nid
) {
2100 /* Keep track of the offsets in numa_faults array */
2101 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
2102 unsigned long faults
= 0, group_faults
= 0;
2105 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
2106 long diff
, f_diff
, f_weight
;
2108 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
2109 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
2110 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
2111 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
2113 /* Decay existing window, copy faults since last scan */
2114 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
2115 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
2116 p
->numa_faults
[membuf_idx
] = 0;
2119 * Normalize the faults_from, so all tasks in a group
2120 * count according to CPU use, instead of by the raw
2121 * number of faults. Tasks with little runtime have
2122 * little over-all impact on throughput, and thus their
2123 * faults are less important.
2125 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
2126 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
2128 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
2129 p
->numa_faults
[cpubuf_idx
] = 0;
2131 p
->numa_faults
[mem_idx
] += diff
;
2132 p
->numa_faults
[cpu_idx
] += f_diff
;
2133 faults
+= p
->numa_faults
[mem_idx
];
2134 p
->total_numa_faults
+= diff
;
2135 if (p
->numa_group
) {
2137 * safe because we can only change our own group
2139 * mem_idx represents the offset for a given
2140 * nid and priv in a specific region because it
2141 * is at the beginning of the numa_faults array.
2143 p
->numa_group
->faults
[mem_idx
] += diff
;
2144 p
->numa_group
->faults_cpu
[mem_idx
] += f_diff
;
2145 p
->numa_group
->total_faults
+= diff
;
2146 group_faults
+= p
->numa_group
->faults
[mem_idx
];
2150 if (faults
> max_faults
) {
2151 max_faults
= faults
;
2155 if (group_faults
> max_group_faults
) {
2156 max_group_faults
= group_faults
;
2157 max_group_nid
= nid
;
2161 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
2163 if (p
->numa_group
) {
2164 numa_group_count_active_nodes(p
->numa_group
);
2165 spin_unlock_irq(group_lock
);
2166 max_nid
= preferred_group_nid(p
, max_group_nid
);
2170 /* Set the new preferred node */
2171 if (max_nid
!= p
->numa_preferred_nid
)
2172 sched_setnuma(p
, max_nid
);
2174 if (task_node(p
) != p
->numa_preferred_nid
)
2175 numa_migrate_preferred(p
);
2179 static inline int get_numa_group(struct numa_group
*grp
)
2181 return atomic_inc_not_zero(&grp
->refcount
);
2184 static inline void put_numa_group(struct numa_group
*grp
)
2186 if (atomic_dec_and_test(&grp
->refcount
))
2187 kfree_rcu(grp
, rcu
);
2190 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
2193 struct numa_group
*grp
, *my_grp
;
2194 struct task_struct
*tsk
;
2196 int cpu
= cpupid_to_cpu(cpupid
);
2199 if (unlikely(!p
->numa_group
)) {
2200 unsigned int size
= sizeof(struct numa_group
) +
2201 4*nr_node_ids
*sizeof(unsigned long);
2203 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
2207 atomic_set(&grp
->refcount
, 1);
2208 grp
->active_nodes
= 1;
2209 grp
->max_faults_cpu
= 0;
2210 spin_lock_init(&grp
->lock
);
2212 /* Second half of the array tracks nids where faults happen */
2213 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
2216 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2217 grp
->faults
[i
] = p
->numa_faults
[i
];
2219 grp
->total_faults
= p
->total_numa_faults
;
2222 rcu_assign_pointer(p
->numa_group
, grp
);
2226 tsk
= READ_ONCE(cpu_rq(cpu
)->curr
);
2228 if (!cpupid_match_pid(tsk
, cpupid
))
2231 grp
= rcu_dereference(tsk
->numa_group
);
2235 my_grp
= p
->numa_group
;
2240 * Only join the other group if its bigger; if we're the bigger group,
2241 * the other task will join us.
2243 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
2247 * Tie-break on the grp address.
2249 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
2252 /* Always join threads in the same process. */
2253 if (tsk
->mm
== current
->mm
)
2256 /* Simple filter to avoid false positives due to PID collisions */
2257 if (flags
& TNF_SHARED
)
2260 /* Update priv based on whether false sharing was detected */
2263 if (join
&& !get_numa_group(grp
))
2271 BUG_ON(irqs_disabled());
2272 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
2274 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
2275 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
2276 grp
->faults
[i
] += p
->numa_faults
[i
];
2278 my_grp
->total_faults
-= p
->total_numa_faults
;
2279 grp
->total_faults
+= p
->total_numa_faults
;
2284 spin_unlock(&my_grp
->lock
);
2285 spin_unlock_irq(&grp
->lock
);
2287 rcu_assign_pointer(p
->numa_group
, grp
);
2289 put_numa_group(my_grp
);
2297 void task_numa_free(struct task_struct
*p
)
2299 struct numa_group
*grp
= p
->numa_group
;
2300 void *numa_faults
= p
->numa_faults
;
2301 unsigned long flags
;
2305 spin_lock_irqsave(&grp
->lock
, flags
);
2306 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2307 grp
->faults
[i
] -= p
->numa_faults
[i
];
2308 grp
->total_faults
-= p
->total_numa_faults
;
2311 spin_unlock_irqrestore(&grp
->lock
, flags
);
2312 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2313 put_numa_group(grp
);
2316 p
->numa_faults
= NULL
;
2321 * Got a PROT_NONE fault for a page on @node.
2323 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2325 struct task_struct
*p
= current
;
2326 bool migrated
= flags
& TNF_MIGRATED
;
2327 int cpu_node
= task_node(current
);
2328 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2329 struct numa_group
*ng
;
2332 if (!static_branch_likely(&sched_numa_balancing
))
2335 /* for example, ksmd faulting in a user's mm */
2339 /* Allocate buffer to track faults on a per-node basis */
2340 if (unlikely(!p
->numa_faults
)) {
2341 int size
= sizeof(*p
->numa_faults
) *
2342 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2344 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2345 if (!p
->numa_faults
)
2348 p
->total_numa_faults
= 0;
2349 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2353 * First accesses are treated as private, otherwise consider accesses
2354 * to be private if the accessing pid has not changed
2356 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2359 priv
= cpupid_match_pid(p
, last_cpupid
);
2360 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2361 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2365 * If a workload spans multiple NUMA nodes, a shared fault that
2366 * occurs wholly within the set of nodes that the workload is
2367 * actively using should be counted as local. This allows the
2368 * scan rate to slow down when a workload has settled down.
2371 if (!priv
&& !local
&& ng
&& ng
->active_nodes
> 1 &&
2372 numa_is_active_node(cpu_node
, ng
) &&
2373 numa_is_active_node(mem_node
, ng
))
2376 task_numa_placement(p
);
2379 * Retry task to preferred node migration periodically, in case it
2380 * case it previously failed, or the scheduler moved us.
2382 if (time_after(jiffies
, p
->numa_migrate_retry
))
2383 numa_migrate_preferred(p
);
2386 p
->numa_pages_migrated
+= pages
;
2387 if (flags
& TNF_MIGRATE_FAIL
)
2388 p
->numa_faults_locality
[2] += pages
;
2390 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2391 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2392 p
->numa_faults_locality
[local
] += pages
;
2395 static void reset_ptenuma_scan(struct task_struct
*p
)
2398 * We only did a read acquisition of the mmap sem, so
2399 * p->mm->numa_scan_seq is written to without exclusive access
2400 * and the update is not guaranteed to be atomic. That's not
2401 * much of an issue though, since this is just used for
2402 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2403 * expensive, to avoid any form of compiler optimizations:
2405 WRITE_ONCE(p
->mm
->numa_scan_seq
, READ_ONCE(p
->mm
->numa_scan_seq
) + 1);
2406 p
->mm
->numa_scan_offset
= 0;
2410 * The expensive part of numa migration is done from task_work context.
2411 * Triggered from task_tick_numa().
2413 void task_numa_work(struct callback_head
*work
)
2415 unsigned long migrate
, next_scan
, now
= jiffies
;
2416 struct task_struct
*p
= current
;
2417 struct mm_struct
*mm
= p
->mm
;
2418 u64 runtime
= p
->se
.sum_exec_runtime
;
2419 struct vm_area_struct
*vma
;
2420 unsigned long start
, end
;
2421 unsigned long nr_pte_updates
= 0;
2422 long pages
, virtpages
;
2424 SCHED_WARN_ON(p
!= container_of(work
, struct task_struct
, numa_work
));
2426 work
->next
= work
; /* protect against double add */
2428 * Who cares about NUMA placement when they're dying.
2430 * NOTE: make sure not to dereference p->mm before this check,
2431 * exit_task_work() happens _after_ exit_mm() so we could be called
2432 * without p->mm even though we still had it when we enqueued this
2435 if (p
->flags
& PF_EXITING
)
2438 if (!mm
->numa_next_scan
) {
2439 mm
->numa_next_scan
= now
+
2440 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2444 * Enforce maximal scan/migration frequency..
2446 migrate
= mm
->numa_next_scan
;
2447 if (time_before(now
, migrate
))
2450 if (p
->numa_scan_period
== 0) {
2451 p
->numa_scan_period_max
= task_scan_max(p
);
2452 p
->numa_scan_period
= task_scan_min(p
);
2455 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2456 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2460 * Delay this task enough that another task of this mm will likely win
2461 * the next time around.
2463 p
->node_stamp
+= 2 * TICK_NSEC
;
2465 start
= mm
->numa_scan_offset
;
2466 pages
= sysctl_numa_balancing_scan_size
;
2467 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2468 virtpages
= pages
* 8; /* Scan up to this much virtual space */
2473 if (!down_read_trylock(&mm
->mmap_sem
))
2475 vma
= find_vma(mm
, start
);
2477 reset_ptenuma_scan(p
);
2481 for (; vma
; vma
= vma
->vm_next
) {
2482 if (!vma_migratable(vma
) || !vma_policy_mof(vma
) ||
2483 is_vm_hugetlb_page(vma
) || (vma
->vm_flags
& VM_MIXEDMAP
)) {
2488 * Shared library pages mapped by multiple processes are not
2489 * migrated as it is expected they are cache replicated. Avoid
2490 * hinting faults in read-only file-backed mappings or the vdso
2491 * as migrating the pages will be of marginal benefit.
2494 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2498 * Skip inaccessible VMAs to avoid any confusion between
2499 * PROT_NONE and NUMA hinting ptes
2501 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
2505 start
= max(start
, vma
->vm_start
);
2506 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2507 end
= min(end
, vma
->vm_end
);
2508 nr_pte_updates
= change_prot_numa(vma
, start
, end
);
2511 * Try to scan sysctl_numa_balancing_size worth of
2512 * hpages that have at least one present PTE that
2513 * is not already pte-numa. If the VMA contains
2514 * areas that are unused or already full of prot_numa
2515 * PTEs, scan up to virtpages, to skip through those
2519 pages
-= (end
- start
) >> PAGE_SHIFT
;
2520 virtpages
-= (end
- start
) >> PAGE_SHIFT
;
2523 if (pages
<= 0 || virtpages
<= 0)
2527 } while (end
!= vma
->vm_end
);
2532 * It is possible to reach the end of the VMA list but the last few
2533 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2534 * would find the !migratable VMA on the next scan but not reset the
2535 * scanner to the start so check it now.
2538 mm
->numa_scan_offset
= start
;
2540 reset_ptenuma_scan(p
);
2541 up_read(&mm
->mmap_sem
);
2544 * Make sure tasks use at least 32x as much time to run other code
2545 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2546 * Usually update_task_scan_period slows down scanning enough; on an
2547 * overloaded system we need to limit overhead on a per task basis.
2549 if (unlikely(p
->se
.sum_exec_runtime
!= runtime
)) {
2550 u64 diff
= p
->se
.sum_exec_runtime
- runtime
;
2551 p
->node_stamp
+= 32 * diff
;
2556 * Drive the periodic memory faults..
2558 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2560 struct callback_head
*work
= &curr
->numa_work
;
2564 * We don't care about NUMA placement if we don't have memory.
2566 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2570 * Using runtime rather than walltime has the dual advantage that
2571 * we (mostly) drive the selection from busy threads and that the
2572 * task needs to have done some actual work before we bother with
2575 now
= curr
->se
.sum_exec_runtime
;
2576 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2578 if (now
> curr
->node_stamp
+ period
) {
2579 if (!curr
->node_stamp
)
2580 curr
->numa_scan_period
= task_scan_min(curr
);
2581 curr
->node_stamp
+= period
;
2583 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2584 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2585 task_work_add(curr
, work
, true);
2590 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2594 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2598 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2601 #endif /* CONFIG_NUMA_BALANCING */
2604 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2606 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2607 if (!parent_entity(se
))
2608 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2610 if (entity_is_task(se
)) {
2611 struct rq
*rq
= rq_of(cfs_rq
);
2613 account_numa_enqueue(rq
, task_of(se
));
2614 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2617 cfs_rq
->nr_running
++;
2621 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2623 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2624 if (!parent_entity(se
))
2625 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2627 if (entity_is_task(se
)) {
2628 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2629 list_del_init(&se
->group_node
);
2632 cfs_rq
->nr_running
--;
2635 #ifdef CONFIG_FAIR_GROUP_SCHED
2637 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2639 long tg_weight
, load
, shares
;
2642 * This really should be: cfs_rq->avg.load_avg, but instead we use
2643 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2644 * the shares for small weight interactive tasks.
2646 load
= scale_load_down(cfs_rq
->load
.weight
);
2648 tg_weight
= atomic_long_read(&tg
->load_avg
);
2650 /* Ensure tg_weight >= load */
2651 tg_weight
-= cfs_rq
->tg_load_avg_contrib
;
2654 shares
= (tg
->shares
* load
);
2656 shares
/= tg_weight
;
2659 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2660 * of a group with small tg->shares value. It is a floor value which is
2661 * assigned as a minimum load.weight to the sched_entity representing
2662 * the group on a CPU.
2664 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2665 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2666 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2667 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2670 if (shares
< MIN_SHARES
)
2671 shares
= MIN_SHARES
;
2672 if (shares
> tg
->shares
)
2673 shares
= tg
->shares
;
2677 # else /* CONFIG_SMP */
2678 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2682 # endif /* CONFIG_SMP */
2684 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2685 unsigned long weight
)
2688 /* commit outstanding execution time */
2689 if (cfs_rq
->curr
== se
)
2690 update_curr(cfs_rq
);
2691 account_entity_dequeue(cfs_rq
, se
);
2694 update_load_set(&se
->load
, weight
);
2697 account_entity_enqueue(cfs_rq
, se
);
2700 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2702 static void update_cfs_shares(struct sched_entity
*se
)
2704 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2705 struct task_group
*tg
;
2711 if (throttled_hierarchy(cfs_rq
))
2717 if (likely(se
->load
.weight
== tg
->shares
))
2720 shares
= calc_cfs_shares(cfs_rq
, tg
);
2722 reweight_entity(cfs_rq_of(se
), se
, shares
);
2725 #else /* CONFIG_FAIR_GROUP_SCHED */
2726 static inline void update_cfs_shares(struct sched_entity
*se
)
2729 #endif /* CONFIG_FAIR_GROUP_SCHED */
2734 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2736 static u64
decay_load(u64 val
, u64 n
)
2738 unsigned int local_n
;
2740 if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2743 /* after bounds checking we can collapse to 32-bit */
2747 * As y^PERIOD = 1/2, we can combine
2748 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2749 * With a look-up table which covers y^n (n<PERIOD)
2751 * To achieve constant time decay_load.
2753 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2754 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2755 local_n
%= LOAD_AVG_PERIOD
;
2758 val
= mul_u64_u32_shr(val
, runnable_avg_yN_inv
[local_n
], 32);
2762 static u32
__accumulate_pelt_segments(u64 periods
, u32 d1
, u32 d3
)
2764 u32 c1
, c2
, c3
= d3
; /* y^0 == 1 */
2769 c1
= decay_load((u64
)d1
, periods
);
2773 * c2 = 1024 \Sum y^n
2777 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
2780 c2
= LOAD_AVG_MAX
- decay_load(LOAD_AVG_MAX
, periods
) - 1024;
2782 return c1
+ c2
+ c3
;
2785 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2788 * Accumulate the three separate parts of the sum; d1 the remainder
2789 * of the last (incomplete) period, d2 the span of full periods and d3
2790 * the remainder of the (incomplete) current period.
2795 * |<->|<----------------->|<--->|
2796 * ... |---x---|------| ... |------|-----x (now)
2799 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
2802 * = u y^p + (Step 1)
2805 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
2808 static __always_inline u32
2809 accumulate_sum(u64 delta
, int cpu
, struct sched_avg
*sa
,
2810 unsigned long weight
, int running
, struct cfs_rq
*cfs_rq
)
2812 unsigned long scale_freq
, scale_cpu
;
2813 u32 contrib
= (u32
)delta
; /* p == 0 -> delta < 1024 */
2816 scale_freq
= arch_scale_freq_capacity(NULL
, cpu
);
2817 scale_cpu
= arch_scale_cpu_capacity(NULL
, cpu
);
2819 delta
+= sa
->period_contrib
;
2820 periods
= delta
/ 1024; /* A period is 1024us (~1ms) */
2823 * Step 1: decay old *_sum if we crossed period boundaries.
2826 sa
->load_sum
= decay_load(sa
->load_sum
, periods
);
2828 cfs_rq
->runnable_load_sum
=
2829 decay_load(cfs_rq
->runnable_load_sum
, periods
);
2831 sa
->util_sum
= decay_load((u64
)(sa
->util_sum
), periods
);
2837 contrib
= __accumulate_pelt_segments(periods
,
2838 1024 - sa
->period_contrib
, delta
);
2840 sa
->period_contrib
= delta
;
2842 contrib
= cap_scale(contrib
, scale_freq
);
2844 sa
->load_sum
+= weight
* contrib
;
2846 cfs_rq
->runnable_load_sum
+= weight
* contrib
;
2849 sa
->util_sum
+= contrib
* scale_cpu
;
2855 * We can represent the historical contribution to runnable average as the
2856 * coefficients of a geometric series. To do this we sub-divide our runnable
2857 * history into segments of approximately 1ms (1024us); label the segment that
2858 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2860 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2862 * (now) (~1ms ago) (~2ms ago)
2864 * Let u_i denote the fraction of p_i that the entity was runnable.
2866 * We then designate the fractions u_i as our co-efficients, yielding the
2867 * following representation of historical load:
2868 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2870 * We choose y based on the with of a reasonably scheduling period, fixing:
2873 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2874 * approximately half as much as the contribution to load within the last ms
2877 * When a period "rolls over" and we have new u_0`, multiplying the previous
2878 * sum again by y is sufficient to update:
2879 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2880 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2882 static __always_inline
int
2883 ___update_load_avg(u64 now
, int cpu
, struct sched_avg
*sa
,
2884 unsigned long weight
, int running
, struct cfs_rq
*cfs_rq
)
2888 delta
= now
- sa
->last_update_time
;
2890 * This should only happen when time goes backwards, which it
2891 * unfortunately does during sched clock init when we swap over to TSC.
2893 if ((s64
)delta
< 0) {
2894 sa
->last_update_time
= now
;
2899 * Use 1024ns as the unit of measurement since it's a reasonable
2900 * approximation of 1us and fast to compute.
2906 sa
->last_update_time
+= delta
<< 10;
2909 * Now we know we crossed measurement unit boundaries. The *_avg
2910 * accrues by two steps:
2912 * Step 1: accumulate *_sum since last_update_time. If we haven't
2913 * crossed period boundaries, finish.
2915 if (!accumulate_sum(delta
, cpu
, sa
, weight
, running
, cfs_rq
))
2919 * Step 2: update *_avg.
2921 sa
->load_avg
= div_u64(sa
->load_sum
, LOAD_AVG_MAX
- 1024 + sa
->period_contrib
);
2923 cfs_rq
->runnable_load_avg
=
2924 div_u64(cfs_rq
->runnable_load_sum
, LOAD_AVG_MAX
- 1024 + sa
->period_contrib
);
2926 sa
->util_avg
= sa
->util_sum
/ (LOAD_AVG_MAX
- 1024 + sa
->period_contrib
);
2932 __update_load_avg_blocked_se(u64 now
, int cpu
, struct sched_entity
*se
)
2934 return ___update_load_avg(now
, cpu
, &se
->avg
, 0, 0, NULL
);
2938 __update_load_avg_se(u64 now
, int cpu
, struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2940 return ___update_load_avg(now
, cpu
, &se
->avg
,
2941 se
->on_rq
* scale_load_down(se
->load
.weight
),
2942 cfs_rq
->curr
== se
, NULL
);
2946 __update_load_avg_cfs_rq(u64 now
, int cpu
, struct cfs_rq
*cfs_rq
)
2948 return ___update_load_avg(now
, cpu
, &cfs_rq
->avg
,
2949 scale_load_down(cfs_rq
->load
.weight
),
2950 cfs_rq
->curr
!= NULL
, cfs_rq
);
2954 * Signed add and clamp on underflow.
2956 * Explicitly do a load-store to ensure the intermediate value never hits
2957 * memory. This allows lockless observations without ever seeing the negative
2960 #define add_positive(_ptr, _val) do { \
2961 typeof(_ptr) ptr = (_ptr); \
2962 typeof(_val) val = (_val); \
2963 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2967 if (val < 0 && res > var) \
2970 WRITE_ONCE(*ptr, res); \
2973 #ifdef CONFIG_FAIR_GROUP_SCHED
2975 * update_tg_load_avg - update the tg's load avg
2976 * @cfs_rq: the cfs_rq whose avg changed
2977 * @force: update regardless of how small the difference
2979 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
2980 * However, because tg->load_avg is a global value there are performance
2983 * In order to avoid having to look at the other cfs_rq's, we use a
2984 * differential update where we store the last value we propagated. This in
2985 * turn allows skipping updates if the differential is 'small'.
2987 * Updating tg's load_avg is necessary before update_cfs_share() (which is
2988 * done) and effective_load() (which is not done because it is too costly).
2990 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
)
2992 long delta
= cfs_rq
->avg
.load_avg
- cfs_rq
->tg_load_avg_contrib
;
2995 * No need to update load_avg for root_task_group as it is not used.
2997 if (cfs_rq
->tg
== &root_task_group
)
3000 if (force
|| abs(delta
) > cfs_rq
->tg_load_avg_contrib
/ 64) {
3001 atomic_long_add(delta
, &cfs_rq
->tg
->load_avg
);
3002 cfs_rq
->tg_load_avg_contrib
= cfs_rq
->avg
.load_avg
;
3007 * Called within set_task_rq() right before setting a task's cpu. The
3008 * caller only guarantees p->pi_lock is held; no other assumptions,
3009 * including the state of rq->lock, should be made.
3011 void set_task_rq_fair(struct sched_entity
*se
,
3012 struct cfs_rq
*prev
, struct cfs_rq
*next
)
3014 u64 p_last_update_time
;
3015 u64 n_last_update_time
;
3017 if (!sched_feat(ATTACH_AGE_LOAD
))
3021 * We are supposed to update the task to "current" time, then its up to
3022 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3023 * getting what current time is, so simply throw away the out-of-date
3024 * time. This will result in the wakee task is less decayed, but giving
3025 * the wakee more load sounds not bad.
3027 if (!(se
->avg
.last_update_time
&& prev
))
3030 #ifndef CONFIG_64BIT
3032 u64 p_last_update_time_copy
;
3033 u64 n_last_update_time_copy
;
3036 p_last_update_time_copy
= prev
->load_last_update_time_copy
;
3037 n_last_update_time_copy
= next
->load_last_update_time_copy
;
3041 p_last_update_time
= prev
->avg
.last_update_time
;
3042 n_last_update_time
= next
->avg
.last_update_time
;
3044 } while (p_last_update_time
!= p_last_update_time_copy
||
3045 n_last_update_time
!= n_last_update_time_copy
);
3048 p_last_update_time
= prev
->avg
.last_update_time
;
3049 n_last_update_time
= next
->avg
.last_update_time
;
3051 __update_load_avg_blocked_se(p_last_update_time
, cpu_of(rq_of(prev
)), se
);
3052 se
->avg
.last_update_time
= n_last_update_time
;
3055 /* Take into account change of utilization of a child task group */
3057 update_tg_cfs_util(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3059 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3060 long delta
= gcfs_rq
->avg
.util_avg
- se
->avg
.util_avg
;
3062 /* Nothing to update */
3066 /* Set new sched_entity's utilization */
3067 se
->avg
.util_avg
= gcfs_rq
->avg
.util_avg
;
3068 se
->avg
.util_sum
= se
->avg
.util_avg
* LOAD_AVG_MAX
;
3070 /* Update parent cfs_rq utilization */
3071 add_positive(&cfs_rq
->avg
.util_avg
, delta
);
3072 cfs_rq
->avg
.util_sum
= cfs_rq
->avg
.util_avg
* LOAD_AVG_MAX
;
3075 /* Take into account change of load of a child task group */
3077 update_tg_cfs_load(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3079 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3080 long delta
, load
= gcfs_rq
->avg
.load_avg
;
3083 * If the load of group cfs_rq is null, the load of the
3084 * sched_entity will also be null so we can skip the formula
3089 /* Get tg's load and ensure tg_load > 0 */
3090 tg_load
= atomic_long_read(&gcfs_rq
->tg
->load_avg
) + 1;
3092 /* Ensure tg_load >= load and updated with current load*/
3093 tg_load
-= gcfs_rq
->tg_load_avg_contrib
;
3097 * We need to compute a correction term in the case that the
3098 * task group is consuming more CPU than a task of equal
3099 * weight. A task with a weight equals to tg->shares will have
3100 * a load less or equal to scale_load_down(tg->shares).
3101 * Similarly, the sched_entities that represent the task group
3102 * at parent level, can't have a load higher than
3103 * scale_load_down(tg->shares). And the Sum of sched_entities'
3104 * load must be <= scale_load_down(tg->shares).
3106 if (tg_load
> scale_load_down(gcfs_rq
->tg
->shares
)) {
3107 /* scale gcfs_rq's load into tg's shares*/
3108 load
*= scale_load_down(gcfs_rq
->tg
->shares
);
3113 delta
= load
- se
->avg
.load_avg
;
3115 /* Nothing to update */
3119 /* Set new sched_entity's load */
3120 se
->avg
.load_avg
= load
;
3121 se
->avg
.load_sum
= se
->avg
.load_avg
* LOAD_AVG_MAX
;
3123 /* Update parent cfs_rq load */
3124 add_positive(&cfs_rq
->avg
.load_avg
, delta
);
3125 cfs_rq
->avg
.load_sum
= cfs_rq
->avg
.load_avg
* LOAD_AVG_MAX
;
3128 * If the sched_entity is already enqueued, we also have to update the
3129 * runnable load avg.
3132 /* Update parent cfs_rq runnable_load_avg */
3133 add_positive(&cfs_rq
->runnable_load_avg
, delta
);
3134 cfs_rq
->runnable_load_sum
= cfs_rq
->runnable_load_avg
* LOAD_AVG_MAX
;
3138 static inline void set_tg_cfs_propagate(struct cfs_rq
*cfs_rq
)
3140 cfs_rq
->propagate_avg
= 1;
3143 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity
*se
)
3145 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
3147 if (!cfs_rq
->propagate_avg
)
3150 cfs_rq
->propagate_avg
= 0;
3154 /* Update task and its cfs_rq load average */
3155 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3157 struct cfs_rq
*cfs_rq
;
3159 if (entity_is_task(se
))
3162 if (!test_and_clear_tg_cfs_propagate(se
))
3165 cfs_rq
= cfs_rq_of(se
);
3167 set_tg_cfs_propagate(cfs_rq
);
3169 update_tg_cfs_util(cfs_rq
, se
);
3170 update_tg_cfs_load(cfs_rq
, se
);
3176 * Check if we need to update the load and the utilization of a blocked
3179 static inline bool skip_blocked_update(struct sched_entity
*se
)
3181 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3184 * If sched_entity still have not zero load or utilization, we have to
3187 if (se
->avg
.load_avg
|| se
->avg
.util_avg
)
3191 * If there is a pending propagation, we have to update the load and
3192 * the utilization of the sched_entity:
3194 if (gcfs_rq
->propagate_avg
)
3198 * Otherwise, the load and the utilization of the sched_entity is
3199 * already zero and there is no pending propagation, so it will be a
3200 * waste of time to try to decay it:
3205 #else /* CONFIG_FAIR_GROUP_SCHED */
3207 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
) {}
3209 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3214 static inline void set_tg_cfs_propagate(struct cfs_rq
*cfs_rq
) {}
3216 #endif /* CONFIG_FAIR_GROUP_SCHED */
3218 static inline void cfs_rq_util_change(struct cfs_rq
*cfs_rq
)
3220 if (&this_rq()->cfs
== cfs_rq
) {
3222 * There are a few boundary cases this might miss but it should
3223 * get called often enough that that should (hopefully) not be
3224 * a real problem -- added to that it only calls on the local
3225 * CPU, so if we enqueue remotely we'll miss an update, but
3226 * the next tick/schedule should update.
3228 * It will not get called when we go idle, because the idle
3229 * thread is a different class (!fair), nor will the utilization
3230 * number include things like RT tasks.
3232 * As is, the util number is not freq-invariant (we'd have to
3233 * implement arch_scale_freq_capacity() for that).
3237 cpufreq_update_util(rq_of(cfs_rq
), 0);
3242 * Unsigned subtract and clamp on underflow.
3244 * Explicitly do a load-store to ensure the intermediate value never hits
3245 * memory. This allows lockless observations without ever seeing the negative
3248 #define sub_positive(_ptr, _val) do { \
3249 typeof(_ptr) ptr = (_ptr); \
3250 typeof(*ptr) val = (_val); \
3251 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3255 WRITE_ONCE(*ptr, res); \
3259 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3260 * @now: current time, as per cfs_rq_clock_task()
3261 * @cfs_rq: cfs_rq to update
3262 * @update_freq: should we call cfs_rq_util_change() or will the call do so
3264 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3265 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3266 * post_init_entity_util_avg().
3268 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3270 * Returns true if the load decayed or we removed load.
3272 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3273 * call update_tg_load_avg() when this function returns true.
3276 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
, bool update_freq
)
3278 struct sched_avg
*sa
= &cfs_rq
->avg
;
3279 int decayed
, removed_load
= 0, removed_util
= 0;
3281 if (atomic_long_read(&cfs_rq
->removed_load_avg
)) {
3282 s64 r
= atomic_long_xchg(&cfs_rq
->removed_load_avg
, 0);
3283 sub_positive(&sa
->load_avg
, r
);
3284 sub_positive(&sa
->load_sum
, r
* LOAD_AVG_MAX
);
3286 set_tg_cfs_propagate(cfs_rq
);
3289 if (atomic_long_read(&cfs_rq
->removed_util_avg
)) {
3290 long r
= atomic_long_xchg(&cfs_rq
->removed_util_avg
, 0);
3291 sub_positive(&sa
->util_avg
, r
);
3292 sub_positive(&sa
->util_sum
, r
* LOAD_AVG_MAX
);
3294 set_tg_cfs_propagate(cfs_rq
);
3297 decayed
= __update_load_avg_cfs_rq(now
, cpu_of(rq_of(cfs_rq
)), cfs_rq
);
3299 #ifndef CONFIG_64BIT
3301 cfs_rq
->load_last_update_time_copy
= sa
->last_update_time
;
3304 if (update_freq
&& (decayed
|| removed_util
))
3305 cfs_rq_util_change(cfs_rq
);
3307 return decayed
|| removed_load
;
3311 * Optional action to be done while updating the load average
3313 #define UPDATE_TG 0x1
3314 #define SKIP_AGE_LOAD 0x2
3316 /* Update task and its cfs_rq load average */
3317 static inline void update_load_avg(struct sched_entity
*se
, int flags
)
3319 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3320 u64 now
= cfs_rq_clock_task(cfs_rq
);
3321 struct rq
*rq
= rq_of(cfs_rq
);
3322 int cpu
= cpu_of(rq
);
3326 * Track task load average for carrying it to new CPU after migrated, and
3327 * track group sched_entity load average for task_h_load calc in migration
3329 if (se
->avg
.last_update_time
&& !(flags
& SKIP_AGE_LOAD
))
3330 __update_load_avg_se(now
, cpu
, cfs_rq
, se
);
3332 decayed
= update_cfs_rq_load_avg(now
, cfs_rq
, true);
3333 decayed
|= propagate_entity_load_avg(se
);
3335 if (decayed
&& (flags
& UPDATE_TG
))
3336 update_tg_load_avg(cfs_rq
, 0);
3340 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3341 * @cfs_rq: cfs_rq to attach to
3342 * @se: sched_entity to attach
3344 * Must call update_cfs_rq_load_avg() before this, since we rely on
3345 * cfs_rq->avg.last_update_time being current.
3347 static void attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3349 se
->avg
.last_update_time
= cfs_rq
->avg
.last_update_time
;
3350 cfs_rq
->avg
.load_avg
+= se
->avg
.load_avg
;
3351 cfs_rq
->avg
.load_sum
+= se
->avg
.load_sum
;
3352 cfs_rq
->avg
.util_avg
+= se
->avg
.util_avg
;
3353 cfs_rq
->avg
.util_sum
+= se
->avg
.util_sum
;
3354 set_tg_cfs_propagate(cfs_rq
);
3356 cfs_rq_util_change(cfs_rq
);
3360 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3361 * @cfs_rq: cfs_rq to detach from
3362 * @se: sched_entity to detach
3364 * Must call update_cfs_rq_load_avg() before this, since we rely on
3365 * cfs_rq->avg.last_update_time being current.
3367 static void detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3370 sub_positive(&cfs_rq
->avg
.load_avg
, se
->avg
.load_avg
);
3371 sub_positive(&cfs_rq
->avg
.load_sum
, se
->avg
.load_sum
);
3372 sub_positive(&cfs_rq
->avg
.util_avg
, se
->avg
.util_avg
);
3373 sub_positive(&cfs_rq
->avg
.util_sum
, se
->avg
.util_sum
);
3374 set_tg_cfs_propagate(cfs_rq
);
3376 cfs_rq_util_change(cfs_rq
);
3379 /* Add the load generated by se into cfs_rq's load average */
3381 enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3383 struct sched_avg
*sa
= &se
->avg
;
3385 cfs_rq
->runnable_load_avg
+= sa
->load_avg
;
3386 cfs_rq
->runnable_load_sum
+= sa
->load_sum
;
3388 if (!sa
->last_update_time
) {
3389 attach_entity_load_avg(cfs_rq
, se
);
3390 update_tg_load_avg(cfs_rq
, 0);
3394 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3396 dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3398 cfs_rq
->runnable_load_avg
=
3399 max_t(long, cfs_rq
->runnable_load_avg
- se
->avg
.load_avg
, 0);
3400 cfs_rq
->runnable_load_sum
=
3401 max_t(s64
, cfs_rq
->runnable_load_sum
- se
->avg
.load_sum
, 0);
3404 #ifndef CONFIG_64BIT
3405 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3407 u64 last_update_time_copy
;
3408 u64 last_update_time
;
3411 last_update_time_copy
= cfs_rq
->load_last_update_time_copy
;
3413 last_update_time
= cfs_rq
->avg
.last_update_time
;
3414 } while (last_update_time
!= last_update_time_copy
);
3416 return last_update_time
;
3419 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3421 return cfs_rq
->avg
.last_update_time
;
3426 * Synchronize entity load avg of dequeued entity without locking
3429 void sync_entity_load_avg(struct sched_entity
*se
)
3431 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3432 u64 last_update_time
;
3434 last_update_time
= cfs_rq_last_update_time(cfs_rq
);
3435 __update_load_avg_blocked_se(last_update_time
, cpu_of(rq_of(cfs_rq
)), se
);
3439 * Task first catches up with cfs_rq, and then subtract
3440 * itself from the cfs_rq (task must be off the queue now).
3442 void remove_entity_load_avg(struct sched_entity
*se
)
3444 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3447 * tasks cannot exit without having gone through wake_up_new_task() ->
3448 * post_init_entity_util_avg() which will have added things to the
3449 * cfs_rq, so we can remove unconditionally.
3451 * Similarly for groups, they will have passed through
3452 * post_init_entity_util_avg() before unregister_sched_fair_group()
3456 sync_entity_load_avg(se
);
3457 atomic_long_add(se
->avg
.load_avg
, &cfs_rq
->removed_load_avg
);
3458 atomic_long_add(se
->avg
.util_avg
, &cfs_rq
->removed_util_avg
);
3461 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq
*cfs_rq
)
3463 return cfs_rq
->runnable_load_avg
;
3466 static inline unsigned long cfs_rq_load_avg(struct cfs_rq
*cfs_rq
)
3468 return cfs_rq
->avg
.load_avg
;
3471 static int idle_balance(struct rq
*this_rq
, struct rq_flags
*rf
);
3473 #else /* CONFIG_SMP */
3476 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
, bool update_freq
)
3481 #define UPDATE_TG 0x0
3482 #define SKIP_AGE_LOAD 0x0
3484 static inline void update_load_avg(struct sched_entity
*se
, int not_used1
)
3486 cpufreq_update_util(rq_of(cfs_rq_of(se
)), 0);
3490 enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3492 dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3493 static inline void remove_entity_load_avg(struct sched_entity
*se
) {}
3496 attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3498 detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3500 static inline int idle_balance(struct rq
*rq
, struct rq_flags
*rf
)
3505 #endif /* CONFIG_SMP */
3507 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3509 #ifdef CONFIG_SCHED_DEBUG
3510 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
3515 if (d
> 3*sysctl_sched_latency
)
3516 schedstat_inc(cfs_rq
->nr_spread_over
);
3521 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
3523 u64 vruntime
= cfs_rq
->min_vruntime
;
3526 * The 'current' period is already promised to the current tasks,
3527 * however the extra weight of the new task will slow them down a
3528 * little, place the new task so that it fits in the slot that
3529 * stays open at the end.
3531 if (initial
&& sched_feat(START_DEBIT
))
3532 vruntime
+= sched_vslice(cfs_rq
, se
);
3534 /* sleeps up to a single latency don't count. */
3536 unsigned long thresh
= sysctl_sched_latency
;
3539 * Halve their sleep time's effect, to allow
3540 * for a gentler effect of sleepers:
3542 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
3548 /* ensure we never gain time by being placed backwards. */
3549 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
3552 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
3554 static inline void check_schedstat_required(void)
3556 #ifdef CONFIG_SCHEDSTATS
3557 if (schedstat_enabled())
3560 /* Force schedstat enabled if a dependent tracepoint is active */
3561 if (trace_sched_stat_wait_enabled() ||
3562 trace_sched_stat_sleep_enabled() ||
3563 trace_sched_stat_iowait_enabled() ||
3564 trace_sched_stat_blocked_enabled() ||
3565 trace_sched_stat_runtime_enabled()) {
3566 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3567 "stat_blocked and stat_runtime require the "
3568 "kernel parameter schedstats=enabled or "
3569 "kernel.sched_schedstats=1\n");
3580 * update_min_vruntime()
3581 * vruntime -= min_vruntime
3585 * update_min_vruntime()
3586 * vruntime += min_vruntime
3588 * this way the vruntime transition between RQs is done when both
3589 * min_vruntime are up-to-date.
3593 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3594 * vruntime -= min_vruntime
3598 * update_min_vruntime()
3599 * vruntime += min_vruntime
3601 * this way we don't have the most up-to-date min_vruntime on the originating
3602 * CPU and an up-to-date min_vruntime on the destination CPU.
3606 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3608 bool renorm
= !(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_MIGRATED
);
3609 bool curr
= cfs_rq
->curr
== se
;
3612 * If we're the current task, we must renormalise before calling
3616 se
->vruntime
+= cfs_rq
->min_vruntime
;
3618 update_curr(cfs_rq
);
3621 * Otherwise, renormalise after, such that we're placed at the current
3622 * moment in time, instead of some random moment in the past. Being
3623 * placed in the past could significantly boost this task to the
3624 * fairness detriment of existing tasks.
3626 if (renorm
&& !curr
)
3627 se
->vruntime
+= cfs_rq
->min_vruntime
;
3630 * When enqueuing a sched_entity, we must:
3631 * - Update loads to have both entity and cfs_rq synced with now.
3632 * - Add its load to cfs_rq->runnable_avg
3633 * - For group_entity, update its weight to reflect the new share of
3635 * - Add its new weight to cfs_rq->load.weight
3637 update_load_avg(se
, UPDATE_TG
);
3638 enqueue_entity_load_avg(cfs_rq
, se
);
3639 update_cfs_shares(se
);
3640 account_entity_enqueue(cfs_rq
, se
);
3642 if (flags
& ENQUEUE_WAKEUP
)
3643 place_entity(cfs_rq
, se
, 0);
3645 check_schedstat_required();
3646 update_stats_enqueue(cfs_rq
, se
, flags
);
3647 check_spread(cfs_rq
, se
);
3649 __enqueue_entity(cfs_rq
, se
);
3652 if (cfs_rq
->nr_running
== 1) {
3653 list_add_leaf_cfs_rq(cfs_rq
);
3654 check_enqueue_throttle(cfs_rq
);
3658 static void __clear_buddies_last(struct sched_entity
*se
)
3660 for_each_sched_entity(se
) {
3661 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3662 if (cfs_rq
->last
!= se
)
3665 cfs_rq
->last
= NULL
;
3669 static void __clear_buddies_next(struct sched_entity
*se
)
3671 for_each_sched_entity(se
) {
3672 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3673 if (cfs_rq
->next
!= se
)
3676 cfs_rq
->next
= NULL
;
3680 static void __clear_buddies_skip(struct sched_entity
*se
)
3682 for_each_sched_entity(se
) {
3683 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3684 if (cfs_rq
->skip
!= se
)
3687 cfs_rq
->skip
= NULL
;
3691 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3693 if (cfs_rq
->last
== se
)
3694 __clear_buddies_last(se
);
3696 if (cfs_rq
->next
== se
)
3697 __clear_buddies_next(se
);
3699 if (cfs_rq
->skip
== se
)
3700 __clear_buddies_skip(se
);
3703 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3706 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3709 * Update run-time statistics of the 'current'.
3711 update_curr(cfs_rq
);
3714 * When dequeuing a sched_entity, we must:
3715 * - Update loads to have both entity and cfs_rq synced with now.
3716 * - Substract its load from the cfs_rq->runnable_avg.
3717 * - Substract its previous weight from cfs_rq->load.weight.
3718 * - For group entity, update its weight to reflect the new share
3719 * of its group cfs_rq.
3721 update_load_avg(se
, UPDATE_TG
);
3722 dequeue_entity_load_avg(cfs_rq
, se
);
3724 update_stats_dequeue(cfs_rq
, se
, flags
);
3726 clear_buddies(cfs_rq
, se
);
3728 if (se
!= cfs_rq
->curr
)
3729 __dequeue_entity(cfs_rq
, se
);
3731 account_entity_dequeue(cfs_rq
, se
);
3734 * Normalize after update_curr(); which will also have moved
3735 * min_vruntime if @se is the one holding it back. But before doing
3736 * update_min_vruntime() again, which will discount @se's position and
3737 * can move min_vruntime forward still more.
3739 if (!(flags
& DEQUEUE_SLEEP
))
3740 se
->vruntime
-= cfs_rq
->min_vruntime
;
3742 /* return excess runtime on last dequeue */
3743 return_cfs_rq_runtime(cfs_rq
);
3745 update_cfs_shares(se
);
3748 * Now advance min_vruntime if @se was the entity holding it back,
3749 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3750 * put back on, and if we advance min_vruntime, we'll be placed back
3751 * further than we started -- ie. we'll be penalized.
3753 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) == DEQUEUE_SAVE
)
3754 update_min_vruntime(cfs_rq
);
3758 * Preempt the current task with a newly woken task if needed:
3761 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3763 unsigned long ideal_runtime
, delta_exec
;
3764 struct sched_entity
*se
;
3767 ideal_runtime
= sched_slice(cfs_rq
, curr
);
3768 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
3769 if (delta_exec
> ideal_runtime
) {
3770 resched_curr(rq_of(cfs_rq
));
3772 * The current task ran long enough, ensure it doesn't get
3773 * re-elected due to buddy favours.
3775 clear_buddies(cfs_rq
, curr
);
3780 * Ensure that a task that missed wakeup preemption by a
3781 * narrow margin doesn't have to wait for a full slice.
3782 * This also mitigates buddy induced latencies under load.
3784 if (delta_exec
< sysctl_sched_min_granularity
)
3787 se
= __pick_first_entity(cfs_rq
);
3788 delta
= curr
->vruntime
- se
->vruntime
;
3793 if (delta
> ideal_runtime
)
3794 resched_curr(rq_of(cfs_rq
));
3798 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3800 /* 'current' is not kept within the tree. */
3803 * Any task has to be enqueued before it get to execute on
3804 * a CPU. So account for the time it spent waiting on the
3807 update_stats_wait_end(cfs_rq
, se
);
3808 __dequeue_entity(cfs_rq
, se
);
3809 update_load_avg(se
, UPDATE_TG
);
3812 update_stats_curr_start(cfs_rq
, se
);
3816 * Track our maximum slice length, if the CPU's load is at
3817 * least twice that of our own weight (i.e. dont track it
3818 * when there are only lesser-weight tasks around):
3820 if (schedstat_enabled() && rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
3821 schedstat_set(se
->statistics
.slice_max
,
3822 max((u64
)schedstat_val(se
->statistics
.slice_max
),
3823 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
));
3826 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
3830 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
3833 * Pick the next process, keeping these things in mind, in this order:
3834 * 1) keep things fair between processes/task groups
3835 * 2) pick the "next" process, since someone really wants that to run
3836 * 3) pick the "last" process, for cache locality
3837 * 4) do not run the "skip" process, if something else is available
3839 static struct sched_entity
*
3840 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3842 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
3843 struct sched_entity
*se
;
3846 * If curr is set we have to see if its left of the leftmost entity
3847 * still in the tree, provided there was anything in the tree at all.
3849 if (!left
|| (curr
&& entity_before(curr
, left
)))
3852 se
= left
; /* ideally we run the leftmost entity */
3855 * Avoid running the skip buddy, if running something else can
3856 * be done without getting too unfair.
3858 if (cfs_rq
->skip
== se
) {
3859 struct sched_entity
*second
;
3862 second
= __pick_first_entity(cfs_rq
);
3864 second
= __pick_next_entity(se
);
3865 if (!second
|| (curr
&& entity_before(curr
, second
)))
3869 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3874 * Prefer last buddy, try to return the CPU to a preempted task.
3876 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3880 * Someone really wants this to run. If it's not unfair, run it.
3882 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3885 clear_buddies(cfs_rq
, se
);
3890 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3892 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3895 * If still on the runqueue then deactivate_task()
3896 * was not called and update_curr() has to be done:
3899 update_curr(cfs_rq
);
3901 /* throttle cfs_rqs exceeding runtime */
3902 check_cfs_rq_runtime(cfs_rq
);
3904 check_spread(cfs_rq
, prev
);
3907 update_stats_wait_start(cfs_rq
, prev
);
3908 /* Put 'current' back into the tree. */
3909 __enqueue_entity(cfs_rq
, prev
);
3910 /* in !on_rq case, update occurred at dequeue */
3911 update_load_avg(prev
, 0);
3913 cfs_rq
->curr
= NULL
;
3917 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3920 * Update run-time statistics of the 'current'.
3922 update_curr(cfs_rq
);
3925 * Ensure that runnable average is periodically updated.
3927 update_load_avg(curr
, UPDATE_TG
);
3928 update_cfs_shares(curr
);
3930 #ifdef CONFIG_SCHED_HRTICK
3932 * queued ticks are scheduled to match the slice, so don't bother
3933 * validating it and just reschedule.
3936 resched_curr(rq_of(cfs_rq
));
3940 * don't let the period tick interfere with the hrtick preemption
3942 if (!sched_feat(DOUBLE_TICK
) &&
3943 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3947 if (cfs_rq
->nr_running
> 1)
3948 check_preempt_tick(cfs_rq
, curr
);
3952 /**************************************************
3953 * CFS bandwidth control machinery
3956 #ifdef CONFIG_CFS_BANDWIDTH
3958 #ifdef HAVE_JUMP_LABEL
3959 static struct static_key __cfs_bandwidth_used
;
3961 static inline bool cfs_bandwidth_used(void)
3963 return static_key_false(&__cfs_bandwidth_used
);
3966 void cfs_bandwidth_usage_inc(void)
3968 static_key_slow_inc(&__cfs_bandwidth_used
);
3971 void cfs_bandwidth_usage_dec(void)
3973 static_key_slow_dec(&__cfs_bandwidth_used
);
3975 #else /* HAVE_JUMP_LABEL */
3976 static bool cfs_bandwidth_used(void)
3981 void cfs_bandwidth_usage_inc(void) {}
3982 void cfs_bandwidth_usage_dec(void) {}
3983 #endif /* HAVE_JUMP_LABEL */
3986 * default period for cfs group bandwidth.
3987 * default: 0.1s, units: nanoseconds
3989 static inline u64
default_cfs_period(void)
3991 return 100000000ULL;
3994 static inline u64
sched_cfs_bandwidth_slice(void)
3996 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
4000 * Replenish runtime according to assigned quota and update expiration time.
4001 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4002 * additional synchronization around rq->lock.
4004 * requires cfs_b->lock
4006 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
4010 if (cfs_b
->quota
== RUNTIME_INF
)
4013 now
= sched_clock_cpu(smp_processor_id());
4014 cfs_b
->runtime
= cfs_b
->quota
;
4015 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
4018 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4020 return &tg
->cfs_bandwidth
;
4023 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4024 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
4026 if (unlikely(cfs_rq
->throttle_count
))
4027 return cfs_rq
->throttled_clock_task
- cfs_rq
->throttled_clock_task_time
;
4029 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
4032 /* returns 0 on failure to allocate runtime */
4033 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4035 struct task_group
*tg
= cfs_rq
->tg
;
4036 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
4037 u64 amount
= 0, min_amount
, expires
;
4039 /* note: this is a positive sum as runtime_remaining <= 0 */
4040 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
4042 raw_spin_lock(&cfs_b
->lock
);
4043 if (cfs_b
->quota
== RUNTIME_INF
)
4044 amount
= min_amount
;
4046 start_cfs_bandwidth(cfs_b
);
4048 if (cfs_b
->runtime
> 0) {
4049 amount
= min(cfs_b
->runtime
, min_amount
);
4050 cfs_b
->runtime
-= amount
;
4054 expires
= cfs_b
->runtime_expires
;
4055 raw_spin_unlock(&cfs_b
->lock
);
4057 cfs_rq
->runtime_remaining
+= amount
;
4059 * we may have advanced our local expiration to account for allowed
4060 * spread between our sched_clock and the one on which runtime was
4063 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
4064 cfs_rq
->runtime_expires
= expires
;
4066 return cfs_rq
->runtime_remaining
> 0;
4070 * Note: This depends on the synchronization provided by sched_clock and the
4071 * fact that rq->clock snapshots this value.
4073 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4075 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4077 /* if the deadline is ahead of our clock, nothing to do */
4078 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
4081 if (cfs_rq
->runtime_remaining
< 0)
4085 * If the local deadline has passed we have to consider the
4086 * possibility that our sched_clock is 'fast' and the global deadline
4087 * has not truly expired.
4089 * Fortunately we can check determine whether this the case by checking
4090 * whether the global deadline has advanced. It is valid to compare
4091 * cfs_b->runtime_expires without any locks since we only care about
4092 * exact equality, so a partial write will still work.
4095 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
4096 /* extend local deadline, drift is bounded above by 2 ticks */
4097 cfs_rq
->runtime_expires
+= TICK_NSEC
;
4099 /* global deadline is ahead, expiration has passed */
4100 cfs_rq
->runtime_remaining
= 0;
4104 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4106 /* dock delta_exec before expiring quota (as it could span periods) */
4107 cfs_rq
->runtime_remaining
-= delta_exec
;
4108 expire_cfs_rq_runtime(cfs_rq
);
4110 if (likely(cfs_rq
->runtime_remaining
> 0))
4114 * if we're unable to extend our runtime we resched so that the active
4115 * hierarchy can be throttled
4117 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
4118 resched_curr(rq_of(cfs_rq
));
4121 static __always_inline
4122 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4124 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
4127 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
4130 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4132 return cfs_bandwidth_used() && cfs_rq
->throttled
;
4135 /* check whether cfs_rq, or any parent, is throttled */
4136 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4138 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
4142 * Ensure that neither of the group entities corresponding to src_cpu or
4143 * dest_cpu are members of a throttled hierarchy when performing group
4144 * load-balance operations.
4146 static inline int throttled_lb_pair(struct task_group
*tg
,
4147 int src_cpu
, int dest_cpu
)
4149 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
4151 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
4152 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
4154 return throttled_hierarchy(src_cfs_rq
) ||
4155 throttled_hierarchy(dest_cfs_rq
);
4158 /* updated child weight may affect parent so we have to do this bottom up */
4159 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
4161 struct rq
*rq
= data
;
4162 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4164 cfs_rq
->throttle_count
--;
4165 if (!cfs_rq
->throttle_count
) {
4166 /* adjust cfs_rq_clock_task() */
4167 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
4168 cfs_rq
->throttled_clock_task
;
4174 static int tg_throttle_down(struct task_group
*tg
, void *data
)
4176 struct rq
*rq
= data
;
4177 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4179 /* group is entering throttled state, stop time */
4180 if (!cfs_rq
->throttle_count
)
4181 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
4182 cfs_rq
->throttle_count
++;
4187 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
4189 struct rq
*rq
= rq_of(cfs_rq
);
4190 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4191 struct sched_entity
*se
;
4192 long task_delta
, dequeue
= 1;
4195 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
4197 /* freeze hierarchy runnable averages while throttled */
4199 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
4202 task_delta
= cfs_rq
->h_nr_running
;
4203 for_each_sched_entity(se
) {
4204 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4205 /* throttled entity or throttle-on-deactivate */
4210 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
4211 qcfs_rq
->h_nr_running
-= task_delta
;
4213 if (qcfs_rq
->load
.weight
)
4218 sub_nr_running(rq
, task_delta
);
4220 cfs_rq
->throttled
= 1;
4221 cfs_rq
->throttled_clock
= rq_clock(rq
);
4222 raw_spin_lock(&cfs_b
->lock
);
4223 empty
= list_empty(&cfs_b
->throttled_cfs_rq
);
4226 * Add to the _head_ of the list, so that an already-started
4227 * distribute_cfs_runtime will not see us
4229 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
4232 * If we're the first throttled task, make sure the bandwidth
4236 start_cfs_bandwidth(cfs_b
);
4238 raw_spin_unlock(&cfs_b
->lock
);
4241 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
4243 struct rq
*rq
= rq_of(cfs_rq
);
4244 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4245 struct sched_entity
*se
;
4249 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
4251 cfs_rq
->throttled
= 0;
4253 update_rq_clock(rq
);
4255 raw_spin_lock(&cfs_b
->lock
);
4256 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
4257 list_del_rcu(&cfs_rq
->throttled_list
);
4258 raw_spin_unlock(&cfs_b
->lock
);
4260 /* update hierarchical throttle state */
4261 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
4263 if (!cfs_rq
->load
.weight
)
4266 task_delta
= cfs_rq
->h_nr_running
;
4267 for_each_sched_entity(se
) {
4271 cfs_rq
= cfs_rq_of(se
);
4273 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
4274 cfs_rq
->h_nr_running
+= task_delta
;
4276 if (cfs_rq_throttled(cfs_rq
))
4281 add_nr_running(rq
, task_delta
);
4283 /* determine whether we need to wake up potentially idle cpu */
4284 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
4288 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
4289 u64 remaining
, u64 expires
)
4291 struct cfs_rq
*cfs_rq
;
4293 u64 starting_runtime
= remaining
;
4296 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
4298 struct rq
*rq
= rq_of(cfs_rq
);
4302 if (!cfs_rq_throttled(cfs_rq
))
4305 runtime
= -cfs_rq
->runtime_remaining
+ 1;
4306 if (runtime
> remaining
)
4307 runtime
= remaining
;
4308 remaining
-= runtime
;
4310 cfs_rq
->runtime_remaining
+= runtime
;
4311 cfs_rq
->runtime_expires
= expires
;
4313 /* we check whether we're throttled above */
4314 if (cfs_rq
->runtime_remaining
> 0)
4315 unthrottle_cfs_rq(cfs_rq
);
4325 return starting_runtime
- remaining
;
4329 * Responsible for refilling a task_group's bandwidth and unthrottling its
4330 * cfs_rqs as appropriate. If there has been no activity within the last
4331 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4332 * used to track this state.
4334 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
4336 u64 runtime
, runtime_expires
;
4339 /* no need to continue the timer with no bandwidth constraint */
4340 if (cfs_b
->quota
== RUNTIME_INF
)
4341 goto out_deactivate
;
4343 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4344 cfs_b
->nr_periods
+= overrun
;
4347 * idle depends on !throttled (for the case of a large deficit), and if
4348 * we're going inactive then everything else can be deferred
4350 if (cfs_b
->idle
&& !throttled
)
4351 goto out_deactivate
;
4353 __refill_cfs_bandwidth_runtime(cfs_b
);
4356 /* mark as potentially idle for the upcoming period */
4361 /* account preceding periods in which throttling occurred */
4362 cfs_b
->nr_throttled
+= overrun
;
4364 runtime_expires
= cfs_b
->runtime_expires
;
4367 * This check is repeated as we are holding onto the new bandwidth while
4368 * we unthrottle. This can potentially race with an unthrottled group
4369 * trying to acquire new bandwidth from the global pool. This can result
4370 * in us over-using our runtime if it is all used during this loop, but
4371 * only by limited amounts in that extreme case.
4373 while (throttled
&& cfs_b
->runtime
> 0) {
4374 runtime
= cfs_b
->runtime
;
4375 raw_spin_unlock(&cfs_b
->lock
);
4376 /* we can't nest cfs_b->lock while distributing bandwidth */
4377 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
4379 raw_spin_lock(&cfs_b
->lock
);
4381 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4383 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
4387 * While we are ensured activity in the period following an
4388 * unthrottle, this also covers the case in which the new bandwidth is
4389 * insufficient to cover the existing bandwidth deficit. (Forcing the
4390 * timer to remain active while there are any throttled entities.)
4400 /* a cfs_rq won't donate quota below this amount */
4401 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
4402 /* minimum remaining period time to redistribute slack quota */
4403 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
4404 /* how long we wait to gather additional slack before distributing */
4405 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
4408 * Are we near the end of the current quota period?
4410 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4411 * hrtimer base being cleared by hrtimer_start. In the case of
4412 * migrate_hrtimers, base is never cleared, so we are fine.
4414 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
4416 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
4419 /* if the call-back is running a quota refresh is already occurring */
4420 if (hrtimer_callback_running(refresh_timer
))
4423 /* is a quota refresh about to occur? */
4424 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
4425 if (remaining
< min_expire
)
4431 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
4433 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
4435 /* if there's a quota refresh soon don't bother with slack */
4436 if (runtime_refresh_within(cfs_b
, min_left
))
4439 hrtimer_start(&cfs_b
->slack_timer
,
4440 ns_to_ktime(cfs_bandwidth_slack_period
),
4444 /* we know any runtime found here is valid as update_curr() precedes return */
4445 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4447 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4448 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
4450 if (slack_runtime
<= 0)
4453 raw_spin_lock(&cfs_b
->lock
);
4454 if (cfs_b
->quota
!= RUNTIME_INF
&&
4455 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
4456 cfs_b
->runtime
+= slack_runtime
;
4458 /* we are under rq->lock, defer unthrottling using a timer */
4459 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
4460 !list_empty(&cfs_b
->throttled_cfs_rq
))
4461 start_cfs_slack_bandwidth(cfs_b
);
4463 raw_spin_unlock(&cfs_b
->lock
);
4465 /* even if it's not valid for return we don't want to try again */
4466 cfs_rq
->runtime_remaining
-= slack_runtime
;
4469 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4471 if (!cfs_bandwidth_used())
4474 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
4477 __return_cfs_rq_runtime(cfs_rq
);
4481 * This is done with a timer (instead of inline with bandwidth return) since
4482 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4484 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
4486 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
4489 /* confirm we're still not at a refresh boundary */
4490 raw_spin_lock(&cfs_b
->lock
);
4491 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
4492 raw_spin_unlock(&cfs_b
->lock
);
4496 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
4497 runtime
= cfs_b
->runtime
;
4499 expires
= cfs_b
->runtime_expires
;
4500 raw_spin_unlock(&cfs_b
->lock
);
4505 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
4507 raw_spin_lock(&cfs_b
->lock
);
4508 if (expires
== cfs_b
->runtime_expires
)
4509 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
4510 raw_spin_unlock(&cfs_b
->lock
);
4514 * When a group wakes up we want to make sure that its quota is not already
4515 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4516 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4518 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
4520 if (!cfs_bandwidth_used())
4523 /* an active group must be handled by the update_curr()->put() path */
4524 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
4527 /* ensure the group is not already throttled */
4528 if (cfs_rq_throttled(cfs_rq
))
4531 /* update runtime allocation */
4532 account_cfs_rq_runtime(cfs_rq
, 0);
4533 if (cfs_rq
->runtime_remaining
<= 0)
4534 throttle_cfs_rq(cfs_rq
);
4537 static void sync_throttle(struct task_group
*tg
, int cpu
)
4539 struct cfs_rq
*pcfs_rq
, *cfs_rq
;
4541 if (!cfs_bandwidth_used())
4547 cfs_rq
= tg
->cfs_rq
[cpu
];
4548 pcfs_rq
= tg
->parent
->cfs_rq
[cpu
];
4550 cfs_rq
->throttle_count
= pcfs_rq
->throttle_count
;
4551 cfs_rq
->throttled_clock_task
= rq_clock_task(cpu_rq(cpu
));
4554 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4555 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4557 if (!cfs_bandwidth_used())
4560 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
4564 * it's possible for a throttled entity to be forced into a running
4565 * state (e.g. set_curr_task), in this case we're finished.
4567 if (cfs_rq_throttled(cfs_rq
))
4570 throttle_cfs_rq(cfs_rq
);
4574 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
4576 struct cfs_bandwidth
*cfs_b
=
4577 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
4579 do_sched_cfs_slack_timer(cfs_b
);
4581 return HRTIMER_NORESTART
;
4584 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
4586 struct cfs_bandwidth
*cfs_b
=
4587 container_of(timer
, struct cfs_bandwidth
, period_timer
);
4591 raw_spin_lock(&cfs_b
->lock
);
4593 overrun
= hrtimer_forward_now(timer
, cfs_b
->period
);
4597 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
4600 cfs_b
->period_active
= 0;
4601 raw_spin_unlock(&cfs_b
->lock
);
4603 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
4606 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4608 raw_spin_lock_init(&cfs_b
->lock
);
4610 cfs_b
->quota
= RUNTIME_INF
;
4611 cfs_b
->period
= ns_to_ktime(default_cfs_period());
4613 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
4614 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_PINNED
);
4615 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
4616 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4617 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
4620 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4622 cfs_rq
->runtime_enabled
= 0;
4623 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
4626 void start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4628 lockdep_assert_held(&cfs_b
->lock
);
4630 if (!cfs_b
->period_active
) {
4631 cfs_b
->period_active
= 1;
4632 hrtimer_forward_now(&cfs_b
->period_timer
, cfs_b
->period
);
4633 hrtimer_start_expires(&cfs_b
->period_timer
, HRTIMER_MODE_ABS_PINNED
);
4637 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4639 /* init_cfs_bandwidth() was not called */
4640 if (!cfs_b
->throttled_cfs_rq
.next
)
4643 hrtimer_cancel(&cfs_b
->period_timer
);
4644 hrtimer_cancel(&cfs_b
->slack_timer
);
4648 * Both these cpu hotplug callbacks race against unregister_fair_sched_group()
4650 * The race is harmless, since modifying bandwidth settings of unhooked group
4651 * bits doesn't do much.
4654 /* cpu online calback */
4655 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
4657 struct task_group
*tg
;
4659 lockdep_assert_held(&rq
->lock
);
4662 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
4663 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
4664 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4666 raw_spin_lock(&cfs_b
->lock
);
4667 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
4668 raw_spin_unlock(&cfs_b
->lock
);
4673 /* cpu offline callback */
4674 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
4676 struct task_group
*tg
;
4678 lockdep_assert_held(&rq
->lock
);
4681 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
4682 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4684 if (!cfs_rq
->runtime_enabled
)
4688 * clock_task is not advancing so we just need to make sure
4689 * there's some valid quota amount
4691 cfs_rq
->runtime_remaining
= 1;
4693 * Offline rq is schedulable till cpu is completely disabled
4694 * in take_cpu_down(), so we prevent new cfs throttling here.
4696 cfs_rq
->runtime_enabled
= 0;
4698 if (cfs_rq_throttled(cfs_rq
))
4699 unthrottle_cfs_rq(cfs_rq
);
4704 #else /* CONFIG_CFS_BANDWIDTH */
4705 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
4707 return rq_clock_task(rq_of(cfs_rq
));
4710 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
4711 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
4712 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
4713 static inline void sync_throttle(struct task_group
*tg
, int cpu
) {}
4714 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4716 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4721 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4726 static inline int throttled_lb_pair(struct task_group
*tg
,
4727 int src_cpu
, int dest_cpu
)
4732 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4734 #ifdef CONFIG_FAIR_GROUP_SCHED
4735 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4738 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4742 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4743 static inline void update_runtime_enabled(struct rq
*rq
) {}
4744 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
4746 #endif /* CONFIG_CFS_BANDWIDTH */
4748 /**************************************************
4749 * CFS operations on tasks:
4752 #ifdef CONFIG_SCHED_HRTICK
4753 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4755 struct sched_entity
*se
= &p
->se
;
4756 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4758 SCHED_WARN_ON(task_rq(p
) != rq
);
4760 if (rq
->cfs
.h_nr_running
> 1) {
4761 u64 slice
= sched_slice(cfs_rq
, se
);
4762 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
4763 s64 delta
= slice
- ran
;
4770 hrtick_start(rq
, delta
);
4775 * called from enqueue/dequeue and updates the hrtick when the
4776 * current task is from our class and nr_running is low enough
4779 static void hrtick_update(struct rq
*rq
)
4781 struct task_struct
*curr
= rq
->curr
;
4783 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
4786 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
4787 hrtick_start_fair(rq
, curr
);
4789 #else /* !CONFIG_SCHED_HRTICK */
4791 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4795 static inline void hrtick_update(struct rq
*rq
)
4801 * The enqueue_task method is called before nr_running is
4802 * increased. Here we update the fair scheduling stats and
4803 * then put the task into the rbtree:
4806 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4808 struct cfs_rq
*cfs_rq
;
4809 struct sched_entity
*se
= &p
->se
;
4812 * If in_iowait is set, the code below may not trigger any cpufreq
4813 * utilization updates, so do it here explicitly with the IOWAIT flag
4817 cpufreq_update_this_cpu(rq
, SCHED_CPUFREQ_IOWAIT
);
4819 for_each_sched_entity(se
) {
4822 cfs_rq
= cfs_rq_of(se
);
4823 enqueue_entity(cfs_rq
, se
, flags
);
4826 * end evaluation on encountering a throttled cfs_rq
4828 * note: in the case of encountering a throttled cfs_rq we will
4829 * post the final h_nr_running increment below.
4831 if (cfs_rq_throttled(cfs_rq
))
4833 cfs_rq
->h_nr_running
++;
4835 flags
= ENQUEUE_WAKEUP
;
4838 for_each_sched_entity(se
) {
4839 cfs_rq
= cfs_rq_of(se
);
4840 cfs_rq
->h_nr_running
++;
4842 if (cfs_rq_throttled(cfs_rq
))
4845 update_load_avg(se
, UPDATE_TG
);
4846 update_cfs_shares(se
);
4850 add_nr_running(rq
, 1);
4855 static void set_next_buddy(struct sched_entity
*se
);
4858 * The dequeue_task method is called before nr_running is
4859 * decreased. We remove the task from the rbtree and
4860 * update the fair scheduling stats:
4862 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4864 struct cfs_rq
*cfs_rq
;
4865 struct sched_entity
*se
= &p
->se
;
4866 int task_sleep
= flags
& DEQUEUE_SLEEP
;
4868 for_each_sched_entity(se
) {
4869 cfs_rq
= cfs_rq_of(se
);
4870 dequeue_entity(cfs_rq
, se
, flags
);
4873 * end evaluation on encountering a throttled cfs_rq
4875 * note: in the case of encountering a throttled cfs_rq we will
4876 * post the final h_nr_running decrement below.
4878 if (cfs_rq_throttled(cfs_rq
))
4880 cfs_rq
->h_nr_running
--;
4882 /* Don't dequeue parent if it has other entities besides us */
4883 if (cfs_rq
->load
.weight
) {
4884 /* Avoid re-evaluating load for this entity: */
4885 se
= parent_entity(se
);
4887 * Bias pick_next to pick a task from this cfs_rq, as
4888 * p is sleeping when it is within its sched_slice.
4890 if (task_sleep
&& se
&& !throttled_hierarchy(cfs_rq
))
4894 flags
|= DEQUEUE_SLEEP
;
4897 for_each_sched_entity(se
) {
4898 cfs_rq
= cfs_rq_of(se
);
4899 cfs_rq
->h_nr_running
--;
4901 if (cfs_rq_throttled(cfs_rq
))
4904 update_load_avg(se
, UPDATE_TG
);
4905 update_cfs_shares(se
);
4909 sub_nr_running(rq
, 1);
4916 /* Working cpumask for: load_balance, load_balance_newidle. */
4917 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
4918 DEFINE_PER_CPU(cpumask_var_t
, select_idle_mask
);
4920 #ifdef CONFIG_NO_HZ_COMMON
4922 * per rq 'load' arrray crap; XXX kill this.
4926 * The exact cpuload calculated at every tick would be:
4928 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4930 * If a cpu misses updates for n ticks (as it was idle) and update gets
4931 * called on the n+1-th tick when cpu may be busy, then we have:
4933 * load_n = (1 - 1/2^i)^n * load_0
4934 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4936 * decay_load_missed() below does efficient calculation of
4938 * load' = (1 - 1/2^i)^n * load
4940 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4941 * This allows us to precompute the above in said factors, thereby allowing the
4942 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4943 * fixed_power_int())
4945 * The calculation is approximated on a 128 point scale.
4947 #define DEGRADE_SHIFT 7
4949 static const u8 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
4950 static const u8 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
4951 { 0, 0, 0, 0, 0, 0, 0, 0 },
4952 { 64, 32, 8, 0, 0, 0, 0, 0 },
4953 { 96, 72, 40, 12, 1, 0, 0, 0 },
4954 { 112, 98, 75, 43, 15, 1, 0, 0 },
4955 { 120, 112, 98, 76, 45, 16, 2, 0 }
4959 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4960 * would be when CPU is idle and so we just decay the old load without
4961 * adding any new load.
4963 static unsigned long
4964 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
4968 if (!missed_updates
)
4971 if (missed_updates
>= degrade_zero_ticks
[idx
])
4975 return load
>> missed_updates
;
4977 while (missed_updates
) {
4978 if (missed_updates
% 2)
4979 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
4981 missed_updates
>>= 1;
4986 #endif /* CONFIG_NO_HZ_COMMON */
4989 * __cpu_load_update - update the rq->cpu_load[] statistics
4990 * @this_rq: The rq to update statistics for
4991 * @this_load: The current load
4992 * @pending_updates: The number of missed updates
4994 * Update rq->cpu_load[] statistics. This function is usually called every
4995 * scheduler tick (TICK_NSEC).
4997 * This function computes a decaying average:
4999 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5001 * Because of NOHZ it might not get called on every tick which gives need for
5002 * the @pending_updates argument.
5004 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5005 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5006 * = A * (A * load[i]_n-2 + B) + B
5007 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5008 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5009 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5010 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5011 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5013 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5014 * any change in load would have resulted in the tick being turned back on.
5016 * For regular NOHZ, this reduces to:
5018 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5020 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5023 static void cpu_load_update(struct rq
*this_rq
, unsigned long this_load
,
5024 unsigned long pending_updates
)
5026 unsigned long __maybe_unused tickless_load
= this_rq
->cpu_load
[0];
5029 this_rq
->nr_load_updates
++;
5031 /* Update our load: */
5032 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
5033 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
5034 unsigned long old_load
, new_load
;
5036 /* scale is effectively 1 << i now, and >> i divides by scale */
5038 old_load
= this_rq
->cpu_load
[i
];
5039 #ifdef CONFIG_NO_HZ_COMMON
5040 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
5041 if (tickless_load
) {
5042 old_load
-= decay_load_missed(tickless_load
, pending_updates
- 1, i
);
5044 * old_load can never be a negative value because a
5045 * decayed tickless_load cannot be greater than the
5046 * original tickless_load.
5048 old_load
+= tickless_load
;
5051 new_load
= this_load
;
5053 * Round up the averaging division if load is increasing. This
5054 * prevents us from getting stuck on 9 if the load is 10, for
5057 if (new_load
> old_load
)
5058 new_load
+= scale
- 1;
5060 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
5063 sched_avg_update(this_rq
);
5066 /* Used instead of source_load when we know the type == 0 */
5067 static unsigned long weighted_cpuload(const int cpu
)
5069 return cfs_rq_runnable_load_avg(&cpu_rq(cpu
)->cfs
);
5072 #ifdef CONFIG_NO_HZ_COMMON
5074 * There is no sane way to deal with nohz on smp when using jiffies because the
5075 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5076 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5078 * Therefore we need to avoid the delta approach from the regular tick when
5079 * possible since that would seriously skew the load calculation. This is why we
5080 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5081 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5082 * loop exit, nohz_idle_balance, nohz full exit...)
5084 * This means we might still be one tick off for nohz periods.
5087 static void cpu_load_update_nohz(struct rq
*this_rq
,
5088 unsigned long curr_jiffies
,
5091 unsigned long pending_updates
;
5093 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
5094 if (pending_updates
) {
5095 this_rq
->last_load_update_tick
= curr_jiffies
;
5097 * In the regular NOHZ case, we were idle, this means load 0.
5098 * In the NOHZ_FULL case, we were non-idle, we should consider
5099 * its weighted load.
5101 cpu_load_update(this_rq
, load
, pending_updates
);
5106 * Called from nohz_idle_balance() to update the load ratings before doing the
5109 static void cpu_load_update_idle(struct rq
*this_rq
)
5112 * bail if there's load or we're actually up-to-date.
5114 if (weighted_cpuload(cpu_of(this_rq
)))
5117 cpu_load_update_nohz(this_rq
, READ_ONCE(jiffies
), 0);
5121 * Record CPU load on nohz entry so we know the tickless load to account
5122 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5123 * than other cpu_load[idx] but it should be fine as cpu_load readers
5124 * shouldn't rely into synchronized cpu_load[*] updates.
5126 void cpu_load_update_nohz_start(void)
5128 struct rq
*this_rq
= this_rq();
5131 * This is all lockless but should be fine. If weighted_cpuload changes
5132 * concurrently we'll exit nohz. And cpu_load write can race with
5133 * cpu_load_update_idle() but both updater would be writing the same.
5135 this_rq
->cpu_load
[0] = weighted_cpuload(cpu_of(this_rq
));
5139 * Account the tickless load in the end of a nohz frame.
5141 void cpu_load_update_nohz_stop(void)
5143 unsigned long curr_jiffies
= READ_ONCE(jiffies
);
5144 struct rq
*this_rq
= this_rq();
5148 if (curr_jiffies
== this_rq
->last_load_update_tick
)
5151 load
= weighted_cpuload(cpu_of(this_rq
));
5152 rq_lock(this_rq
, &rf
);
5153 update_rq_clock(this_rq
);
5154 cpu_load_update_nohz(this_rq
, curr_jiffies
, load
);
5155 rq_unlock(this_rq
, &rf
);
5157 #else /* !CONFIG_NO_HZ_COMMON */
5158 static inline void cpu_load_update_nohz(struct rq
*this_rq
,
5159 unsigned long curr_jiffies
,
5160 unsigned long load
) { }
5161 #endif /* CONFIG_NO_HZ_COMMON */
5163 static void cpu_load_update_periodic(struct rq
*this_rq
, unsigned long load
)
5165 #ifdef CONFIG_NO_HZ_COMMON
5166 /* See the mess around cpu_load_update_nohz(). */
5167 this_rq
->last_load_update_tick
= READ_ONCE(jiffies
);
5169 cpu_load_update(this_rq
, load
, 1);
5173 * Called from scheduler_tick()
5175 void cpu_load_update_active(struct rq
*this_rq
)
5177 unsigned long load
= weighted_cpuload(cpu_of(this_rq
));
5179 if (tick_nohz_tick_stopped())
5180 cpu_load_update_nohz(this_rq
, READ_ONCE(jiffies
), load
);
5182 cpu_load_update_periodic(this_rq
, load
);
5186 * Return a low guess at the load of a migration-source cpu weighted
5187 * according to the scheduling class and "nice" value.
5189 * We want to under-estimate the load of migration sources, to
5190 * balance conservatively.
5192 static unsigned long source_load(int cpu
, int type
)
5194 struct rq
*rq
= cpu_rq(cpu
);
5195 unsigned long total
= weighted_cpuload(cpu
);
5197 if (type
== 0 || !sched_feat(LB_BIAS
))
5200 return min(rq
->cpu_load
[type
-1], total
);
5204 * Return a high guess at the load of a migration-target cpu weighted
5205 * according to the scheduling class and "nice" value.
5207 static unsigned long target_load(int cpu
, int type
)
5209 struct rq
*rq
= cpu_rq(cpu
);
5210 unsigned long total
= weighted_cpuload(cpu
);
5212 if (type
== 0 || !sched_feat(LB_BIAS
))
5215 return max(rq
->cpu_load
[type
-1], total
);
5218 static unsigned long capacity_of(int cpu
)
5220 return cpu_rq(cpu
)->cpu_capacity
;
5223 static unsigned long capacity_orig_of(int cpu
)
5225 return cpu_rq(cpu
)->cpu_capacity_orig
;
5228 static unsigned long cpu_avg_load_per_task(int cpu
)
5230 struct rq
*rq
= cpu_rq(cpu
);
5231 unsigned long nr_running
= READ_ONCE(rq
->cfs
.h_nr_running
);
5232 unsigned long load_avg
= weighted_cpuload(cpu
);
5235 return load_avg
/ nr_running
;
5240 #ifdef CONFIG_FAIR_GROUP_SCHED
5242 * effective_load() calculates the load change as seen from the root_task_group
5244 * Adding load to a group doesn't make a group heavier, but can cause movement
5245 * of group shares between cpus. Assuming the shares were perfectly aligned one
5246 * can calculate the shift in shares.
5248 * Calculate the effective load difference if @wl is added (subtracted) to @tg
5249 * on this @cpu and results in a total addition (subtraction) of @wg to the
5250 * total group weight.
5252 * Given a runqueue weight distribution (rw_i) we can compute a shares
5253 * distribution (s_i) using:
5255 * s_i = rw_i / \Sum rw_j (1)
5257 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
5258 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
5259 * shares distribution (s_i):
5261 * rw_i = { 2, 4, 1, 0 }
5262 * s_i = { 2/7, 4/7, 1/7, 0 }
5264 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
5265 * task used to run on and the CPU the waker is running on), we need to
5266 * compute the effect of waking a task on either CPU and, in case of a sync
5267 * wakeup, compute the effect of the current task going to sleep.
5269 * So for a change of @wl to the local @cpu with an overall group weight change
5270 * of @wl we can compute the new shares distribution (s'_i) using:
5272 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
5274 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
5275 * differences in waking a task to CPU 0. The additional task changes the
5276 * weight and shares distributions like:
5278 * rw'_i = { 3, 4, 1, 0 }
5279 * s'_i = { 3/8, 4/8, 1/8, 0 }
5281 * We can then compute the difference in effective weight by using:
5283 * dw_i = S * (s'_i - s_i) (3)
5285 * Where 'S' is the group weight as seen by its parent.
5287 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
5288 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
5289 * 4/7) times the weight of the group.
5291 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
5293 struct sched_entity
*se
= tg
->se
[cpu
];
5295 if (!tg
->parent
) /* the trivial, non-cgroup case */
5298 for_each_sched_entity(se
) {
5299 struct cfs_rq
*cfs_rq
= se
->my_q
;
5300 long W
, w
= cfs_rq_load_avg(cfs_rq
);
5305 * W = @wg + \Sum rw_j
5307 W
= wg
+ atomic_long_read(&tg
->load_avg
);
5309 /* Ensure \Sum rw_j >= rw_i */
5310 W
-= cfs_rq
->tg_load_avg_contrib
;
5319 * wl = S * s'_i; see (2)
5322 wl
= (w
* (long)scale_load_down(tg
->shares
)) / W
;
5324 wl
= scale_load_down(tg
->shares
);
5327 * Per the above, wl is the new se->load.weight value; since
5328 * those are clipped to [MIN_SHARES, ...) do so now. See
5329 * calc_cfs_shares().
5331 if (wl
< MIN_SHARES
)
5335 * wl = dw_i = S * (s'_i - s_i); see (3)
5337 wl
-= se
->avg
.load_avg
;
5340 * Recursively apply this logic to all parent groups to compute
5341 * the final effective load change on the root group. Since
5342 * only the @tg group gets extra weight, all parent groups can
5343 * only redistribute existing shares. @wl is the shift in shares
5344 * resulting from this level per the above.
5353 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
5360 static void record_wakee(struct task_struct
*p
)
5363 * Only decay a single time; tasks that have less then 1 wakeup per
5364 * jiffy will not have built up many flips.
5366 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
5367 current
->wakee_flips
>>= 1;
5368 current
->wakee_flip_decay_ts
= jiffies
;
5371 if (current
->last_wakee
!= p
) {
5372 current
->last_wakee
= p
;
5373 current
->wakee_flips
++;
5378 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5380 * A waker of many should wake a different task than the one last awakened
5381 * at a frequency roughly N times higher than one of its wakees.
5383 * In order to determine whether we should let the load spread vs consolidating
5384 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5385 * partner, and a factor of lls_size higher frequency in the other.
5387 * With both conditions met, we can be relatively sure that the relationship is
5388 * non-monogamous, with partner count exceeding socket size.
5390 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5391 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5394 static int wake_wide(struct task_struct
*p
)
5396 unsigned int master
= current
->wakee_flips
;
5397 unsigned int slave
= p
->wakee_flips
;
5398 int factor
= this_cpu_read(sd_llc_size
);
5401 swap(master
, slave
);
5402 if (slave
< factor
|| master
< slave
* factor
)
5407 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
,
5408 int prev_cpu
, int sync
)
5410 s64 this_load
, load
;
5411 s64 this_eff_load
, prev_eff_load
;
5413 struct task_group
*tg
;
5414 unsigned long weight
;
5418 this_cpu
= smp_processor_id();
5419 load
= source_load(prev_cpu
, idx
);
5420 this_load
= target_load(this_cpu
, idx
);
5423 * If sync wakeup then subtract the (maximum possible)
5424 * effect of the currently running task from the load
5425 * of the current CPU:
5428 tg
= task_group(current
);
5429 weight
= current
->se
.avg
.load_avg
;
5431 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
5432 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
5436 weight
= p
->se
.avg
.load_avg
;
5439 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5440 * due to the sync cause above having dropped this_load to 0, we'll
5441 * always have an imbalance, but there's really nothing you can do
5442 * about that, so that's good too.
5444 * Otherwise check if either cpus are near enough in load to allow this
5445 * task to be woken on this_cpu.
5447 this_eff_load
= 100;
5448 this_eff_load
*= capacity_of(prev_cpu
);
5450 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
5451 prev_eff_load
*= capacity_of(this_cpu
);
5453 if (this_load
> 0) {
5454 this_eff_load
*= this_load
+
5455 effective_load(tg
, this_cpu
, weight
, weight
);
5457 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
5460 balanced
= this_eff_load
<= prev_eff_load
;
5462 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine_attempts
);
5467 schedstat_inc(sd
->ttwu_move_affine
);
5468 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine
);
5473 static inline int task_util(struct task_struct
*p
);
5474 static int cpu_util_wake(int cpu
, struct task_struct
*p
);
5476 static unsigned long capacity_spare_wake(int cpu
, struct task_struct
*p
)
5478 return capacity_orig_of(cpu
) - cpu_util_wake(cpu
, p
);
5482 * find_idlest_group finds and returns the least busy CPU group within the
5485 static struct sched_group
*
5486 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
5487 int this_cpu
, int sd_flag
)
5489 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
5490 struct sched_group
*most_spare_sg
= NULL
;
5491 unsigned long min_runnable_load
= ULONG_MAX
, this_runnable_load
= 0;
5492 unsigned long min_avg_load
= ULONG_MAX
, this_avg_load
= 0;
5493 unsigned long most_spare
= 0, this_spare
= 0;
5494 int load_idx
= sd
->forkexec_idx
;
5495 int imbalance_scale
= 100 + (sd
->imbalance_pct
-100)/2;
5496 unsigned long imbalance
= scale_load_down(NICE_0_LOAD
) *
5497 (sd
->imbalance_pct
-100) / 100;
5499 if (sd_flag
& SD_BALANCE_WAKE
)
5500 load_idx
= sd
->wake_idx
;
5503 unsigned long load
, avg_load
, runnable_load
;
5504 unsigned long spare_cap
, max_spare_cap
;
5508 /* Skip over this group if it has no CPUs allowed */
5509 if (!cpumask_intersects(sched_group_span(group
),
5513 local_group
= cpumask_test_cpu(this_cpu
,
5514 sched_group_span(group
));
5517 * Tally up the load of all CPUs in the group and find
5518 * the group containing the CPU with most spare capacity.
5524 for_each_cpu(i
, sched_group_span(group
)) {
5525 /* Bias balancing toward cpus of our domain */
5527 load
= source_load(i
, load_idx
);
5529 load
= target_load(i
, load_idx
);
5531 runnable_load
+= load
;
5533 avg_load
+= cfs_rq_load_avg(&cpu_rq(i
)->cfs
);
5535 spare_cap
= capacity_spare_wake(i
, p
);
5537 if (spare_cap
> max_spare_cap
)
5538 max_spare_cap
= spare_cap
;
5541 /* Adjust by relative CPU capacity of the group */
5542 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) /
5543 group
->sgc
->capacity
;
5544 runnable_load
= (runnable_load
* SCHED_CAPACITY_SCALE
) /
5545 group
->sgc
->capacity
;
5548 this_runnable_load
= runnable_load
;
5549 this_avg_load
= avg_load
;
5550 this_spare
= max_spare_cap
;
5552 if (min_runnable_load
> (runnable_load
+ imbalance
)) {
5554 * The runnable load is significantly smaller
5555 * so we can pick this new cpu
5557 min_runnable_load
= runnable_load
;
5558 min_avg_load
= avg_load
;
5560 } else if ((runnable_load
< (min_runnable_load
+ imbalance
)) &&
5561 (100*min_avg_load
> imbalance_scale
*avg_load
)) {
5563 * The runnable loads are close so take the
5564 * blocked load into account through avg_load.
5566 min_avg_load
= avg_load
;
5570 if (most_spare
< max_spare_cap
) {
5571 most_spare
= max_spare_cap
;
5572 most_spare_sg
= group
;
5575 } while (group
= group
->next
, group
!= sd
->groups
);
5578 * The cross-over point between using spare capacity or least load
5579 * is too conservative for high utilization tasks on partially
5580 * utilized systems if we require spare_capacity > task_util(p),
5581 * so we allow for some task stuffing by using
5582 * spare_capacity > task_util(p)/2.
5584 * Spare capacity can't be used for fork because the utilization has
5585 * not been set yet, we must first select a rq to compute the initial
5588 if (sd_flag
& SD_BALANCE_FORK
)
5591 if (this_spare
> task_util(p
) / 2 &&
5592 imbalance_scale
*this_spare
> 100*most_spare
)
5595 if (most_spare
> task_util(p
) / 2)
5596 return most_spare_sg
;
5602 if (min_runnable_load
> (this_runnable_load
+ imbalance
))
5605 if ((this_runnable_load
< (min_runnable_load
+ imbalance
)) &&
5606 (100*this_avg_load
< imbalance_scale
*min_avg_load
))
5613 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5616 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
5618 unsigned long load
, min_load
= ULONG_MAX
;
5619 unsigned int min_exit_latency
= UINT_MAX
;
5620 u64 latest_idle_timestamp
= 0;
5621 int least_loaded_cpu
= this_cpu
;
5622 int shallowest_idle_cpu
= -1;
5625 /* Check if we have any choice: */
5626 if (group
->group_weight
== 1)
5627 return cpumask_first(sched_group_span(group
));
5629 /* Traverse only the allowed CPUs */
5630 for_each_cpu_and(i
, sched_group_span(group
), &p
->cpus_allowed
) {
5632 struct rq
*rq
= cpu_rq(i
);
5633 struct cpuidle_state
*idle
= idle_get_state(rq
);
5634 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
5636 * We give priority to a CPU whose idle state
5637 * has the smallest exit latency irrespective
5638 * of any idle timestamp.
5640 min_exit_latency
= idle
->exit_latency
;
5641 latest_idle_timestamp
= rq
->idle_stamp
;
5642 shallowest_idle_cpu
= i
;
5643 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
5644 rq
->idle_stamp
> latest_idle_timestamp
) {
5646 * If equal or no active idle state, then
5647 * the most recently idled CPU might have
5650 latest_idle_timestamp
= rq
->idle_stamp
;
5651 shallowest_idle_cpu
= i
;
5653 } else if (shallowest_idle_cpu
== -1) {
5654 load
= weighted_cpuload(i
);
5655 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
5657 least_loaded_cpu
= i
;
5662 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
5665 #ifdef CONFIG_SCHED_SMT
5667 static inline void set_idle_cores(int cpu
, int val
)
5669 struct sched_domain_shared
*sds
;
5671 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
5673 WRITE_ONCE(sds
->has_idle_cores
, val
);
5676 static inline bool test_idle_cores(int cpu
, bool def
)
5678 struct sched_domain_shared
*sds
;
5680 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
5682 return READ_ONCE(sds
->has_idle_cores
);
5688 * Scans the local SMT mask to see if the entire core is idle, and records this
5689 * information in sd_llc_shared->has_idle_cores.
5691 * Since SMT siblings share all cache levels, inspecting this limited remote
5692 * state should be fairly cheap.
5694 void __update_idle_core(struct rq
*rq
)
5696 int core
= cpu_of(rq
);
5700 if (test_idle_cores(core
, true))
5703 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
5711 set_idle_cores(core
, 1);
5717 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5718 * there are no idle cores left in the system; tracked through
5719 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5721 static int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5723 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
5726 if (!static_branch_likely(&sched_smt_present
))
5729 if (!test_idle_cores(target
, false))
5732 cpumask_and(cpus
, sched_domain_span(sd
), &p
->cpus_allowed
);
5734 for_each_cpu_wrap(core
, cpus
, target
) {
5737 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
5738 cpumask_clear_cpu(cpu
, cpus
);
5748 * Failed to find an idle core; stop looking for one.
5750 set_idle_cores(target
, 0);
5756 * Scan the local SMT mask for idle CPUs.
5758 static int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5762 if (!static_branch_likely(&sched_smt_present
))
5765 for_each_cpu(cpu
, cpu_smt_mask(target
)) {
5766 if (!cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
5775 #else /* CONFIG_SCHED_SMT */
5777 static inline int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5782 static inline int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5787 #endif /* CONFIG_SCHED_SMT */
5790 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5791 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5792 * average idle time for this rq (as found in rq->avg_idle).
5794 static int select_idle_cpu(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5796 struct sched_domain
*this_sd
;
5797 u64 avg_cost
, avg_idle
;
5800 int cpu
, nr
= INT_MAX
;
5802 this_sd
= rcu_dereference(*this_cpu_ptr(&sd_llc
));
5807 * Due to large variance we need a large fuzz factor; hackbench in
5808 * particularly is sensitive here.
5810 avg_idle
= this_rq()->avg_idle
/ 512;
5811 avg_cost
= this_sd
->avg_scan_cost
+ 1;
5813 if (sched_feat(SIS_AVG_CPU
) && avg_idle
< avg_cost
)
5816 if (sched_feat(SIS_PROP
)) {
5817 u64 span_avg
= sd
->span_weight
* avg_idle
;
5818 if (span_avg
> 4*avg_cost
)
5819 nr
= div_u64(span_avg
, avg_cost
);
5824 time
= local_clock();
5826 for_each_cpu_wrap(cpu
, sched_domain_span(sd
), target
) {
5829 if (!cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
5835 time
= local_clock() - time
;
5836 cost
= this_sd
->avg_scan_cost
;
5837 delta
= (s64
)(time
- cost
) / 8;
5838 this_sd
->avg_scan_cost
+= delta
;
5844 * Try and locate an idle core/thread in the LLC cache domain.
5846 static int select_idle_sibling(struct task_struct
*p
, int prev
, int target
)
5848 struct sched_domain
*sd
;
5851 if (idle_cpu(target
))
5855 * If the previous cpu is cache affine and idle, don't be stupid.
5857 if (prev
!= target
&& cpus_share_cache(prev
, target
) && idle_cpu(prev
))
5860 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
5864 i
= select_idle_core(p
, sd
, target
);
5865 if ((unsigned)i
< nr_cpumask_bits
)
5868 i
= select_idle_cpu(p
, sd
, target
);
5869 if ((unsigned)i
< nr_cpumask_bits
)
5872 i
= select_idle_smt(p
, sd
, target
);
5873 if ((unsigned)i
< nr_cpumask_bits
)
5880 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5881 * tasks. The unit of the return value must be the one of capacity so we can
5882 * compare the utilization with the capacity of the CPU that is available for
5883 * CFS task (ie cpu_capacity).
5885 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5886 * recent utilization of currently non-runnable tasks on a CPU. It represents
5887 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5888 * capacity_orig is the cpu_capacity available at the highest frequency
5889 * (arch_scale_freq_capacity()).
5890 * The utilization of a CPU converges towards a sum equal to or less than the
5891 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5892 * the running time on this CPU scaled by capacity_curr.
5894 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5895 * higher than capacity_orig because of unfortunate rounding in
5896 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5897 * the average stabilizes with the new running time. We need to check that the
5898 * utilization stays within the range of [0..capacity_orig] and cap it if
5899 * necessary. Without utilization capping, a group could be seen as overloaded
5900 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5901 * available capacity. We allow utilization to overshoot capacity_curr (but not
5902 * capacity_orig) as it useful for predicting the capacity required after task
5903 * migrations (scheduler-driven DVFS).
5905 static int cpu_util(int cpu
)
5907 unsigned long util
= cpu_rq(cpu
)->cfs
.avg
.util_avg
;
5908 unsigned long capacity
= capacity_orig_of(cpu
);
5910 return (util
>= capacity
) ? capacity
: util
;
5913 static inline int task_util(struct task_struct
*p
)
5915 return p
->se
.avg
.util_avg
;
5919 * cpu_util_wake: Compute cpu utilization with any contributions from
5920 * the waking task p removed.
5922 static int cpu_util_wake(int cpu
, struct task_struct
*p
)
5924 unsigned long util
, capacity
;
5926 /* Task has no contribution or is new */
5927 if (cpu
!= task_cpu(p
) || !p
->se
.avg
.last_update_time
)
5928 return cpu_util(cpu
);
5930 capacity
= capacity_orig_of(cpu
);
5931 util
= max_t(long, cpu_rq(cpu
)->cfs
.avg
.util_avg
- task_util(p
), 0);
5933 return (util
>= capacity
) ? capacity
: util
;
5937 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5938 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5940 * In that case WAKE_AFFINE doesn't make sense and we'll let
5941 * BALANCE_WAKE sort things out.
5943 static int wake_cap(struct task_struct
*p
, int cpu
, int prev_cpu
)
5945 long min_cap
, max_cap
;
5947 min_cap
= min(capacity_orig_of(prev_cpu
), capacity_orig_of(cpu
));
5948 max_cap
= cpu_rq(cpu
)->rd
->max_cpu_capacity
;
5950 /* Minimum capacity is close to max, no need to abort wake_affine */
5951 if (max_cap
- min_cap
< max_cap
>> 3)
5954 /* Bring task utilization in sync with prev_cpu */
5955 sync_entity_load_avg(&p
->se
);
5957 return min_cap
* 1024 < task_util(p
) * capacity_margin
;
5961 * select_task_rq_fair: Select target runqueue for the waking task in domains
5962 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5963 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5965 * Balances load by selecting the idlest cpu in the idlest group, or under
5966 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5968 * Returns the target cpu number.
5970 * preempt must be disabled.
5973 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
5975 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
5976 int cpu
= smp_processor_id();
5977 int new_cpu
= prev_cpu
;
5978 int want_affine
= 0;
5979 int sync
= wake_flags
& WF_SYNC
;
5981 if (sd_flag
& SD_BALANCE_WAKE
) {
5983 want_affine
= !wake_wide(p
) && !wake_cap(p
, cpu
, prev_cpu
)
5984 && cpumask_test_cpu(cpu
, &p
->cpus_allowed
);
5988 for_each_domain(cpu
, tmp
) {
5989 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
5993 * If both cpu and prev_cpu are part of this domain,
5994 * cpu is a valid SD_WAKE_AFFINE target.
5996 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
5997 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
6002 if (tmp
->flags
& sd_flag
)
6004 else if (!want_affine
)
6009 sd
= NULL
; /* Prefer wake_affine over balance flags */
6010 if (cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, prev_cpu
, sync
))
6015 if (sd_flag
& SD_BALANCE_WAKE
) /* XXX always ? */
6016 new_cpu
= select_idle_sibling(p
, prev_cpu
, new_cpu
);
6019 struct sched_group
*group
;
6022 if (!(sd
->flags
& sd_flag
)) {
6027 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
6033 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
6034 if (new_cpu
== -1 || new_cpu
== cpu
) {
6035 /* Now try balancing at a lower domain level of cpu */
6040 /* Now try balancing at a lower domain level of new_cpu */
6042 weight
= sd
->span_weight
;
6044 for_each_domain(cpu
, tmp
) {
6045 if (weight
<= tmp
->span_weight
)
6047 if (tmp
->flags
& sd_flag
)
6050 /* while loop will break here if sd == NULL */
6058 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6059 * cfs_rq_of(p) references at time of call are still valid and identify the
6060 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6062 static void migrate_task_rq_fair(struct task_struct
*p
)
6065 * As blocked tasks retain absolute vruntime the migration needs to
6066 * deal with this by subtracting the old and adding the new
6067 * min_vruntime -- the latter is done by enqueue_entity() when placing
6068 * the task on the new runqueue.
6070 if (p
->state
== TASK_WAKING
) {
6071 struct sched_entity
*se
= &p
->se
;
6072 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6075 #ifndef CONFIG_64BIT
6076 u64 min_vruntime_copy
;
6079 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
6081 min_vruntime
= cfs_rq
->min_vruntime
;
6082 } while (min_vruntime
!= min_vruntime_copy
);
6084 min_vruntime
= cfs_rq
->min_vruntime
;
6087 se
->vruntime
-= min_vruntime
;
6091 * We are supposed to update the task to "current" time, then its up to date
6092 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6093 * what current time is, so simply throw away the out-of-date time. This
6094 * will result in the wakee task is less decayed, but giving the wakee more
6095 * load sounds not bad.
6097 remove_entity_load_avg(&p
->se
);
6099 /* Tell new CPU we are migrated */
6100 p
->se
.avg
.last_update_time
= 0;
6102 /* We have migrated, no longer consider this task hot */
6103 p
->se
.exec_start
= 0;
6106 static void task_dead_fair(struct task_struct
*p
)
6108 remove_entity_load_avg(&p
->se
);
6110 #endif /* CONFIG_SMP */
6112 static unsigned long
6113 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
6115 unsigned long gran
= sysctl_sched_wakeup_granularity
;
6118 * Since its curr running now, convert the gran from real-time
6119 * to virtual-time in his units.
6121 * By using 'se' instead of 'curr' we penalize light tasks, so
6122 * they get preempted easier. That is, if 'se' < 'curr' then
6123 * the resulting gran will be larger, therefore penalizing the
6124 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6125 * be smaller, again penalizing the lighter task.
6127 * This is especially important for buddies when the leftmost
6128 * task is higher priority than the buddy.
6130 return calc_delta_fair(gran
, se
);
6134 * Should 'se' preempt 'curr'.
6148 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
6150 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
6155 gran
= wakeup_gran(curr
, se
);
6162 static void set_last_buddy(struct sched_entity
*se
)
6164 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
6167 for_each_sched_entity(se
)
6168 cfs_rq_of(se
)->last
= se
;
6171 static void set_next_buddy(struct sched_entity
*se
)
6173 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
6176 for_each_sched_entity(se
)
6177 cfs_rq_of(se
)->next
= se
;
6180 static void set_skip_buddy(struct sched_entity
*se
)
6182 for_each_sched_entity(se
)
6183 cfs_rq_of(se
)->skip
= se
;
6187 * Preempt the current task with a newly woken task if needed:
6189 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
6191 struct task_struct
*curr
= rq
->curr
;
6192 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
6193 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
6194 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
6195 int next_buddy_marked
= 0;
6197 if (unlikely(se
== pse
))
6201 * This is possible from callers such as attach_tasks(), in which we
6202 * unconditionally check_prempt_curr() after an enqueue (which may have
6203 * lead to a throttle). This both saves work and prevents false
6204 * next-buddy nomination below.
6206 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
6209 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
6210 set_next_buddy(pse
);
6211 next_buddy_marked
= 1;
6215 * We can come here with TIF_NEED_RESCHED already set from new task
6218 * Note: this also catches the edge-case of curr being in a throttled
6219 * group (e.g. via set_curr_task), since update_curr() (in the
6220 * enqueue of curr) will have resulted in resched being set. This
6221 * prevents us from potentially nominating it as a false LAST_BUDDY
6224 if (test_tsk_need_resched(curr
))
6227 /* Idle tasks are by definition preempted by non-idle tasks. */
6228 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
6229 likely(p
->policy
!= SCHED_IDLE
))
6233 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6234 * is driven by the tick):
6236 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
6239 find_matching_se(&se
, &pse
);
6240 update_curr(cfs_rq_of(se
));
6242 if (wakeup_preempt_entity(se
, pse
) == 1) {
6244 * Bias pick_next to pick the sched entity that is
6245 * triggering this preemption.
6247 if (!next_buddy_marked
)
6248 set_next_buddy(pse
);
6257 * Only set the backward buddy when the current task is still
6258 * on the rq. This can happen when a wakeup gets interleaved
6259 * with schedule on the ->pre_schedule() or idle_balance()
6260 * point, either of which can * drop the rq lock.
6262 * Also, during early boot the idle thread is in the fair class,
6263 * for obvious reasons its a bad idea to schedule back to it.
6265 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
6268 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
6272 static struct task_struct
*
6273 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
6275 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
6276 struct sched_entity
*se
;
6277 struct task_struct
*p
;
6281 #ifdef CONFIG_FAIR_GROUP_SCHED
6282 if (!cfs_rq
->nr_running
)
6285 if (prev
->sched_class
!= &fair_sched_class
)
6289 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6290 * likely that a next task is from the same cgroup as the current.
6292 * Therefore attempt to avoid putting and setting the entire cgroup
6293 * hierarchy, only change the part that actually changes.
6297 struct sched_entity
*curr
= cfs_rq
->curr
;
6300 * Since we got here without doing put_prev_entity() we also
6301 * have to consider cfs_rq->curr. If it is still a runnable
6302 * entity, update_curr() will update its vruntime, otherwise
6303 * forget we've ever seen it.
6307 update_curr(cfs_rq
);
6312 * This call to check_cfs_rq_runtime() will do the
6313 * throttle and dequeue its entity in the parent(s).
6314 * Therefore the 'simple' nr_running test will indeed
6317 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
6321 se
= pick_next_entity(cfs_rq
, curr
);
6322 cfs_rq
= group_cfs_rq(se
);
6328 * Since we haven't yet done put_prev_entity and if the selected task
6329 * is a different task than we started out with, try and touch the
6330 * least amount of cfs_rqs.
6333 struct sched_entity
*pse
= &prev
->se
;
6335 while (!(cfs_rq
= is_same_group(se
, pse
))) {
6336 int se_depth
= se
->depth
;
6337 int pse_depth
= pse
->depth
;
6339 if (se_depth
<= pse_depth
) {
6340 put_prev_entity(cfs_rq_of(pse
), pse
);
6341 pse
= parent_entity(pse
);
6343 if (se_depth
>= pse_depth
) {
6344 set_next_entity(cfs_rq_of(se
), se
);
6345 se
= parent_entity(se
);
6349 put_prev_entity(cfs_rq
, pse
);
6350 set_next_entity(cfs_rq
, se
);
6353 if (hrtick_enabled(rq
))
6354 hrtick_start_fair(rq
, p
);
6361 if (!cfs_rq
->nr_running
)
6364 put_prev_task(rq
, prev
);
6367 se
= pick_next_entity(cfs_rq
, NULL
);
6368 set_next_entity(cfs_rq
, se
);
6369 cfs_rq
= group_cfs_rq(se
);
6374 if (hrtick_enabled(rq
))
6375 hrtick_start_fair(rq
, p
);
6380 new_tasks
= idle_balance(rq
, rf
);
6383 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6384 * possible for any higher priority task to appear. In that case we
6385 * must re-start the pick_next_entity() loop.
6397 * Account for a descheduled task:
6399 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
6401 struct sched_entity
*se
= &prev
->se
;
6402 struct cfs_rq
*cfs_rq
;
6404 for_each_sched_entity(se
) {
6405 cfs_rq
= cfs_rq_of(se
);
6406 put_prev_entity(cfs_rq
, se
);
6411 * sched_yield() is very simple
6413 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6415 static void yield_task_fair(struct rq
*rq
)
6417 struct task_struct
*curr
= rq
->curr
;
6418 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
6419 struct sched_entity
*se
= &curr
->se
;
6422 * Are we the only task in the tree?
6424 if (unlikely(rq
->nr_running
== 1))
6427 clear_buddies(cfs_rq
, se
);
6429 if (curr
->policy
!= SCHED_BATCH
) {
6430 update_rq_clock(rq
);
6432 * Update run-time statistics of the 'current'.
6434 update_curr(cfs_rq
);
6436 * Tell update_rq_clock() that we've just updated,
6437 * so we don't do microscopic update in schedule()
6438 * and double the fastpath cost.
6440 rq_clock_skip_update(rq
, true);
6446 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
6448 struct sched_entity
*se
= &p
->se
;
6450 /* throttled hierarchies are not runnable */
6451 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
6454 /* Tell the scheduler that we'd really like pse to run next. */
6457 yield_task_fair(rq
);
6463 /**************************************************
6464 * Fair scheduling class load-balancing methods.
6468 * The purpose of load-balancing is to achieve the same basic fairness the
6469 * per-cpu scheduler provides, namely provide a proportional amount of compute
6470 * time to each task. This is expressed in the following equation:
6472 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6474 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6475 * W_i,0 is defined as:
6477 * W_i,0 = \Sum_j w_i,j (2)
6479 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6480 * is derived from the nice value as per sched_prio_to_weight[].
6482 * The weight average is an exponential decay average of the instantaneous
6485 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6487 * C_i is the compute capacity of cpu i, typically it is the
6488 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6489 * can also include other factors [XXX].
6491 * To achieve this balance we define a measure of imbalance which follows
6492 * directly from (1):
6494 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6496 * We them move tasks around to minimize the imbalance. In the continuous
6497 * function space it is obvious this converges, in the discrete case we get
6498 * a few fun cases generally called infeasible weight scenarios.
6501 * - infeasible weights;
6502 * - local vs global optima in the discrete case. ]
6507 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6508 * for all i,j solution, we create a tree of cpus that follows the hardware
6509 * topology where each level pairs two lower groups (or better). This results
6510 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6511 * tree to only the first of the previous level and we decrease the frequency
6512 * of load-balance at each level inv. proportional to the number of cpus in
6518 * \Sum { --- * --- * 2^i } = O(n) (5)
6520 * `- size of each group
6521 * | | `- number of cpus doing load-balance
6523 * `- sum over all levels
6525 * Coupled with a limit on how many tasks we can migrate every balance pass,
6526 * this makes (5) the runtime complexity of the balancer.
6528 * An important property here is that each CPU is still (indirectly) connected
6529 * to every other cpu in at most O(log n) steps:
6531 * The adjacency matrix of the resulting graph is given by:
6534 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6537 * And you'll find that:
6539 * A^(log_2 n)_i,j != 0 for all i,j (7)
6541 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6542 * The task movement gives a factor of O(m), giving a convergence complexity
6545 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6550 * In order to avoid CPUs going idle while there's still work to do, new idle
6551 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6552 * tree itself instead of relying on other CPUs to bring it work.
6554 * This adds some complexity to both (5) and (8) but it reduces the total idle
6562 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6565 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6570 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6572 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6574 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6577 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6578 * rewrite all of this once again.]
6581 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
6583 enum fbq_type
{ regular
, remote
, all
};
6585 #define LBF_ALL_PINNED 0x01
6586 #define LBF_NEED_BREAK 0x02
6587 #define LBF_DST_PINNED 0x04
6588 #define LBF_SOME_PINNED 0x08
6591 struct sched_domain
*sd
;
6599 struct cpumask
*dst_grpmask
;
6601 enum cpu_idle_type idle
;
6603 /* The set of CPUs under consideration for load-balancing */
6604 struct cpumask
*cpus
;
6609 unsigned int loop_break
;
6610 unsigned int loop_max
;
6612 enum fbq_type fbq_type
;
6613 struct list_head tasks
;
6617 * Is this task likely cache-hot:
6619 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
6623 lockdep_assert_held(&env
->src_rq
->lock
);
6625 if (p
->sched_class
!= &fair_sched_class
)
6628 if (unlikely(p
->policy
== SCHED_IDLE
))
6632 * Buddy candidates are cache hot:
6634 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
6635 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
6636 &p
->se
== cfs_rq_of(&p
->se
)->last
))
6639 if (sysctl_sched_migration_cost
== -1)
6641 if (sysctl_sched_migration_cost
== 0)
6644 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
6646 return delta
< (s64
)sysctl_sched_migration_cost
;
6649 #ifdef CONFIG_NUMA_BALANCING
6651 * Returns 1, if task migration degrades locality
6652 * Returns 0, if task migration improves locality i.e migration preferred.
6653 * Returns -1, if task migration is not affected by locality.
6655 static int migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
6657 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
6658 unsigned long src_faults
, dst_faults
;
6659 int src_nid
, dst_nid
;
6661 if (!static_branch_likely(&sched_numa_balancing
))
6664 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
6667 src_nid
= cpu_to_node(env
->src_cpu
);
6668 dst_nid
= cpu_to_node(env
->dst_cpu
);
6670 if (src_nid
== dst_nid
)
6673 /* Migrating away from the preferred node is always bad. */
6674 if (src_nid
== p
->numa_preferred_nid
) {
6675 if (env
->src_rq
->nr_running
> env
->src_rq
->nr_preferred_running
)
6681 /* Encourage migration to the preferred node. */
6682 if (dst_nid
== p
->numa_preferred_nid
)
6686 src_faults
= group_faults(p
, src_nid
);
6687 dst_faults
= group_faults(p
, dst_nid
);
6689 src_faults
= task_faults(p
, src_nid
);
6690 dst_faults
= task_faults(p
, dst_nid
);
6693 return dst_faults
< src_faults
;
6697 static inline int migrate_degrades_locality(struct task_struct
*p
,
6705 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6708 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
6712 lockdep_assert_held(&env
->src_rq
->lock
);
6715 * We do not migrate tasks that are:
6716 * 1) throttled_lb_pair, or
6717 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6718 * 3) running (obviously), or
6719 * 4) are cache-hot on their current CPU.
6721 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
6724 if (!cpumask_test_cpu(env
->dst_cpu
, &p
->cpus_allowed
)) {
6727 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_affine
);
6729 env
->flags
|= LBF_SOME_PINNED
;
6732 * Remember if this task can be migrated to any other cpu in
6733 * our sched_group. We may want to revisit it if we couldn't
6734 * meet load balance goals by pulling other tasks on src_cpu.
6736 * Also avoid computing new_dst_cpu if we have already computed
6737 * one in current iteration.
6739 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
6742 /* Prevent to re-select dst_cpu via env's cpus */
6743 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
6744 if (cpumask_test_cpu(cpu
, &p
->cpus_allowed
)) {
6745 env
->flags
|= LBF_DST_PINNED
;
6746 env
->new_dst_cpu
= cpu
;
6754 /* Record that we found atleast one task that could run on dst_cpu */
6755 env
->flags
&= ~LBF_ALL_PINNED
;
6757 if (task_running(env
->src_rq
, p
)) {
6758 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_running
);
6763 * Aggressive migration if:
6764 * 1) destination numa is preferred
6765 * 2) task is cache cold, or
6766 * 3) too many balance attempts have failed.
6768 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
6769 if (tsk_cache_hot
== -1)
6770 tsk_cache_hot
= task_hot(p
, env
);
6772 if (tsk_cache_hot
<= 0 ||
6773 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
6774 if (tsk_cache_hot
== 1) {
6775 schedstat_inc(env
->sd
->lb_hot_gained
[env
->idle
]);
6776 schedstat_inc(p
->se
.statistics
.nr_forced_migrations
);
6781 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_hot
);
6786 * detach_task() -- detach the task for the migration specified in env
6788 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
6790 lockdep_assert_held(&env
->src_rq
->lock
);
6792 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
6793 deactivate_task(env
->src_rq
, p
, DEQUEUE_NOCLOCK
);
6794 set_task_cpu(p
, env
->dst_cpu
);
6798 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6799 * part of active balancing operations within "domain".
6801 * Returns a task if successful and NULL otherwise.
6803 static struct task_struct
*detach_one_task(struct lb_env
*env
)
6805 struct task_struct
*p
, *n
;
6807 lockdep_assert_held(&env
->src_rq
->lock
);
6809 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
6810 if (!can_migrate_task(p
, env
))
6813 detach_task(p
, env
);
6816 * Right now, this is only the second place where
6817 * lb_gained[env->idle] is updated (other is detach_tasks)
6818 * so we can safely collect stats here rather than
6819 * inside detach_tasks().
6821 schedstat_inc(env
->sd
->lb_gained
[env
->idle
]);
6827 static const unsigned int sched_nr_migrate_break
= 32;
6830 * detach_tasks() -- tries to detach up to imbalance weighted load from
6831 * busiest_rq, as part of a balancing operation within domain "sd".
6833 * Returns number of detached tasks if successful and 0 otherwise.
6835 static int detach_tasks(struct lb_env
*env
)
6837 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
6838 struct task_struct
*p
;
6842 lockdep_assert_held(&env
->src_rq
->lock
);
6844 if (env
->imbalance
<= 0)
6847 while (!list_empty(tasks
)) {
6849 * We don't want to steal all, otherwise we may be treated likewise,
6850 * which could at worst lead to a livelock crash.
6852 if (env
->idle
!= CPU_NOT_IDLE
&& env
->src_rq
->nr_running
<= 1)
6855 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
6858 /* We've more or less seen every task there is, call it quits */
6859 if (env
->loop
> env
->loop_max
)
6862 /* take a breather every nr_migrate tasks */
6863 if (env
->loop
> env
->loop_break
) {
6864 env
->loop_break
+= sched_nr_migrate_break
;
6865 env
->flags
|= LBF_NEED_BREAK
;
6869 if (!can_migrate_task(p
, env
))
6872 load
= task_h_load(p
);
6874 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
6877 if ((load
/ 2) > env
->imbalance
)
6880 detach_task(p
, env
);
6881 list_add(&p
->se
.group_node
, &env
->tasks
);
6884 env
->imbalance
-= load
;
6886 #ifdef CONFIG_PREEMPT
6888 * NEWIDLE balancing is a source of latency, so preemptible
6889 * kernels will stop after the first task is detached to minimize
6890 * the critical section.
6892 if (env
->idle
== CPU_NEWLY_IDLE
)
6897 * We only want to steal up to the prescribed amount of
6900 if (env
->imbalance
<= 0)
6905 list_move_tail(&p
->se
.group_node
, tasks
);
6909 * Right now, this is one of only two places we collect this stat
6910 * so we can safely collect detach_one_task() stats here rather
6911 * than inside detach_one_task().
6913 schedstat_add(env
->sd
->lb_gained
[env
->idle
], detached
);
6919 * attach_task() -- attach the task detached by detach_task() to its new rq.
6921 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
6923 lockdep_assert_held(&rq
->lock
);
6925 BUG_ON(task_rq(p
) != rq
);
6926 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
6927 p
->on_rq
= TASK_ON_RQ_QUEUED
;
6928 check_preempt_curr(rq
, p
, 0);
6932 * attach_one_task() -- attaches the task returned from detach_one_task() to
6935 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
6940 update_rq_clock(rq
);
6946 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6949 static void attach_tasks(struct lb_env
*env
)
6951 struct list_head
*tasks
= &env
->tasks
;
6952 struct task_struct
*p
;
6955 rq_lock(env
->dst_rq
, &rf
);
6956 update_rq_clock(env
->dst_rq
);
6958 while (!list_empty(tasks
)) {
6959 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
6960 list_del_init(&p
->se
.group_node
);
6962 attach_task(env
->dst_rq
, p
);
6965 rq_unlock(env
->dst_rq
, &rf
);
6968 #ifdef CONFIG_FAIR_GROUP_SCHED
6970 static inline bool cfs_rq_is_decayed(struct cfs_rq
*cfs_rq
)
6972 if (cfs_rq
->load
.weight
)
6975 if (cfs_rq
->avg
.load_sum
)
6978 if (cfs_rq
->avg
.util_sum
)
6981 if (cfs_rq
->runnable_load_sum
)
6987 static void update_blocked_averages(int cpu
)
6989 struct rq
*rq
= cpu_rq(cpu
);
6990 struct cfs_rq
*cfs_rq
, *pos
;
6993 rq_lock_irqsave(rq
, &rf
);
6994 update_rq_clock(rq
);
6997 * Iterates the task_group tree in a bottom up fashion, see
6998 * list_add_leaf_cfs_rq() for details.
7000 for_each_leaf_cfs_rq_safe(rq
, cfs_rq
, pos
) {
7001 struct sched_entity
*se
;
7003 /* throttled entities do not contribute to load */
7004 if (throttled_hierarchy(cfs_rq
))
7007 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq
), cfs_rq
, true))
7008 update_tg_load_avg(cfs_rq
, 0);
7010 /* Propagate pending load changes to the parent, if any: */
7011 se
= cfs_rq
->tg
->se
[cpu
];
7012 if (se
&& !skip_blocked_update(se
))
7013 update_load_avg(se
, 0);
7016 * There can be a lot of idle CPU cgroups. Don't let fully
7017 * decayed cfs_rqs linger on the list.
7019 if (cfs_rq_is_decayed(cfs_rq
))
7020 list_del_leaf_cfs_rq(cfs_rq
);
7022 rq_unlock_irqrestore(rq
, &rf
);
7026 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7027 * This needs to be done in a top-down fashion because the load of a child
7028 * group is a fraction of its parents load.
7030 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
7032 struct rq
*rq
= rq_of(cfs_rq
);
7033 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
7034 unsigned long now
= jiffies
;
7037 if (cfs_rq
->last_h_load_update
== now
)
7040 cfs_rq
->h_load_next
= NULL
;
7041 for_each_sched_entity(se
) {
7042 cfs_rq
= cfs_rq_of(se
);
7043 cfs_rq
->h_load_next
= se
;
7044 if (cfs_rq
->last_h_load_update
== now
)
7049 cfs_rq
->h_load
= cfs_rq_load_avg(cfs_rq
);
7050 cfs_rq
->last_h_load_update
= now
;
7053 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
7054 load
= cfs_rq
->h_load
;
7055 load
= div64_ul(load
* se
->avg
.load_avg
,
7056 cfs_rq_load_avg(cfs_rq
) + 1);
7057 cfs_rq
= group_cfs_rq(se
);
7058 cfs_rq
->h_load
= load
;
7059 cfs_rq
->last_h_load_update
= now
;
7063 static unsigned long task_h_load(struct task_struct
*p
)
7065 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
7067 update_cfs_rq_h_load(cfs_rq
);
7068 return div64_ul(p
->se
.avg
.load_avg
* cfs_rq
->h_load
,
7069 cfs_rq_load_avg(cfs_rq
) + 1);
7072 static inline void update_blocked_averages(int cpu
)
7074 struct rq
*rq
= cpu_rq(cpu
);
7075 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
7078 rq_lock_irqsave(rq
, &rf
);
7079 update_rq_clock(rq
);
7080 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq
), cfs_rq
, true);
7081 rq_unlock_irqrestore(rq
, &rf
);
7084 static unsigned long task_h_load(struct task_struct
*p
)
7086 return p
->se
.avg
.load_avg
;
7090 /********** Helpers for find_busiest_group ************************/
7099 * sg_lb_stats - stats of a sched_group required for load_balancing
7101 struct sg_lb_stats
{
7102 unsigned long avg_load
; /*Avg load across the CPUs of the group */
7103 unsigned long group_load
; /* Total load over the CPUs of the group */
7104 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
7105 unsigned long load_per_task
;
7106 unsigned long group_capacity
;
7107 unsigned long group_util
; /* Total utilization of the group */
7108 unsigned int sum_nr_running
; /* Nr tasks running in the group */
7109 unsigned int idle_cpus
;
7110 unsigned int group_weight
;
7111 enum group_type group_type
;
7112 int group_no_capacity
;
7113 #ifdef CONFIG_NUMA_BALANCING
7114 unsigned int nr_numa_running
;
7115 unsigned int nr_preferred_running
;
7120 * sd_lb_stats - Structure to store the statistics of a sched_domain
7121 * during load balancing.
7123 struct sd_lb_stats
{
7124 struct sched_group
*busiest
; /* Busiest group in this sd */
7125 struct sched_group
*local
; /* Local group in this sd */
7126 unsigned long total_load
; /* Total load of all groups in sd */
7127 unsigned long total_capacity
; /* Total capacity of all groups in sd */
7128 unsigned long avg_load
; /* Average load across all groups in sd */
7130 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
7131 struct sg_lb_stats local_stat
; /* Statistics of the local group */
7134 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
7137 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7138 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7139 * We must however clear busiest_stat::avg_load because
7140 * update_sd_pick_busiest() reads this before assignment.
7142 *sds
= (struct sd_lb_stats
){
7146 .total_capacity
= 0UL,
7149 .sum_nr_running
= 0,
7150 .group_type
= group_other
,
7156 * get_sd_load_idx - Obtain the load index for a given sched domain.
7157 * @sd: The sched_domain whose load_idx is to be obtained.
7158 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7160 * Return: The load index.
7162 static inline int get_sd_load_idx(struct sched_domain
*sd
,
7163 enum cpu_idle_type idle
)
7169 load_idx
= sd
->busy_idx
;
7172 case CPU_NEWLY_IDLE
:
7173 load_idx
= sd
->newidle_idx
;
7176 load_idx
= sd
->idle_idx
;
7183 static unsigned long scale_rt_capacity(int cpu
)
7185 struct rq
*rq
= cpu_rq(cpu
);
7186 u64 total
, used
, age_stamp
, avg
;
7190 * Since we're reading these variables without serialization make sure
7191 * we read them once before doing sanity checks on them.
7193 age_stamp
= READ_ONCE(rq
->age_stamp
);
7194 avg
= READ_ONCE(rq
->rt_avg
);
7195 delta
= __rq_clock_broken(rq
) - age_stamp
;
7197 if (unlikely(delta
< 0))
7200 total
= sched_avg_period() + delta
;
7202 used
= div_u64(avg
, total
);
7204 if (likely(used
< SCHED_CAPACITY_SCALE
))
7205 return SCHED_CAPACITY_SCALE
- used
;
7210 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
7212 unsigned long capacity
= arch_scale_cpu_capacity(sd
, cpu
);
7213 struct sched_group
*sdg
= sd
->groups
;
7215 cpu_rq(cpu
)->cpu_capacity_orig
= capacity
;
7217 capacity
*= scale_rt_capacity(cpu
);
7218 capacity
>>= SCHED_CAPACITY_SHIFT
;
7223 cpu_rq(cpu
)->cpu_capacity
= capacity
;
7224 sdg
->sgc
->capacity
= capacity
;
7225 sdg
->sgc
->min_capacity
= capacity
;
7228 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
7230 struct sched_domain
*child
= sd
->child
;
7231 struct sched_group
*group
, *sdg
= sd
->groups
;
7232 unsigned long capacity
, min_capacity
;
7233 unsigned long interval
;
7235 interval
= msecs_to_jiffies(sd
->balance_interval
);
7236 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
7237 sdg
->sgc
->next_update
= jiffies
+ interval
;
7240 update_cpu_capacity(sd
, cpu
);
7245 min_capacity
= ULONG_MAX
;
7247 if (child
->flags
& SD_OVERLAP
) {
7249 * SD_OVERLAP domains cannot assume that child groups
7250 * span the current group.
7253 for_each_cpu(cpu
, sched_group_span(sdg
)) {
7254 struct sched_group_capacity
*sgc
;
7255 struct rq
*rq
= cpu_rq(cpu
);
7258 * build_sched_domains() -> init_sched_groups_capacity()
7259 * gets here before we've attached the domains to the
7262 * Use capacity_of(), which is set irrespective of domains
7263 * in update_cpu_capacity().
7265 * This avoids capacity from being 0 and
7266 * causing divide-by-zero issues on boot.
7268 if (unlikely(!rq
->sd
)) {
7269 capacity
+= capacity_of(cpu
);
7271 sgc
= rq
->sd
->groups
->sgc
;
7272 capacity
+= sgc
->capacity
;
7275 min_capacity
= min(capacity
, min_capacity
);
7279 * !SD_OVERLAP domains can assume that child groups
7280 * span the current group.
7283 group
= child
->groups
;
7285 struct sched_group_capacity
*sgc
= group
->sgc
;
7287 capacity
+= sgc
->capacity
;
7288 min_capacity
= min(sgc
->min_capacity
, min_capacity
);
7289 group
= group
->next
;
7290 } while (group
!= child
->groups
);
7293 sdg
->sgc
->capacity
= capacity
;
7294 sdg
->sgc
->min_capacity
= min_capacity
;
7298 * Check whether the capacity of the rq has been noticeably reduced by side
7299 * activity. The imbalance_pct is used for the threshold.
7300 * Return true is the capacity is reduced
7303 check_cpu_capacity(struct rq
*rq
, struct sched_domain
*sd
)
7305 return ((rq
->cpu_capacity
* sd
->imbalance_pct
) <
7306 (rq
->cpu_capacity_orig
* 100));
7310 * Group imbalance indicates (and tries to solve) the problem where balancing
7311 * groups is inadequate due to ->cpus_allowed constraints.
7313 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7314 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7317 * { 0 1 2 3 } { 4 5 6 7 }
7320 * If we were to balance group-wise we'd place two tasks in the first group and
7321 * two tasks in the second group. Clearly this is undesired as it will overload
7322 * cpu 3 and leave one of the cpus in the second group unused.
7324 * The current solution to this issue is detecting the skew in the first group
7325 * by noticing the lower domain failed to reach balance and had difficulty
7326 * moving tasks due to affinity constraints.
7328 * When this is so detected; this group becomes a candidate for busiest; see
7329 * update_sd_pick_busiest(). And calculate_imbalance() and
7330 * find_busiest_group() avoid some of the usual balance conditions to allow it
7331 * to create an effective group imbalance.
7333 * This is a somewhat tricky proposition since the next run might not find the
7334 * group imbalance and decide the groups need to be balanced again. A most
7335 * subtle and fragile situation.
7338 static inline int sg_imbalanced(struct sched_group
*group
)
7340 return group
->sgc
->imbalance
;
7344 * group_has_capacity returns true if the group has spare capacity that could
7345 * be used by some tasks.
7346 * We consider that a group has spare capacity if the * number of task is
7347 * smaller than the number of CPUs or if the utilization is lower than the
7348 * available capacity for CFS tasks.
7349 * For the latter, we use a threshold to stabilize the state, to take into
7350 * account the variance of the tasks' load and to return true if the available
7351 * capacity in meaningful for the load balancer.
7352 * As an example, an available capacity of 1% can appear but it doesn't make
7353 * any benefit for the load balance.
7356 group_has_capacity(struct lb_env
*env
, struct sg_lb_stats
*sgs
)
7358 if (sgs
->sum_nr_running
< sgs
->group_weight
)
7361 if ((sgs
->group_capacity
* 100) >
7362 (sgs
->group_util
* env
->sd
->imbalance_pct
))
7369 * group_is_overloaded returns true if the group has more tasks than it can
7371 * group_is_overloaded is not equals to !group_has_capacity because a group
7372 * with the exact right number of tasks, has no more spare capacity but is not
7373 * overloaded so both group_has_capacity and group_is_overloaded return
7377 group_is_overloaded(struct lb_env
*env
, struct sg_lb_stats
*sgs
)
7379 if (sgs
->sum_nr_running
<= sgs
->group_weight
)
7382 if ((sgs
->group_capacity
* 100) <
7383 (sgs
->group_util
* env
->sd
->imbalance_pct
))
7390 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7391 * per-CPU capacity than sched_group ref.
7394 group_smaller_cpu_capacity(struct sched_group
*sg
, struct sched_group
*ref
)
7396 return sg
->sgc
->min_capacity
* capacity_margin
<
7397 ref
->sgc
->min_capacity
* 1024;
7401 group_type
group_classify(struct sched_group
*group
,
7402 struct sg_lb_stats
*sgs
)
7404 if (sgs
->group_no_capacity
)
7405 return group_overloaded
;
7407 if (sg_imbalanced(group
))
7408 return group_imbalanced
;
7414 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7415 * @env: The load balancing environment.
7416 * @group: sched_group whose statistics are to be updated.
7417 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7418 * @local_group: Does group contain this_cpu.
7419 * @sgs: variable to hold the statistics for this group.
7420 * @overload: Indicate more than one runnable task for any CPU.
7422 static inline void update_sg_lb_stats(struct lb_env
*env
,
7423 struct sched_group
*group
, int load_idx
,
7424 int local_group
, struct sg_lb_stats
*sgs
,
7430 memset(sgs
, 0, sizeof(*sgs
));
7432 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
7433 struct rq
*rq
= cpu_rq(i
);
7435 /* Bias balancing toward cpus of our domain */
7437 load
= target_load(i
, load_idx
);
7439 load
= source_load(i
, load_idx
);
7441 sgs
->group_load
+= load
;
7442 sgs
->group_util
+= cpu_util(i
);
7443 sgs
->sum_nr_running
+= rq
->cfs
.h_nr_running
;
7445 nr_running
= rq
->nr_running
;
7449 #ifdef CONFIG_NUMA_BALANCING
7450 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
7451 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
7453 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
7455 * No need to call idle_cpu() if nr_running is not 0
7457 if (!nr_running
&& idle_cpu(i
))
7461 /* Adjust by relative CPU capacity of the group */
7462 sgs
->group_capacity
= group
->sgc
->capacity
;
7463 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
7465 if (sgs
->sum_nr_running
)
7466 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
7468 sgs
->group_weight
= group
->group_weight
;
7470 sgs
->group_no_capacity
= group_is_overloaded(env
, sgs
);
7471 sgs
->group_type
= group_classify(group
, sgs
);
7475 * update_sd_pick_busiest - return 1 on busiest group
7476 * @env: The load balancing environment.
7477 * @sds: sched_domain statistics
7478 * @sg: sched_group candidate to be checked for being the busiest
7479 * @sgs: sched_group statistics
7481 * Determine if @sg is a busier group than the previously selected
7484 * Return: %true if @sg is a busier group than the previously selected
7485 * busiest group. %false otherwise.
7487 static bool update_sd_pick_busiest(struct lb_env
*env
,
7488 struct sd_lb_stats
*sds
,
7489 struct sched_group
*sg
,
7490 struct sg_lb_stats
*sgs
)
7492 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
7494 if (sgs
->group_type
> busiest
->group_type
)
7497 if (sgs
->group_type
< busiest
->group_type
)
7500 if (sgs
->avg_load
<= busiest
->avg_load
)
7503 if (!(env
->sd
->flags
& SD_ASYM_CPUCAPACITY
))
7507 * Candidate sg has no more than one task per CPU and
7508 * has higher per-CPU capacity. Migrating tasks to less
7509 * capable CPUs may harm throughput. Maximize throughput,
7510 * power/energy consequences are not considered.
7512 if (sgs
->sum_nr_running
<= sgs
->group_weight
&&
7513 group_smaller_cpu_capacity(sds
->local
, sg
))
7517 /* This is the busiest node in its class. */
7518 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
7521 /* No ASYM_PACKING if target cpu is already busy */
7522 if (env
->idle
== CPU_NOT_IDLE
)
7525 * ASYM_PACKING needs to move all the work to the highest
7526 * prority CPUs in the group, therefore mark all groups
7527 * of lower priority than ourself as busy.
7529 if (sgs
->sum_nr_running
&&
7530 sched_asym_prefer(env
->dst_cpu
, sg
->asym_prefer_cpu
)) {
7534 /* Prefer to move from lowest priority cpu's work */
7535 if (sched_asym_prefer(sds
->busiest
->asym_prefer_cpu
,
7536 sg
->asym_prefer_cpu
))
7543 #ifdef CONFIG_NUMA_BALANCING
7544 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
7546 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
7548 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
7553 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
7555 if (rq
->nr_running
> rq
->nr_numa_running
)
7557 if (rq
->nr_running
> rq
->nr_preferred_running
)
7562 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
7567 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
7571 #endif /* CONFIG_NUMA_BALANCING */
7574 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7575 * @env: The load balancing environment.
7576 * @sds: variable to hold the statistics for this sched_domain.
7578 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
7580 struct sched_domain
*child
= env
->sd
->child
;
7581 struct sched_group
*sg
= env
->sd
->groups
;
7582 struct sg_lb_stats
*local
= &sds
->local_stat
;
7583 struct sg_lb_stats tmp_sgs
;
7584 int load_idx
, prefer_sibling
= 0;
7585 bool overload
= false;
7587 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
7590 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
7593 struct sg_lb_stats
*sgs
= &tmp_sgs
;
7596 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_span(sg
));
7601 if (env
->idle
!= CPU_NEWLY_IDLE
||
7602 time_after_eq(jiffies
, sg
->sgc
->next_update
))
7603 update_group_capacity(env
->sd
, env
->dst_cpu
);
7606 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
,
7613 * In case the child domain prefers tasks go to siblings
7614 * first, lower the sg capacity so that we'll try
7615 * and move all the excess tasks away. We lower the capacity
7616 * of a group only if the local group has the capacity to fit
7617 * these excess tasks. The extra check prevents the case where
7618 * you always pull from the heaviest group when it is already
7619 * under-utilized (possible with a large weight task outweighs
7620 * the tasks on the system).
7622 if (prefer_sibling
&& sds
->local
&&
7623 group_has_capacity(env
, local
) &&
7624 (sgs
->sum_nr_running
> local
->sum_nr_running
+ 1)) {
7625 sgs
->group_no_capacity
= 1;
7626 sgs
->group_type
= group_classify(sg
, sgs
);
7629 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
7631 sds
->busiest_stat
= *sgs
;
7635 /* Now, start updating sd_lb_stats */
7636 sds
->total_load
+= sgs
->group_load
;
7637 sds
->total_capacity
+= sgs
->group_capacity
;
7640 } while (sg
!= env
->sd
->groups
);
7642 if (env
->sd
->flags
& SD_NUMA
)
7643 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
7645 if (!env
->sd
->parent
) {
7646 /* update overload indicator if we are at root domain */
7647 if (env
->dst_rq
->rd
->overload
!= overload
)
7648 env
->dst_rq
->rd
->overload
= overload
;
7654 * check_asym_packing - Check to see if the group is packed into the
7657 * This is primarily intended to used at the sibling level. Some
7658 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7659 * case of POWER7, it can move to lower SMT modes only when higher
7660 * threads are idle. When in lower SMT modes, the threads will
7661 * perform better since they share less core resources. Hence when we
7662 * have idle threads, we want them to be the higher ones.
7664 * This packing function is run on idle threads. It checks to see if
7665 * the busiest CPU in this domain (core in the P7 case) has a higher
7666 * CPU number than the packing function is being run on. Here we are
7667 * assuming lower CPU number will be equivalent to lower a SMT thread
7670 * Return: 1 when packing is required and a task should be moved to
7671 * this CPU. The amount of the imbalance is returned in *imbalance.
7673 * @env: The load balancing environment.
7674 * @sds: Statistics of the sched_domain which is to be packed
7676 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
7680 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
7683 if (env
->idle
== CPU_NOT_IDLE
)
7689 busiest_cpu
= sds
->busiest
->asym_prefer_cpu
;
7690 if (sched_asym_prefer(busiest_cpu
, env
->dst_cpu
))
7693 env
->imbalance
= DIV_ROUND_CLOSEST(
7694 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
7695 SCHED_CAPACITY_SCALE
);
7701 * fix_small_imbalance - Calculate the minor imbalance that exists
7702 * amongst the groups of a sched_domain, during
7704 * @env: The load balancing environment.
7705 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7708 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
7710 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
7711 unsigned int imbn
= 2;
7712 unsigned long scaled_busy_load_per_task
;
7713 struct sg_lb_stats
*local
, *busiest
;
7715 local
= &sds
->local_stat
;
7716 busiest
= &sds
->busiest_stat
;
7718 if (!local
->sum_nr_running
)
7719 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
7720 else if (busiest
->load_per_task
> local
->load_per_task
)
7723 scaled_busy_load_per_task
=
7724 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
7725 busiest
->group_capacity
;
7727 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
7728 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
7729 env
->imbalance
= busiest
->load_per_task
;
7734 * OK, we don't have enough imbalance to justify moving tasks,
7735 * however we may be able to increase total CPU capacity used by
7739 capa_now
+= busiest
->group_capacity
*
7740 min(busiest
->load_per_task
, busiest
->avg_load
);
7741 capa_now
+= local
->group_capacity
*
7742 min(local
->load_per_task
, local
->avg_load
);
7743 capa_now
/= SCHED_CAPACITY_SCALE
;
7745 /* Amount of load we'd subtract */
7746 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
7747 capa_move
+= busiest
->group_capacity
*
7748 min(busiest
->load_per_task
,
7749 busiest
->avg_load
- scaled_busy_load_per_task
);
7752 /* Amount of load we'd add */
7753 if (busiest
->avg_load
* busiest
->group_capacity
<
7754 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
7755 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
7756 local
->group_capacity
;
7758 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
7759 local
->group_capacity
;
7761 capa_move
+= local
->group_capacity
*
7762 min(local
->load_per_task
, local
->avg_load
+ tmp
);
7763 capa_move
/= SCHED_CAPACITY_SCALE
;
7765 /* Move if we gain throughput */
7766 if (capa_move
> capa_now
)
7767 env
->imbalance
= busiest
->load_per_task
;
7771 * calculate_imbalance - Calculate the amount of imbalance present within the
7772 * groups of a given sched_domain during load balance.
7773 * @env: load balance environment
7774 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7776 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
7778 unsigned long max_pull
, load_above_capacity
= ~0UL;
7779 struct sg_lb_stats
*local
, *busiest
;
7781 local
= &sds
->local_stat
;
7782 busiest
= &sds
->busiest_stat
;
7784 if (busiest
->group_type
== group_imbalanced
) {
7786 * In the group_imb case we cannot rely on group-wide averages
7787 * to ensure cpu-load equilibrium, look at wider averages. XXX
7789 busiest
->load_per_task
=
7790 min(busiest
->load_per_task
, sds
->avg_load
);
7794 * Avg load of busiest sg can be less and avg load of local sg can
7795 * be greater than avg load across all sgs of sd because avg load
7796 * factors in sg capacity and sgs with smaller group_type are
7797 * skipped when updating the busiest sg:
7799 if (busiest
->avg_load
<= sds
->avg_load
||
7800 local
->avg_load
>= sds
->avg_load
) {
7802 return fix_small_imbalance(env
, sds
);
7806 * If there aren't any idle cpus, avoid creating some.
7808 if (busiest
->group_type
== group_overloaded
&&
7809 local
->group_type
== group_overloaded
) {
7810 load_above_capacity
= busiest
->sum_nr_running
* SCHED_CAPACITY_SCALE
;
7811 if (load_above_capacity
> busiest
->group_capacity
) {
7812 load_above_capacity
-= busiest
->group_capacity
;
7813 load_above_capacity
*= scale_load_down(NICE_0_LOAD
);
7814 load_above_capacity
/= busiest
->group_capacity
;
7816 load_above_capacity
= ~0UL;
7820 * We're trying to get all the cpus to the average_load, so we don't
7821 * want to push ourselves above the average load, nor do we wish to
7822 * reduce the max loaded cpu below the average load. At the same time,
7823 * we also don't want to reduce the group load below the group
7824 * capacity. Thus we look for the minimum possible imbalance.
7826 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
7828 /* How much load to actually move to equalise the imbalance */
7829 env
->imbalance
= min(
7830 max_pull
* busiest
->group_capacity
,
7831 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
7832 ) / SCHED_CAPACITY_SCALE
;
7835 * if *imbalance is less than the average load per runnable task
7836 * there is no guarantee that any tasks will be moved so we'll have
7837 * a think about bumping its value to force at least one task to be
7840 if (env
->imbalance
< busiest
->load_per_task
)
7841 return fix_small_imbalance(env
, sds
);
7844 /******* find_busiest_group() helpers end here *********************/
7847 * find_busiest_group - Returns the busiest group within the sched_domain
7848 * if there is an imbalance.
7850 * Also calculates the amount of weighted load which should be moved
7851 * to restore balance.
7853 * @env: The load balancing environment.
7855 * Return: - The busiest group if imbalance exists.
7857 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
7859 struct sg_lb_stats
*local
, *busiest
;
7860 struct sd_lb_stats sds
;
7862 init_sd_lb_stats(&sds
);
7865 * Compute the various statistics relavent for load balancing at
7868 update_sd_lb_stats(env
, &sds
);
7869 local
= &sds
.local_stat
;
7870 busiest
= &sds
.busiest_stat
;
7872 /* ASYM feature bypasses nice load balance check */
7873 if (check_asym_packing(env
, &sds
))
7876 /* There is no busy sibling group to pull tasks from */
7877 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
7880 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
7881 / sds
.total_capacity
;
7884 * If the busiest group is imbalanced the below checks don't
7885 * work because they assume all things are equal, which typically
7886 * isn't true due to cpus_allowed constraints and the like.
7888 if (busiest
->group_type
== group_imbalanced
)
7891 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7892 if (env
->idle
== CPU_NEWLY_IDLE
&& group_has_capacity(env
, local
) &&
7893 busiest
->group_no_capacity
)
7897 * If the local group is busier than the selected busiest group
7898 * don't try and pull any tasks.
7900 if (local
->avg_load
>= busiest
->avg_load
)
7904 * Don't pull any tasks if this group is already above the domain
7907 if (local
->avg_load
>= sds
.avg_load
)
7910 if (env
->idle
== CPU_IDLE
) {
7912 * This cpu is idle. If the busiest group is not overloaded
7913 * and there is no imbalance between this and busiest group
7914 * wrt idle cpus, it is balanced. The imbalance becomes
7915 * significant if the diff is greater than 1 otherwise we
7916 * might end up to just move the imbalance on another group
7918 if ((busiest
->group_type
!= group_overloaded
) &&
7919 (local
->idle_cpus
<= (busiest
->idle_cpus
+ 1)))
7923 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7924 * imbalance_pct to be conservative.
7926 if (100 * busiest
->avg_load
<=
7927 env
->sd
->imbalance_pct
* local
->avg_load
)
7932 /* Looks like there is an imbalance. Compute it */
7933 calculate_imbalance(env
, &sds
);
7942 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7944 static struct rq
*find_busiest_queue(struct lb_env
*env
,
7945 struct sched_group
*group
)
7947 struct rq
*busiest
= NULL
, *rq
;
7948 unsigned long busiest_load
= 0, busiest_capacity
= 1;
7951 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
7952 unsigned long capacity
, wl
;
7956 rt
= fbq_classify_rq(rq
);
7959 * We classify groups/runqueues into three groups:
7960 * - regular: there are !numa tasks
7961 * - remote: there are numa tasks that run on the 'wrong' node
7962 * - all: there is no distinction
7964 * In order to avoid migrating ideally placed numa tasks,
7965 * ignore those when there's better options.
7967 * If we ignore the actual busiest queue to migrate another
7968 * task, the next balance pass can still reduce the busiest
7969 * queue by moving tasks around inside the node.
7971 * If we cannot move enough load due to this classification
7972 * the next pass will adjust the group classification and
7973 * allow migration of more tasks.
7975 * Both cases only affect the total convergence complexity.
7977 if (rt
> env
->fbq_type
)
7980 capacity
= capacity_of(i
);
7982 wl
= weighted_cpuload(i
);
7985 * When comparing with imbalance, use weighted_cpuload()
7986 * which is not scaled with the cpu capacity.
7989 if (rq
->nr_running
== 1 && wl
> env
->imbalance
&&
7990 !check_cpu_capacity(rq
, env
->sd
))
7994 * For the load comparisons with the other cpu's, consider
7995 * the weighted_cpuload() scaled with the cpu capacity, so
7996 * that the load can be moved away from the cpu that is
7997 * potentially running at a lower capacity.
7999 * Thus we're looking for max(wl_i / capacity_i), crosswise
8000 * multiplication to rid ourselves of the division works out
8001 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8002 * our previous maximum.
8004 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
8006 busiest_capacity
= capacity
;
8015 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8016 * so long as it is large enough.
8018 #define MAX_PINNED_INTERVAL 512
8020 static int need_active_balance(struct lb_env
*env
)
8022 struct sched_domain
*sd
= env
->sd
;
8024 if (env
->idle
== CPU_NEWLY_IDLE
) {
8027 * ASYM_PACKING needs to force migrate tasks from busy but
8028 * lower priority CPUs in order to pack all tasks in the
8029 * highest priority CPUs.
8031 if ((sd
->flags
& SD_ASYM_PACKING
) &&
8032 sched_asym_prefer(env
->dst_cpu
, env
->src_cpu
))
8037 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8038 * It's worth migrating the task if the src_cpu's capacity is reduced
8039 * because of other sched_class or IRQs if more capacity stays
8040 * available on dst_cpu.
8042 if ((env
->idle
!= CPU_NOT_IDLE
) &&
8043 (env
->src_rq
->cfs
.h_nr_running
== 1)) {
8044 if ((check_cpu_capacity(env
->src_rq
, sd
)) &&
8045 (capacity_of(env
->src_cpu
)*sd
->imbalance_pct
< capacity_of(env
->dst_cpu
)*100))
8049 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
8052 static int active_load_balance_cpu_stop(void *data
);
8054 static int should_we_balance(struct lb_env
*env
)
8056 struct sched_group
*sg
= env
->sd
->groups
;
8057 int cpu
, balance_cpu
= -1;
8060 * In the newly idle case, we will allow all the cpu's
8061 * to do the newly idle load balance.
8063 if (env
->idle
== CPU_NEWLY_IDLE
)
8066 /* Try to find first idle cpu */
8067 for_each_cpu_and(cpu
, group_balance_mask(sg
), env
->cpus
) {
8075 if (balance_cpu
== -1)
8076 balance_cpu
= group_balance_cpu(sg
);
8079 * First idle cpu or the first cpu(busiest) in this sched group
8080 * is eligible for doing load balancing at this and above domains.
8082 return balance_cpu
== env
->dst_cpu
;
8086 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8087 * tasks if there is an imbalance.
8089 static int load_balance(int this_cpu
, struct rq
*this_rq
,
8090 struct sched_domain
*sd
, enum cpu_idle_type idle
,
8091 int *continue_balancing
)
8093 int ld_moved
, cur_ld_moved
, active_balance
= 0;
8094 struct sched_domain
*sd_parent
= sd
->parent
;
8095 struct sched_group
*group
;
8098 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
8100 struct lb_env env
= {
8102 .dst_cpu
= this_cpu
,
8104 .dst_grpmask
= sched_group_span(sd
->groups
),
8106 .loop_break
= sched_nr_migrate_break
,
8109 .tasks
= LIST_HEAD_INIT(env
.tasks
),
8113 * For NEWLY_IDLE load_balancing, we don't need to consider
8114 * other cpus in our group
8116 if (idle
== CPU_NEWLY_IDLE
)
8117 env
.dst_grpmask
= NULL
;
8119 cpumask_copy(cpus
, cpu_active_mask
);
8121 schedstat_inc(sd
->lb_count
[idle
]);
8124 if (!should_we_balance(&env
)) {
8125 *continue_balancing
= 0;
8129 group
= find_busiest_group(&env
);
8131 schedstat_inc(sd
->lb_nobusyg
[idle
]);
8135 busiest
= find_busiest_queue(&env
, group
);
8137 schedstat_inc(sd
->lb_nobusyq
[idle
]);
8141 BUG_ON(busiest
== env
.dst_rq
);
8143 schedstat_add(sd
->lb_imbalance
[idle
], env
.imbalance
);
8145 env
.src_cpu
= busiest
->cpu
;
8146 env
.src_rq
= busiest
;
8149 if (busiest
->nr_running
> 1) {
8151 * Attempt to move tasks. If find_busiest_group has found
8152 * an imbalance but busiest->nr_running <= 1, the group is
8153 * still unbalanced. ld_moved simply stays zero, so it is
8154 * correctly treated as an imbalance.
8156 env
.flags
|= LBF_ALL_PINNED
;
8157 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
8160 rq_lock_irqsave(busiest
, &rf
);
8161 update_rq_clock(busiest
);
8164 * cur_ld_moved - load moved in current iteration
8165 * ld_moved - cumulative load moved across iterations
8167 cur_ld_moved
= detach_tasks(&env
);
8170 * We've detached some tasks from busiest_rq. Every
8171 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8172 * unlock busiest->lock, and we are able to be sure
8173 * that nobody can manipulate the tasks in parallel.
8174 * See task_rq_lock() family for the details.
8177 rq_unlock(busiest
, &rf
);
8181 ld_moved
+= cur_ld_moved
;
8184 local_irq_restore(rf
.flags
);
8186 if (env
.flags
& LBF_NEED_BREAK
) {
8187 env
.flags
&= ~LBF_NEED_BREAK
;
8192 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8193 * us and move them to an alternate dst_cpu in our sched_group
8194 * where they can run. The upper limit on how many times we
8195 * iterate on same src_cpu is dependent on number of cpus in our
8198 * This changes load balance semantics a bit on who can move
8199 * load to a given_cpu. In addition to the given_cpu itself
8200 * (or a ilb_cpu acting on its behalf where given_cpu is
8201 * nohz-idle), we now have balance_cpu in a position to move
8202 * load to given_cpu. In rare situations, this may cause
8203 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8204 * _independently_ and at _same_ time to move some load to
8205 * given_cpu) causing exceess load to be moved to given_cpu.
8206 * This however should not happen so much in practice and
8207 * moreover subsequent load balance cycles should correct the
8208 * excess load moved.
8210 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
8212 /* Prevent to re-select dst_cpu via env's cpus */
8213 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
8215 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
8216 env
.dst_cpu
= env
.new_dst_cpu
;
8217 env
.flags
&= ~LBF_DST_PINNED
;
8219 env
.loop_break
= sched_nr_migrate_break
;
8222 * Go back to "more_balance" rather than "redo" since we
8223 * need to continue with same src_cpu.
8229 * We failed to reach balance because of affinity.
8232 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
8234 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
8235 *group_imbalance
= 1;
8238 /* All tasks on this runqueue were pinned by CPU affinity */
8239 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
8240 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
8241 if (!cpumask_empty(cpus
)) {
8243 env
.loop_break
= sched_nr_migrate_break
;
8246 goto out_all_pinned
;
8251 schedstat_inc(sd
->lb_failed
[idle
]);
8253 * Increment the failure counter only on periodic balance.
8254 * We do not want newidle balance, which can be very
8255 * frequent, pollute the failure counter causing
8256 * excessive cache_hot migrations and active balances.
8258 if (idle
!= CPU_NEWLY_IDLE
)
8259 sd
->nr_balance_failed
++;
8261 if (need_active_balance(&env
)) {
8262 unsigned long flags
;
8264 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
8266 /* don't kick the active_load_balance_cpu_stop,
8267 * if the curr task on busiest cpu can't be
8270 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
8271 raw_spin_unlock_irqrestore(&busiest
->lock
,
8273 env
.flags
|= LBF_ALL_PINNED
;
8274 goto out_one_pinned
;
8278 * ->active_balance synchronizes accesses to
8279 * ->active_balance_work. Once set, it's cleared
8280 * only after active load balance is finished.
8282 if (!busiest
->active_balance
) {
8283 busiest
->active_balance
= 1;
8284 busiest
->push_cpu
= this_cpu
;
8287 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
8289 if (active_balance
) {
8290 stop_one_cpu_nowait(cpu_of(busiest
),
8291 active_load_balance_cpu_stop
, busiest
,
8292 &busiest
->active_balance_work
);
8295 /* We've kicked active balancing, force task migration. */
8296 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
8299 sd
->nr_balance_failed
= 0;
8301 if (likely(!active_balance
)) {
8302 /* We were unbalanced, so reset the balancing interval */
8303 sd
->balance_interval
= sd
->min_interval
;
8306 * If we've begun active balancing, start to back off. This
8307 * case may not be covered by the all_pinned logic if there
8308 * is only 1 task on the busy runqueue (because we don't call
8311 if (sd
->balance_interval
< sd
->max_interval
)
8312 sd
->balance_interval
*= 2;
8319 * We reach balance although we may have faced some affinity
8320 * constraints. Clear the imbalance flag if it was set.
8323 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
8325 if (*group_imbalance
)
8326 *group_imbalance
= 0;
8331 * We reach balance because all tasks are pinned at this level so
8332 * we can't migrate them. Let the imbalance flag set so parent level
8333 * can try to migrate them.
8335 schedstat_inc(sd
->lb_balanced
[idle
]);
8337 sd
->nr_balance_failed
= 0;
8340 /* tune up the balancing interval */
8341 if (((env
.flags
& LBF_ALL_PINNED
) &&
8342 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
8343 (sd
->balance_interval
< sd
->max_interval
))
8344 sd
->balance_interval
*= 2;
8351 static inline unsigned long
8352 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
8354 unsigned long interval
= sd
->balance_interval
;
8357 interval
*= sd
->busy_factor
;
8359 /* scale ms to jiffies */
8360 interval
= msecs_to_jiffies(interval
);
8361 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
8367 update_next_balance(struct sched_domain
*sd
, unsigned long *next_balance
)
8369 unsigned long interval
, next
;
8371 /* used by idle balance, so cpu_busy = 0 */
8372 interval
= get_sd_balance_interval(sd
, 0);
8373 next
= sd
->last_balance
+ interval
;
8375 if (time_after(*next_balance
, next
))
8376 *next_balance
= next
;
8380 * idle_balance is called by schedule() if this_cpu is about to become
8381 * idle. Attempts to pull tasks from other CPUs.
8383 static int idle_balance(struct rq
*this_rq
, struct rq_flags
*rf
)
8385 unsigned long next_balance
= jiffies
+ HZ
;
8386 int this_cpu
= this_rq
->cpu
;
8387 struct sched_domain
*sd
;
8388 int pulled_task
= 0;
8392 * We must set idle_stamp _before_ calling idle_balance(), such that we
8393 * measure the duration of idle_balance() as idle time.
8395 this_rq
->idle_stamp
= rq_clock(this_rq
);
8398 * This is OK, because current is on_cpu, which avoids it being picked
8399 * for load-balance and preemption/IRQs are still disabled avoiding
8400 * further scheduler activity on it and we're being very careful to
8401 * re-start the picking loop.
8403 rq_unpin_lock(this_rq
, rf
);
8405 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
8406 !this_rq
->rd
->overload
) {
8408 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
8410 update_next_balance(sd
, &next_balance
);
8416 raw_spin_unlock(&this_rq
->lock
);
8418 update_blocked_averages(this_cpu
);
8420 for_each_domain(this_cpu
, sd
) {
8421 int continue_balancing
= 1;
8422 u64 t0
, domain_cost
;
8424 if (!(sd
->flags
& SD_LOAD_BALANCE
))
8427 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
8428 update_next_balance(sd
, &next_balance
);
8432 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
8433 t0
= sched_clock_cpu(this_cpu
);
8435 pulled_task
= load_balance(this_cpu
, this_rq
,
8437 &continue_balancing
);
8439 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
8440 if (domain_cost
> sd
->max_newidle_lb_cost
)
8441 sd
->max_newidle_lb_cost
= domain_cost
;
8443 curr_cost
+= domain_cost
;
8446 update_next_balance(sd
, &next_balance
);
8449 * Stop searching for tasks to pull if there are
8450 * now runnable tasks on this rq.
8452 if (pulled_task
|| this_rq
->nr_running
> 0)
8457 raw_spin_lock(&this_rq
->lock
);
8459 if (curr_cost
> this_rq
->max_idle_balance_cost
)
8460 this_rq
->max_idle_balance_cost
= curr_cost
;
8463 * While browsing the domains, we released the rq lock, a task could
8464 * have been enqueued in the meantime. Since we're not going idle,
8465 * pretend we pulled a task.
8467 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
8471 /* Move the next balance forward */
8472 if (time_after(this_rq
->next_balance
, next_balance
))
8473 this_rq
->next_balance
= next_balance
;
8475 /* Is there a task of a high priority class? */
8476 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
8480 this_rq
->idle_stamp
= 0;
8482 rq_repin_lock(this_rq
, rf
);
8488 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8489 * running tasks off the busiest CPU onto idle CPUs. It requires at
8490 * least 1 task to be running on each physical CPU where possible, and
8491 * avoids physical / logical imbalances.
8493 static int active_load_balance_cpu_stop(void *data
)
8495 struct rq
*busiest_rq
= data
;
8496 int busiest_cpu
= cpu_of(busiest_rq
);
8497 int target_cpu
= busiest_rq
->push_cpu
;
8498 struct rq
*target_rq
= cpu_rq(target_cpu
);
8499 struct sched_domain
*sd
;
8500 struct task_struct
*p
= NULL
;
8503 rq_lock_irq(busiest_rq
, &rf
);
8505 /* make sure the requested cpu hasn't gone down in the meantime */
8506 if (unlikely(busiest_cpu
!= smp_processor_id() ||
8507 !busiest_rq
->active_balance
))
8510 /* Is there any task to move? */
8511 if (busiest_rq
->nr_running
<= 1)
8515 * This condition is "impossible", if it occurs
8516 * we need to fix it. Originally reported by
8517 * Bjorn Helgaas on a 128-cpu setup.
8519 BUG_ON(busiest_rq
== target_rq
);
8521 /* Search for an sd spanning us and the target CPU. */
8523 for_each_domain(target_cpu
, sd
) {
8524 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
8525 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
8530 struct lb_env env
= {
8532 .dst_cpu
= target_cpu
,
8533 .dst_rq
= target_rq
,
8534 .src_cpu
= busiest_rq
->cpu
,
8535 .src_rq
= busiest_rq
,
8539 schedstat_inc(sd
->alb_count
);
8540 update_rq_clock(busiest_rq
);
8542 p
= detach_one_task(&env
);
8544 schedstat_inc(sd
->alb_pushed
);
8545 /* Active balancing done, reset the failure counter. */
8546 sd
->nr_balance_failed
= 0;
8548 schedstat_inc(sd
->alb_failed
);
8553 busiest_rq
->active_balance
= 0;
8554 rq_unlock(busiest_rq
, &rf
);
8557 attach_one_task(target_rq
, p
);
8564 static inline int on_null_domain(struct rq
*rq
)
8566 return unlikely(!rcu_dereference_sched(rq
->sd
));
8569 #ifdef CONFIG_NO_HZ_COMMON
8571 * idle load balancing details
8572 * - When one of the busy CPUs notice that there may be an idle rebalancing
8573 * needed, they will kick the idle load balancer, which then does idle
8574 * load balancing for all the idle CPUs.
8577 cpumask_var_t idle_cpus_mask
;
8579 unsigned long next_balance
; /* in jiffy units */
8580 } nohz ____cacheline_aligned
;
8582 static inline int find_new_ilb(void)
8584 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
8586 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
8593 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8594 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8595 * CPU (if there is one).
8597 static void nohz_balancer_kick(void)
8601 nohz
.next_balance
++;
8603 ilb_cpu
= find_new_ilb();
8605 if (ilb_cpu
>= nr_cpu_ids
)
8608 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
8611 * Use smp_send_reschedule() instead of resched_cpu().
8612 * This way we generate a sched IPI on the target cpu which
8613 * is idle. And the softirq performing nohz idle load balance
8614 * will be run before returning from the IPI.
8616 smp_send_reschedule(ilb_cpu
);
8620 void nohz_balance_exit_idle(unsigned int cpu
)
8622 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
8624 * Completely isolated CPUs don't ever set, so we must test.
8626 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
8627 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
8628 atomic_dec(&nohz
.nr_cpus
);
8630 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
8634 static inline void set_cpu_sd_state_busy(void)
8636 struct sched_domain
*sd
;
8637 int cpu
= smp_processor_id();
8640 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
8642 if (!sd
|| !sd
->nohz_idle
)
8646 atomic_inc(&sd
->shared
->nr_busy_cpus
);
8651 void set_cpu_sd_state_idle(void)
8653 struct sched_domain
*sd
;
8654 int cpu
= smp_processor_id();
8657 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
8659 if (!sd
|| sd
->nohz_idle
)
8663 atomic_dec(&sd
->shared
->nr_busy_cpus
);
8669 * This routine will record that the cpu is going idle with tick stopped.
8670 * This info will be used in performing idle load balancing in the future.
8672 void nohz_balance_enter_idle(int cpu
)
8675 * If this cpu is going down, then nothing needs to be done.
8677 if (!cpu_active(cpu
))
8680 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
8684 * If we're a completely isolated CPU, we don't play.
8686 if (on_null_domain(cpu_rq(cpu
)))
8689 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
8690 atomic_inc(&nohz
.nr_cpus
);
8691 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
8695 static DEFINE_SPINLOCK(balancing
);
8698 * Scale the max load_balance interval with the number of CPUs in the system.
8699 * This trades load-balance latency on larger machines for less cross talk.
8701 void update_max_interval(void)
8703 max_load_balance_interval
= HZ
*num_online_cpus()/10;
8707 * It checks each scheduling domain to see if it is due to be balanced,
8708 * and initiates a balancing operation if so.
8710 * Balancing parameters are set up in init_sched_domains.
8712 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
8714 int continue_balancing
= 1;
8716 unsigned long interval
;
8717 struct sched_domain
*sd
;
8718 /* Earliest time when we have to do rebalance again */
8719 unsigned long next_balance
= jiffies
+ 60*HZ
;
8720 int update_next_balance
= 0;
8721 int need_serialize
, need_decay
= 0;
8724 update_blocked_averages(cpu
);
8727 for_each_domain(cpu
, sd
) {
8729 * Decay the newidle max times here because this is a regular
8730 * visit to all the domains. Decay ~1% per second.
8732 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
8733 sd
->max_newidle_lb_cost
=
8734 (sd
->max_newidle_lb_cost
* 253) / 256;
8735 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
8738 max_cost
+= sd
->max_newidle_lb_cost
;
8740 if (!(sd
->flags
& SD_LOAD_BALANCE
))
8744 * Stop the load balance at this level. There is another
8745 * CPU in our sched group which is doing load balancing more
8748 if (!continue_balancing
) {
8754 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
8756 need_serialize
= sd
->flags
& SD_SERIALIZE
;
8757 if (need_serialize
) {
8758 if (!spin_trylock(&balancing
))
8762 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
8763 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
8765 * The LBF_DST_PINNED logic could have changed
8766 * env->dst_cpu, so we can't know our idle
8767 * state even if we migrated tasks. Update it.
8769 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
8771 sd
->last_balance
= jiffies
;
8772 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
8775 spin_unlock(&balancing
);
8777 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
8778 next_balance
= sd
->last_balance
+ interval
;
8779 update_next_balance
= 1;
8784 * Ensure the rq-wide value also decays but keep it at a
8785 * reasonable floor to avoid funnies with rq->avg_idle.
8787 rq
->max_idle_balance_cost
=
8788 max((u64
)sysctl_sched_migration_cost
, max_cost
);
8793 * next_balance will be updated only when there is a need.
8794 * When the cpu is attached to null domain for ex, it will not be
8797 if (likely(update_next_balance
)) {
8798 rq
->next_balance
= next_balance
;
8800 #ifdef CONFIG_NO_HZ_COMMON
8802 * If this CPU has been elected to perform the nohz idle
8803 * balance. Other idle CPUs have already rebalanced with
8804 * nohz_idle_balance() and nohz.next_balance has been
8805 * updated accordingly. This CPU is now running the idle load
8806 * balance for itself and we need to update the
8807 * nohz.next_balance accordingly.
8809 if ((idle
== CPU_IDLE
) && time_after(nohz
.next_balance
, rq
->next_balance
))
8810 nohz
.next_balance
= rq
->next_balance
;
8815 #ifdef CONFIG_NO_HZ_COMMON
8817 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8818 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8820 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
8822 int this_cpu
= this_rq
->cpu
;
8825 /* Earliest time when we have to do rebalance again */
8826 unsigned long next_balance
= jiffies
+ 60*HZ
;
8827 int update_next_balance
= 0;
8829 if (idle
!= CPU_IDLE
||
8830 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
8833 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
8834 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
8838 * If this cpu gets work to do, stop the load balancing
8839 * work being done for other cpus. Next load
8840 * balancing owner will pick it up.
8845 rq
= cpu_rq(balance_cpu
);
8848 * If time for next balance is due,
8851 if (time_after_eq(jiffies
, rq
->next_balance
)) {
8854 rq_lock_irq(rq
, &rf
);
8855 update_rq_clock(rq
);
8856 cpu_load_update_idle(rq
);
8857 rq_unlock_irq(rq
, &rf
);
8859 rebalance_domains(rq
, CPU_IDLE
);
8862 if (time_after(next_balance
, rq
->next_balance
)) {
8863 next_balance
= rq
->next_balance
;
8864 update_next_balance
= 1;
8869 * next_balance will be updated only when there is a need.
8870 * When the CPU is attached to null domain for ex, it will not be
8873 if (likely(update_next_balance
))
8874 nohz
.next_balance
= next_balance
;
8876 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
8880 * Current heuristic for kicking the idle load balancer in the presence
8881 * of an idle cpu in the system.
8882 * - This rq has more than one task.
8883 * - This rq has at least one CFS task and the capacity of the CPU is
8884 * significantly reduced because of RT tasks or IRQs.
8885 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8886 * multiple busy cpu.
8887 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8888 * domain span are idle.
8890 static inline bool nohz_kick_needed(struct rq
*rq
)
8892 unsigned long now
= jiffies
;
8893 struct sched_domain_shared
*sds
;
8894 struct sched_domain
*sd
;
8895 int nr_busy
, i
, cpu
= rq
->cpu
;
8898 if (unlikely(rq
->idle_balance
))
8902 * We may be recently in ticked or tickless idle mode. At the first
8903 * busy tick after returning from idle, we will update the busy stats.
8905 set_cpu_sd_state_busy();
8906 nohz_balance_exit_idle(cpu
);
8909 * None are in tickless mode and hence no need for NOHZ idle load
8912 if (likely(!atomic_read(&nohz
.nr_cpus
)))
8915 if (time_before(now
, nohz
.next_balance
))
8918 if (rq
->nr_running
>= 2)
8922 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
8925 * XXX: write a coherent comment on why we do this.
8926 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
8928 nr_busy
= atomic_read(&sds
->nr_busy_cpus
);
8936 sd
= rcu_dereference(rq
->sd
);
8938 if ((rq
->cfs
.h_nr_running
>= 1) &&
8939 check_cpu_capacity(rq
, sd
)) {
8945 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
8947 for_each_cpu(i
, sched_domain_span(sd
)) {
8949 !cpumask_test_cpu(i
, nohz
.idle_cpus_mask
))
8952 if (sched_asym_prefer(i
, cpu
)) {
8963 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
8967 * run_rebalance_domains is triggered when needed from the scheduler tick.
8968 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8970 static __latent_entropy
void run_rebalance_domains(struct softirq_action
*h
)
8972 struct rq
*this_rq
= this_rq();
8973 enum cpu_idle_type idle
= this_rq
->idle_balance
?
8974 CPU_IDLE
: CPU_NOT_IDLE
;
8977 * If this cpu has a pending nohz_balance_kick, then do the
8978 * balancing on behalf of the other idle cpus whose ticks are
8979 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8980 * give the idle cpus a chance to load balance. Else we may
8981 * load balance only within the local sched_domain hierarchy
8982 * and abort nohz_idle_balance altogether if we pull some load.
8984 nohz_idle_balance(this_rq
, idle
);
8985 rebalance_domains(this_rq
, idle
);
8989 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8991 void trigger_load_balance(struct rq
*rq
)
8993 /* Don't need to rebalance while attached to NULL domain */
8994 if (unlikely(on_null_domain(rq
)))
8997 if (time_after_eq(jiffies
, rq
->next_balance
))
8998 raise_softirq(SCHED_SOFTIRQ
);
8999 #ifdef CONFIG_NO_HZ_COMMON
9000 if (nohz_kick_needed(rq
))
9001 nohz_balancer_kick();
9005 static void rq_online_fair(struct rq
*rq
)
9009 update_runtime_enabled(rq
);
9012 static void rq_offline_fair(struct rq
*rq
)
9016 /* Ensure any throttled groups are reachable by pick_next_task */
9017 unthrottle_offline_cfs_rqs(rq
);
9020 #endif /* CONFIG_SMP */
9023 * scheduler tick hitting a task of our scheduling class:
9025 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
9027 struct cfs_rq
*cfs_rq
;
9028 struct sched_entity
*se
= &curr
->se
;
9030 for_each_sched_entity(se
) {
9031 cfs_rq
= cfs_rq_of(se
);
9032 entity_tick(cfs_rq
, se
, queued
);
9035 if (static_branch_unlikely(&sched_numa_balancing
))
9036 task_tick_numa(rq
, curr
);
9040 * called on fork with the child task as argument from the parent's context
9041 * - child not yet on the tasklist
9042 * - preemption disabled
9044 static void task_fork_fair(struct task_struct
*p
)
9046 struct cfs_rq
*cfs_rq
;
9047 struct sched_entity
*se
= &p
->se
, *curr
;
9048 struct rq
*rq
= this_rq();
9052 update_rq_clock(rq
);
9054 cfs_rq
= task_cfs_rq(current
);
9055 curr
= cfs_rq
->curr
;
9057 update_curr(cfs_rq
);
9058 se
->vruntime
= curr
->vruntime
;
9060 place_entity(cfs_rq
, se
, 1);
9062 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
9064 * Upon rescheduling, sched_class::put_prev_task() will place
9065 * 'current' within the tree based on its new key value.
9067 swap(curr
->vruntime
, se
->vruntime
);
9071 se
->vruntime
-= cfs_rq
->min_vruntime
;
9076 * Priority of the task has changed. Check to see if we preempt
9080 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
9082 if (!task_on_rq_queued(p
))
9086 * Reschedule if we are currently running on this runqueue and
9087 * our priority decreased, or if we are not currently running on
9088 * this runqueue and our priority is higher than the current's
9090 if (rq
->curr
== p
) {
9091 if (p
->prio
> oldprio
)
9094 check_preempt_curr(rq
, p
, 0);
9097 static inline bool vruntime_normalized(struct task_struct
*p
)
9099 struct sched_entity
*se
= &p
->se
;
9102 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9103 * the dequeue_entity(.flags=0) will already have normalized the
9110 * When !on_rq, vruntime of the task has usually NOT been normalized.
9111 * But there are some cases where it has already been normalized:
9113 * - A forked child which is waiting for being woken up by
9114 * wake_up_new_task().
9115 * - A task which has been woken up by try_to_wake_up() and
9116 * waiting for actually being woken up by sched_ttwu_pending().
9118 if (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
)
9124 #ifdef CONFIG_FAIR_GROUP_SCHED
9126 * Propagate the changes of the sched_entity across the tg tree to make it
9127 * visible to the root
9129 static void propagate_entity_cfs_rq(struct sched_entity
*se
)
9131 struct cfs_rq
*cfs_rq
;
9133 /* Start to propagate at parent */
9136 for_each_sched_entity(se
) {
9137 cfs_rq
= cfs_rq_of(se
);
9139 if (cfs_rq_throttled(cfs_rq
))
9142 update_load_avg(se
, UPDATE_TG
);
9146 static void propagate_entity_cfs_rq(struct sched_entity
*se
) { }
9149 static void detach_entity_cfs_rq(struct sched_entity
*se
)
9151 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9153 /* Catch up with the cfs_rq and remove our load when we leave */
9154 update_load_avg(se
, 0);
9155 detach_entity_load_avg(cfs_rq
, se
);
9156 update_tg_load_avg(cfs_rq
, false);
9157 propagate_entity_cfs_rq(se
);
9160 static void attach_entity_cfs_rq(struct sched_entity
*se
)
9162 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9164 #ifdef CONFIG_FAIR_GROUP_SCHED
9166 * Since the real-depth could have been changed (only FAIR
9167 * class maintain depth value), reset depth properly.
9169 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
9172 /* Synchronize entity with its cfs_rq */
9173 update_load_avg(se
, sched_feat(ATTACH_AGE_LOAD
) ? 0 : SKIP_AGE_LOAD
);
9174 attach_entity_load_avg(cfs_rq
, se
);
9175 update_tg_load_avg(cfs_rq
, false);
9176 propagate_entity_cfs_rq(se
);
9179 static void detach_task_cfs_rq(struct task_struct
*p
)
9181 struct sched_entity
*se
= &p
->se
;
9182 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9184 if (!vruntime_normalized(p
)) {
9186 * Fix up our vruntime so that the current sleep doesn't
9187 * cause 'unlimited' sleep bonus.
9189 place_entity(cfs_rq
, se
, 0);
9190 se
->vruntime
-= cfs_rq
->min_vruntime
;
9193 detach_entity_cfs_rq(se
);
9196 static void attach_task_cfs_rq(struct task_struct
*p
)
9198 struct sched_entity
*se
= &p
->se
;
9199 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9201 attach_entity_cfs_rq(se
);
9203 if (!vruntime_normalized(p
))
9204 se
->vruntime
+= cfs_rq
->min_vruntime
;
9207 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
9209 detach_task_cfs_rq(p
);
9212 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
9214 attach_task_cfs_rq(p
);
9216 if (task_on_rq_queued(p
)) {
9218 * We were most likely switched from sched_rt, so
9219 * kick off the schedule if running, otherwise just see
9220 * if we can still preempt the current task.
9225 check_preempt_curr(rq
, p
, 0);
9229 /* Account for a task changing its policy or group.
9231 * This routine is mostly called to set cfs_rq->curr field when a task
9232 * migrates between groups/classes.
9234 static void set_curr_task_fair(struct rq
*rq
)
9236 struct sched_entity
*se
= &rq
->curr
->se
;
9238 for_each_sched_entity(se
) {
9239 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9241 set_next_entity(cfs_rq
, se
);
9242 /* ensure bandwidth has been allocated on our new cfs_rq */
9243 account_cfs_rq_runtime(cfs_rq
, 0);
9247 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
9249 cfs_rq
->tasks_timeline
= RB_ROOT
;
9250 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9251 #ifndef CONFIG_64BIT
9252 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
9255 #ifdef CONFIG_FAIR_GROUP_SCHED
9256 cfs_rq
->propagate_avg
= 0;
9258 atomic_long_set(&cfs_rq
->removed_load_avg
, 0);
9259 atomic_long_set(&cfs_rq
->removed_util_avg
, 0);
9263 #ifdef CONFIG_FAIR_GROUP_SCHED
9264 static void task_set_group_fair(struct task_struct
*p
)
9266 struct sched_entity
*se
= &p
->se
;
9268 set_task_rq(p
, task_cpu(p
));
9269 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
9272 static void task_move_group_fair(struct task_struct
*p
)
9274 detach_task_cfs_rq(p
);
9275 set_task_rq(p
, task_cpu(p
));
9278 /* Tell se's cfs_rq has been changed -- migrated */
9279 p
->se
.avg
.last_update_time
= 0;
9281 attach_task_cfs_rq(p
);
9284 static void task_change_group_fair(struct task_struct
*p
, int type
)
9287 case TASK_SET_GROUP
:
9288 task_set_group_fair(p
);
9291 case TASK_MOVE_GROUP
:
9292 task_move_group_fair(p
);
9297 void free_fair_sched_group(struct task_group
*tg
)
9301 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
9303 for_each_possible_cpu(i
) {
9305 kfree(tg
->cfs_rq
[i
]);
9314 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9316 struct sched_entity
*se
;
9317 struct cfs_rq
*cfs_rq
;
9320 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9323 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9327 tg
->shares
= NICE_0_LOAD
;
9329 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
9331 for_each_possible_cpu(i
) {
9332 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9333 GFP_KERNEL
, cpu_to_node(i
));
9337 se
= kzalloc_node(sizeof(struct sched_entity
),
9338 GFP_KERNEL
, cpu_to_node(i
));
9342 init_cfs_rq(cfs_rq
);
9343 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
9344 init_entity_runnable_average(se
);
9355 void online_fair_sched_group(struct task_group
*tg
)
9357 struct sched_entity
*se
;
9361 for_each_possible_cpu(i
) {
9365 raw_spin_lock_irq(&rq
->lock
);
9366 update_rq_clock(rq
);
9367 attach_entity_cfs_rq(se
);
9368 sync_throttle(tg
, i
);
9369 raw_spin_unlock_irq(&rq
->lock
);
9373 void unregister_fair_sched_group(struct task_group
*tg
)
9375 unsigned long flags
;
9379 for_each_possible_cpu(cpu
) {
9381 remove_entity_load_avg(tg
->se
[cpu
]);
9384 * Only empty task groups can be destroyed; so we can speculatively
9385 * check on_list without danger of it being re-added.
9387 if (!tg
->cfs_rq
[cpu
]->on_list
)
9392 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9393 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
9394 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9398 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9399 struct sched_entity
*se
, int cpu
,
9400 struct sched_entity
*parent
)
9402 struct rq
*rq
= cpu_rq(cpu
);
9406 init_cfs_rq_runtime(cfs_rq
);
9408 tg
->cfs_rq
[cpu
] = cfs_rq
;
9411 /* se could be NULL for root_task_group */
9416 se
->cfs_rq
= &rq
->cfs
;
9419 se
->cfs_rq
= parent
->my_q
;
9420 se
->depth
= parent
->depth
+ 1;
9424 /* guarantee group entities always have weight */
9425 update_load_set(&se
->load
, NICE_0_LOAD
);
9426 se
->parent
= parent
;
9429 static DEFINE_MUTEX(shares_mutex
);
9431 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9436 * We can't change the weight of the root cgroup.
9441 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
9443 mutex_lock(&shares_mutex
);
9444 if (tg
->shares
== shares
)
9447 tg
->shares
= shares
;
9448 for_each_possible_cpu(i
) {
9449 struct rq
*rq
= cpu_rq(i
);
9450 struct sched_entity
*se
= tg
->se
[i
];
9453 /* Propagate contribution to hierarchy */
9454 rq_lock_irqsave(rq
, &rf
);
9455 update_rq_clock(rq
);
9456 for_each_sched_entity(se
) {
9457 update_load_avg(se
, UPDATE_TG
);
9458 update_cfs_shares(se
);
9460 rq_unlock_irqrestore(rq
, &rf
);
9464 mutex_unlock(&shares_mutex
);
9467 #else /* CONFIG_FAIR_GROUP_SCHED */
9469 void free_fair_sched_group(struct task_group
*tg
) { }
9471 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9476 void online_fair_sched_group(struct task_group
*tg
) { }
9478 void unregister_fair_sched_group(struct task_group
*tg
) { }
9480 #endif /* CONFIG_FAIR_GROUP_SCHED */
9483 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
9485 struct sched_entity
*se
= &task
->se
;
9486 unsigned int rr_interval
= 0;
9489 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9492 if (rq
->cfs
.load
.weight
)
9493 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
9499 * All the scheduling class methods:
9501 const struct sched_class fair_sched_class
= {
9502 .next
= &idle_sched_class
,
9503 .enqueue_task
= enqueue_task_fair
,
9504 .dequeue_task
= dequeue_task_fair
,
9505 .yield_task
= yield_task_fair
,
9506 .yield_to_task
= yield_to_task_fair
,
9508 .check_preempt_curr
= check_preempt_wakeup
,
9510 .pick_next_task
= pick_next_task_fair
,
9511 .put_prev_task
= put_prev_task_fair
,
9514 .select_task_rq
= select_task_rq_fair
,
9515 .migrate_task_rq
= migrate_task_rq_fair
,
9517 .rq_online
= rq_online_fair
,
9518 .rq_offline
= rq_offline_fair
,
9520 .task_dead
= task_dead_fair
,
9521 .set_cpus_allowed
= set_cpus_allowed_common
,
9524 .set_curr_task
= set_curr_task_fair
,
9525 .task_tick
= task_tick_fair
,
9526 .task_fork
= task_fork_fair
,
9528 .prio_changed
= prio_changed_fair
,
9529 .switched_from
= switched_from_fair
,
9530 .switched_to
= switched_to_fair
,
9532 .get_rr_interval
= get_rr_interval_fair
,
9534 .update_curr
= update_curr_fair
,
9536 #ifdef CONFIG_FAIR_GROUP_SCHED
9537 .task_change_group
= task_change_group_fair
,
9541 #ifdef CONFIG_SCHED_DEBUG
9542 void print_cfs_stats(struct seq_file
*m
, int cpu
)
9544 struct cfs_rq
*cfs_rq
, *pos
;
9547 for_each_leaf_cfs_rq_safe(cpu_rq(cpu
), cfs_rq
, pos
)
9548 print_cfs_rq(m
, cpu
, cfs_rq
);
9552 #ifdef CONFIG_NUMA_BALANCING
9553 void show_numa_stats(struct task_struct
*p
, struct seq_file
*m
)
9556 unsigned long tsf
= 0, tpf
= 0, gsf
= 0, gpf
= 0;
9558 for_each_online_node(node
) {
9559 if (p
->numa_faults
) {
9560 tsf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
9561 tpf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
9563 if (p
->numa_group
) {
9564 gsf
= p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)],
9565 gpf
= p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
9567 print_numa_stats(m
, node
, tsf
, tpf
, gsf
, gpf
);
9570 #endif /* CONFIG_NUMA_BALANCING */
9571 #endif /* CONFIG_SCHED_DEBUG */
9573 __init
void init_sched_fair_class(void)
9576 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9578 #ifdef CONFIG_NO_HZ_COMMON
9579 nohz
.next_balance
= jiffies
;
9580 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);