2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 #include <linux/sched/mm.h>
24 #include <linux/sched/topology.h>
26 #include <linux/latencytop.h>
27 #include <linux/cpumask.h>
28 #include <linux/cpuidle.h>
29 #include <linux/slab.h>
30 #include <linux/profile.h>
31 #include <linux/interrupt.h>
32 #include <linux/mempolicy.h>
33 #include <linux/migrate.h>
34 #include <linux/task_work.h>
36 #include <trace/events/sched.h>
41 * Targeted preemption latency for CPU-bound tasks:
43 * NOTE: this latency value is not the same as the concept of
44 * 'timeslice length' - timeslices in CFS are of variable length
45 * and have no persistent notion like in traditional, time-slice
46 * based scheduling concepts.
48 * (to see the precise effective timeslice length of your workload,
49 * run vmstat and monitor the context-switches (cs) field)
51 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
53 unsigned int sysctl_sched_latency
= 6000000ULL;
54 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
57 * The initial- and re-scaling of tunables is configurable
61 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
62 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
63 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
65 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
67 enum sched_tunable_scaling sysctl_sched_tunable_scaling
= SCHED_TUNABLESCALING_LOG
;
70 * Minimal preemption granularity for CPU-bound tasks:
72 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
74 unsigned int sysctl_sched_min_granularity
= 750000ULL;
75 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
78 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
80 static unsigned int sched_nr_latency
= 8;
83 * After fork, child runs first. If set to 0 (default) then
84 * parent will (try to) run first.
86 unsigned int sysctl_sched_child_runs_first __read_mostly
;
89 * SCHED_OTHER wake-up granularity.
91 * This option delays the preemption effects of decoupled workloads
92 * and reduces their over-scheduling. Synchronous workloads will still
93 * have immediate wakeup/sleep latencies.
95 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
97 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
98 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
100 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
104 * For asym packing, by default the lower numbered cpu has higher priority.
106 int __weak
arch_asym_cpu_priority(int cpu
)
112 #ifdef CONFIG_CFS_BANDWIDTH
114 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
115 * each time a cfs_rq requests quota.
117 * Note: in the case that the slice exceeds the runtime remaining (either due
118 * to consumption or the quota being specified to be smaller than the slice)
119 * we will always only issue the remaining available time.
121 * (default: 5 msec, units: microseconds)
123 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
127 * The margin used when comparing utilization with CPU capacity:
128 * util * margin < capacity * 1024
132 unsigned int capacity_margin
= 1280;
134 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
140 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
146 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
153 * Increase the granularity value when there are more CPUs,
154 * because with more CPUs the 'effective latency' as visible
155 * to users decreases. But the relationship is not linear,
156 * so pick a second-best guess by going with the log2 of the
159 * This idea comes from the SD scheduler of Con Kolivas:
161 static unsigned int get_update_sysctl_factor(void)
163 unsigned int cpus
= min_t(unsigned int, num_online_cpus(), 8);
166 switch (sysctl_sched_tunable_scaling
) {
167 case SCHED_TUNABLESCALING_NONE
:
170 case SCHED_TUNABLESCALING_LINEAR
:
173 case SCHED_TUNABLESCALING_LOG
:
175 factor
= 1 + ilog2(cpus
);
182 static void update_sysctl(void)
184 unsigned int factor
= get_update_sysctl_factor();
186 #define SET_SYSCTL(name) \
187 (sysctl_##name = (factor) * normalized_sysctl_##name)
188 SET_SYSCTL(sched_min_granularity
);
189 SET_SYSCTL(sched_latency
);
190 SET_SYSCTL(sched_wakeup_granularity
);
194 void sched_init_granularity(void)
199 #define WMULT_CONST (~0U)
200 #define WMULT_SHIFT 32
202 static void __update_inv_weight(struct load_weight
*lw
)
206 if (likely(lw
->inv_weight
))
209 w
= scale_load_down(lw
->weight
);
211 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
213 else if (unlikely(!w
))
214 lw
->inv_weight
= WMULT_CONST
;
216 lw
->inv_weight
= WMULT_CONST
/ w
;
220 * delta_exec * weight / lw.weight
222 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
224 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
225 * we're guaranteed shift stays positive because inv_weight is guaranteed to
226 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
228 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
229 * weight/lw.weight <= 1, and therefore our shift will also be positive.
231 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
233 u64 fact
= scale_load_down(weight
);
234 int shift
= WMULT_SHIFT
;
236 __update_inv_weight(lw
);
238 if (unlikely(fact
>> 32)) {
245 /* hint to use a 32x32->64 mul */
246 fact
= (u64
)(u32
)fact
* lw
->inv_weight
;
253 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
257 const struct sched_class fair_sched_class
;
259 /**************************************************************
260 * CFS operations on generic schedulable entities:
263 #ifdef CONFIG_FAIR_GROUP_SCHED
265 /* cpu runqueue to which this cfs_rq is attached */
266 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
271 /* An entity is a task if it doesn't "own" a runqueue */
272 #define entity_is_task(se) (!se->my_q)
274 static inline struct task_struct
*task_of(struct sched_entity
*se
)
276 SCHED_WARN_ON(!entity_is_task(se
));
277 return container_of(se
, struct task_struct
, se
);
280 /* Walk up scheduling entities hierarchy */
281 #define for_each_sched_entity(se) \
282 for (; se; se = se->parent)
284 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
289 /* runqueue on which this entity is (to be) queued */
290 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
295 /* runqueue "owned" by this group */
296 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
301 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
303 if (!cfs_rq
->on_list
) {
304 struct rq
*rq
= rq_of(cfs_rq
);
305 int cpu
= cpu_of(rq
);
307 * Ensure we either appear before our parent (if already
308 * enqueued) or force our parent to appear after us when it is
309 * enqueued. The fact that we always enqueue bottom-up
310 * reduces this to two cases and a special case for the root
311 * cfs_rq. Furthermore, it also means that we will always reset
312 * tmp_alone_branch either when the branch is connected
313 * to a tree or when we reach the beg of the tree
315 if (cfs_rq
->tg
->parent
&&
316 cfs_rq
->tg
->parent
->cfs_rq
[cpu
]->on_list
) {
318 * If parent is already on the list, we add the child
319 * just before. Thanks to circular linked property of
320 * the list, this means to put the child at the tail
321 * of the list that starts by parent.
323 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
324 &(cfs_rq
->tg
->parent
->cfs_rq
[cpu
]->leaf_cfs_rq_list
));
326 * The branch is now connected to its tree so we can
327 * reset tmp_alone_branch to the beginning of the
330 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
331 } else if (!cfs_rq
->tg
->parent
) {
333 * cfs rq without parent should be put
334 * at the tail of the list.
336 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
337 &rq
->leaf_cfs_rq_list
);
339 * We have reach the beg of a tree so we can reset
340 * tmp_alone_branch to the beginning of the list.
342 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
345 * The parent has not already been added so we want to
346 * make sure that it will be put after us.
347 * tmp_alone_branch points to the beg of the branch
348 * where we will add parent.
350 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
351 rq
->tmp_alone_branch
);
353 * update tmp_alone_branch to points to the new beg
356 rq
->tmp_alone_branch
= &cfs_rq
->leaf_cfs_rq_list
;
363 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
365 if (cfs_rq
->on_list
) {
366 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
371 /* Iterate thr' all leaf cfs_rq's on a runqueue */
372 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
373 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
376 /* Do the two (enqueued) entities belong to the same group ? */
377 static inline struct cfs_rq
*
378 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
380 if (se
->cfs_rq
== pse
->cfs_rq
)
386 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
392 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
394 int se_depth
, pse_depth
;
397 * preemption test can be made between sibling entities who are in the
398 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
399 * both tasks until we find their ancestors who are siblings of common
403 /* First walk up until both entities are at same depth */
404 se_depth
= (*se
)->depth
;
405 pse_depth
= (*pse
)->depth
;
407 while (se_depth
> pse_depth
) {
409 *se
= parent_entity(*se
);
412 while (pse_depth
> se_depth
) {
414 *pse
= parent_entity(*pse
);
417 while (!is_same_group(*se
, *pse
)) {
418 *se
= parent_entity(*se
);
419 *pse
= parent_entity(*pse
);
423 #else /* !CONFIG_FAIR_GROUP_SCHED */
425 static inline struct task_struct
*task_of(struct sched_entity
*se
)
427 return container_of(se
, struct task_struct
, se
);
430 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
432 return container_of(cfs_rq
, struct rq
, cfs
);
435 #define entity_is_task(se) 1
437 #define for_each_sched_entity(se) \
438 for (; se; se = NULL)
440 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
442 return &task_rq(p
)->cfs
;
445 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
447 struct task_struct
*p
= task_of(se
);
448 struct rq
*rq
= task_rq(p
);
453 /* runqueue "owned" by this group */
454 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
459 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
463 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
467 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
468 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
470 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
476 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
480 #endif /* CONFIG_FAIR_GROUP_SCHED */
482 static __always_inline
483 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
485 /**************************************************************
486 * Scheduling class tree data structure manipulation methods:
489 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
491 s64 delta
= (s64
)(vruntime
- max_vruntime
);
493 max_vruntime
= vruntime
;
498 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
500 s64 delta
= (s64
)(vruntime
- min_vruntime
);
502 min_vruntime
= vruntime
;
507 static inline int entity_before(struct sched_entity
*a
,
508 struct sched_entity
*b
)
510 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
513 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
515 struct sched_entity
*curr
= cfs_rq
->curr
;
517 u64 vruntime
= cfs_rq
->min_vruntime
;
521 vruntime
= curr
->vruntime
;
526 if (cfs_rq
->rb_leftmost
) {
527 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
532 vruntime
= se
->vruntime
;
534 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
537 /* ensure we never gain time by being placed backwards. */
538 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
541 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
546 * Enqueue an entity into the rb-tree:
548 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
550 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
551 struct rb_node
*parent
= NULL
;
552 struct sched_entity
*entry
;
556 * Find the right place in the rbtree:
560 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
562 * We dont care about collisions. Nodes with
563 * the same key stay together.
565 if (entity_before(se
, entry
)) {
566 link
= &parent
->rb_left
;
568 link
= &parent
->rb_right
;
574 * Maintain a cache of leftmost tree entries (it is frequently
578 cfs_rq
->rb_leftmost
= &se
->run_node
;
580 rb_link_node(&se
->run_node
, parent
, link
);
581 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
584 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
586 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
587 struct rb_node
*next_node
;
589 next_node
= rb_next(&se
->run_node
);
590 cfs_rq
->rb_leftmost
= next_node
;
593 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
596 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
598 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
603 return rb_entry(left
, struct sched_entity
, run_node
);
606 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
608 struct rb_node
*next
= rb_next(&se
->run_node
);
613 return rb_entry(next
, struct sched_entity
, run_node
);
616 #ifdef CONFIG_SCHED_DEBUG
617 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
619 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
624 return rb_entry(last
, struct sched_entity
, run_node
);
627 /**************************************************************
628 * Scheduling class statistics methods:
631 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
632 void __user
*buffer
, size_t *lenp
,
635 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
636 unsigned int factor
= get_update_sysctl_factor();
641 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
642 sysctl_sched_min_granularity
);
644 #define WRT_SYSCTL(name) \
645 (normalized_sysctl_##name = sysctl_##name / (factor))
646 WRT_SYSCTL(sched_min_granularity
);
647 WRT_SYSCTL(sched_latency
);
648 WRT_SYSCTL(sched_wakeup_granularity
);
658 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
660 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
661 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
667 * The idea is to set a period in which each task runs once.
669 * When there are too many tasks (sched_nr_latency) we have to stretch
670 * this period because otherwise the slices get too small.
672 * p = (nr <= nl) ? l : l*nr/nl
674 static u64
__sched_period(unsigned long nr_running
)
676 if (unlikely(nr_running
> sched_nr_latency
))
677 return nr_running
* sysctl_sched_min_granularity
;
679 return sysctl_sched_latency
;
683 * We calculate the wall-time slice from the period by taking a part
684 * proportional to the weight.
688 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
690 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
692 for_each_sched_entity(se
) {
693 struct load_weight
*load
;
694 struct load_weight lw
;
696 cfs_rq
= cfs_rq_of(se
);
697 load
= &cfs_rq
->load
;
699 if (unlikely(!se
->on_rq
)) {
702 update_load_add(&lw
, se
->load
.weight
);
705 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
711 * We calculate the vruntime slice of a to-be-inserted task.
715 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
717 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
722 #include "sched-pelt.h"
724 static int select_idle_sibling(struct task_struct
*p
, int prev_cpu
, int cpu
);
725 static unsigned long task_h_load(struct task_struct
*p
);
727 /* Give new sched_entity start runnable values to heavy its load in infant time */
728 void init_entity_runnable_average(struct sched_entity
*se
)
730 struct sched_avg
*sa
= &se
->avg
;
732 sa
->last_update_time
= 0;
734 * sched_avg's period_contrib should be strictly less then 1024, so
735 * we give it 1023 to make sure it is almost a period (1024us), and
736 * will definitely be update (after enqueue).
738 sa
->period_contrib
= 1023;
740 * Tasks are intialized with full load to be seen as heavy tasks until
741 * they get a chance to stabilize to their real load level.
742 * Group entities are intialized with zero load to reflect the fact that
743 * nothing has been attached to the task group yet.
745 if (entity_is_task(se
))
746 sa
->load_avg
= scale_load_down(se
->load
.weight
);
747 sa
->load_sum
= sa
->load_avg
* LOAD_AVG_MAX
;
749 * At this point, util_avg won't be used in select_task_rq_fair anyway
753 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
756 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
757 static void attach_entity_cfs_rq(struct sched_entity
*se
);
760 * With new tasks being created, their initial util_avgs are extrapolated
761 * based on the cfs_rq's current util_avg:
763 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
765 * However, in many cases, the above util_avg does not give a desired
766 * value. Moreover, the sum of the util_avgs may be divergent, such
767 * as when the series is a harmonic series.
769 * To solve this problem, we also cap the util_avg of successive tasks to
770 * only 1/2 of the left utilization budget:
772 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
774 * where n denotes the nth task.
776 * For example, a simplest series from the beginning would be like:
778 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
779 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
781 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
782 * if util_avg > util_avg_cap.
784 void post_init_entity_util_avg(struct sched_entity
*se
)
786 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
787 struct sched_avg
*sa
= &se
->avg
;
788 long cap
= (long)(SCHED_CAPACITY_SCALE
- cfs_rq
->avg
.util_avg
) / 2;
791 if (cfs_rq
->avg
.util_avg
!= 0) {
792 sa
->util_avg
= cfs_rq
->avg
.util_avg
* se
->load
.weight
;
793 sa
->util_avg
/= (cfs_rq
->avg
.load_avg
+ 1);
795 if (sa
->util_avg
> cap
)
800 sa
->util_sum
= sa
->util_avg
* LOAD_AVG_MAX
;
803 if (entity_is_task(se
)) {
804 struct task_struct
*p
= task_of(se
);
805 if (p
->sched_class
!= &fair_sched_class
) {
807 * For !fair tasks do:
809 update_cfs_rq_load_avg(now, cfs_rq);
810 attach_entity_load_avg(cfs_rq, se);
811 switched_from_fair(rq, p);
813 * such that the next switched_to_fair() has the
816 se
->avg
.last_update_time
= cfs_rq_clock_task(cfs_rq
);
821 attach_entity_cfs_rq(se
);
824 #else /* !CONFIG_SMP */
825 void init_entity_runnable_average(struct sched_entity
*se
)
828 void post_init_entity_util_avg(struct sched_entity
*se
)
831 static void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
)
834 #endif /* CONFIG_SMP */
837 * Update the current task's runtime statistics.
839 static void update_curr(struct cfs_rq
*cfs_rq
)
841 struct sched_entity
*curr
= cfs_rq
->curr
;
842 u64 now
= rq_clock_task(rq_of(cfs_rq
));
848 delta_exec
= now
- curr
->exec_start
;
849 if (unlikely((s64
)delta_exec
<= 0))
852 curr
->exec_start
= now
;
854 schedstat_set(curr
->statistics
.exec_max
,
855 max(delta_exec
, curr
->statistics
.exec_max
));
857 curr
->sum_exec_runtime
+= delta_exec
;
858 schedstat_add(cfs_rq
->exec_clock
, delta_exec
);
860 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
861 update_min_vruntime(cfs_rq
);
863 if (entity_is_task(curr
)) {
864 struct task_struct
*curtask
= task_of(curr
);
866 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
867 cpuacct_charge(curtask
, delta_exec
);
868 account_group_exec_runtime(curtask
, delta_exec
);
871 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
874 static void update_curr_fair(struct rq
*rq
)
876 update_curr(cfs_rq_of(&rq
->curr
->se
));
880 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
882 u64 wait_start
, prev_wait_start
;
884 if (!schedstat_enabled())
887 wait_start
= rq_clock(rq_of(cfs_rq
));
888 prev_wait_start
= schedstat_val(se
->statistics
.wait_start
);
890 if (entity_is_task(se
) && task_on_rq_migrating(task_of(se
)) &&
891 likely(wait_start
> prev_wait_start
))
892 wait_start
-= prev_wait_start
;
894 schedstat_set(se
->statistics
.wait_start
, wait_start
);
898 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
900 struct task_struct
*p
;
903 if (!schedstat_enabled())
906 delta
= rq_clock(rq_of(cfs_rq
)) - schedstat_val(se
->statistics
.wait_start
);
908 if (entity_is_task(se
)) {
910 if (task_on_rq_migrating(p
)) {
912 * Preserve migrating task's wait time so wait_start
913 * time stamp can be adjusted to accumulate wait time
914 * prior to migration.
916 schedstat_set(se
->statistics
.wait_start
, delta
);
919 trace_sched_stat_wait(p
, delta
);
922 schedstat_set(se
->statistics
.wait_max
,
923 max(schedstat_val(se
->statistics
.wait_max
), delta
));
924 schedstat_inc(se
->statistics
.wait_count
);
925 schedstat_add(se
->statistics
.wait_sum
, delta
);
926 schedstat_set(se
->statistics
.wait_start
, 0);
930 update_stats_enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
932 struct task_struct
*tsk
= NULL
;
933 u64 sleep_start
, block_start
;
935 if (!schedstat_enabled())
938 sleep_start
= schedstat_val(se
->statistics
.sleep_start
);
939 block_start
= schedstat_val(se
->statistics
.block_start
);
941 if (entity_is_task(se
))
945 u64 delta
= rq_clock(rq_of(cfs_rq
)) - sleep_start
;
950 if (unlikely(delta
> schedstat_val(se
->statistics
.sleep_max
)))
951 schedstat_set(se
->statistics
.sleep_max
, delta
);
953 schedstat_set(se
->statistics
.sleep_start
, 0);
954 schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
957 account_scheduler_latency(tsk
, delta
>> 10, 1);
958 trace_sched_stat_sleep(tsk
, delta
);
962 u64 delta
= rq_clock(rq_of(cfs_rq
)) - block_start
;
967 if (unlikely(delta
> schedstat_val(se
->statistics
.block_max
)))
968 schedstat_set(se
->statistics
.block_max
, delta
);
970 schedstat_set(se
->statistics
.block_start
, 0);
971 schedstat_add(se
->statistics
.sum_sleep_runtime
, delta
);
974 if (tsk
->in_iowait
) {
975 schedstat_add(se
->statistics
.iowait_sum
, delta
);
976 schedstat_inc(se
->statistics
.iowait_count
);
977 trace_sched_stat_iowait(tsk
, delta
);
980 trace_sched_stat_blocked(tsk
, delta
);
983 * Blocking time is in units of nanosecs, so shift by
984 * 20 to get a milliseconds-range estimation of the
985 * amount of time that the task spent sleeping:
987 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
988 profile_hits(SLEEP_PROFILING
,
989 (void *)get_wchan(tsk
),
992 account_scheduler_latency(tsk
, delta
>> 10, 0);
998 * Task is being enqueued - update stats:
1001 update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1003 if (!schedstat_enabled())
1007 * Are we enqueueing a waiting task? (for current tasks
1008 * a dequeue/enqueue event is a NOP)
1010 if (se
!= cfs_rq
->curr
)
1011 update_stats_wait_start(cfs_rq
, se
);
1013 if (flags
& ENQUEUE_WAKEUP
)
1014 update_stats_enqueue_sleeper(cfs_rq
, se
);
1018 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1021 if (!schedstat_enabled())
1025 * Mark the end of the wait period if dequeueing a
1028 if (se
!= cfs_rq
->curr
)
1029 update_stats_wait_end(cfs_rq
, se
);
1031 if ((flags
& DEQUEUE_SLEEP
) && entity_is_task(se
)) {
1032 struct task_struct
*tsk
= task_of(se
);
1034 if (tsk
->state
& TASK_INTERRUPTIBLE
)
1035 schedstat_set(se
->statistics
.sleep_start
,
1036 rq_clock(rq_of(cfs_rq
)));
1037 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
1038 schedstat_set(se
->statistics
.block_start
,
1039 rq_clock(rq_of(cfs_rq
)));
1044 * We are picking a new current task - update its stats:
1047 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1050 * We are starting a new run period:
1052 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
1055 /**************************************************
1056 * Scheduling class queueing methods:
1059 #ifdef CONFIG_NUMA_BALANCING
1061 * Approximate time to scan a full NUMA task in ms. The task scan period is
1062 * calculated based on the tasks virtual memory size and
1063 * numa_balancing_scan_size.
1065 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
1066 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
1068 /* Portion of address space to scan in MB */
1069 unsigned int sysctl_numa_balancing_scan_size
= 256;
1071 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1072 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
1077 spinlock_t lock
; /* nr_tasks, tasks */
1082 struct rcu_head rcu
;
1083 unsigned long total_faults
;
1084 unsigned long max_faults_cpu
;
1086 * Faults_cpu is used to decide whether memory should move
1087 * towards the CPU. As a consequence, these stats are weighted
1088 * more by CPU use than by memory faults.
1090 unsigned long *faults_cpu
;
1091 unsigned long faults
[0];
1094 static inline unsigned long group_faults_priv(struct numa_group
*ng
);
1095 static inline unsigned long group_faults_shared(struct numa_group
*ng
);
1097 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
1099 unsigned long rss
= 0;
1100 unsigned long nr_scan_pages
;
1103 * Calculations based on RSS as non-present and empty pages are skipped
1104 * by the PTE scanner and NUMA hinting faults should be trapped based
1107 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
1108 rss
= get_mm_rss(p
->mm
);
1110 rss
= nr_scan_pages
;
1112 rss
= round_up(rss
, nr_scan_pages
);
1113 return rss
/ nr_scan_pages
;
1116 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1117 #define MAX_SCAN_WINDOW 2560
1119 static unsigned int task_scan_min(struct task_struct
*p
)
1121 unsigned int scan_size
= READ_ONCE(sysctl_numa_balancing_scan_size
);
1122 unsigned int scan
, floor
;
1123 unsigned int windows
= 1;
1125 if (scan_size
< MAX_SCAN_WINDOW
)
1126 windows
= MAX_SCAN_WINDOW
/ scan_size
;
1127 floor
= 1000 / windows
;
1129 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
1130 return max_t(unsigned int, floor
, scan
);
1133 static unsigned int task_scan_start(struct task_struct
*p
)
1135 unsigned long smin
= task_scan_min(p
);
1136 unsigned long period
= smin
;
1138 /* Scale the maximum scan period with the amount of shared memory. */
1139 if (p
->numa_group
) {
1140 struct numa_group
*ng
= p
->numa_group
;
1141 unsigned long shared
= group_faults_shared(ng
);
1142 unsigned long private = group_faults_priv(ng
);
1144 period
*= atomic_read(&ng
->refcount
);
1145 period
*= shared
+ 1;
1146 period
/= private + shared
+ 1;
1149 return max(smin
, period
);
1152 static unsigned int task_scan_max(struct task_struct
*p
)
1154 unsigned long smin
= task_scan_min(p
);
1157 /* Watch for min being lower than max due to floor calculations */
1158 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
1160 /* Scale the maximum scan period with the amount of shared memory. */
1161 if (p
->numa_group
) {
1162 struct numa_group
*ng
= p
->numa_group
;
1163 unsigned long shared
= group_faults_shared(ng
);
1164 unsigned long private = group_faults_priv(ng
);
1165 unsigned long period
= smax
;
1167 period
*= atomic_read(&ng
->refcount
);
1168 period
*= shared
+ 1;
1169 period
/= private + shared
+ 1;
1171 smax
= max(smax
, period
);
1174 return max(smin
, smax
);
1177 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
1179 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
1180 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
1183 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
1185 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
1186 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
1189 /* Shared or private faults. */
1190 #define NR_NUMA_HINT_FAULT_TYPES 2
1192 /* Memory and CPU locality */
1193 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1195 /* Averaged statistics, and temporary buffers. */
1196 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1198 pid_t
task_numa_group_id(struct task_struct
*p
)
1200 return p
->numa_group
? p
->numa_group
->gid
: 0;
1204 * The averaged statistics, shared & private, memory & cpu,
1205 * occupy the first half of the array. The second half of the
1206 * array is for current counters, which are averaged into the
1207 * first set by task_numa_placement.
1209 static inline int task_faults_idx(enum numa_faults_stats s
, int nid
, int priv
)
1211 return NR_NUMA_HINT_FAULT_TYPES
* (s
* nr_node_ids
+ nid
) + priv
;
1214 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
1216 if (!p
->numa_faults
)
1219 return p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1220 p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1223 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
1228 return p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1229 p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1232 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
1234 return group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1235 group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1238 static inline unsigned long group_faults_priv(struct numa_group
*ng
)
1240 unsigned long faults
= 0;
1243 for_each_online_node(node
) {
1244 faults
+= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
1250 static inline unsigned long group_faults_shared(struct numa_group
*ng
)
1252 unsigned long faults
= 0;
1255 for_each_online_node(node
) {
1256 faults
+= ng
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
1263 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1264 * considered part of a numa group's pseudo-interleaving set. Migrations
1265 * between these nodes are slowed down, to allow things to settle down.
1267 #define ACTIVE_NODE_FRACTION 3
1269 static bool numa_is_active_node(int nid
, struct numa_group
*ng
)
1271 return group_faults_cpu(ng
, nid
) * ACTIVE_NODE_FRACTION
> ng
->max_faults_cpu
;
1274 /* Handle placement on systems where not all nodes are directly connected. */
1275 static unsigned long score_nearby_nodes(struct task_struct
*p
, int nid
,
1276 int maxdist
, bool task
)
1278 unsigned long score
= 0;
1282 * All nodes are directly connected, and the same distance
1283 * from each other. No need for fancy placement algorithms.
1285 if (sched_numa_topology_type
== NUMA_DIRECT
)
1289 * This code is called for each node, introducing N^2 complexity,
1290 * which should be ok given the number of nodes rarely exceeds 8.
1292 for_each_online_node(node
) {
1293 unsigned long faults
;
1294 int dist
= node_distance(nid
, node
);
1297 * The furthest away nodes in the system are not interesting
1298 * for placement; nid was already counted.
1300 if (dist
== sched_max_numa_distance
|| node
== nid
)
1304 * On systems with a backplane NUMA topology, compare groups
1305 * of nodes, and move tasks towards the group with the most
1306 * memory accesses. When comparing two nodes at distance
1307 * "hoplimit", only nodes closer by than "hoplimit" are part
1308 * of each group. Skip other nodes.
1310 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1314 /* Add up the faults from nearby nodes. */
1316 faults
= task_faults(p
, node
);
1318 faults
= group_faults(p
, node
);
1321 * On systems with a glueless mesh NUMA topology, there are
1322 * no fixed "groups of nodes". Instead, nodes that are not
1323 * directly connected bounce traffic through intermediate
1324 * nodes; a numa_group can occupy any set of nodes.
1325 * The further away a node is, the less the faults count.
1326 * This seems to result in good task placement.
1328 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1329 faults
*= (sched_max_numa_distance
- dist
);
1330 faults
/= (sched_max_numa_distance
- LOCAL_DISTANCE
);
1340 * These return the fraction of accesses done by a particular task, or
1341 * task group, on a particular numa node. The group weight is given a
1342 * larger multiplier, in order to group tasks together that are almost
1343 * evenly spread out between numa nodes.
1345 static inline unsigned long task_weight(struct task_struct
*p
, int nid
,
1348 unsigned long faults
, total_faults
;
1350 if (!p
->numa_faults
)
1353 total_faults
= p
->total_numa_faults
;
1358 faults
= task_faults(p
, nid
);
1359 faults
+= score_nearby_nodes(p
, nid
, dist
, true);
1361 return 1000 * faults
/ total_faults
;
1364 static inline unsigned long group_weight(struct task_struct
*p
, int nid
,
1367 unsigned long faults
, total_faults
;
1372 total_faults
= p
->numa_group
->total_faults
;
1377 faults
= group_faults(p
, nid
);
1378 faults
+= score_nearby_nodes(p
, nid
, dist
, false);
1380 return 1000 * faults
/ total_faults
;
1383 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
1384 int src_nid
, int dst_cpu
)
1386 struct numa_group
*ng
= p
->numa_group
;
1387 int dst_nid
= cpu_to_node(dst_cpu
);
1388 int last_cpupid
, this_cpupid
;
1390 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
1393 * Multi-stage node selection is used in conjunction with a periodic
1394 * migration fault to build a temporal task<->page relation. By using
1395 * a two-stage filter we remove short/unlikely relations.
1397 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1398 * a task's usage of a particular page (n_p) per total usage of this
1399 * page (n_t) (in a given time-span) to a probability.
1401 * Our periodic faults will sample this probability and getting the
1402 * same result twice in a row, given these samples are fully
1403 * independent, is then given by P(n)^2, provided our sample period
1404 * is sufficiently short compared to the usage pattern.
1406 * This quadric squishes small probabilities, making it less likely we
1407 * act on an unlikely task<->page relation.
1409 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
1410 if (!cpupid_pid_unset(last_cpupid
) &&
1411 cpupid_to_nid(last_cpupid
) != dst_nid
)
1414 /* Always allow migrate on private faults */
1415 if (cpupid_match_pid(p
, last_cpupid
))
1418 /* A shared fault, but p->numa_group has not been set up yet. */
1423 * Destination node is much more heavily used than the source
1424 * node? Allow migration.
1426 if (group_faults_cpu(ng
, dst_nid
) > group_faults_cpu(ng
, src_nid
) *
1427 ACTIVE_NODE_FRACTION
)
1431 * Distribute memory according to CPU & memory use on each node,
1432 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1434 * faults_cpu(dst) 3 faults_cpu(src)
1435 * --------------- * - > ---------------
1436 * faults_mem(dst) 4 faults_mem(src)
1438 return group_faults_cpu(ng
, dst_nid
) * group_faults(p
, src_nid
) * 3 >
1439 group_faults_cpu(ng
, src_nid
) * group_faults(p
, dst_nid
) * 4;
1442 static unsigned long weighted_cpuload(struct rq
*rq
);
1443 static unsigned long source_load(int cpu
, int type
);
1444 static unsigned long target_load(int cpu
, int type
);
1445 static unsigned long capacity_of(int cpu
);
1447 /* Cached statistics for all CPUs within a node */
1449 unsigned long nr_running
;
1452 /* Total compute capacity of CPUs on a node */
1453 unsigned long compute_capacity
;
1455 /* Approximate capacity in terms of runnable tasks on a node */
1456 unsigned long task_capacity
;
1457 int has_free_capacity
;
1461 * XXX borrowed from update_sg_lb_stats
1463 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1465 int smt
, cpu
, cpus
= 0;
1466 unsigned long capacity
;
1468 memset(ns
, 0, sizeof(*ns
));
1469 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1470 struct rq
*rq
= cpu_rq(cpu
);
1472 ns
->nr_running
+= rq
->nr_running
;
1473 ns
->load
+= weighted_cpuload(rq
);
1474 ns
->compute_capacity
+= capacity_of(cpu
);
1480 * If we raced with hotplug and there are no CPUs left in our mask
1481 * the @ns structure is NULL'ed and task_numa_compare() will
1482 * not find this node attractive.
1484 * We'll either bail at !has_free_capacity, or we'll detect a huge
1485 * imbalance and bail there.
1490 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1491 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, ns
->compute_capacity
);
1492 capacity
= cpus
/ smt
; /* cores */
1494 ns
->task_capacity
= min_t(unsigned, capacity
,
1495 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
));
1496 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1499 struct task_numa_env
{
1500 struct task_struct
*p
;
1502 int src_cpu
, src_nid
;
1503 int dst_cpu
, dst_nid
;
1505 struct numa_stats src_stats
, dst_stats
;
1510 struct task_struct
*best_task
;
1515 static void task_numa_assign(struct task_numa_env
*env
,
1516 struct task_struct
*p
, long imp
)
1519 put_task_struct(env
->best_task
);
1524 env
->best_imp
= imp
;
1525 env
->best_cpu
= env
->dst_cpu
;
1528 static bool load_too_imbalanced(long src_load
, long dst_load
,
1529 struct task_numa_env
*env
)
1532 long orig_src_load
, orig_dst_load
;
1533 long src_capacity
, dst_capacity
;
1536 * The load is corrected for the CPU capacity available on each node.
1539 * ------------ vs ---------
1540 * src_capacity dst_capacity
1542 src_capacity
= env
->src_stats
.compute_capacity
;
1543 dst_capacity
= env
->dst_stats
.compute_capacity
;
1545 /* We care about the slope of the imbalance, not the direction. */
1546 if (dst_load
< src_load
)
1547 swap(dst_load
, src_load
);
1549 /* Is the difference below the threshold? */
1550 imb
= dst_load
* src_capacity
* 100 -
1551 src_load
* dst_capacity
* env
->imbalance_pct
;
1556 * The imbalance is above the allowed threshold.
1557 * Compare it with the old imbalance.
1559 orig_src_load
= env
->src_stats
.load
;
1560 orig_dst_load
= env
->dst_stats
.load
;
1562 if (orig_dst_load
< orig_src_load
)
1563 swap(orig_dst_load
, orig_src_load
);
1565 old_imb
= orig_dst_load
* src_capacity
* 100 -
1566 orig_src_load
* dst_capacity
* env
->imbalance_pct
;
1568 /* Would this change make things worse? */
1569 return (imb
> old_imb
);
1573 * This checks if the overall compute and NUMA accesses of the system would
1574 * be improved if the source tasks was migrated to the target dst_cpu taking
1575 * into account that it might be best if task running on the dst_cpu should
1576 * be exchanged with the source task
1578 static void task_numa_compare(struct task_numa_env
*env
,
1579 long taskimp
, long groupimp
)
1581 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1582 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1583 struct task_struct
*cur
;
1584 long src_load
, dst_load
;
1586 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1588 int dist
= env
->dist
;
1591 cur
= task_rcu_dereference(&dst_rq
->curr
);
1592 if (cur
&& ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
)))
1596 * Because we have preemption enabled we can get migrated around and
1597 * end try selecting ourselves (current == env->p) as a swap candidate.
1603 * "imp" is the fault differential for the source task between the
1604 * source and destination node. Calculate the total differential for
1605 * the source task and potential destination task. The more negative
1606 * the value is, the more rmeote accesses that would be expected to
1607 * be incurred if the tasks were swapped.
1610 /* Skip this swap candidate if cannot move to the source cpu */
1611 if (!cpumask_test_cpu(env
->src_cpu
, &cur
->cpus_allowed
))
1615 * If dst and source tasks are in the same NUMA group, or not
1616 * in any group then look only at task weights.
1618 if (cur
->numa_group
== env
->p
->numa_group
) {
1619 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1620 task_weight(cur
, env
->dst_nid
, dist
);
1622 * Add some hysteresis to prevent swapping the
1623 * tasks within a group over tiny differences.
1625 if (cur
->numa_group
)
1629 * Compare the group weights. If a task is all by
1630 * itself (not part of a group), use the task weight
1633 if (cur
->numa_group
)
1634 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1635 group_weight(cur
, env
->dst_nid
, dist
);
1637 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1638 task_weight(cur
, env
->dst_nid
, dist
);
1642 if (imp
<= env
->best_imp
&& moveimp
<= env
->best_imp
)
1646 /* Is there capacity at our destination? */
1647 if (env
->src_stats
.nr_running
<= env
->src_stats
.task_capacity
&&
1648 !env
->dst_stats
.has_free_capacity
)
1654 /* Balance doesn't matter much if we're running a task per cpu */
1655 if (imp
> env
->best_imp
&& src_rq
->nr_running
== 1 &&
1656 dst_rq
->nr_running
== 1)
1660 * In the overloaded case, try and keep the load balanced.
1663 load
= task_h_load(env
->p
);
1664 dst_load
= env
->dst_stats
.load
+ load
;
1665 src_load
= env
->src_stats
.load
- load
;
1667 if (moveimp
> imp
&& moveimp
> env
->best_imp
) {
1669 * If the improvement from just moving env->p direction is
1670 * better than swapping tasks around, check if a move is
1671 * possible. Store a slightly smaller score than moveimp,
1672 * so an actually idle CPU will win.
1674 if (!load_too_imbalanced(src_load
, dst_load
, env
)) {
1681 if (imp
<= env
->best_imp
)
1685 load
= task_h_load(cur
);
1690 if (load_too_imbalanced(src_load
, dst_load
, env
))
1694 * One idle CPU per node is evaluated for a task numa move.
1695 * Call select_idle_sibling to maybe find a better one.
1699 * select_idle_siblings() uses an per-cpu cpumask that
1700 * can be used from IRQ context.
1702 local_irq_disable();
1703 env
->dst_cpu
= select_idle_sibling(env
->p
, env
->src_cpu
,
1709 task_numa_assign(env
, cur
, imp
);
1714 static void task_numa_find_cpu(struct task_numa_env
*env
,
1715 long taskimp
, long groupimp
)
1719 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1720 /* Skip this CPU if the source task cannot migrate */
1721 if (!cpumask_test_cpu(cpu
, &env
->p
->cpus_allowed
))
1725 task_numa_compare(env
, taskimp
, groupimp
);
1729 /* Only move tasks to a NUMA node less busy than the current node. */
1730 static bool numa_has_capacity(struct task_numa_env
*env
)
1732 struct numa_stats
*src
= &env
->src_stats
;
1733 struct numa_stats
*dst
= &env
->dst_stats
;
1735 if (src
->has_free_capacity
&& !dst
->has_free_capacity
)
1739 * Only consider a task move if the source has a higher load
1740 * than the destination, corrected for CPU capacity on each node.
1742 * src->load dst->load
1743 * --------------------- vs ---------------------
1744 * src->compute_capacity dst->compute_capacity
1746 if (src
->load
* dst
->compute_capacity
* env
->imbalance_pct
>
1748 dst
->load
* src
->compute_capacity
* 100)
1754 static int task_numa_migrate(struct task_struct
*p
)
1756 struct task_numa_env env
= {
1759 .src_cpu
= task_cpu(p
),
1760 .src_nid
= task_node(p
),
1762 .imbalance_pct
= 112,
1768 struct sched_domain
*sd
;
1769 unsigned long taskweight
, groupweight
;
1771 long taskimp
, groupimp
;
1774 * Pick the lowest SD_NUMA domain, as that would have the smallest
1775 * imbalance and would be the first to start moving tasks about.
1777 * And we want to avoid any moving of tasks about, as that would create
1778 * random movement of tasks -- counter the numa conditions we're trying
1782 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1784 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1788 * Cpusets can break the scheduler domain tree into smaller
1789 * balance domains, some of which do not cross NUMA boundaries.
1790 * Tasks that are "trapped" in such domains cannot be migrated
1791 * elsewhere, so there is no point in (re)trying.
1793 if (unlikely(!sd
)) {
1794 p
->numa_preferred_nid
= task_node(p
);
1798 env
.dst_nid
= p
->numa_preferred_nid
;
1799 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1800 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1801 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1802 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1803 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
1804 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
1805 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1807 /* Try to find a spot on the preferred nid. */
1808 if (numa_has_capacity(&env
))
1809 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1812 * Look at other nodes in these cases:
1813 * - there is no space available on the preferred_nid
1814 * - the task is part of a numa_group that is interleaved across
1815 * multiple NUMA nodes; in order to better consolidate the group,
1816 * we need to check other locations.
1818 if (env
.best_cpu
== -1 || (p
->numa_group
&& p
->numa_group
->active_nodes
> 1)) {
1819 for_each_online_node(nid
) {
1820 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1823 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1824 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1826 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1827 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1830 /* Only consider nodes where both task and groups benefit */
1831 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
1832 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
1833 if (taskimp
< 0 && groupimp
< 0)
1838 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1839 if (numa_has_capacity(&env
))
1840 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1845 * If the task is part of a workload that spans multiple NUMA nodes,
1846 * and is migrating into one of the workload's active nodes, remember
1847 * this node as the task's preferred numa node, so the workload can
1849 * A task that migrated to a second choice node will be better off
1850 * trying for a better one later. Do not set the preferred node here.
1852 if (p
->numa_group
) {
1853 struct numa_group
*ng
= p
->numa_group
;
1855 if (env
.best_cpu
== -1)
1860 if (ng
->active_nodes
> 1 && numa_is_active_node(env
.dst_nid
, ng
))
1861 sched_setnuma(p
, env
.dst_nid
);
1864 /* No better CPU than the current one was found. */
1865 if (env
.best_cpu
== -1)
1869 * Reset the scan period if the task is being rescheduled on an
1870 * alternative node to recheck if the tasks is now properly placed.
1872 p
->numa_scan_period
= task_scan_start(p
);
1874 if (env
.best_task
== NULL
) {
1875 ret
= migrate_task_to(p
, env
.best_cpu
);
1877 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1881 ret
= migrate_swap(p
, env
.best_task
);
1883 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1884 put_task_struct(env
.best_task
);
1888 /* Attempt to migrate a task to a CPU on the preferred node. */
1889 static void numa_migrate_preferred(struct task_struct
*p
)
1891 unsigned long interval
= HZ
;
1893 /* This task has no NUMA fault statistics yet */
1894 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults
))
1897 /* Periodically retry migrating the task to the preferred node */
1898 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1899 p
->numa_migrate_retry
= jiffies
+ interval
;
1901 /* Success if task is already running on preferred CPU */
1902 if (task_node(p
) == p
->numa_preferred_nid
)
1905 /* Otherwise, try migrate to a CPU on the preferred node */
1906 task_numa_migrate(p
);
1910 * Find out how many nodes on the workload is actively running on. Do this by
1911 * tracking the nodes from which NUMA hinting faults are triggered. This can
1912 * be different from the set of nodes where the workload's memory is currently
1915 static void numa_group_count_active_nodes(struct numa_group
*numa_group
)
1917 unsigned long faults
, max_faults
= 0;
1918 int nid
, active_nodes
= 0;
1920 for_each_online_node(nid
) {
1921 faults
= group_faults_cpu(numa_group
, nid
);
1922 if (faults
> max_faults
)
1923 max_faults
= faults
;
1926 for_each_online_node(nid
) {
1927 faults
= group_faults_cpu(numa_group
, nid
);
1928 if (faults
* ACTIVE_NODE_FRACTION
> max_faults
)
1932 numa_group
->max_faults_cpu
= max_faults
;
1933 numa_group
->active_nodes
= active_nodes
;
1937 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1938 * increments. The more local the fault statistics are, the higher the scan
1939 * period will be for the next scan window. If local/(local+remote) ratio is
1940 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1941 * the scan period will decrease. Aim for 70% local accesses.
1943 #define NUMA_PERIOD_SLOTS 10
1944 #define NUMA_PERIOD_THRESHOLD 7
1947 * Increase the scan period (slow down scanning) if the majority of
1948 * our memory is already on our local node, or if the majority of
1949 * the page accesses are shared with other processes.
1950 * Otherwise, decrease the scan period.
1952 static void update_task_scan_period(struct task_struct
*p
,
1953 unsigned long shared
, unsigned long private)
1955 unsigned int period_slot
;
1956 int lr_ratio
, ps_ratio
;
1959 unsigned long remote
= p
->numa_faults_locality
[0];
1960 unsigned long local
= p
->numa_faults_locality
[1];
1963 * If there were no record hinting faults then either the task is
1964 * completely idle or all activity is areas that are not of interest
1965 * to automatic numa balancing. Related to that, if there were failed
1966 * migration then it implies we are migrating too quickly or the local
1967 * node is overloaded. In either case, scan slower
1969 if (local
+ shared
== 0 || p
->numa_faults_locality
[2]) {
1970 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1971 p
->numa_scan_period
<< 1);
1973 p
->mm
->numa_next_scan
= jiffies
+
1974 msecs_to_jiffies(p
->numa_scan_period
);
1980 * Prepare to scale scan period relative to the current period.
1981 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1982 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1983 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1985 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1986 lr_ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1987 ps_ratio
= (private * NUMA_PERIOD_SLOTS
) / (private + shared
);
1989 if (ps_ratio
>= NUMA_PERIOD_THRESHOLD
) {
1991 * Most memory accesses are local. There is no need to
1992 * do fast NUMA scanning, since memory is already local.
1994 int slot
= ps_ratio
- NUMA_PERIOD_THRESHOLD
;
1997 diff
= slot
* period_slot
;
1998 } else if (lr_ratio
>= NUMA_PERIOD_THRESHOLD
) {
2000 * Most memory accesses are shared with other tasks.
2001 * There is no point in continuing fast NUMA scanning,
2002 * since other tasks may just move the memory elsewhere.
2004 int slot
= lr_ratio
- NUMA_PERIOD_THRESHOLD
;
2007 diff
= slot
* period_slot
;
2010 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2011 * yet they are not on the local NUMA node. Speed up
2012 * NUMA scanning to get the memory moved over.
2014 int ratio
= max(lr_ratio
, ps_ratio
);
2015 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
2018 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
2019 task_scan_min(p
), task_scan_max(p
));
2020 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2024 * Get the fraction of time the task has been running since the last
2025 * NUMA placement cycle. The scheduler keeps similar statistics, but
2026 * decays those on a 32ms period, which is orders of magnitude off
2027 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2028 * stats only if the task is so new there are no NUMA statistics yet.
2030 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
2032 u64 runtime
, delta
, now
;
2033 /* Use the start of this time slice to avoid calculations. */
2034 now
= p
->se
.exec_start
;
2035 runtime
= p
->se
.sum_exec_runtime
;
2037 if (p
->last_task_numa_placement
) {
2038 delta
= runtime
- p
->last_sum_exec_runtime
;
2039 *period
= now
- p
->last_task_numa_placement
;
2041 delta
= p
->se
.avg
.load_sum
/ p
->se
.load
.weight
;
2042 *period
= LOAD_AVG_MAX
;
2045 p
->last_sum_exec_runtime
= runtime
;
2046 p
->last_task_numa_placement
= now
;
2052 * Determine the preferred nid for a task in a numa_group. This needs to
2053 * be done in a way that produces consistent results with group_weight,
2054 * otherwise workloads might not converge.
2056 static int preferred_group_nid(struct task_struct
*p
, int nid
)
2061 /* Direct connections between all NUMA nodes. */
2062 if (sched_numa_topology_type
== NUMA_DIRECT
)
2066 * On a system with glueless mesh NUMA topology, group_weight
2067 * scores nodes according to the number of NUMA hinting faults on
2068 * both the node itself, and on nearby nodes.
2070 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
2071 unsigned long score
, max_score
= 0;
2072 int node
, max_node
= nid
;
2074 dist
= sched_max_numa_distance
;
2076 for_each_online_node(node
) {
2077 score
= group_weight(p
, node
, dist
);
2078 if (score
> max_score
) {
2087 * Finding the preferred nid in a system with NUMA backplane
2088 * interconnect topology is more involved. The goal is to locate
2089 * tasks from numa_groups near each other in the system, and
2090 * untangle workloads from different sides of the system. This requires
2091 * searching down the hierarchy of node groups, recursively searching
2092 * inside the highest scoring group of nodes. The nodemask tricks
2093 * keep the complexity of the search down.
2095 nodes
= node_online_map
;
2096 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
2097 unsigned long max_faults
= 0;
2098 nodemask_t max_group
= NODE_MASK_NONE
;
2101 /* Are there nodes at this distance from each other? */
2102 if (!find_numa_distance(dist
))
2105 for_each_node_mask(a
, nodes
) {
2106 unsigned long faults
= 0;
2107 nodemask_t this_group
;
2108 nodes_clear(this_group
);
2110 /* Sum group's NUMA faults; includes a==b case. */
2111 for_each_node_mask(b
, nodes
) {
2112 if (node_distance(a
, b
) < dist
) {
2113 faults
+= group_faults(p
, b
);
2114 node_set(b
, this_group
);
2115 node_clear(b
, nodes
);
2119 /* Remember the top group. */
2120 if (faults
> max_faults
) {
2121 max_faults
= faults
;
2122 max_group
= this_group
;
2124 * subtle: at the smallest distance there is
2125 * just one node left in each "group", the
2126 * winner is the preferred nid.
2131 /* Next round, evaluate the nodes within max_group. */
2139 static void task_numa_placement(struct task_struct
*p
)
2141 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
2142 unsigned long max_faults
= 0, max_group_faults
= 0;
2143 unsigned long fault_types
[2] = { 0, 0 };
2144 unsigned long total_faults
;
2145 u64 runtime
, period
;
2146 spinlock_t
*group_lock
= NULL
;
2149 * The p->mm->numa_scan_seq field gets updated without
2150 * exclusive access. Use READ_ONCE() here to ensure
2151 * that the field is read in a single access:
2153 seq
= READ_ONCE(p
->mm
->numa_scan_seq
);
2154 if (p
->numa_scan_seq
== seq
)
2156 p
->numa_scan_seq
= seq
;
2157 p
->numa_scan_period_max
= task_scan_max(p
);
2159 total_faults
= p
->numa_faults_locality
[0] +
2160 p
->numa_faults_locality
[1];
2161 runtime
= numa_get_avg_runtime(p
, &period
);
2163 /* If the task is part of a group prevent parallel updates to group stats */
2164 if (p
->numa_group
) {
2165 group_lock
= &p
->numa_group
->lock
;
2166 spin_lock_irq(group_lock
);
2169 /* Find the node with the highest number of faults */
2170 for_each_online_node(nid
) {
2171 /* Keep track of the offsets in numa_faults array */
2172 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
2173 unsigned long faults
= 0, group_faults
= 0;
2176 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
2177 long diff
, f_diff
, f_weight
;
2179 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
2180 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
2181 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
2182 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
2184 /* Decay existing window, copy faults since last scan */
2185 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
2186 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
2187 p
->numa_faults
[membuf_idx
] = 0;
2190 * Normalize the faults_from, so all tasks in a group
2191 * count according to CPU use, instead of by the raw
2192 * number of faults. Tasks with little runtime have
2193 * little over-all impact on throughput, and thus their
2194 * faults are less important.
2196 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
2197 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
2199 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
2200 p
->numa_faults
[cpubuf_idx
] = 0;
2202 p
->numa_faults
[mem_idx
] += diff
;
2203 p
->numa_faults
[cpu_idx
] += f_diff
;
2204 faults
+= p
->numa_faults
[mem_idx
];
2205 p
->total_numa_faults
+= diff
;
2206 if (p
->numa_group
) {
2208 * safe because we can only change our own group
2210 * mem_idx represents the offset for a given
2211 * nid and priv in a specific region because it
2212 * is at the beginning of the numa_faults array.
2214 p
->numa_group
->faults
[mem_idx
] += diff
;
2215 p
->numa_group
->faults_cpu
[mem_idx
] += f_diff
;
2216 p
->numa_group
->total_faults
+= diff
;
2217 group_faults
+= p
->numa_group
->faults
[mem_idx
];
2221 if (faults
> max_faults
) {
2222 max_faults
= faults
;
2226 if (group_faults
> max_group_faults
) {
2227 max_group_faults
= group_faults
;
2228 max_group_nid
= nid
;
2232 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
2234 if (p
->numa_group
) {
2235 numa_group_count_active_nodes(p
->numa_group
);
2236 spin_unlock_irq(group_lock
);
2237 max_nid
= preferred_group_nid(p
, max_group_nid
);
2241 /* Set the new preferred node */
2242 if (max_nid
!= p
->numa_preferred_nid
)
2243 sched_setnuma(p
, max_nid
);
2245 if (task_node(p
) != p
->numa_preferred_nid
)
2246 numa_migrate_preferred(p
);
2250 static inline int get_numa_group(struct numa_group
*grp
)
2252 return atomic_inc_not_zero(&grp
->refcount
);
2255 static inline void put_numa_group(struct numa_group
*grp
)
2257 if (atomic_dec_and_test(&grp
->refcount
))
2258 kfree_rcu(grp
, rcu
);
2261 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
2264 struct numa_group
*grp
, *my_grp
;
2265 struct task_struct
*tsk
;
2267 int cpu
= cpupid_to_cpu(cpupid
);
2270 if (unlikely(!p
->numa_group
)) {
2271 unsigned int size
= sizeof(struct numa_group
) +
2272 4*nr_node_ids
*sizeof(unsigned long);
2274 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
2278 atomic_set(&grp
->refcount
, 1);
2279 grp
->active_nodes
= 1;
2280 grp
->max_faults_cpu
= 0;
2281 spin_lock_init(&grp
->lock
);
2283 /* Second half of the array tracks nids where faults happen */
2284 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
2287 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2288 grp
->faults
[i
] = p
->numa_faults
[i
];
2290 grp
->total_faults
= p
->total_numa_faults
;
2293 rcu_assign_pointer(p
->numa_group
, grp
);
2297 tsk
= READ_ONCE(cpu_rq(cpu
)->curr
);
2299 if (!cpupid_match_pid(tsk
, cpupid
))
2302 grp
= rcu_dereference(tsk
->numa_group
);
2306 my_grp
= p
->numa_group
;
2311 * Only join the other group if its bigger; if we're the bigger group,
2312 * the other task will join us.
2314 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
2318 * Tie-break on the grp address.
2320 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
2323 /* Always join threads in the same process. */
2324 if (tsk
->mm
== current
->mm
)
2327 /* Simple filter to avoid false positives due to PID collisions */
2328 if (flags
& TNF_SHARED
)
2331 /* Update priv based on whether false sharing was detected */
2334 if (join
&& !get_numa_group(grp
))
2342 BUG_ON(irqs_disabled());
2343 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
2345 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
2346 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
2347 grp
->faults
[i
] += p
->numa_faults
[i
];
2349 my_grp
->total_faults
-= p
->total_numa_faults
;
2350 grp
->total_faults
+= p
->total_numa_faults
;
2355 spin_unlock(&my_grp
->lock
);
2356 spin_unlock_irq(&grp
->lock
);
2358 rcu_assign_pointer(p
->numa_group
, grp
);
2360 put_numa_group(my_grp
);
2368 void task_numa_free(struct task_struct
*p
)
2370 struct numa_group
*grp
= p
->numa_group
;
2371 void *numa_faults
= p
->numa_faults
;
2372 unsigned long flags
;
2376 spin_lock_irqsave(&grp
->lock
, flags
);
2377 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2378 grp
->faults
[i
] -= p
->numa_faults
[i
];
2379 grp
->total_faults
-= p
->total_numa_faults
;
2382 spin_unlock_irqrestore(&grp
->lock
, flags
);
2383 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2384 put_numa_group(grp
);
2387 p
->numa_faults
= NULL
;
2392 * Got a PROT_NONE fault for a page on @node.
2394 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2396 struct task_struct
*p
= current
;
2397 bool migrated
= flags
& TNF_MIGRATED
;
2398 int cpu_node
= task_node(current
);
2399 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2400 struct numa_group
*ng
;
2403 if (!static_branch_likely(&sched_numa_balancing
))
2406 /* for example, ksmd faulting in a user's mm */
2410 /* Allocate buffer to track faults on a per-node basis */
2411 if (unlikely(!p
->numa_faults
)) {
2412 int size
= sizeof(*p
->numa_faults
) *
2413 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2415 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2416 if (!p
->numa_faults
)
2419 p
->total_numa_faults
= 0;
2420 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2424 * First accesses are treated as private, otherwise consider accesses
2425 * to be private if the accessing pid has not changed
2427 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2430 priv
= cpupid_match_pid(p
, last_cpupid
);
2431 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2432 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2436 * If a workload spans multiple NUMA nodes, a shared fault that
2437 * occurs wholly within the set of nodes that the workload is
2438 * actively using should be counted as local. This allows the
2439 * scan rate to slow down when a workload has settled down.
2442 if (!priv
&& !local
&& ng
&& ng
->active_nodes
> 1 &&
2443 numa_is_active_node(cpu_node
, ng
) &&
2444 numa_is_active_node(mem_node
, ng
))
2447 task_numa_placement(p
);
2450 * Retry task to preferred node migration periodically, in case it
2451 * case it previously failed, or the scheduler moved us.
2453 if (time_after(jiffies
, p
->numa_migrate_retry
))
2454 numa_migrate_preferred(p
);
2457 p
->numa_pages_migrated
+= pages
;
2458 if (flags
& TNF_MIGRATE_FAIL
)
2459 p
->numa_faults_locality
[2] += pages
;
2461 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2462 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2463 p
->numa_faults_locality
[local
] += pages
;
2466 static void reset_ptenuma_scan(struct task_struct
*p
)
2469 * We only did a read acquisition of the mmap sem, so
2470 * p->mm->numa_scan_seq is written to without exclusive access
2471 * and the update is not guaranteed to be atomic. That's not
2472 * much of an issue though, since this is just used for
2473 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2474 * expensive, to avoid any form of compiler optimizations:
2476 WRITE_ONCE(p
->mm
->numa_scan_seq
, READ_ONCE(p
->mm
->numa_scan_seq
) + 1);
2477 p
->mm
->numa_scan_offset
= 0;
2481 * The expensive part of numa migration is done from task_work context.
2482 * Triggered from task_tick_numa().
2484 void task_numa_work(struct callback_head
*work
)
2486 unsigned long migrate
, next_scan
, now
= jiffies
;
2487 struct task_struct
*p
= current
;
2488 struct mm_struct
*mm
= p
->mm
;
2489 u64 runtime
= p
->se
.sum_exec_runtime
;
2490 struct vm_area_struct
*vma
;
2491 unsigned long start
, end
;
2492 unsigned long nr_pte_updates
= 0;
2493 long pages
, virtpages
;
2495 SCHED_WARN_ON(p
!= container_of(work
, struct task_struct
, numa_work
));
2497 work
->next
= work
; /* protect against double add */
2499 * Who cares about NUMA placement when they're dying.
2501 * NOTE: make sure not to dereference p->mm before this check,
2502 * exit_task_work() happens _after_ exit_mm() so we could be called
2503 * without p->mm even though we still had it when we enqueued this
2506 if (p
->flags
& PF_EXITING
)
2509 if (!mm
->numa_next_scan
) {
2510 mm
->numa_next_scan
= now
+
2511 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2515 * Enforce maximal scan/migration frequency..
2517 migrate
= mm
->numa_next_scan
;
2518 if (time_before(now
, migrate
))
2521 if (p
->numa_scan_period
== 0) {
2522 p
->numa_scan_period_max
= task_scan_max(p
);
2523 p
->numa_scan_period
= task_scan_start(p
);
2526 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2527 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2531 * Delay this task enough that another task of this mm will likely win
2532 * the next time around.
2534 p
->node_stamp
+= 2 * TICK_NSEC
;
2536 start
= mm
->numa_scan_offset
;
2537 pages
= sysctl_numa_balancing_scan_size
;
2538 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2539 virtpages
= pages
* 8; /* Scan up to this much virtual space */
2544 if (!down_read_trylock(&mm
->mmap_sem
))
2546 vma
= find_vma(mm
, start
);
2548 reset_ptenuma_scan(p
);
2552 for (; vma
; vma
= vma
->vm_next
) {
2553 if (!vma_migratable(vma
) || !vma_policy_mof(vma
) ||
2554 is_vm_hugetlb_page(vma
) || (vma
->vm_flags
& VM_MIXEDMAP
)) {
2559 * Shared library pages mapped by multiple processes are not
2560 * migrated as it is expected they are cache replicated. Avoid
2561 * hinting faults in read-only file-backed mappings or the vdso
2562 * as migrating the pages will be of marginal benefit.
2565 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2569 * Skip inaccessible VMAs to avoid any confusion between
2570 * PROT_NONE and NUMA hinting ptes
2572 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
2576 start
= max(start
, vma
->vm_start
);
2577 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2578 end
= min(end
, vma
->vm_end
);
2579 nr_pte_updates
= change_prot_numa(vma
, start
, end
);
2582 * Try to scan sysctl_numa_balancing_size worth of
2583 * hpages that have at least one present PTE that
2584 * is not already pte-numa. If the VMA contains
2585 * areas that are unused or already full of prot_numa
2586 * PTEs, scan up to virtpages, to skip through those
2590 pages
-= (end
- start
) >> PAGE_SHIFT
;
2591 virtpages
-= (end
- start
) >> PAGE_SHIFT
;
2594 if (pages
<= 0 || virtpages
<= 0)
2598 } while (end
!= vma
->vm_end
);
2603 * It is possible to reach the end of the VMA list but the last few
2604 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2605 * would find the !migratable VMA on the next scan but not reset the
2606 * scanner to the start so check it now.
2609 mm
->numa_scan_offset
= start
;
2611 reset_ptenuma_scan(p
);
2612 up_read(&mm
->mmap_sem
);
2615 * Make sure tasks use at least 32x as much time to run other code
2616 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2617 * Usually update_task_scan_period slows down scanning enough; on an
2618 * overloaded system we need to limit overhead on a per task basis.
2620 if (unlikely(p
->se
.sum_exec_runtime
!= runtime
)) {
2621 u64 diff
= p
->se
.sum_exec_runtime
- runtime
;
2622 p
->node_stamp
+= 32 * diff
;
2627 * Drive the periodic memory faults..
2629 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2631 struct callback_head
*work
= &curr
->numa_work
;
2635 * We don't care about NUMA placement if we don't have memory.
2637 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2641 * Using runtime rather than walltime has the dual advantage that
2642 * we (mostly) drive the selection from busy threads and that the
2643 * task needs to have done some actual work before we bother with
2646 now
= curr
->se
.sum_exec_runtime
;
2647 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2649 if (now
> curr
->node_stamp
+ period
) {
2650 if (!curr
->node_stamp
)
2651 curr
->numa_scan_period
= task_scan_start(curr
);
2652 curr
->node_stamp
+= period
;
2654 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2655 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2656 task_work_add(curr
, work
, true);
2662 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2666 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2670 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2674 #endif /* CONFIG_NUMA_BALANCING */
2677 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2679 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2680 if (!parent_entity(se
))
2681 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2683 if (entity_is_task(se
)) {
2684 struct rq
*rq
= rq_of(cfs_rq
);
2686 account_numa_enqueue(rq
, task_of(se
));
2687 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2690 cfs_rq
->nr_running
++;
2694 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2696 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2697 if (!parent_entity(se
))
2698 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2700 if (entity_is_task(se
)) {
2701 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2702 list_del_init(&se
->group_node
);
2705 cfs_rq
->nr_running
--;
2708 #ifdef CONFIG_FAIR_GROUP_SCHED
2710 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2712 long tg_weight
, load
, shares
;
2715 * This really should be: cfs_rq->avg.load_avg, but instead we use
2716 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2717 * the shares for small weight interactive tasks.
2719 load
= scale_load_down(cfs_rq
->load
.weight
);
2721 tg_weight
= atomic_long_read(&tg
->load_avg
);
2723 /* Ensure tg_weight >= load */
2724 tg_weight
-= cfs_rq
->tg_load_avg_contrib
;
2727 shares
= (tg
->shares
* load
);
2729 shares
/= tg_weight
;
2732 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2733 * of a group with small tg->shares value. It is a floor value which is
2734 * assigned as a minimum load.weight to the sched_entity representing
2735 * the group on a CPU.
2737 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2738 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2739 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2740 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2743 if (shares
< MIN_SHARES
)
2744 shares
= MIN_SHARES
;
2745 if (shares
> tg
->shares
)
2746 shares
= tg
->shares
;
2750 # else /* CONFIG_SMP */
2751 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2755 # endif /* CONFIG_SMP */
2757 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2758 unsigned long weight
)
2761 /* commit outstanding execution time */
2762 if (cfs_rq
->curr
== se
)
2763 update_curr(cfs_rq
);
2764 account_entity_dequeue(cfs_rq
, se
);
2767 update_load_set(&se
->load
, weight
);
2770 account_entity_enqueue(cfs_rq
, se
);
2773 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2775 static void update_cfs_shares(struct sched_entity
*se
)
2777 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2778 struct task_group
*tg
;
2784 if (throttled_hierarchy(cfs_rq
))
2790 if (likely(se
->load
.weight
== tg
->shares
))
2793 shares
= calc_cfs_shares(cfs_rq
, tg
);
2795 reweight_entity(cfs_rq_of(se
), se
, shares
);
2798 #else /* CONFIG_FAIR_GROUP_SCHED */
2799 static inline void update_cfs_shares(struct sched_entity
*se
)
2802 #endif /* CONFIG_FAIR_GROUP_SCHED */
2804 static inline void cfs_rq_util_change(struct cfs_rq
*cfs_rq
)
2806 if (&this_rq()->cfs
== cfs_rq
) {
2808 * There are a few boundary cases this might miss but it should
2809 * get called often enough that that should (hopefully) not be
2810 * a real problem -- added to that it only calls on the local
2811 * CPU, so if we enqueue remotely we'll miss an update, but
2812 * the next tick/schedule should update.
2814 * It will not get called when we go idle, because the idle
2815 * thread is a different class (!fair), nor will the utilization
2816 * number include things like RT tasks.
2818 * As is, the util number is not freq-invariant (we'd have to
2819 * implement arch_scale_freq_capacity() for that).
2823 cpufreq_update_util(rq_of(cfs_rq
), 0);
2830 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2832 static u64
decay_load(u64 val
, u64 n
)
2834 unsigned int local_n
;
2836 if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2839 /* after bounds checking we can collapse to 32-bit */
2843 * As y^PERIOD = 1/2, we can combine
2844 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2845 * With a look-up table which covers y^n (n<PERIOD)
2847 * To achieve constant time decay_load.
2849 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2850 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2851 local_n
%= LOAD_AVG_PERIOD
;
2854 val
= mul_u64_u32_shr(val
, runnable_avg_yN_inv
[local_n
], 32);
2858 static u32
__accumulate_pelt_segments(u64 periods
, u32 d1
, u32 d3
)
2860 u32 c1
, c2
, c3
= d3
; /* y^0 == 1 */
2865 c1
= decay_load((u64
)d1
, periods
);
2869 * c2 = 1024 \Sum y^n
2873 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
2876 c2
= LOAD_AVG_MAX
- decay_load(LOAD_AVG_MAX
, periods
) - 1024;
2878 return c1
+ c2
+ c3
;
2881 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2884 * Accumulate the three separate parts of the sum; d1 the remainder
2885 * of the last (incomplete) period, d2 the span of full periods and d3
2886 * the remainder of the (incomplete) current period.
2891 * |<->|<----------------->|<--->|
2892 * ... |---x---|------| ... |------|-----x (now)
2895 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
2898 * = u y^p + (Step 1)
2901 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
2904 static __always_inline u32
2905 accumulate_sum(u64 delta
, int cpu
, struct sched_avg
*sa
,
2906 unsigned long weight
, int running
, struct cfs_rq
*cfs_rq
)
2908 unsigned long scale_freq
, scale_cpu
;
2909 u32 contrib
= (u32
)delta
; /* p == 0 -> delta < 1024 */
2912 scale_freq
= arch_scale_freq_capacity(NULL
, cpu
);
2913 scale_cpu
= arch_scale_cpu_capacity(NULL
, cpu
);
2915 delta
+= sa
->period_contrib
;
2916 periods
= delta
/ 1024; /* A period is 1024us (~1ms) */
2919 * Step 1: decay old *_sum if we crossed period boundaries.
2922 sa
->load_sum
= decay_load(sa
->load_sum
, periods
);
2924 cfs_rq
->runnable_load_sum
=
2925 decay_load(cfs_rq
->runnable_load_sum
, periods
);
2927 sa
->util_sum
= decay_load((u64
)(sa
->util_sum
), periods
);
2933 contrib
= __accumulate_pelt_segments(periods
,
2934 1024 - sa
->period_contrib
, delta
);
2936 sa
->period_contrib
= delta
;
2938 contrib
= cap_scale(contrib
, scale_freq
);
2940 sa
->load_sum
+= weight
* contrib
;
2942 cfs_rq
->runnable_load_sum
+= weight
* contrib
;
2945 sa
->util_sum
+= contrib
* scale_cpu
;
2951 * We can represent the historical contribution to runnable average as the
2952 * coefficients of a geometric series. To do this we sub-divide our runnable
2953 * history into segments of approximately 1ms (1024us); label the segment that
2954 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2956 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2958 * (now) (~1ms ago) (~2ms ago)
2960 * Let u_i denote the fraction of p_i that the entity was runnable.
2962 * We then designate the fractions u_i as our co-efficients, yielding the
2963 * following representation of historical load:
2964 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2966 * We choose y based on the with of a reasonably scheduling period, fixing:
2969 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2970 * approximately half as much as the contribution to load within the last ms
2973 * When a period "rolls over" and we have new u_0`, multiplying the previous
2974 * sum again by y is sufficient to update:
2975 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2976 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2978 static __always_inline
int
2979 ___update_load_avg(u64 now
, int cpu
, struct sched_avg
*sa
,
2980 unsigned long weight
, int running
, struct cfs_rq
*cfs_rq
)
2984 delta
= now
- sa
->last_update_time
;
2986 * This should only happen when time goes backwards, which it
2987 * unfortunately does during sched clock init when we swap over to TSC.
2989 if ((s64
)delta
< 0) {
2990 sa
->last_update_time
= now
;
2995 * Use 1024ns as the unit of measurement since it's a reasonable
2996 * approximation of 1us and fast to compute.
3002 sa
->last_update_time
+= delta
<< 10;
3005 * running is a subset of runnable (weight) so running can't be set if
3006 * runnable is clear. But there are some corner cases where the current
3007 * se has been already dequeued but cfs_rq->curr still points to it.
3008 * This means that weight will be 0 but not running for a sched_entity
3009 * but also for a cfs_rq if the latter becomes idle. As an example,
3010 * this happens during idle_balance() which calls
3011 * update_blocked_averages()
3017 * Now we know we crossed measurement unit boundaries. The *_avg
3018 * accrues by two steps:
3020 * Step 1: accumulate *_sum since last_update_time. If we haven't
3021 * crossed period boundaries, finish.
3023 if (!accumulate_sum(delta
, cpu
, sa
, weight
, running
, cfs_rq
))
3027 * Step 2: update *_avg.
3029 sa
->load_avg
= div_u64(sa
->load_sum
, LOAD_AVG_MAX
- 1024 + sa
->period_contrib
);
3031 cfs_rq
->runnable_load_avg
=
3032 div_u64(cfs_rq
->runnable_load_sum
, LOAD_AVG_MAX
- 1024 + sa
->period_contrib
);
3034 sa
->util_avg
= sa
->util_sum
/ (LOAD_AVG_MAX
- 1024 + sa
->period_contrib
);
3040 __update_load_avg_blocked_se(u64 now
, int cpu
, struct sched_entity
*se
)
3042 return ___update_load_avg(now
, cpu
, &se
->avg
, 0, 0, NULL
);
3046 __update_load_avg_se(u64 now
, int cpu
, struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3048 return ___update_load_avg(now
, cpu
, &se
->avg
,
3049 se
->on_rq
* scale_load_down(se
->load
.weight
),
3050 cfs_rq
->curr
== se
, NULL
);
3054 __update_load_avg_cfs_rq(u64 now
, int cpu
, struct cfs_rq
*cfs_rq
)
3056 return ___update_load_avg(now
, cpu
, &cfs_rq
->avg
,
3057 scale_load_down(cfs_rq
->load
.weight
),
3058 cfs_rq
->curr
!= NULL
, cfs_rq
);
3062 * Signed add and clamp on underflow.
3064 * Explicitly do a load-store to ensure the intermediate value never hits
3065 * memory. This allows lockless observations without ever seeing the negative
3068 #define add_positive(_ptr, _val) do { \
3069 typeof(_ptr) ptr = (_ptr); \
3070 typeof(_val) val = (_val); \
3071 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3075 if (val < 0 && res > var) \
3078 WRITE_ONCE(*ptr, res); \
3081 #ifdef CONFIG_FAIR_GROUP_SCHED
3083 * update_tg_load_avg - update the tg's load avg
3084 * @cfs_rq: the cfs_rq whose avg changed
3085 * @force: update regardless of how small the difference
3087 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3088 * However, because tg->load_avg is a global value there are performance
3091 * In order to avoid having to look at the other cfs_rq's, we use a
3092 * differential update where we store the last value we propagated. This in
3093 * turn allows skipping updates if the differential is 'small'.
3095 * Updating tg's load_avg is necessary before update_cfs_share().
3097 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
)
3099 long delta
= cfs_rq
->avg
.load_avg
- cfs_rq
->tg_load_avg_contrib
;
3102 * No need to update load_avg for root_task_group as it is not used.
3104 if (cfs_rq
->tg
== &root_task_group
)
3107 if (force
|| abs(delta
) > cfs_rq
->tg_load_avg_contrib
/ 64) {
3108 atomic_long_add(delta
, &cfs_rq
->tg
->load_avg
);
3109 cfs_rq
->tg_load_avg_contrib
= cfs_rq
->avg
.load_avg
;
3114 * Called within set_task_rq() right before setting a task's cpu. The
3115 * caller only guarantees p->pi_lock is held; no other assumptions,
3116 * including the state of rq->lock, should be made.
3118 void set_task_rq_fair(struct sched_entity
*se
,
3119 struct cfs_rq
*prev
, struct cfs_rq
*next
)
3121 u64 p_last_update_time
;
3122 u64 n_last_update_time
;
3124 if (!sched_feat(ATTACH_AGE_LOAD
))
3128 * We are supposed to update the task to "current" time, then its up to
3129 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3130 * getting what current time is, so simply throw away the out-of-date
3131 * time. This will result in the wakee task is less decayed, but giving
3132 * the wakee more load sounds not bad.
3134 if (!(se
->avg
.last_update_time
&& prev
))
3137 #ifndef CONFIG_64BIT
3139 u64 p_last_update_time_copy
;
3140 u64 n_last_update_time_copy
;
3143 p_last_update_time_copy
= prev
->load_last_update_time_copy
;
3144 n_last_update_time_copy
= next
->load_last_update_time_copy
;
3148 p_last_update_time
= prev
->avg
.last_update_time
;
3149 n_last_update_time
= next
->avg
.last_update_time
;
3151 } while (p_last_update_time
!= p_last_update_time_copy
||
3152 n_last_update_time
!= n_last_update_time_copy
);
3155 p_last_update_time
= prev
->avg
.last_update_time
;
3156 n_last_update_time
= next
->avg
.last_update_time
;
3158 __update_load_avg_blocked_se(p_last_update_time
, cpu_of(rq_of(prev
)), se
);
3159 se
->avg
.last_update_time
= n_last_update_time
;
3162 /* Take into account change of utilization of a child task group */
3164 update_tg_cfs_util(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3166 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3167 long delta
= gcfs_rq
->avg
.util_avg
- se
->avg
.util_avg
;
3169 /* Nothing to update */
3173 /* Set new sched_entity's utilization */
3174 se
->avg
.util_avg
= gcfs_rq
->avg
.util_avg
;
3175 se
->avg
.util_sum
= se
->avg
.util_avg
* LOAD_AVG_MAX
;
3177 /* Update parent cfs_rq utilization */
3178 add_positive(&cfs_rq
->avg
.util_avg
, delta
);
3179 cfs_rq
->avg
.util_sum
= cfs_rq
->avg
.util_avg
* LOAD_AVG_MAX
;
3182 /* Take into account change of load of a child task group */
3184 update_tg_cfs_load(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3186 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3187 long delta
, load
= gcfs_rq
->avg
.load_avg
;
3190 * If the load of group cfs_rq is null, the load of the
3191 * sched_entity will also be null so we can skip the formula
3196 /* Get tg's load and ensure tg_load > 0 */
3197 tg_load
= atomic_long_read(&gcfs_rq
->tg
->load_avg
) + 1;
3199 /* Ensure tg_load >= load and updated with current load*/
3200 tg_load
-= gcfs_rq
->tg_load_avg_contrib
;
3204 * We need to compute a correction term in the case that the
3205 * task group is consuming more CPU than a task of equal
3206 * weight. A task with a weight equals to tg->shares will have
3207 * a load less or equal to scale_load_down(tg->shares).
3208 * Similarly, the sched_entities that represent the task group
3209 * at parent level, can't have a load higher than
3210 * scale_load_down(tg->shares). And the Sum of sched_entities'
3211 * load must be <= scale_load_down(tg->shares).
3213 if (tg_load
> scale_load_down(gcfs_rq
->tg
->shares
)) {
3214 /* scale gcfs_rq's load into tg's shares*/
3215 load
*= scale_load_down(gcfs_rq
->tg
->shares
);
3220 delta
= load
- se
->avg
.load_avg
;
3222 /* Nothing to update */
3226 /* Set new sched_entity's load */
3227 se
->avg
.load_avg
= load
;
3228 se
->avg
.load_sum
= se
->avg
.load_avg
* LOAD_AVG_MAX
;
3230 /* Update parent cfs_rq load */
3231 add_positive(&cfs_rq
->avg
.load_avg
, delta
);
3232 cfs_rq
->avg
.load_sum
= cfs_rq
->avg
.load_avg
* LOAD_AVG_MAX
;
3235 * If the sched_entity is already enqueued, we also have to update the
3236 * runnable load avg.
3239 /* Update parent cfs_rq runnable_load_avg */
3240 add_positive(&cfs_rq
->runnable_load_avg
, delta
);
3241 cfs_rq
->runnable_load_sum
= cfs_rq
->runnable_load_avg
* LOAD_AVG_MAX
;
3245 static inline void set_tg_cfs_propagate(struct cfs_rq
*cfs_rq
)
3247 cfs_rq
->propagate_avg
= 1;
3250 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity
*se
)
3252 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
3254 if (!cfs_rq
->propagate_avg
)
3257 cfs_rq
->propagate_avg
= 0;
3261 /* Update task and its cfs_rq load average */
3262 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3264 struct cfs_rq
*cfs_rq
;
3266 if (entity_is_task(se
))
3269 if (!test_and_clear_tg_cfs_propagate(se
))
3272 cfs_rq
= cfs_rq_of(se
);
3274 set_tg_cfs_propagate(cfs_rq
);
3276 update_tg_cfs_util(cfs_rq
, se
);
3277 update_tg_cfs_load(cfs_rq
, se
);
3283 * Check if we need to update the load and the utilization of a blocked
3286 static inline bool skip_blocked_update(struct sched_entity
*se
)
3288 struct cfs_rq
*gcfs_rq
= group_cfs_rq(se
);
3291 * If sched_entity still have not zero load or utilization, we have to
3294 if (se
->avg
.load_avg
|| se
->avg
.util_avg
)
3298 * If there is a pending propagation, we have to update the load and
3299 * the utilization of the sched_entity:
3301 if (gcfs_rq
->propagate_avg
)
3305 * Otherwise, the load and the utilization of the sched_entity is
3306 * already zero and there is no pending propagation, so it will be a
3307 * waste of time to try to decay it:
3312 #else /* CONFIG_FAIR_GROUP_SCHED */
3314 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
) {}
3316 static inline int propagate_entity_load_avg(struct sched_entity
*se
)
3321 static inline void set_tg_cfs_propagate(struct cfs_rq
*cfs_rq
) {}
3323 #endif /* CONFIG_FAIR_GROUP_SCHED */
3326 * Unsigned subtract and clamp on underflow.
3328 * Explicitly do a load-store to ensure the intermediate value never hits
3329 * memory. This allows lockless observations without ever seeing the negative
3332 #define sub_positive(_ptr, _val) do { \
3333 typeof(_ptr) ptr = (_ptr); \
3334 typeof(*ptr) val = (_val); \
3335 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3339 WRITE_ONCE(*ptr, res); \
3343 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3344 * @now: current time, as per cfs_rq_clock_task()
3345 * @cfs_rq: cfs_rq to update
3347 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3348 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3349 * post_init_entity_util_avg().
3351 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3353 * Returns true if the load decayed or we removed load.
3355 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3356 * call update_tg_load_avg() when this function returns true.
3359 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
)
3361 struct sched_avg
*sa
= &cfs_rq
->avg
;
3362 int decayed
, removed_load
= 0, removed_util
= 0;
3364 if (atomic_long_read(&cfs_rq
->removed_load_avg
)) {
3365 s64 r
= atomic_long_xchg(&cfs_rq
->removed_load_avg
, 0);
3366 sub_positive(&sa
->load_avg
, r
);
3367 sub_positive(&sa
->load_sum
, r
* LOAD_AVG_MAX
);
3369 set_tg_cfs_propagate(cfs_rq
);
3372 if (atomic_long_read(&cfs_rq
->removed_util_avg
)) {
3373 long r
= atomic_long_xchg(&cfs_rq
->removed_util_avg
, 0);
3374 sub_positive(&sa
->util_avg
, r
);
3375 sub_positive(&sa
->util_sum
, r
* LOAD_AVG_MAX
);
3377 set_tg_cfs_propagate(cfs_rq
);
3380 decayed
= __update_load_avg_cfs_rq(now
, cpu_of(rq_of(cfs_rq
)), cfs_rq
);
3382 #ifndef CONFIG_64BIT
3384 cfs_rq
->load_last_update_time_copy
= sa
->last_update_time
;
3387 if (decayed
|| removed_util
)
3388 cfs_rq_util_change(cfs_rq
);
3390 return decayed
|| removed_load
;
3394 * Optional action to be done while updating the load average
3396 #define UPDATE_TG 0x1
3397 #define SKIP_AGE_LOAD 0x2
3399 /* Update task and its cfs_rq load average */
3400 static inline void update_load_avg(struct sched_entity
*se
, int flags
)
3402 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3403 u64 now
= cfs_rq_clock_task(cfs_rq
);
3404 struct rq
*rq
= rq_of(cfs_rq
);
3405 int cpu
= cpu_of(rq
);
3409 * Track task load average for carrying it to new CPU after migrated, and
3410 * track group sched_entity load average for task_h_load calc in migration
3412 if (se
->avg
.last_update_time
&& !(flags
& SKIP_AGE_LOAD
))
3413 __update_load_avg_se(now
, cpu
, cfs_rq
, se
);
3415 decayed
= update_cfs_rq_load_avg(now
, cfs_rq
);
3416 decayed
|= propagate_entity_load_avg(se
);
3418 if (decayed
&& (flags
& UPDATE_TG
))
3419 update_tg_load_avg(cfs_rq
, 0);
3423 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3424 * @cfs_rq: cfs_rq to attach to
3425 * @se: sched_entity to attach
3427 * Must call update_cfs_rq_load_avg() before this, since we rely on
3428 * cfs_rq->avg.last_update_time being current.
3430 static void attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3432 se
->avg
.last_update_time
= cfs_rq
->avg
.last_update_time
;
3433 cfs_rq
->avg
.load_avg
+= se
->avg
.load_avg
;
3434 cfs_rq
->avg
.load_sum
+= se
->avg
.load_sum
;
3435 cfs_rq
->avg
.util_avg
+= se
->avg
.util_avg
;
3436 cfs_rq
->avg
.util_sum
+= se
->avg
.util_sum
;
3437 set_tg_cfs_propagate(cfs_rq
);
3439 cfs_rq_util_change(cfs_rq
);
3443 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3444 * @cfs_rq: cfs_rq to detach from
3445 * @se: sched_entity to detach
3447 * Must call update_cfs_rq_load_avg() before this, since we rely on
3448 * cfs_rq->avg.last_update_time being current.
3450 static void detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3453 sub_positive(&cfs_rq
->avg
.load_avg
, se
->avg
.load_avg
);
3454 sub_positive(&cfs_rq
->avg
.load_sum
, se
->avg
.load_sum
);
3455 sub_positive(&cfs_rq
->avg
.util_avg
, se
->avg
.util_avg
);
3456 sub_positive(&cfs_rq
->avg
.util_sum
, se
->avg
.util_sum
);
3457 set_tg_cfs_propagate(cfs_rq
);
3459 cfs_rq_util_change(cfs_rq
);
3462 /* Add the load generated by se into cfs_rq's load average */
3464 enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3466 struct sched_avg
*sa
= &se
->avg
;
3468 cfs_rq
->runnable_load_avg
+= sa
->load_avg
;
3469 cfs_rq
->runnable_load_sum
+= sa
->load_sum
;
3471 if (!sa
->last_update_time
) {
3472 attach_entity_load_avg(cfs_rq
, se
);
3473 update_tg_load_avg(cfs_rq
, 0);
3477 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3479 dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3481 cfs_rq
->runnable_load_avg
=
3482 max_t(long, cfs_rq
->runnable_load_avg
- se
->avg
.load_avg
, 0);
3483 cfs_rq
->runnable_load_sum
=
3484 max_t(s64
, cfs_rq
->runnable_load_sum
- se
->avg
.load_sum
, 0);
3487 #ifndef CONFIG_64BIT
3488 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3490 u64 last_update_time_copy
;
3491 u64 last_update_time
;
3494 last_update_time_copy
= cfs_rq
->load_last_update_time_copy
;
3496 last_update_time
= cfs_rq
->avg
.last_update_time
;
3497 } while (last_update_time
!= last_update_time_copy
);
3499 return last_update_time
;
3502 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3504 return cfs_rq
->avg
.last_update_time
;
3509 * Synchronize entity load avg of dequeued entity without locking
3512 void sync_entity_load_avg(struct sched_entity
*se
)
3514 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3515 u64 last_update_time
;
3517 last_update_time
= cfs_rq_last_update_time(cfs_rq
);
3518 __update_load_avg_blocked_se(last_update_time
, cpu_of(rq_of(cfs_rq
)), se
);
3522 * Task first catches up with cfs_rq, and then subtract
3523 * itself from the cfs_rq (task must be off the queue now).
3525 void remove_entity_load_avg(struct sched_entity
*se
)
3527 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3530 * tasks cannot exit without having gone through wake_up_new_task() ->
3531 * post_init_entity_util_avg() which will have added things to the
3532 * cfs_rq, so we can remove unconditionally.
3534 * Similarly for groups, they will have passed through
3535 * post_init_entity_util_avg() before unregister_sched_fair_group()
3539 sync_entity_load_avg(se
);
3540 atomic_long_add(se
->avg
.load_avg
, &cfs_rq
->removed_load_avg
);
3541 atomic_long_add(se
->avg
.util_avg
, &cfs_rq
->removed_util_avg
);
3544 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq
*cfs_rq
)
3546 return cfs_rq
->runnable_load_avg
;
3549 static inline unsigned long cfs_rq_load_avg(struct cfs_rq
*cfs_rq
)
3551 return cfs_rq
->avg
.load_avg
;
3554 static int idle_balance(struct rq
*this_rq
, struct rq_flags
*rf
);
3556 #else /* CONFIG_SMP */
3559 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
)
3564 #define UPDATE_TG 0x0
3565 #define SKIP_AGE_LOAD 0x0
3567 static inline void update_load_avg(struct sched_entity
*se
, int not_used1
)
3569 cfs_rq_util_change(cfs_rq_of(se
));
3573 enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3575 dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3576 static inline void remove_entity_load_avg(struct sched_entity
*se
) {}
3579 attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3581 detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3583 static inline int idle_balance(struct rq
*rq
, struct rq_flags
*rf
)
3588 #endif /* CONFIG_SMP */
3590 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3592 #ifdef CONFIG_SCHED_DEBUG
3593 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
3598 if (d
> 3*sysctl_sched_latency
)
3599 schedstat_inc(cfs_rq
->nr_spread_over
);
3604 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
3606 u64 vruntime
= cfs_rq
->min_vruntime
;
3609 * The 'current' period is already promised to the current tasks,
3610 * however the extra weight of the new task will slow them down a
3611 * little, place the new task so that it fits in the slot that
3612 * stays open at the end.
3614 if (initial
&& sched_feat(START_DEBIT
))
3615 vruntime
+= sched_vslice(cfs_rq
, se
);
3617 /* sleeps up to a single latency don't count. */
3619 unsigned long thresh
= sysctl_sched_latency
;
3622 * Halve their sleep time's effect, to allow
3623 * for a gentler effect of sleepers:
3625 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
3631 /* ensure we never gain time by being placed backwards. */
3632 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
3635 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
3637 static inline void check_schedstat_required(void)
3639 #ifdef CONFIG_SCHEDSTATS
3640 if (schedstat_enabled())
3643 /* Force schedstat enabled if a dependent tracepoint is active */
3644 if (trace_sched_stat_wait_enabled() ||
3645 trace_sched_stat_sleep_enabled() ||
3646 trace_sched_stat_iowait_enabled() ||
3647 trace_sched_stat_blocked_enabled() ||
3648 trace_sched_stat_runtime_enabled()) {
3649 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3650 "stat_blocked and stat_runtime require the "
3651 "kernel parameter schedstats=enable or "
3652 "kernel.sched_schedstats=1\n");
3663 * update_min_vruntime()
3664 * vruntime -= min_vruntime
3668 * update_min_vruntime()
3669 * vruntime += min_vruntime
3671 * this way the vruntime transition between RQs is done when both
3672 * min_vruntime are up-to-date.
3676 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3677 * vruntime -= min_vruntime
3681 * update_min_vruntime()
3682 * vruntime += min_vruntime
3684 * this way we don't have the most up-to-date min_vruntime on the originating
3685 * CPU and an up-to-date min_vruntime on the destination CPU.
3689 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3691 bool renorm
= !(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_MIGRATED
);
3692 bool curr
= cfs_rq
->curr
== se
;
3695 * If we're the current task, we must renormalise before calling
3699 se
->vruntime
+= cfs_rq
->min_vruntime
;
3701 update_curr(cfs_rq
);
3704 * Otherwise, renormalise after, such that we're placed at the current
3705 * moment in time, instead of some random moment in the past. Being
3706 * placed in the past could significantly boost this task to the
3707 * fairness detriment of existing tasks.
3709 if (renorm
&& !curr
)
3710 se
->vruntime
+= cfs_rq
->min_vruntime
;
3713 * When enqueuing a sched_entity, we must:
3714 * - Update loads to have both entity and cfs_rq synced with now.
3715 * - Add its load to cfs_rq->runnable_avg
3716 * - For group_entity, update its weight to reflect the new share of
3718 * - Add its new weight to cfs_rq->load.weight
3720 update_load_avg(se
, UPDATE_TG
);
3721 enqueue_entity_load_avg(cfs_rq
, se
);
3722 update_cfs_shares(se
);
3723 account_entity_enqueue(cfs_rq
, se
);
3725 if (flags
& ENQUEUE_WAKEUP
)
3726 place_entity(cfs_rq
, se
, 0);
3728 check_schedstat_required();
3729 update_stats_enqueue(cfs_rq
, se
, flags
);
3730 check_spread(cfs_rq
, se
);
3732 __enqueue_entity(cfs_rq
, se
);
3735 if (cfs_rq
->nr_running
== 1) {
3736 list_add_leaf_cfs_rq(cfs_rq
);
3737 check_enqueue_throttle(cfs_rq
);
3741 static void __clear_buddies_last(struct sched_entity
*se
)
3743 for_each_sched_entity(se
) {
3744 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3745 if (cfs_rq
->last
!= se
)
3748 cfs_rq
->last
= NULL
;
3752 static void __clear_buddies_next(struct sched_entity
*se
)
3754 for_each_sched_entity(se
) {
3755 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3756 if (cfs_rq
->next
!= se
)
3759 cfs_rq
->next
= NULL
;
3763 static void __clear_buddies_skip(struct sched_entity
*se
)
3765 for_each_sched_entity(se
) {
3766 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3767 if (cfs_rq
->skip
!= se
)
3770 cfs_rq
->skip
= NULL
;
3774 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3776 if (cfs_rq
->last
== se
)
3777 __clear_buddies_last(se
);
3779 if (cfs_rq
->next
== se
)
3780 __clear_buddies_next(se
);
3782 if (cfs_rq
->skip
== se
)
3783 __clear_buddies_skip(se
);
3786 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3789 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3792 * Update run-time statistics of the 'current'.
3794 update_curr(cfs_rq
);
3797 * When dequeuing a sched_entity, we must:
3798 * - Update loads to have both entity and cfs_rq synced with now.
3799 * - Substract its load from the cfs_rq->runnable_avg.
3800 * - Substract its previous weight from cfs_rq->load.weight.
3801 * - For group entity, update its weight to reflect the new share
3802 * of its group cfs_rq.
3804 update_load_avg(se
, UPDATE_TG
);
3805 dequeue_entity_load_avg(cfs_rq
, se
);
3807 update_stats_dequeue(cfs_rq
, se
, flags
);
3809 clear_buddies(cfs_rq
, se
);
3811 if (se
!= cfs_rq
->curr
)
3812 __dequeue_entity(cfs_rq
, se
);
3814 account_entity_dequeue(cfs_rq
, se
);
3817 * Normalize after update_curr(); which will also have moved
3818 * min_vruntime if @se is the one holding it back. But before doing
3819 * update_min_vruntime() again, which will discount @se's position and
3820 * can move min_vruntime forward still more.
3822 if (!(flags
& DEQUEUE_SLEEP
))
3823 se
->vruntime
-= cfs_rq
->min_vruntime
;
3825 /* return excess runtime on last dequeue */
3826 return_cfs_rq_runtime(cfs_rq
);
3828 update_cfs_shares(se
);
3831 * Now advance min_vruntime if @se was the entity holding it back,
3832 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3833 * put back on, and if we advance min_vruntime, we'll be placed back
3834 * further than we started -- ie. we'll be penalized.
3836 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) == DEQUEUE_SAVE
)
3837 update_min_vruntime(cfs_rq
);
3841 * Preempt the current task with a newly woken task if needed:
3844 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3846 unsigned long ideal_runtime
, delta_exec
;
3847 struct sched_entity
*se
;
3850 ideal_runtime
= sched_slice(cfs_rq
, curr
);
3851 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
3852 if (delta_exec
> ideal_runtime
) {
3853 resched_curr(rq_of(cfs_rq
));
3855 * The current task ran long enough, ensure it doesn't get
3856 * re-elected due to buddy favours.
3858 clear_buddies(cfs_rq
, curr
);
3863 * Ensure that a task that missed wakeup preemption by a
3864 * narrow margin doesn't have to wait for a full slice.
3865 * This also mitigates buddy induced latencies under load.
3867 if (delta_exec
< sysctl_sched_min_granularity
)
3870 se
= __pick_first_entity(cfs_rq
);
3871 delta
= curr
->vruntime
- se
->vruntime
;
3876 if (delta
> ideal_runtime
)
3877 resched_curr(rq_of(cfs_rq
));
3881 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3883 /* 'current' is not kept within the tree. */
3886 * Any task has to be enqueued before it get to execute on
3887 * a CPU. So account for the time it spent waiting on the
3890 update_stats_wait_end(cfs_rq
, se
);
3891 __dequeue_entity(cfs_rq
, se
);
3892 update_load_avg(se
, UPDATE_TG
);
3895 update_stats_curr_start(cfs_rq
, se
);
3899 * Track our maximum slice length, if the CPU's load is at
3900 * least twice that of our own weight (i.e. dont track it
3901 * when there are only lesser-weight tasks around):
3903 if (schedstat_enabled() && rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
3904 schedstat_set(se
->statistics
.slice_max
,
3905 max((u64
)schedstat_val(se
->statistics
.slice_max
),
3906 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
));
3909 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
3913 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
3916 * Pick the next process, keeping these things in mind, in this order:
3917 * 1) keep things fair between processes/task groups
3918 * 2) pick the "next" process, since someone really wants that to run
3919 * 3) pick the "last" process, for cache locality
3920 * 4) do not run the "skip" process, if something else is available
3922 static struct sched_entity
*
3923 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3925 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
3926 struct sched_entity
*se
;
3929 * If curr is set we have to see if its left of the leftmost entity
3930 * still in the tree, provided there was anything in the tree at all.
3932 if (!left
|| (curr
&& entity_before(curr
, left
)))
3935 se
= left
; /* ideally we run the leftmost entity */
3938 * Avoid running the skip buddy, if running something else can
3939 * be done without getting too unfair.
3941 if (cfs_rq
->skip
== se
) {
3942 struct sched_entity
*second
;
3945 second
= __pick_first_entity(cfs_rq
);
3947 second
= __pick_next_entity(se
);
3948 if (!second
|| (curr
&& entity_before(curr
, second
)))
3952 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3957 * Prefer last buddy, try to return the CPU to a preempted task.
3959 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3963 * Someone really wants this to run. If it's not unfair, run it.
3965 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3968 clear_buddies(cfs_rq
, se
);
3973 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3975 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3978 * If still on the runqueue then deactivate_task()
3979 * was not called and update_curr() has to be done:
3982 update_curr(cfs_rq
);
3984 /* throttle cfs_rqs exceeding runtime */
3985 check_cfs_rq_runtime(cfs_rq
);
3987 check_spread(cfs_rq
, prev
);
3990 update_stats_wait_start(cfs_rq
, prev
);
3991 /* Put 'current' back into the tree. */
3992 __enqueue_entity(cfs_rq
, prev
);
3993 /* in !on_rq case, update occurred at dequeue */
3994 update_load_avg(prev
, 0);
3996 cfs_rq
->curr
= NULL
;
4000 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
4003 * Update run-time statistics of the 'current'.
4005 update_curr(cfs_rq
);
4008 * Ensure that runnable average is periodically updated.
4010 update_load_avg(curr
, UPDATE_TG
);
4011 update_cfs_shares(curr
);
4013 #ifdef CONFIG_SCHED_HRTICK
4015 * queued ticks are scheduled to match the slice, so don't bother
4016 * validating it and just reschedule.
4019 resched_curr(rq_of(cfs_rq
));
4023 * don't let the period tick interfere with the hrtick preemption
4025 if (!sched_feat(DOUBLE_TICK
) &&
4026 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
4030 if (cfs_rq
->nr_running
> 1)
4031 check_preempt_tick(cfs_rq
, curr
);
4035 /**************************************************
4036 * CFS bandwidth control machinery
4039 #ifdef CONFIG_CFS_BANDWIDTH
4041 #ifdef HAVE_JUMP_LABEL
4042 static struct static_key __cfs_bandwidth_used
;
4044 static inline bool cfs_bandwidth_used(void)
4046 return static_key_false(&__cfs_bandwidth_used
);
4049 void cfs_bandwidth_usage_inc(void)
4051 static_key_slow_inc(&__cfs_bandwidth_used
);
4054 void cfs_bandwidth_usage_dec(void)
4056 static_key_slow_dec(&__cfs_bandwidth_used
);
4058 #else /* HAVE_JUMP_LABEL */
4059 static bool cfs_bandwidth_used(void)
4064 void cfs_bandwidth_usage_inc(void) {}
4065 void cfs_bandwidth_usage_dec(void) {}
4066 #endif /* HAVE_JUMP_LABEL */
4069 * default period for cfs group bandwidth.
4070 * default: 0.1s, units: nanoseconds
4072 static inline u64
default_cfs_period(void)
4074 return 100000000ULL;
4077 static inline u64
sched_cfs_bandwidth_slice(void)
4079 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
4083 * Replenish runtime according to assigned quota and update expiration time.
4084 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4085 * additional synchronization around rq->lock.
4087 * requires cfs_b->lock
4089 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
4093 if (cfs_b
->quota
== RUNTIME_INF
)
4096 now
= sched_clock_cpu(smp_processor_id());
4097 cfs_b
->runtime
= cfs_b
->quota
;
4098 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
4101 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4103 return &tg
->cfs_bandwidth
;
4106 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4107 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
4109 if (unlikely(cfs_rq
->throttle_count
))
4110 return cfs_rq
->throttled_clock_task
- cfs_rq
->throttled_clock_task_time
;
4112 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
4115 /* returns 0 on failure to allocate runtime */
4116 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4118 struct task_group
*tg
= cfs_rq
->tg
;
4119 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
4120 u64 amount
= 0, min_amount
, expires
;
4122 /* note: this is a positive sum as runtime_remaining <= 0 */
4123 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
4125 raw_spin_lock(&cfs_b
->lock
);
4126 if (cfs_b
->quota
== RUNTIME_INF
)
4127 amount
= min_amount
;
4129 start_cfs_bandwidth(cfs_b
);
4131 if (cfs_b
->runtime
> 0) {
4132 amount
= min(cfs_b
->runtime
, min_amount
);
4133 cfs_b
->runtime
-= amount
;
4137 expires
= cfs_b
->runtime_expires
;
4138 raw_spin_unlock(&cfs_b
->lock
);
4140 cfs_rq
->runtime_remaining
+= amount
;
4142 * we may have advanced our local expiration to account for allowed
4143 * spread between our sched_clock and the one on which runtime was
4146 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
4147 cfs_rq
->runtime_expires
= expires
;
4149 return cfs_rq
->runtime_remaining
> 0;
4153 * Note: This depends on the synchronization provided by sched_clock and the
4154 * fact that rq->clock snapshots this value.
4156 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4158 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4160 /* if the deadline is ahead of our clock, nothing to do */
4161 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
4164 if (cfs_rq
->runtime_remaining
< 0)
4168 * If the local deadline has passed we have to consider the
4169 * possibility that our sched_clock is 'fast' and the global deadline
4170 * has not truly expired.
4172 * Fortunately we can check determine whether this the case by checking
4173 * whether the global deadline has advanced. It is valid to compare
4174 * cfs_b->runtime_expires without any locks since we only care about
4175 * exact equality, so a partial write will still work.
4178 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
4179 /* extend local deadline, drift is bounded above by 2 ticks */
4180 cfs_rq
->runtime_expires
+= TICK_NSEC
;
4182 /* global deadline is ahead, expiration has passed */
4183 cfs_rq
->runtime_remaining
= 0;
4187 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4189 /* dock delta_exec before expiring quota (as it could span periods) */
4190 cfs_rq
->runtime_remaining
-= delta_exec
;
4191 expire_cfs_rq_runtime(cfs_rq
);
4193 if (likely(cfs_rq
->runtime_remaining
> 0))
4197 * if we're unable to extend our runtime we resched so that the active
4198 * hierarchy can be throttled
4200 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
4201 resched_curr(rq_of(cfs_rq
));
4204 static __always_inline
4205 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
4207 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
4210 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
4213 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4215 return cfs_bandwidth_used() && cfs_rq
->throttled
;
4218 /* check whether cfs_rq, or any parent, is throttled */
4219 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4221 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
4225 * Ensure that neither of the group entities corresponding to src_cpu or
4226 * dest_cpu are members of a throttled hierarchy when performing group
4227 * load-balance operations.
4229 static inline int throttled_lb_pair(struct task_group
*tg
,
4230 int src_cpu
, int dest_cpu
)
4232 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
4234 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
4235 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
4237 return throttled_hierarchy(src_cfs_rq
) ||
4238 throttled_hierarchy(dest_cfs_rq
);
4241 /* updated child weight may affect parent so we have to do this bottom up */
4242 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
4244 struct rq
*rq
= data
;
4245 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4247 cfs_rq
->throttle_count
--;
4248 if (!cfs_rq
->throttle_count
) {
4249 /* adjust cfs_rq_clock_task() */
4250 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
4251 cfs_rq
->throttled_clock_task
;
4257 static int tg_throttle_down(struct task_group
*tg
, void *data
)
4259 struct rq
*rq
= data
;
4260 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4262 /* group is entering throttled state, stop time */
4263 if (!cfs_rq
->throttle_count
)
4264 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
4265 cfs_rq
->throttle_count
++;
4270 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
4272 struct rq
*rq
= rq_of(cfs_rq
);
4273 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4274 struct sched_entity
*se
;
4275 long task_delta
, dequeue
= 1;
4278 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
4280 /* freeze hierarchy runnable averages while throttled */
4282 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
4285 task_delta
= cfs_rq
->h_nr_running
;
4286 for_each_sched_entity(se
) {
4287 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
4288 /* throttled entity or throttle-on-deactivate */
4293 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
4294 qcfs_rq
->h_nr_running
-= task_delta
;
4296 if (qcfs_rq
->load
.weight
)
4301 sub_nr_running(rq
, task_delta
);
4303 cfs_rq
->throttled
= 1;
4304 cfs_rq
->throttled_clock
= rq_clock(rq
);
4305 raw_spin_lock(&cfs_b
->lock
);
4306 empty
= list_empty(&cfs_b
->throttled_cfs_rq
);
4309 * Add to the _head_ of the list, so that an already-started
4310 * distribute_cfs_runtime will not see us
4312 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
4315 * If we're the first throttled task, make sure the bandwidth
4319 start_cfs_bandwidth(cfs_b
);
4321 raw_spin_unlock(&cfs_b
->lock
);
4324 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
4326 struct rq
*rq
= rq_of(cfs_rq
);
4327 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4328 struct sched_entity
*se
;
4332 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
4334 cfs_rq
->throttled
= 0;
4336 update_rq_clock(rq
);
4338 raw_spin_lock(&cfs_b
->lock
);
4339 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
4340 list_del_rcu(&cfs_rq
->throttled_list
);
4341 raw_spin_unlock(&cfs_b
->lock
);
4343 /* update hierarchical throttle state */
4344 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
4346 if (!cfs_rq
->load
.weight
)
4349 task_delta
= cfs_rq
->h_nr_running
;
4350 for_each_sched_entity(se
) {
4354 cfs_rq
= cfs_rq_of(se
);
4356 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
4357 cfs_rq
->h_nr_running
+= task_delta
;
4359 if (cfs_rq_throttled(cfs_rq
))
4364 add_nr_running(rq
, task_delta
);
4366 /* determine whether we need to wake up potentially idle cpu */
4367 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
4371 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
4372 u64 remaining
, u64 expires
)
4374 struct cfs_rq
*cfs_rq
;
4376 u64 starting_runtime
= remaining
;
4379 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
4381 struct rq
*rq
= rq_of(cfs_rq
);
4385 if (!cfs_rq_throttled(cfs_rq
))
4388 runtime
= -cfs_rq
->runtime_remaining
+ 1;
4389 if (runtime
> remaining
)
4390 runtime
= remaining
;
4391 remaining
-= runtime
;
4393 cfs_rq
->runtime_remaining
+= runtime
;
4394 cfs_rq
->runtime_expires
= expires
;
4396 /* we check whether we're throttled above */
4397 if (cfs_rq
->runtime_remaining
> 0)
4398 unthrottle_cfs_rq(cfs_rq
);
4408 return starting_runtime
- remaining
;
4412 * Responsible for refilling a task_group's bandwidth and unthrottling its
4413 * cfs_rqs as appropriate. If there has been no activity within the last
4414 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4415 * used to track this state.
4417 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
4419 u64 runtime
, runtime_expires
;
4422 /* no need to continue the timer with no bandwidth constraint */
4423 if (cfs_b
->quota
== RUNTIME_INF
)
4424 goto out_deactivate
;
4426 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4427 cfs_b
->nr_periods
+= overrun
;
4430 * idle depends on !throttled (for the case of a large deficit), and if
4431 * we're going inactive then everything else can be deferred
4433 if (cfs_b
->idle
&& !throttled
)
4434 goto out_deactivate
;
4436 __refill_cfs_bandwidth_runtime(cfs_b
);
4439 /* mark as potentially idle for the upcoming period */
4444 /* account preceding periods in which throttling occurred */
4445 cfs_b
->nr_throttled
+= overrun
;
4447 runtime_expires
= cfs_b
->runtime_expires
;
4450 * This check is repeated as we are holding onto the new bandwidth while
4451 * we unthrottle. This can potentially race with an unthrottled group
4452 * trying to acquire new bandwidth from the global pool. This can result
4453 * in us over-using our runtime if it is all used during this loop, but
4454 * only by limited amounts in that extreme case.
4456 while (throttled
&& cfs_b
->runtime
> 0) {
4457 runtime
= cfs_b
->runtime
;
4458 raw_spin_unlock(&cfs_b
->lock
);
4459 /* we can't nest cfs_b->lock while distributing bandwidth */
4460 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
4462 raw_spin_lock(&cfs_b
->lock
);
4464 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4466 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
4470 * While we are ensured activity in the period following an
4471 * unthrottle, this also covers the case in which the new bandwidth is
4472 * insufficient to cover the existing bandwidth deficit. (Forcing the
4473 * timer to remain active while there are any throttled entities.)
4483 /* a cfs_rq won't donate quota below this amount */
4484 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
4485 /* minimum remaining period time to redistribute slack quota */
4486 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
4487 /* how long we wait to gather additional slack before distributing */
4488 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
4491 * Are we near the end of the current quota period?
4493 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4494 * hrtimer base being cleared by hrtimer_start. In the case of
4495 * migrate_hrtimers, base is never cleared, so we are fine.
4497 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
4499 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
4502 /* if the call-back is running a quota refresh is already occurring */
4503 if (hrtimer_callback_running(refresh_timer
))
4506 /* is a quota refresh about to occur? */
4507 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
4508 if (remaining
< min_expire
)
4514 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
4516 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
4518 /* if there's a quota refresh soon don't bother with slack */
4519 if (runtime_refresh_within(cfs_b
, min_left
))
4522 hrtimer_start(&cfs_b
->slack_timer
,
4523 ns_to_ktime(cfs_bandwidth_slack_period
),
4527 /* we know any runtime found here is valid as update_curr() precedes return */
4528 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4530 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4531 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
4533 if (slack_runtime
<= 0)
4536 raw_spin_lock(&cfs_b
->lock
);
4537 if (cfs_b
->quota
!= RUNTIME_INF
&&
4538 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
4539 cfs_b
->runtime
+= slack_runtime
;
4541 /* we are under rq->lock, defer unthrottling using a timer */
4542 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
4543 !list_empty(&cfs_b
->throttled_cfs_rq
))
4544 start_cfs_slack_bandwidth(cfs_b
);
4546 raw_spin_unlock(&cfs_b
->lock
);
4548 /* even if it's not valid for return we don't want to try again */
4549 cfs_rq
->runtime_remaining
-= slack_runtime
;
4552 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4554 if (!cfs_bandwidth_used())
4557 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
4560 __return_cfs_rq_runtime(cfs_rq
);
4564 * This is done with a timer (instead of inline with bandwidth return) since
4565 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4567 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
4569 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
4572 /* confirm we're still not at a refresh boundary */
4573 raw_spin_lock(&cfs_b
->lock
);
4574 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
4575 raw_spin_unlock(&cfs_b
->lock
);
4579 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
4580 runtime
= cfs_b
->runtime
;
4582 expires
= cfs_b
->runtime_expires
;
4583 raw_spin_unlock(&cfs_b
->lock
);
4588 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
4590 raw_spin_lock(&cfs_b
->lock
);
4591 if (expires
== cfs_b
->runtime_expires
)
4592 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
4593 raw_spin_unlock(&cfs_b
->lock
);
4597 * When a group wakes up we want to make sure that its quota is not already
4598 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4599 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4601 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
4603 if (!cfs_bandwidth_used())
4606 /* an active group must be handled by the update_curr()->put() path */
4607 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
4610 /* ensure the group is not already throttled */
4611 if (cfs_rq_throttled(cfs_rq
))
4614 /* update runtime allocation */
4615 account_cfs_rq_runtime(cfs_rq
, 0);
4616 if (cfs_rq
->runtime_remaining
<= 0)
4617 throttle_cfs_rq(cfs_rq
);
4620 static void sync_throttle(struct task_group
*tg
, int cpu
)
4622 struct cfs_rq
*pcfs_rq
, *cfs_rq
;
4624 if (!cfs_bandwidth_used())
4630 cfs_rq
= tg
->cfs_rq
[cpu
];
4631 pcfs_rq
= tg
->parent
->cfs_rq
[cpu
];
4633 cfs_rq
->throttle_count
= pcfs_rq
->throttle_count
;
4634 cfs_rq
->throttled_clock_task
= rq_clock_task(cpu_rq(cpu
));
4637 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4638 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4640 if (!cfs_bandwidth_used())
4643 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
4647 * it's possible for a throttled entity to be forced into a running
4648 * state (e.g. set_curr_task), in this case we're finished.
4650 if (cfs_rq_throttled(cfs_rq
))
4653 throttle_cfs_rq(cfs_rq
);
4657 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
4659 struct cfs_bandwidth
*cfs_b
=
4660 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
4662 do_sched_cfs_slack_timer(cfs_b
);
4664 return HRTIMER_NORESTART
;
4667 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
4669 struct cfs_bandwidth
*cfs_b
=
4670 container_of(timer
, struct cfs_bandwidth
, period_timer
);
4674 raw_spin_lock(&cfs_b
->lock
);
4676 overrun
= hrtimer_forward_now(timer
, cfs_b
->period
);
4680 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
4683 cfs_b
->period_active
= 0;
4684 raw_spin_unlock(&cfs_b
->lock
);
4686 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
4689 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4691 raw_spin_lock_init(&cfs_b
->lock
);
4693 cfs_b
->quota
= RUNTIME_INF
;
4694 cfs_b
->period
= ns_to_ktime(default_cfs_period());
4696 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
4697 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_PINNED
);
4698 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
4699 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4700 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
4703 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4705 cfs_rq
->runtime_enabled
= 0;
4706 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
4709 void start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4711 lockdep_assert_held(&cfs_b
->lock
);
4713 if (!cfs_b
->period_active
) {
4714 cfs_b
->period_active
= 1;
4715 hrtimer_forward_now(&cfs_b
->period_timer
, cfs_b
->period
);
4716 hrtimer_start_expires(&cfs_b
->period_timer
, HRTIMER_MODE_ABS_PINNED
);
4720 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4722 /* init_cfs_bandwidth() was not called */
4723 if (!cfs_b
->throttled_cfs_rq
.next
)
4726 hrtimer_cancel(&cfs_b
->period_timer
);
4727 hrtimer_cancel(&cfs_b
->slack_timer
);
4731 * Both these cpu hotplug callbacks race against unregister_fair_sched_group()
4733 * The race is harmless, since modifying bandwidth settings of unhooked group
4734 * bits doesn't do much.
4737 /* cpu online calback */
4738 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
4740 struct task_group
*tg
;
4742 lockdep_assert_held(&rq
->lock
);
4745 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
4746 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
4747 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4749 raw_spin_lock(&cfs_b
->lock
);
4750 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
4751 raw_spin_unlock(&cfs_b
->lock
);
4756 /* cpu offline callback */
4757 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
4759 struct task_group
*tg
;
4761 lockdep_assert_held(&rq
->lock
);
4764 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
4765 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
4767 if (!cfs_rq
->runtime_enabled
)
4771 * clock_task is not advancing so we just need to make sure
4772 * there's some valid quota amount
4774 cfs_rq
->runtime_remaining
= 1;
4776 * Offline rq is schedulable till cpu is completely disabled
4777 * in take_cpu_down(), so we prevent new cfs throttling here.
4779 cfs_rq
->runtime_enabled
= 0;
4781 if (cfs_rq_throttled(cfs_rq
))
4782 unthrottle_cfs_rq(cfs_rq
);
4787 #else /* CONFIG_CFS_BANDWIDTH */
4788 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
4790 return rq_clock_task(rq_of(cfs_rq
));
4793 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
4794 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
4795 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
4796 static inline void sync_throttle(struct task_group
*tg
, int cpu
) {}
4797 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4799 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4804 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4809 static inline int throttled_lb_pair(struct task_group
*tg
,
4810 int src_cpu
, int dest_cpu
)
4815 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4817 #ifdef CONFIG_FAIR_GROUP_SCHED
4818 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4821 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4825 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4826 static inline void update_runtime_enabled(struct rq
*rq
) {}
4827 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
4829 #endif /* CONFIG_CFS_BANDWIDTH */
4831 /**************************************************
4832 * CFS operations on tasks:
4835 #ifdef CONFIG_SCHED_HRTICK
4836 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4838 struct sched_entity
*se
= &p
->se
;
4839 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4841 SCHED_WARN_ON(task_rq(p
) != rq
);
4843 if (rq
->cfs
.h_nr_running
> 1) {
4844 u64 slice
= sched_slice(cfs_rq
, se
);
4845 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
4846 s64 delta
= slice
- ran
;
4853 hrtick_start(rq
, delta
);
4858 * called from enqueue/dequeue and updates the hrtick when the
4859 * current task is from our class and nr_running is low enough
4862 static void hrtick_update(struct rq
*rq
)
4864 struct task_struct
*curr
= rq
->curr
;
4866 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
4869 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
4870 hrtick_start_fair(rq
, curr
);
4872 #else /* !CONFIG_SCHED_HRTICK */
4874 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4878 static inline void hrtick_update(struct rq
*rq
)
4884 * The enqueue_task method is called before nr_running is
4885 * increased. Here we update the fair scheduling stats and
4886 * then put the task into the rbtree:
4889 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4891 struct cfs_rq
*cfs_rq
;
4892 struct sched_entity
*se
= &p
->se
;
4895 * If in_iowait is set, the code below may not trigger any cpufreq
4896 * utilization updates, so do it here explicitly with the IOWAIT flag
4900 cpufreq_update_this_cpu(rq
, SCHED_CPUFREQ_IOWAIT
);
4902 for_each_sched_entity(se
) {
4905 cfs_rq
= cfs_rq_of(se
);
4906 enqueue_entity(cfs_rq
, se
, flags
);
4909 * end evaluation on encountering a throttled cfs_rq
4911 * note: in the case of encountering a throttled cfs_rq we will
4912 * post the final h_nr_running increment below.
4914 if (cfs_rq_throttled(cfs_rq
))
4916 cfs_rq
->h_nr_running
++;
4918 flags
= ENQUEUE_WAKEUP
;
4921 for_each_sched_entity(se
) {
4922 cfs_rq
= cfs_rq_of(se
);
4923 cfs_rq
->h_nr_running
++;
4925 if (cfs_rq_throttled(cfs_rq
))
4928 update_load_avg(se
, UPDATE_TG
);
4929 update_cfs_shares(se
);
4933 add_nr_running(rq
, 1);
4938 static void set_next_buddy(struct sched_entity
*se
);
4941 * The dequeue_task method is called before nr_running is
4942 * decreased. We remove the task from the rbtree and
4943 * update the fair scheduling stats:
4945 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4947 struct cfs_rq
*cfs_rq
;
4948 struct sched_entity
*se
= &p
->se
;
4949 int task_sleep
= flags
& DEQUEUE_SLEEP
;
4951 for_each_sched_entity(se
) {
4952 cfs_rq
= cfs_rq_of(se
);
4953 dequeue_entity(cfs_rq
, se
, flags
);
4956 * end evaluation on encountering a throttled cfs_rq
4958 * note: in the case of encountering a throttled cfs_rq we will
4959 * post the final h_nr_running decrement below.
4961 if (cfs_rq_throttled(cfs_rq
))
4963 cfs_rq
->h_nr_running
--;
4965 /* Don't dequeue parent if it has other entities besides us */
4966 if (cfs_rq
->load
.weight
) {
4967 /* Avoid re-evaluating load for this entity: */
4968 se
= parent_entity(se
);
4970 * Bias pick_next to pick a task from this cfs_rq, as
4971 * p is sleeping when it is within its sched_slice.
4973 if (task_sleep
&& se
&& !throttled_hierarchy(cfs_rq
))
4977 flags
|= DEQUEUE_SLEEP
;
4980 for_each_sched_entity(se
) {
4981 cfs_rq
= cfs_rq_of(se
);
4982 cfs_rq
->h_nr_running
--;
4984 if (cfs_rq_throttled(cfs_rq
))
4987 update_load_avg(se
, UPDATE_TG
);
4988 update_cfs_shares(se
);
4992 sub_nr_running(rq
, 1);
4999 /* Working cpumask for: load_balance, load_balance_newidle. */
5000 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5001 DEFINE_PER_CPU(cpumask_var_t
, select_idle_mask
);
5003 #ifdef CONFIG_NO_HZ_COMMON
5005 * per rq 'load' arrray crap; XXX kill this.
5009 * The exact cpuload calculated at every tick would be:
5011 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5013 * If a cpu misses updates for n ticks (as it was idle) and update gets
5014 * called on the n+1-th tick when cpu may be busy, then we have:
5016 * load_n = (1 - 1/2^i)^n * load_0
5017 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5019 * decay_load_missed() below does efficient calculation of
5021 * load' = (1 - 1/2^i)^n * load
5023 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5024 * This allows us to precompute the above in said factors, thereby allowing the
5025 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5026 * fixed_power_int())
5028 * The calculation is approximated on a 128 point scale.
5030 #define DEGRADE_SHIFT 7
5032 static const u8 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
5033 static const u8 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
5034 { 0, 0, 0, 0, 0, 0, 0, 0 },
5035 { 64, 32, 8, 0, 0, 0, 0, 0 },
5036 { 96, 72, 40, 12, 1, 0, 0, 0 },
5037 { 112, 98, 75, 43, 15, 1, 0, 0 },
5038 { 120, 112, 98, 76, 45, 16, 2, 0 }
5042 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5043 * would be when CPU is idle and so we just decay the old load without
5044 * adding any new load.
5046 static unsigned long
5047 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
5051 if (!missed_updates
)
5054 if (missed_updates
>= degrade_zero_ticks
[idx
])
5058 return load
>> missed_updates
;
5060 while (missed_updates
) {
5061 if (missed_updates
% 2)
5062 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
5064 missed_updates
>>= 1;
5069 #endif /* CONFIG_NO_HZ_COMMON */
5072 * __cpu_load_update - update the rq->cpu_load[] statistics
5073 * @this_rq: The rq to update statistics for
5074 * @this_load: The current load
5075 * @pending_updates: The number of missed updates
5077 * Update rq->cpu_load[] statistics. This function is usually called every
5078 * scheduler tick (TICK_NSEC).
5080 * This function computes a decaying average:
5082 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5084 * Because of NOHZ it might not get called on every tick which gives need for
5085 * the @pending_updates argument.
5087 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5088 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5089 * = A * (A * load[i]_n-2 + B) + B
5090 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5091 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5092 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5093 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5094 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5096 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5097 * any change in load would have resulted in the tick being turned back on.
5099 * For regular NOHZ, this reduces to:
5101 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5103 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5106 static void cpu_load_update(struct rq
*this_rq
, unsigned long this_load
,
5107 unsigned long pending_updates
)
5109 unsigned long __maybe_unused tickless_load
= this_rq
->cpu_load
[0];
5112 this_rq
->nr_load_updates
++;
5114 /* Update our load: */
5115 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
5116 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
5117 unsigned long old_load
, new_load
;
5119 /* scale is effectively 1 << i now, and >> i divides by scale */
5121 old_load
= this_rq
->cpu_load
[i
];
5122 #ifdef CONFIG_NO_HZ_COMMON
5123 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
5124 if (tickless_load
) {
5125 old_load
-= decay_load_missed(tickless_load
, pending_updates
- 1, i
);
5127 * old_load can never be a negative value because a
5128 * decayed tickless_load cannot be greater than the
5129 * original tickless_load.
5131 old_load
+= tickless_load
;
5134 new_load
= this_load
;
5136 * Round up the averaging division if load is increasing. This
5137 * prevents us from getting stuck on 9 if the load is 10, for
5140 if (new_load
> old_load
)
5141 new_load
+= scale
- 1;
5143 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
5146 sched_avg_update(this_rq
);
5149 /* Used instead of source_load when we know the type == 0 */
5150 static unsigned long weighted_cpuload(struct rq
*rq
)
5152 return cfs_rq_runnable_load_avg(&rq
->cfs
);
5155 #ifdef CONFIG_NO_HZ_COMMON
5157 * There is no sane way to deal with nohz on smp when using jiffies because the
5158 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5159 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5161 * Therefore we need to avoid the delta approach from the regular tick when
5162 * possible since that would seriously skew the load calculation. This is why we
5163 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5164 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5165 * loop exit, nohz_idle_balance, nohz full exit...)
5167 * This means we might still be one tick off for nohz periods.
5170 static void cpu_load_update_nohz(struct rq
*this_rq
,
5171 unsigned long curr_jiffies
,
5174 unsigned long pending_updates
;
5176 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
5177 if (pending_updates
) {
5178 this_rq
->last_load_update_tick
= curr_jiffies
;
5180 * In the regular NOHZ case, we were idle, this means load 0.
5181 * In the NOHZ_FULL case, we were non-idle, we should consider
5182 * its weighted load.
5184 cpu_load_update(this_rq
, load
, pending_updates
);
5189 * Called from nohz_idle_balance() to update the load ratings before doing the
5192 static void cpu_load_update_idle(struct rq
*this_rq
)
5195 * bail if there's load or we're actually up-to-date.
5197 if (weighted_cpuload(this_rq
))
5200 cpu_load_update_nohz(this_rq
, READ_ONCE(jiffies
), 0);
5204 * Record CPU load on nohz entry so we know the tickless load to account
5205 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5206 * than other cpu_load[idx] but it should be fine as cpu_load readers
5207 * shouldn't rely into synchronized cpu_load[*] updates.
5209 void cpu_load_update_nohz_start(void)
5211 struct rq
*this_rq
= this_rq();
5214 * This is all lockless but should be fine. If weighted_cpuload changes
5215 * concurrently we'll exit nohz. And cpu_load write can race with
5216 * cpu_load_update_idle() but both updater would be writing the same.
5218 this_rq
->cpu_load
[0] = weighted_cpuload(this_rq
);
5222 * Account the tickless load in the end of a nohz frame.
5224 void cpu_load_update_nohz_stop(void)
5226 unsigned long curr_jiffies
= READ_ONCE(jiffies
);
5227 struct rq
*this_rq
= this_rq();
5231 if (curr_jiffies
== this_rq
->last_load_update_tick
)
5234 load
= weighted_cpuload(this_rq
);
5235 rq_lock(this_rq
, &rf
);
5236 update_rq_clock(this_rq
);
5237 cpu_load_update_nohz(this_rq
, curr_jiffies
, load
);
5238 rq_unlock(this_rq
, &rf
);
5240 #else /* !CONFIG_NO_HZ_COMMON */
5241 static inline void cpu_load_update_nohz(struct rq
*this_rq
,
5242 unsigned long curr_jiffies
,
5243 unsigned long load
) { }
5244 #endif /* CONFIG_NO_HZ_COMMON */
5246 static void cpu_load_update_periodic(struct rq
*this_rq
, unsigned long load
)
5248 #ifdef CONFIG_NO_HZ_COMMON
5249 /* See the mess around cpu_load_update_nohz(). */
5250 this_rq
->last_load_update_tick
= READ_ONCE(jiffies
);
5252 cpu_load_update(this_rq
, load
, 1);
5256 * Called from scheduler_tick()
5258 void cpu_load_update_active(struct rq
*this_rq
)
5260 unsigned long load
= weighted_cpuload(this_rq
);
5262 if (tick_nohz_tick_stopped())
5263 cpu_load_update_nohz(this_rq
, READ_ONCE(jiffies
), load
);
5265 cpu_load_update_periodic(this_rq
, load
);
5269 * Return a low guess at the load of a migration-source cpu weighted
5270 * according to the scheduling class and "nice" value.
5272 * We want to under-estimate the load of migration sources, to
5273 * balance conservatively.
5275 static unsigned long source_load(int cpu
, int type
)
5277 struct rq
*rq
= cpu_rq(cpu
);
5278 unsigned long total
= weighted_cpuload(rq
);
5280 if (type
== 0 || !sched_feat(LB_BIAS
))
5283 return min(rq
->cpu_load
[type
-1], total
);
5287 * Return a high guess at the load of a migration-target cpu weighted
5288 * according to the scheduling class and "nice" value.
5290 static unsigned long target_load(int cpu
, int type
)
5292 struct rq
*rq
= cpu_rq(cpu
);
5293 unsigned long total
= weighted_cpuload(rq
);
5295 if (type
== 0 || !sched_feat(LB_BIAS
))
5298 return max(rq
->cpu_load
[type
-1], total
);
5301 static unsigned long capacity_of(int cpu
)
5303 return cpu_rq(cpu
)->cpu_capacity
;
5306 static unsigned long capacity_orig_of(int cpu
)
5308 return cpu_rq(cpu
)->cpu_capacity_orig
;
5311 static unsigned long cpu_avg_load_per_task(int cpu
)
5313 struct rq
*rq
= cpu_rq(cpu
);
5314 unsigned long nr_running
= READ_ONCE(rq
->cfs
.h_nr_running
);
5315 unsigned long load_avg
= weighted_cpuload(rq
);
5318 return load_avg
/ nr_running
;
5323 static void record_wakee(struct task_struct
*p
)
5326 * Only decay a single time; tasks that have less then 1 wakeup per
5327 * jiffy will not have built up many flips.
5329 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
5330 current
->wakee_flips
>>= 1;
5331 current
->wakee_flip_decay_ts
= jiffies
;
5334 if (current
->last_wakee
!= p
) {
5335 current
->last_wakee
= p
;
5336 current
->wakee_flips
++;
5341 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5343 * A waker of many should wake a different task than the one last awakened
5344 * at a frequency roughly N times higher than one of its wakees.
5346 * In order to determine whether we should let the load spread vs consolidating
5347 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5348 * partner, and a factor of lls_size higher frequency in the other.
5350 * With both conditions met, we can be relatively sure that the relationship is
5351 * non-monogamous, with partner count exceeding socket size.
5353 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5354 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5357 static int wake_wide(struct task_struct
*p
)
5359 unsigned int master
= current
->wakee_flips
;
5360 unsigned int slave
= p
->wakee_flips
;
5361 int factor
= this_cpu_read(sd_llc_size
);
5364 swap(master
, slave
);
5365 if (slave
< factor
|| master
< slave
* factor
)
5371 unsigned long nr_running
;
5373 unsigned long capacity
;
5377 static bool get_llc_stats(struct llc_stats
*stats
, int cpu
)
5379 struct sched_domain_shared
*sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
5384 stats
->nr_running
= READ_ONCE(sds
->nr_running
);
5385 stats
->load
= READ_ONCE(sds
->load
);
5386 stats
->capacity
= READ_ONCE(sds
->capacity
);
5387 stats
->has_capacity
= stats
->nr_running
< per_cpu(sd_llc_size
, cpu
);
5393 * Can a task be moved from prev_cpu to this_cpu without causing a load
5394 * imbalance that would trigger the load balancer?
5396 * Since we're running on 'stale' values, we might in fact create an imbalance
5397 * but recomputing these values is expensive, as that'd mean iteration 2 cache
5398 * domains worth of CPUs.
5401 wake_affine_llc(struct sched_domain
*sd
, struct task_struct
*p
,
5402 int this_cpu
, int prev_cpu
, int sync
)
5404 struct llc_stats prev_stats
, this_stats
;
5405 s64 this_eff_load
, prev_eff_load
;
5406 unsigned long task_load
;
5408 if (!get_llc_stats(&prev_stats
, prev_cpu
) ||
5409 !get_llc_stats(&this_stats
, this_cpu
))
5413 * If sync wakeup then subtract the (maximum possible)
5414 * effect of the currently running task from the load
5415 * of the current LLC.
5418 unsigned long current_load
= task_h_load(current
);
5420 /* in this case load hits 0 and this LLC is considered 'idle' */
5421 if (current_load
> this_stats
.load
)
5424 this_stats
.load
-= current_load
;
5428 * The has_capacity stuff is not SMT aware, but by trying to balance
5429 * the nr_running on both ends we try and fill the domain at equal
5430 * rates, thereby first consuming cores before siblings.
5433 /* if the old cache has capacity, stay there */
5434 if (prev_stats
.has_capacity
&& prev_stats
.nr_running
< this_stats
.nr_running
+1)
5437 /* if this cache has capacity, come here */
5438 if (this_stats
.has_capacity
&& this_stats
.nr_running
+1 < prev_stats
.nr_running
)
5442 * Check to see if we can move the load without causing too much
5445 task_load
= task_h_load(p
);
5447 this_eff_load
= 100;
5448 this_eff_load
*= prev_stats
.capacity
;
5450 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
5451 prev_eff_load
*= this_stats
.capacity
;
5453 this_eff_load
*= this_stats
.load
+ task_load
;
5454 prev_eff_load
*= prev_stats
.load
- task_load
;
5456 return this_eff_load
<= prev_eff_load
;
5459 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
,
5460 int prev_cpu
, int sync
)
5462 int this_cpu
= smp_processor_id();
5466 * Default to no affine wakeups; wake_affine() should not effect a task
5467 * placement the load-balancer feels inclined to undo. The conservative
5468 * option is therefore to not move tasks when they wake up.
5473 * If the wakeup is across cache domains, try to evaluate if movement
5474 * makes sense, otherwise rely on select_idle_siblings() to do
5475 * placement inside the cache domain.
5477 if (!cpus_share_cache(prev_cpu
, this_cpu
))
5478 affine
= wake_affine_llc(sd
, p
, this_cpu
, prev_cpu
, sync
);
5480 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine_attempts
);
5482 schedstat_inc(sd
->ttwu_move_affine
);
5483 schedstat_inc(p
->se
.statistics
.nr_wakeups_affine
);
5489 static inline int task_util(struct task_struct
*p
);
5490 static int cpu_util_wake(int cpu
, struct task_struct
*p
);
5492 static unsigned long capacity_spare_wake(int cpu
, struct task_struct
*p
)
5494 return capacity_orig_of(cpu
) - cpu_util_wake(cpu
, p
);
5498 * find_idlest_group finds and returns the least busy CPU group within the
5501 static struct sched_group
*
5502 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
5503 int this_cpu
, int sd_flag
)
5505 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
5506 struct sched_group
*most_spare_sg
= NULL
;
5507 unsigned long min_runnable_load
= ULONG_MAX
, this_runnable_load
= 0;
5508 unsigned long min_avg_load
= ULONG_MAX
, this_avg_load
= 0;
5509 unsigned long most_spare
= 0, this_spare
= 0;
5510 int load_idx
= sd
->forkexec_idx
;
5511 int imbalance_scale
= 100 + (sd
->imbalance_pct
-100)/2;
5512 unsigned long imbalance
= scale_load_down(NICE_0_LOAD
) *
5513 (sd
->imbalance_pct
-100) / 100;
5515 if (sd_flag
& SD_BALANCE_WAKE
)
5516 load_idx
= sd
->wake_idx
;
5519 unsigned long load
, avg_load
, runnable_load
;
5520 unsigned long spare_cap
, max_spare_cap
;
5524 /* Skip over this group if it has no CPUs allowed */
5525 if (!cpumask_intersects(sched_group_span(group
),
5529 local_group
= cpumask_test_cpu(this_cpu
,
5530 sched_group_span(group
));
5533 * Tally up the load of all CPUs in the group and find
5534 * the group containing the CPU with most spare capacity.
5540 for_each_cpu(i
, sched_group_span(group
)) {
5541 /* Bias balancing toward cpus of our domain */
5543 load
= source_load(i
, load_idx
);
5545 load
= target_load(i
, load_idx
);
5547 runnable_load
+= load
;
5549 avg_load
+= cfs_rq_load_avg(&cpu_rq(i
)->cfs
);
5551 spare_cap
= capacity_spare_wake(i
, p
);
5553 if (spare_cap
> max_spare_cap
)
5554 max_spare_cap
= spare_cap
;
5557 /* Adjust by relative CPU capacity of the group */
5558 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) /
5559 group
->sgc
->capacity
;
5560 runnable_load
= (runnable_load
* SCHED_CAPACITY_SCALE
) /
5561 group
->sgc
->capacity
;
5564 this_runnable_load
= runnable_load
;
5565 this_avg_load
= avg_load
;
5566 this_spare
= max_spare_cap
;
5568 if (min_runnable_load
> (runnable_load
+ imbalance
)) {
5570 * The runnable load is significantly smaller
5571 * so we can pick this new cpu
5573 min_runnable_load
= runnable_load
;
5574 min_avg_load
= avg_load
;
5576 } else if ((runnable_load
< (min_runnable_load
+ imbalance
)) &&
5577 (100*min_avg_load
> imbalance_scale
*avg_load
)) {
5579 * The runnable loads are close so take the
5580 * blocked load into account through avg_load.
5582 min_avg_load
= avg_load
;
5586 if (most_spare
< max_spare_cap
) {
5587 most_spare
= max_spare_cap
;
5588 most_spare_sg
= group
;
5591 } while (group
= group
->next
, group
!= sd
->groups
);
5594 * The cross-over point between using spare capacity or least load
5595 * is too conservative for high utilization tasks on partially
5596 * utilized systems if we require spare_capacity > task_util(p),
5597 * so we allow for some task stuffing by using
5598 * spare_capacity > task_util(p)/2.
5600 * Spare capacity can't be used for fork because the utilization has
5601 * not been set yet, we must first select a rq to compute the initial
5604 if (sd_flag
& SD_BALANCE_FORK
)
5607 if (this_spare
> task_util(p
) / 2 &&
5608 imbalance_scale
*this_spare
> 100*most_spare
)
5611 if (most_spare
> task_util(p
) / 2)
5612 return most_spare_sg
;
5618 if (min_runnable_load
> (this_runnable_load
+ imbalance
))
5621 if ((this_runnable_load
< (min_runnable_load
+ imbalance
)) &&
5622 (100*this_avg_load
< imbalance_scale
*min_avg_load
))
5629 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5632 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
5634 unsigned long load
, min_load
= ULONG_MAX
;
5635 unsigned int min_exit_latency
= UINT_MAX
;
5636 u64 latest_idle_timestamp
= 0;
5637 int least_loaded_cpu
= this_cpu
;
5638 int shallowest_idle_cpu
= -1;
5641 /* Check if we have any choice: */
5642 if (group
->group_weight
== 1)
5643 return cpumask_first(sched_group_span(group
));
5645 /* Traverse only the allowed CPUs */
5646 for_each_cpu_and(i
, sched_group_span(group
), &p
->cpus_allowed
) {
5648 struct rq
*rq
= cpu_rq(i
);
5649 struct cpuidle_state
*idle
= idle_get_state(rq
);
5650 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
5652 * We give priority to a CPU whose idle state
5653 * has the smallest exit latency irrespective
5654 * of any idle timestamp.
5656 min_exit_latency
= idle
->exit_latency
;
5657 latest_idle_timestamp
= rq
->idle_stamp
;
5658 shallowest_idle_cpu
= i
;
5659 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
5660 rq
->idle_stamp
> latest_idle_timestamp
) {
5662 * If equal or no active idle state, then
5663 * the most recently idled CPU might have
5666 latest_idle_timestamp
= rq
->idle_stamp
;
5667 shallowest_idle_cpu
= i
;
5669 } else if (shallowest_idle_cpu
== -1) {
5670 load
= weighted_cpuload(cpu_rq(i
));
5671 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
5673 least_loaded_cpu
= i
;
5678 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
5681 #ifdef CONFIG_SCHED_SMT
5683 static inline void set_idle_cores(int cpu
, int val
)
5685 struct sched_domain_shared
*sds
;
5687 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
5689 WRITE_ONCE(sds
->has_idle_cores
, val
);
5692 static inline bool test_idle_cores(int cpu
, bool def
)
5694 struct sched_domain_shared
*sds
;
5696 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
5698 return READ_ONCE(sds
->has_idle_cores
);
5704 * Scans the local SMT mask to see if the entire core is idle, and records this
5705 * information in sd_llc_shared->has_idle_cores.
5707 * Since SMT siblings share all cache levels, inspecting this limited remote
5708 * state should be fairly cheap.
5710 void __update_idle_core(struct rq
*rq
)
5712 int core
= cpu_of(rq
);
5716 if (test_idle_cores(core
, true))
5719 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
5727 set_idle_cores(core
, 1);
5733 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5734 * there are no idle cores left in the system; tracked through
5735 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5737 static int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5739 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(select_idle_mask
);
5742 if (!static_branch_likely(&sched_smt_present
))
5745 if (!test_idle_cores(target
, false))
5748 cpumask_and(cpus
, sched_domain_span(sd
), &p
->cpus_allowed
);
5750 for_each_cpu_wrap(core
, cpus
, target
) {
5753 for_each_cpu(cpu
, cpu_smt_mask(core
)) {
5754 cpumask_clear_cpu(cpu
, cpus
);
5764 * Failed to find an idle core; stop looking for one.
5766 set_idle_cores(target
, 0);
5772 * Scan the local SMT mask for idle CPUs.
5774 static int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5778 if (!static_branch_likely(&sched_smt_present
))
5781 for_each_cpu(cpu
, cpu_smt_mask(target
)) {
5782 if (!cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
5791 #else /* CONFIG_SCHED_SMT */
5793 static inline int select_idle_core(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5798 static inline int select_idle_smt(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5803 #endif /* CONFIG_SCHED_SMT */
5806 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5807 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5808 * average idle time for this rq (as found in rq->avg_idle).
5810 static int select_idle_cpu(struct task_struct
*p
, struct sched_domain
*sd
, int target
)
5812 struct sched_domain
*this_sd
;
5813 u64 avg_cost
, avg_idle
;
5816 int cpu
, nr
= INT_MAX
;
5818 this_sd
= rcu_dereference(*this_cpu_ptr(&sd_llc
));
5823 * Due to large variance we need a large fuzz factor; hackbench in
5824 * particularly is sensitive here.
5826 avg_idle
= this_rq()->avg_idle
/ 512;
5827 avg_cost
= this_sd
->avg_scan_cost
+ 1;
5829 if (sched_feat(SIS_AVG_CPU
) && avg_idle
< avg_cost
)
5832 if (sched_feat(SIS_PROP
)) {
5833 u64 span_avg
= sd
->span_weight
* avg_idle
;
5834 if (span_avg
> 4*avg_cost
)
5835 nr
= div_u64(span_avg
, avg_cost
);
5840 time
= local_clock();
5842 for_each_cpu_wrap(cpu
, sched_domain_span(sd
), target
) {
5845 if (!cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
5851 time
= local_clock() - time
;
5852 cost
= this_sd
->avg_scan_cost
;
5853 delta
= (s64
)(time
- cost
) / 8;
5854 this_sd
->avg_scan_cost
+= delta
;
5860 * Try and locate an idle core/thread in the LLC cache domain.
5862 static int select_idle_sibling(struct task_struct
*p
, int prev
, int target
)
5864 struct sched_domain
*sd
;
5867 if (idle_cpu(target
))
5871 * If the previous cpu is cache affine and idle, don't be stupid.
5873 if (prev
!= target
&& cpus_share_cache(prev
, target
) && idle_cpu(prev
))
5876 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
5880 i
= select_idle_core(p
, sd
, target
);
5881 if ((unsigned)i
< nr_cpumask_bits
)
5884 i
= select_idle_cpu(p
, sd
, target
);
5885 if ((unsigned)i
< nr_cpumask_bits
)
5888 i
= select_idle_smt(p
, sd
, target
);
5889 if ((unsigned)i
< nr_cpumask_bits
)
5896 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5897 * tasks. The unit of the return value must be the one of capacity so we can
5898 * compare the utilization with the capacity of the CPU that is available for
5899 * CFS task (ie cpu_capacity).
5901 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5902 * recent utilization of currently non-runnable tasks on a CPU. It represents
5903 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5904 * capacity_orig is the cpu_capacity available at the highest frequency
5905 * (arch_scale_freq_capacity()).
5906 * The utilization of a CPU converges towards a sum equal to or less than the
5907 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5908 * the running time on this CPU scaled by capacity_curr.
5910 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5911 * higher than capacity_orig because of unfortunate rounding in
5912 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5913 * the average stabilizes with the new running time. We need to check that the
5914 * utilization stays within the range of [0..capacity_orig] and cap it if
5915 * necessary. Without utilization capping, a group could be seen as overloaded
5916 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5917 * available capacity. We allow utilization to overshoot capacity_curr (but not
5918 * capacity_orig) as it useful for predicting the capacity required after task
5919 * migrations (scheduler-driven DVFS).
5921 static int cpu_util(int cpu
)
5923 unsigned long util
= cpu_rq(cpu
)->cfs
.avg
.util_avg
;
5924 unsigned long capacity
= capacity_orig_of(cpu
);
5926 return (util
>= capacity
) ? capacity
: util
;
5929 static inline int task_util(struct task_struct
*p
)
5931 return p
->se
.avg
.util_avg
;
5935 * cpu_util_wake: Compute cpu utilization with any contributions from
5936 * the waking task p removed.
5938 static int cpu_util_wake(int cpu
, struct task_struct
*p
)
5940 unsigned long util
, capacity
;
5942 /* Task has no contribution or is new */
5943 if (cpu
!= task_cpu(p
) || !p
->se
.avg
.last_update_time
)
5944 return cpu_util(cpu
);
5946 capacity
= capacity_orig_of(cpu
);
5947 util
= max_t(long, cpu_rq(cpu
)->cfs
.avg
.util_avg
- task_util(p
), 0);
5949 return (util
>= capacity
) ? capacity
: util
;
5953 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5954 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5956 * In that case WAKE_AFFINE doesn't make sense and we'll let
5957 * BALANCE_WAKE sort things out.
5959 static int wake_cap(struct task_struct
*p
, int cpu
, int prev_cpu
)
5961 long min_cap
, max_cap
;
5963 min_cap
= min(capacity_orig_of(prev_cpu
), capacity_orig_of(cpu
));
5964 max_cap
= cpu_rq(cpu
)->rd
->max_cpu_capacity
;
5966 /* Minimum capacity is close to max, no need to abort wake_affine */
5967 if (max_cap
- min_cap
< max_cap
>> 3)
5970 /* Bring task utilization in sync with prev_cpu */
5971 sync_entity_load_avg(&p
->se
);
5973 return min_cap
* 1024 < task_util(p
) * capacity_margin
;
5977 * select_task_rq_fair: Select target runqueue for the waking task in domains
5978 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5979 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5981 * Balances load by selecting the idlest cpu in the idlest group, or under
5982 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5984 * Returns the target cpu number.
5986 * preempt must be disabled.
5989 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
5991 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
5992 int cpu
= smp_processor_id();
5993 int new_cpu
= prev_cpu
;
5994 int want_affine
= 0;
5995 int sync
= wake_flags
& WF_SYNC
;
5997 if (sd_flag
& SD_BALANCE_WAKE
) {
5999 want_affine
= !wake_wide(p
) && !wake_cap(p
, cpu
, prev_cpu
)
6000 && cpumask_test_cpu(cpu
, &p
->cpus_allowed
);
6004 for_each_domain(cpu
, tmp
) {
6005 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
6009 * If both cpu and prev_cpu are part of this domain,
6010 * cpu is a valid SD_WAKE_AFFINE target.
6012 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
6013 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
6018 if (tmp
->flags
& sd_flag
)
6020 else if (!want_affine
)
6025 sd
= NULL
; /* Prefer wake_affine over balance flags */
6026 if (cpu
== prev_cpu
)
6029 if (wake_affine(affine_sd
, p
, prev_cpu
, sync
))
6035 if (sd_flag
& SD_BALANCE_WAKE
) /* XXX always ? */
6036 new_cpu
= select_idle_sibling(p
, prev_cpu
, new_cpu
);
6039 struct sched_group
*group
;
6042 if (!(sd
->flags
& sd_flag
)) {
6047 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
6053 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
6054 if (new_cpu
== -1 || new_cpu
== cpu
) {
6055 /* Now try balancing at a lower domain level of cpu */
6060 /* Now try balancing at a lower domain level of new_cpu */
6062 weight
= sd
->span_weight
;
6064 for_each_domain(cpu
, tmp
) {
6065 if (weight
<= tmp
->span_weight
)
6067 if (tmp
->flags
& sd_flag
)
6070 /* while loop will break here if sd == NULL */
6078 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6079 * cfs_rq_of(p) references at time of call are still valid and identify the
6080 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6082 static void migrate_task_rq_fair(struct task_struct
*p
)
6085 * As blocked tasks retain absolute vruntime the migration needs to
6086 * deal with this by subtracting the old and adding the new
6087 * min_vruntime -- the latter is done by enqueue_entity() when placing
6088 * the task on the new runqueue.
6090 if (p
->state
== TASK_WAKING
) {
6091 struct sched_entity
*se
= &p
->se
;
6092 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6095 #ifndef CONFIG_64BIT
6096 u64 min_vruntime_copy
;
6099 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
6101 min_vruntime
= cfs_rq
->min_vruntime
;
6102 } while (min_vruntime
!= min_vruntime_copy
);
6104 min_vruntime
= cfs_rq
->min_vruntime
;
6107 se
->vruntime
-= min_vruntime
;
6111 * We are supposed to update the task to "current" time, then its up to date
6112 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6113 * what current time is, so simply throw away the out-of-date time. This
6114 * will result in the wakee task is less decayed, but giving the wakee more
6115 * load sounds not bad.
6117 remove_entity_load_avg(&p
->se
);
6119 /* Tell new CPU we are migrated */
6120 p
->se
.avg
.last_update_time
= 0;
6122 /* We have migrated, no longer consider this task hot */
6123 p
->se
.exec_start
= 0;
6126 static void task_dead_fair(struct task_struct
*p
)
6128 remove_entity_load_avg(&p
->se
);
6130 #endif /* CONFIG_SMP */
6132 static unsigned long
6133 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
6135 unsigned long gran
= sysctl_sched_wakeup_granularity
;
6138 * Since its curr running now, convert the gran from real-time
6139 * to virtual-time in his units.
6141 * By using 'se' instead of 'curr' we penalize light tasks, so
6142 * they get preempted easier. That is, if 'se' < 'curr' then
6143 * the resulting gran will be larger, therefore penalizing the
6144 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6145 * be smaller, again penalizing the lighter task.
6147 * This is especially important for buddies when the leftmost
6148 * task is higher priority than the buddy.
6150 return calc_delta_fair(gran
, se
);
6154 * Should 'se' preempt 'curr'.
6168 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
6170 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
6175 gran
= wakeup_gran(curr
, se
);
6182 static void set_last_buddy(struct sched_entity
*se
)
6184 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
6187 for_each_sched_entity(se
) {
6188 if (SCHED_WARN_ON(!se
->on_rq
))
6190 cfs_rq_of(se
)->last
= se
;
6194 static void set_next_buddy(struct sched_entity
*se
)
6196 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
6199 for_each_sched_entity(se
) {
6200 if (SCHED_WARN_ON(!se
->on_rq
))
6202 cfs_rq_of(se
)->next
= se
;
6206 static void set_skip_buddy(struct sched_entity
*se
)
6208 for_each_sched_entity(se
)
6209 cfs_rq_of(se
)->skip
= se
;
6213 * Preempt the current task with a newly woken task if needed:
6215 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
6217 struct task_struct
*curr
= rq
->curr
;
6218 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
6219 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
6220 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
6221 int next_buddy_marked
= 0;
6223 if (unlikely(se
== pse
))
6227 * This is possible from callers such as attach_tasks(), in which we
6228 * unconditionally check_prempt_curr() after an enqueue (which may have
6229 * lead to a throttle). This both saves work and prevents false
6230 * next-buddy nomination below.
6232 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
6235 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
6236 set_next_buddy(pse
);
6237 next_buddy_marked
= 1;
6241 * We can come here with TIF_NEED_RESCHED already set from new task
6244 * Note: this also catches the edge-case of curr being in a throttled
6245 * group (e.g. via set_curr_task), since update_curr() (in the
6246 * enqueue of curr) will have resulted in resched being set. This
6247 * prevents us from potentially nominating it as a false LAST_BUDDY
6250 if (test_tsk_need_resched(curr
))
6253 /* Idle tasks are by definition preempted by non-idle tasks. */
6254 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
6255 likely(p
->policy
!= SCHED_IDLE
))
6259 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6260 * is driven by the tick):
6262 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
6265 find_matching_se(&se
, &pse
);
6266 update_curr(cfs_rq_of(se
));
6268 if (wakeup_preempt_entity(se
, pse
) == 1) {
6270 * Bias pick_next to pick the sched entity that is
6271 * triggering this preemption.
6273 if (!next_buddy_marked
)
6274 set_next_buddy(pse
);
6283 * Only set the backward buddy when the current task is still
6284 * on the rq. This can happen when a wakeup gets interleaved
6285 * with schedule on the ->pre_schedule() or idle_balance()
6286 * point, either of which can * drop the rq lock.
6288 * Also, during early boot the idle thread is in the fair class,
6289 * for obvious reasons its a bad idea to schedule back to it.
6291 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
6294 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
6298 static struct task_struct
*
6299 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
6301 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
6302 struct sched_entity
*se
;
6303 struct task_struct
*p
;
6307 if (!cfs_rq
->nr_running
)
6310 #ifdef CONFIG_FAIR_GROUP_SCHED
6311 if (prev
->sched_class
!= &fair_sched_class
)
6315 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6316 * likely that a next task is from the same cgroup as the current.
6318 * Therefore attempt to avoid putting and setting the entire cgroup
6319 * hierarchy, only change the part that actually changes.
6323 struct sched_entity
*curr
= cfs_rq
->curr
;
6326 * Since we got here without doing put_prev_entity() we also
6327 * have to consider cfs_rq->curr. If it is still a runnable
6328 * entity, update_curr() will update its vruntime, otherwise
6329 * forget we've ever seen it.
6333 update_curr(cfs_rq
);
6338 * This call to check_cfs_rq_runtime() will do the
6339 * throttle and dequeue its entity in the parent(s).
6340 * Therefore the nr_running test will indeed
6343 if (unlikely(check_cfs_rq_runtime(cfs_rq
))) {
6346 if (!cfs_rq
->nr_running
)
6353 se
= pick_next_entity(cfs_rq
, curr
);
6354 cfs_rq
= group_cfs_rq(se
);
6360 * Since we haven't yet done put_prev_entity and if the selected task
6361 * is a different task than we started out with, try and touch the
6362 * least amount of cfs_rqs.
6365 struct sched_entity
*pse
= &prev
->se
;
6367 while (!(cfs_rq
= is_same_group(se
, pse
))) {
6368 int se_depth
= se
->depth
;
6369 int pse_depth
= pse
->depth
;
6371 if (se_depth
<= pse_depth
) {
6372 put_prev_entity(cfs_rq_of(pse
), pse
);
6373 pse
= parent_entity(pse
);
6375 if (se_depth
>= pse_depth
) {
6376 set_next_entity(cfs_rq_of(se
), se
);
6377 se
= parent_entity(se
);
6381 put_prev_entity(cfs_rq
, pse
);
6382 set_next_entity(cfs_rq
, se
);
6385 if (hrtick_enabled(rq
))
6386 hrtick_start_fair(rq
, p
);
6392 put_prev_task(rq
, prev
);
6395 se
= pick_next_entity(cfs_rq
, NULL
);
6396 set_next_entity(cfs_rq
, se
);
6397 cfs_rq
= group_cfs_rq(se
);
6402 if (hrtick_enabled(rq
))
6403 hrtick_start_fair(rq
, p
);
6408 new_tasks
= idle_balance(rq
, rf
);
6411 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6412 * possible for any higher priority task to appear. In that case we
6413 * must re-start the pick_next_entity() loop.
6425 * Account for a descheduled task:
6427 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
6429 struct sched_entity
*se
= &prev
->se
;
6430 struct cfs_rq
*cfs_rq
;
6432 for_each_sched_entity(se
) {
6433 cfs_rq
= cfs_rq_of(se
);
6434 put_prev_entity(cfs_rq
, se
);
6439 * sched_yield() is very simple
6441 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6443 static void yield_task_fair(struct rq
*rq
)
6445 struct task_struct
*curr
= rq
->curr
;
6446 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
6447 struct sched_entity
*se
= &curr
->se
;
6450 * Are we the only task in the tree?
6452 if (unlikely(rq
->nr_running
== 1))
6455 clear_buddies(cfs_rq
, se
);
6457 if (curr
->policy
!= SCHED_BATCH
) {
6458 update_rq_clock(rq
);
6460 * Update run-time statistics of the 'current'.
6462 update_curr(cfs_rq
);
6464 * Tell update_rq_clock() that we've just updated,
6465 * so we don't do microscopic update in schedule()
6466 * and double the fastpath cost.
6468 rq_clock_skip_update(rq
, true);
6474 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
6476 struct sched_entity
*se
= &p
->se
;
6478 /* throttled hierarchies are not runnable */
6479 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
6482 /* Tell the scheduler that we'd really like pse to run next. */
6485 yield_task_fair(rq
);
6491 /**************************************************
6492 * Fair scheduling class load-balancing methods.
6496 * The purpose of load-balancing is to achieve the same basic fairness the
6497 * per-cpu scheduler provides, namely provide a proportional amount of compute
6498 * time to each task. This is expressed in the following equation:
6500 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6502 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6503 * W_i,0 is defined as:
6505 * W_i,0 = \Sum_j w_i,j (2)
6507 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6508 * is derived from the nice value as per sched_prio_to_weight[].
6510 * The weight average is an exponential decay average of the instantaneous
6513 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6515 * C_i is the compute capacity of cpu i, typically it is the
6516 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6517 * can also include other factors [XXX].
6519 * To achieve this balance we define a measure of imbalance which follows
6520 * directly from (1):
6522 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6524 * We them move tasks around to minimize the imbalance. In the continuous
6525 * function space it is obvious this converges, in the discrete case we get
6526 * a few fun cases generally called infeasible weight scenarios.
6529 * - infeasible weights;
6530 * - local vs global optima in the discrete case. ]
6535 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6536 * for all i,j solution, we create a tree of cpus that follows the hardware
6537 * topology where each level pairs two lower groups (or better). This results
6538 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6539 * tree to only the first of the previous level and we decrease the frequency
6540 * of load-balance at each level inv. proportional to the number of cpus in
6546 * \Sum { --- * --- * 2^i } = O(n) (5)
6548 * `- size of each group
6549 * | | `- number of cpus doing load-balance
6551 * `- sum over all levels
6553 * Coupled with a limit on how many tasks we can migrate every balance pass,
6554 * this makes (5) the runtime complexity of the balancer.
6556 * An important property here is that each CPU is still (indirectly) connected
6557 * to every other cpu in at most O(log n) steps:
6559 * The adjacency matrix of the resulting graph is given by:
6562 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6565 * And you'll find that:
6567 * A^(log_2 n)_i,j != 0 for all i,j (7)
6569 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6570 * The task movement gives a factor of O(m), giving a convergence complexity
6573 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6578 * In order to avoid CPUs going idle while there's still work to do, new idle
6579 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6580 * tree itself instead of relying on other CPUs to bring it work.
6582 * This adds some complexity to both (5) and (8) but it reduces the total idle
6590 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6593 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6598 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6600 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6602 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6605 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6606 * rewrite all of this once again.]
6609 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
6611 enum fbq_type
{ regular
, remote
, all
};
6613 #define LBF_ALL_PINNED 0x01
6614 #define LBF_NEED_BREAK 0x02
6615 #define LBF_DST_PINNED 0x04
6616 #define LBF_SOME_PINNED 0x08
6619 struct sched_domain
*sd
;
6627 struct cpumask
*dst_grpmask
;
6629 enum cpu_idle_type idle
;
6631 /* The set of CPUs under consideration for load-balancing */
6632 struct cpumask
*cpus
;
6637 unsigned int loop_break
;
6638 unsigned int loop_max
;
6640 enum fbq_type fbq_type
;
6641 struct list_head tasks
;
6645 * Is this task likely cache-hot:
6647 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
6651 lockdep_assert_held(&env
->src_rq
->lock
);
6653 if (p
->sched_class
!= &fair_sched_class
)
6656 if (unlikely(p
->policy
== SCHED_IDLE
))
6660 * Buddy candidates are cache hot:
6662 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
6663 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
6664 &p
->se
== cfs_rq_of(&p
->se
)->last
))
6667 if (sysctl_sched_migration_cost
== -1)
6669 if (sysctl_sched_migration_cost
== 0)
6672 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
6674 return delta
< (s64
)sysctl_sched_migration_cost
;
6677 #ifdef CONFIG_NUMA_BALANCING
6679 * Returns 1, if task migration degrades locality
6680 * Returns 0, if task migration improves locality i.e migration preferred.
6681 * Returns -1, if task migration is not affected by locality.
6683 static int migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
6685 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
6686 unsigned long src_faults
, dst_faults
;
6687 int src_nid
, dst_nid
;
6689 if (!static_branch_likely(&sched_numa_balancing
))
6692 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
6695 src_nid
= cpu_to_node(env
->src_cpu
);
6696 dst_nid
= cpu_to_node(env
->dst_cpu
);
6698 if (src_nid
== dst_nid
)
6701 /* Migrating away from the preferred node is always bad. */
6702 if (src_nid
== p
->numa_preferred_nid
) {
6703 if (env
->src_rq
->nr_running
> env
->src_rq
->nr_preferred_running
)
6709 /* Encourage migration to the preferred node. */
6710 if (dst_nid
== p
->numa_preferred_nid
)
6713 /* Leaving a core idle is often worse than degrading locality. */
6714 if (env
->idle
!= CPU_NOT_IDLE
)
6718 src_faults
= group_faults(p
, src_nid
);
6719 dst_faults
= group_faults(p
, dst_nid
);
6721 src_faults
= task_faults(p
, src_nid
);
6722 dst_faults
= task_faults(p
, dst_nid
);
6725 return dst_faults
< src_faults
;
6729 static inline int migrate_degrades_locality(struct task_struct
*p
,
6737 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6740 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
6744 lockdep_assert_held(&env
->src_rq
->lock
);
6747 * We do not migrate tasks that are:
6748 * 1) throttled_lb_pair, or
6749 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6750 * 3) running (obviously), or
6751 * 4) are cache-hot on their current CPU.
6753 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
6756 if (!cpumask_test_cpu(env
->dst_cpu
, &p
->cpus_allowed
)) {
6759 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_affine
);
6761 env
->flags
|= LBF_SOME_PINNED
;
6764 * Remember if this task can be migrated to any other cpu in
6765 * our sched_group. We may want to revisit it if we couldn't
6766 * meet load balance goals by pulling other tasks on src_cpu.
6768 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
6769 * already computed one in current iteration.
6771 if (env
->idle
== CPU_NEWLY_IDLE
|| (env
->flags
& LBF_DST_PINNED
))
6774 /* Prevent to re-select dst_cpu via env's cpus */
6775 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
6776 if (cpumask_test_cpu(cpu
, &p
->cpus_allowed
)) {
6777 env
->flags
|= LBF_DST_PINNED
;
6778 env
->new_dst_cpu
= cpu
;
6786 /* Record that we found atleast one task that could run on dst_cpu */
6787 env
->flags
&= ~LBF_ALL_PINNED
;
6789 if (task_running(env
->src_rq
, p
)) {
6790 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_running
);
6795 * Aggressive migration if:
6796 * 1) destination numa is preferred
6797 * 2) task is cache cold, or
6798 * 3) too many balance attempts have failed.
6800 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
6801 if (tsk_cache_hot
== -1)
6802 tsk_cache_hot
= task_hot(p
, env
);
6804 if (tsk_cache_hot
<= 0 ||
6805 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
6806 if (tsk_cache_hot
== 1) {
6807 schedstat_inc(env
->sd
->lb_hot_gained
[env
->idle
]);
6808 schedstat_inc(p
->se
.statistics
.nr_forced_migrations
);
6813 schedstat_inc(p
->se
.statistics
.nr_failed_migrations_hot
);
6818 * detach_task() -- detach the task for the migration specified in env
6820 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
6822 lockdep_assert_held(&env
->src_rq
->lock
);
6824 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
6825 deactivate_task(env
->src_rq
, p
, DEQUEUE_NOCLOCK
);
6826 set_task_cpu(p
, env
->dst_cpu
);
6830 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6831 * part of active balancing operations within "domain".
6833 * Returns a task if successful and NULL otherwise.
6835 static struct task_struct
*detach_one_task(struct lb_env
*env
)
6837 struct task_struct
*p
, *n
;
6839 lockdep_assert_held(&env
->src_rq
->lock
);
6841 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
6842 if (!can_migrate_task(p
, env
))
6845 detach_task(p
, env
);
6848 * Right now, this is only the second place where
6849 * lb_gained[env->idle] is updated (other is detach_tasks)
6850 * so we can safely collect stats here rather than
6851 * inside detach_tasks().
6853 schedstat_inc(env
->sd
->lb_gained
[env
->idle
]);
6859 static const unsigned int sched_nr_migrate_break
= 32;
6862 * detach_tasks() -- tries to detach up to imbalance weighted load from
6863 * busiest_rq, as part of a balancing operation within domain "sd".
6865 * Returns number of detached tasks if successful and 0 otherwise.
6867 static int detach_tasks(struct lb_env
*env
)
6869 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
6870 struct task_struct
*p
;
6874 lockdep_assert_held(&env
->src_rq
->lock
);
6876 if (env
->imbalance
<= 0)
6879 while (!list_empty(tasks
)) {
6881 * We don't want to steal all, otherwise we may be treated likewise,
6882 * which could at worst lead to a livelock crash.
6884 if (env
->idle
!= CPU_NOT_IDLE
&& env
->src_rq
->nr_running
<= 1)
6887 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
6890 /* We've more or less seen every task there is, call it quits */
6891 if (env
->loop
> env
->loop_max
)
6894 /* take a breather every nr_migrate tasks */
6895 if (env
->loop
> env
->loop_break
) {
6896 env
->loop_break
+= sched_nr_migrate_break
;
6897 env
->flags
|= LBF_NEED_BREAK
;
6901 if (!can_migrate_task(p
, env
))
6904 load
= task_h_load(p
);
6906 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
6909 if ((load
/ 2) > env
->imbalance
)
6912 detach_task(p
, env
);
6913 list_add(&p
->se
.group_node
, &env
->tasks
);
6916 env
->imbalance
-= load
;
6918 #ifdef CONFIG_PREEMPT
6920 * NEWIDLE balancing is a source of latency, so preemptible
6921 * kernels will stop after the first task is detached to minimize
6922 * the critical section.
6924 if (env
->idle
== CPU_NEWLY_IDLE
)
6929 * We only want to steal up to the prescribed amount of
6932 if (env
->imbalance
<= 0)
6937 list_move_tail(&p
->se
.group_node
, tasks
);
6941 * Right now, this is one of only two places we collect this stat
6942 * so we can safely collect detach_one_task() stats here rather
6943 * than inside detach_one_task().
6945 schedstat_add(env
->sd
->lb_gained
[env
->idle
], detached
);
6951 * attach_task() -- attach the task detached by detach_task() to its new rq.
6953 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
6955 lockdep_assert_held(&rq
->lock
);
6957 BUG_ON(task_rq(p
) != rq
);
6958 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
6959 p
->on_rq
= TASK_ON_RQ_QUEUED
;
6960 check_preempt_curr(rq
, p
, 0);
6964 * attach_one_task() -- attaches the task returned from detach_one_task() to
6967 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
6972 update_rq_clock(rq
);
6978 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6981 static void attach_tasks(struct lb_env
*env
)
6983 struct list_head
*tasks
= &env
->tasks
;
6984 struct task_struct
*p
;
6987 rq_lock(env
->dst_rq
, &rf
);
6988 update_rq_clock(env
->dst_rq
);
6990 while (!list_empty(tasks
)) {
6991 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
6992 list_del_init(&p
->se
.group_node
);
6994 attach_task(env
->dst_rq
, p
);
6997 rq_unlock(env
->dst_rq
, &rf
);
7000 #ifdef CONFIG_FAIR_GROUP_SCHED
7002 static inline bool cfs_rq_is_decayed(struct cfs_rq
*cfs_rq
)
7004 if (cfs_rq
->load
.weight
)
7007 if (cfs_rq
->avg
.load_sum
)
7010 if (cfs_rq
->avg
.util_sum
)
7013 if (cfs_rq
->runnable_load_sum
)
7019 static void update_blocked_averages(int cpu
)
7021 struct rq
*rq
= cpu_rq(cpu
);
7022 struct cfs_rq
*cfs_rq
, *pos
;
7025 rq_lock_irqsave(rq
, &rf
);
7026 update_rq_clock(rq
);
7029 * Iterates the task_group tree in a bottom up fashion, see
7030 * list_add_leaf_cfs_rq() for details.
7032 for_each_leaf_cfs_rq_safe(rq
, cfs_rq
, pos
) {
7033 struct sched_entity
*se
;
7035 /* throttled entities do not contribute to load */
7036 if (throttled_hierarchy(cfs_rq
))
7039 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq
), cfs_rq
))
7040 update_tg_load_avg(cfs_rq
, 0);
7042 /* Propagate pending load changes to the parent, if any: */
7043 se
= cfs_rq
->tg
->se
[cpu
];
7044 if (se
&& !skip_blocked_update(se
))
7045 update_load_avg(se
, 0);
7048 * There can be a lot of idle CPU cgroups. Don't let fully
7049 * decayed cfs_rqs linger on the list.
7051 if (cfs_rq_is_decayed(cfs_rq
))
7052 list_del_leaf_cfs_rq(cfs_rq
);
7054 rq_unlock_irqrestore(rq
, &rf
);
7058 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7059 * This needs to be done in a top-down fashion because the load of a child
7060 * group is a fraction of its parents load.
7062 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
7064 struct rq
*rq
= rq_of(cfs_rq
);
7065 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
7066 unsigned long now
= jiffies
;
7069 if (cfs_rq
->last_h_load_update
== now
)
7072 cfs_rq
->h_load_next
= NULL
;
7073 for_each_sched_entity(se
) {
7074 cfs_rq
= cfs_rq_of(se
);
7075 cfs_rq
->h_load_next
= se
;
7076 if (cfs_rq
->last_h_load_update
== now
)
7081 cfs_rq
->h_load
= cfs_rq_load_avg(cfs_rq
);
7082 cfs_rq
->last_h_load_update
= now
;
7085 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
7086 load
= cfs_rq
->h_load
;
7087 load
= div64_ul(load
* se
->avg
.load_avg
,
7088 cfs_rq_load_avg(cfs_rq
) + 1);
7089 cfs_rq
= group_cfs_rq(se
);
7090 cfs_rq
->h_load
= load
;
7091 cfs_rq
->last_h_load_update
= now
;
7095 static unsigned long task_h_load(struct task_struct
*p
)
7097 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
7099 update_cfs_rq_h_load(cfs_rq
);
7100 return div64_ul(p
->se
.avg
.load_avg
* cfs_rq
->h_load
,
7101 cfs_rq_load_avg(cfs_rq
) + 1);
7104 static inline void update_blocked_averages(int cpu
)
7106 struct rq
*rq
= cpu_rq(cpu
);
7107 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
7110 rq_lock_irqsave(rq
, &rf
);
7111 update_rq_clock(rq
);
7112 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq
), cfs_rq
);
7113 rq_unlock_irqrestore(rq
, &rf
);
7116 static unsigned long task_h_load(struct task_struct
*p
)
7118 return p
->se
.avg
.load_avg
;
7122 /********** Helpers for find_busiest_group ************************/
7131 * sg_lb_stats - stats of a sched_group required for load_balancing
7133 struct sg_lb_stats
{
7134 unsigned long avg_load
; /*Avg load across the CPUs of the group */
7135 unsigned long group_load
; /* Total load over the CPUs of the group */
7136 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
7137 unsigned long load_per_task
;
7138 unsigned long group_capacity
;
7139 unsigned long group_util
; /* Total utilization of the group */
7140 unsigned int sum_nr_running
; /* Nr tasks running in the group */
7141 unsigned int idle_cpus
;
7142 unsigned int group_weight
;
7143 enum group_type group_type
;
7144 int group_no_capacity
;
7145 #ifdef CONFIG_NUMA_BALANCING
7146 unsigned int nr_numa_running
;
7147 unsigned int nr_preferred_running
;
7152 * sd_lb_stats - Structure to store the statistics of a sched_domain
7153 * during load balancing.
7155 struct sd_lb_stats
{
7156 struct sched_group
*busiest
; /* Busiest group in this sd */
7157 struct sched_group
*local
; /* Local group in this sd */
7158 unsigned long total_running
;
7159 unsigned long total_load
; /* Total load of all groups in sd */
7160 unsigned long total_capacity
; /* Total capacity of all groups in sd */
7161 unsigned long avg_load
; /* Average load across all groups in sd */
7163 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
7164 struct sg_lb_stats local_stat
; /* Statistics of the local group */
7167 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
7170 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7171 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7172 * We must however clear busiest_stat::avg_load because
7173 * update_sd_pick_busiest() reads this before assignment.
7175 *sds
= (struct sd_lb_stats
){
7178 .total_running
= 0UL,
7180 .total_capacity
= 0UL,
7183 .sum_nr_running
= 0,
7184 .group_type
= group_other
,
7190 * get_sd_load_idx - Obtain the load index for a given sched domain.
7191 * @sd: The sched_domain whose load_idx is to be obtained.
7192 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7194 * Return: The load index.
7196 static inline int get_sd_load_idx(struct sched_domain
*sd
,
7197 enum cpu_idle_type idle
)
7203 load_idx
= sd
->busy_idx
;
7206 case CPU_NEWLY_IDLE
:
7207 load_idx
= sd
->newidle_idx
;
7210 load_idx
= sd
->idle_idx
;
7217 static unsigned long scale_rt_capacity(int cpu
)
7219 struct rq
*rq
= cpu_rq(cpu
);
7220 u64 total
, used
, age_stamp
, avg
;
7224 * Since we're reading these variables without serialization make sure
7225 * we read them once before doing sanity checks on them.
7227 age_stamp
= READ_ONCE(rq
->age_stamp
);
7228 avg
= READ_ONCE(rq
->rt_avg
);
7229 delta
= __rq_clock_broken(rq
) - age_stamp
;
7231 if (unlikely(delta
< 0))
7234 total
= sched_avg_period() + delta
;
7236 used
= div_u64(avg
, total
);
7238 if (likely(used
< SCHED_CAPACITY_SCALE
))
7239 return SCHED_CAPACITY_SCALE
- used
;
7244 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
7246 unsigned long capacity
= arch_scale_cpu_capacity(sd
, cpu
);
7247 struct sched_group
*sdg
= sd
->groups
;
7249 cpu_rq(cpu
)->cpu_capacity_orig
= capacity
;
7251 capacity
*= scale_rt_capacity(cpu
);
7252 capacity
>>= SCHED_CAPACITY_SHIFT
;
7257 cpu_rq(cpu
)->cpu_capacity
= capacity
;
7258 sdg
->sgc
->capacity
= capacity
;
7259 sdg
->sgc
->min_capacity
= capacity
;
7262 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
7264 struct sched_domain
*child
= sd
->child
;
7265 struct sched_group
*group
, *sdg
= sd
->groups
;
7266 unsigned long capacity
, min_capacity
;
7267 unsigned long interval
;
7269 interval
= msecs_to_jiffies(sd
->balance_interval
);
7270 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
7271 sdg
->sgc
->next_update
= jiffies
+ interval
;
7274 update_cpu_capacity(sd
, cpu
);
7279 min_capacity
= ULONG_MAX
;
7281 if (child
->flags
& SD_OVERLAP
) {
7283 * SD_OVERLAP domains cannot assume that child groups
7284 * span the current group.
7287 for_each_cpu(cpu
, sched_group_span(sdg
)) {
7288 struct sched_group_capacity
*sgc
;
7289 struct rq
*rq
= cpu_rq(cpu
);
7292 * build_sched_domains() -> init_sched_groups_capacity()
7293 * gets here before we've attached the domains to the
7296 * Use capacity_of(), which is set irrespective of domains
7297 * in update_cpu_capacity().
7299 * This avoids capacity from being 0 and
7300 * causing divide-by-zero issues on boot.
7302 if (unlikely(!rq
->sd
)) {
7303 capacity
+= capacity_of(cpu
);
7305 sgc
= rq
->sd
->groups
->sgc
;
7306 capacity
+= sgc
->capacity
;
7309 min_capacity
= min(capacity
, min_capacity
);
7313 * !SD_OVERLAP domains can assume that child groups
7314 * span the current group.
7317 group
= child
->groups
;
7319 struct sched_group_capacity
*sgc
= group
->sgc
;
7321 capacity
+= sgc
->capacity
;
7322 min_capacity
= min(sgc
->min_capacity
, min_capacity
);
7323 group
= group
->next
;
7324 } while (group
!= child
->groups
);
7327 sdg
->sgc
->capacity
= capacity
;
7328 sdg
->sgc
->min_capacity
= min_capacity
;
7332 * Check whether the capacity of the rq has been noticeably reduced by side
7333 * activity. The imbalance_pct is used for the threshold.
7334 * Return true is the capacity is reduced
7337 check_cpu_capacity(struct rq
*rq
, struct sched_domain
*sd
)
7339 return ((rq
->cpu_capacity
* sd
->imbalance_pct
) <
7340 (rq
->cpu_capacity_orig
* 100));
7344 * Group imbalance indicates (and tries to solve) the problem where balancing
7345 * groups is inadequate due to ->cpus_allowed constraints.
7347 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7348 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7351 * { 0 1 2 3 } { 4 5 6 7 }
7354 * If we were to balance group-wise we'd place two tasks in the first group and
7355 * two tasks in the second group. Clearly this is undesired as it will overload
7356 * cpu 3 and leave one of the cpus in the second group unused.
7358 * The current solution to this issue is detecting the skew in the first group
7359 * by noticing the lower domain failed to reach balance and had difficulty
7360 * moving tasks due to affinity constraints.
7362 * When this is so detected; this group becomes a candidate for busiest; see
7363 * update_sd_pick_busiest(). And calculate_imbalance() and
7364 * find_busiest_group() avoid some of the usual balance conditions to allow it
7365 * to create an effective group imbalance.
7367 * This is a somewhat tricky proposition since the next run might not find the
7368 * group imbalance and decide the groups need to be balanced again. A most
7369 * subtle and fragile situation.
7372 static inline int sg_imbalanced(struct sched_group
*group
)
7374 return group
->sgc
->imbalance
;
7378 * group_has_capacity returns true if the group has spare capacity that could
7379 * be used by some tasks.
7380 * We consider that a group has spare capacity if the * number of task is
7381 * smaller than the number of CPUs or if the utilization is lower than the
7382 * available capacity for CFS tasks.
7383 * For the latter, we use a threshold to stabilize the state, to take into
7384 * account the variance of the tasks' load and to return true if the available
7385 * capacity in meaningful for the load balancer.
7386 * As an example, an available capacity of 1% can appear but it doesn't make
7387 * any benefit for the load balance.
7390 group_has_capacity(struct lb_env
*env
, struct sg_lb_stats
*sgs
)
7392 if (sgs
->sum_nr_running
< sgs
->group_weight
)
7395 if ((sgs
->group_capacity
* 100) >
7396 (sgs
->group_util
* env
->sd
->imbalance_pct
))
7403 * group_is_overloaded returns true if the group has more tasks than it can
7405 * group_is_overloaded is not equals to !group_has_capacity because a group
7406 * with the exact right number of tasks, has no more spare capacity but is not
7407 * overloaded so both group_has_capacity and group_is_overloaded return
7411 group_is_overloaded(struct lb_env
*env
, struct sg_lb_stats
*sgs
)
7413 if (sgs
->sum_nr_running
<= sgs
->group_weight
)
7416 if ((sgs
->group_capacity
* 100) <
7417 (sgs
->group_util
* env
->sd
->imbalance_pct
))
7424 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7425 * per-CPU capacity than sched_group ref.
7428 group_smaller_cpu_capacity(struct sched_group
*sg
, struct sched_group
*ref
)
7430 return sg
->sgc
->min_capacity
* capacity_margin
<
7431 ref
->sgc
->min_capacity
* 1024;
7435 group_type
group_classify(struct sched_group
*group
,
7436 struct sg_lb_stats
*sgs
)
7438 if (sgs
->group_no_capacity
)
7439 return group_overloaded
;
7441 if (sg_imbalanced(group
))
7442 return group_imbalanced
;
7448 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7449 * @env: The load balancing environment.
7450 * @group: sched_group whose statistics are to be updated.
7451 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7452 * @local_group: Does group contain this_cpu.
7453 * @sgs: variable to hold the statistics for this group.
7454 * @overload: Indicate more than one runnable task for any CPU.
7456 static inline void update_sg_lb_stats(struct lb_env
*env
,
7457 struct sched_group
*group
, int load_idx
,
7458 int local_group
, struct sg_lb_stats
*sgs
,
7464 memset(sgs
, 0, sizeof(*sgs
));
7466 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
7467 struct rq
*rq
= cpu_rq(i
);
7469 /* Bias balancing toward cpus of our domain */
7471 load
= target_load(i
, load_idx
);
7473 load
= source_load(i
, load_idx
);
7475 sgs
->group_load
+= load
;
7476 sgs
->group_util
+= cpu_util(i
);
7477 sgs
->sum_nr_running
+= rq
->cfs
.h_nr_running
;
7479 nr_running
= rq
->nr_running
;
7483 #ifdef CONFIG_NUMA_BALANCING
7484 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
7485 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
7487 sgs
->sum_weighted_load
+= weighted_cpuload(rq
);
7489 * No need to call idle_cpu() if nr_running is not 0
7491 if (!nr_running
&& idle_cpu(i
))
7495 /* Adjust by relative CPU capacity of the group */
7496 sgs
->group_capacity
= group
->sgc
->capacity
;
7497 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
7499 if (sgs
->sum_nr_running
)
7500 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
7502 sgs
->group_weight
= group
->group_weight
;
7504 sgs
->group_no_capacity
= group_is_overloaded(env
, sgs
);
7505 sgs
->group_type
= group_classify(group
, sgs
);
7509 * update_sd_pick_busiest - return 1 on busiest group
7510 * @env: The load balancing environment.
7511 * @sds: sched_domain statistics
7512 * @sg: sched_group candidate to be checked for being the busiest
7513 * @sgs: sched_group statistics
7515 * Determine if @sg is a busier group than the previously selected
7518 * Return: %true if @sg is a busier group than the previously selected
7519 * busiest group. %false otherwise.
7521 static bool update_sd_pick_busiest(struct lb_env
*env
,
7522 struct sd_lb_stats
*sds
,
7523 struct sched_group
*sg
,
7524 struct sg_lb_stats
*sgs
)
7526 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
7528 if (sgs
->group_type
> busiest
->group_type
)
7531 if (sgs
->group_type
< busiest
->group_type
)
7534 if (sgs
->avg_load
<= busiest
->avg_load
)
7537 if (!(env
->sd
->flags
& SD_ASYM_CPUCAPACITY
))
7541 * Candidate sg has no more than one task per CPU and
7542 * has higher per-CPU capacity. Migrating tasks to less
7543 * capable CPUs may harm throughput. Maximize throughput,
7544 * power/energy consequences are not considered.
7546 if (sgs
->sum_nr_running
<= sgs
->group_weight
&&
7547 group_smaller_cpu_capacity(sds
->local
, sg
))
7551 /* This is the busiest node in its class. */
7552 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
7555 /* No ASYM_PACKING if target cpu is already busy */
7556 if (env
->idle
== CPU_NOT_IDLE
)
7559 * ASYM_PACKING needs to move all the work to the highest
7560 * prority CPUs in the group, therefore mark all groups
7561 * of lower priority than ourself as busy.
7563 if (sgs
->sum_nr_running
&&
7564 sched_asym_prefer(env
->dst_cpu
, sg
->asym_prefer_cpu
)) {
7568 /* Prefer to move from lowest priority cpu's work */
7569 if (sched_asym_prefer(sds
->busiest
->asym_prefer_cpu
,
7570 sg
->asym_prefer_cpu
))
7577 #ifdef CONFIG_NUMA_BALANCING
7578 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
7580 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
7582 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
7587 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
7589 if (rq
->nr_running
> rq
->nr_numa_running
)
7591 if (rq
->nr_running
> rq
->nr_preferred_running
)
7596 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
7601 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
7605 #endif /* CONFIG_NUMA_BALANCING */
7608 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7609 * @env: The load balancing environment.
7610 * @sds: variable to hold the statistics for this sched_domain.
7612 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
7614 struct sched_domain_shared
*shared
= env
->sd
->shared
;
7615 struct sched_domain
*child
= env
->sd
->child
;
7616 struct sched_group
*sg
= env
->sd
->groups
;
7617 struct sg_lb_stats
*local
= &sds
->local_stat
;
7618 struct sg_lb_stats tmp_sgs
;
7619 int load_idx
, prefer_sibling
= 0;
7620 bool overload
= false;
7622 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
7625 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
7628 struct sg_lb_stats
*sgs
= &tmp_sgs
;
7631 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_span(sg
));
7636 if (env
->idle
!= CPU_NEWLY_IDLE
||
7637 time_after_eq(jiffies
, sg
->sgc
->next_update
))
7638 update_group_capacity(env
->sd
, env
->dst_cpu
);
7641 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
,
7648 * In case the child domain prefers tasks go to siblings
7649 * first, lower the sg capacity so that we'll try
7650 * and move all the excess tasks away. We lower the capacity
7651 * of a group only if the local group has the capacity to fit
7652 * these excess tasks. The extra check prevents the case where
7653 * you always pull from the heaviest group when it is already
7654 * under-utilized (possible with a large weight task outweighs
7655 * the tasks on the system).
7657 if (prefer_sibling
&& sds
->local
&&
7658 group_has_capacity(env
, local
) &&
7659 (sgs
->sum_nr_running
> local
->sum_nr_running
+ 1)) {
7660 sgs
->group_no_capacity
= 1;
7661 sgs
->group_type
= group_classify(sg
, sgs
);
7664 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
7666 sds
->busiest_stat
= *sgs
;
7670 /* Now, start updating sd_lb_stats */
7671 sds
->total_running
+= sgs
->sum_nr_running
;
7672 sds
->total_load
+= sgs
->group_load
;
7673 sds
->total_capacity
+= sgs
->group_capacity
;
7676 } while (sg
!= env
->sd
->groups
);
7678 if (env
->sd
->flags
& SD_NUMA
)
7679 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
7681 if (!env
->sd
->parent
) {
7682 /* update overload indicator if we are at root domain */
7683 if (env
->dst_rq
->rd
->overload
!= overload
)
7684 env
->dst_rq
->rd
->overload
= overload
;
7691 * Since these are sums over groups they can contain some CPUs
7692 * multiple times for the NUMA domains.
7694 * Currently only wake_affine_llc() and find_busiest_group()
7695 * uses these numbers, only the last is affected by this problem.
7699 WRITE_ONCE(shared
->nr_running
, sds
->total_running
);
7700 WRITE_ONCE(shared
->load
, sds
->total_load
);
7701 WRITE_ONCE(shared
->capacity
, sds
->total_capacity
);
7705 * check_asym_packing - Check to see if the group is packed into the
7708 * This is primarily intended to used at the sibling level. Some
7709 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7710 * case of POWER7, it can move to lower SMT modes only when higher
7711 * threads are idle. When in lower SMT modes, the threads will
7712 * perform better since they share less core resources. Hence when we
7713 * have idle threads, we want them to be the higher ones.
7715 * This packing function is run on idle threads. It checks to see if
7716 * the busiest CPU in this domain (core in the P7 case) has a higher
7717 * CPU number than the packing function is being run on. Here we are
7718 * assuming lower CPU number will be equivalent to lower a SMT thread
7721 * Return: 1 when packing is required and a task should be moved to
7722 * this CPU. The amount of the imbalance is returned in env->imbalance.
7724 * @env: The load balancing environment.
7725 * @sds: Statistics of the sched_domain which is to be packed
7727 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
7731 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
7734 if (env
->idle
== CPU_NOT_IDLE
)
7740 busiest_cpu
= sds
->busiest
->asym_prefer_cpu
;
7741 if (sched_asym_prefer(busiest_cpu
, env
->dst_cpu
))
7744 env
->imbalance
= DIV_ROUND_CLOSEST(
7745 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
7746 SCHED_CAPACITY_SCALE
);
7752 * fix_small_imbalance - Calculate the minor imbalance that exists
7753 * amongst the groups of a sched_domain, during
7755 * @env: The load balancing environment.
7756 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7759 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
7761 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
7762 unsigned int imbn
= 2;
7763 unsigned long scaled_busy_load_per_task
;
7764 struct sg_lb_stats
*local
, *busiest
;
7766 local
= &sds
->local_stat
;
7767 busiest
= &sds
->busiest_stat
;
7769 if (!local
->sum_nr_running
)
7770 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
7771 else if (busiest
->load_per_task
> local
->load_per_task
)
7774 scaled_busy_load_per_task
=
7775 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
7776 busiest
->group_capacity
;
7778 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
7779 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
7780 env
->imbalance
= busiest
->load_per_task
;
7785 * OK, we don't have enough imbalance to justify moving tasks,
7786 * however we may be able to increase total CPU capacity used by
7790 capa_now
+= busiest
->group_capacity
*
7791 min(busiest
->load_per_task
, busiest
->avg_load
);
7792 capa_now
+= local
->group_capacity
*
7793 min(local
->load_per_task
, local
->avg_load
);
7794 capa_now
/= SCHED_CAPACITY_SCALE
;
7796 /* Amount of load we'd subtract */
7797 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
7798 capa_move
+= busiest
->group_capacity
*
7799 min(busiest
->load_per_task
,
7800 busiest
->avg_load
- scaled_busy_load_per_task
);
7803 /* Amount of load we'd add */
7804 if (busiest
->avg_load
* busiest
->group_capacity
<
7805 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
7806 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
7807 local
->group_capacity
;
7809 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
7810 local
->group_capacity
;
7812 capa_move
+= local
->group_capacity
*
7813 min(local
->load_per_task
, local
->avg_load
+ tmp
);
7814 capa_move
/= SCHED_CAPACITY_SCALE
;
7816 /* Move if we gain throughput */
7817 if (capa_move
> capa_now
)
7818 env
->imbalance
= busiest
->load_per_task
;
7822 * calculate_imbalance - Calculate the amount of imbalance present within the
7823 * groups of a given sched_domain during load balance.
7824 * @env: load balance environment
7825 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7827 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
7829 unsigned long max_pull
, load_above_capacity
= ~0UL;
7830 struct sg_lb_stats
*local
, *busiest
;
7832 local
= &sds
->local_stat
;
7833 busiest
= &sds
->busiest_stat
;
7835 if (busiest
->group_type
== group_imbalanced
) {
7837 * In the group_imb case we cannot rely on group-wide averages
7838 * to ensure cpu-load equilibrium, look at wider averages. XXX
7840 busiest
->load_per_task
=
7841 min(busiest
->load_per_task
, sds
->avg_load
);
7845 * Avg load of busiest sg can be less and avg load of local sg can
7846 * be greater than avg load across all sgs of sd because avg load
7847 * factors in sg capacity and sgs with smaller group_type are
7848 * skipped when updating the busiest sg:
7850 if (busiest
->avg_load
<= sds
->avg_load
||
7851 local
->avg_load
>= sds
->avg_load
) {
7853 return fix_small_imbalance(env
, sds
);
7857 * If there aren't any idle cpus, avoid creating some.
7859 if (busiest
->group_type
== group_overloaded
&&
7860 local
->group_type
== group_overloaded
) {
7861 load_above_capacity
= busiest
->sum_nr_running
* SCHED_CAPACITY_SCALE
;
7862 if (load_above_capacity
> busiest
->group_capacity
) {
7863 load_above_capacity
-= busiest
->group_capacity
;
7864 load_above_capacity
*= scale_load_down(NICE_0_LOAD
);
7865 load_above_capacity
/= busiest
->group_capacity
;
7867 load_above_capacity
= ~0UL;
7871 * We're trying to get all the cpus to the average_load, so we don't
7872 * want to push ourselves above the average load, nor do we wish to
7873 * reduce the max loaded cpu below the average load. At the same time,
7874 * we also don't want to reduce the group load below the group
7875 * capacity. Thus we look for the minimum possible imbalance.
7877 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
7879 /* How much load to actually move to equalise the imbalance */
7880 env
->imbalance
= min(
7881 max_pull
* busiest
->group_capacity
,
7882 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
7883 ) / SCHED_CAPACITY_SCALE
;
7886 * if *imbalance is less than the average load per runnable task
7887 * there is no guarantee that any tasks will be moved so we'll have
7888 * a think about bumping its value to force at least one task to be
7891 if (env
->imbalance
< busiest
->load_per_task
)
7892 return fix_small_imbalance(env
, sds
);
7895 /******* find_busiest_group() helpers end here *********************/
7898 * find_busiest_group - Returns the busiest group within the sched_domain
7899 * if there is an imbalance.
7901 * Also calculates the amount of weighted load which should be moved
7902 * to restore balance.
7904 * @env: The load balancing environment.
7906 * Return: - The busiest group if imbalance exists.
7908 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
7910 struct sg_lb_stats
*local
, *busiest
;
7911 struct sd_lb_stats sds
;
7913 init_sd_lb_stats(&sds
);
7916 * Compute the various statistics relavent for load balancing at
7919 update_sd_lb_stats(env
, &sds
);
7920 local
= &sds
.local_stat
;
7921 busiest
= &sds
.busiest_stat
;
7923 /* ASYM feature bypasses nice load balance check */
7924 if (check_asym_packing(env
, &sds
))
7927 /* There is no busy sibling group to pull tasks from */
7928 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
7931 /* XXX broken for overlapping NUMA groups */
7932 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
7933 / sds
.total_capacity
;
7936 * If the busiest group is imbalanced the below checks don't
7937 * work because they assume all things are equal, which typically
7938 * isn't true due to cpus_allowed constraints and the like.
7940 if (busiest
->group_type
== group_imbalanced
)
7943 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7944 if (env
->idle
== CPU_NEWLY_IDLE
&& group_has_capacity(env
, local
) &&
7945 busiest
->group_no_capacity
)
7949 * If the local group is busier than the selected busiest group
7950 * don't try and pull any tasks.
7952 if (local
->avg_load
>= busiest
->avg_load
)
7956 * Don't pull any tasks if this group is already above the domain
7959 if (local
->avg_load
>= sds
.avg_load
)
7962 if (env
->idle
== CPU_IDLE
) {
7964 * This cpu is idle. If the busiest group is not overloaded
7965 * and there is no imbalance between this and busiest group
7966 * wrt idle cpus, it is balanced. The imbalance becomes
7967 * significant if the diff is greater than 1 otherwise we
7968 * might end up to just move the imbalance on another group
7970 if ((busiest
->group_type
!= group_overloaded
) &&
7971 (local
->idle_cpus
<= (busiest
->idle_cpus
+ 1)))
7975 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7976 * imbalance_pct to be conservative.
7978 if (100 * busiest
->avg_load
<=
7979 env
->sd
->imbalance_pct
* local
->avg_load
)
7984 /* Looks like there is an imbalance. Compute it */
7985 calculate_imbalance(env
, &sds
);
7994 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7996 static struct rq
*find_busiest_queue(struct lb_env
*env
,
7997 struct sched_group
*group
)
7999 struct rq
*busiest
= NULL
, *rq
;
8000 unsigned long busiest_load
= 0, busiest_capacity
= 1;
8003 for_each_cpu_and(i
, sched_group_span(group
), env
->cpus
) {
8004 unsigned long capacity
, wl
;
8008 rt
= fbq_classify_rq(rq
);
8011 * We classify groups/runqueues into three groups:
8012 * - regular: there are !numa tasks
8013 * - remote: there are numa tasks that run on the 'wrong' node
8014 * - all: there is no distinction
8016 * In order to avoid migrating ideally placed numa tasks,
8017 * ignore those when there's better options.
8019 * If we ignore the actual busiest queue to migrate another
8020 * task, the next balance pass can still reduce the busiest
8021 * queue by moving tasks around inside the node.
8023 * If we cannot move enough load due to this classification
8024 * the next pass will adjust the group classification and
8025 * allow migration of more tasks.
8027 * Both cases only affect the total convergence complexity.
8029 if (rt
> env
->fbq_type
)
8032 capacity
= capacity_of(i
);
8034 wl
= weighted_cpuload(rq
);
8037 * When comparing with imbalance, use weighted_cpuload()
8038 * which is not scaled with the cpu capacity.
8041 if (rq
->nr_running
== 1 && wl
> env
->imbalance
&&
8042 !check_cpu_capacity(rq
, env
->sd
))
8046 * For the load comparisons with the other cpu's, consider
8047 * the weighted_cpuload() scaled with the cpu capacity, so
8048 * that the load can be moved away from the cpu that is
8049 * potentially running at a lower capacity.
8051 * Thus we're looking for max(wl_i / capacity_i), crosswise
8052 * multiplication to rid ourselves of the division works out
8053 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8054 * our previous maximum.
8056 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
8058 busiest_capacity
= capacity
;
8067 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8068 * so long as it is large enough.
8070 #define MAX_PINNED_INTERVAL 512
8072 static int need_active_balance(struct lb_env
*env
)
8074 struct sched_domain
*sd
= env
->sd
;
8076 if (env
->idle
== CPU_NEWLY_IDLE
) {
8079 * ASYM_PACKING needs to force migrate tasks from busy but
8080 * lower priority CPUs in order to pack all tasks in the
8081 * highest priority CPUs.
8083 if ((sd
->flags
& SD_ASYM_PACKING
) &&
8084 sched_asym_prefer(env
->dst_cpu
, env
->src_cpu
))
8089 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8090 * It's worth migrating the task if the src_cpu's capacity is reduced
8091 * because of other sched_class or IRQs if more capacity stays
8092 * available on dst_cpu.
8094 if ((env
->idle
!= CPU_NOT_IDLE
) &&
8095 (env
->src_rq
->cfs
.h_nr_running
== 1)) {
8096 if ((check_cpu_capacity(env
->src_rq
, sd
)) &&
8097 (capacity_of(env
->src_cpu
)*sd
->imbalance_pct
< capacity_of(env
->dst_cpu
)*100))
8101 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
8104 static int active_load_balance_cpu_stop(void *data
);
8106 static int should_we_balance(struct lb_env
*env
)
8108 struct sched_group
*sg
= env
->sd
->groups
;
8109 int cpu
, balance_cpu
= -1;
8112 * In the newly idle case, we will allow all the cpu's
8113 * to do the newly idle load balance.
8115 if (env
->idle
== CPU_NEWLY_IDLE
)
8118 /* Try to find first idle cpu */
8119 for_each_cpu_and(cpu
, group_balance_mask(sg
), env
->cpus
) {
8127 if (balance_cpu
== -1)
8128 balance_cpu
= group_balance_cpu(sg
);
8131 * First idle cpu or the first cpu(busiest) in this sched group
8132 * is eligible for doing load balancing at this and above domains.
8134 return balance_cpu
== env
->dst_cpu
;
8138 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8139 * tasks if there is an imbalance.
8141 static int load_balance(int this_cpu
, struct rq
*this_rq
,
8142 struct sched_domain
*sd
, enum cpu_idle_type idle
,
8143 int *continue_balancing
)
8145 int ld_moved
, cur_ld_moved
, active_balance
= 0;
8146 struct sched_domain
*sd_parent
= sd
->parent
;
8147 struct sched_group
*group
;
8150 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
8152 struct lb_env env
= {
8154 .dst_cpu
= this_cpu
,
8156 .dst_grpmask
= sched_group_span(sd
->groups
),
8158 .loop_break
= sched_nr_migrate_break
,
8161 .tasks
= LIST_HEAD_INIT(env
.tasks
),
8164 cpumask_and(cpus
, sched_domain_span(sd
), cpu_active_mask
);
8166 schedstat_inc(sd
->lb_count
[idle
]);
8169 if (!should_we_balance(&env
)) {
8170 *continue_balancing
= 0;
8174 group
= find_busiest_group(&env
);
8176 schedstat_inc(sd
->lb_nobusyg
[idle
]);
8180 busiest
= find_busiest_queue(&env
, group
);
8182 schedstat_inc(sd
->lb_nobusyq
[idle
]);
8186 BUG_ON(busiest
== env
.dst_rq
);
8188 schedstat_add(sd
->lb_imbalance
[idle
], env
.imbalance
);
8190 env
.src_cpu
= busiest
->cpu
;
8191 env
.src_rq
= busiest
;
8194 if (busiest
->nr_running
> 1) {
8196 * Attempt to move tasks. If find_busiest_group has found
8197 * an imbalance but busiest->nr_running <= 1, the group is
8198 * still unbalanced. ld_moved simply stays zero, so it is
8199 * correctly treated as an imbalance.
8201 env
.flags
|= LBF_ALL_PINNED
;
8202 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
8205 rq_lock_irqsave(busiest
, &rf
);
8206 update_rq_clock(busiest
);
8209 * cur_ld_moved - load moved in current iteration
8210 * ld_moved - cumulative load moved across iterations
8212 cur_ld_moved
= detach_tasks(&env
);
8215 * We've detached some tasks from busiest_rq. Every
8216 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8217 * unlock busiest->lock, and we are able to be sure
8218 * that nobody can manipulate the tasks in parallel.
8219 * See task_rq_lock() family for the details.
8222 rq_unlock(busiest
, &rf
);
8226 ld_moved
+= cur_ld_moved
;
8229 local_irq_restore(rf
.flags
);
8231 if (env
.flags
& LBF_NEED_BREAK
) {
8232 env
.flags
&= ~LBF_NEED_BREAK
;
8237 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8238 * us and move them to an alternate dst_cpu in our sched_group
8239 * where they can run. The upper limit on how many times we
8240 * iterate on same src_cpu is dependent on number of cpus in our
8243 * This changes load balance semantics a bit on who can move
8244 * load to a given_cpu. In addition to the given_cpu itself
8245 * (or a ilb_cpu acting on its behalf where given_cpu is
8246 * nohz-idle), we now have balance_cpu in a position to move
8247 * load to given_cpu. In rare situations, this may cause
8248 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8249 * _independently_ and at _same_ time to move some load to
8250 * given_cpu) causing exceess load to be moved to given_cpu.
8251 * This however should not happen so much in practice and
8252 * moreover subsequent load balance cycles should correct the
8253 * excess load moved.
8255 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
8257 /* Prevent to re-select dst_cpu via env's cpus */
8258 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
8260 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
8261 env
.dst_cpu
= env
.new_dst_cpu
;
8262 env
.flags
&= ~LBF_DST_PINNED
;
8264 env
.loop_break
= sched_nr_migrate_break
;
8267 * Go back to "more_balance" rather than "redo" since we
8268 * need to continue with same src_cpu.
8274 * We failed to reach balance because of affinity.
8277 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
8279 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
8280 *group_imbalance
= 1;
8283 /* All tasks on this runqueue were pinned by CPU affinity */
8284 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
8285 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
8287 * Attempting to continue load balancing at the current
8288 * sched_domain level only makes sense if there are
8289 * active CPUs remaining as possible busiest CPUs to
8290 * pull load from which are not contained within the
8291 * destination group that is receiving any migrated
8294 if (!cpumask_subset(cpus
, env
.dst_grpmask
)) {
8296 env
.loop_break
= sched_nr_migrate_break
;
8299 goto out_all_pinned
;
8304 schedstat_inc(sd
->lb_failed
[idle
]);
8306 * Increment the failure counter only on periodic balance.
8307 * We do not want newidle balance, which can be very
8308 * frequent, pollute the failure counter causing
8309 * excessive cache_hot migrations and active balances.
8311 if (idle
!= CPU_NEWLY_IDLE
)
8312 sd
->nr_balance_failed
++;
8314 if (need_active_balance(&env
)) {
8315 unsigned long flags
;
8317 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
8319 /* don't kick the active_load_balance_cpu_stop,
8320 * if the curr task on busiest cpu can't be
8323 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
8324 raw_spin_unlock_irqrestore(&busiest
->lock
,
8326 env
.flags
|= LBF_ALL_PINNED
;
8327 goto out_one_pinned
;
8331 * ->active_balance synchronizes accesses to
8332 * ->active_balance_work. Once set, it's cleared
8333 * only after active load balance is finished.
8335 if (!busiest
->active_balance
) {
8336 busiest
->active_balance
= 1;
8337 busiest
->push_cpu
= this_cpu
;
8340 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
8342 if (active_balance
) {
8343 stop_one_cpu_nowait(cpu_of(busiest
),
8344 active_load_balance_cpu_stop
, busiest
,
8345 &busiest
->active_balance_work
);
8348 /* We've kicked active balancing, force task migration. */
8349 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
8352 sd
->nr_balance_failed
= 0;
8354 if (likely(!active_balance
)) {
8355 /* We were unbalanced, so reset the balancing interval */
8356 sd
->balance_interval
= sd
->min_interval
;
8359 * If we've begun active balancing, start to back off. This
8360 * case may not be covered by the all_pinned logic if there
8361 * is only 1 task on the busy runqueue (because we don't call
8364 if (sd
->balance_interval
< sd
->max_interval
)
8365 sd
->balance_interval
*= 2;
8372 * We reach balance although we may have faced some affinity
8373 * constraints. Clear the imbalance flag if it was set.
8376 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
8378 if (*group_imbalance
)
8379 *group_imbalance
= 0;
8384 * We reach balance because all tasks are pinned at this level so
8385 * we can't migrate them. Let the imbalance flag set so parent level
8386 * can try to migrate them.
8388 schedstat_inc(sd
->lb_balanced
[idle
]);
8390 sd
->nr_balance_failed
= 0;
8393 /* tune up the balancing interval */
8394 if (((env
.flags
& LBF_ALL_PINNED
) &&
8395 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
8396 (sd
->balance_interval
< sd
->max_interval
))
8397 sd
->balance_interval
*= 2;
8404 static inline unsigned long
8405 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
8407 unsigned long interval
= sd
->balance_interval
;
8410 interval
*= sd
->busy_factor
;
8412 /* scale ms to jiffies */
8413 interval
= msecs_to_jiffies(interval
);
8414 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
8420 update_next_balance(struct sched_domain
*sd
, unsigned long *next_balance
)
8422 unsigned long interval
, next
;
8424 /* used by idle balance, so cpu_busy = 0 */
8425 interval
= get_sd_balance_interval(sd
, 0);
8426 next
= sd
->last_balance
+ interval
;
8428 if (time_after(*next_balance
, next
))
8429 *next_balance
= next
;
8433 * idle_balance is called by schedule() if this_cpu is about to become
8434 * idle. Attempts to pull tasks from other CPUs.
8436 static int idle_balance(struct rq
*this_rq
, struct rq_flags
*rf
)
8438 unsigned long next_balance
= jiffies
+ HZ
;
8439 int this_cpu
= this_rq
->cpu
;
8440 struct sched_domain
*sd
;
8441 int pulled_task
= 0;
8445 * We must set idle_stamp _before_ calling idle_balance(), such that we
8446 * measure the duration of idle_balance() as idle time.
8448 this_rq
->idle_stamp
= rq_clock(this_rq
);
8451 * This is OK, because current is on_cpu, which avoids it being picked
8452 * for load-balance and preemption/IRQs are still disabled avoiding
8453 * further scheduler activity on it and we're being very careful to
8454 * re-start the picking loop.
8456 rq_unpin_lock(this_rq
, rf
);
8458 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
8459 !this_rq
->rd
->overload
) {
8461 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
8463 update_next_balance(sd
, &next_balance
);
8469 raw_spin_unlock(&this_rq
->lock
);
8471 update_blocked_averages(this_cpu
);
8473 for_each_domain(this_cpu
, sd
) {
8474 int continue_balancing
= 1;
8475 u64 t0
, domain_cost
;
8477 if (!(sd
->flags
& SD_LOAD_BALANCE
))
8480 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
8481 update_next_balance(sd
, &next_balance
);
8485 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
8486 t0
= sched_clock_cpu(this_cpu
);
8488 pulled_task
= load_balance(this_cpu
, this_rq
,
8490 &continue_balancing
);
8492 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
8493 if (domain_cost
> sd
->max_newidle_lb_cost
)
8494 sd
->max_newidle_lb_cost
= domain_cost
;
8496 curr_cost
+= domain_cost
;
8499 update_next_balance(sd
, &next_balance
);
8502 * Stop searching for tasks to pull if there are
8503 * now runnable tasks on this rq.
8505 if (pulled_task
|| this_rq
->nr_running
> 0)
8510 raw_spin_lock(&this_rq
->lock
);
8512 if (curr_cost
> this_rq
->max_idle_balance_cost
)
8513 this_rq
->max_idle_balance_cost
= curr_cost
;
8516 * While browsing the domains, we released the rq lock, a task could
8517 * have been enqueued in the meantime. Since we're not going idle,
8518 * pretend we pulled a task.
8520 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
8524 /* Move the next balance forward */
8525 if (time_after(this_rq
->next_balance
, next_balance
))
8526 this_rq
->next_balance
= next_balance
;
8528 /* Is there a task of a high priority class? */
8529 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
8533 this_rq
->idle_stamp
= 0;
8535 rq_repin_lock(this_rq
, rf
);
8541 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8542 * running tasks off the busiest CPU onto idle CPUs. It requires at
8543 * least 1 task to be running on each physical CPU where possible, and
8544 * avoids physical / logical imbalances.
8546 static int active_load_balance_cpu_stop(void *data
)
8548 struct rq
*busiest_rq
= data
;
8549 int busiest_cpu
= cpu_of(busiest_rq
);
8550 int target_cpu
= busiest_rq
->push_cpu
;
8551 struct rq
*target_rq
= cpu_rq(target_cpu
);
8552 struct sched_domain
*sd
;
8553 struct task_struct
*p
= NULL
;
8556 rq_lock_irq(busiest_rq
, &rf
);
8558 /* make sure the requested cpu hasn't gone down in the meantime */
8559 if (unlikely(busiest_cpu
!= smp_processor_id() ||
8560 !busiest_rq
->active_balance
))
8563 /* Is there any task to move? */
8564 if (busiest_rq
->nr_running
<= 1)
8568 * This condition is "impossible", if it occurs
8569 * we need to fix it. Originally reported by
8570 * Bjorn Helgaas on a 128-cpu setup.
8572 BUG_ON(busiest_rq
== target_rq
);
8574 /* Search for an sd spanning us and the target CPU. */
8576 for_each_domain(target_cpu
, sd
) {
8577 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
8578 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
8583 struct lb_env env
= {
8585 .dst_cpu
= target_cpu
,
8586 .dst_rq
= target_rq
,
8587 .src_cpu
= busiest_rq
->cpu
,
8588 .src_rq
= busiest_rq
,
8591 * can_migrate_task() doesn't need to compute new_dst_cpu
8592 * for active balancing. Since we have CPU_IDLE, but no
8593 * @dst_grpmask we need to make that test go away with lying
8596 .flags
= LBF_DST_PINNED
,
8599 schedstat_inc(sd
->alb_count
);
8600 update_rq_clock(busiest_rq
);
8602 p
= detach_one_task(&env
);
8604 schedstat_inc(sd
->alb_pushed
);
8605 /* Active balancing done, reset the failure counter. */
8606 sd
->nr_balance_failed
= 0;
8608 schedstat_inc(sd
->alb_failed
);
8613 busiest_rq
->active_balance
= 0;
8614 rq_unlock(busiest_rq
, &rf
);
8617 attach_one_task(target_rq
, p
);
8624 static inline int on_null_domain(struct rq
*rq
)
8626 return unlikely(!rcu_dereference_sched(rq
->sd
));
8629 #ifdef CONFIG_NO_HZ_COMMON
8631 * idle load balancing details
8632 * - When one of the busy CPUs notice that there may be an idle rebalancing
8633 * needed, they will kick the idle load balancer, which then does idle
8634 * load balancing for all the idle CPUs.
8637 cpumask_var_t idle_cpus_mask
;
8639 unsigned long next_balance
; /* in jiffy units */
8640 } nohz ____cacheline_aligned
;
8642 static inline int find_new_ilb(void)
8644 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
8646 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
8653 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8654 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8655 * CPU (if there is one).
8657 static void nohz_balancer_kick(void)
8661 nohz
.next_balance
++;
8663 ilb_cpu
= find_new_ilb();
8665 if (ilb_cpu
>= nr_cpu_ids
)
8668 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
8671 * Use smp_send_reschedule() instead of resched_cpu().
8672 * This way we generate a sched IPI on the target cpu which
8673 * is idle. And the softirq performing nohz idle load balance
8674 * will be run before returning from the IPI.
8676 smp_send_reschedule(ilb_cpu
);
8680 void nohz_balance_exit_idle(unsigned int cpu
)
8682 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
8684 * Completely isolated CPUs don't ever set, so we must test.
8686 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
8687 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
8688 atomic_dec(&nohz
.nr_cpus
);
8690 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
8694 static inline void set_cpu_sd_state_busy(void)
8696 struct sched_domain
*sd
;
8697 int cpu
= smp_processor_id();
8700 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
8702 if (!sd
|| !sd
->nohz_idle
)
8706 atomic_inc(&sd
->shared
->nr_busy_cpus
);
8711 void set_cpu_sd_state_idle(void)
8713 struct sched_domain
*sd
;
8714 int cpu
= smp_processor_id();
8717 sd
= rcu_dereference(per_cpu(sd_llc
, cpu
));
8719 if (!sd
|| sd
->nohz_idle
)
8723 atomic_dec(&sd
->shared
->nr_busy_cpus
);
8729 * This routine will record that the cpu is going idle with tick stopped.
8730 * This info will be used in performing idle load balancing in the future.
8732 void nohz_balance_enter_idle(int cpu
)
8735 * If this cpu is going down, then nothing needs to be done.
8737 if (!cpu_active(cpu
))
8740 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
8741 if (!is_housekeeping_cpu(cpu
))
8744 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
8748 * If we're a completely isolated CPU, we don't play.
8750 if (on_null_domain(cpu_rq(cpu
)))
8753 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
8754 atomic_inc(&nohz
.nr_cpus
);
8755 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
8759 static DEFINE_SPINLOCK(balancing
);
8762 * Scale the max load_balance interval with the number of CPUs in the system.
8763 * This trades load-balance latency on larger machines for less cross talk.
8765 void update_max_interval(void)
8767 max_load_balance_interval
= HZ
*num_online_cpus()/10;
8771 * It checks each scheduling domain to see if it is due to be balanced,
8772 * and initiates a balancing operation if so.
8774 * Balancing parameters are set up in init_sched_domains.
8776 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
8778 int continue_balancing
= 1;
8780 unsigned long interval
;
8781 struct sched_domain
*sd
;
8782 /* Earliest time when we have to do rebalance again */
8783 unsigned long next_balance
= jiffies
+ 60*HZ
;
8784 int update_next_balance
= 0;
8785 int need_serialize
, need_decay
= 0;
8788 update_blocked_averages(cpu
);
8791 for_each_domain(cpu
, sd
) {
8793 * Decay the newidle max times here because this is a regular
8794 * visit to all the domains. Decay ~1% per second.
8796 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
8797 sd
->max_newidle_lb_cost
=
8798 (sd
->max_newidle_lb_cost
* 253) / 256;
8799 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
8802 max_cost
+= sd
->max_newidle_lb_cost
;
8804 if (!(sd
->flags
& SD_LOAD_BALANCE
))
8808 * Stop the load balance at this level. There is another
8809 * CPU in our sched group which is doing load balancing more
8812 if (!continue_balancing
) {
8818 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
8820 need_serialize
= sd
->flags
& SD_SERIALIZE
;
8821 if (need_serialize
) {
8822 if (!spin_trylock(&balancing
))
8826 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
8827 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
8829 * The LBF_DST_PINNED logic could have changed
8830 * env->dst_cpu, so we can't know our idle
8831 * state even if we migrated tasks. Update it.
8833 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
8835 sd
->last_balance
= jiffies
;
8836 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
8839 spin_unlock(&balancing
);
8841 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
8842 next_balance
= sd
->last_balance
+ interval
;
8843 update_next_balance
= 1;
8848 * Ensure the rq-wide value also decays but keep it at a
8849 * reasonable floor to avoid funnies with rq->avg_idle.
8851 rq
->max_idle_balance_cost
=
8852 max((u64
)sysctl_sched_migration_cost
, max_cost
);
8857 * next_balance will be updated only when there is a need.
8858 * When the cpu is attached to null domain for ex, it will not be
8861 if (likely(update_next_balance
)) {
8862 rq
->next_balance
= next_balance
;
8864 #ifdef CONFIG_NO_HZ_COMMON
8866 * If this CPU has been elected to perform the nohz idle
8867 * balance. Other idle CPUs have already rebalanced with
8868 * nohz_idle_balance() and nohz.next_balance has been
8869 * updated accordingly. This CPU is now running the idle load
8870 * balance for itself and we need to update the
8871 * nohz.next_balance accordingly.
8873 if ((idle
== CPU_IDLE
) && time_after(nohz
.next_balance
, rq
->next_balance
))
8874 nohz
.next_balance
= rq
->next_balance
;
8879 #ifdef CONFIG_NO_HZ_COMMON
8881 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8882 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8884 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
8886 int this_cpu
= this_rq
->cpu
;
8889 /* Earliest time when we have to do rebalance again */
8890 unsigned long next_balance
= jiffies
+ 60*HZ
;
8891 int update_next_balance
= 0;
8893 if (idle
!= CPU_IDLE
||
8894 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
8897 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
8898 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
8902 * If this cpu gets work to do, stop the load balancing
8903 * work being done for other cpus. Next load
8904 * balancing owner will pick it up.
8909 rq
= cpu_rq(balance_cpu
);
8912 * If time for next balance is due,
8915 if (time_after_eq(jiffies
, rq
->next_balance
)) {
8918 rq_lock_irq(rq
, &rf
);
8919 update_rq_clock(rq
);
8920 cpu_load_update_idle(rq
);
8921 rq_unlock_irq(rq
, &rf
);
8923 rebalance_domains(rq
, CPU_IDLE
);
8926 if (time_after(next_balance
, rq
->next_balance
)) {
8927 next_balance
= rq
->next_balance
;
8928 update_next_balance
= 1;
8933 * next_balance will be updated only when there is a need.
8934 * When the CPU is attached to null domain for ex, it will not be
8937 if (likely(update_next_balance
))
8938 nohz
.next_balance
= next_balance
;
8940 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
8944 * Current heuristic for kicking the idle load balancer in the presence
8945 * of an idle cpu in the system.
8946 * - This rq has more than one task.
8947 * - This rq has at least one CFS task and the capacity of the CPU is
8948 * significantly reduced because of RT tasks or IRQs.
8949 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8950 * multiple busy cpu.
8951 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8952 * domain span are idle.
8954 static inline bool nohz_kick_needed(struct rq
*rq
)
8956 unsigned long now
= jiffies
;
8957 struct sched_domain_shared
*sds
;
8958 struct sched_domain
*sd
;
8959 int nr_busy
, i
, cpu
= rq
->cpu
;
8962 if (unlikely(rq
->idle_balance
))
8966 * We may be recently in ticked or tickless idle mode. At the first
8967 * busy tick after returning from idle, we will update the busy stats.
8969 set_cpu_sd_state_busy();
8970 nohz_balance_exit_idle(cpu
);
8973 * None are in tickless mode and hence no need for NOHZ idle load
8976 if (likely(!atomic_read(&nohz
.nr_cpus
)))
8979 if (time_before(now
, nohz
.next_balance
))
8982 if (rq
->nr_running
>= 2)
8986 sds
= rcu_dereference(per_cpu(sd_llc_shared
, cpu
));
8989 * XXX: write a coherent comment on why we do this.
8990 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
8992 nr_busy
= atomic_read(&sds
->nr_busy_cpus
);
9000 sd
= rcu_dereference(rq
->sd
);
9002 if ((rq
->cfs
.h_nr_running
>= 1) &&
9003 check_cpu_capacity(rq
, sd
)) {
9009 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
9011 for_each_cpu(i
, sched_domain_span(sd
)) {
9013 !cpumask_test_cpu(i
, nohz
.idle_cpus_mask
))
9016 if (sched_asym_prefer(i
, cpu
)) {
9027 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
9031 * run_rebalance_domains is triggered when needed from the scheduler tick.
9032 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9034 static __latent_entropy
void run_rebalance_domains(struct softirq_action
*h
)
9036 struct rq
*this_rq
= this_rq();
9037 enum cpu_idle_type idle
= this_rq
->idle_balance
?
9038 CPU_IDLE
: CPU_NOT_IDLE
;
9041 * If this cpu has a pending nohz_balance_kick, then do the
9042 * balancing on behalf of the other idle cpus whose ticks are
9043 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9044 * give the idle cpus a chance to load balance. Else we may
9045 * load balance only within the local sched_domain hierarchy
9046 * and abort nohz_idle_balance altogether if we pull some load.
9048 nohz_idle_balance(this_rq
, idle
);
9049 rebalance_domains(this_rq
, idle
);
9053 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9055 void trigger_load_balance(struct rq
*rq
)
9057 /* Don't need to rebalance while attached to NULL domain */
9058 if (unlikely(on_null_domain(rq
)))
9061 if (time_after_eq(jiffies
, rq
->next_balance
))
9062 raise_softirq(SCHED_SOFTIRQ
);
9063 #ifdef CONFIG_NO_HZ_COMMON
9064 if (nohz_kick_needed(rq
))
9065 nohz_balancer_kick();
9069 static void rq_online_fair(struct rq
*rq
)
9073 update_runtime_enabled(rq
);
9076 static void rq_offline_fair(struct rq
*rq
)
9080 /* Ensure any throttled groups are reachable by pick_next_task */
9081 unthrottle_offline_cfs_rqs(rq
);
9084 #endif /* CONFIG_SMP */
9087 * scheduler tick hitting a task of our scheduling class:
9089 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
9091 struct cfs_rq
*cfs_rq
;
9092 struct sched_entity
*se
= &curr
->se
;
9094 for_each_sched_entity(se
) {
9095 cfs_rq
= cfs_rq_of(se
);
9096 entity_tick(cfs_rq
, se
, queued
);
9099 if (static_branch_unlikely(&sched_numa_balancing
))
9100 task_tick_numa(rq
, curr
);
9104 * called on fork with the child task as argument from the parent's context
9105 * - child not yet on the tasklist
9106 * - preemption disabled
9108 static void task_fork_fair(struct task_struct
*p
)
9110 struct cfs_rq
*cfs_rq
;
9111 struct sched_entity
*se
= &p
->se
, *curr
;
9112 struct rq
*rq
= this_rq();
9116 update_rq_clock(rq
);
9118 cfs_rq
= task_cfs_rq(current
);
9119 curr
= cfs_rq
->curr
;
9121 update_curr(cfs_rq
);
9122 se
->vruntime
= curr
->vruntime
;
9124 place_entity(cfs_rq
, se
, 1);
9126 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
9128 * Upon rescheduling, sched_class::put_prev_task() will place
9129 * 'current' within the tree based on its new key value.
9131 swap(curr
->vruntime
, se
->vruntime
);
9135 se
->vruntime
-= cfs_rq
->min_vruntime
;
9140 * Priority of the task has changed. Check to see if we preempt
9144 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
9146 if (!task_on_rq_queued(p
))
9150 * Reschedule if we are currently running on this runqueue and
9151 * our priority decreased, or if we are not currently running on
9152 * this runqueue and our priority is higher than the current's
9154 if (rq
->curr
== p
) {
9155 if (p
->prio
> oldprio
)
9158 check_preempt_curr(rq
, p
, 0);
9161 static inline bool vruntime_normalized(struct task_struct
*p
)
9163 struct sched_entity
*se
= &p
->se
;
9166 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9167 * the dequeue_entity(.flags=0) will already have normalized the
9174 * When !on_rq, vruntime of the task has usually NOT been normalized.
9175 * But there are some cases where it has already been normalized:
9177 * - A forked child which is waiting for being woken up by
9178 * wake_up_new_task().
9179 * - A task which has been woken up by try_to_wake_up() and
9180 * waiting for actually being woken up by sched_ttwu_pending().
9182 if (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
)
9188 #ifdef CONFIG_FAIR_GROUP_SCHED
9190 * Propagate the changes of the sched_entity across the tg tree to make it
9191 * visible to the root
9193 static void propagate_entity_cfs_rq(struct sched_entity
*se
)
9195 struct cfs_rq
*cfs_rq
;
9197 /* Start to propagate at parent */
9200 for_each_sched_entity(se
) {
9201 cfs_rq
= cfs_rq_of(se
);
9203 if (cfs_rq_throttled(cfs_rq
))
9206 update_load_avg(se
, UPDATE_TG
);
9210 static void propagate_entity_cfs_rq(struct sched_entity
*se
) { }
9213 static void detach_entity_cfs_rq(struct sched_entity
*se
)
9215 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9217 /* Catch up with the cfs_rq and remove our load when we leave */
9218 update_load_avg(se
, 0);
9219 detach_entity_load_avg(cfs_rq
, se
);
9220 update_tg_load_avg(cfs_rq
, false);
9221 propagate_entity_cfs_rq(se
);
9224 static void attach_entity_cfs_rq(struct sched_entity
*se
)
9226 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9228 #ifdef CONFIG_FAIR_GROUP_SCHED
9230 * Since the real-depth could have been changed (only FAIR
9231 * class maintain depth value), reset depth properly.
9233 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
9236 /* Synchronize entity with its cfs_rq */
9237 update_load_avg(se
, sched_feat(ATTACH_AGE_LOAD
) ? 0 : SKIP_AGE_LOAD
);
9238 attach_entity_load_avg(cfs_rq
, se
);
9239 update_tg_load_avg(cfs_rq
, false);
9240 propagate_entity_cfs_rq(se
);
9243 static void detach_task_cfs_rq(struct task_struct
*p
)
9245 struct sched_entity
*se
= &p
->se
;
9246 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9248 if (!vruntime_normalized(p
)) {
9250 * Fix up our vruntime so that the current sleep doesn't
9251 * cause 'unlimited' sleep bonus.
9253 place_entity(cfs_rq
, se
, 0);
9254 se
->vruntime
-= cfs_rq
->min_vruntime
;
9257 detach_entity_cfs_rq(se
);
9260 static void attach_task_cfs_rq(struct task_struct
*p
)
9262 struct sched_entity
*se
= &p
->se
;
9263 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9265 attach_entity_cfs_rq(se
);
9267 if (!vruntime_normalized(p
))
9268 se
->vruntime
+= cfs_rq
->min_vruntime
;
9271 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
9273 detach_task_cfs_rq(p
);
9276 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
9278 attach_task_cfs_rq(p
);
9280 if (task_on_rq_queued(p
)) {
9282 * We were most likely switched from sched_rt, so
9283 * kick off the schedule if running, otherwise just see
9284 * if we can still preempt the current task.
9289 check_preempt_curr(rq
, p
, 0);
9293 /* Account for a task changing its policy or group.
9295 * This routine is mostly called to set cfs_rq->curr field when a task
9296 * migrates between groups/classes.
9298 static void set_curr_task_fair(struct rq
*rq
)
9300 struct sched_entity
*se
= &rq
->curr
->se
;
9302 for_each_sched_entity(se
) {
9303 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
9305 set_next_entity(cfs_rq
, se
);
9306 /* ensure bandwidth has been allocated on our new cfs_rq */
9307 account_cfs_rq_runtime(cfs_rq
, 0);
9311 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
9313 cfs_rq
->tasks_timeline
= RB_ROOT
;
9314 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9315 #ifndef CONFIG_64BIT
9316 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
9319 #ifdef CONFIG_FAIR_GROUP_SCHED
9320 cfs_rq
->propagate_avg
= 0;
9322 atomic_long_set(&cfs_rq
->removed_load_avg
, 0);
9323 atomic_long_set(&cfs_rq
->removed_util_avg
, 0);
9327 #ifdef CONFIG_FAIR_GROUP_SCHED
9328 static void task_set_group_fair(struct task_struct
*p
)
9330 struct sched_entity
*se
= &p
->se
;
9332 set_task_rq(p
, task_cpu(p
));
9333 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
9336 static void task_move_group_fair(struct task_struct
*p
)
9338 detach_task_cfs_rq(p
);
9339 set_task_rq(p
, task_cpu(p
));
9342 /* Tell se's cfs_rq has been changed -- migrated */
9343 p
->se
.avg
.last_update_time
= 0;
9345 attach_task_cfs_rq(p
);
9348 static void task_change_group_fair(struct task_struct
*p
, int type
)
9351 case TASK_SET_GROUP
:
9352 task_set_group_fair(p
);
9355 case TASK_MOVE_GROUP
:
9356 task_move_group_fair(p
);
9361 void free_fair_sched_group(struct task_group
*tg
)
9365 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
9367 for_each_possible_cpu(i
) {
9369 kfree(tg
->cfs_rq
[i
]);
9378 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9380 struct sched_entity
*se
;
9381 struct cfs_rq
*cfs_rq
;
9384 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9387 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9391 tg
->shares
= NICE_0_LOAD
;
9393 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
9395 for_each_possible_cpu(i
) {
9396 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9397 GFP_KERNEL
, cpu_to_node(i
));
9401 se
= kzalloc_node(sizeof(struct sched_entity
),
9402 GFP_KERNEL
, cpu_to_node(i
));
9406 init_cfs_rq(cfs_rq
);
9407 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
9408 init_entity_runnable_average(se
);
9419 void online_fair_sched_group(struct task_group
*tg
)
9421 struct sched_entity
*se
;
9425 for_each_possible_cpu(i
) {
9429 raw_spin_lock_irq(&rq
->lock
);
9430 update_rq_clock(rq
);
9431 attach_entity_cfs_rq(se
);
9432 sync_throttle(tg
, i
);
9433 raw_spin_unlock_irq(&rq
->lock
);
9437 void unregister_fair_sched_group(struct task_group
*tg
)
9439 unsigned long flags
;
9443 for_each_possible_cpu(cpu
) {
9445 remove_entity_load_avg(tg
->se
[cpu
]);
9448 * Only empty task groups can be destroyed; so we can speculatively
9449 * check on_list without danger of it being re-added.
9451 if (!tg
->cfs_rq
[cpu
]->on_list
)
9456 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9457 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
9458 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9462 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9463 struct sched_entity
*se
, int cpu
,
9464 struct sched_entity
*parent
)
9466 struct rq
*rq
= cpu_rq(cpu
);
9470 init_cfs_rq_runtime(cfs_rq
);
9472 tg
->cfs_rq
[cpu
] = cfs_rq
;
9475 /* se could be NULL for root_task_group */
9480 se
->cfs_rq
= &rq
->cfs
;
9483 se
->cfs_rq
= parent
->my_q
;
9484 se
->depth
= parent
->depth
+ 1;
9488 /* guarantee group entities always have weight */
9489 update_load_set(&se
->load
, NICE_0_LOAD
);
9490 se
->parent
= parent
;
9493 static DEFINE_MUTEX(shares_mutex
);
9495 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9500 * We can't change the weight of the root cgroup.
9505 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
9507 mutex_lock(&shares_mutex
);
9508 if (tg
->shares
== shares
)
9511 tg
->shares
= shares
;
9512 for_each_possible_cpu(i
) {
9513 struct rq
*rq
= cpu_rq(i
);
9514 struct sched_entity
*se
= tg
->se
[i
];
9517 /* Propagate contribution to hierarchy */
9518 rq_lock_irqsave(rq
, &rf
);
9519 update_rq_clock(rq
);
9520 for_each_sched_entity(se
) {
9521 update_load_avg(se
, UPDATE_TG
);
9522 update_cfs_shares(se
);
9524 rq_unlock_irqrestore(rq
, &rf
);
9528 mutex_unlock(&shares_mutex
);
9531 #else /* CONFIG_FAIR_GROUP_SCHED */
9533 void free_fair_sched_group(struct task_group
*tg
) { }
9535 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9540 void online_fair_sched_group(struct task_group
*tg
) { }
9542 void unregister_fair_sched_group(struct task_group
*tg
) { }
9544 #endif /* CONFIG_FAIR_GROUP_SCHED */
9547 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
9549 struct sched_entity
*se
= &task
->se
;
9550 unsigned int rr_interval
= 0;
9553 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9556 if (rq
->cfs
.load
.weight
)
9557 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
9563 * All the scheduling class methods:
9565 const struct sched_class fair_sched_class
= {
9566 .next
= &idle_sched_class
,
9567 .enqueue_task
= enqueue_task_fair
,
9568 .dequeue_task
= dequeue_task_fair
,
9569 .yield_task
= yield_task_fair
,
9570 .yield_to_task
= yield_to_task_fair
,
9572 .check_preempt_curr
= check_preempt_wakeup
,
9574 .pick_next_task
= pick_next_task_fair
,
9575 .put_prev_task
= put_prev_task_fair
,
9578 .select_task_rq
= select_task_rq_fair
,
9579 .migrate_task_rq
= migrate_task_rq_fair
,
9581 .rq_online
= rq_online_fair
,
9582 .rq_offline
= rq_offline_fair
,
9584 .task_dead
= task_dead_fair
,
9585 .set_cpus_allowed
= set_cpus_allowed_common
,
9588 .set_curr_task
= set_curr_task_fair
,
9589 .task_tick
= task_tick_fair
,
9590 .task_fork
= task_fork_fair
,
9592 .prio_changed
= prio_changed_fair
,
9593 .switched_from
= switched_from_fair
,
9594 .switched_to
= switched_to_fair
,
9596 .get_rr_interval
= get_rr_interval_fair
,
9598 .update_curr
= update_curr_fair
,
9600 #ifdef CONFIG_FAIR_GROUP_SCHED
9601 .task_change_group
= task_change_group_fair
,
9605 #ifdef CONFIG_SCHED_DEBUG
9606 void print_cfs_stats(struct seq_file
*m
, int cpu
)
9608 struct cfs_rq
*cfs_rq
, *pos
;
9611 for_each_leaf_cfs_rq_safe(cpu_rq(cpu
), cfs_rq
, pos
)
9612 print_cfs_rq(m
, cpu
, cfs_rq
);
9616 #ifdef CONFIG_NUMA_BALANCING
9617 void show_numa_stats(struct task_struct
*p
, struct seq_file
*m
)
9620 unsigned long tsf
= 0, tpf
= 0, gsf
= 0, gpf
= 0;
9622 for_each_online_node(node
) {
9623 if (p
->numa_faults
) {
9624 tsf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
9625 tpf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
9627 if (p
->numa_group
) {
9628 gsf
= p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)],
9629 gpf
= p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
9631 print_numa_stats(m
, node
, tsf
, tpf
, gsf
, gpf
);
9634 #endif /* CONFIG_NUMA_BALANCING */
9635 #endif /* CONFIG_SCHED_DEBUG */
9637 __init
void init_sched_fair_class(void)
9640 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9642 #ifdef CONFIG_NO_HZ_COMMON
9643 nohz
.next_balance
= jiffies
;
9644 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);