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(rq, cfs_rq) \
373 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
375 /* Do the two (enqueued) entities belong to the same group ? */
376 static inline struct cfs_rq
*
377 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
379 if (se
->cfs_rq
== pse
->cfs_rq
)
385 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
391 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
393 int se_depth
, pse_depth
;
396 * preemption test can be made between sibling entities who are in the
397 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
398 * both tasks until we find their ancestors who are siblings of common
402 /* First walk up until both entities are at same depth */
403 se_depth
= (*se
)->depth
;
404 pse_depth
= (*pse
)->depth
;
406 while (se_depth
> pse_depth
) {
408 *se
= parent_entity(*se
);
411 while (pse_depth
> se_depth
) {
413 *pse
= parent_entity(*pse
);
416 while (!is_same_group(*se
, *pse
)) {
417 *se
= parent_entity(*se
);
418 *pse
= parent_entity(*pse
);
422 #else /* !CONFIG_FAIR_GROUP_SCHED */
424 static inline struct task_struct
*task_of(struct sched_entity
*se
)
426 return container_of(se
, struct task_struct
, se
);
429 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
431 return container_of(cfs_rq
, struct rq
, cfs
);
434 #define entity_is_task(se) 1
436 #define for_each_sched_entity(se) \
437 for (; se; se = NULL)
439 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
441 return &task_rq(p
)->cfs
;
444 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
446 struct task_struct
*p
= task_of(se
);
447 struct rq
*rq
= task_rq(p
);
452 /* runqueue "owned" by this group */
453 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
458 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
462 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
466 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
467 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
469 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
475 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
479 #endif /* CONFIG_FAIR_GROUP_SCHED */
481 static __always_inline
482 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
484 /**************************************************************
485 * Scheduling class tree data structure manipulation methods:
488 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
490 s64 delta
= (s64
)(vruntime
- max_vruntime
);
492 max_vruntime
= vruntime
;
497 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
499 s64 delta
= (s64
)(vruntime
- min_vruntime
);
501 min_vruntime
= vruntime
;
506 static inline int entity_before(struct sched_entity
*a
,
507 struct sched_entity
*b
)
509 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
512 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
514 struct sched_entity
*curr
= cfs_rq
->curr
;
516 u64 vruntime
= cfs_rq
->min_vruntime
;
520 vruntime
= curr
->vruntime
;
525 if (cfs_rq
->rb_leftmost
) {
526 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
531 vruntime
= se
->vruntime
;
533 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
536 /* ensure we never gain time by being placed backwards. */
537 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
540 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
545 * Enqueue an entity into the rb-tree:
547 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
549 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
550 struct rb_node
*parent
= NULL
;
551 struct sched_entity
*entry
;
555 * Find the right place in the rbtree:
559 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
561 * We dont care about collisions. Nodes with
562 * the same key stay together.
564 if (entity_before(se
, entry
)) {
565 link
= &parent
->rb_left
;
567 link
= &parent
->rb_right
;
573 * Maintain a cache of leftmost tree entries (it is frequently
577 cfs_rq
->rb_leftmost
= &se
->run_node
;
579 rb_link_node(&se
->run_node
, parent
, link
);
580 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
583 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
585 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
586 struct rb_node
*next_node
;
588 next_node
= rb_next(&se
->run_node
);
589 cfs_rq
->rb_leftmost
= next_node
;
592 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
595 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
597 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
602 return rb_entry(left
, struct sched_entity
, run_node
);
605 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
607 struct rb_node
*next
= rb_next(&se
->run_node
);
612 return rb_entry(next
, struct sched_entity
, run_node
);
615 #ifdef CONFIG_SCHED_DEBUG
616 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
618 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
623 return rb_entry(last
, struct sched_entity
, run_node
);
626 /**************************************************************
627 * Scheduling class statistics methods:
630 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
631 void __user
*buffer
, size_t *lenp
,
634 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
635 unsigned int factor
= get_update_sysctl_factor();
640 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
641 sysctl_sched_min_granularity
);
643 #define WRT_SYSCTL(name) \
644 (normalized_sysctl_##name = sysctl_##name / (factor))
645 WRT_SYSCTL(sched_min_granularity
);
646 WRT_SYSCTL(sched_latency
);
647 WRT_SYSCTL(sched_wakeup_granularity
);
657 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
659 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
660 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
666 * The idea is to set a period in which each task runs once.
668 * When there are too many tasks (sched_nr_latency) we have to stretch
669 * this period because otherwise the slices get too small.
671 * p = (nr <= nl) ? l : l*nr/nl
673 static u64
__sched_period(unsigned long nr_running
)
675 if (unlikely(nr_running
> sched_nr_latency
))
676 return nr_running
* sysctl_sched_min_granularity
;
678 return sysctl_sched_latency
;
682 * We calculate the wall-time slice from the period by taking a part
683 * proportional to the weight.
687 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
689 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
691 for_each_sched_entity(se
) {
692 struct load_weight
*load
;
693 struct load_weight lw
;
695 cfs_rq
= cfs_rq_of(se
);
696 load
= &cfs_rq
->load
;
698 if (unlikely(!se
->on_rq
)) {
701 update_load_add(&lw
, se
->load
.weight
);
704 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
710 * We calculate the vruntime slice of a to-be-inserted task.
714 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
716 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
721 #include "sched-pelt.h"
723 static int select_idle_sibling(struct task_struct
*p
, int prev_cpu
, int cpu
);
724 static unsigned long task_h_load(struct task_struct
*p
);
726 /* Give new sched_entity start runnable values to heavy its load in infant time */
727 void init_entity_runnable_average(struct sched_entity
*se
)
729 struct sched_avg
*sa
= &se
->avg
;
731 sa
->last_update_time
= 0;
733 * sched_avg's period_contrib should be strictly less then 1024, so
734 * we give it 1023 to make sure it is almost a period (1024us), and
735 * will definitely be update (after enqueue).
737 sa
->period_contrib
= 1023;
739 * Tasks are intialized with full load to be seen as heavy tasks until
740 * they get a chance to stabilize to their real load level.
741 * Group entities are intialized with zero load to reflect the fact that
742 * nothing has been attached to the task group yet.
744 if (entity_is_task(se
))
745 sa
->load_avg
= scale_load_down(se
->load
.weight
);
746 sa
->load_sum
= sa
->load_avg
* LOAD_AVG_MAX
;
748 * At this point, util_avg won't be used in select_task_rq_fair anyway
752 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
755 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
756 static void attach_entity_cfs_rq(struct sched_entity
*se
);
759 * With new tasks being created, their initial util_avgs are extrapolated
760 * based on the cfs_rq's current util_avg:
762 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
764 * However, in many cases, the above util_avg does not give a desired
765 * value. Moreover, the sum of the util_avgs may be divergent, such
766 * as when the series is a harmonic series.
768 * To solve this problem, we also cap the util_avg of successive tasks to
769 * only 1/2 of the left utilization budget:
771 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
773 * where n denotes the nth task.
775 * For example, a simplest series from the beginning would be like:
777 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
778 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
780 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
781 * if util_avg > util_avg_cap.
783 void post_init_entity_util_avg(struct sched_entity
*se
)
785 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
786 struct sched_avg
*sa
= &se
->avg
;
787 long cap
= (long)(SCHED_CAPACITY_SCALE
- cfs_rq
->avg
.util_avg
) / 2;
790 if (cfs_rq
->avg
.util_avg
!= 0) {
791 sa
->util_avg
= cfs_rq
->avg
.util_avg
* se
->load
.weight
;
792 sa
->util_avg
/= (cfs_rq
->avg
.load_avg
+ 1);
794 if (sa
->util_avg
> cap
)
799 sa
->util_sum
= sa
->util_avg
* LOAD_AVG_MAX
;
802 if (entity_is_task(se
)) {
803 struct task_struct
*p
= task_of(se
);
804 if (p
->sched_class
!= &fair_sched_class
) {
806 * For !fair tasks do:
808 update_cfs_rq_load_avg(now, cfs_rq, false);
809 attach_entity_load_avg(cfs_rq, se);
810 switched_from_fair(rq, p);
812 * such that the next switched_to_fair() has the
815 se
->avg
.last_update_time
= cfs_rq_clock_task(cfs_rq
);
820 attach_entity_cfs_rq(se
);
823 #else /* !CONFIG_SMP */
824 void init_entity_runnable_average(struct sched_entity
*se
)
827 void post_init_entity_util_avg(struct sched_entity
*se
)
830 static void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
)
833 #endif /* CONFIG_SMP */
836 * Update the current task's runtime statistics.
838 static void update_curr(struct cfs_rq
*cfs_rq
)
840 struct sched_entity
*curr
= cfs_rq
->curr
;
841 u64 now
= rq_clock_task(rq_of(cfs_rq
));
847 delta_exec
= now
- curr
->exec_start
;
848 if (unlikely((s64
)delta_exec
<= 0))
851 curr
->exec_start
= now
;
853 schedstat_set(curr
->statistics
.exec_max
,
854 max(delta_exec
, curr
->statistics
.exec_max
));
856 curr
->sum_exec_runtime
+= delta_exec
;
857 schedstat_add(cfs_rq
->exec_clock
, delta_exec
);
859 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
860 update_min_vruntime(cfs_rq
);
862 if (entity_is_task(curr
)) {
863 struct task_struct
*curtask
= task_of(curr
);
865 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
866 cpuacct_charge(curtask
, delta_exec
);
867 account_group_exec_runtime(curtask
, delta_exec
);
870 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
873 static void update_curr_fair(struct rq
*rq
)
875 update_curr(cfs_rq_of(&rq
->curr
->se
));
879 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
881 u64 wait_start
, prev_wait_start
;
883 if (!schedstat_enabled())
886 wait_start
= rq_clock(rq_of(cfs_rq
));
887 prev_wait_start
= schedstat_val(se
->statistics
.wait_start
);
889 if (entity_is_task(se
) && task_on_rq_migrating(task_of(se
)) &&
890 likely(wait_start
> prev_wait_start
))
891 wait_start
-= prev_wait_start
;
893 schedstat_set(se
->statistics
.wait_start
, wait_start
);
897 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
899 struct task_struct
*p
;
902 if (!schedstat_enabled())
905 delta
= rq_clock(rq_of(cfs_rq
)) - schedstat_val(se
->statistics
.wait_start
);
907 if (entity_is_task(se
)) {
909 if (task_on_rq_migrating(p
)) {
911 * Preserve migrating task's wait time so wait_start
912 * time stamp can be adjusted to accumulate wait time
913 * prior to migration.
915 schedstat_set(se
->statistics
.wait_start
, delta
);
918 trace_sched_stat_wait(p
, delta
);
921 schedstat_set(se
->statistics
.wait_max
,
922 max(schedstat_val(se
->statistics
.wait_max
), delta
));
923 schedstat_inc(se
->statistics
.wait_count
);
924 schedstat_add(se
->statistics
.wait_sum
, delta
);
925 schedstat_set(se
->statistics
.wait_start
, 0);
929 update_stats_enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
931 struct task_struct
*tsk
= NULL
;
932 u64 sleep_start
, block_start
;
934 if (!schedstat_enabled())
937 sleep_start
= schedstat_val(se
->statistics
.sleep_start
);
938 block_start
= schedstat_val(se
->statistics
.block_start
);
940 if (entity_is_task(se
))
944 u64 delta
= rq_clock(rq_of(cfs_rq
)) - sleep_start
;
949 if (unlikely(delta
> schedstat_val(se
->statistics
.sleep_max
)))
950 schedstat_set(se
->statistics
.sleep_max
, delta
);
952 schedstat_set(se
->statistics
.sleep_start
, 0);
953 schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
956 account_scheduler_latency(tsk
, delta
>> 10, 1);
957 trace_sched_stat_sleep(tsk
, delta
);
961 u64 delta
= rq_clock(rq_of(cfs_rq
)) - block_start
;
966 if (unlikely(delta
> schedstat_val(se
->statistics
.block_max
)))
967 schedstat_set(se
->statistics
.block_max
, delta
);
969 schedstat_set(se
->statistics
.block_start
, 0);
970 schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
973 if (tsk
->in_iowait
) {
974 schedstat_add(se
->statistics
.iowait_sum
, delta
);
975 schedstat_inc(se
->statistics
.iowait_count
);
976 trace_sched_stat_iowait(tsk
, delta
);
979 trace_sched_stat_blocked(tsk
, delta
);
982 * Blocking time is in units of nanosecs, so shift by
983 * 20 to get a milliseconds-range estimation of the
984 * amount of time that the task spent sleeping:
986 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
987 profile_hits(SLEEP_PROFILING
,
988 (void *)get_wchan(tsk
),
991 account_scheduler_latency(tsk
, delta
>> 10, 0);
997 * Task is being enqueued - update stats:
1000 update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1002 if (!schedstat_enabled())
1006 * Are we enqueueing a waiting task? (for current tasks
1007 * a dequeue/enqueue event is a NOP)
1009 if (se
!= cfs_rq
->curr
)
1010 update_stats_wait_start(cfs_rq
, se
);
1012 if (flags
& ENQUEUE_WAKEUP
)
1013 update_stats_enqueue_sleeper(cfs_rq
, se
);
1017 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1020 if (!schedstat_enabled())
1024 * Mark the end of the wait period if dequeueing a
1027 if (se
!= cfs_rq
->curr
)
1028 update_stats_wait_end(cfs_rq
, se
);
1030 if ((flags
& DEQUEUE_SLEEP
) && entity_is_task(se
)) {
1031 struct task_struct
*tsk
= task_of(se
);
1033 if (tsk
->state
& TASK_INTERRUPTIBLE
)
1034 schedstat_set(se
->statistics
.sleep_start
,
1035 rq_clock(rq_of(cfs_rq
)));
1036 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
1037 schedstat_set(se
->statistics
.block_start
,
1038 rq_clock(rq_of(cfs_rq
)));
1043 * We are picking a new current task - update its stats:
1046 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1049 * We are starting a new run period:
1051 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
1054 /**************************************************
1055 * Scheduling class queueing methods:
1058 #ifdef CONFIG_NUMA_BALANCING
1060 * Approximate time to scan a full NUMA task in ms. The task scan period is
1061 * calculated based on the tasks virtual memory size and
1062 * numa_balancing_scan_size.
1064 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
1065 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
1067 /* Portion of address space to scan in MB */
1068 unsigned int sysctl_numa_balancing_scan_size
= 256;
1070 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1071 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
1073 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
1075 unsigned long rss
= 0;
1076 unsigned long nr_scan_pages
;
1079 * Calculations based on RSS as non-present and empty pages are skipped
1080 * by the PTE scanner and NUMA hinting faults should be trapped based
1083 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
1084 rss
= get_mm_rss(p
->mm
);
1086 rss
= nr_scan_pages
;
1088 rss
= round_up(rss
, nr_scan_pages
);
1089 return rss
/ nr_scan_pages
;
1092 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1093 #define MAX_SCAN_WINDOW 2560
1095 static unsigned int task_scan_min(struct task_struct
*p
)
1097 unsigned int scan_size
= READ_ONCE(sysctl_numa_balancing_scan_size
);
1098 unsigned int scan
, floor
;
1099 unsigned int windows
= 1;
1101 if (scan_size
< MAX_SCAN_WINDOW
)
1102 windows
= MAX_SCAN_WINDOW
/ scan_size
;
1103 floor
= 1000 / windows
;
1105 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
1106 return max_t(unsigned int, floor
, scan
);
1109 static unsigned int task_scan_max(struct task_struct
*p
)
1111 unsigned int smin
= task_scan_min(p
);
1114 /* Watch for min being lower than max due to floor calculations */
1115 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
1116 return max(smin
, smax
);
1119 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
1121 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
1122 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
1125 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
1127 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
1128 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
1134 spinlock_t lock
; /* nr_tasks, tasks */
1139 struct rcu_head rcu
;
1140 unsigned long total_faults
;
1141 unsigned long max_faults_cpu
;
1143 * Faults_cpu is used to decide whether memory should move
1144 * towards the CPU. As a consequence, these stats are weighted
1145 * more by CPU use than by memory faults.
1147 unsigned long *faults_cpu
;
1148 unsigned long faults
[0];
1151 /* Shared or private faults. */
1152 #define NR_NUMA_HINT_FAULT_TYPES 2
1154 /* Memory and CPU locality */
1155 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1157 /* Averaged statistics, and temporary buffers. */
1158 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1160 pid_t
task_numa_group_id(struct task_struct
*p
)
1162 return p
->numa_group
? p
->numa_group
->gid
: 0;
1166 * The averaged statistics, shared & private, memory & cpu,
1167 * occupy the first half of the array. The second half of the
1168 * array is for current counters, which are averaged into the
1169 * first set by task_numa_placement.
1171 static inline int task_faults_idx(enum numa_faults_stats s
, int nid
, int priv
)
1173 return NR_NUMA_HINT_FAULT_TYPES
* (s
* nr_node_ids
+ nid
) + priv
;
1176 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
1178 if (!p
->numa_faults
)
1181 return p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1182 p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1185 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
1190 return p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1191 p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1194 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
1196 return group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1197 group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1201 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1202 * considered part of a numa group's pseudo-interleaving set. Migrations
1203 * between these nodes are slowed down, to allow things to settle down.
1205 #define ACTIVE_NODE_FRACTION 3
1207 static bool numa_is_active_node(int nid
, struct numa_group
*ng
)
1209 return group_faults_cpu(ng
, nid
) * ACTIVE_NODE_FRACTION
> ng
->max_faults_cpu
;
1212 /* Handle placement on systems where not all nodes are directly connected. */
1213 static unsigned long score_nearby_nodes(struct task_struct
*p
, int nid
,
1214 int maxdist
, bool task
)
1216 unsigned long score
= 0;
1220 * All nodes are directly connected, and the same distance
1221 * from each other. No need for fancy placement algorithms.
1223 if (sched_numa_topology_type
== NUMA_DIRECT
)
1227 * This code is called for each node, introducing N^2 complexity,
1228 * which should be ok given the number of nodes rarely exceeds 8.
1230 for_each_online_node(node
) {
1231 unsigned long faults
;
1232 int dist
= node_distance(nid
, node
);
1235 * The furthest away nodes in the system are not interesting
1236 * for placement; nid was already counted.
1238 if (dist
== sched_max_numa_distance
|| node
== nid
)
1242 * On systems with a backplane NUMA topology, compare groups
1243 * of nodes, and move tasks towards the group with the most
1244 * memory accesses. When comparing two nodes at distance
1245 * "hoplimit", only nodes closer by than "hoplimit" are part
1246 * of each group. Skip other nodes.
1248 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1252 /* Add up the faults from nearby nodes. */
1254 faults
= task_faults(p
, node
);
1256 faults
= group_faults(p
, node
);
1259 * On systems with a glueless mesh NUMA topology, there are
1260 * no fixed "groups of nodes". Instead, nodes that are not
1261 * directly connected bounce traffic through intermediate
1262 * nodes; a numa_group can occupy any set of nodes.
1263 * The further away a node is, the less the faults count.
1264 * This seems to result in good task placement.
1266 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1267 faults
*= (sched_max_numa_distance
- dist
);
1268 faults
/= (sched_max_numa_distance
- LOCAL_DISTANCE
);
1278 * These return the fraction of accesses done by a particular task, or
1279 * task group, on a particular numa node. The group weight is given a
1280 * larger multiplier, in order to group tasks together that are almost
1281 * evenly spread out between numa nodes.
1283 static inline unsigned long task_weight(struct task_struct
*p
, int nid
,
1286 unsigned long faults
, total_faults
;
1288 if (!p
->numa_faults
)
1291 total_faults
= p
->total_numa_faults
;
1296 faults
= task_faults(p
, nid
);
1297 faults
+= score_nearby_nodes(p
, nid
, dist
, true);
1299 return 1000 * faults
/ total_faults
;
1302 static inline unsigned long group_weight(struct task_struct
*p
, int nid
,
1305 unsigned long faults
, total_faults
;
1310 total_faults
= p
->numa_group
->total_faults
;
1315 faults
= group_faults(p
, nid
);
1316 faults
+= score_nearby_nodes(p
, nid
, dist
, false);
1318 return 1000 * faults
/ total_faults
;
1321 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
1322 int src_nid
, int dst_cpu
)
1324 struct numa_group
*ng
= p
->numa_group
;
1325 int dst_nid
= cpu_to_node(dst_cpu
);
1326 int last_cpupid
, this_cpupid
;
1328 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
1331 * Multi-stage node selection is used in conjunction with a periodic
1332 * migration fault to build a temporal task<->page relation. By using
1333 * a two-stage filter we remove short/unlikely relations.
1335 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1336 * a task's usage of a particular page (n_p) per total usage of this
1337 * page (n_t) (in a given time-span) to a probability.
1339 * Our periodic faults will sample this probability and getting the
1340 * same result twice in a row, given these samples are fully
1341 * independent, is then given by P(n)^2, provided our sample period
1342 * is sufficiently short compared to the usage pattern.
1344 * This quadric squishes small probabilities, making it less likely we
1345 * act on an unlikely task<->page relation.
1347 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
1348 if (!cpupid_pid_unset(last_cpupid
) &&
1349 cpupid_to_nid(last_cpupid
) != dst_nid
)
1352 /* Always allow migrate on private faults */
1353 if (cpupid_match_pid(p
, last_cpupid
))
1356 /* A shared fault, but p->numa_group has not been set up yet. */
1361 * Destination node is much more heavily used than the source
1362 * node? Allow migration.
1364 if (group_faults_cpu(ng
, dst_nid
) > group_faults_cpu(ng
, src_nid
) *
1365 ACTIVE_NODE_FRACTION
)
1369 * Distribute memory according to CPU & memory use on each node,
1370 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1372 * faults_cpu(dst) 3 faults_cpu(src)
1373 * --------------- * - > ---------------
1374 * faults_mem(dst) 4 faults_mem(src)
1376 return group_faults_cpu(ng
, dst_nid
) * group_faults(p
, src_nid
) * 3 >
1377 group_faults_cpu(ng
, src_nid
) * group_faults(p
, dst_nid
) * 4;
1380 static unsigned long weighted_cpuload(const int cpu
);
1381 static unsigned long source_load(int cpu
, int type
);
1382 static unsigned long target_load(int cpu
, int type
);
1383 static unsigned long capacity_of(int cpu
);
1384 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
1386 /* Cached statistics for all CPUs within a node */
1388 unsigned long nr_running
;
1391 /* Total compute capacity of CPUs on a node */
1392 unsigned long compute_capacity
;
1394 /* Approximate capacity in terms of runnable tasks on a node */
1395 unsigned long task_capacity
;
1396 int has_free_capacity
;
1400 * XXX borrowed from update_sg_lb_stats
1402 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1404 int smt
, cpu
, cpus
= 0;
1405 unsigned long capacity
;
1407 memset(ns
, 0, sizeof(*ns
));
1408 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1409 struct rq
*rq
= cpu_rq(cpu
);
1411 ns
->nr_running
+= rq
->nr_running
;
1412 ns
->load
+= weighted_cpuload(cpu
);
1413 ns
->compute_capacity
+= capacity_of(cpu
);
1419 * If we raced with hotplug and there are no CPUs left in our mask
1420 * the @ns structure is NULL'ed and task_numa_compare() will
1421 * not find this node attractive.
1423 * We'll either bail at !has_free_capacity, or we'll detect a huge
1424 * imbalance and bail there.
1429 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1430 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, ns
->compute_capacity
);
1431 capacity
= cpus
/ smt
; /* cores */
1433 ns
->task_capacity
= min_t(unsigned, capacity
,
1434 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
));
1435 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1438 struct task_numa_env
{
1439 struct task_struct
*p
;
1441 int src_cpu
, src_nid
;
1442 int dst_cpu
, dst_nid
;
1444 struct numa_stats src_stats
, dst_stats
;
1449 struct task_struct
*best_task
;
1454 static void task_numa_assign(struct task_numa_env
*env
,
1455 struct task_struct
*p
, long imp
)
1458 put_task_struct(env
->best_task
);
1463 env
->best_imp
= imp
;
1464 env
->best_cpu
= env
->dst_cpu
;
1467 static bool load_too_imbalanced(long src_load
, long dst_load
,
1468 struct task_numa_env
*env
)
1471 long orig_src_load
, orig_dst_load
;
1472 long src_capacity
, dst_capacity
;
1475 * The load is corrected for the CPU capacity available on each node.
1478 * ------------ vs ---------
1479 * src_capacity dst_capacity
1481 src_capacity
= env
->src_stats
.compute_capacity
;
1482 dst_capacity
= env
->dst_stats
.compute_capacity
;
1484 /* We care about the slope of the imbalance, not the direction. */
1485 if (dst_load
< src_load
)
1486 swap(dst_load
, src_load
);
1488 /* Is the difference below the threshold? */
1489 imb
= dst_load
* src_capacity
* 100 -
1490 src_load
* dst_capacity
* env
->imbalance_pct
;
1495 * The imbalance is above the allowed threshold.
1496 * Compare it with the old imbalance.
1498 orig_src_load
= env
->src_stats
.load
;
1499 orig_dst_load
= env
->dst_stats
.load
;
1501 if (orig_dst_load
< orig_src_load
)
1502 swap(orig_dst_load
, orig_src_load
);
1504 old_imb
= orig_dst_load
* src_capacity
* 100 -
1505 orig_src_load
* dst_capacity
* env
->imbalance_pct
;
1507 /* Would this change make things worse? */
1508 return (imb
> old_imb
);
1512 * This checks if the overall compute and NUMA accesses of the system would
1513 * be improved if the source tasks was migrated to the target dst_cpu taking
1514 * into account that it might be best if task running on the dst_cpu should
1515 * be exchanged with the source task
1517 static void task_numa_compare(struct task_numa_env
*env
,
1518 long taskimp
, long groupimp
)
1520 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1521 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1522 struct task_struct
*cur
;
1523 long src_load
, dst_load
;
1525 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1527 int dist
= env
->dist
;
1530 cur
= task_rcu_dereference(&dst_rq
->curr
);
1531 if (cur
&& ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
)))
1535 * Because we have preemption enabled we can get migrated around and
1536 * end try selecting ourselves (current == env->p) as a swap candidate.
1542 * "imp" is the fault differential for the source task between the
1543 * source and destination node. Calculate the total differential for
1544 * the source task and potential destination task. The more negative
1545 * the value is, the more rmeote accesses that would be expected to
1546 * be incurred if the tasks were swapped.
1549 /* Skip this swap candidate if cannot move to the source cpu */
1550 if (!cpumask_test_cpu(env
->src_cpu
, &cur
->cpus_allowed
))
1554 * If dst and source tasks are in the same NUMA group, or not
1555 * in any group then look only at task weights.
1557 if (cur
->numa_group
== env
->p
->numa_group
) {
1558 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1559 task_weight(cur
, env
->dst_nid
, dist
);
1561 * Add some hysteresis to prevent swapping the
1562 * tasks within a group over tiny differences.
1564 if (cur
->numa_group
)
1568 * Compare the group weights. If a task is all by
1569 * itself (not part of a group), use the task weight
1572 if (cur
->numa_group
)
1573 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1574 group_weight(cur
, env
->dst_nid
, dist
);
1576 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1577 task_weight(cur
, env
->dst_nid
, dist
);
1581 if (imp
<= env
->best_imp
&& moveimp
<= env
->best_imp
)
1585 /* Is there capacity at our destination? */
1586 if (env
->src_stats
.nr_running
<= env
->src_stats
.task_capacity
&&
1587 !env
->dst_stats
.has_free_capacity
)
1593 /* Balance doesn't matter much if we're running a task per cpu */
1594 if (imp
> env
->best_imp
&& src_rq
->nr_running
== 1 &&
1595 dst_rq
->nr_running
== 1)
1599 * In the overloaded case, try and keep the load balanced.
1602 load
= task_h_load(env
->p
);
1603 dst_load
= env
->dst_stats
.load
+ load
;
1604 src_load
= env
->src_stats
.load
- load
;
1606 if (moveimp
> imp
&& moveimp
> env
->best_imp
) {
1608 * If the improvement from just moving env->p direction is
1609 * better than swapping tasks around, check if a move is
1610 * possible. Store a slightly smaller score than moveimp,
1611 * so an actually idle CPU will win.
1613 if (!load_too_imbalanced(src_load
, dst_load
, env
)) {
1620 if (imp
<= env
->best_imp
)
1624 load
= task_h_load(cur
);
1629 if (load_too_imbalanced(src_load
, dst_load
, env
))
1633 * One idle CPU per node is evaluated for a task numa move.
1634 * Call select_idle_sibling to maybe find a better one.
1638 * select_idle_siblings() uses an per-cpu cpumask that
1639 * can be used from IRQ context.
1641 local_irq_disable();
1642 env
->dst_cpu
= select_idle_sibling(env
->p
, env
->src_cpu
,
1648 task_numa_assign(env
, cur
, imp
);
1653 static void task_numa_find_cpu(struct task_numa_env
*env
,
1654 long taskimp
, long groupimp
)
1658 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1659 /* Skip this CPU if the source task cannot migrate */
1660 if (!cpumask_test_cpu(cpu
, &env
->p
->cpus_allowed
))
1664 task_numa_compare(env
, taskimp
, groupimp
);
1668 /* Only move tasks to a NUMA node less busy than the current node. */
1669 static bool numa_has_capacity(struct task_numa_env
*env
)
1671 struct numa_stats
*src
= &env
->src_stats
;
1672 struct numa_stats
*dst
= &env
->dst_stats
;
1674 if (src
->has_free_capacity
&& !dst
->has_free_capacity
)
1678 * Only consider a task move if the source has a higher load
1679 * than the destination, corrected for CPU capacity on each node.
1681 * src->load dst->load
1682 * --------------------- vs ---------------------
1683 * src->compute_capacity dst->compute_capacity
1685 if (src
->load
* dst
->compute_capacity
* env
->imbalance_pct
>
1687 dst
->load
* src
->compute_capacity
* 100)
1693 static int task_numa_migrate(struct task_struct
*p
)
1695 struct task_numa_env env
= {
1698 .src_cpu
= task_cpu(p
),
1699 .src_nid
= task_node(p
),
1701 .imbalance_pct
= 112,
1707 struct sched_domain
*sd
;
1708 unsigned long taskweight
, groupweight
;
1710 long taskimp
, groupimp
;
1713 * Pick the lowest SD_NUMA domain, as that would have the smallest
1714 * imbalance and would be the first to start moving tasks about.
1716 * And we want to avoid any moving of tasks about, as that would create
1717 * random movement of tasks -- counter the numa conditions we're trying
1721 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1723 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1727 * Cpusets can break the scheduler domain tree into smaller
1728 * balance domains, some of which do not cross NUMA boundaries.
1729 * Tasks that are "trapped" in such domains cannot be migrated
1730 * elsewhere, so there is no point in (re)trying.
1732 if (unlikely(!sd
)) {
1733 p
->numa_preferred_nid
= task_node(p
);
1737 env
.dst_nid
= p
->numa_preferred_nid
;
1738 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1739 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1740 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1741 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1742 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
1743 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
1744 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1746 /* Try to find a spot on the preferred nid. */
1747 if (numa_has_capacity(&env
))
1748 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1751 * Look at other nodes in these cases:
1752 * - there is no space available on the preferred_nid
1753 * - the task is part of a numa_group that is interleaved across
1754 * multiple NUMA nodes; in order to better consolidate the group,
1755 * we need to check other locations.
1757 if (env
.best_cpu
== -1 || (p
->numa_group
&& p
->numa_group
->active_nodes
> 1)) {
1758 for_each_online_node(nid
) {
1759 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1762 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1763 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1765 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1766 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1769 /* Only consider nodes where both task and groups benefit */
1770 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
1771 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
1772 if (taskimp
< 0 && groupimp
< 0)
1777 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1778 if (numa_has_capacity(&env
))
1779 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1784 * If the task is part of a workload that spans multiple NUMA nodes,
1785 * and is migrating into one of the workload's active nodes, remember
1786 * this node as the task's preferred numa node, so the workload can
1788 * A task that migrated to a second choice node will be better off
1789 * trying for a better one later. Do not set the preferred node here.
1791 if (p
->numa_group
) {
1792 struct numa_group
*ng
= p
->numa_group
;
1794 if (env
.best_cpu
== -1)
1799 if (ng
->active_nodes
> 1 && numa_is_active_node(env
.dst_nid
, ng
))
1800 sched_setnuma(p
, env
.dst_nid
);
1803 /* No better CPU than the current one was found. */
1804 if (env
.best_cpu
== -1)
1808 * Reset the scan period if the task is being rescheduled on an
1809 * alternative node to recheck if the tasks is now properly placed.
1811 p
->numa_scan_period
= task_scan_min(p
);
1813 if (env
.best_task
== NULL
) {
1814 ret
= migrate_task_to(p
, env
.best_cpu
);
1816 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1820 ret
= migrate_swap(p
, env
.best_task
);
1822 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1823 put_task_struct(env
.best_task
);
1827 /* Attempt to migrate a task to a CPU on the preferred node. */
1828 static void numa_migrate_preferred(struct task_struct
*p
)
1830 unsigned long interval
= HZ
;
1832 /* This task has no NUMA fault statistics yet */
1833 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults
))
1836 /* Periodically retry migrating the task to the preferred node */
1837 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1838 p
->numa_migrate_retry
= jiffies
+ interval
;
1840 /* Success if task is already running on preferred CPU */
1841 if (task_node(p
) == p
->numa_preferred_nid
)
1844 /* Otherwise, try migrate to a CPU on the preferred node */
1845 task_numa_migrate(p
);
1849 * Find out how many nodes on the workload is actively running on. Do this by
1850 * tracking the nodes from which NUMA hinting faults are triggered. This can
1851 * be different from the set of nodes where the workload's memory is currently
1854 static void numa_group_count_active_nodes(struct numa_group
*numa_group
)
1856 unsigned long faults
, max_faults
= 0;
1857 int nid
, active_nodes
= 0;
1859 for_each_online_node(nid
) {
1860 faults
= group_faults_cpu(numa_group
, nid
);
1861 if (faults
> max_faults
)
1862 max_faults
= faults
;
1865 for_each_online_node(nid
) {
1866 faults
= group_faults_cpu(numa_group
, nid
);
1867 if (faults
* ACTIVE_NODE_FRACTION
> max_faults
)
1871 numa_group
->max_faults_cpu
= max_faults
;
1872 numa_group
->active_nodes
= active_nodes
;
1876 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1877 * increments. The more local the fault statistics are, the higher the scan
1878 * period will be for the next scan window. If local/(local+remote) ratio is
1879 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1880 * the scan period will decrease. Aim for 70% local accesses.
1882 #define NUMA_PERIOD_SLOTS 10
1883 #define NUMA_PERIOD_THRESHOLD 7
1886 * Increase the scan period (slow down scanning) if the majority of
1887 * our memory is already on our local node, or if the majority of
1888 * the page accesses are shared with other processes.
1889 * Otherwise, decrease the scan period.
1891 static void update_task_scan_period(struct task_struct
*p
,
1892 unsigned long shared
, unsigned long private)
1894 unsigned int period_slot
;
1898 unsigned long remote
= p
->numa_faults_locality
[0];
1899 unsigned long local
= p
->numa_faults_locality
[1];
1902 * If there were no record hinting faults then either the task is
1903 * completely idle or all activity is areas that are not of interest
1904 * to automatic numa balancing. Related to that, if there were failed
1905 * migration then it implies we are migrating too quickly or the local
1906 * node is overloaded. In either case, scan slower
1908 if (local
+ shared
== 0 || p
->numa_faults_locality
[2]) {
1909 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1910 p
->numa_scan_period
<< 1);
1912 p
->mm
->numa_next_scan
= jiffies
+
1913 msecs_to_jiffies(p
->numa_scan_period
);
1919 * Prepare to scale scan period relative to the current period.
1920 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1921 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1922 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1924 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1925 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1926 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1927 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1930 diff
= slot
* period_slot
;
1932 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1935 * Scale scan rate increases based on sharing. There is an
1936 * inverse relationship between the degree of sharing and
1937 * the adjustment made to the scanning period. Broadly
1938 * speaking the intent is that there is little point
1939 * scanning faster if shared accesses dominate as it may
1940 * simply bounce migrations uselessly
1942 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
+ 1));
1943 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1946 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1947 task_scan_min(p
), task_scan_max(p
));
1948 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1952 * Get the fraction of time the task has been running since the last
1953 * NUMA placement cycle. The scheduler keeps similar statistics, but
1954 * decays those on a 32ms period, which is orders of magnitude off
1955 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1956 * stats only if the task is so new there are no NUMA statistics yet.
1958 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1960 u64 runtime
, delta
, now
;
1961 /* Use the start of this time slice to avoid calculations. */
1962 now
= p
->se
.exec_start
;
1963 runtime
= p
->se
.sum_exec_runtime
;
1965 if (p
->last_task_numa_placement
) {
1966 delta
= runtime
- p
->last_sum_exec_runtime
;
1967 *period
= now
- p
->last_task_numa_placement
;
1969 delta
= p
->se
.avg
.load_sum
/ p
->se
.load
.weight
;
1970 *period
= LOAD_AVG_MAX
;
1973 p
->last_sum_exec_runtime
= runtime
;
1974 p
->last_task_numa_placement
= now
;
1980 * Determine the preferred nid for a task in a numa_group. This needs to
1981 * be done in a way that produces consistent results with group_weight,
1982 * otherwise workloads might not converge.
1984 static int preferred_group_nid(struct task_struct
*p
, int nid
)
1989 /* Direct connections between all NUMA nodes. */
1990 if (sched_numa_topology_type
== NUMA_DIRECT
)
1994 * On a system with glueless mesh NUMA topology, group_weight
1995 * scores nodes according to the number of NUMA hinting faults on
1996 * both the node itself, and on nearby nodes.
1998 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1999 unsigned long score
, max_score
= 0;
2000 int node
, max_node
= nid
;
2002 dist
= sched_max_numa_distance
;
2004 for_each_online_node(node
) {
2005 score
= group_weight(p
, node
, dist
);
2006 if (score
> max_score
) {
2015 * Finding the preferred nid in a system with NUMA backplane
2016 * interconnect topology is more involved. The goal is to locate
2017 * tasks from numa_groups near each other in the system, and
2018 * untangle workloads from different sides of the system. This requires
2019 * searching down the hierarchy of node groups, recursively searching
2020 * inside the highest scoring group of nodes. The nodemask tricks
2021 * keep the complexity of the search down.
2023 nodes
= node_online_map
;
2024 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
2025 unsigned long max_faults
= 0;
2026 nodemask_t max_group
= NODE_MASK_NONE
;
2029 /* Are there nodes at this distance from each other? */
2030 if (!find_numa_distance(dist
))
2033 for_each_node_mask(a
, nodes
) {
2034 unsigned long faults
= 0;
2035 nodemask_t this_group
;
2036 nodes_clear(this_group
);
2038 /* Sum group's NUMA faults; includes a==b case. */
2039 for_each_node_mask(b
, nodes
) {
2040 if (node_distance(a
, b
) < dist
) {
2041 faults
+= group_faults(p
, b
);
2042 node_set(b
, this_group
);
2043 node_clear(b
, nodes
);
2047 /* Remember the top group. */
2048 if (faults
> max_faults
) {
2049 max_faults
= faults
;
2050 max_group
= this_group
;
2052 * subtle: at the smallest distance there is
2053 * just one node left in each "group", the
2054 * winner is the preferred nid.
2059 /* Next round, evaluate the nodes within max_group. */
2067 static void task_numa_placement(struct task_struct
*p
)
2069 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
2070 unsigned long max_faults
= 0, max_group_faults
= 0;
2071 unsigned long fault_types
[2] = { 0, 0 };
2072 unsigned long total_faults
;
2073 u64 runtime
, period
;
2074 spinlock_t
*group_lock
= NULL
;
2077 * The p->mm->numa_scan_seq field gets updated without
2078 * exclusive access. Use READ_ONCE() here to ensure
2079 * that the field is read in a single access:
2081 seq
= READ_ONCE(p
->mm
->numa_scan_seq
);
2082 if (p
->numa_scan_seq
== seq
)
2084 p
->numa_scan_seq
= seq
;
2085 p
->numa_scan_period_max
= task_scan_max(p
);
2087 total_faults
= p
->numa_faults_locality
[0] +
2088 p
->numa_faults_locality
[1];
2089 runtime
= numa_get_avg_runtime(p
, &period
);
2091 /* If the task is part of a group prevent parallel updates to group stats */
2092 if (p
->numa_group
) {
2093 group_lock
= &p
->numa_group
->lock
;
2094 spin_lock_irq(group_lock
);
2097 /* Find the node with the highest number of faults */
2098 for_each_online_node(nid
) {
2099 /* Keep track of the offsets in numa_faults array */
2100 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
2101 unsigned long faults
= 0, group_faults
= 0;
2104 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
2105 long diff
, f_diff
, f_weight
;
2107 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
2108 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
2109 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
2110 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
2112 /* Decay existing window, copy faults since last scan */
2113 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
2114 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
2115 p
->numa_faults
[membuf_idx
] = 0;
2118 * Normalize the faults_from, so all tasks in a group
2119 * count according to CPU use, instead of by the raw
2120 * number of faults. Tasks with little runtime have
2121 * little over-all impact on throughput, and thus their
2122 * faults are less important.
2124 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
2125 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
2127 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
2128 p
->numa_faults
[cpubuf_idx
] = 0;
2130 p
->numa_faults
[mem_idx
] += diff
;
2131 p
->numa_faults
[cpu_idx
] += f_diff
;
2132 faults
+= p
->numa_faults
[mem_idx
];
2133 p
->total_numa_faults
+= diff
;
2134 if (p
->numa_group
) {
2136 * safe because we can only change our own group
2138 * mem_idx represents the offset for a given
2139 * nid and priv in a specific region because it
2140 * is at the beginning of the numa_faults array.
2142 p
->numa_group
->faults
[mem_idx
] += diff
;
2143 p
->numa_group
->faults_cpu
[mem_idx
] += f_diff
;
2144 p
->numa_group
->total_faults
+= diff
;
2145 group_faults
+= p
->numa_group
->faults
[mem_idx
];
2149 if (faults
> max_faults
) {
2150 max_faults
= faults
;
2154 if (group_faults
> max_group_faults
) {
2155 max_group_faults
= group_faults
;
2156 max_group_nid
= nid
;
2160 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
2162 if (p
->numa_group
) {
2163 numa_group_count_active_nodes(p
->numa_group
);
2164 spin_unlock_irq(group_lock
);
2165 max_nid
= preferred_group_nid(p
, max_group_nid
);
2169 /* Set the new preferred node */
2170 if (max_nid
!= p
->numa_preferred_nid
)
2171 sched_setnuma(p
, max_nid
);
2173 if (task_node(p
) != p
->numa_preferred_nid
)
2174 numa_migrate_preferred(p
);
2178 static inline int get_numa_group(struct numa_group
*grp
)
2180 return atomic_inc_not_zero(&grp
->refcount
);
2183 static inline void put_numa_group(struct numa_group
*grp
)
2185 if (atomic_dec_and_test(&grp
->refcount
))
2186 kfree_rcu(grp
, rcu
);
2189 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
2192 struct numa_group
*grp
, *my_grp
;
2193 struct task_struct
*tsk
;
2195 int cpu
= cpupid_to_cpu(cpupid
);
2198 if (unlikely(!p
->numa_group
)) {
2199 unsigned int size
= sizeof(struct numa_group
) +
2200 4*nr_node_ids
*sizeof(unsigned long);
2202 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
2206 atomic_set(&grp
->refcount
, 1);
2207 grp
->active_nodes
= 1;
2208 grp
->max_faults_cpu
= 0;
2209 spin_lock_init(&grp
->lock
);
2211 /* Second half of the array tracks nids where faults happen */
2212 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
2215 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2216 grp
->faults
[i
] = p
->numa_faults
[i
];
2218 grp
->total_faults
= p
->total_numa_faults
;
2221 rcu_assign_pointer(p
->numa_group
, grp
);
2225 tsk
= READ_ONCE(cpu_rq(cpu
)->curr
);
2227 if (!cpupid_match_pid(tsk
, cpupid
))
2230 grp
= rcu_dereference(tsk
->numa_group
);
2234 my_grp
= p
->numa_group
;
2239 * Only join the other group if its bigger; if we're the bigger group,
2240 * the other task will join us.
2242 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
2246 * Tie-break on the grp address.
2248 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
2251 /* Always join threads in the same process. */
2252 if (tsk
->mm
== current
->mm
)
2255 /* Simple filter to avoid false positives due to PID collisions */
2256 if (flags
& TNF_SHARED
)
2259 /* Update priv based on whether false sharing was detected */
2262 if (join
&& !get_numa_group(grp
))
2270 BUG_ON(irqs_disabled());
2271 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
2273 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
2274 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
2275 grp
->faults
[i
] += p
->numa_faults
[i
];
2277 my_grp
->total_faults
-= p
->total_numa_faults
;
2278 grp
->total_faults
+= p
->total_numa_faults
;
2283 spin_unlock(&my_grp
->lock
);
2284 spin_unlock_irq(&grp
->lock
);
2286 rcu_assign_pointer(p
->numa_group
, grp
);
2288 put_numa_group(my_grp
);
2296 void task_numa_free(struct task_struct
*p
)
2298 struct numa_group
*grp
= p
->numa_group
;
2299 void *numa_faults
= p
->numa_faults
;
2300 unsigned long flags
;
2304 spin_lock_irqsave(&grp
->lock
, flags
);
2305 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2306 grp
->faults
[i
] -= p
->numa_faults
[i
];
2307 grp
->total_faults
-= p
->total_numa_faults
;
2310 spin_unlock_irqrestore(&grp
->lock
, flags
);
2311 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2312 put_numa_group(grp
);
2315 p
->numa_faults
= NULL
;
2320 * Got a PROT_NONE fault for a page on @node.
2322 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2324 struct task_struct
*p
= current
;
2325 bool migrated
= flags
& TNF_MIGRATED
;
2326 int cpu_node
= task_node(current
);
2327 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2328 struct numa_group
*ng
;
2331 if (!static_branch_likely(&sched_numa_balancing
))
2334 /* for example, ksmd faulting in a user's mm */
2338 /* Allocate buffer to track faults on a per-node basis */
2339 if (unlikely(!p
->numa_faults
)) {
2340 int size
= sizeof(*p
->numa_faults
) *
2341 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2343 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2344 if (!p
->numa_faults
)
2347 p
->total_numa_faults
= 0;
2348 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2352 * First accesses are treated as private, otherwise consider accesses
2353 * to be private if the accessing pid has not changed
2355 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2358 priv
= cpupid_match_pid(p
, last_cpupid
);
2359 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2360 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2364 * If a workload spans multiple NUMA nodes, a shared fault that
2365 * occurs wholly within the set of nodes that the workload is
2366 * actively using should be counted as local. This allows the
2367 * scan rate to slow down when a workload has settled down.
2370 if (!priv
&& !local
&& ng
&& ng
->active_nodes
> 1 &&
2371 numa_is_active_node(cpu_node
, ng
) &&
2372 numa_is_active_node(mem_node
, ng
))
2375 task_numa_placement(p
);
2378 * Retry task to preferred node migration periodically, in case it
2379 * case it previously failed, or the scheduler moved us.
2381 if (time_after(jiffies
, p
->numa_migrate_retry
))
2382 numa_migrate_preferred(p
);
2385 p
->numa_pages_migrated
+= pages
;
2386 if (flags
& TNF_MIGRATE_FAIL
)
2387 p
->numa_faults_locality
[2] += pages
;
2389 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2390 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2391 p
->numa_faults_locality
[local
] += pages
;
2394 static void reset_ptenuma_scan(struct task_struct
*p
)
2397 * We only did a read acquisition of the mmap sem, so
2398 * p->mm->numa_scan_seq is written to without exclusive access
2399 * and the update is not guaranteed to be atomic. That's not
2400 * much of an issue though, since this is just used for
2401 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2402 * expensive, to avoid any form of compiler optimizations:
2404 WRITE_ONCE(p
->mm
->numa_scan_seq
, READ_ONCE(p
->mm
->numa_scan_seq
) + 1);
2405 p
->mm
->numa_scan_offset
= 0;
2409 * The expensive part of numa migration is done from task_work context.
2410 * Triggered from task_tick_numa().
2412 void task_numa_work(struct callback_head
*work
)
2414 unsigned long migrate
, next_scan
, now
= jiffies
;
2415 struct task_struct
*p
= current
;
2416 struct mm_struct
*mm
= p
->mm
;
2417 u64 runtime
= p
->se
.sum_exec_runtime
;
2418 struct vm_area_struct
*vma
;
2419 unsigned long start
, end
;
2420 unsigned long nr_pte_updates
= 0;
2421 long pages
, virtpages
;
2423 SCHED_WARN_ON(p
!= container_of(work
, struct task_struct
, numa_work
));
2425 work
->next
= work
; /* protect against double add */
2427 * Who cares about NUMA placement when they're dying.
2429 * NOTE: make sure not to dereference p->mm before this check,
2430 * exit_task_work() happens _after_ exit_mm() so we could be called
2431 * without p->mm even though we still had it when we enqueued this
2434 if (p
->flags
& PF_EXITING
)
2437 if (!mm
->numa_next_scan
) {
2438 mm
->numa_next_scan
= now
+
2439 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2443 * Enforce maximal scan/migration frequency..
2445 migrate
= mm
->numa_next_scan
;
2446 if (time_before(now
, migrate
))
2449 if (p
->numa_scan_period
== 0) {
2450 p
->numa_scan_period_max
= task_scan_max(p
);
2451 p
->numa_scan_period
= task_scan_min(p
);
2454 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2455 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2459 * Delay this task enough that another task of this mm will likely win
2460 * the next time around.
2462 p
->node_stamp
+= 2 * TICK_NSEC
;
2464 start
= mm
->numa_scan_offset
;
2465 pages
= sysctl_numa_balancing_scan_size
;
2466 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2467 virtpages
= pages
* 8; /* Scan up to this much virtual space */
2472 down_read(&mm
->mmap_sem
);
2473 vma
= find_vma(mm
, start
);
2475 reset_ptenuma_scan(p
);
2479 for (; vma
; vma
= vma
->vm_next
) {
2480 if (!vma_migratable(vma
) || !vma_policy_mof(vma
) ||
2481 is_vm_hugetlb_page(vma
) || (vma
->vm_flags
& VM_MIXEDMAP
)) {
2486 * Shared library pages mapped by multiple processes are not
2487 * migrated as it is expected they are cache replicated. Avoid
2488 * hinting faults in read-only file-backed mappings or the vdso
2489 * as migrating the pages will be of marginal benefit.
2492 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2496 * Skip inaccessible VMAs to avoid any confusion between
2497 * PROT_NONE and NUMA hinting ptes
2499 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
2503 start
= max(start
, vma
->vm_start
);
2504 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2505 end
= min(end
, vma
->vm_end
);
2506 nr_pte_updates
= change_prot_numa(vma
, start
, end
);
2509 * Try to scan sysctl_numa_balancing_size worth of
2510 * hpages that have at least one present PTE that
2511 * is not already pte-numa. If the VMA contains
2512 * areas that are unused or already full of prot_numa
2513 * PTEs, scan up to virtpages, to skip through those
2517 pages
-= (end
- start
) >> PAGE_SHIFT
;
2518 virtpages
-= (end
- start
) >> PAGE_SHIFT
;
2521 if (pages
<= 0 || virtpages
<= 0)
2525 } while (end
!= vma
->vm_end
);
2530 * It is possible to reach the end of the VMA list but the last few
2531 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2532 * would find the !migratable VMA on the next scan but not reset the
2533 * scanner to the start so check it now.
2536 mm
->numa_scan_offset
= start
;
2538 reset_ptenuma_scan(p
);
2539 up_read(&mm
->mmap_sem
);
2542 * Make sure tasks use at least 32x as much time to run other code
2543 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2544 * Usually update_task_scan_period slows down scanning enough; on an
2545 * overloaded system we need to limit overhead on a per task basis.
2547 if (unlikely(p
->se
.sum_exec_runtime
!= runtime
)) {
2548 u64 diff
= p
->se
.sum_exec_runtime
- runtime
;
2549 p
->node_stamp
+= 32 * diff
;
2554 * Drive the periodic memory faults..
2556 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2558 struct callback_head
*work
= &curr
->numa_work
;
2562 * We don't care about NUMA placement if we don't have memory.
2564 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2568 * Using runtime rather than walltime has the dual advantage that
2569 * we (mostly) drive the selection from busy threads and that the
2570 * task needs to have done some actual work before we bother with
2573 now
= curr
->se
.sum_exec_runtime
;
2574 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2576 if (now
> curr
->node_stamp
+ period
) {
2577 if (!curr
->node_stamp
)
2578 curr
->numa_scan_period
= task_scan_min(curr
);
2579 curr
->node_stamp
+= period
;
2581 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2582 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2583 task_work_add(curr
, work
, true);
2588 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2592 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2596 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2599 #endif /* CONFIG_NUMA_BALANCING */
2602 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2604 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2605 if (!parent_entity(se
))
2606 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2608 if (entity_is_task(se
)) {
2609 struct rq
*rq
= rq_of(cfs_rq
);
2611 account_numa_enqueue(rq
, task_of(se
));
2612 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2615 cfs_rq
->nr_running
++;
2619 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2621 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2622 if (!parent_entity(se
))
2623 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2625 if (entity_is_task(se
)) {
2626 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2627 list_del_init(&se
->group_node
);
2630 cfs_rq
->nr_running
--;
2633 #ifdef CONFIG_FAIR_GROUP_SCHED
2635 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2637 long tg_weight
, load
, shares
;
2640 * This really should be: cfs_rq->avg.load_avg, but instead we use
2641 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2642 * the shares for small weight interactive tasks.
2644 load
= scale_load_down(cfs_rq
->load
.weight
);
2646 tg_weight
= atomic_long_read(&tg
->load_avg
);
2648 /* Ensure tg_weight >= load */
2649 tg_weight
-= cfs_rq
->tg_load_avg_contrib
;
2652 shares
= (tg
->shares
* load
);
2654 shares
/= tg_weight
;
2657 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2658 * of a group with small tg->shares value. It is a floor value which is
2659 * assigned as a minimum load.weight to the sched_entity representing
2660 * the group on a CPU.
2662 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2663 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2664 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2665 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2668 if (shares
< MIN_SHARES
)
2669 shares
= MIN_SHARES
;
2670 if (shares
> tg
->shares
)
2671 shares
= tg
->shares
;
2675 # else /* CONFIG_SMP */
2676 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2680 # endif /* CONFIG_SMP */
2682 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2683 unsigned long weight
)
2686 /* commit outstanding execution time */
2687 if (cfs_rq
->curr
== se
)
2688 update_curr(cfs_rq
);
2689 account_entity_dequeue(cfs_rq
, se
);
2692 update_load_set(&se
->load
, weight
);
2695 account_entity_enqueue(cfs_rq
, se
);
2698 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2700 static void update_cfs_shares(struct sched_entity
*se
)
2702 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2703 struct task_group
*tg
;
2709 if (throttled_hierarchy(cfs_rq
))
2715 if (likely(se
->load
.weight
== tg
->shares
))
2718 shares
= calc_cfs_shares(cfs_rq
, tg
);
2720 reweight_entity(cfs_rq_of(se
), se
, shares
);
2723 #else /* CONFIG_FAIR_GROUP_SCHED */
2724 static inline void update_cfs_shares(struct sched_entity
*se
)
2727 #endif /* CONFIG_FAIR_GROUP_SCHED */
2732 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2734 static u64
decay_load(u64 val
, u64 n
)
2736 unsigned int local_n
;
2738 if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2741 /* after bounds checking we can collapse to 32-bit */
2745 * As y^PERIOD = 1/2, we can combine
2746 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2747 * With a look-up table which covers y^n (n<PERIOD)
2749 * To achieve constant time decay_load.
2751 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2752 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2753 local_n
%= LOAD_AVG_PERIOD
;
2756 val
= mul_u64_u32_shr(val
, runnable_avg_yN_inv
[local_n
], 32);
2760 static u32
__accumulate_pelt_segments(u64 periods
, u32 d1
, u32 d3
)
2762 u32 c1
, c2
, c3
= d3
; /* y^0 == 1 */
2767 c1
= decay_load((u64
)d1
, periods
);
2771 * c2 = 1024 \Sum y^n
2775 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
2778 c2
= LOAD_AVG_MAX
- decay_load(LOAD_AVG_MAX
, periods
) - 1024;
2780 return c1
+ c2
+ c3
;
2783 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2786 * Accumulate the three separate parts of the sum; d1 the remainder
2787 * of the last (incomplete) period, d2 the span of full periods and d3
2788 * the remainder of the (incomplete) current period.
2793 * |<->|<----------------->|<--->|
2794 * ... |---x---|------| ... |------|-----x (now)
2797 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
2800 * = u y^p + (Step 1)
2803 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
2806 static __always_inline u32
2807 accumulate_sum(u64 delta
, int cpu
, struct sched_avg
*sa
,
2808 unsigned long weight
, int running
, struct cfs_rq
*cfs_rq
)
2810 unsigned long scale_freq
, scale_cpu
;
2811 u32 contrib
= (u32
)delta
; /* p == 0 -> delta < 1024 */
2814 scale_freq
= arch_scale_freq_capacity(NULL
, cpu
);
2815 scale_cpu
= arch_scale_cpu_capacity(NULL
, cpu
);
2817 delta
+= sa
->period_contrib
;
2818 periods
= delta
/ 1024; /* A period is 1024us (~1ms) */
2821 * Step 1: decay old *_sum if we crossed period boundaries.
2824 sa
->load_sum
= decay_load(sa
->load_sum
, periods
);
2826 cfs_rq
->runnable_load_sum
=
2827 decay_load(cfs_rq
->runnable_load_sum
, periods
);
2829 sa
->util_sum
= decay_load((u64
)(sa
->util_sum
), periods
);
2835 contrib
= __accumulate_pelt_segments(periods
,
2836 1024 - sa
->period_contrib
, delta
);
2838 sa
->period_contrib
= delta
;
2840 contrib
= cap_scale(contrib
, scale_freq
);
2842 sa
->load_sum
+= weight
* contrib
;
2844 cfs_rq
->runnable_load_sum
+= weight
* contrib
;
2847 sa
->util_sum
+= contrib
* scale_cpu
;
2853 * We can represent the historical contribution to runnable average as the
2854 * coefficients of a geometric series. To do this we sub-divide our runnable
2855 * history into segments of approximately 1ms (1024us); label the segment that
2856 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2858 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2860 * (now) (~1ms ago) (~2ms ago)
2862 * Let u_i denote the fraction of p_i that the entity was runnable.
2864 * We then designate the fractions u_i as our co-efficients, yielding the
2865 * following representation of historical load:
2866 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2868 * We choose y based on the with of a reasonably scheduling period, fixing:
2871 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2872 * approximately half as much as the contribution to load within the last ms
2875 * When a period "rolls over" and we have new u_0`, multiplying the previous
2876 * sum again by y is sufficient to update:
2877 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2878 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2880 static __always_inline
int
2881 ___update_load_avg(u64 now
, int cpu
, struct sched_avg
*sa
,
2882 unsigned long weight
, int running
, struct cfs_rq
*cfs_rq
)
2886 delta
= now
- sa
->last_update_time
;
2888 * This should only happen when time goes backwards, which it
2889 * unfortunately does during sched clock init when we swap over to TSC.
2891 if ((s64
)delta
< 0) {
2892 sa
->last_update_time
= now
;
2897 * Use 1024ns as the unit of measurement since it's a reasonable
2898 * approximation of 1us and fast to compute.
2904 sa
->last_update_time
+= delta
<< 10;
2907 * Now we know we crossed measurement unit boundaries. The *_avg
2908 * accrues by two steps:
2910 * Step 1: accumulate *_sum since last_update_time. If we haven't
2911 * crossed period boundaries, finish.
2913 if (!accumulate_sum(delta
, cpu
, sa
, weight
, running
, cfs_rq
))
2917 * Step 2: update *_avg.
2919 sa
->load_avg
= div_u64(sa
->load_sum
, LOAD_AVG_MAX
);
2921 cfs_rq
->runnable_load_avg
=
2922 div_u64(cfs_rq
->runnable_load_sum
, LOAD_AVG_MAX
);
2924 sa
->util_avg
= sa
->util_sum
/ LOAD_AVG_MAX
;
2930 __update_load_avg_blocked_se(u64 now
, int cpu
, struct sched_entity
*se
)
2932 return ___update_load_avg(now
, cpu
, &se
->avg
, 0, 0, NULL
);
2936 __update_load_avg_se(u64 now
, int cpu
, struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2938 return ___update_load_avg(now
, cpu
, &se
->avg
,
2939 se
->on_rq
* scale_load_down(se
->load
.weight
),
2940 cfs_rq
->curr
== se
, NULL
);
2944 __update_load_avg_cfs_rq(u64 now
, int cpu
, struct cfs_rq
*cfs_rq
)
2946 return ___update_load_avg(now
, cpu
, &cfs_rq
->avg
,
2947 scale_load_down(cfs_rq
->load
.weight
),
2948 cfs_rq
->curr
!= NULL
, cfs_rq
);
2952 * Signed add and clamp on underflow.
2954 * Explicitly do a load-store to ensure the intermediate value never hits
2955 * memory. This allows lockless observations without ever seeing the negative
2958 #define add_positive(_ptr, _val) do { \
2959 typeof(_ptr) ptr = (_ptr); \
2960 typeof(_val) val = (_val); \
2961 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2965 if (val < 0 && res > var) \
2968 WRITE_ONCE(*ptr, res); \
2971 #ifdef CONFIG_FAIR_GROUP_SCHED
2973 * update_tg_load_avg - update the tg's load avg
2974 * @cfs_rq: the cfs_rq whose avg changed
2975 * @force: update regardless of how small the difference
2977 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
2978 * However, because tg->load_avg is a global value there are performance
2981 * In order to avoid having to look at the other cfs_rq's, we use a
2982 * differential update where we store the last value we propagated. This in
2983 * turn allows skipping updates if the differential is 'small'.
2985 * Updating tg's load_avg is necessary before update_cfs_share() (which is
2986 * done) and effective_load() (which is not done because it is too costly).
2988 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
)
2990 long delta
= cfs_rq
->avg
.load_avg
- cfs_rq
->tg_load_avg_contrib
;
2993 * No need to update load_avg for root_task_group as it is not used.
2995 if (cfs_rq
->tg
== &root_task_group
)
2998 if (force
|| abs(delta
) > cfs_rq
->tg_load_avg_contrib
/ 64) {
2999 atomic_long_add(delta
, &cfs_rq
->tg
->load_avg
);
3000 cfs_rq
->tg_load_avg_contrib
= cfs_rq
->avg
.load_avg
;
3005 * Called within set_task_rq() right before setting a task's cpu. The
3006 * caller only guarantees p->pi_lock is held; no other assumptions,
3007 * including the state of rq->lock, should be made.
3009 void set_task_rq_fair(struct sched_entity
*se
,
3010 struct cfs_rq
*prev
, struct cfs_rq
*next
)
3012 u64 p_last_update_time
;
3013 u64 n_last_update_time
;
3015 if (!sched_feat(ATTACH_AGE_LOAD
))
3019 * We are supposed to update the task to "current" time, then its up to
3020 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3021 * getting what current time is, so simply throw away the out-of-date
3022 * time. This will result in the wakee task is less decayed, but giving
3023 * the wakee more load sounds not bad.
3025 if (!(se
->avg
.last_update_time
&& prev
))
3028 #ifndef CONFIG_64BIT
3030 u64 p_last_update_time_copy
;
3031 u64 n_last_update_time_copy
;
3034 p_last_update_time_copy
= prev
->load_last_update_time_copy
;
3035 n_last_update_time_copy
= next
->load_last_update_time_copy
;
3039 p_last_update_time
= prev
->avg
.last_update_time
;
3040 n_last_update_time
= next
->avg
.last_update_time
;
3042 } while (p_last_update_time
!= p_last_update_time_copy
||
3043 n_last_update_time
!= n_last_update_time_copy
);
3046 p_last_update_time
= prev
->avg
.last_update_time
;
3047 n_last_update_time
= next
->avg
.last_update_time
;
3049 __update_load_avg_blocked_se(p_last_update_time
, cpu_of(rq_of(prev
)), se
);
3050 se
->avg
.last_update_time
= n_last_update_time
;
3053 /* Take into account change of utilization of a child task group */
3055 update_tg_cfs_util(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3057 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3058 long delta
= gcfs_rq
->avg
.util_avg
- se
->avg
.util_avg
;
3060 /* Nothing to update */
3064 /* Set new sched_entity's utilization */
3065 se
->avg
.util_avg
= gcfs_rq
->avg
.util_avg
;
3066 se
->avg
.util_sum
= se
->avg
.util_avg
* LOAD_AVG_MAX
;
3068 /* Update parent cfs_rq utilization */
3069 add_positive(&cfs_rq
->avg
.util_avg
, delta
);
3070 cfs_rq
->avg
.util_sum
= cfs_rq
->avg
.util_avg
* LOAD_AVG_MAX
;
3073 /* Take into account change of load of a child task group */
3075 update_tg_cfs_load(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3077 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3078 long delta
, load
= gcfs_rq
->avg
.load_avg
;
3081 * If the load of group cfs_rq is null, the load of the
3082 * sched_entity will also be null so we can skip the formula
3087 /* Get tg's load and ensure tg_load > 0 */
3088 tg_load
= atomic_long_read(&gcfs_rq
->tg
->load_avg
) + 1;
3090 /* Ensure tg_load >= load and updated with current load*/
3091 tg_load
-= gcfs_rq
->tg_load_avg_contrib
;
3095 * We need to compute a correction term in the case that the
3096 * task group is consuming more CPU than a task of equal
3097 * weight. A task with a weight equals to tg->shares will have
3098 * a load less or equal to scale_load_down(tg->shares).
3099 * Similarly, the sched_entities that represent the task group
3100 * at parent level, can't have a load higher than
3101 * scale_load_down(tg->shares). And the Sum of sched_entities'
3102 * load must be <= scale_load_down(tg->shares).
3104 if (tg_load
> scale_load_down(gcfs_rq
->tg
->shares
)) {
3105 /* scale gcfs_rq's load into tg's shares*/
3106 load
*= scale_load_down(gcfs_rq
->tg
->shares
);
3111 delta
= load
- se
->avg
.load_avg
;
3113 /* Nothing to update */
3117 /* Set new sched_entity's load */
3118 se
->avg
.load_avg
= load
;
3119 se
->avg
.load_sum
= se
->avg
.load_avg
* LOAD_AVG_MAX
;
3121 /* Update parent cfs_rq load */
3122 add_positive(&cfs_rq
->avg
.load_avg
, delta
);
3123 cfs_rq
->avg
.load_sum
= cfs_rq
->avg
.load_avg
* LOAD_AVG_MAX
;
3126 * If the sched_entity is already enqueued, we also have to update the
3127 * runnable load avg.
3130 /* Update parent cfs_rq runnable_load_avg */
3131 add_positive(&cfs_rq
->runnable_load_avg
, delta
);
3132 cfs_rq
->runnable_load_sum
= cfs_rq
->runnable_load_avg
* LOAD_AVG_MAX
;
3136 static inline void set_tg_cfs_propagate(struct cfs_rq
*cfs_rq
)
3138 cfs_rq
->propagate_avg
= 1;
3141 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity
*se
)
3143 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
3145 if (!cfs_rq
->propagate_avg
)
3148 cfs_rq
->propagate_avg
= 0;
3152 /* Update task and its cfs_rq load average */
3153 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3155 struct cfs_rq
*cfs_rq
;
3157 if (entity_is_task(se
))
3160 if (!test_and_clear_tg_cfs_propagate(se
))
3163 cfs_rq
= cfs_rq_of(se
);
3165 set_tg_cfs_propagate(cfs_rq
);
3167 update_tg_cfs_util(cfs_rq
, se
);
3168 update_tg_cfs_load(cfs_rq
, se
);
3174 * Check if we need to update the load and the utilization of a blocked
3177 static inline bool skip_blocked_update(struct sched_entity
*se
)
3179 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3182 * If sched_entity still have not zero load or utilization, we have to
3185 if (se
->avg
.load_avg
|| se
->avg
.util_avg
)
3189 * If there is a pending propagation, we have to update the load and
3190 * the utilization of the sched_entity:
3192 if (gcfs_rq
->propagate_avg
)
3196 * Otherwise, the load and the utilization of the sched_entity is
3197 * already zero and there is no pending propagation, so it will be a
3198 * waste of time to try to decay it:
3203 #else /* CONFIG_FAIR_GROUP_SCHED */
3205 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
) {}
3207 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3212 static inline void set_tg_cfs_propagate(struct cfs_rq
*cfs_rq
) {}
3214 #endif /* CONFIG_FAIR_GROUP_SCHED */
3216 static inline void cfs_rq_util_change(struct cfs_rq
*cfs_rq
)
3218 if (&this_rq()->cfs
== cfs_rq
) {
3220 * There are a few boundary cases this might miss but it should
3221 * get called often enough that that should (hopefully) not be
3222 * a real problem -- added to that it only calls on the local
3223 * CPU, so if we enqueue remotely we'll miss an update, but
3224 * the next tick/schedule should update.
3226 * It will not get called when we go idle, because the idle
3227 * thread is a different class (!fair), nor will the utilization
3228 * number include things like RT tasks.
3230 * As is, the util number is not freq-invariant (we'd have to
3231 * implement arch_scale_freq_capacity() for that).
3235 cpufreq_update_util(rq_of(cfs_rq
), 0);
3240 * Unsigned subtract and clamp on underflow.
3242 * Explicitly do a load-store to ensure the intermediate value never hits
3243 * memory. This allows lockless observations without ever seeing the negative
3246 #define sub_positive(_ptr, _val) do { \
3247 typeof(_ptr) ptr = (_ptr); \
3248 typeof(*ptr) val = (_val); \
3249 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3253 WRITE_ONCE(*ptr, res); \
3257 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3258 * @now: current time, as per cfs_rq_clock_task()
3259 * @cfs_rq: cfs_rq to update
3260 * @update_freq: should we call cfs_rq_util_change() or will the call do so
3262 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3263 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3264 * post_init_entity_util_avg().
3266 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3268 * Returns true if the load decayed or we removed load.
3270 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3271 * call update_tg_load_avg() when this function returns true.
3274 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
, bool update_freq
)
3276 struct sched_avg
*sa
= &cfs_rq
->avg
;
3277 int decayed
, removed_load
= 0, removed_util
= 0;
3279 if (atomic_long_read(&cfs_rq
->removed_load_avg
)) {
3280 s64 r
= atomic_long_xchg(&cfs_rq
->removed_load_avg
, 0);
3281 sub_positive(&sa
->load_avg
, r
);
3282 sub_positive(&sa
->load_sum
, r
* LOAD_AVG_MAX
);
3284 set_tg_cfs_propagate(cfs_rq
);
3287 if (atomic_long_read(&cfs_rq
->removed_util_avg
)) {
3288 long r
= atomic_long_xchg(&cfs_rq
->removed_util_avg
, 0);
3289 sub_positive(&sa
->util_avg
, r
);
3290 sub_positive(&sa
->util_sum
, r
* LOAD_AVG_MAX
);
3292 set_tg_cfs_propagate(cfs_rq
);
3295 decayed
= __update_load_avg_cfs_rq(now
, cpu_of(rq_of(cfs_rq
)), cfs_rq
);
3297 #ifndef CONFIG_64BIT
3299 cfs_rq
->load_last_update_time_copy
= sa
->last_update_time
;
3302 if (update_freq
&& (decayed
|| removed_util
))
3303 cfs_rq_util_change(cfs_rq
);
3305 return decayed
|| removed_load
;
3309 * Optional action to be done while updating the load average
3311 #define UPDATE_TG 0x1
3312 #define SKIP_AGE_LOAD 0x2
3314 /* Update task and its cfs_rq load average */
3315 static inline void update_load_avg(struct sched_entity
*se
, int flags
)
3317 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3318 u64 now
= cfs_rq_clock_task(cfs_rq
);
3319 struct rq
*rq
= rq_of(cfs_rq
);
3320 int cpu
= cpu_of(rq
);
3324 * Track task load average for carrying it to new CPU after migrated, and
3325 * track group sched_entity load average for task_h_load calc in migration
3327 if (se
->avg
.last_update_time
&& !(flags
& SKIP_AGE_LOAD
))
3328 __update_load_avg_se(now
, cpu
, cfs_rq
, se
);
3330 decayed
= update_cfs_rq_load_avg(now
, cfs_rq
, true);
3331 decayed
|= propagate_entity_load_avg(se
);
3333 if (decayed
&& (flags
& UPDATE_TG
))
3334 update_tg_load_avg(cfs_rq
, 0);
3338 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3339 * @cfs_rq: cfs_rq to attach to
3340 * @se: sched_entity to attach
3342 * Must call update_cfs_rq_load_avg() before this, since we rely on
3343 * cfs_rq->avg.last_update_time being current.
3345 static void attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3347 se
->avg
.last_update_time
= cfs_rq
->avg
.last_update_time
;
3348 cfs_rq
->avg
.load_avg
+= se
->avg
.load_avg
;
3349 cfs_rq
->avg
.load_sum
+= se
->avg
.load_sum
;
3350 cfs_rq
->avg
.util_avg
+= se
->avg
.util_avg
;
3351 cfs_rq
->avg
.util_sum
+= se
->avg
.util_sum
;
3352 set_tg_cfs_propagate(cfs_rq
);
3354 cfs_rq_util_change(cfs_rq
);
3358 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3359 * @cfs_rq: cfs_rq to detach from
3360 * @se: sched_entity to detach
3362 * Must call update_cfs_rq_load_avg() before this, since we rely on
3363 * cfs_rq->avg.last_update_time being current.
3365 static void detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3368 sub_positive(&cfs_rq
->avg
.load_avg
, se
->avg
.load_avg
);
3369 sub_positive(&cfs_rq
->avg
.load_sum
, se
->avg
.load_sum
);
3370 sub_positive(&cfs_rq
->avg
.util_avg
, se
->avg
.util_avg
);
3371 sub_positive(&cfs_rq
->avg
.util_sum
, se
->avg
.util_sum
);
3372 set_tg_cfs_propagate(cfs_rq
);
3374 cfs_rq_util_change(cfs_rq
);
3377 /* Add the load generated by se into cfs_rq's load average */
3379 enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3381 struct sched_avg
*sa
= &se
->avg
;
3383 cfs_rq
->runnable_load_avg
+= sa
->load_avg
;
3384 cfs_rq
->runnable_load_sum
+= sa
->load_sum
;
3386 if (!sa
->last_update_time
) {
3387 attach_entity_load_avg(cfs_rq
, se
);
3388 update_tg_load_avg(cfs_rq
, 0);
3392 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3394 dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3396 cfs_rq
->runnable_load_avg
=
3397 max_t(long, cfs_rq
->runnable_load_avg
- se
->avg
.load_avg
, 0);
3398 cfs_rq
->runnable_load_sum
=
3399 max_t(s64
, cfs_rq
->runnable_load_sum
- se
->avg
.load_sum
, 0);
3402 #ifndef CONFIG_64BIT
3403 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3405 u64 last_update_time_copy
;
3406 u64 last_update_time
;
3409 last_update_time_copy
= cfs_rq
->load_last_update_time_copy
;
3411 last_update_time
= cfs_rq
->avg
.last_update_time
;
3412 } while (last_update_time
!= last_update_time_copy
);
3414 return last_update_time
;
3417 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3419 return cfs_rq
->avg
.last_update_time
;
3424 * Synchronize entity load avg of dequeued entity without locking
3427 void sync_entity_load_avg(struct sched_entity
*se
)
3429 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3430 u64 last_update_time
;
3432 last_update_time
= cfs_rq_last_update_time(cfs_rq
);
3433 __update_load_avg_blocked_se(last_update_time
, cpu_of(rq_of(cfs_rq
)), se
);
3437 * Task first catches up with cfs_rq, and then subtract
3438 * itself from the cfs_rq (task must be off the queue now).
3440 void remove_entity_load_avg(struct sched_entity
*se
)
3442 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3445 * tasks cannot exit without having gone through wake_up_new_task() ->
3446 * post_init_entity_util_avg() which will have added things to the
3447 * cfs_rq, so we can remove unconditionally.
3449 * Similarly for groups, they will have passed through
3450 * post_init_entity_util_avg() before unregister_sched_fair_group()
3454 sync_entity_load_avg(se
);
3455 atomic_long_add(se
->avg
.load_avg
, &cfs_rq
->removed_load_avg
);
3456 atomic_long_add(se
->avg
.util_avg
, &cfs_rq
->removed_util_avg
);
3459 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq
*cfs_rq
)
3461 return cfs_rq
->runnable_load_avg
;
3464 static inline unsigned long cfs_rq_load_avg(struct cfs_rq
*cfs_rq
)
3466 return cfs_rq
->avg
.load_avg
;
3469 static int idle_balance(struct rq
*this_rq
, struct rq_flags
*rf
);
3471 #else /* CONFIG_SMP */
3474 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
, bool update_freq
)
3479 #define UPDATE_TG 0x0
3480 #define SKIP_AGE_LOAD 0x0
3482 static inline void update_load_avg(struct sched_entity
*se
, int not_used1
)
3484 cpufreq_update_util(rq_of(cfs_rq_of(se
)), 0);
3488 enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3490 dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3491 static inline void remove_entity_load_avg(struct sched_entity
*se
) {}
3494 attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3496 detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3498 static inline int idle_balance(struct rq
*rq
, struct rq_flags
*rf
)
3503 #endif /* CONFIG_SMP */
3505 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3507 #ifdef CONFIG_SCHED_DEBUG
3508 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
3513 if (d
> 3*sysctl_sched_latency
)
3514 schedstat_inc(cfs_rq
->nr_spread_over
);
3519 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
3521 u64 vruntime
= cfs_rq
->min_vruntime
;
3524 * The 'current' period is already promised to the current tasks,
3525 * however the extra weight of the new task will slow them down a
3526 * little, place the new task so that it fits in the slot that
3527 * stays open at the end.
3529 if (initial
&& sched_feat(START_DEBIT
))
3530 vruntime
+= sched_vslice(cfs_rq
, se
);
3532 /* sleeps up to a single latency don't count. */
3534 unsigned long thresh
= sysctl_sched_latency
;
3537 * Halve their sleep time's effect, to allow
3538 * for a gentler effect of sleepers:
3540 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
3546 /* ensure we never gain time by being placed backwards. */
3547 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
3550 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
3552 static inline void check_schedstat_required(void)
3554 #ifdef CONFIG_SCHEDSTATS
3555 if (schedstat_enabled())
3558 /* Force schedstat enabled if a dependent tracepoint is active */
3559 if (trace_sched_stat_wait_enabled() ||
3560 trace_sched_stat_sleep_enabled() ||
3561 trace_sched_stat_iowait_enabled() ||
3562 trace_sched_stat_blocked_enabled() ||
3563 trace_sched_stat_runtime_enabled()) {
3564 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3565 "stat_blocked and stat_runtime require the "
3566 "kernel parameter schedstats=enable or "
3567 "kernel.sched_schedstats=1\n");
3578 * update_min_vruntime()
3579 * vruntime -= min_vruntime
3583 * update_min_vruntime()
3584 * vruntime += min_vruntime
3586 * this way the vruntime transition between RQs is done when both
3587 * min_vruntime are up-to-date.
3591 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3592 * vruntime -= min_vruntime
3596 * update_min_vruntime()
3597 * vruntime += min_vruntime
3599 * this way we don't have the most up-to-date min_vruntime on the originating
3600 * CPU and an up-to-date min_vruntime on the destination CPU.
3604 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3606 bool renorm
= !(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_MIGRATED
);
3607 bool curr
= cfs_rq
->curr
== se
;
3610 * If we're the current task, we must renormalise before calling
3614 se
->vruntime
+= cfs_rq
->min_vruntime
;
3616 update_curr(cfs_rq
);
3619 * Otherwise, renormalise after, such that we're placed at the current
3620 * moment in time, instead of some random moment in the past. Being
3621 * placed in the past could significantly boost this task to the
3622 * fairness detriment of existing tasks.
3624 if (renorm
&& !curr
)
3625 se
->vruntime
+= cfs_rq
->min_vruntime
;
3628 * When enqueuing a sched_entity, we must:
3629 * - Update loads to have both entity and cfs_rq synced with now.
3630 * - Add its load to cfs_rq->runnable_avg
3631 * - For group_entity, update its weight to reflect the new share of
3633 * - Add its new weight to cfs_rq->load.weight
3635 update_load_avg(se
, UPDATE_TG
);
3636 enqueue_entity_load_avg(cfs_rq
, se
);
3637 update_cfs_shares(se
);
3638 account_entity_enqueue(cfs_rq
, se
);
3640 if (flags
& ENQUEUE_WAKEUP
)
3641 place_entity(cfs_rq
, se
, 0);
3643 check_schedstat_required();
3644 update_stats_enqueue(cfs_rq
, se
, flags
);
3645 check_spread(cfs_rq
, se
);
3647 __enqueue_entity(cfs_rq
, se
);
3650 if (cfs_rq
->nr_running
== 1) {
3651 list_add_leaf_cfs_rq(cfs_rq
);
3652 check_enqueue_throttle(cfs_rq
);
3656 static void __clear_buddies_last(struct sched_entity
*se
)
3658 for_each_sched_entity(se
) {
3659 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3660 if (cfs_rq
->last
!= se
)
3663 cfs_rq
->last
= NULL
;
3667 static void __clear_buddies_next(struct sched_entity
*se
)
3669 for_each_sched_entity(se
) {
3670 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3671 if (cfs_rq
->next
!= se
)
3674 cfs_rq
->next
= NULL
;
3678 static void __clear_buddies_skip(struct sched_entity
*se
)
3680 for_each_sched_entity(se
) {
3681 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3682 if (cfs_rq
->skip
!= se
)
3685 cfs_rq
->skip
= NULL
;
3689 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3691 if (cfs_rq
->last
== se
)
3692 __clear_buddies_last(se
);
3694 if (cfs_rq
->next
== se
)
3695 __clear_buddies_next(se
);
3697 if (cfs_rq
->skip
== se
)
3698 __clear_buddies_skip(se
);
3701 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3704 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3707 * Update run-time statistics of the 'current'.
3709 update_curr(cfs_rq
);
3712 * When dequeuing a sched_entity, we must:
3713 * - Update loads to have both entity and cfs_rq synced with now.
3714 * - Substract its load from the cfs_rq->runnable_avg.
3715 * - Substract its previous weight from cfs_rq->load.weight.
3716 * - For group entity, update its weight to reflect the new share
3717 * of its group cfs_rq.
3719 update_load_avg(se
, UPDATE_TG
);
3720 dequeue_entity_load_avg(cfs_rq
, se
);
3722 update_stats_dequeue(cfs_rq
, se
, flags
);
3724 clear_buddies(cfs_rq
, se
);
3726 if (se
!= cfs_rq
->curr
)
3727 __dequeue_entity(cfs_rq
, se
);
3729 account_entity_dequeue(cfs_rq
, se
);
3732 * Normalize after update_curr(); which will also have moved
3733 * min_vruntime if @se is the one holding it back. But before doing
3734 * update_min_vruntime() again, which will discount @se's position and
3735 * can move min_vruntime forward still more.
3737 if (!(flags
& DEQUEUE_SLEEP
))
3738 se
->vruntime
-= cfs_rq
->min_vruntime
;
3740 /* return excess runtime on last dequeue */
3741 return_cfs_rq_runtime(cfs_rq
);
3743 update_cfs_shares(se
);
3746 * Now advance min_vruntime if @se was the entity holding it back,
3747 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3748 * put back on, and if we advance min_vruntime, we'll be placed back
3749 * further than we started -- ie. we'll be penalized.
3751 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) == DEQUEUE_SAVE
)
3752 update_min_vruntime(cfs_rq
);
3756 * Preempt the current task with a newly woken task if needed:
3759 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3761 unsigned long ideal_runtime
, delta_exec
;
3762 struct sched_entity
*se
;
3765 ideal_runtime
= sched_slice(cfs_rq
, curr
);
3766 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
3767 if (delta_exec
> ideal_runtime
) {
3768 resched_curr(rq_of(cfs_rq
));
3770 * The current task ran long enough, ensure it doesn't get
3771 * re-elected due to buddy favours.
3773 clear_buddies(cfs_rq
, curr
);
3778 * Ensure that a task that missed wakeup preemption by a
3779 * narrow margin doesn't have to wait for a full slice.
3780 * This also mitigates buddy induced latencies under load.
3782 if (delta_exec
< sysctl_sched_min_granularity
)
3785 se
= __pick_first_entity(cfs_rq
);
3786 delta
= curr
->vruntime
- se
->vruntime
;
3791 if (delta
> ideal_runtime
)
3792 resched_curr(rq_of(cfs_rq
));
3796 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3798 /* 'current' is not kept within the tree. */
3801 * Any task has to be enqueued before it get to execute on
3802 * a CPU. So account for the time it spent waiting on the
3805 update_stats_wait_end(cfs_rq
, se
);
3806 __dequeue_entity(cfs_rq
, se
);
3807 update_load_avg(se
, UPDATE_TG
);
3810 update_stats_curr_start(cfs_rq
, se
);
3814 * Track our maximum slice length, if the CPU's load is at
3815 * least twice that of our own weight (i.e. dont track it
3816 * when there are only lesser-weight tasks around):
3818 if (schedstat_enabled() && rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
3819 schedstat_set(se
->statistics
.slice_max
,
3820 max((u64
)schedstat_val(se
->statistics
.slice_max
),
3821 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
));
3824 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
3828 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
3831 * Pick the next process, keeping these things in mind, in this order:
3832 * 1) keep things fair between processes/task groups
3833 * 2) pick the "next" process, since someone really wants that to run
3834 * 3) pick the "last" process, for cache locality
3835 * 4) do not run the "skip" process, if something else is available
3837 static struct sched_entity
*
3838 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3840 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
3841 struct sched_entity
*se
;
3844 * If curr is set we have to see if its left of the leftmost entity
3845 * still in the tree, provided there was anything in the tree at all.
3847 if (!left
|| (curr
&& entity_before(curr
, left
)))
3850 se
= left
; /* ideally we run the leftmost entity */
3853 * Avoid running the skip buddy, if running something else can
3854 * be done without getting too unfair.
3856 if (cfs_rq
->skip
== se
) {
3857 struct sched_entity
*second
;
3860 second
= __pick_first_entity(cfs_rq
);
3862 second
= __pick_next_entity(se
);
3863 if (!second
|| (curr
&& entity_before(curr
, second
)))
3867 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3872 * Prefer last buddy, try to return the CPU to a preempted task.
3874 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3878 * Someone really wants this to run. If it's not unfair, run it.
3880 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3883 clear_buddies(cfs_rq
, se
);
3888 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3890 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3893 * If still on the runqueue then deactivate_task()
3894 * was not called and update_curr() has to be done:
3897 update_curr(cfs_rq
);
3899 /* throttle cfs_rqs exceeding runtime */
3900 check_cfs_rq_runtime(cfs_rq
);
3902 check_spread(cfs_rq
, prev
);
3905 update_stats_wait_start(cfs_rq
, prev
);
3906 /* Put 'current' back into the tree. */
3907 __enqueue_entity(cfs_rq
, prev
);
3908 /* in !on_rq case, update occurred at dequeue */
3909 update_load_avg(prev
, 0);
3911 cfs_rq
->curr
= NULL
;
3915 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3918 * Update run-time statistics of the 'current'.
3920 update_curr(cfs_rq
);
3923 * Ensure that runnable average is periodically updated.
3925 update_load_avg(curr
, UPDATE_TG
);
3926 update_cfs_shares(curr
);
3928 #ifdef CONFIG_SCHED_HRTICK
3930 * queued ticks are scheduled to match the slice, so don't bother
3931 * validating it and just reschedule.
3934 resched_curr(rq_of(cfs_rq
));
3938 * don't let the period tick interfere with the hrtick preemption
3940 if (!sched_feat(DOUBLE_TICK
) &&
3941 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3945 if (cfs_rq
->nr_running
> 1)
3946 check_preempt_tick(cfs_rq
, curr
);
3950 /**************************************************
3951 * CFS bandwidth control machinery
3954 #ifdef CONFIG_CFS_BANDWIDTH
3956 #ifdef HAVE_JUMP_LABEL
3957 static struct static_key __cfs_bandwidth_used
;
3959 static inline bool cfs_bandwidth_used(void)
3961 return static_key_false(&__cfs_bandwidth_used
);
3964 void cfs_bandwidth_usage_inc(void)
3966 static_key_slow_inc(&__cfs_bandwidth_used
);
3969 void cfs_bandwidth_usage_dec(void)
3971 static_key_slow_dec(&__cfs_bandwidth_used
);
3973 #else /* HAVE_JUMP_LABEL */
3974 static bool cfs_bandwidth_used(void)
3979 void cfs_bandwidth_usage_inc(void) {}
3980 void cfs_bandwidth_usage_dec(void) {}
3981 #endif /* HAVE_JUMP_LABEL */
3984 * default period for cfs group bandwidth.
3985 * default: 0.1s, units: nanoseconds
3987 static inline u64
default_cfs_period(void)
3989 return 100000000ULL;
3992 static inline u64
sched_cfs_bandwidth_slice(void)
3994 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3998 * Replenish runtime according to assigned quota and update expiration time.
3999 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4000 * additional synchronization around rq->lock.
4002 * requires cfs_b->lock
4004 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
4008 if (cfs_b
->quota
== RUNTIME_INF
)
4011 now
= sched_clock_cpu(smp_processor_id());
4012 cfs_b
->runtime
= cfs_b
->quota
;
4013 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
4016 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4018 return &tg
->cfs_bandwidth
;
4021 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4022 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
4024 if (unlikely(cfs_rq
->throttle_count
))
4025 return cfs_rq
->throttled_clock_task
- cfs_rq
->throttled_clock_task_time
;
4027 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
4030 /* returns 0 on failure to allocate runtime */
4031 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4033 struct task_group
*tg
= cfs_rq
->tg
;
4034 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
4035 u64 amount
= 0, min_amount
, expires
;
4037 /* note: this is a positive sum as runtime_remaining <= 0 */
4038 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
4040 raw_spin_lock(&cfs_b
->lock
);
4041 if (cfs_b
->quota
== RUNTIME_INF
)
4042 amount
= min_amount
;
4044 start_cfs_bandwidth(cfs_b
);
4046 if (cfs_b
->runtime
> 0) {
4047 amount
= min(cfs_b
->runtime
, min_amount
);
4048 cfs_b
->runtime
-= amount
;
4052 expires
= cfs_b
->runtime_expires
;
4053 raw_spin_unlock(&cfs_b
->lock
);
4055 cfs_rq
->runtime_remaining
+= amount
;
4057 * we may have advanced our local expiration to account for allowed
4058 * spread between our sched_clock and the one on which runtime was
4061 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
4062 cfs_rq
->runtime_expires
= expires
;
4064 return cfs_rq
->runtime_remaining
> 0;
4068 * Note: This depends on the synchronization provided by sched_clock and the
4069 * fact that rq->clock snapshots this value.
4071 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4073 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4075 /* if the deadline is ahead of our clock, nothing to do */
4076 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
4079 if (cfs_rq
->runtime_remaining
< 0)
4083 * If the local deadline has passed we have to consider the
4084 * possibility that our sched_clock is 'fast' and the global deadline
4085 * has not truly expired.
4087 * Fortunately we can check determine whether this the case by checking
4088 * whether the global deadline has advanced. It is valid to compare
4089 * cfs_b->runtime_expires without any locks since we only care about
4090 * exact equality, so a partial write will still work.
4093 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
4094 /* extend local deadline, drift is bounded above by 2 ticks */
4095 cfs_rq
->runtime_expires
+= TICK_NSEC
;
4097 /* global deadline is ahead, expiration has passed */
4098 cfs_rq
->runtime_remaining
= 0;
4102 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4104 /* dock delta_exec before expiring quota (as it could span periods) */
4105 cfs_rq
->runtime_remaining
-= delta_exec
;
4106 expire_cfs_rq_runtime(cfs_rq
);
4108 if (likely(cfs_rq
->runtime_remaining
> 0))
4112 * if we're unable to extend our runtime we resched so that the active
4113 * hierarchy can be throttled
4115 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
4116 resched_curr(rq_of(cfs_rq
));
4119 static __always_inline
4120 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4122 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
4125 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
4128 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4130 return cfs_bandwidth_used() && cfs_rq
->throttled
;
4133 /* check whether cfs_rq, or any parent, is throttled */
4134 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4136 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
4140 * Ensure that neither of the group entities corresponding to src_cpu or
4141 * dest_cpu are members of a throttled hierarchy when performing group
4142 * load-balance operations.
4144 static inline int throttled_lb_pair(struct task_group
*tg
,
4145 int src_cpu
, int dest_cpu
)
4147 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
4149 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
4150 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
4152 return throttled_hierarchy(src_cfs_rq
) ||
4153 throttled_hierarchy(dest_cfs_rq
);
4156 /* updated child weight may affect parent so we have to do this bottom up */
4157 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
4159 struct rq
*rq
= data
;
4160 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4162 cfs_rq
->throttle_count
--;
4163 if (!cfs_rq
->throttle_count
) {
4164 /* adjust cfs_rq_clock_task() */
4165 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
4166 cfs_rq
->throttled_clock_task
;
4172 static int tg_throttle_down(struct task_group
*tg
, void *data
)
4174 struct rq
*rq
= data
;
4175 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4177 /* group is entering throttled state, stop time */
4178 if (!cfs_rq
->throttle_count
)
4179 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
4180 cfs_rq
->throttle_count
++;
4185 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
4187 struct rq
*rq
= rq_of(cfs_rq
);
4188 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4189 struct sched_entity
*se
;
4190 long task_delta
, dequeue
= 1;
4193 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
4195 /* freeze hierarchy runnable averages while throttled */
4197 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
4200 task_delta
= cfs_rq
->h_nr_running
;
4201 for_each_sched_entity(se
) {
4202 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4203 /* throttled entity or throttle-on-deactivate */
4208 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
4209 qcfs_rq
->h_nr_running
-= task_delta
;
4211 if (qcfs_rq
->load
.weight
)
4216 sub_nr_running(rq
, task_delta
);
4218 cfs_rq
->throttled
= 1;
4219 cfs_rq
->throttled_clock
= rq_clock(rq
);
4220 raw_spin_lock(&cfs_b
->lock
);
4221 empty
= list_empty(&cfs_b
->throttled_cfs_rq
);
4224 * Add to the _head_ of the list, so that an already-started
4225 * distribute_cfs_runtime will not see us
4227 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
4230 * If we're the first throttled task, make sure the bandwidth
4234 start_cfs_bandwidth(cfs_b
);
4236 raw_spin_unlock(&cfs_b
->lock
);
4239 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
4241 struct rq
*rq
= rq_of(cfs_rq
);
4242 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4243 struct sched_entity
*se
;
4247 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
4249 cfs_rq
->throttled
= 0;
4251 update_rq_clock(rq
);
4253 raw_spin_lock(&cfs_b
->lock
);
4254 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
4255 list_del_rcu(&cfs_rq
->throttled_list
);
4256 raw_spin_unlock(&cfs_b
->lock
);
4258 /* update hierarchical throttle state */
4259 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
4261 if (!cfs_rq
->load
.weight
)
4264 task_delta
= cfs_rq
->h_nr_running
;
4265 for_each_sched_entity(se
) {
4269 cfs_rq
= cfs_rq_of(se
);
4271 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
4272 cfs_rq
->h_nr_running
+= task_delta
;
4274 if (cfs_rq_throttled(cfs_rq
))
4279 add_nr_running(rq
, task_delta
);
4281 /* determine whether we need to wake up potentially idle cpu */
4282 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
4286 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
4287 u64 remaining
, u64 expires
)
4289 struct cfs_rq
*cfs_rq
;
4291 u64 starting_runtime
= remaining
;
4294 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
4296 struct rq
*rq
= rq_of(cfs_rq
);
4300 if (!cfs_rq_throttled(cfs_rq
))
4303 runtime
= -cfs_rq
->runtime_remaining
+ 1;
4304 if (runtime
> remaining
)
4305 runtime
= remaining
;
4306 remaining
-= runtime
;
4308 cfs_rq
->runtime_remaining
+= runtime
;
4309 cfs_rq
->runtime_expires
= expires
;
4311 /* we check whether we're throttled above */
4312 if (cfs_rq
->runtime_remaining
> 0)
4313 unthrottle_cfs_rq(cfs_rq
);
4323 return starting_runtime
- remaining
;
4327 * Responsible for refilling a task_group's bandwidth and unthrottling its
4328 * cfs_rqs as appropriate. If there has been no activity within the last
4329 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4330 * used to track this state.
4332 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
4334 u64 runtime
, runtime_expires
;
4337 /* no need to continue the timer with no bandwidth constraint */
4338 if (cfs_b
->quota
== RUNTIME_INF
)
4339 goto out_deactivate
;
4341 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4342 cfs_b
->nr_periods
+= overrun
;
4345 * idle depends on !throttled (for the case of a large deficit), and if
4346 * we're going inactive then everything else can be deferred
4348 if (cfs_b
->idle
&& !throttled
)
4349 goto out_deactivate
;
4351 __refill_cfs_bandwidth_runtime(cfs_b
);
4354 /* mark as potentially idle for the upcoming period */
4359 /* account preceding periods in which throttling occurred */
4360 cfs_b
->nr_throttled
+= overrun
;
4362 runtime_expires
= cfs_b
->runtime_expires
;
4365 * This check is repeated as we are holding onto the new bandwidth while
4366 * we unthrottle. This can potentially race with an unthrottled group
4367 * trying to acquire new bandwidth from the global pool. This can result
4368 * in us over-using our runtime if it is all used during this loop, but
4369 * only by limited amounts in that extreme case.
4371 while (throttled
&& cfs_b
->runtime
> 0) {
4372 runtime
= cfs_b
->runtime
;
4373 raw_spin_unlock(&cfs_b
->lock
);
4374 /* we can't nest cfs_b->lock while distributing bandwidth */
4375 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
4377 raw_spin_lock(&cfs_b
->lock
);
4379 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4381 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
4385 * While we are ensured activity in the period following an
4386 * unthrottle, this also covers the case in which the new bandwidth is
4387 * insufficient to cover the existing bandwidth deficit. (Forcing the
4388 * timer to remain active while there are any throttled entities.)
4398 /* a cfs_rq won't donate quota below this amount */
4399 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
4400 /* minimum remaining period time to redistribute slack quota */
4401 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
4402 /* how long we wait to gather additional slack before distributing */
4403 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
4406 * Are we near the end of the current quota period?
4408 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4409 * hrtimer base being cleared by hrtimer_start. In the case of
4410 * migrate_hrtimers, base is never cleared, so we are fine.
4412 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
4414 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
4417 /* if the call-back is running a quota refresh is already occurring */
4418 if (hrtimer_callback_running(refresh_timer
))
4421 /* is a quota refresh about to occur? */
4422 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
4423 if (remaining
< min_expire
)
4429 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
4431 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
4433 /* if there's a quota refresh soon don't bother with slack */
4434 if (runtime_refresh_within(cfs_b
, min_left
))
4437 hrtimer_start(&cfs_b
->slack_timer
,
4438 ns_to_ktime(cfs_bandwidth_slack_period
),
4442 /* we know any runtime found here is valid as update_curr() precedes return */
4443 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4445 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4446 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
4448 if (slack_runtime
<= 0)
4451 raw_spin_lock(&cfs_b
->lock
);
4452 if (cfs_b
->quota
!= RUNTIME_INF
&&
4453 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
4454 cfs_b
->runtime
+= slack_runtime
;
4456 /* we are under rq->lock, defer unthrottling using a timer */
4457 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
4458 !list_empty(&cfs_b
->throttled_cfs_rq
))
4459 start_cfs_slack_bandwidth(cfs_b
);
4461 raw_spin_unlock(&cfs_b
->lock
);
4463 /* even if it's not valid for return we don't want to try again */
4464 cfs_rq
->runtime_remaining
-= slack_runtime
;
4467 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4469 if (!cfs_bandwidth_used())
4472 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
4475 __return_cfs_rq_runtime(cfs_rq
);
4479 * This is done with a timer (instead of inline with bandwidth return) since
4480 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4482 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
4484 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
4487 /* confirm we're still not at a refresh boundary */
4488 raw_spin_lock(&cfs_b
->lock
);
4489 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
4490 raw_spin_unlock(&cfs_b
->lock
);
4494 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
4495 runtime
= cfs_b
->runtime
;
4497 expires
= cfs_b
->runtime_expires
;
4498 raw_spin_unlock(&cfs_b
->lock
);
4503 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
4505 raw_spin_lock(&cfs_b
->lock
);
4506 if (expires
== cfs_b
->runtime_expires
)
4507 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
4508 raw_spin_unlock(&cfs_b
->lock
);
4512 * When a group wakes up we want to make sure that its quota is not already
4513 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4514 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4516 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
4518 if (!cfs_bandwidth_used())
4521 /* an active group must be handled by the update_curr()->put() path */
4522 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
4525 /* ensure the group is not already throttled */
4526 if (cfs_rq_throttled(cfs_rq
))
4529 /* update runtime allocation */
4530 account_cfs_rq_runtime(cfs_rq
, 0);
4531 if (cfs_rq
->runtime_remaining
<= 0)
4532 throttle_cfs_rq(cfs_rq
);
4535 static void sync_throttle(struct task_group
*tg
, int cpu
)
4537 struct cfs_rq
*pcfs_rq
, *cfs_rq
;
4539 if (!cfs_bandwidth_used())
4545 cfs_rq
= tg
->cfs_rq
[cpu
];
4546 pcfs_rq
= tg
->parent
->cfs_rq
[cpu
];
4548 cfs_rq
->throttle_count
= pcfs_rq
->throttle_count
;
4549 cfs_rq
->throttled_clock_task
= rq_clock_task(cpu_rq(cpu
));
4552 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4553 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4555 if (!cfs_bandwidth_used())
4558 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
4562 * it's possible for a throttled entity to be forced into a running
4563 * state (e.g. set_curr_task), in this case we're finished.
4565 if (cfs_rq_throttled(cfs_rq
))
4568 throttle_cfs_rq(cfs_rq
);
4572 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
4574 struct cfs_bandwidth
*cfs_b
=
4575 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
4577 do_sched_cfs_slack_timer(cfs_b
);
4579 return HRTIMER_NORESTART
;
4582 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
4584 struct cfs_bandwidth
*cfs_b
=
4585 container_of(timer
, struct cfs_bandwidth
, period_timer
);
4589 raw_spin_lock(&cfs_b
->lock
);
4591 overrun
= hrtimer_forward_now(timer
, cfs_b
->period
);
4595 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
4598 cfs_b
->period_active
= 0;
4599 raw_spin_unlock(&cfs_b
->lock
);
4601 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
4604 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4606 raw_spin_lock_init(&cfs_b
->lock
);
4608 cfs_b
->quota
= RUNTIME_INF
;
4609 cfs_b
->period
= ns_to_ktime(default_cfs_period());
4611 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
4612 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_PINNED
);
4613 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
4614 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4615 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
4618 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4620 cfs_rq
->runtime_enabled
= 0;
4621 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
4624 void start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4626 lockdep_assert_held(&cfs_b
->lock
);
4628 if (!cfs_b
->period_active
) {
4629 cfs_b
->period_active
= 1;
4630 hrtimer_forward_now(&cfs_b
->period_timer
, cfs_b
->period
);
4631 hrtimer_start_expires(&cfs_b
->period_timer
, HRTIMER_MODE_ABS_PINNED
);
4635 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4637 /* init_cfs_bandwidth() was not called */
4638 if (!cfs_b
->throttled_cfs_rq
.next
)
4641 hrtimer_cancel(&cfs_b
->period_timer
);
4642 hrtimer_cancel(&cfs_b
->slack_timer
);
4645 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
4647 struct cfs_rq
*cfs_rq
;
4649 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4650 struct cfs_bandwidth
*cfs_b
= &cfs_rq
->tg
->cfs_bandwidth
;
4652 raw_spin_lock(&cfs_b
->lock
);
4653 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
4654 raw_spin_unlock(&cfs_b
->lock
);
4658 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
4660 struct cfs_rq
*cfs_rq
;
4662 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4663 if (!cfs_rq
->runtime_enabled
)
4667 * clock_task is not advancing so we just need to make sure
4668 * there's some valid quota amount
4670 cfs_rq
->runtime_remaining
= 1;
4672 * Offline rq is schedulable till cpu is completely disabled
4673 * in take_cpu_down(), so we prevent new cfs throttling here.
4675 cfs_rq
->runtime_enabled
= 0;
4677 if (cfs_rq_throttled(cfs_rq
))
4678 unthrottle_cfs_rq(cfs_rq
);
4682 #else /* CONFIG_CFS_BANDWIDTH */
4683 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
4685 return rq_clock_task(rq_of(cfs_rq
));
4688 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
4689 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
4690 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
4691 static inline void sync_throttle(struct task_group
*tg
, int cpu
) {}
4692 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4694 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4699 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4704 static inline int throttled_lb_pair(struct task_group
*tg
,
4705 int src_cpu
, int dest_cpu
)
4710 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4712 #ifdef CONFIG_FAIR_GROUP_SCHED
4713 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4716 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4720 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4721 static inline void update_runtime_enabled(struct rq
*rq
) {}
4722 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
4724 #endif /* CONFIG_CFS_BANDWIDTH */
4726 /**************************************************
4727 * CFS operations on tasks:
4730 #ifdef CONFIG_SCHED_HRTICK
4731 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4733 struct sched_entity
*se
= &p
->se
;
4734 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4736 SCHED_WARN_ON(task_rq(p
) != rq
);
4738 if (rq
->cfs
.h_nr_running
> 1) {
4739 u64 slice
= sched_slice(cfs_rq
, se
);
4740 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
4741 s64 delta
= slice
- ran
;
4748 hrtick_start(rq
, delta
);
4753 * called from enqueue/dequeue and updates the hrtick when the
4754 * current task is from our class and nr_running is low enough
4757 static void hrtick_update(struct rq
*rq
)
4759 struct task_struct
*curr
= rq
->curr
;
4761 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
4764 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
4765 hrtick_start_fair(rq
, curr
);
4767 #else /* !CONFIG_SCHED_HRTICK */
4769 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4773 static inline void hrtick_update(struct rq
*rq
)
4779 * The enqueue_task method is called before nr_running is
4780 * increased. Here we update the fair scheduling stats and
4781 * then put the task into the rbtree:
4784 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4786 struct cfs_rq
*cfs_rq
;
4787 struct sched_entity
*se
= &p
->se
;
4790 * If in_iowait is set, the code below may not trigger any cpufreq
4791 * utilization updates, so do it here explicitly with the IOWAIT flag
4795 cpufreq_update_this_cpu(rq
, SCHED_CPUFREQ_IOWAIT
);
4797 for_each_sched_entity(se
) {
4800 cfs_rq
= cfs_rq_of(se
);
4801 enqueue_entity(cfs_rq
, se
, flags
);
4804 * end evaluation on encountering a throttled cfs_rq
4806 * note: in the case of encountering a throttled cfs_rq we will
4807 * post the final h_nr_running increment below.
4809 if (cfs_rq_throttled(cfs_rq
))
4811 cfs_rq
->h_nr_running
++;
4813 flags
= ENQUEUE_WAKEUP
;
4816 for_each_sched_entity(se
) {
4817 cfs_rq
= cfs_rq_of(se
);
4818 cfs_rq
->h_nr_running
++;
4820 if (cfs_rq_throttled(cfs_rq
))
4823 update_load_avg(se
, UPDATE_TG
);
4824 update_cfs_shares(se
);
4828 add_nr_running(rq
, 1);
4833 static void set_next_buddy(struct sched_entity
*se
);
4836 * The dequeue_task method is called before nr_running is
4837 * decreased. We remove the task from the rbtree and
4838 * update the fair scheduling stats:
4840 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4842 struct cfs_rq
*cfs_rq
;
4843 struct sched_entity
*se
= &p
->se
;
4844 int task_sleep
= flags
& DEQUEUE_SLEEP
;
4846 for_each_sched_entity(se
) {
4847 cfs_rq
= cfs_rq_of(se
);
4848 dequeue_entity(cfs_rq
, se
, flags
);
4851 * end evaluation on encountering a throttled cfs_rq
4853 * note: in the case of encountering a throttled cfs_rq we will
4854 * post the final h_nr_running decrement below.
4856 if (cfs_rq_throttled(cfs_rq
))
4858 cfs_rq
->h_nr_running
--;
4860 /* Don't dequeue parent if it has other entities besides us */
4861 if (cfs_rq
->load
.weight
) {
4862 /* Avoid re-evaluating load for this entity: */
4863 se
= parent_entity(se
);
4865 * Bias pick_next to pick a task from this cfs_rq, as
4866 * p is sleeping when it is within its sched_slice.
4868 if (task_sleep
&& se
&& !throttled_hierarchy(cfs_rq
))
4872 flags
|= DEQUEUE_SLEEP
;
4875 for_each_sched_entity(se
) {
4876 cfs_rq
= cfs_rq_of(se
);
4877 cfs_rq
->h_nr_running
--;
4879 if (cfs_rq_throttled(cfs_rq
))
4882 update_load_avg(se
, UPDATE_TG
);
4883 update_cfs_shares(se
);
4887 sub_nr_running(rq
, 1);
4894 /* Working cpumask for: load_balance, load_balance_newidle. */
4895 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
4896 DEFINE_PER_CPU(cpumask_var_t
, select_idle_mask
);
4898 #ifdef CONFIG_NO_HZ_COMMON
4900 * per rq 'load' arrray crap; XXX kill this.
4904 * The exact cpuload calculated at every tick would be:
4906 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4908 * If a cpu misses updates for n ticks (as it was idle) and update gets
4909 * called on the n+1-th tick when cpu may be busy, then we have:
4911 * load_n = (1 - 1/2^i)^n * load_0
4912 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4914 * decay_load_missed() below does efficient calculation of
4916 * load' = (1 - 1/2^i)^n * load
4918 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4919 * This allows us to precompute the above in said factors, thereby allowing the
4920 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4921 * fixed_power_int())
4923 * The calculation is approximated on a 128 point scale.
4925 #define DEGRADE_SHIFT 7
4927 static const u8 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
4928 static const u8 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
4929 { 0, 0, 0, 0, 0, 0, 0, 0 },
4930 { 64, 32, 8, 0, 0, 0, 0, 0 },
4931 { 96, 72, 40, 12, 1, 0, 0, 0 },
4932 { 112, 98, 75, 43, 15, 1, 0, 0 },
4933 { 120, 112, 98, 76, 45, 16, 2, 0 }
4937 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4938 * would be when CPU is idle and so we just decay the old load without
4939 * adding any new load.
4941 static unsigned long
4942 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
4946 if (!missed_updates
)
4949 if (missed_updates
>= degrade_zero_ticks
[idx
])
4953 return load
>> missed_updates
;
4955 while (missed_updates
) {
4956 if (missed_updates
% 2)
4957 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
4959 missed_updates
>>= 1;
4964 #endif /* CONFIG_NO_HZ_COMMON */
4967 * __cpu_load_update - update the rq->cpu_load[] statistics
4968 * @this_rq: The rq to update statistics for
4969 * @this_load: The current load
4970 * @pending_updates: The number of missed updates
4972 * Update rq->cpu_load[] statistics. This function is usually called every
4973 * scheduler tick (TICK_NSEC).
4975 * This function computes a decaying average:
4977 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4979 * Because of NOHZ it might not get called on every tick which gives need for
4980 * the @pending_updates argument.
4982 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4983 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4984 * = A * (A * load[i]_n-2 + B) + B
4985 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4986 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4987 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4988 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4989 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4991 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4992 * any change in load would have resulted in the tick being turned back on.
4994 * For regular NOHZ, this reduces to:
4996 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4998 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5001 static void cpu_load_update(struct rq
*this_rq
, unsigned long this_load
,
5002 unsigned long pending_updates
)
5004 unsigned long __maybe_unused tickless_load
= this_rq
->cpu_load
[0];
5007 this_rq
->nr_load_updates
++;
5009 /* Update our load: */
5010 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
5011 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
5012 unsigned long old_load
, new_load
;
5014 /* scale is effectively 1 << i now, and >> i divides by scale */
5016 old_load
= this_rq
->cpu_load
[i
];
5017 #ifdef CONFIG_NO_HZ_COMMON
5018 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
5019 if (tickless_load
) {
5020 old_load
-= decay_load_missed(tickless_load
, pending_updates
- 1, i
);
5022 * old_load can never be a negative value because a
5023 * decayed tickless_load cannot be greater than the
5024 * original tickless_load.
5026 old_load
+= tickless_load
;
5029 new_load
= this_load
;
5031 * Round up the averaging division if load is increasing. This
5032 * prevents us from getting stuck on 9 if the load is 10, for
5035 if (new_load
> old_load
)
5036 new_load
+= scale
- 1;
5038 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
5041 sched_avg_update(this_rq
);
5044 /* Used instead of source_load when we know the type == 0 */
5045 static unsigned long weighted_cpuload(const int cpu
)
5047 return cfs_rq_runnable_load_avg(&cpu_rq(cpu
)->cfs
);
5050 #ifdef CONFIG_NO_HZ_COMMON
5052 * There is no sane way to deal with nohz on smp when using jiffies because the
5053 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5054 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5056 * Therefore we need to avoid the delta approach from the regular tick when
5057 * possible since that would seriously skew the load calculation. This is why we
5058 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5059 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5060 * loop exit, nohz_idle_balance, nohz full exit...)
5062 * This means we might still be one tick off for nohz periods.
5065 static void cpu_load_update_nohz(struct rq
*this_rq
,
5066 unsigned long curr_jiffies
,
5069 unsigned long pending_updates
;
5071 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
5072 if (pending_updates
) {
5073 this_rq
->last_load_update_tick
= curr_jiffies
;
5075 * In the regular NOHZ case, we were idle, this means load 0.
5076 * In the NOHZ_FULL case, we were non-idle, we should consider
5077 * its weighted load.
5079 cpu_load_update(this_rq
, load
, pending_updates
);
5084 * Called from nohz_idle_balance() to update the load ratings before doing the
5087 static void cpu_load_update_idle(struct rq
*this_rq
)
5090 * bail if there's load or we're actually up-to-date.
5092 if (weighted_cpuload(cpu_of(this_rq
)))
5095 cpu_load_update_nohz(this_rq
, READ_ONCE(jiffies
), 0);
5099 * Record CPU load on nohz entry so we know the tickless load to account
5100 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5101 * than other cpu_load[idx] but it should be fine as cpu_load readers
5102 * shouldn't rely into synchronized cpu_load[*] updates.
5104 void cpu_load_update_nohz_start(void)
5106 struct rq
*this_rq
= this_rq();
5109 * This is all lockless but should be fine. If weighted_cpuload changes
5110 * concurrently we'll exit nohz. And cpu_load write can race with
5111 * cpu_load_update_idle() but both updater would be writing the same.
5113 this_rq
->cpu_load
[0] = weighted_cpuload(cpu_of(this_rq
));
5117 * Account the tickless load in the end of a nohz frame.
5119 void cpu_load_update_nohz_stop(void)
5121 unsigned long curr_jiffies
= READ_ONCE(jiffies
);
5122 struct rq
*this_rq
= this_rq();
5126 if (curr_jiffies
== this_rq
->last_load_update_tick
)
5129 load
= weighted_cpuload(cpu_of(this_rq
));
5130 rq_lock(this_rq
, &rf
);
5131 update_rq_clock(this_rq
);
5132 cpu_load_update_nohz(this_rq
, curr_jiffies
, load
);
5133 rq_unlock(this_rq
, &rf
);
5135 #else /* !CONFIG_NO_HZ_COMMON */
5136 static inline void cpu_load_update_nohz(struct rq
*this_rq
,
5137 unsigned long curr_jiffies
,
5138 unsigned long load
) { }
5139 #endif /* CONFIG_NO_HZ_COMMON */
5141 static void cpu_load_update_periodic(struct rq
*this_rq
, unsigned long load
)
5143 #ifdef CONFIG_NO_HZ_COMMON
5144 /* See the mess around cpu_load_update_nohz(). */
5145 this_rq
->last_load_update_tick
= READ_ONCE(jiffies
);
5147 cpu_load_update(this_rq
, load
, 1);
5151 * Called from scheduler_tick()
5153 void cpu_load_update_active(struct rq
*this_rq
)
5155 unsigned long load
= weighted_cpuload(cpu_of(this_rq
));
5157 if (tick_nohz_tick_stopped())
5158 cpu_load_update_nohz(this_rq
, READ_ONCE(jiffies
), load
);
5160 cpu_load_update_periodic(this_rq
, load
);
5164 * Return a low guess at the load of a migration-source cpu weighted
5165 * according to the scheduling class and "nice" value.
5167 * We want to under-estimate the load of migration sources, to
5168 * balance conservatively.
5170 static unsigned long source_load(int cpu
, int type
)
5172 struct rq
*rq
= cpu_rq(cpu
);
5173 unsigned long total
= weighted_cpuload(cpu
);
5175 if (type
== 0 || !sched_feat(LB_BIAS
))
5178 return min(rq
->cpu_load
[type
-1], total
);
5182 * Return a high guess at the load of a migration-target cpu weighted
5183 * according to the scheduling class and "nice" value.
5185 static unsigned long target_load(int cpu
, int type
)
5187 struct rq
*rq
= cpu_rq(cpu
);
5188 unsigned long total
= weighted_cpuload(cpu
);
5190 if (type
== 0 || !sched_feat(LB_BIAS
))
5193 return max(rq
->cpu_load
[type
-1], total
);
5196 static unsigned long capacity_of(int cpu
)
5198 return cpu_rq(cpu
)->cpu_capacity
;
5201 static unsigned long capacity_orig_of(int cpu
)
5203 return cpu_rq(cpu
)->cpu_capacity_orig
;
5206 static unsigned long cpu_avg_load_per_task(int cpu
)
5208 struct rq
*rq
= cpu_rq(cpu
);
5209 unsigned long nr_running
= READ_ONCE(rq
->cfs
.h_nr_running
);
5210 unsigned long load_avg
= weighted_cpuload(cpu
);
5213 return load_avg
/ nr_running
;
5218 #ifdef CONFIG_FAIR_GROUP_SCHED
5220 * effective_load() calculates the load change as seen from the root_task_group
5222 * Adding load to a group doesn't make a group heavier, but can cause movement
5223 * of group shares between cpus. Assuming the shares were perfectly aligned one
5224 * can calculate the shift in shares.
5226 * Calculate the effective load difference if @wl is added (subtracted) to @tg
5227 * on this @cpu and results in a total addition (subtraction) of @wg to the
5228 * total group weight.
5230 * Given a runqueue weight distribution (rw_i) we can compute a shares
5231 * distribution (s_i) using:
5233 * s_i = rw_i / \Sum rw_j (1)
5235 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
5236 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
5237 * shares distribution (s_i):
5239 * rw_i = { 2, 4, 1, 0 }
5240 * s_i = { 2/7, 4/7, 1/7, 0 }
5242 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
5243 * task used to run on and the CPU the waker is running on), we need to
5244 * compute the effect of waking a task on either CPU and, in case of a sync
5245 * wakeup, compute the effect of the current task going to sleep.
5247 * So for a change of @wl to the local @cpu with an overall group weight change
5248 * of @wl we can compute the new shares distribution (s'_i) using:
5250 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
5252 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
5253 * differences in waking a task to CPU 0. The additional task changes the
5254 * weight and shares distributions like:
5256 * rw'_i = { 3, 4, 1, 0 }
5257 * s'_i = { 3/8, 4/8, 1/8, 0 }
5259 * We can then compute the difference in effective weight by using:
5261 * dw_i = S * (s'_i - s_i) (3)
5263 * Where 'S' is the group weight as seen by its parent.
5265 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
5266 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
5267 * 4/7) times the weight of the group.
5269 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
5271 struct sched_entity
*se
= tg
->se
[cpu
];
5273 if (!tg
->parent
) /* the trivial, non-cgroup case */
5276 for_each_sched_entity(se
) {
5277 struct cfs_rq
*cfs_rq
= se
->my_q
;
5278 long W
, w
= cfs_rq_load_avg(cfs_rq
);
5283 * W = @wg + \Sum rw_j
5285 W
= wg
+ atomic_long_read(&tg
->load_avg
);
5287 /* Ensure \Sum rw_j >= rw_i */
5288 W
-= cfs_rq
->tg_load_avg_contrib
;
5297 * wl = S * s'_i; see (2)
5300 wl
= (w
* (long)scale_load_down(tg
->shares
)) / W
;
5302 wl
= scale_load_down(tg
->shares
);
5305 * Per the above, wl is the new se->load.weight value; since
5306 * those are clipped to [MIN_SHARES, ...) do so now. See
5307 * calc_cfs_shares().
5309 if (wl
< MIN_SHARES
)
5313 * wl = dw_i = S * (s'_i - s_i); see (3)
5315 wl
-= se
->avg
.load_avg
;
5318 * Recursively apply this logic to all parent groups to compute
5319 * the final effective load change on the root group. Since
5320 * only the @tg group gets extra weight, all parent groups can
5321 * only redistribute existing shares. @wl is the shift in shares
5322 * resulting from this level per the above.
5331 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
5338 static void record_wakee(struct task_struct
*p
)
5341 * Only decay a single time; tasks that have less then 1 wakeup per
5342 * jiffy will not have built up many flips.
5344 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
5345 current
->wakee_flips
>>= 1;
5346 current
->wakee_flip_decay_ts
= jiffies
;
5349 if (current
->last_wakee
!= p
) {
5350 current
->last_wakee
= p
;
5351 current
->wakee_flips
++;
5356 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5358 * A waker of many should wake a different task than the one last awakened
5359 * at a frequency roughly N times higher than one of its wakees.
5361 * In order to determine whether we should let the load spread vs consolidating
5362 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5363 * partner, and a factor of lls_size higher frequency in the other.
5365 * With both conditions met, we can be relatively sure that the relationship is
5366 * non-monogamous, with partner count exceeding socket size.
5368 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5369 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5372 static int wake_wide(struct task_struct
*p
)
5374 unsigned int master
= current
->wakee_flips
;
5375 unsigned int slave
= p
->wakee_flips
;
5376 int factor
= this_cpu_read(sd_llc_size
);
5379 swap(master
, slave
);
5380 if (slave
< factor
|| master
< slave
* factor
)
5385 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
,
5386 int prev_cpu
, int sync
)
5388 s64 this_load
, load
;
5389 s64 this_eff_load
, prev_eff_load
;
5391 struct task_group
*tg
;
5392 unsigned long weight
;
5396 this_cpu
= smp_processor_id();
5397 load
= source_load(prev_cpu
, idx
);
5398 this_load
= target_load(this_cpu
, idx
);
5401 * If sync wakeup then subtract the (maximum possible)
5402 * effect of the currently running task from the load
5403 * of the current CPU:
5406 tg
= task_group(current
);
5407 weight
= current
->se
.avg
.load_avg
;
5409 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
5410 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
5414 weight
= p
->se
.avg
.load_avg
;
5417 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5418 * due to the sync cause above having dropped this_load to 0, we'll
5419 * always have an imbalance, but there's really nothing you can do
5420 * about that, so that's good too.
5422 * Otherwise check if either cpus are near enough in load to allow this
5423 * task to be woken on this_cpu.
5425 this_eff_load
= 100;
5426 this_eff_load
*= capacity_of(prev_cpu
);
5428 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
5429 prev_eff_load
*= capacity_of(this_cpu
);
5431 if (this_load
> 0) {
5432 this_eff_load
*= this_load
+
5433 effective_load(tg
, this_cpu
, weight
, weight
);
5435 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
5438 balanced
= this_eff_load
<= prev_eff_load
;
5440 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine_attempts
);
5445 schedstat_inc(sd
->ttwu_move_affine
);
5446 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine
);
5451 static inline int task_util(struct task_struct
*p
);
5452 static int cpu_util_wake(int cpu
, struct task_struct
*p
);
5454 static unsigned long capacity_spare_wake(int cpu
, struct task_struct
*p
)
5456 return capacity_orig_of(cpu
) - cpu_util_wake(cpu
, p
);
5460 * find_idlest_group finds and returns the least busy CPU group within the
5463 static struct sched_group
*
5464 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
5465 int this_cpu
, int sd_flag
)
5467 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
5468 struct sched_group
*most_spare_sg
= NULL
;
5469 unsigned long min_runnable_load
= ULONG_MAX
, this_runnable_load
= 0;
5470 unsigned long min_avg_load
= ULONG_MAX
, this_avg_load
= 0;
5471 unsigned long most_spare
= 0, this_spare
= 0;
5472 int load_idx
= sd
->forkexec_idx
;
5473 int imbalance_scale
= 100 + (sd
->imbalance_pct
-100)/2;
5474 unsigned long imbalance
= scale_load_down(NICE_0_LOAD
) *
5475 (sd
->imbalance_pct
-100) / 100;
5477 if (sd_flag
& SD_BALANCE_WAKE
)
5478 load_idx
= sd
->wake_idx
;
5481 unsigned long load
, avg_load
, runnable_load
;
5482 unsigned long spare_cap
, max_spare_cap
;
5486 /* Skip over this group if it has no CPUs allowed */
5487 if (!cpumask_intersects(sched_group_cpus(group
),
5491 local_group
= cpumask_test_cpu(this_cpu
,
5492 sched_group_cpus(group
));
5495 * Tally up the load of all CPUs in the group and find
5496 * the group containing the CPU with most spare capacity.
5502 for_each_cpu(i
, sched_group_cpus(group
)) {
5503 /* Bias balancing toward cpus of our domain */
5505 load
= source_load(i
, load_idx
);
5507 load
= target_load(i
, load_idx
);
5509 runnable_load
+= load
;
5511 avg_load
+= cfs_rq_load_avg(&cpu_rq(i
)->cfs
);
5513 spare_cap
= capacity_spare_wake(i
, p
);
5515 if (spare_cap
> max_spare_cap
)
5516 max_spare_cap
= spare_cap
;
5519 /* Adjust by relative CPU capacity of the group */
5520 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) /
5521 group
->sgc
->capacity
;
5522 runnable_load
= (runnable_load
* SCHED_CAPACITY_SCALE
) /
5523 group
->sgc
->capacity
;
5526 this_runnable_load
= runnable_load
;
5527 this_avg_load
= avg_load
;
5528 this_spare
= max_spare_cap
;
5530 if (min_runnable_load
> (runnable_load
+ imbalance
)) {
5532 * The runnable load is significantly smaller
5533 * so we can pick this new cpu
5535 min_runnable_load
= runnable_load
;
5536 min_avg_load
= avg_load
;
5538 } else if ((runnable_load
< (min_runnable_load
+ imbalance
)) &&
5539 (100*min_avg_load
> imbalance_scale
*avg_load
)) {
5541 * The runnable loads are close so take the
5542 * blocked load into account through avg_load.
5544 min_avg_load
= avg_load
;
5548 if (most_spare
< max_spare_cap
) {
5549 most_spare
= max_spare_cap
;
5550 most_spare_sg
= group
;
5553 } while (group
= group
->next
, group
!= sd
->groups
);
5556 * The cross-over point between using spare capacity or least load
5557 * is too conservative for high utilization tasks on partially
5558 * utilized systems if we require spare_capacity > task_util(p),
5559 * so we allow for some task stuffing by using
5560 * spare_capacity > task_util(p)/2.
5562 * Spare capacity can't be used for fork because the utilization has
5563 * not been set yet, we must first select a rq to compute the initial
5566 if (sd_flag
& SD_BALANCE_FORK
)
5569 if (this_spare
> task_util(p
) / 2 &&
5570 imbalance_scale
*this_spare
> 100*most_spare
)
5573 if (most_spare
> task_util(p
) / 2)
5574 return most_spare_sg
;
5580 if (min_runnable_load
> (this_runnable_load
+ imbalance
))
5583 if ((this_runnable_load
< (min_runnable_load
+ imbalance
)) &&
5584 (100*this_avg_load
< imbalance_scale
*min_avg_load
))
5591 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5594 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
5596 unsigned long load
, min_load
= ULONG_MAX
;
5597 unsigned int min_exit_latency
= UINT_MAX
;
5598 u64 latest_idle_timestamp
= 0;
5599 int least_loaded_cpu
= this_cpu
;
5600 int shallowest_idle_cpu
= -1;
5603 /* Check if we have any choice: */
5604 if (group
->group_weight
== 1)
5605 return cpumask_first(sched_group_cpus(group
));
5607 /* Traverse only the allowed CPUs */
5608 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
5610 struct rq
*rq
= cpu_rq(i
);
5611 struct cpuidle_state
*idle
= idle_get_state(rq
);
5612 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
5614 * We give priority to a CPU whose idle state
5615 * has the smallest exit latency irrespective
5616 * of any idle timestamp.
5618 min_exit_latency
= idle
->exit_latency
;
5619 latest_idle_timestamp
= rq
->idle_stamp
;
5620 shallowest_idle_cpu
= i
;
5621 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
5622 rq
->idle_stamp
> latest_idle_timestamp
) {
5624 * If equal or no active idle state, then
5625 * the most recently idled CPU might have
5628 latest_idle_timestamp
= rq
->idle_stamp
;
5629 shallowest_idle_cpu
= i
;
5631 } else if (shallowest_idle_cpu
== -1) {
5632 load
= weighted_cpuload(i
);
5633 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
5635 least_loaded_cpu
= i
;
5640 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
5644 * Implement a for_each_cpu() variant that starts the scan at a given cpu
5645 * (@start), and wraps around.
5647 * This is used to scan for idle CPUs; such that not all CPUs looking for an
5648 * idle CPU find the same CPU. The down-side is that tasks tend to cycle
5649 * through the LLC domain.
5651 * Especially tbench is found sensitive to this.
5654 static int cpumask_next_wrap(int n
, const struct cpumask
*mask
, int start
, int *wrapped
)
5659 next
= find_next_bit(cpumask_bits(mask
), nr_cpumask_bits
, n
+1);
5663 return nr_cpumask_bits
;
5665 if (next
>= nr_cpumask_bits
) {
5675 #define for_each_cpu_wrap(cpu, mask, start, wrap) \
5676 for ((wrap) = 0, (cpu) = (start)-1; \
5677 (cpu) = cpumask_next_wrap((cpu), (mask), (start), &(wrap)), \
5678 (cpu) < nr_cpumask_bits; )
5680 #ifdef CONFIG_SCHED_SMT
5682 static inline void set_idle_cores(int cpu
, int val
)
5684 struct sched_domain_shared
*sds
;
5686 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
5688 WRITE_ONCE(sds
->has_idle_cores
, val
);
5691 static inline bool test_idle_cores(int cpu
, bool def
)
5693 struct sched_domain_shared
*sds
;
5695 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
5697 return READ_ONCE(sds
->has_idle_cores
);
5703 * Scans the local SMT mask to see if the entire core is idle, and records this
5704 * information in sd_llc_shared->has_idle_cores.
5706 * Since SMT siblings share all cache levels, inspecting this limited remote
5707 * state should be fairly cheap.
5709 void __update_idle_core(struct rq
*rq
)
5711 int core
= cpu_of(rq
);
5715 if (test_idle_cores(core
, true))
5718 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
5726 set_idle_cores(core
, 1);
5732 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5733 * there are no idle cores left in the system; tracked through
5734 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5736 static int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5738 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
5739 int core
, cpu
, wrap
;
5741 if (!static_branch_likely(&sched_smt_present
))
5744 if (!test_idle_cores(target
, false))
5747 cpumask_and(cpus
, sched_domain_span(sd
), &p
->cpus_allowed
);
5749 for_each_cpu_wrap(core
, cpus
, target
, wrap
) {
5752 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
5753 cpumask_clear_cpu(cpu
, cpus
);
5763 * Failed to find an idle core; stop looking for one.
5765 set_idle_cores(target
, 0);
5771 * Scan the local SMT mask for idle CPUs.
5773 static int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5777 if (!static_branch_likely(&sched_smt_present
))
5780 for_each_cpu(cpu
, cpu_smt_mask(target
)) {
5781 if (!cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
5790 #else /* CONFIG_SCHED_SMT */
5792 static inline int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5797 static inline int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5802 #endif /* CONFIG_SCHED_SMT */
5805 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5806 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5807 * average idle time for this rq (as found in rq->avg_idle).
5809 static int select_idle_cpu(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5811 struct sched_domain
*this_sd
;
5812 u64 avg_cost
, avg_idle
= this_rq()->avg_idle
;
5817 this_sd
= rcu_dereference(*this_cpu_ptr(&sd_llc
));
5821 avg_cost
= this_sd
->avg_scan_cost
;
5824 * Due to large variance we need a large fuzz factor; hackbench in
5825 * particularly is sensitive here.
5827 if (sched_feat(SIS_AVG_CPU
) && (avg_idle
/ 512) < avg_cost
)
5830 time
= local_clock();
5832 for_each_cpu_wrap(cpu
, sched_domain_span(sd
), target
, wrap
) {
5833 if (!cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
5839 time
= local_clock() - time
;
5840 cost
= this_sd
->avg_scan_cost
;
5841 delta
= (s64
)(time
- cost
) / 8;
5842 this_sd
->avg_scan_cost
+= delta
;
5848 * Try and locate an idle core/thread in the LLC cache domain.
5850 static int select_idle_sibling(struct task_struct
*p
, int prev
, int target
)
5852 struct sched_domain
*sd
;
5855 if (idle_cpu(target
))
5859 * If the previous cpu is cache affine and idle, don't be stupid.
5861 if (prev
!= target
&& cpus_share_cache(prev
, target
) && idle_cpu(prev
))
5864 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
5868 i
= select_idle_core(p
, sd
, target
);
5869 if ((unsigned)i
< nr_cpumask_bits
)
5872 i
= select_idle_cpu(p
, sd
, target
);
5873 if ((unsigned)i
< nr_cpumask_bits
)
5876 i
= select_idle_smt(p
, sd
, target
);
5877 if ((unsigned)i
< nr_cpumask_bits
)
5884 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5885 * tasks. The unit of the return value must be the one of capacity so we can
5886 * compare the utilization with the capacity of the CPU that is available for
5887 * CFS task (ie cpu_capacity).
5889 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5890 * recent utilization of currently non-runnable tasks on a CPU. It represents
5891 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5892 * capacity_orig is the cpu_capacity available at the highest frequency
5893 * (arch_scale_freq_capacity()).
5894 * The utilization of a CPU converges towards a sum equal to or less than the
5895 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5896 * the running time on this CPU scaled by capacity_curr.
5898 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5899 * higher than capacity_orig because of unfortunate rounding in
5900 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5901 * the average stabilizes with the new running time. We need to check that the
5902 * utilization stays within the range of [0..capacity_orig] and cap it if
5903 * necessary. Without utilization capping, a group could be seen as overloaded
5904 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5905 * available capacity. We allow utilization to overshoot capacity_curr (but not
5906 * capacity_orig) as it useful for predicting the capacity required after task
5907 * migrations (scheduler-driven DVFS).
5909 static int cpu_util(int cpu
)
5911 unsigned long util
= cpu_rq(cpu
)->cfs
.avg
.util_avg
;
5912 unsigned long capacity
= capacity_orig_of(cpu
);
5914 return (util
>= capacity
) ? capacity
: util
;
5917 static inline int task_util(struct task_struct
*p
)
5919 return p
->se
.avg
.util_avg
;
5923 * cpu_util_wake: Compute cpu utilization with any contributions from
5924 * the waking task p removed.
5926 static int cpu_util_wake(int cpu
, struct task_struct
*p
)
5928 unsigned long util
, capacity
;
5930 /* Task has no contribution or is new */
5931 if (cpu
!= task_cpu(p
) || !p
->se
.avg
.last_update_time
)
5932 return cpu_util(cpu
);
5934 capacity
= capacity_orig_of(cpu
);
5935 util
= max_t(long, cpu_rq(cpu
)->cfs
.avg
.util_avg
- task_util(p
), 0);
5937 return (util
>= capacity
) ? capacity
: util
;
5941 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5942 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5944 * In that case WAKE_AFFINE doesn't make sense and we'll let
5945 * BALANCE_WAKE sort things out.
5947 static int wake_cap(struct task_struct
*p
, int cpu
, int prev_cpu
)
5949 long min_cap
, max_cap
;
5951 min_cap
= min(capacity_orig_of(prev_cpu
), capacity_orig_of(cpu
));
5952 max_cap
= cpu_rq(cpu
)->rd
->max_cpu_capacity
;
5954 /* Minimum capacity is close to max, no need to abort wake_affine */
5955 if (max_cap
- min_cap
< max_cap
>> 3)
5958 /* Bring task utilization in sync with prev_cpu */
5959 sync_entity_load_avg(&p
->se
);
5961 return min_cap
* 1024 < task_util(p
) * capacity_margin
;
5965 * select_task_rq_fair: Select target runqueue for the waking task in domains
5966 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5967 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5969 * Balances load by selecting the idlest cpu in the idlest group, or under
5970 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5972 * Returns the target cpu number.
5974 * preempt must be disabled.
5977 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
5979 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
5980 int cpu
= smp_processor_id();
5981 int new_cpu
= prev_cpu
;
5982 int want_affine
= 0;
5983 int sync
= wake_flags
& WF_SYNC
;
5985 if (sd_flag
& SD_BALANCE_WAKE
) {
5987 want_affine
= !wake_wide(p
) && !wake_cap(p
, cpu
, prev_cpu
)
5988 && cpumask_test_cpu(cpu
, &p
->cpus_allowed
);
5992 for_each_domain(cpu
, tmp
) {
5993 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
5997 * If both cpu and prev_cpu are part of this domain,
5998 * cpu is a valid SD_WAKE_AFFINE target.
6000 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
6001 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
6006 if (tmp
->flags
& sd_flag
)
6008 else if (!want_affine
)
6013 sd
= NULL
; /* Prefer wake_affine over balance flags */
6014 if (cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, prev_cpu
, sync
))
6019 if (sd_flag
& SD_BALANCE_WAKE
) /* XXX always ? */
6020 new_cpu
= select_idle_sibling(p
, prev_cpu
, new_cpu
);
6023 struct sched_group
*group
;
6026 if (!(sd
->flags
& sd_flag
)) {
6031 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
6037 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
6038 if (new_cpu
== -1 || new_cpu
== cpu
) {
6039 /* Now try balancing at a lower domain level of cpu */
6044 /* Now try balancing at a lower domain level of new_cpu */
6046 weight
= sd
->span_weight
;
6048 for_each_domain(cpu
, tmp
) {
6049 if (weight
<= tmp
->span_weight
)
6051 if (tmp
->flags
& sd_flag
)
6054 /* while loop will break here if sd == NULL */
6062 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6063 * cfs_rq_of(p) references at time of call are still valid and identify the
6064 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6066 static void migrate_task_rq_fair(struct task_struct
*p
)
6069 * As blocked tasks retain absolute vruntime the migration needs to
6070 * deal with this by subtracting the old and adding the new
6071 * min_vruntime -- the latter is done by enqueue_entity() when placing
6072 * the task on the new runqueue.
6074 if (p
->state
== TASK_WAKING
) {
6075 struct sched_entity
*se
= &p
->se
;
6076 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6079 #ifndef CONFIG_64BIT
6080 u64 min_vruntime_copy
;
6083 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
6085 min_vruntime
= cfs_rq
->min_vruntime
;
6086 } while (min_vruntime
!= min_vruntime_copy
);
6088 min_vruntime
= cfs_rq
->min_vruntime
;
6091 se
->vruntime
-= min_vruntime
;
6095 * We are supposed to update the task to "current" time, then its up to date
6096 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6097 * what current time is, so simply throw away the out-of-date time. This
6098 * will result in the wakee task is less decayed, but giving the wakee more
6099 * load sounds not bad.
6101 remove_entity_load_avg(&p
->se
);
6103 /* Tell new CPU we are migrated */
6104 p
->se
.avg
.last_update_time
= 0;
6106 /* We have migrated, no longer consider this task hot */
6107 p
->se
.exec_start
= 0;
6110 static void task_dead_fair(struct task_struct
*p
)
6112 remove_entity_load_avg(&p
->se
);
6114 #endif /* CONFIG_SMP */
6116 static unsigned long
6117 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
6119 unsigned long gran
= sysctl_sched_wakeup_granularity
;
6122 * Since its curr running now, convert the gran from real-time
6123 * to virtual-time in his units.
6125 * By using 'se' instead of 'curr' we penalize light tasks, so
6126 * they get preempted easier. That is, if 'se' < 'curr' then
6127 * the resulting gran will be larger, therefore penalizing the
6128 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6129 * be smaller, again penalizing the lighter task.
6131 * This is especially important for buddies when the leftmost
6132 * task is higher priority than the buddy.
6134 return calc_delta_fair(gran
, se
);
6138 * Should 'se' preempt 'curr'.
6152 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
6154 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
6159 gran
= wakeup_gran(curr
, se
);
6166 static void set_last_buddy(struct sched_entity
*se
)
6168 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
6171 for_each_sched_entity(se
)
6172 cfs_rq_of(se
)->last
= se
;
6175 static void set_next_buddy(struct sched_entity
*se
)
6177 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
6180 for_each_sched_entity(se
)
6181 cfs_rq_of(se
)->next
= se
;
6184 static void set_skip_buddy(struct sched_entity
*se
)
6186 for_each_sched_entity(se
)
6187 cfs_rq_of(se
)->skip
= se
;
6191 * Preempt the current task with a newly woken task if needed:
6193 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
6195 struct task_struct
*curr
= rq
->curr
;
6196 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
6197 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
6198 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
6199 int next_buddy_marked
= 0;
6201 if (unlikely(se
== pse
))
6205 * This is possible from callers such as attach_tasks(), in which we
6206 * unconditionally check_prempt_curr() after an enqueue (which may have
6207 * lead to a throttle). This both saves work and prevents false
6208 * next-buddy nomination below.
6210 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
6213 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
6214 set_next_buddy(pse
);
6215 next_buddy_marked
= 1;
6219 * We can come here with TIF_NEED_RESCHED already set from new task
6222 * Note: this also catches the edge-case of curr being in a throttled
6223 * group (e.g. via set_curr_task), since update_curr() (in the
6224 * enqueue of curr) will have resulted in resched being set. This
6225 * prevents us from potentially nominating it as a false LAST_BUDDY
6228 if (test_tsk_need_resched(curr
))
6231 /* Idle tasks are by definition preempted by non-idle tasks. */
6232 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
6233 likely(p
->policy
!= SCHED_IDLE
))
6237 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6238 * is driven by the tick):
6240 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
6243 find_matching_se(&se
, &pse
);
6244 update_curr(cfs_rq_of(se
));
6246 if (wakeup_preempt_entity(se
, pse
) == 1) {
6248 * Bias pick_next to pick the sched entity that is
6249 * triggering this preemption.
6251 if (!next_buddy_marked
)
6252 set_next_buddy(pse
);
6261 * Only set the backward buddy when the current task is still
6262 * on the rq. This can happen when a wakeup gets interleaved
6263 * with schedule on the ->pre_schedule() or idle_balance()
6264 * point, either of which can * drop the rq lock.
6266 * Also, during early boot the idle thread is in the fair class,
6267 * for obvious reasons its a bad idea to schedule back to it.
6269 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
6272 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
6276 static struct task_struct
*
6277 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
6279 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
6280 struct sched_entity
*se
;
6281 struct task_struct
*p
;
6285 #ifdef CONFIG_FAIR_GROUP_SCHED
6286 if (!cfs_rq
->nr_running
)
6289 if (prev
->sched_class
!= &fair_sched_class
)
6293 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6294 * likely that a next task is from the same cgroup as the current.
6296 * Therefore attempt to avoid putting and setting the entire cgroup
6297 * hierarchy, only change the part that actually changes.
6301 struct sched_entity
*curr
= cfs_rq
->curr
;
6304 * Since we got here without doing put_prev_entity() we also
6305 * have to consider cfs_rq->curr. If it is still a runnable
6306 * entity, update_curr() will update its vruntime, otherwise
6307 * forget we've ever seen it.
6311 update_curr(cfs_rq
);
6316 * This call to check_cfs_rq_runtime() will do the
6317 * throttle and dequeue its entity in the parent(s).
6318 * Therefore the 'simple' nr_running test will indeed
6321 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
6325 se
= pick_next_entity(cfs_rq
, curr
);
6326 cfs_rq
= group_cfs_rq(se
);
6332 * Since we haven't yet done put_prev_entity and if the selected task
6333 * is a different task than we started out with, try and touch the
6334 * least amount of cfs_rqs.
6337 struct sched_entity
*pse
= &prev
->se
;
6339 while (!(cfs_rq
= is_same_group(se
, pse
))) {
6340 int se_depth
= se
->depth
;
6341 int pse_depth
= pse
->depth
;
6343 if (se_depth
<= pse_depth
) {
6344 put_prev_entity(cfs_rq_of(pse
), pse
);
6345 pse
= parent_entity(pse
);
6347 if (se_depth
>= pse_depth
) {
6348 set_next_entity(cfs_rq_of(se
), se
);
6349 se
= parent_entity(se
);
6353 put_prev_entity(cfs_rq
, pse
);
6354 set_next_entity(cfs_rq
, se
);
6357 if (hrtick_enabled(rq
))
6358 hrtick_start_fair(rq
, p
);
6365 if (!cfs_rq
->nr_running
)
6368 put_prev_task(rq
, prev
);
6371 se
= pick_next_entity(cfs_rq
, NULL
);
6372 set_next_entity(cfs_rq
, se
);
6373 cfs_rq
= group_cfs_rq(se
);
6378 if (hrtick_enabled(rq
))
6379 hrtick_start_fair(rq
, p
);
6384 new_tasks
= idle_balance(rq
, rf
);
6387 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6388 * possible for any higher priority task to appear. In that case we
6389 * must re-start the pick_next_entity() loop.
6401 * Account for a descheduled task:
6403 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
6405 struct sched_entity
*se
= &prev
->se
;
6406 struct cfs_rq
*cfs_rq
;
6408 for_each_sched_entity(se
) {
6409 cfs_rq
= cfs_rq_of(se
);
6410 put_prev_entity(cfs_rq
, se
);
6415 * sched_yield() is very simple
6417 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6419 static void yield_task_fair(struct rq
*rq
)
6421 struct task_struct
*curr
= rq
->curr
;
6422 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
6423 struct sched_entity
*se
= &curr
->se
;
6426 * Are we the only task in the tree?
6428 if (unlikely(rq
->nr_running
== 1))
6431 clear_buddies(cfs_rq
, se
);
6433 if (curr
->policy
!= SCHED_BATCH
) {
6434 update_rq_clock(rq
);
6436 * Update run-time statistics of the 'current'.
6438 update_curr(cfs_rq
);
6440 * Tell update_rq_clock() that we've just updated,
6441 * so we don't do microscopic update in schedule()
6442 * and double the fastpath cost.
6444 rq_clock_skip_update(rq
, true);
6450 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
6452 struct sched_entity
*se
= &p
->se
;
6454 /* throttled hierarchies are not runnable */
6455 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
6458 /* Tell the scheduler that we'd really like pse to run next. */
6461 yield_task_fair(rq
);
6467 /**************************************************
6468 * Fair scheduling class load-balancing methods.
6472 * The purpose of load-balancing is to achieve the same basic fairness the
6473 * per-cpu scheduler provides, namely provide a proportional amount of compute
6474 * time to each task. This is expressed in the following equation:
6476 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6478 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6479 * W_i,0 is defined as:
6481 * W_i,0 = \Sum_j w_i,j (2)
6483 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6484 * is derived from the nice value as per sched_prio_to_weight[].
6486 * The weight average is an exponential decay average of the instantaneous
6489 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6491 * C_i is the compute capacity of cpu i, typically it is the
6492 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6493 * can also include other factors [XXX].
6495 * To achieve this balance we define a measure of imbalance which follows
6496 * directly from (1):
6498 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6500 * We them move tasks around to minimize the imbalance. In the continuous
6501 * function space it is obvious this converges, in the discrete case we get
6502 * a few fun cases generally called infeasible weight scenarios.
6505 * - infeasible weights;
6506 * - local vs global optima in the discrete case. ]
6511 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6512 * for all i,j solution, we create a tree of cpus that follows the hardware
6513 * topology where each level pairs two lower groups (or better). This results
6514 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6515 * tree to only the first of the previous level and we decrease the frequency
6516 * of load-balance at each level inv. proportional to the number of cpus in
6522 * \Sum { --- * --- * 2^i } = O(n) (5)
6524 * `- size of each group
6525 * | | `- number of cpus doing load-balance
6527 * `- sum over all levels
6529 * Coupled with a limit on how many tasks we can migrate every balance pass,
6530 * this makes (5) the runtime complexity of the balancer.
6532 * An important property here is that each CPU is still (indirectly) connected
6533 * to every other cpu in at most O(log n) steps:
6535 * The adjacency matrix of the resulting graph is given by:
6538 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6541 * And you'll find that:
6543 * A^(log_2 n)_i,j != 0 for all i,j (7)
6545 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6546 * The task movement gives a factor of O(m), giving a convergence complexity
6549 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6554 * In order to avoid CPUs going idle while there's still work to do, new idle
6555 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6556 * tree itself instead of relying on other CPUs to bring it work.
6558 * This adds some complexity to both (5) and (8) but it reduces the total idle
6566 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6569 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6574 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6576 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6578 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6581 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6582 * rewrite all of this once again.]
6585 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
6587 enum fbq_type
{ regular
, remote
, all
};
6589 #define LBF_ALL_PINNED 0x01
6590 #define LBF_NEED_BREAK 0x02
6591 #define LBF_DST_PINNED 0x04
6592 #define LBF_SOME_PINNED 0x08
6595 struct sched_domain
*sd
;
6603 struct cpumask
*dst_grpmask
;
6605 enum cpu_idle_type idle
;
6607 /* The set of CPUs under consideration for load-balancing */
6608 struct cpumask
*cpus
;
6613 unsigned int loop_break
;
6614 unsigned int loop_max
;
6616 enum fbq_type fbq_type
;
6617 struct list_head tasks
;
6621 * Is this task likely cache-hot:
6623 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
6627 lockdep_assert_held(&env
->src_rq
->lock
);
6629 if (p
->sched_class
!= &fair_sched_class
)
6632 if (unlikely(p
->policy
== SCHED_IDLE
))
6636 * Buddy candidates are cache hot:
6638 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
6639 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
6640 &p
->se
== cfs_rq_of(&p
->se
)->last
))
6643 if (sysctl_sched_migration_cost
== -1)
6645 if (sysctl_sched_migration_cost
== 0)
6648 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
6650 return delta
< (s64
)sysctl_sched_migration_cost
;
6653 #ifdef CONFIG_NUMA_BALANCING
6655 * Returns 1, if task migration degrades locality
6656 * Returns 0, if task migration improves locality i.e migration preferred.
6657 * Returns -1, if task migration is not affected by locality.
6659 static int migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
6661 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
6662 unsigned long src_faults
, dst_faults
;
6663 int src_nid
, dst_nid
;
6665 if (!static_branch_likely(&sched_numa_balancing
))
6668 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
6671 src_nid
= cpu_to_node(env
->src_cpu
);
6672 dst_nid
= cpu_to_node(env
->dst_cpu
);
6674 if (src_nid
== dst_nid
)
6677 /* Migrating away from the preferred node is always bad. */
6678 if (src_nid
== p
->numa_preferred_nid
) {
6679 if (env
->src_rq
->nr_running
> env
->src_rq
->nr_preferred_running
)
6685 /* Encourage migration to the preferred node. */
6686 if (dst_nid
== p
->numa_preferred_nid
)
6690 src_faults
= group_faults(p
, src_nid
);
6691 dst_faults
= group_faults(p
, dst_nid
);
6693 src_faults
= task_faults(p
, src_nid
);
6694 dst_faults
= task_faults(p
, dst_nid
);
6697 return dst_faults
< src_faults
;
6701 static inline int migrate_degrades_locality(struct task_struct
*p
,
6709 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6712 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
6716 lockdep_assert_held(&env
->src_rq
->lock
);
6719 * We do not migrate tasks that are:
6720 * 1) throttled_lb_pair, or
6721 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6722 * 3) running (obviously), or
6723 * 4) are cache-hot on their current CPU.
6725 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
6728 if (!cpumask_test_cpu(env
->dst_cpu
, &p
->cpus_allowed
)) {
6731 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_affine
);
6733 env
->flags
|= LBF_SOME_PINNED
;
6736 * Remember if this task can be migrated to any other cpu in
6737 * our sched_group. We may want to revisit it if we couldn't
6738 * meet load balance goals by pulling other tasks on src_cpu.
6740 * Also avoid computing new_dst_cpu if we have already computed
6741 * one in current iteration.
6743 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
6746 /* Prevent to re-select dst_cpu via env's cpus */
6747 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
6748 if (cpumask_test_cpu(cpu
, &p
->cpus_allowed
)) {
6749 env
->flags
|= LBF_DST_PINNED
;
6750 env
->new_dst_cpu
= cpu
;
6758 /* Record that we found atleast one task that could run on dst_cpu */
6759 env
->flags
&= ~LBF_ALL_PINNED
;
6761 if (task_running(env
->src_rq
, p
)) {
6762 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_running
);
6767 * Aggressive migration if:
6768 * 1) destination numa is preferred
6769 * 2) task is cache cold, or
6770 * 3) too many balance attempts have failed.
6772 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
6773 if (tsk_cache_hot
== -1)
6774 tsk_cache_hot
= task_hot(p
, env
);
6776 if (tsk_cache_hot
<= 0 ||
6777 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
6778 if (tsk_cache_hot
== 1) {
6779 schedstat_inc(env
->sd
->lb_hot_gained
[env
->idle
]);
6780 schedstat_inc(p
->se
.statistics
.nr_forced_migrations
);
6785 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_hot
);
6790 * detach_task() -- detach the task for the migration specified in env
6792 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
6794 lockdep_assert_held(&env
->src_rq
->lock
);
6796 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
6797 deactivate_task(env
->src_rq
, p
, DEQUEUE_NOCLOCK
);
6798 set_task_cpu(p
, env
->dst_cpu
);
6802 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6803 * part of active balancing operations within "domain".
6805 * Returns a task if successful and NULL otherwise.
6807 static struct task_struct
*detach_one_task(struct lb_env
*env
)
6809 struct task_struct
*p
, *n
;
6811 lockdep_assert_held(&env
->src_rq
->lock
);
6813 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
6814 if (!can_migrate_task(p
, env
))
6817 detach_task(p
, env
);
6820 * Right now, this is only the second place where
6821 * lb_gained[env->idle] is updated (other is detach_tasks)
6822 * so we can safely collect stats here rather than
6823 * inside detach_tasks().
6825 schedstat_inc(env
->sd
->lb_gained
[env
->idle
]);
6831 static const unsigned int sched_nr_migrate_break
= 32;
6834 * detach_tasks() -- tries to detach up to imbalance weighted load from
6835 * busiest_rq, as part of a balancing operation within domain "sd".
6837 * Returns number of detached tasks if successful and 0 otherwise.
6839 static int detach_tasks(struct lb_env
*env
)
6841 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
6842 struct task_struct
*p
;
6846 lockdep_assert_held(&env
->src_rq
->lock
);
6848 if (env
->imbalance
<= 0)
6851 while (!list_empty(tasks
)) {
6853 * We don't want to steal all, otherwise we may be treated likewise,
6854 * which could at worst lead to a livelock crash.
6856 if (env
->idle
!= CPU_NOT_IDLE
&& env
->src_rq
->nr_running
<= 1)
6859 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
6862 /* We've more or less seen every task there is, call it quits */
6863 if (env
->loop
> env
->loop_max
)
6866 /* take a breather every nr_migrate tasks */
6867 if (env
->loop
> env
->loop_break
) {
6868 env
->loop_break
+= sched_nr_migrate_break
;
6869 env
->flags
|= LBF_NEED_BREAK
;
6873 if (!can_migrate_task(p
, env
))
6876 load
= task_h_load(p
);
6878 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
6881 if ((load
/ 2) > env
->imbalance
)
6884 detach_task(p
, env
);
6885 list_add(&p
->se
.group_node
, &env
->tasks
);
6888 env
->imbalance
-= load
;
6890 #ifdef CONFIG_PREEMPT
6892 * NEWIDLE balancing is a source of latency, so preemptible
6893 * kernels will stop after the first task is detached to minimize
6894 * the critical section.
6896 if (env
->idle
== CPU_NEWLY_IDLE
)
6901 * We only want to steal up to the prescribed amount of
6904 if (env
->imbalance
<= 0)
6909 list_move_tail(&p
->se
.group_node
, tasks
);
6913 * Right now, this is one of only two places we collect this stat
6914 * so we can safely collect detach_one_task() stats here rather
6915 * than inside detach_one_task().
6917 schedstat_add(env
->sd
->lb_gained
[env
->idle
], detached
);
6923 * attach_task() -- attach the task detached by detach_task() to its new rq.
6925 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
6927 lockdep_assert_held(&rq
->lock
);
6929 BUG_ON(task_rq(p
) != rq
);
6930 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
6931 p
->on_rq
= TASK_ON_RQ_QUEUED
;
6932 check_preempt_curr(rq
, p
, 0);
6936 * attach_one_task() -- attaches the task returned from detach_one_task() to
6939 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
6944 update_rq_clock(rq
);
6950 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6953 static void attach_tasks(struct lb_env
*env
)
6955 struct list_head
*tasks
= &env
->tasks
;
6956 struct task_struct
*p
;
6959 rq_lock(env
->dst_rq
, &rf
);
6960 update_rq_clock(env
->dst_rq
);
6962 while (!list_empty(tasks
)) {
6963 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
6964 list_del_init(&p
->se
.group_node
);
6966 attach_task(env
->dst_rq
, p
);
6969 rq_unlock(env
->dst_rq
, &rf
);
6972 #ifdef CONFIG_FAIR_GROUP_SCHED
6973 static void update_blocked_averages(int cpu
)
6975 struct rq
*rq
= cpu_rq(cpu
);
6976 struct cfs_rq
*cfs_rq
;
6979 rq_lock_irqsave(rq
, &rf
);
6980 update_rq_clock(rq
);
6983 * Iterates the task_group tree in a bottom up fashion, see
6984 * list_add_leaf_cfs_rq() for details.
6986 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
6987 struct sched_entity
*se
;
6989 /* throttled entities do not contribute to load */
6990 if (throttled_hierarchy(cfs_rq
))
6993 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq
), cfs_rq
, true))
6994 update_tg_load_avg(cfs_rq
, 0);
6996 /* Propagate pending load changes to the parent, if any: */
6997 se
= cfs_rq
->tg
->se
[cpu
];
6998 if (se
&& !skip_blocked_update(se
))
6999 update_load_avg(se
, 0);
7001 rq_unlock_irqrestore(rq
, &rf
);
7005 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7006 * This needs to be done in a top-down fashion because the load of a child
7007 * group is a fraction of its parents load.
7009 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
7011 struct rq
*rq
= rq_of(cfs_rq
);
7012 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
7013 unsigned long now
= jiffies
;
7016 if (cfs_rq
->last_h_load_update
== now
)
7019 cfs_rq
->h_load_next
= NULL
;
7020 for_each_sched_entity(se
) {
7021 cfs_rq
= cfs_rq_of(se
);
7022 cfs_rq
->h_load_next
= se
;
7023 if (cfs_rq
->last_h_load_update
== now
)
7028 cfs_rq
->h_load
= cfs_rq_load_avg(cfs_rq
);
7029 cfs_rq
->last_h_load_update
= now
;
7032 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
7033 load
= cfs_rq
->h_load
;
7034 load
= div64_ul(load
* se
->avg
.load_avg
,
7035 cfs_rq_load_avg(cfs_rq
) + 1);
7036 cfs_rq
= group_cfs_rq(se
);
7037 cfs_rq
->h_load
= load
;
7038 cfs_rq
->last_h_load_update
= now
;
7042 static unsigned long task_h_load(struct task_struct
*p
)
7044 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
7046 update_cfs_rq_h_load(cfs_rq
);
7047 return div64_ul(p
->se
.avg
.load_avg
* cfs_rq
->h_load
,
7048 cfs_rq_load_avg(cfs_rq
) + 1);
7051 static inline void update_blocked_averages(int cpu
)
7053 struct rq
*rq
= cpu_rq(cpu
);
7054 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
7057 rq_lock_irqsave(rq
, &rf
);
7058 update_rq_clock(rq
);
7059 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq
), cfs_rq
, true);
7060 rq_unlock_irqrestore(rq
, &rf
);
7063 static unsigned long task_h_load(struct task_struct
*p
)
7065 return p
->se
.avg
.load_avg
;
7069 /********** Helpers for find_busiest_group ************************/
7078 * sg_lb_stats - stats of a sched_group required for load_balancing
7080 struct sg_lb_stats
{
7081 unsigned long avg_load
; /*Avg load across the CPUs of the group */
7082 unsigned long group_load
; /* Total load over the CPUs of the group */
7083 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
7084 unsigned long load_per_task
;
7085 unsigned long group_capacity
;
7086 unsigned long group_util
; /* Total utilization of the group */
7087 unsigned int sum_nr_running
; /* Nr tasks running in the group */
7088 unsigned int idle_cpus
;
7089 unsigned int group_weight
;
7090 enum group_type group_type
;
7091 int group_no_capacity
;
7092 #ifdef CONFIG_NUMA_BALANCING
7093 unsigned int nr_numa_running
;
7094 unsigned int nr_preferred_running
;
7099 * sd_lb_stats - Structure to store the statistics of a sched_domain
7100 * during load balancing.
7102 struct sd_lb_stats
{
7103 struct sched_group
*busiest
; /* Busiest group in this sd */
7104 struct sched_group
*local
; /* Local group in this sd */
7105 unsigned long total_load
; /* Total load of all groups in sd */
7106 unsigned long total_capacity
; /* Total capacity of all groups in sd */
7107 unsigned long avg_load
; /* Average load across all groups in sd */
7109 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
7110 struct sg_lb_stats local_stat
; /* Statistics of the local group */
7113 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
7116 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7117 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7118 * We must however clear busiest_stat::avg_load because
7119 * update_sd_pick_busiest() reads this before assignment.
7121 *sds
= (struct sd_lb_stats
){
7125 .total_capacity
= 0UL,
7128 .sum_nr_running
= 0,
7129 .group_type
= group_other
,
7135 * get_sd_load_idx - Obtain the load index for a given sched domain.
7136 * @sd: The sched_domain whose load_idx is to be obtained.
7137 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7139 * Return: The load index.
7141 static inline int get_sd_load_idx(struct sched_domain
*sd
,
7142 enum cpu_idle_type idle
)
7148 load_idx
= sd
->busy_idx
;
7151 case CPU_NEWLY_IDLE
:
7152 load_idx
= sd
->newidle_idx
;
7155 load_idx
= sd
->idle_idx
;
7162 static unsigned long scale_rt_capacity(int cpu
)
7164 struct rq
*rq
= cpu_rq(cpu
);
7165 u64 total
, used
, age_stamp
, avg
;
7169 * Since we're reading these variables without serialization make sure
7170 * we read them once before doing sanity checks on them.
7172 age_stamp
= READ_ONCE(rq
->age_stamp
);
7173 avg
= READ_ONCE(rq
->rt_avg
);
7174 delta
= __rq_clock_broken(rq
) - age_stamp
;
7176 if (unlikely(delta
< 0))
7179 total
= sched_avg_period() + delta
;
7181 used
= div_u64(avg
, total
);
7183 if (likely(used
< SCHED_CAPACITY_SCALE
))
7184 return SCHED_CAPACITY_SCALE
- used
;
7189 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
7191 unsigned long capacity
= arch_scale_cpu_capacity(sd
, cpu
);
7192 struct sched_group
*sdg
= sd
->groups
;
7194 cpu_rq(cpu
)->cpu_capacity_orig
= capacity
;
7196 capacity
*= scale_rt_capacity(cpu
);
7197 capacity
>>= SCHED_CAPACITY_SHIFT
;
7202 cpu_rq(cpu
)->cpu_capacity
= capacity
;
7203 sdg
->sgc
->capacity
= capacity
;
7204 sdg
->sgc
->min_capacity
= capacity
;
7207 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
7209 struct sched_domain
*child
= sd
->child
;
7210 struct sched_group
*group
, *sdg
= sd
->groups
;
7211 unsigned long capacity
, min_capacity
;
7212 unsigned long interval
;
7214 interval
= msecs_to_jiffies(sd
->balance_interval
);
7215 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
7216 sdg
->sgc
->next_update
= jiffies
+ interval
;
7219 update_cpu_capacity(sd
, cpu
);
7224 min_capacity
= ULONG_MAX
;
7226 if (child
->flags
& SD_OVERLAP
) {
7228 * SD_OVERLAP domains cannot assume that child groups
7229 * span the current group.
7232 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
7233 struct sched_group_capacity
*sgc
;
7234 struct rq
*rq
= cpu_rq(cpu
);
7237 * build_sched_domains() -> init_sched_groups_capacity()
7238 * gets here before we've attached the domains to the
7241 * Use capacity_of(), which is set irrespective of domains
7242 * in update_cpu_capacity().
7244 * This avoids capacity from being 0 and
7245 * causing divide-by-zero issues on boot.
7247 if (unlikely(!rq
->sd
)) {
7248 capacity
+= capacity_of(cpu
);
7250 sgc
= rq
->sd
->groups
->sgc
;
7251 capacity
+= sgc
->capacity
;
7254 min_capacity
= min(capacity
, min_capacity
);
7258 * !SD_OVERLAP domains can assume that child groups
7259 * span the current group.
7262 group
= child
->groups
;
7264 struct sched_group_capacity
*sgc
= group
->sgc
;
7266 capacity
+= sgc
->capacity
;
7267 min_capacity
= min(sgc
->min_capacity
, min_capacity
);
7268 group
= group
->next
;
7269 } while (group
!= child
->groups
);
7272 sdg
->sgc
->capacity
= capacity
;
7273 sdg
->sgc
->min_capacity
= min_capacity
;
7277 * Check whether the capacity of the rq has been noticeably reduced by side
7278 * activity. The imbalance_pct is used for the threshold.
7279 * Return true is the capacity is reduced
7282 check_cpu_capacity(struct rq
*rq
, struct sched_domain
*sd
)
7284 return ((rq
->cpu_capacity
* sd
->imbalance_pct
) <
7285 (rq
->cpu_capacity_orig
* 100));
7289 * Group imbalance indicates (and tries to solve) the problem where balancing
7290 * groups is inadequate due to ->cpus_allowed constraints.
7292 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7293 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7296 * { 0 1 2 3 } { 4 5 6 7 }
7299 * If we were to balance group-wise we'd place two tasks in the first group and
7300 * two tasks in the second group. Clearly this is undesired as it will overload
7301 * cpu 3 and leave one of the cpus in the second group unused.
7303 * The current solution to this issue is detecting the skew in the first group
7304 * by noticing the lower domain failed to reach balance and had difficulty
7305 * moving tasks due to affinity constraints.
7307 * When this is so detected; this group becomes a candidate for busiest; see
7308 * update_sd_pick_busiest(). And calculate_imbalance() and
7309 * find_busiest_group() avoid some of the usual balance conditions to allow it
7310 * to create an effective group imbalance.
7312 * This is a somewhat tricky proposition since the next run might not find the
7313 * group imbalance and decide the groups need to be balanced again. A most
7314 * subtle and fragile situation.
7317 static inline int sg_imbalanced(struct sched_group
*group
)
7319 return group
->sgc
->imbalance
;
7323 * group_has_capacity returns true if the group has spare capacity that could
7324 * be used by some tasks.
7325 * We consider that a group has spare capacity if the * number of task is
7326 * smaller than the number of CPUs or if the utilization is lower than the
7327 * available capacity for CFS tasks.
7328 * For the latter, we use a threshold to stabilize the state, to take into
7329 * account the variance of the tasks' load and to return true if the available
7330 * capacity in meaningful for the load balancer.
7331 * As an example, an available capacity of 1% can appear but it doesn't make
7332 * any benefit for the load balance.
7335 group_has_capacity(struct lb_env
*env
, struct sg_lb_stats
*sgs
)
7337 if (sgs
->sum_nr_running
< sgs
->group_weight
)
7340 if ((sgs
->group_capacity
* 100) >
7341 (sgs
->group_util
* env
->sd
->imbalance_pct
))
7348 * group_is_overloaded returns true if the group has more tasks than it can
7350 * group_is_overloaded is not equals to !group_has_capacity because a group
7351 * with the exact right number of tasks, has no more spare capacity but is not
7352 * overloaded so both group_has_capacity and group_is_overloaded return
7356 group_is_overloaded(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_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7370 * per-CPU capacity than sched_group ref.
7373 group_smaller_cpu_capacity(struct sched_group
*sg
, struct sched_group
*ref
)
7375 return sg
->sgc
->min_capacity
* capacity_margin
<
7376 ref
->sgc
->min_capacity
* 1024;
7380 group_type
group_classify(struct sched_group
*group
,
7381 struct sg_lb_stats
*sgs
)
7383 if (sgs
->group_no_capacity
)
7384 return group_overloaded
;
7386 if (sg_imbalanced(group
))
7387 return group_imbalanced
;
7393 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7394 * @env: The load balancing environment.
7395 * @group: sched_group whose statistics are to be updated.
7396 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7397 * @local_group: Does group contain this_cpu.
7398 * @sgs: variable to hold the statistics for this group.
7399 * @overload: Indicate more than one runnable task for any CPU.
7401 static inline void update_sg_lb_stats(struct lb_env
*env
,
7402 struct sched_group
*group
, int load_idx
,
7403 int local_group
, struct sg_lb_stats
*sgs
,
7409 memset(sgs
, 0, sizeof(*sgs
));
7411 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
7412 struct rq
*rq
= cpu_rq(i
);
7414 /* Bias balancing toward cpus of our domain */
7416 load
= target_load(i
, load_idx
);
7418 load
= source_load(i
, load_idx
);
7420 sgs
->group_load
+= load
;
7421 sgs
->group_util
+= cpu_util(i
);
7422 sgs
->sum_nr_running
+= rq
->cfs
.h_nr_running
;
7424 nr_running
= rq
->nr_running
;
7428 #ifdef CONFIG_NUMA_BALANCING
7429 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
7430 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
7432 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
7434 * No need to call idle_cpu() if nr_running is not 0
7436 if (!nr_running
&& idle_cpu(i
))
7440 /* Adjust by relative CPU capacity of the group */
7441 sgs
->group_capacity
= group
->sgc
->capacity
;
7442 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
7444 if (sgs
->sum_nr_running
)
7445 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
7447 sgs
->group_weight
= group
->group_weight
;
7449 sgs
->group_no_capacity
= group_is_overloaded(env
, sgs
);
7450 sgs
->group_type
= group_classify(group
, sgs
);
7454 * update_sd_pick_busiest - return 1 on busiest group
7455 * @env: The load balancing environment.
7456 * @sds: sched_domain statistics
7457 * @sg: sched_group candidate to be checked for being the busiest
7458 * @sgs: sched_group statistics
7460 * Determine if @sg is a busier group than the previously selected
7463 * Return: %true if @sg is a busier group than the previously selected
7464 * busiest group. %false otherwise.
7466 static bool update_sd_pick_busiest(struct lb_env
*env
,
7467 struct sd_lb_stats
*sds
,
7468 struct sched_group
*sg
,
7469 struct sg_lb_stats
*sgs
)
7471 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
7473 if (sgs
->group_type
> busiest
->group_type
)
7476 if (sgs
->group_type
< busiest
->group_type
)
7479 if (sgs
->avg_load
<= busiest
->avg_load
)
7482 if (!(env
->sd
->flags
& SD_ASYM_CPUCAPACITY
))
7486 * Candidate sg has no more than one task per CPU and
7487 * has higher per-CPU capacity. Migrating tasks to less
7488 * capable CPUs may harm throughput. Maximize throughput,
7489 * power/energy consequences are not considered.
7491 if (sgs
->sum_nr_running
<= sgs
->group_weight
&&
7492 group_smaller_cpu_capacity(sds
->local
, sg
))
7496 /* This is the busiest node in its class. */
7497 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
7500 /* No ASYM_PACKING if target cpu is already busy */
7501 if (env
->idle
== CPU_NOT_IDLE
)
7504 * ASYM_PACKING needs to move all the work to the highest
7505 * prority CPUs in the group, therefore mark all groups
7506 * of lower priority than ourself as busy.
7508 if (sgs
->sum_nr_running
&&
7509 sched_asym_prefer(env
->dst_cpu
, sg
->asym_prefer_cpu
)) {
7513 /* Prefer to move from lowest priority cpu's work */
7514 if (sched_asym_prefer(sds
->busiest
->asym_prefer_cpu
,
7515 sg
->asym_prefer_cpu
))
7522 #ifdef CONFIG_NUMA_BALANCING
7523 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
7525 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
7527 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
7532 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
7534 if (rq
->nr_running
> rq
->nr_numa_running
)
7536 if (rq
->nr_running
> rq
->nr_preferred_running
)
7541 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
7546 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
7550 #endif /* CONFIG_NUMA_BALANCING */
7553 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7554 * @env: The load balancing environment.
7555 * @sds: variable to hold the statistics for this sched_domain.
7557 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
7559 struct sched_domain
*child
= env
->sd
->child
;
7560 struct sched_group
*sg
= env
->sd
->groups
;
7561 struct sg_lb_stats
*local
= &sds
->local_stat
;
7562 struct sg_lb_stats tmp_sgs
;
7563 int load_idx
, prefer_sibling
= 0;
7564 bool overload
= false;
7566 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
7569 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
7572 struct sg_lb_stats
*sgs
= &tmp_sgs
;
7575 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
7580 if (env
->idle
!= CPU_NEWLY_IDLE
||
7581 time_after_eq(jiffies
, sg
->sgc
->next_update
))
7582 update_group_capacity(env
->sd
, env
->dst_cpu
);
7585 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
,
7592 * In case the child domain prefers tasks go to siblings
7593 * first, lower the sg capacity so that we'll try
7594 * and move all the excess tasks away. We lower the capacity
7595 * of a group only if the local group has the capacity to fit
7596 * these excess tasks. The extra check prevents the case where
7597 * you always pull from the heaviest group when it is already
7598 * under-utilized (possible with a large weight task outweighs
7599 * the tasks on the system).
7601 if (prefer_sibling
&& sds
->local
&&
7602 group_has_capacity(env
, local
) &&
7603 (sgs
->sum_nr_running
> local
->sum_nr_running
+ 1)) {
7604 sgs
->group_no_capacity
= 1;
7605 sgs
->group_type
= group_classify(sg
, sgs
);
7608 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
7610 sds
->busiest_stat
= *sgs
;
7614 /* Now, start updating sd_lb_stats */
7615 sds
->total_load
+= sgs
->group_load
;
7616 sds
->total_capacity
+= sgs
->group_capacity
;
7619 } while (sg
!= env
->sd
->groups
);
7621 if (env
->sd
->flags
& SD_NUMA
)
7622 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
7624 if (!env
->sd
->parent
) {
7625 /* update overload indicator if we are at root domain */
7626 if (env
->dst_rq
->rd
->overload
!= overload
)
7627 env
->dst_rq
->rd
->overload
= overload
;
7633 * check_asym_packing - Check to see if the group is packed into the
7636 * This is primarily intended to used at the sibling level. Some
7637 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7638 * case of POWER7, it can move to lower SMT modes only when higher
7639 * threads are idle. When in lower SMT modes, the threads will
7640 * perform better since they share less core resources. Hence when we
7641 * have idle threads, we want them to be the higher ones.
7643 * This packing function is run on idle threads. It checks to see if
7644 * the busiest CPU in this domain (core in the P7 case) has a higher
7645 * CPU number than the packing function is being run on. Here we are
7646 * assuming lower CPU number will be equivalent to lower a SMT thread
7649 * Return: 1 when packing is required and a task should be moved to
7650 * this CPU. The amount of the imbalance is returned in *imbalance.
7652 * @env: The load balancing environment.
7653 * @sds: Statistics of the sched_domain which is to be packed
7655 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
7659 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
7662 if (env
->idle
== CPU_NOT_IDLE
)
7668 busiest_cpu
= sds
->busiest
->asym_prefer_cpu
;
7669 if (sched_asym_prefer(busiest_cpu
, env
->dst_cpu
))
7672 env
->imbalance
= DIV_ROUND_CLOSEST(
7673 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
7674 SCHED_CAPACITY_SCALE
);
7680 * fix_small_imbalance - Calculate the minor imbalance that exists
7681 * amongst the groups of a sched_domain, during
7683 * @env: The load balancing environment.
7684 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7687 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
7689 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
7690 unsigned int imbn
= 2;
7691 unsigned long scaled_busy_load_per_task
;
7692 struct sg_lb_stats
*local
, *busiest
;
7694 local
= &sds
->local_stat
;
7695 busiest
= &sds
->busiest_stat
;
7697 if (!local
->sum_nr_running
)
7698 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
7699 else if (busiest
->load_per_task
> local
->load_per_task
)
7702 scaled_busy_load_per_task
=
7703 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
7704 busiest
->group_capacity
;
7706 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
7707 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
7708 env
->imbalance
= busiest
->load_per_task
;
7713 * OK, we don't have enough imbalance to justify moving tasks,
7714 * however we may be able to increase total CPU capacity used by
7718 capa_now
+= busiest
->group_capacity
*
7719 min(busiest
->load_per_task
, busiest
->avg_load
);
7720 capa_now
+= local
->group_capacity
*
7721 min(local
->load_per_task
, local
->avg_load
);
7722 capa_now
/= SCHED_CAPACITY_SCALE
;
7724 /* Amount of load we'd subtract */
7725 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
7726 capa_move
+= busiest
->group_capacity
*
7727 min(busiest
->load_per_task
,
7728 busiest
->avg_load
- scaled_busy_load_per_task
);
7731 /* Amount of load we'd add */
7732 if (busiest
->avg_load
* busiest
->group_capacity
<
7733 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
7734 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
7735 local
->group_capacity
;
7737 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
7738 local
->group_capacity
;
7740 capa_move
+= local
->group_capacity
*
7741 min(local
->load_per_task
, local
->avg_load
+ tmp
);
7742 capa_move
/= SCHED_CAPACITY_SCALE
;
7744 /* Move if we gain throughput */
7745 if (capa_move
> capa_now
)
7746 env
->imbalance
= busiest
->load_per_task
;
7750 * calculate_imbalance - Calculate the amount of imbalance present within the
7751 * groups of a given sched_domain during load balance.
7752 * @env: load balance environment
7753 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7755 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
7757 unsigned long max_pull
, load_above_capacity
= ~0UL;
7758 struct sg_lb_stats
*local
, *busiest
;
7760 local
= &sds
->local_stat
;
7761 busiest
= &sds
->busiest_stat
;
7763 if (busiest
->group_type
== group_imbalanced
) {
7765 * In the group_imb case we cannot rely on group-wide averages
7766 * to ensure cpu-load equilibrium, look at wider averages. XXX
7768 busiest
->load_per_task
=
7769 min(busiest
->load_per_task
, sds
->avg_load
);
7773 * Avg load of busiest sg can be less and avg load of local sg can
7774 * be greater than avg load across all sgs of sd because avg load
7775 * factors in sg capacity and sgs with smaller group_type are
7776 * skipped when updating the busiest sg:
7778 if (busiest
->avg_load
<= sds
->avg_load
||
7779 local
->avg_load
>= sds
->avg_load
) {
7781 return fix_small_imbalance(env
, sds
);
7785 * If there aren't any idle cpus, avoid creating some.
7787 if (busiest
->group_type
== group_overloaded
&&
7788 local
->group_type
== group_overloaded
) {
7789 load_above_capacity
= busiest
->sum_nr_running
* SCHED_CAPACITY_SCALE
;
7790 if (load_above_capacity
> busiest
->group_capacity
) {
7791 load_above_capacity
-= busiest
->group_capacity
;
7792 load_above_capacity
*= scale_load_down(NICE_0_LOAD
);
7793 load_above_capacity
/= busiest
->group_capacity
;
7795 load_above_capacity
= ~0UL;
7799 * We're trying to get all the cpus to the average_load, so we don't
7800 * want to push ourselves above the average load, nor do we wish to
7801 * reduce the max loaded cpu below the average load. At the same time,
7802 * we also don't want to reduce the group load below the group
7803 * capacity. Thus we look for the minimum possible imbalance.
7805 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
7807 /* How much load to actually move to equalise the imbalance */
7808 env
->imbalance
= min(
7809 max_pull
* busiest
->group_capacity
,
7810 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
7811 ) / SCHED_CAPACITY_SCALE
;
7814 * if *imbalance is less than the average load per runnable task
7815 * there is no guarantee that any tasks will be moved so we'll have
7816 * a think about bumping its value to force at least one task to be
7819 if (env
->imbalance
< busiest
->load_per_task
)
7820 return fix_small_imbalance(env
, sds
);
7823 /******* find_busiest_group() helpers end here *********************/
7826 * find_busiest_group - Returns the busiest group within the sched_domain
7827 * if there is an imbalance.
7829 * Also calculates the amount of weighted load which should be moved
7830 * to restore balance.
7832 * @env: The load balancing environment.
7834 * Return: - The busiest group if imbalance exists.
7836 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
7838 struct sg_lb_stats
*local
, *busiest
;
7839 struct sd_lb_stats sds
;
7841 init_sd_lb_stats(&sds
);
7844 * Compute the various statistics relavent for load balancing at
7847 update_sd_lb_stats(env
, &sds
);
7848 local
= &sds
.local_stat
;
7849 busiest
= &sds
.busiest_stat
;
7851 /* ASYM feature bypasses nice load balance check */
7852 if (check_asym_packing(env
, &sds
))
7855 /* There is no busy sibling group to pull tasks from */
7856 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
7859 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
7860 / sds
.total_capacity
;
7863 * If the busiest group is imbalanced the below checks don't
7864 * work because they assume all things are equal, which typically
7865 * isn't true due to cpus_allowed constraints and the like.
7867 if (busiest
->group_type
== group_imbalanced
)
7870 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7871 if (env
->idle
== CPU_NEWLY_IDLE
&& group_has_capacity(env
, local
) &&
7872 busiest
->group_no_capacity
)
7876 * If the local group is busier than the selected busiest group
7877 * don't try and pull any tasks.
7879 if (local
->avg_load
>= busiest
->avg_load
)
7883 * Don't pull any tasks if this group is already above the domain
7886 if (local
->avg_load
>= sds
.avg_load
)
7889 if (env
->idle
== CPU_IDLE
) {
7891 * This cpu is idle. If the busiest group is not overloaded
7892 * and there is no imbalance between this and busiest group
7893 * wrt idle cpus, it is balanced. The imbalance becomes
7894 * significant if the diff is greater than 1 otherwise we
7895 * might end up to just move the imbalance on another group
7897 if ((busiest
->group_type
!= group_overloaded
) &&
7898 (local
->idle_cpus
<= (busiest
->idle_cpus
+ 1)))
7902 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7903 * imbalance_pct to be conservative.
7905 if (100 * busiest
->avg_load
<=
7906 env
->sd
->imbalance_pct
* local
->avg_load
)
7911 /* Looks like there is an imbalance. Compute it */
7912 calculate_imbalance(env
, &sds
);
7921 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7923 static struct rq
*find_busiest_queue(struct lb_env
*env
,
7924 struct sched_group
*group
)
7926 struct rq
*busiest
= NULL
, *rq
;
7927 unsigned long busiest_load
= 0, busiest_capacity
= 1;
7930 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
7931 unsigned long capacity
, wl
;
7935 rt
= fbq_classify_rq(rq
);
7938 * We classify groups/runqueues into three groups:
7939 * - regular: there are !numa tasks
7940 * - remote: there are numa tasks that run on the 'wrong' node
7941 * - all: there is no distinction
7943 * In order to avoid migrating ideally placed numa tasks,
7944 * ignore those when there's better options.
7946 * If we ignore the actual busiest queue to migrate another
7947 * task, the next balance pass can still reduce the busiest
7948 * queue by moving tasks around inside the node.
7950 * If we cannot move enough load due to this classification
7951 * the next pass will adjust the group classification and
7952 * allow migration of more tasks.
7954 * Both cases only affect the total convergence complexity.
7956 if (rt
> env
->fbq_type
)
7959 capacity
= capacity_of(i
);
7961 wl
= weighted_cpuload(i
);
7964 * When comparing with imbalance, use weighted_cpuload()
7965 * which is not scaled with the cpu capacity.
7968 if (rq
->nr_running
== 1 && wl
> env
->imbalance
&&
7969 !check_cpu_capacity(rq
, env
->sd
))
7973 * For the load comparisons with the other cpu's, consider
7974 * the weighted_cpuload() scaled with the cpu capacity, so
7975 * that the load can be moved away from the cpu that is
7976 * potentially running at a lower capacity.
7978 * Thus we're looking for max(wl_i / capacity_i), crosswise
7979 * multiplication to rid ourselves of the division works out
7980 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7981 * our previous maximum.
7983 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
7985 busiest_capacity
= capacity
;
7994 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7995 * so long as it is large enough.
7997 #define MAX_PINNED_INTERVAL 512
7999 static int need_active_balance(struct lb_env
*env
)
8001 struct sched_domain
*sd
= env
->sd
;
8003 if (env
->idle
== CPU_NEWLY_IDLE
) {
8006 * ASYM_PACKING needs to force migrate tasks from busy but
8007 * lower priority CPUs in order to pack all tasks in the
8008 * highest priority CPUs.
8010 if ((sd
->flags
& SD_ASYM_PACKING
) &&
8011 sched_asym_prefer(env
->dst_cpu
, env
->src_cpu
))
8016 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8017 * It's worth migrating the task if the src_cpu's capacity is reduced
8018 * because of other sched_class or IRQs if more capacity stays
8019 * available on dst_cpu.
8021 if ((env
->idle
!= CPU_NOT_IDLE
) &&
8022 (env
->src_rq
->cfs
.h_nr_running
== 1)) {
8023 if ((check_cpu_capacity(env
->src_rq
, sd
)) &&
8024 (capacity_of(env
->src_cpu
)*sd
->imbalance_pct
< capacity_of(env
->dst_cpu
)*100))
8028 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
8031 static int active_load_balance_cpu_stop(void *data
);
8033 static int should_we_balance(struct lb_env
*env
)
8035 struct sched_group
*sg
= env
->sd
->groups
;
8036 struct cpumask
*sg_cpus
, *sg_mask
;
8037 int cpu
, balance_cpu
= -1;
8040 * In the newly idle case, we will allow all the cpu's
8041 * to do the newly idle load balance.
8043 if (env
->idle
== CPU_NEWLY_IDLE
)
8046 sg_cpus
= sched_group_cpus(sg
);
8047 sg_mask
= sched_group_mask(sg
);
8048 /* Try to find first idle cpu */
8049 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
8050 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
8057 if (balance_cpu
== -1)
8058 balance_cpu
= group_balance_cpu(sg
);
8061 * First idle cpu or the first cpu(busiest) in this sched group
8062 * is eligible for doing load balancing at this and above domains.
8064 return balance_cpu
== env
->dst_cpu
;
8068 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8069 * tasks if there is an imbalance.
8071 static int load_balance(int this_cpu
, struct rq
*this_rq
,
8072 struct sched_domain
*sd
, enum cpu_idle_type idle
,
8073 int *continue_balancing
)
8075 int ld_moved
, cur_ld_moved
, active_balance
= 0;
8076 struct sched_domain
*sd_parent
= sd
->parent
;
8077 struct sched_group
*group
;
8080 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
8082 struct lb_env env
= {
8084 .dst_cpu
= this_cpu
,
8086 .dst_grpmask
= sched_group_cpus(sd
->groups
),
8088 .loop_break
= sched_nr_migrate_break
,
8091 .tasks
= LIST_HEAD_INIT(env
.tasks
),
8095 * For NEWLY_IDLE load_balancing, we don't need to consider
8096 * other cpus in our group
8098 if (idle
== CPU_NEWLY_IDLE
)
8099 env
.dst_grpmask
= NULL
;
8101 cpumask_copy(cpus
, cpu_active_mask
);
8103 schedstat_inc(sd
->lb_count
[idle
]);
8106 if (!should_we_balance(&env
)) {
8107 *continue_balancing
= 0;
8111 group
= find_busiest_group(&env
);
8113 schedstat_inc(sd
->lb_nobusyg
[idle
]);
8117 busiest
= find_busiest_queue(&env
, group
);
8119 schedstat_inc(sd
->lb_nobusyq
[idle
]);
8123 BUG_ON(busiest
== env
.dst_rq
);
8125 schedstat_add(sd
->lb_imbalance
[idle
], env
.imbalance
);
8127 env
.src_cpu
= busiest
->cpu
;
8128 env
.src_rq
= busiest
;
8131 if (busiest
->nr_running
> 1) {
8133 * Attempt to move tasks. If find_busiest_group has found
8134 * an imbalance but busiest->nr_running <= 1, the group is
8135 * still unbalanced. ld_moved simply stays zero, so it is
8136 * correctly treated as an imbalance.
8138 env
.flags
|= LBF_ALL_PINNED
;
8139 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
8142 rq_lock_irqsave(busiest
, &rf
);
8143 update_rq_clock(busiest
);
8146 * cur_ld_moved - load moved in current iteration
8147 * ld_moved - cumulative load moved across iterations
8149 cur_ld_moved
= detach_tasks(&env
);
8152 * We've detached some tasks from busiest_rq. Every
8153 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8154 * unlock busiest->lock, and we are able to be sure
8155 * that nobody can manipulate the tasks in parallel.
8156 * See task_rq_lock() family for the details.
8159 rq_unlock(busiest
, &rf
);
8163 ld_moved
+= cur_ld_moved
;
8166 local_irq_restore(rf
.flags
);
8168 if (env
.flags
& LBF_NEED_BREAK
) {
8169 env
.flags
&= ~LBF_NEED_BREAK
;
8174 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8175 * us and move them to an alternate dst_cpu in our sched_group
8176 * where they can run. The upper limit on how many times we
8177 * iterate on same src_cpu is dependent on number of cpus in our
8180 * This changes load balance semantics a bit on who can move
8181 * load to a given_cpu. In addition to the given_cpu itself
8182 * (or a ilb_cpu acting on its behalf where given_cpu is
8183 * nohz-idle), we now have balance_cpu in a position to move
8184 * load to given_cpu. In rare situations, this may cause
8185 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8186 * _independently_ and at _same_ time to move some load to
8187 * given_cpu) causing exceess load to be moved to given_cpu.
8188 * This however should not happen so much in practice and
8189 * moreover subsequent load balance cycles should correct the
8190 * excess load moved.
8192 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
8194 /* Prevent to re-select dst_cpu via env's cpus */
8195 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
8197 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
8198 env
.dst_cpu
= env
.new_dst_cpu
;
8199 env
.flags
&= ~LBF_DST_PINNED
;
8201 env
.loop_break
= sched_nr_migrate_break
;
8204 * Go back to "more_balance" rather than "redo" since we
8205 * need to continue with same src_cpu.
8211 * We failed to reach balance because of affinity.
8214 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
8216 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
8217 *group_imbalance
= 1;
8220 /* All tasks on this runqueue were pinned by CPU affinity */
8221 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
8222 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
8223 if (!cpumask_empty(cpus
)) {
8225 env
.loop_break
= sched_nr_migrate_break
;
8228 goto out_all_pinned
;
8233 schedstat_inc(sd
->lb_failed
[idle
]);
8235 * Increment the failure counter only on periodic balance.
8236 * We do not want newidle balance, which can be very
8237 * frequent, pollute the failure counter causing
8238 * excessive cache_hot migrations and active balances.
8240 if (idle
!= CPU_NEWLY_IDLE
)
8241 sd
->nr_balance_failed
++;
8243 if (need_active_balance(&env
)) {
8244 unsigned long flags
;
8246 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
8248 /* don't kick the active_load_balance_cpu_stop,
8249 * if the curr task on busiest cpu can't be
8252 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
8253 raw_spin_unlock_irqrestore(&busiest
->lock
,
8255 env
.flags
|= LBF_ALL_PINNED
;
8256 goto out_one_pinned
;
8260 * ->active_balance synchronizes accesses to
8261 * ->active_balance_work. Once set, it's cleared
8262 * only after active load balance is finished.
8264 if (!busiest
->active_balance
) {
8265 busiest
->active_balance
= 1;
8266 busiest
->push_cpu
= this_cpu
;
8269 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
8271 if (active_balance
) {
8272 stop_one_cpu_nowait(cpu_of(busiest
),
8273 active_load_balance_cpu_stop
, busiest
,
8274 &busiest
->active_balance_work
);
8277 /* We've kicked active balancing, force task migration. */
8278 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
8281 sd
->nr_balance_failed
= 0;
8283 if (likely(!active_balance
)) {
8284 /* We were unbalanced, so reset the balancing interval */
8285 sd
->balance_interval
= sd
->min_interval
;
8288 * If we've begun active balancing, start to back off. This
8289 * case may not be covered by the all_pinned logic if there
8290 * is only 1 task on the busy runqueue (because we don't call
8293 if (sd
->balance_interval
< sd
->max_interval
)
8294 sd
->balance_interval
*= 2;
8301 * We reach balance although we may have faced some affinity
8302 * constraints. Clear the imbalance flag if it was set.
8305 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
8307 if (*group_imbalance
)
8308 *group_imbalance
= 0;
8313 * We reach balance because all tasks are pinned at this level so
8314 * we can't migrate them. Let the imbalance flag set so parent level
8315 * can try to migrate them.
8317 schedstat_inc(sd
->lb_balanced
[idle
]);
8319 sd
->nr_balance_failed
= 0;
8322 /* tune up the balancing interval */
8323 if (((env
.flags
& LBF_ALL_PINNED
) &&
8324 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
8325 (sd
->balance_interval
< sd
->max_interval
))
8326 sd
->balance_interval
*= 2;
8333 static inline unsigned long
8334 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
8336 unsigned long interval
= sd
->balance_interval
;
8339 interval
*= sd
->busy_factor
;
8341 /* scale ms to jiffies */
8342 interval
= msecs_to_jiffies(interval
);
8343 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
8349 update_next_balance(struct sched_domain
*sd
, unsigned long *next_balance
)
8351 unsigned long interval
, next
;
8353 /* used by idle balance, so cpu_busy = 0 */
8354 interval
= get_sd_balance_interval(sd
, 0);
8355 next
= sd
->last_balance
+ interval
;
8357 if (time_after(*next_balance
, next
))
8358 *next_balance
= next
;
8362 * idle_balance is called by schedule() if this_cpu is about to become
8363 * idle. Attempts to pull tasks from other CPUs.
8365 static int idle_balance(struct rq
*this_rq
, struct rq_flags
*rf
)
8367 unsigned long next_balance
= jiffies
+ HZ
;
8368 int this_cpu
= this_rq
->cpu
;
8369 struct sched_domain
*sd
;
8370 int pulled_task
= 0;
8374 * We must set idle_stamp _before_ calling idle_balance(), such that we
8375 * measure the duration of idle_balance() as idle time.
8377 this_rq
->idle_stamp
= rq_clock(this_rq
);
8380 * This is OK, because current is on_cpu, which avoids it being picked
8381 * for load-balance and preemption/IRQs are still disabled avoiding
8382 * further scheduler activity on it and we're being very careful to
8383 * re-start the picking loop.
8385 rq_unpin_lock(this_rq
, rf
);
8387 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
8388 !this_rq
->rd
->overload
) {
8390 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
8392 update_next_balance(sd
, &next_balance
);
8398 raw_spin_unlock(&this_rq
->lock
);
8400 update_blocked_averages(this_cpu
);
8402 for_each_domain(this_cpu
, sd
) {
8403 int continue_balancing
= 1;
8404 u64 t0
, domain_cost
;
8406 if (!(sd
->flags
& SD_LOAD_BALANCE
))
8409 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
8410 update_next_balance(sd
, &next_balance
);
8414 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
8415 t0
= sched_clock_cpu(this_cpu
);
8417 pulled_task
= load_balance(this_cpu
, this_rq
,
8419 &continue_balancing
);
8421 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
8422 if (domain_cost
> sd
->max_newidle_lb_cost
)
8423 sd
->max_newidle_lb_cost
= domain_cost
;
8425 curr_cost
+= domain_cost
;
8428 update_next_balance(sd
, &next_balance
);
8431 * Stop searching for tasks to pull if there are
8432 * now runnable tasks on this rq.
8434 if (pulled_task
|| this_rq
->nr_running
> 0)
8439 raw_spin_lock(&this_rq
->lock
);
8441 if (curr_cost
> this_rq
->max_idle_balance_cost
)
8442 this_rq
->max_idle_balance_cost
= curr_cost
;
8445 * While browsing the domains, we released the rq lock, a task could
8446 * have been enqueued in the meantime. Since we're not going idle,
8447 * pretend we pulled a task.
8449 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
8453 /* Move the next balance forward */
8454 if (time_after(this_rq
->next_balance
, next_balance
))
8455 this_rq
->next_balance
= next_balance
;
8457 /* Is there a task of a high priority class? */
8458 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
8462 this_rq
->idle_stamp
= 0;
8464 rq_repin_lock(this_rq
, rf
);
8470 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8471 * running tasks off the busiest CPU onto idle CPUs. It requires at
8472 * least 1 task to be running on each physical CPU where possible, and
8473 * avoids physical / logical imbalances.
8475 static int active_load_balance_cpu_stop(void *data
)
8477 struct rq
*busiest_rq
= data
;
8478 int busiest_cpu
= cpu_of(busiest_rq
);
8479 int target_cpu
= busiest_rq
->push_cpu
;
8480 struct rq
*target_rq
= cpu_rq(target_cpu
);
8481 struct sched_domain
*sd
;
8482 struct task_struct
*p
= NULL
;
8485 rq_lock_irq(busiest_rq
, &rf
);
8487 /* make sure the requested cpu hasn't gone down in the meantime */
8488 if (unlikely(busiest_cpu
!= smp_processor_id() ||
8489 !busiest_rq
->active_balance
))
8492 /* Is there any task to move? */
8493 if (busiest_rq
->nr_running
<= 1)
8497 * This condition is "impossible", if it occurs
8498 * we need to fix it. Originally reported by
8499 * Bjorn Helgaas on a 128-cpu setup.
8501 BUG_ON(busiest_rq
== target_rq
);
8503 /* Search for an sd spanning us and the target CPU. */
8505 for_each_domain(target_cpu
, sd
) {
8506 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
8507 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
8512 struct lb_env env
= {
8514 .dst_cpu
= target_cpu
,
8515 .dst_rq
= target_rq
,
8516 .src_cpu
= busiest_rq
->cpu
,
8517 .src_rq
= busiest_rq
,
8521 schedstat_inc(sd
->alb_count
);
8522 update_rq_clock(busiest_rq
);
8524 p
= detach_one_task(&env
);
8526 schedstat_inc(sd
->alb_pushed
);
8527 /* Active balancing done, reset the failure counter. */
8528 sd
->nr_balance_failed
= 0;
8530 schedstat_inc(sd
->alb_failed
);
8535 busiest_rq
->active_balance
= 0;
8536 rq_unlock(busiest_rq
, &rf
);
8539 attach_one_task(target_rq
, p
);
8546 static inline int on_null_domain(struct rq
*rq
)
8548 return unlikely(!rcu_dereference_sched(rq
->sd
));
8551 #ifdef CONFIG_NO_HZ_COMMON
8553 * idle load balancing details
8554 * - When one of the busy CPUs notice that there may be an idle rebalancing
8555 * needed, they will kick the idle load balancer, which then does idle
8556 * load balancing for all the idle CPUs.
8559 cpumask_var_t idle_cpus_mask
;
8561 unsigned long next_balance
; /* in jiffy units */
8562 } nohz ____cacheline_aligned
;
8564 static inline int find_new_ilb(void)
8566 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
8568 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
8575 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8576 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8577 * CPU (if there is one).
8579 static void nohz_balancer_kick(void)
8583 nohz
.next_balance
++;
8585 ilb_cpu
= find_new_ilb();
8587 if (ilb_cpu
>= nr_cpu_ids
)
8590 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
8593 * Use smp_send_reschedule() instead of resched_cpu().
8594 * This way we generate a sched IPI on the target cpu which
8595 * is idle. And the softirq performing nohz idle load balance
8596 * will be run before returning from the IPI.
8598 smp_send_reschedule(ilb_cpu
);
8602 void nohz_balance_exit_idle(unsigned int cpu
)
8604 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
8606 * Completely isolated CPUs don't ever set, so we must test.
8608 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
8609 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
8610 atomic_dec(&nohz
.nr_cpus
);
8612 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
8616 static inline void set_cpu_sd_state_busy(void)
8618 struct sched_domain
*sd
;
8619 int cpu
= smp_processor_id();
8622 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
8624 if (!sd
|| !sd
->nohz_idle
)
8628 atomic_inc(&sd
->shared
->nr_busy_cpus
);
8633 void set_cpu_sd_state_idle(void)
8635 struct sched_domain
*sd
;
8636 int cpu
= smp_processor_id();
8639 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
8641 if (!sd
|| sd
->nohz_idle
)
8645 atomic_dec(&sd
->shared
->nr_busy_cpus
);
8651 * This routine will record that the cpu is going idle with tick stopped.
8652 * This info will be used in performing idle load balancing in the future.
8654 void nohz_balance_enter_idle(int cpu
)
8657 * If this cpu is going down, then nothing needs to be done.
8659 if (!cpu_active(cpu
))
8662 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
8666 * If we're a completely isolated CPU, we don't play.
8668 if (on_null_domain(cpu_rq(cpu
)))
8671 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
8672 atomic_inc(&nohz
.nr_cpus
);
8673 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
8677 static DEFINE_SPINLOCK(balancing
);
8680 * Scale the max load_balance interval with the number of CPUs in the system.
8681 * This trades load-balance latency on larger machines for less cross talk.
8683 void update_max_interval(void)
8685 max_load_balance_interval
= HZ
*num_online_cpus()/10;
8689 * It checks each scheduling domain to see if it is due to be balanced,
8690 * and initiates a balancing operation if so.
8692 * Balancing parameters are set up in init_sched_domains.
8694 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
8696 int continue_balancing
= 1;
8698 unsigned long interval
;
8699 struct sched_domain
*sd
;
8700 /* Earliest time when we have to do rebalance again */
8701 unsigned long next_balance
= jiffies
+ 60*HZ
;
8702 int update_next_balance
= 0;
8703 int need_serialize
, need_decay
= 0;
8706 update_blocked_averages(cpu
);
8709 for_each_domain(cpu
, sd
) {
8711 * Decay the newidle max times here because this is a regular
8712 * visit to all the domains. Decay ~1% per second.
8714 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
8715 sd
->max_newidle_lb_cost
=
8716 (sd
->max_newidle_lb_cost
* 253) / 256;
8717 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
8720 max_cost
+= sd
->max_newidle_lb_cost
;
8722 if (!(sd
->flags
& SD_LOAD_BALANCE
))
8726 * Stop the load balance at this level. There is another
8727 * CPU in our sched group which is doing load balancing more
8730 if (!continue_balancing
) {
8736 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
8738 need_serialize
= sd
->flags
& SD_SERIALIZE
;
8739 if (need_serialize
) {
8740 if (!spin_trylock(&balancing
))
8744 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
8745 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
8747 * The LBF_DST_PINNED logic could have changed
8748 * env->dst_cpu, so we can't know our idle
8749 * state even if we migrated tasks. Update it.
8751 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
8753 sd
->last_balance
= jiffies
;
8754 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
8757 spin_unlock(&balancing
);
8759 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
8760 next_balance
= sd
->last_balance
+ interval
;
8761 update_next_balance
= 1;
8766 * Ensure the rq-wide value also decays but keep it at a
8767 * reasonable floor to avoid funnies with rq->avg_idle.
8769 rq
->max_idle_balance_cost
=
8770 max((u64
)sysctl_sched_migration_cost
, max_cost
);
8775 * next_balance will be updated only when there is a need.
8776 * When the cpu is attached to null domain for ex, it will not be
8779 if (likely(update_next_balance
)) {
8780 rq
->next_balance
= next_balance
;
8782 #ifdef CONFIG_NO_HZ_COMMON
8784 * If this CPU has been elected to perform the nohz idle
8785 * balance. Other idle CPUs have already rebalanced with
8786 * nohz_idle_balance() and nohz.next_balance has been
8787 * updated accordingly. This CPU is now running the idle load
8788 * balance for itself and we need to update the
8789 * nohz.next_balance accordingly.
8791 if ((idle
== CPU_IDLE
) && time_after(nohz
.next_balance
, rq
->next_balance
))
8792 nohz
.next_balance
= rq
->next_balance
;
8797 #ifdef CONFIG_NO_HZ_COMMON
8799 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8800 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8802 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
8804 int this_cpu
= this_rq
->cpu
;
8807 /* Earliest time when we have to do rebalance again */
8808 unsigned long next_balance
= jiffies
+ 60*HZ
;
8809 int update_next_balance
= 0;
8811 if (idle
!= CPU_IDLE
||
8812 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
8815 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
8816 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
8820 * If this cpu gets work to do, stop the load balancing
8821 * work being done for other cpus. Next load
8822 * balancing owner will pick it up.
8827 rq
= cpu_rq(balance_cpu
);
8830 * If time for next balance is due,
8833 if (time_after_eq(jiffies
, rq
->next_balance
)) {
8836 rq_lock_irq(rq
, &rf
);
8837 update_rq_clock(rq
);
8838 cpu_load_update_idle(rq
);
8839 rq_unlock_irq(rq
, &rf
);
8841 rebalance_domains(rq
, CPU_IDLE
);
8844 if (time_after(next_balance
, rq
->next_balance
)) {
8845 next_balance
= rq
->next_balance
;
8846 update_next_balance
= 1;
8851 * next_balance will be updated only when there is a need.
8852 * When the CPU is attached to null domain for ex, it will not be
8855 if (likely(update_next_balance
))
8856 nohz
.next_balance
= next_balance
;
8858 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
8862 * Current heuristic for kicking the idle load balancer in the presence
8863 * of an idle cpu in the system.
8864 * - This rq has more than one task.
8865 * - This rq has at least one CFS task and the capacity of the CPU is
8866 * significantly reduced because of RT tasks or IRQs.
8867 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8868 * multiple busy cpu.
8869 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8870 * domain span are idle.
8872 static inline bool nohz_kick_needed(struct rq
*rq
)
8874 unsigned long now
= jiffies
;
8875 struct sched_domain_shared
*sds
;
8876 struct sched_domain
*sd
;
8877 int nr_busy
, i
, cpu
= rq
->cpu
;
8880 if (unlikely(rq
->idle_balance
))
8884 * We may be recently in ticked or tickless idle mode. At the first
8885 * busy tick after returning from idle, we will update the busy stats.
8887 set_cpu_sd_state_busy();
8888 nohz_balance_exit_idle(cpu
);
8891 * None are in tickless mode and hence no need for NOHZ idle load
8894 if (likely(!atomic_read(&nohz
.nr_cpus
)))
8897 if (time_before(now
, nohz
.next_balance
))
8900 if (rq
->nr_running
>= 2)
8904 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
8907 * XXX: write a coherent comment on why we do this.
8908 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
8910 nr_busy
= atomic_read(&sds
->nr_busy_cpus
);
8918 sd
= rcu_dereference(rq
->sd
);
8920 if ((rq
->cfs
.h_nr_running
>= 1) &&
8921 check_cpu_capacity(rq
, sd
)) {
8927 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
8929 for_each_cpu(i
, sched_domain_span(sd
)) {
8931 !cpumask_test_cpu(i
, nohz
.idle_cpus_mask
))
8934 if (sched_asym_prefer(i
, cpu
)) {
8945 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
8949 * run_rebalance_domains is triggered when needed from the scheduler tick.
8950 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8952 static __latent_entropy
void run_rebalance_domains(struct softirq_action
*h
)
8954 struct rq
*this_rq
= this_rq();
8955 enum cpu_idle_type idle
= this_rq
->idle_balance
?
8956 CPU_IDLE
: CPU_NOT_IDLE
;
8959 * If this cpu has a pending nohz_balance_kick, then do the
8960 * balancing on behalf of the other idle cpus whose ticks are
8961 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8962 * give the idle cpus a chance to load balance. Else we may
8963 * load balance only within the local sched_domain hierarchy
8964 * and abort nohz_idle_balance altogether if we pull some load.
8966 nohz_idle_balance(this_rq
, idle
);
8967 rebalance_domains(this_rq
, idle
);
8971 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8973 void trigger_load_balance(struct rq
*rq
)
8975 /* Don't need to rebalance while attached to NULL domain */
8976 if (unlikely(on_null_domain(rq
)))
8979 if (time_after_eq(jiffies
, rq
->next_balance
))
8980 raise_softirq(SCHED_SOFTIRQ
);
8981 #ifdef CONFIG_NO_HZ_COMMON
8982 if (nohz_kick_needed(rq
))
8983 nohz_balancer_kick();
8987 static void rq_online_fair(struct rq
*rq
)
8991 update_runtime_enabled(rq
);
8994 static void rq_offline_fair(struct rq
*rq
)
8998 /* Ensure any throttled groups are reachable by pick_next_task */
8999 unthrottle_offline_cfs_rqs(rq
);
9002 #endif /* CONFIG_SMP */
9005 * scheduler tick hitting a task of our scheduling class:
9007 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
9009 struct cfs_rq
*cfs_rq
;
9010 struct sched_entity
*se
= &curr
->se
;
9012 for_each_sched_entity(se
) {
9013 cfs_rq
= cfs_rq_of(se
);
9014 entity_tick(cfs_rq
, se
, queued
);
9017 if (static_branch_unlikely(&sched_numa_balancing
))
9018 task_tick_numa(rq
, curr
);
9022 * called on fork with the child task as argument from the parent's context
9023 * - child not yet on the tasklist
9024 * - preemption disabled
9026 static void task_fork_fair(struct task_struct
*p
)
9028 struct cfs_rq
*cfs_rq
;
9029 struct sched_entity
*se
= &p
->se
, *curr
;
9030 struct rq
*rq
= this_rq();
9034 update_rq_clock(rq
);
9036 cfs_rq
= task_cfs_rq(current
);
9037 curr
= cfs_rq
->curr
;
9039 update_curr(cfs_rq
);
9040 se
->vruntime
= curr
->vruntime
;
9042 place_entity(cfs_rq
, se
, 1);
9044 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
9046 * Upon rescheduling, sched_class::put_prev_task() will place
9047 * 'current' within the tree based on its new key value.
9049 swap(curr
->vruntime
, se
->vruntime
);
9053 se
->vruntime
-= cfs_rq
->min_vruntime
;
9058 * Priority of the task has changed. Check to see if we preempt
9062 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
9064 if (!task_on_rq_queued(p
))
9068 * Reschedule if we are currently running on this runqueue and
9069 * our priority decreased, or if we are not currently running on
9070 * this runqueue and our priority is higher than the current's
9072 if (rq
->curr
== p
) {
9073 if (p
->prio
> oldprio
)
9076 check_preempt_curr(rq
, p
, 0);
9079 static inline bool vruntime_normalized(struct task_struct
*p
)
9081 struct sched_entity
*se
= &p
->se
;
9084 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9085 * the dequeue_entity(.flags=0) will already have normalized the
9092 * When !on_rq, vruntime of the task has usually NOT been normalized.
9093 * But there are some cases where it has already been normalized:
9095 * - A forked child which is waiting for being woken up by
9096 * wake_up_new_task().
9097 * - A task which has been woken up by try_to_wake_up() and
9098 * waiting for actually being woken up by sched_ttwu_pending().
9100 if (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
)
9106 #ifdef CONFIG_FAIR_GROUP_SCHED
9108 * Propagate the changes of the sched_entity across the tg tree to make it
9109 * visible to the root
9111 static void propagate_entity_cfs_rq(struct sched_entity
*se
)
9113 struct cfs_rq
*cfs_rq
;
9115 /* Start to propagate at parent */
9118 for_each_sched_entity(se
) {
9119 cfs_rq
= cfs_rq_of(se
);
9121 if (cfs_rq_throttled(cfs_rq
))
9124 update_load_avg(se
, UPDATE_TG
);
9128 static void propagate_entity_cfs_rq(struct sched_entity
*se
) { }
9131 static void detach_entity_cfs_rq(struct sched_entity
*se
)
9133 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9135 /* Catch up with the cfs_rq and remove our load when we leave */
9136 update_load_avg(se
, 0);
9137 detach_entity_load_avg(cfs_rq
, se
);
9138 update_tg_load_avg(cfs_rq
, false);
9139 propagate_entity_cfs_rq(se
);
9142 static void attach_entity_cfs_rq(struct sched_entity
*se
)
9144 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9146 #ifdef CONFIG_FAIR_GROUP_SCHED
9148 * Since the real-depth could have been changed (only FAIR
9149 * class maintain depth value), reset depth properly.
9151 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
9154 /* Synchronize entity with its cfs_rq */
9155 update_load_avg(se
, sched_feat(ATTACH_AGE_LOAD
) ? 0 : SKIP_AGE_LOAD
);
9156 attach_entity_load_avg(cfs_rq
, se
);
9157 update_tg_load_avg(cfs_rq
, false);
9158 propagate_entity_cfs_rq(se
);
9161 static void detach_task_cfs_rq(struct task_struct
*p
)
9163 struct sched_entity
*se
= &p
->se
;
9164 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9166 if (!vruntime_normalized(p
)) {
9168 * Fix up our vruntime so that the current sleep doesn't
9169 * cause 'unlimited' sleep bonus.
9171 place_entity(cfs_rq
, se
, 0);
9172 se
->vruntime
-= cfs_rq
->min_vruntime
;
9175 detach_entity_cfs_rq(se
);
9178 static void attach_task_cfs_rq(struct task_struct
*p
)
9180 struct sched_entity
*se
= &p
->se
;
9181 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9183 attach_entity_cfs_rq(se
);
9185 if (!vruntime_normalized(p
))
9186 se
->vruntime
+= cfs_rq
->min_vruntime
;
9189 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
9191 detach_task_cfs_rq(p
);
9194 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
9196 attach_task_cfs_rq(p
);
9198 if (task_on_rq_queued(p
)) {
9200 * We were most likely switched from sched_rt, so
9201 * kick off the schedule if running, otherwise just see
9202 * if we can still preempt the current task.
9207 check_preempt_curr(rq
, p
, 0);
9211 /* Account for a task changing its policy or group.
9213 * This routine is mostly called to set cfs_rq->curr field when a task
9214 * migrates between groups/classes.
9216 static void set_curr_task_fair(struct rq
*rq
)
9218 struct sched_entity
*se
= &rq
->curr
->se
;
9220 for_each_sched_entity(se
) {
9221 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9223 set_next_entity(cfs_rq
, se
);
9224 /* ensure bandwidth has been allocated on our new cfs_rq */
9225 account_cfs_rq_runtime(cfs_rq
, 0);
9229 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
9231 cfs_rq
->tasks_timeline
= RB_ROOT
;
9232 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9233 #ifndef CONFIG_64BIT
9234 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
9237 #ifdef CONFIG_FAIR_GROUP_SCHED
9238 cfs_rq
->propagate_avg
= 0;
9240 atomic_long_set(&cfs_rq
->removed_load_avg
, 0);
9241 atomic_long_set(&cfs_rq
->removed_util_avg
, 0);
9245 #ifdef CONFIG_FAIR_GROUP_SCHED
9246 static void task_set_group_fair(struct task_struct
*p
)
9248 struct sched_entity
*se
= &p
->se
;
9250 set_task_rq(p
, task_cpu(p
));
9251 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
9254 static void task_move_group_fair(struct task_struct
*p
)
9256 detach_task_cfs_rq(p
);
9257 set_task_rq(p
, task_cpu(p
));
9260 /* Tell se's cfs_rq has been changed -- migrated */
9261 p
->se
.avg
.last_update_time
= 0;
9263 attach_task_cfs_rq(p
);
9266 static void task_change_group_fair(struct task_struct
*p
, int type
)
9269 case TASK_SET_GROUP
:
9270 task_set_group_fair(p
);
9273 case TASK_MOVE_GROUP
:
9274 task_move_group_fair(p
);
9279 void free_fair_sched_group(struct task_group
*tg
)
9283 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
9285 for_each_possible_cpu(i
) {
9287 kfree(tg
->cfs_rq
[i
]);
9296 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9298 struct sched_entity
*se
;
9299 struct cfs_rq
*cfs_rq
;
9302 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9305 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9309 tg
->shares
= NICE_0_LOAD
;
9311 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
9313 for_each_possible_cpu(i
) {
9314 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9315 GFP_KERNEL
, cpu_to_node(i
));
9319 se
= kzalloc_node(sizeof(struct sched_entity
),
9320 GFP_KERNEL
, cpu_to_node(i
));
9324 init_cfs_rq(cfs_rq
);
9325 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
9326 init_entity_runnable_average(se
);
9337 void online_fair_sched_group(struct task_group
*tg
)
9339 struct sched_entity
*se
;
9343 for_each_possible_cpu(i
) {
9347 raw_spin_lock_irq(&rq
->lock
);
9348 update_rq_clock(rq
);
9349 attach_entity_cfs_rq(se
);
9350 sync_throttle(tg
, i
);
9351 raw_spin_unlock_irq(&rq
->lock
);
9355 void unregister_fair_sched_group(struct task_group
*tg
)
9357 unsigned long flags
;
9361 for_each_possible_cpu(cpu
) {
9363 remove_entity_load_avg(tg
->se
[cpu
]);
9366 * Only empty task groups can be destroyed; so we can speculatively
9367 * check on_list without danger of it being re-added.
9369 if (!tg
->cfs_rq
[cpu
]->on_list
)
9374 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9375 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
9376 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9380 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9381 struct sched_entity
*se
, int cpu
,
9382 struct sched_entity
*parent
)
9384 struct rq
*rq
= cpu_rq(cpu
);
9388 init_cfs_rq_runtime(cfs_rq
);
9390 tg
->cfs_rq
[cpu
] = cfs_rq
;
9393 /* se could be NULL for root_task_group */
9398 se
->cfs_rq
= &rq
->cfs
;
9401 se
->cfs_rq
= parent
->my_q
;
9402 se
->depth
= parent
->depth
+ 1;
9406 /* guarantee group entities always have weight */
9407 update_load_set(&se
->load
, NICE_0_LOAD
);
9408 se
->parent
= parent
;
9411 static DEFINE_MUTEX(shares_mutex
);
9413 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9418 * We can't change the weight of the root cgroup.
9423 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
9425 mutex_lock(&shares_mutex
);
9426 if (tg
->shares
== shares
)
9429 tg
->shares
= shares
;
9430 for_each_possible_cpu(i
) {
9431 struct rq
*rq
= cpu_rq(i
);
9432 struct sched_entity
*se
= tg
->se
[i
];
9435 /* Propagate contribution to hierarchy */
9436 rq_lock_irqsave(rq
, &rf
);
9437 update_rq_clock(rq
);
9438 for_each_sched_entity(se
) {
9439 update_load_avg(se
, UPDATE_TG
);
9440 update_cfs_shares(se
);
9442 rq_unlock_irqrestore(rq
, &rf
);
9446 mutex_unlock(&shares_mutex
);
9449 #else /* CONFIG_FAIR_GROUP_SCHED */
9451 void free_fair_sched_group(struct task_group
*tg
) { }
9453 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9458 void online_fair_sched_group(struct task_group
*tg
) { }
9460 void unregister_fair_sched_group(struct task_group
*tg
) { }
9462 #endif /* CONFIG_FAIR_GROUP_SCHED */
9465 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
9467 struct sched_entity
*se
= &task
->se
;
9468 unsigned int rr_interval
= 0;
9471 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9474 if (rq
->cfs
.load
.weight
)
9475 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
9481 * All the scheduling class methods:
9483 const struct sched_class fair_sched_class
= {
9484 .next
= &idle_sched_class
,
9485 .enqueue_task
= enqueue_task_fair
,
9486 .dequeue_task
= dequeue_task_fair
,
9487 .yield_task
= yield_task_fair
,
9488 .yield_to_task
= yield_to_task_fair
,
9490 .check_preempt_curr
= check_preempt_wakeup
,
9492 .pick_next_task
= pick_next_task_fair
,
9493 .put_prev_task
= put_prev_task_fair
,
9496 .select_task_rq
= select_task_rq_fair
,
9497 .migrate_task_rq
= migrate_task_rq_fair
,
9499 .rq_online
= rq_online_fair
,
9500 .rq_offline
= rq_offline_fair
,
9502 .task_dead
= task_dead_fair
,
9503 .set_cpus_allowed
= set_cpus_allowed_common
,
9506 .set_curr_task
= set_curr_task_fair
,
9507 .task_tick
= task_tick_fair
,
9508 .task_fork
= task_fork_fair
,
9510 .prio_changed
= prio_changed_fair
,
9511 .switched_from
= switched_from_fair
,
9512 .switched_to
= switched_to_fair
,
9514 .get_rr_interval
= get_rr_interval_fair
,
9516 .update_curr
= update_curr_fair
,
9518 #ifdef CONFIG_FAIR_GROUP_SCHED
9519 .task_change_group
= task_change_group_fair
,
9523 #ifdef CONFIG_SCHED_DEBUG
9524 void print_cfs_stats(struct seq_file
*m
, int cpu
)
9526 struct cfs_rq
*cfs_rq
;
9529 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
9530 print_cfs_rq(m
, cpu
, cfs_rq
);
9534 #ifdef CONFIG_NUMA_BALANCING
9535 void show_numa_stats(struct task_struct
*p
, struct seq_file
*m
)
9538 unsigned long tsf
= 0, tpf
= 0, gsf
= 0, gpf
= 0;
9540 for_each_online_node(node
) {
9541 if (p
->numa_faults
) {
9542 tsf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
9543 tpf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
9545 if (p
->numa_group
) {
9546 gsf
= p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)],
9547 gpf
= p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
9549 print_numa_stats(m
, node
, tsf
, tpf
, gsf
, gpf
);
9552 #endif /* CONFIG_NUMA_BALANCING */
9553 #endif /* CONFIG_SCHED_DEBUG */
9555 __init
void init_sched_fair_class(void)
9558 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9560 #ifdef CONFIG_NO_HZ_COMMON
9561 nohz
.next_balance
= jiffies
;
9562 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);