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.h>
24 #include <linux/latencytop.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency
= 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG
;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity
= 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency
= 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly
;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
94 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window
= 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
117 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
123 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
129 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static unsigned int get_update_sysctl_factor(void)
146 unsigned int cpus
= min_t(unsigned int, num_online_cpus(), 8);
149 switch (sysctl_sched_tunable_scaling
) {
150 case SCHED_TUNABLESCALING_NONE
:
153 case SCHED_TUNABLESCALING_LINEAR
:
156 case SCHED_TUNABLESCALING_LOG
:
158 factor
= 1 + ilog2(cpus
);
165 static void update_sysctl(void)
167 unsigned int factor
= get_update_sysctl_factor();
169 #define SET_SYSCTL(name) \
170 (sysctl_##name = (factor) * normalized_sysctl_##name)
171 SET_SYSCTL(sched_min_granularity
);
172 SET_SYSCTL(sched_latency
);
173 SET_SYSCTL(sched_wakeup_granularity
);
177 void sched_init_granularity(void)
182 #define WMULT_CONST (~0U)
183 #define WMULT_SHIFT 32
185 static void __update_inv_weight(struct load_weight
*lw
)
189 if (likely(lw
->inv_weight
))
192 w
= scale_load_down(lw
->weight
);
194 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
196 else if (unlikely(!w
))
197 lw
->inv_weight
= WMULT_CONST
;
199 lw
->inv_weight
= WMULT_CONST
/ w
;
203 * delta_exec * weight / lw.weight
205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
207 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
208 * we're guaranteed shift stays positive because inv_weight is guaranteed to
209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212 * weight/lw.weight <= 1, and therefore our shift will also be positive.
214 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
216 u64 fact
= scale_load_down(weight
);
217 int shift
= WMULT_SHIFT
;
219 __update_inv_weight(lw
);
221 if (unlikely(fact
>> 32)) {
228 /* hint to use a 32x32->64 mul */
229 fact
= (u64
)(u32
)fact
* lw
->inv_weight
;
236 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
240 const struct sched_class fair_sched_class
;
242 /**************************************************************
243 * CFS operations on generic schedulable entities:
246 #ifdef CONFIG_FAIR_GROUP_SCHED
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se) (!se->my_q)
257 static inline struct task_struct
*task_of(struct sched_entity
*se
)
259 #ifdef CONFIG_SCHED_DEBUG
260 WARN_ON_ONCE(!entity_is_task(se
));
262 return container_of(se
, struct task_struct
, se
);
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 for (; se; se = se->parent)
269 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
288 if (!cfs_rq
->on_list
) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq
->tg
->parent
&&
296 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
297 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
298 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
300 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
301 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
308 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
310 if (cfs_rq
->on_list
) {
311 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
316 /* Iterate thr' all leaf cfs_rq's on a runqueue */
317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
318 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
320 /* Do the two (enqueued) entities belong to the same group ? */
321 static inline struct cfs_rq
*
322 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
324 if (se
->cfs_rq
== pse
->cfs_rq
)
330 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
336 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
338 int se_depth
, pse_depth
;
341 * preemption test can be made between sibling entities who are in the
342 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
343 * both tasks until we find their ancestors who are siblings of common
347 /* First walk up until both entities are at same depth */
348 se_depth
= (*se
)->depth
;
349 pse_depth
= (*pse
)->depth
;
351 while (se_depth
> pse_depth
) {
353 *se
= parent_entity(*se
);
356 while (pse_depth
> se_depth
) {
358 *pse
= parent_entity(*pse
);
361 while (!is_same_group(*se
, *pse
)) {
362 *se
= parent_entity(*se
);
363 *pse
= parent_entity(*pse
);
367 #else /* !CONFIG_FAIR_GROUP_SCHED */
369 static inline struct task_struct
*task_of(struct sched_entity
*se
)
371 return container_of(se
, struct task_struct
, se
);
374 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
376 return container_of(cfs_rq
, struct rq
, cfs
);
379 #define entity_is_task(se) 1
381 #define for_each_sched_entity(se) \
382 for (; se; se = NULL)
384 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
386 return &task_rq(p
)->cfs
;
389 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
391 struct task_struct
*p
= task_of(se
);
392 struct rq
*rq
= task_rq(p
);
397 /* runqueue "owned" by this group */
398 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
403 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
407 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
412 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
414 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
420 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
424 #endif /* CONFIG_FAIR_GROUP_SCHED */
426 static __always_inline
427 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
429 /**************************************************************
430 * Scheduling class tree data structure manipulation methods:
433 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
435 s64 delta
= (s64
)(vruntime
- max_vruntime
);
437 max_vruntime
= vruntime
;
442 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
444 s64 delta
= (s64
)(vruntime
- min_vruntime
);
446 min_vruntime
= vruntime
;
451 static inline int entity_before(struct sched_entity
*a
,
452 struct sched_entity
*b
)
454 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
457 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
459 u64 vruntime
= cfs_rq
->min_vruntime
;
462 vruntime
= cfs_rq
->curr
->vruntime
;
464 if (cfs_rq
->rb_leftmost
) {
465 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
470 vruntime
= se
->vruntime
;
472 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
475 /* ensure we never gain time by being placed backwards. */
476 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
479 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
484 * Enqueue an entity into the rb-tree:
486 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
488 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
489 struct rb_node
*parent
= NULL
;
490 struct sched_entity
*entry
;
494 * Find the right place in the rbtree:
498 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
500 * We dont care about collisions. Nodes with
501 * the same key stay together.
503 if (entity_before(se
, entry
)) {
504 link
= &parent
->rb_left
;
506 link
= &parent
->rb_right
;
512 * Maintain a cache of leftmost tree entries (it is frequently
516 cfs_rq
->rb_leftmost
= &se
->run_node
;
518 rb_link_node(&se
->run_node
, parent
, link
);
519 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
522 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
524 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
525 struct rb_node
*next_node
;
527 next_node
= rb_next(&se
->run_node
);
528 cfs_rq
->rb_leftmost
= next_node
;
531 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
534 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
536 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
541 return rb_entry(left
, struct sched_entity
, run_node
);
544 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
546 struct rb_node
*next
= rb_next(&se
->run_node
);
551 return rb_entry(next
, struct sched_entity
, run_node
);
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
557 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
562 return rb_entry(last
, struct sched_entity
, run_node
);
565 /**************************************************************
566 * Scheduling class statistics methods:
569 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
570 void __user
*buffer
, size_t *lenp
,
573 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
574 unsigned int factor
= get_update_sysctl_factor();
579 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
580 sysctl_sched_min_granularity
);
582 #define WRT_SYSCTL(name) \
583 (normalized_sysctl_##name = sysctl_##name / (factor))
584 WRT_SYSCTL(sched_min_granularity
);
585 WRT_SYSCTL(sched_latency
);
586 WRT_SYSCTL(sched_wakeup_granularity
);
596 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
598 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
599 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
605 * The idea is to set a period in which each task runs once.
607 * When there are too many tasks (sched_nr_latency) we have to stretch
608 * this period because otherwise the slices get too small.
610 * p = (nr <= nl) ? l : l*nr/nl
612 static u64
__sched_period(unsigned long nr_running
)
614 if (unlikely(nr_running
> sched_nr_latency
))
615 return nr_running
* sysctl_sched_min_granularity
;
617 return sysctl_sched_latency
;
621 * We calculate the wall-time slice from the period by taking a part
622 * proportional to the weight.
626 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
628 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
630 for_each_sched_entity(se
) {
631 struct load_weight
*load
;
632 struct load_weight lw
;
634 cfs_rq
= cfs_rq_of(se
);
635 load
= &cfs_rq
->load
;
637 if (unlikely(!se
->on_rq
)) {
640 update_load_add(&lw
, se
->load
.weight
);
643 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
649 * We calculate the vruntime slice of a to-be-inserted task.
653 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
655 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
659 static int select_idle_sibling(struct task_struct
*p
, int cpu
);
660 static unsigned long task_h_load(struct task_struct
*p
);
663 * We choose a half-life close to 1 scheduling period.
664 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
665 * dependent on this value.
667 #define LOAD_AVG_PERIOD 32
668 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
669 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
671 /* Give new sched_entity start runnable values to heavy its load in infant time */
672 void init_entity_runnable_average(struct sched_entity
*se
)
674 struct sched_avg
*sa
= &se
->avg
;
676 sa
->last_update_time
= 0;
678 * sched_avg's period_contrib should be strictly less then 1024, so
679 * we give it 1023 to make sure it is almost a period (1024us), and
680 * will definitely be update (after enqueue).
682 sa
->period_contrib
= 1023;
683 sa
->load_avg
= scale_load_down(se
->load
.weight
);
684 sa
->load_sum
= sa
->load_avg
* LOAD_AVG_MAX
;
686 * At this point, util_avg won't be used in select_task_rq_fair anyway
690 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
694 * With new tasks being created, their initial util_avgs are extrapolated
695 * based on the cfs_rq's current util_avg:
697 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
699 * However, in many cases, the above util_avg does not give a desired
700 * value. Moreover, the sum of the util_avgs may be divergent, such
701 * as when the series is a harmonic series.
703 * To solve this problem, we also cap the util_avg of successive tasks to
704 * only 1/2 of the left utilization budget:
706 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
708 * where n denotes the nth task.
710 * For example, a simplest series from the beginning would be like:
712 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
713 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
715 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
716 * if util_avg > util_avg_cap.
718 void post_init_entity_util_avg(struct sched_entity
*se
)
720 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
721 struct sched_avg
*sa
= &se
->avg
;
722 long cap
= (long)(SCHED_CAPACITY_SCALE
- cfs_rq
->avg
.util_avg
) / 2;
725 if (cfs_rq
->avg
.util_avg
!= 0) {
726 sa
->util_avg
= cfs_rq
->avg
.util_avg
* se
->load
.weight
;
727 sa
->util_avg
/= (cfs_rq
->avg
.load_avg
+ 1);
729 if (sa
->util_avg
> cap
)
734 sa
->util_sum
= sa
->util_avg
* LOAD_AVG_MAX
;
738 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq
*cfs_rq
);
739 static inline unsigned long cfs_rq_load_avg(struct cfs_rq
*cfs_rq
);
741 void init_entity_runnable_average(struct sched_entity
*se
)
744 void post_init_entity_util_avg(struct sched_entity
*se
)
750 * Update the current task's runtime statistics.
752 static void update_curr(struct cfs_rq
*cfs_rq
)
754 struct sched_entity
*curr
= cfs_rq
->curr
;
755 u64 now
= rq_clock_task(rq_of(cfs_rq
));
761 delta_exec
= now
- curr
->exec_start
;
762 if (unlikely((s64
)delta_exec
<= 0))
765 curr
->exec_start
= now
;
767 schedstat_set(curr
->statistics
.exec_max
,
768 max(delta_exec
, curr
->statistics
.exec_max
));
770 curr
->sum_exec_runtime
+= delta_exec
;
771 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
773 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
774 update_min_vruntime(cfs_rq
);
776 if (entity_is_task(curr
)) {
777 struct task_struct
*curtask
= task_of(curr
);
779 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
780 cpuacct_charge(curtask
, delta_exec
);
781 account_group_exec_runtime(curtask
, delta_exec
);
784 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
787 static void update_curr_fair(struct rq
*rq
)
789 update_curr(cfs_rq_of(&rq
->curr
->se
));
792 #ifdef CONFIG_SCHEDSTATS
794 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
796 u64 wait_start
= rq_clock(rq_of(cfs_rq
));
798 if (entity_is_task(se
) && task_on_rq_migrating(task_of(se
)) &&
799 likely(wait_start
> se
->statistics
.wait_start
))
800 wait_start
-= se
->statistics
.wait_start
;
802 se
->statistics
.wait_start
= wait_start
;
806 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
808 struct task_struct
*p
;
811 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
;
813 if (entity_is_task(se
)) {
815 if (task_on_rq_migrating(p
)) {
817 * Preserve migrating task's wait time so wait_start
818 * time stamp can be adjusted to accumulate wait time
819 * prior to migration.
821 se
->statistics
.wait_start
= delta
;
824 trace_sched_stat_wait(p
, delta
);
827 se
->statistics
.wait_max
= max(se
->statistics
.wait_max
, delta
);
828 se
->statistics
.wait_count
++;
829 se
->statistics
.wait_sum
+= delta
;
830 se
->statistics
.wait_start
= 0;
834 * Task is being enqueued - update stats:
837 update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
840 * Are we enqueueing a waiting task? (for current tasks
841 * a dequeue/enqueue event is a NOP)
843 if (se
!= cfs_rq
->curr
)
844 update_stats_wait_start(cfs_rq
, se
);
848 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
851 * Mark the end of the wait period if dequeueing a
854 if (se
!= cfs_rq
->curr
)
855 update_stats_wait_end(cfs_rq
, se
);
857 if (flags
& DEQUEUE_SLEEP
) {
858 if (entity_is_task(se
)) {
859 struct task_struct
*tsk
= task_of(se
);
861 if (tsk
->state
& TASK_INTERRUPTIBLE
)
862 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
863 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
864 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
871 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
876 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
881 update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
886 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
892 * We are picking a new current task - update its stats:
895 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
898 * We are starting a new run period:
900 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
903 /**************************************************
904 * Scheduling class queueing methods:
907 #ifdef CONFIG_NUMA_BALANCING
909 * Approximate time to scan a full NUMA task in ms. The task scan period is
910 * calculated based on the tasks virtual memory size and
911 * numa_balancing_scan_size.
913 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
914 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
916 /* Portion of address space to scan in MB */
917 unsigned int sysctl_numa_balancing_scan_size
= 256;
919 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
920 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
922 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
924 unsigned long rss
= 0;
925 unsigned long nr_scan_pages
;
928 * Calculations based on RSS as non-present and empty pages are skipped
929 * by the PTE scanner and NUMA hinting faults should be trapped based
932 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
933 rss
= get_mm_rss(p
->mm
);
937 rss
= round_up(rss
, nr_scan_pages
);
938 return rss
/ nr_scan_pages
;
941 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
942 #define MAX_SCAN_WINDOW 2560
944 static unsigned int task_scan_min(struct task_struct
*p
)
946 unsigned int scan_size
= READ_ONCE(sysctl_numa_balancing_scan_size
);
947 unsigned int scan
, floor
;
948 unsigned int windows
= 1;
950 if (scan_size
< MAX_SCAN_WINDOW
)
951 windows
= MAX_SCAN_WINDOW
/ scan_size
;
952 floor
= 1000 / windows
;
954 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
955 return max_t(unsigned int, floor
, scan
);
958 static unsigned int task_scan_max(struct task_struct
*p
)
960 unsigned int smin
= task_scan_min(p
);
963 /* Watch for min being lower than max due to floor calculations */
964 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
965 return max(smin
, smax
);
968 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
970 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
971 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
974 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
976 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
977 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
983 spinlock_t lock
; /* nr_tasks, tasks */
989 unsigned long total_faults
;
990 unsigned long max_faults_cpu
;
992 * Faults_cpu is used to decide whether memory should move
993 * towards the CPU. As a consequence, these stats are weighted
994 * more by CPU use than by memory faults.
996 unsigned long *faults_cpu
;
997 unsigned long faults
[0];
1000 /* Shared or private faults. */
1001 #define NR_NUMA_HINT_FAULT_TYPES 2
1003 /* Memory and CPU locality */
1004 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1006 /* Averaged statistics, and temporary buffers. */
1007 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1009 pid_t
task_numa_group_id(struct task_struct
*p
)
1011 return p
->numa_group
? p
->numa_group
->gid
: 0;
1015 * The averaged statistics, shared & private, memory & cpu,
1016 * occupy the first half of the array. The second half of the
1017 * array is for current counters, which are averaged into the
1018 * first set by task_numa_placement.
1020 static inline int task_faults_idx(enum numa_faults_stats s
, int nid
, int priv
)
1022 return NR_NUMA_HINT_FAULT_TYPES
* (s
* nr_node_ids
+ nid
) + priv
;
1025 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
1027 if (!p
->numa_faults
)
1030 return p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1031 p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1034 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
1039 return p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1040 p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1043 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
1045 return group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
1046 group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 1)];
1050 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1051 * considered part of a numa group's pseudo-interleaving set. Migrations
1052 * between these nodes are slowed down, to allow things to settle down.
1054 #define ACTIVE_NODE_FRACTION 3
1056 static bool numa_is_active_node(int nid
, struct numa_group
*ng
)
1058 return group_faults_cpu(ng
, nid
) * ACTIVE_NODE_FRACTION
> ng
->max_faults_cpu
;
1061 /* Handle placement on systems where not all nodes are directly connected. */
1062 static unsigned long score_nearby_nodes(struct task_struct
*p
, int nid
,
1063 int maxdist
, bool task
)
1065 unsigned long score
= 0;
1069 * All nodes are directly connected, and the same distance
1070 * from each other. No need for fancy placement algorithms.
1072 if (sched_numa_topology_type
== NUMA_DIRECT
)
1076 * This code is called for each node, introducing N^2 complexity,
1077 * which should be ok given the number of nodes rarely exceeds 8.
1079 for_each_online_node(node
) {
1080 unsigned long faults
;
1081 int dist
= node_distance(nid
, node
);
1084 * The furthest away nodes in the system are not interesting
1085 * for placement; nid was already counted.
1087 if (dist
== sched_max_numa_distance
|| node
== nid
)
1091 * On systems with a backplane NUMA topology, compare groups
1092 * of nodes, and move tasks towards the group with the most
1093 * memory accesses. When comparing two nodes at distance
1094 * "hoplimit", only nodes closer by than "hoplimit" are part
1095 * of each group. Skip other nodes.
1097 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1101 /* Add up the faults from nearby nodes. */
1103 faults
= task_faults(p
, node
);
1105 faults
= group_faults(p
, node
);
1108 * On systems with a glueless mesh NUMA topology, there are
1109 * no fixed "groups of nodes". Instead, nodes that are not
1110 * directly connected bounce traffic through intermediate
1111 * nodes; a numa_group can occupy any set of nodes.
1112 * The further away a node is, the less the faults count.
1113 * This seems to result in good task placement.
1115 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1116 faults
*= (sched_max_numa_distance
- dist
);
1117 faults
/= (sched_max_numa_distance
- LOCAL_DISTANCE
);
1127 * These return the fraction of accesses done by a particular task, or
1128 * task group, on a particular numa node. The group weight is given a
1129 * larger multiplier, in order to group tasks together that are almost
1130 * evenly spread out between numa nodes.
1132 static inline unsigned long task_weight(struct task_struct
*p
, int nid
,
1135 unsigned long faults
, total_faults
;
1137 if (!p
->numa_faults
)
1140 total_faults
= p
->total_numa_faults
;
1145 faults
= task_faults(p
, nid
);
1146 faults
+= score_nearby_nodes(p
, nid
, dist
, true);
1148 return 1000 * faults
/ total_faults
;
1151 static inline unsigned long group_weight(struct task_struct
*p
, int nid
,
1154 unsigned long faults
, total_faults
;
1159 total_faults
= p
->numa_group
->total_faults
;
1164 faults
= group_faults(p
, nid
);
1165 faults
+= score_nearby_nodes(p
, nid
, dist
, false);
1167 return 1000 * faults
/ total_faults
;
1170 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
1171 int src_nid
, int dst_cpu
)
1173 struct numa_group
*ng
= p
->numa_group
;
1174 int dst_nid
= cpu_to_node(dst_cpu
);
1175 int last_cpupid
, this_cpupid
;
1177 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
1180 * Multi-stage node selection is used in conjunction with a periodic
1181 * migration fault to build a temporal task<->page relation. By using
1182 * a two-stage filter we remove short/unlikely relations.
1184 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1185 * a task's usage of a particular page (n_p) per total usage of this
1186 * page (n_t) (in a given time-span) to a probability.
1188 * Our periodic faults will sample this probability and getting the
1189 * same result twice in a row, given these samples are fully
1190 * independent, is then given by P(n)^2, provided our sample period
1191 * is sufficiently short compared to the usage pattern.
1193 * This quadric squishes small probabilities, making it less likely we
1194 * act on an unlikely task<->page relation.
1196 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
1197 if (!cpupid_pid_unset(last_cpupid
) &&
1198 cpupid_to_nid(last_cpupid
) != dst_nid
)
1201 /* Always allow migrate on private faults */
1202 if (cpupid_match_pid(p
, last_cpupid
))
1205 /* A shared fault, but p->numa_group has not been set up yet. */
1210 * Destination node is much more heavily used than the source
1211 * node? Allow migration.
1213 if (group_faults_cpu(ng
, dst_nid
) > group_faults_cpu(ng
, src_nid
) *
1214 ACTIVE_NODE_FRACTION
)
1218 * Distribute memory according to CPU & memory use on each node,
1219 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1221 * faults_cpu(dst) 3 faults_cpu(src)
1222 * --------------- * - > ---------------
1223 * faults_mem(dst) 4 faults_mem(src)
1225 return group_faults_cpu(ng
, dst_nid
) * group_faults(p
, src_nid
) * 3 >
1226 group_faults_cpu(ng
, src_nid
) * group_faults(p
, dst_nid
) * 4;
1229 static unsigned long weighted_cpuload(const int cpu
);
1230 static unsigned long source_load(int cpu
, int type
);
1231 static unsigned long target_load(int cpu
, int type
);
1232 static unsigned long capacity_of(int cpu
);
1233 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
1235 /* Cached statistics for all CPUs within a node */
1237 unsigned long nr_running
;
1240 /* Total compute capacity of CPUs on a node */
1241 unsigned long compute_capacity
;
1243 /* Approximate capacity in terms of runnable tasks on a node */
1244 unsigned long task_capacity
;
1245 int has_free_capacity
;
1249 * XXX borrowed from update_sg_lb_stats
1251 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1253 int smt
, cpu
, cpus
= 0;
1254 unsigned long capacity
;
1256 memset(ns
, 0, sizeof(*ns
));
1257 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1258 struct rq
*rq
= cpu_rq(cpu
);
1260 ns
->nr_running
+= rq
->nr_running
;
1261 ns
->load
+= weighted_cpuload(cpu
);
1262 ns
->compute_capacity
+= capacity_of(cpu
);
1268 * If we raced with hotplug and there are no CPUs left in our mask
1269 * the @ns structure is NULL'ed and task_numa_compare() will
1270 * not find this node attractive.
1272 * We'll either bail at !has_free_capacity, or we'll detect a huge
1273 * imbalance and bail there.
1278 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1279 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, ns
->compute_capacity
);
1280 capacity
= cpus
/ smt
; /* cores */
1282 ns
->task_capacity
= min_t(unsigned, capacity
,
1283 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
));
1284 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1287 struct task_numa_env
{
1288 struct task_struct
*p
;
1290 int src_cpu
, src_nid
;
1291 int dst_cpu
, dst_nid
;
1293 struct numa_stats src_stats
, dst_stats
;
1298 struct task_struct
*best_task
;
1303 static void task_numa_assign(struct task_numa_env
*env
,
1304 struct task_struct
*p
, long imp
)
1307 put_task_struct(env
->best_task
);
1312 env
->best_imp
= imp
;
1313 env
->best_cpu
= env
->dst_cpu
;
1316 static bool load_too_imbalanced(long src_load
, long dst_load
,
1317 struct task_numa_env
*env
)
1320 long orig_src_load
, orig_dst_load
;
1321 long src_capacity
, dst_capacity
;
1324 * The load is corrected for the CPU capacity available on each node.
1327 * ------------ vs ---------
1328 * src_capacity dst_capacity
1330 src_capacity
= env
->src_stats
.compute_capacity
;
1331 dst_capacity
= env
->dst_stats
.compute_capacity
;
1333 /* We care about the slope of the imbalance, not the direction. */
1334 if (dst_load
< src_load
)
1335 swap(dst_load
, src_load
);
1337 /* Is the difference below the threshold? */
1338 imb
= dst_load
* src_capacity
* 100 -
1339 src_load
* dst_capacity
* env
->imbalance_pct
;
1344 * The imbalance is above the allowed threshold.
1345 * Compare it with the old imbalance.
1347 orig_src_load
= env
->src_stats
.load
;
1348 orig_dst_load
= env
->dst_stats
.load
;
1350 if (orig_dst_load
< orig_src_load
)
1351 swap(orig_dst_load
, orig_src_load
);
1353 old_imb
= orig_dst_load
* src_capacity
* 100 -
1354 orig_src_load
* dst_capacity
* env
->imbalance_pct
;
1356 /* Would this change make things worse? */
1357 return (imb
> old_imb
);
1361 * This checks if the overall compute and NUMA accesses of the system would
1362 * be improved if the source tasks was migrated to the target dst_cpu taking
1363 * into account that it might be best if task running on the dst_cpu should
1364 * be exchanged with the source task
1366 static void task_numa_compare(struct task_numa_env
*env
,
1367 long taskimp
, long groupimp
)
1369 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1370 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1371 struct task_struct
*cur
;
1372 long src_load
, dst_load
;
1374 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1376 int dist
= env
->dist
;
1379 cur
= task_rcu_dereference(&dst_rq
->curr
);
1380 if (cur
&& ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
)))
1384 * Because we have preemption enabled we can get migrated around and
1385 * end try selecting ourselves (current == env->p) as a swap candidate.
1391 * "imp" is the fault differential for the source task between the
1392 * source and destination node. Calculate the total differential for
1393 * the source task and potential destination task. The more negative
1394 * the value is, the more rmeote accesses that would be expected to
1395 * be incurred if the tasks were swapped.
1398 /* Skip this swap candidate if cannot move to the source cpu */
1399 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1403 * If dst and source tasks are in the same NUMA group, or not
1404 * in any group then look only at task weights.
1406 if (cur
->numa_group
== env
->p
->numa_group
) {
1407 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1408 task_weight(cur
, env
->dst_nid
, dist
);
1410 * Add some hysteresis to prevent swapping the
1411 * tasks within a group over tiny differences.
1413 if (cur
->numa_group
)
1417 * Compare the group weights. If a task is all by
1418 * itself (not part of a group), use the task weight
1421 if (cur
->numa_group
)
1422 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1423 group_weight(cur
, env
->dst_nid
, dist
);
1425 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1426 task_weight(cur
, env
->dst_nid
, dist
);
1430 if (imp
<= env
->best_imp
&& moveimp
<= env
->best_imp
)
1434 /* Is there capacity at our destination? */
1435 if (env
->src_stats
.nr_running
<= env
->src_stats
.task_capacity
&&
1436 !env
->dst_stats
.has_free_capacity
)
1442 /* Balance doesn't matter much if we're running a task per cpu */
1443 if (imp
> env
->best_imp
&& src_rq
->nr_running
== 1 &&
1444 dst_rq
->nr_running
== 1)
1448 * In the overloaded case, try and keep the load balanced.
1451 load
= task_h_load(env
->p
);
1452 dst_load
= env
->dst_stats
.load
+ load
;
1453 src_load
= env
->src_stats
.load
- load
;
1455 if (moveimp
> imp
&& moveimp
> env
->best_imp
) {
1457 * If the improvement from just moving env->p direction is
1458 * better than swapping tasks around, check if a move is
1459 * possible. Store a slightly smaller score than moveimp,
1460 * so an actually idle CPU will win.
1462 if (!load_too_imbalanced(src_load
, dst_load
, env
)) {
1469 if (imp
<= env
->best_imp
)
1473 load
= task_h_load(cur
);
1478 if (load_too_imbalanced(src_load
, dst_load
, env
))
1482 * One idle CPU per node is evaluated for a task numa move.
1483 * Call select_idle_sibling to maybe find a better one.
1486 env
->dst_cpu
= select_idle_sibling(env
->p
, env
->dst_cpu
);
1489 task_numa_assign(env
, cur
, imp
);
1494 static void task_numa_find_cpu(struct task_numa_env
*env
,
1495 long taskimp
, long groupimp
)
1499 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1500 /* Skip this CPU if the source task cannot migrate */
1501 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1505 task_numa_compare(env
, taskimp
, groupimp
);
1509 /* Only move tasks to a NUMA node less busy than the current node. */
1510 static bool numa_has_capacity(struct task_numa_env
*env
)
1512 struct numa_stats
*src
= &env
->src_stats
;
1513 struct numa_stats
*dst
= &env
->dst_stats
;
1515 if (src
->has_free_capacity
&& !dst
->has_free_capacity
)
1519 * Only consider a task move if the source has a higher load
1520 * than the destination, corrected for CPU capacity on each node.
1522 * src->load dst->load
1523 * --------------------- vs ---------------------
1524 * src->compute_capacity dst->compute_capacity
1526 if (src
->load
* dst
->compute_capacity
* env
->imbalance_pct
>
1528 dst
->load
* src
->compute_capacity
* 100)
1534 static int task_numa_migrate(struct task_struct
*p
)
1536 struct task_numa_env env
= {
1539 .src_cpu
= task_cpu(p
),
1540 .src_nid
= task_node(p
),
1542 .imbalance_pct
= 112,
1548 struct sched_domain
*sd
;
1549 unsigned long taskweight
, groupweight
;
1551 long taskimp
, groupimp
;
1554 * Pick the lowest SD_NUMA domain, as that would have the smallest
1555 * imbalance and would be the first to start moving tasks about.
1557 * And we want to avoid any moving of tasks about, as that would create
1558 * random movement of tasks -- counter the numa conditions we're trying
1562 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1564 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1568 * Cpusets can break the scheduler domain tree into smaller
1569 * balance domains, some of which do not cross NUMA boundaries.
1570 * Tasks that are "trapped" in such domains cannot be migrated
1571 * elsewhere, so there is no point in (re)trying.
1573 if (unlikely(!sd
)) {
1574 p
->numa_preferred_nid
= task_node(p
);
1578 env
.dst_nid
= p
->numa_preferred_nid
;
1579 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1580 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1581 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1582 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1583 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
1584 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
1585 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1587 /* Try to find a spot on the preferred nid. */
1588 if (numa_has_capacity(&env
))
1589 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1592 * Look at other nodes in these cases:
1593 * - there is no space available on the preferred_nid
1594 * - the task is part of a numa_group that is interleaved across
1595 * multiple NUMA nodes; in order to better consolidate the group,
1596 * we need to check other locations.
1598 if (env
.best_cpu
== -1 || (p
->numa_group
&& p
->numa_group
->active_nodes
> 1)) {
1599 for_each_online_node(nid
) {
1600 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1603 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1604 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1606 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1607 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1610 /* Only consider nodes where both task and groups benefit */
1611 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
1612 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
1613 if (taskimp
< 0 && groupimp
< 0)
1618 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1619 if (numa_has_capacity(&env
))
1620 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1625 * If the task is part of a workload that spans multiple NUMA nodes,
1626 * and is migrating into one of the workload's active nodes, remember
1627 * this node as the task's preferred numa node, so the workload can
1629 * A task that migrated to a second choice node will be better off
1630 * trying for a better one later. Do not set the preferred node here.
1632 if (p
->numa_group
) {
1633 struct numa_group
*ng
= p
->numa_group
;
1635 if (env
.best_cpu
== -1)
1640 if (ng
->active_nodes
> 1 && numa_is_active_node(env
.dst_nid
, ng
))
1641 sched_setnuma(p
, env
.dst_nid
);
1644 /* No better CPU than the current one was found. */
1645 if (env
.best_cpu
== -1)
1649 * Reset the scan period if the task is being rescheduled on an
1650 * alternative node to recheck if the tasks is now properly placed.
1652 p
->numa_scan_period
= task_scan_min(p
);
1654 if (env
.best_task
== NULL
) {
1655 ret
= migrate_task_to(p
, env
.best_cpu
);
1657 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1661 ret
= migrate_swap(p
, env
.best_task
);
1663 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1664 put_task_struct(env
.best_task
);
1668 /* Attempt to migrate a task to a CPU on the preferred node. */
1669 static void numa_migrate_preferred(struct task_struct
*p
)
1671 unsigned long interval
= HZ
;
1673 /* This task has no NUMA fault statistics yet */
1674 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults
))
1677 /* Periodically retry migrating the task to the preferred node */
1678 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1679 p
->numa_migrate_retry
= jiffies
+ interval
;
1681 /* Success if task is already running on preferred CPU */
1682 if (task_node(p
) == p
->numa_preferred_nid
)
1685 /* Otherwise, try migrate to a CPU on the preferred node */
1686 task_numa_migrate(p
);
1690 * Find out how many nodes on the workload is actively running on. Do this by
1691 * tracking the nodes from which NUMA hinting faults are triggered. This can
1692 * be different from the set of nodes where the workload's memory is currently
1695 static void numa_group_count_active_nodes(struct numa_group
*numa_group
)
1697 unsigned long faults
, max_faults
= 0;
1698 int nid
, active_nodes
= 0;
1700 for_each_online_node(nid
) {
1701 faults
= group_faults_cpu(numa_group
, nid
);
1702 if (faults
> max_faults
)
1703 max_faults
= faults
;
1706 for_each_online_node(nid
) {
1707 faults
= group_faults_cpu(numa_group
, nid
);
1708 if (faults
* ACTIVE_NODE_FRACTION
> max_faults
)
1712 numa_group
->max_faults_cpu
= max_faults
;
1713 numa_group
->active_nodes
= active_nodes
;
1717 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1718 * increments. The more local the fault statistics are, the higher the scan
1719 * period will be for the next scan window. If local/(local+remote) ratio is
1720 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1721 * the scan period will decrease. Aim for 70% local accesses.
1723 #define NUMA_PERIOD_SLOTS 10
1724 #define NUMA_PERIOD_THRESHOLD 7
1727 * Increase the scan period (slow down scanning) if the majority of
1728 * our memory is already on our local node, or if the majority of
1729 * the page accesses are shared with other processes.
1730 * Otherwise, decrease the scan period.
1732 static void update_task_scan_period(struct task_struct
*p
,
1733 unsigned long shared
, unsigned long private)
1735 unsigned int period_slot
;
1739 unsigned long remote
= p
->numa_faults_locality
[0];
1740 unsigned long local
= p
->numa_faults_locality
[1];
1743 * If there were no record hinting faults then either the task is
1744 * completely idle or all activity is areas that are not of interest
1745 * to automatic numa balancing. Related to that, if there were failed
1746 * migration then it implies we are migrating too quickly or the local
1747 * node is overloaded. In either case, scan slower
1749 if (local
+ shared
== 0 || p
->numa_faults_locality
[2]) {
1750 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1751 p
->numa_scan_period
<< 1);
1753 p
->mm
->numa_next_scan
= jiffies
+
1754 msecs_to_jiffies(p
->numa_scan_period
);
1760 * Prepare to scale scan period relative to the current period.
1761 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1762 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1763 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1765 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1766 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1767 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1768 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1771 diff
= slot
* period_slot
;
1773 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1776 * Scale scan rate increases based on sharing. There is an
1777 * inverse relationship between the degree of sharing and
1778 * the adjustment made to the scanning period. Broadly
1779 * speaking the intent is that there is little point
1780 * scanning faster if shared accesses dominate as it may
1781 * simply bounce migrations uselessly
1783 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
+ 1));
1784 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1787 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1788 task_scan_min(p
), task_scan_max(p
));
1789 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1793 * Get the fraction of time the task has been running since the last
1794 * NUMA placement cycle. The scheduler keeps similar statistics, but
1795 * decays those on a 32ms period, which is orders of magnitude off
1796 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1797 * stats only if the task is so new there are no NUMA statistics yet.
1799 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1801 u64 runtime
, delta
, now
;
1802 /* Use the start of this time slice to avoid calculations. */
1803 now
= p
->se
.exec_start
;
1804 runtime
= p
->se
.sum_exec_runtime
;
1806 if (p
->last_task_numa_placement
) {
1807 delta
= runtime
- p
->last_sum_exec_runtime
;
1808 *period
= now
- p
->last_task_numa_placement
;
1810 delta
= p
->se
.avg
.load_sum
/ p
->se
.load
.weight
;
1811 *period
= LOAD_AVG_MAX
;
1814 p
->last_sum_exec_runtime
= runtime
;
1815 p
->last_task_numa_placement
= now
;
1821 * Determine the preferred nid for a task in a numa_group. This needs to
1822 * be done in a way that produces consistent results with group_weight,
1823 * otherwise workloads might not converge.
1825 static int preferred_group_nid(struct task_struct
*p
, int nid
)
1830 /* Direct connections between all NUMA nodes. */
1831 if (sched_numa_topology_type
== NUMA_DIRECT
)
1835 * On a system with glueless mesh NUMA topology, group_weight
1836 * scores nodes according to the number of NUMA hinting faults on
1837 * both the node itself, and on nearby nodes.
1839 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1840 unsigned long score
, max_score
= 0;
1841 int node
, max_node
= nid
;
1843 dist
= sched_max_numa_distance
;
1845 for_each_online_node(node
) {
1846 score
= group_weight(p
, node
, dist
);
1847 if (score
> max_score
) {
1856 * Finding the preferred nid in a system with NUMA backplane
1857 * interconnect topology is more involved. The goal is to locate
1858 * tasks from numa_groups near each other in the system, and
1859 * untangle workloads from different sides of the system. This requires
1860 * searching down the hierarchy of node groups, recursively searching
1861 * inside the highest scoring group of nodes. The nodemask tricks
1862 * keep the complexity of the search down.
1864 nodes
= node_online_map
;
1865 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
1866 unsigned long max_faults
= 0;
1867 nodemask_t max_group
= NODE_MASK_NONE
;
1870 /* Are there nodes at this distance from each other? */
1871 if (!find_numa_distance(dist
))
1874 for_each_node_mask(a
, nodes
) {
1875 unsigned long faults
= 0;
1876 nodemask_t this_group
;
1877 nodes_clear(this_group
);
1879 /* Sum group's NUMA faults; includes a==b case. */
1880 for_each_node_mask(b
, nodes
) {
1881 if (node_distance(a
, b
) < dist
) {
1882 faults
+= group_faults(p
, b
);
1883 node_set(b
, this_group
);
1884 node_clear(b
, nodes
);
1888 /* Remember the top group. */
1889 if (faults
> max_faults
) {
1890 max_faults
= faults
;
1891 max_group
= this_group
;
1893 * subtle: at the smallest distance there is
1894 * just one node left in each "group", the
1895 * winner is the preferred nid.
1900 /* Next round, evaluate the nodes within max_group. */
1908 static void task_numa_placement(struct task_struct
*p
)
1910 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1911 unsigned long max_faults
= 0, max_group_faults
= 0;
1912 unsigned long fault_types
[2] = { 0, 0 };
1913 unsigned long total_faults
;
1914 u64 runtime
, period
;
1915 spinlock_t
*group_lock
= NULL
;
1918 * The p->mm->numa_scan_seq field gets updated without
1919 * exclusive access. Use READ_ONCE() here to ensure
1920 * that the field is read in a single access:
1922 seq
= READ_ONCE(p
->mm
->numa_scan_seq
);
1923 if (p
->numa_scan_seq
== seq
)
1925 p
->numa_scan_seq
= seq
;
1926 p
->numa_scan_period_max
= task_scan_max(p
);
1928 total_faults
= p
->numa_faults_locality
[0] +
1929 p
->numa_faults_locality
[1];
1930 runtime
= numa_get_avg_runtime(p
, &period
);
1932 /* If the task is part of a group prevent parallel updates to group stats */
1933 if (p
->numa_group
) {
1934 group_lock
= &p
->numa_group
->lock
;
1935 spin_lock_irq(group_lock
);
1938 /* Find the node with the highest number of faults */
1939 for_each_online_node(nid
) {
1940 /* Keep track of the offsets in numa_faults array */
1941 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
1942 unsigned long faults
= 0, group_faults
= 0;
1945 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
1946 long diff
, f_diff
, f_weight
;
1948 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
1949 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
1950 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
1951 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
1953 /* Decay existing window, copy faults since last scan */
1954 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
1955 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
1956 p
->numa_faults
[membuf_idx
] = 0;
1959 * Normalize the faults_from, so all tasks in a group
1960 * count according to CPU use, instead of by the raw
1961 * number of faults. Tasks with little runtime have
1962 * little over-all impact on throughput, and thus their
1963 * faults are less important.
1965 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
1966 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
1968 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
1969 p
->numa_faults
[cpubuf_idx
] = 0;
1971 p
->numa_faults
[mem_idx
] += diff
;
1972 p
->numa_faults
[cpu_idx
] += f_diff
;
1973 faults
+= p
->numa_faults
[mem_idx
];
1974 p
->total_numa_faults
+= diff
;
1975 if (p
->numa_group
) {
1977 * safe because we can only change our own group
1979 * mem_idx represents the offset for a given
1980 * nid and priv in a specific region because it
1981 * is at the beginning of the numa_faults array.
1983 p
->numa_group
->faults
[mem_idx
] += diff
;
1984 p
->numa_group
->faults_cpu
[mem_idx
] += f_diff
;
1985 p
->numa_group
->total_faults
+= diff
;
1986 group_faults
+= p
->numa_group
->faults
[mem_idx
];
1990 if (faults
> max_faults
) {
1991 max_faults
= faults
;
1995 if (group_faults
> max_group_faults
) {
1996 max_group_faults
= group_faults
;
1997 max_group_nid
= nid
;
2001 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
2003 if (p
->numa_group
) {
2004 numa_group_count_active_nodes(p
->numa_group
);
2005 spin_unlock_irq(group_lock
);
2006 max_nid
= preferred_group_nid(p
, max_group_nid
);
2010 /* Set the new preferred node */
2011 if (max_nid
!= p
->numa_preferred_nid
)
2012 sched_setnuma(p
, max_nid
);
2014 if (task_node(p
) != p
->numa_preferred_nid
)
2015 numa_migrate_preferred(p
);
2019 static inline int get_numa_group(struct numa_group
*grp
)
2021 return atomic_inc_not_zero(&grp
->refcount
);
2024 static inline void put_numa_group(struct numa_group
*grp
)
2026 if (atomic_dec_and_test(&grp
->refcount
))
2027 kfree_rcu(grp
, rcu
);
2030 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
2033 struct numa_group
*grp
, *my_grp
;
2034 struct task_struct
*tsk
;
2036 int cpu
= cpupid_to_cpu(cpupid
);
2039 if (unlikely(!p
->numa_group
)) {
2040 unsigned int size
= sizeof(struct numa_group
) +
2041 4*nr_node_ids
*sizeof(unsigned long);
2043 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
2047 atomic_set(&grp
->refcount
, 1);
2048 grp
->active_nodes
= 1;
2049 grp
->max_faults_cpu
= 0;
2050 spin_lock_init(&grp
->lock
);
2052 /* Second half of the array tracks nids where faults happen */
2053 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
2056 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2057 grp
->faults
[i
] = p
->numa_faults
[i
];
2059 grp
->total_faults
= p
->total_numa_faults
;
2062 rcu_assign_pointer(p
->numa_group
, grp
);
2066 tsk
= READ_ONCE(cpu_rq(cpu
)->curr
);
2068 if (!cpupid_match_pid(tsk
, cpupid
))
2071 grp
= rcu_dereference(tsk
->numa_group
);
2075 my_grp
= p
->numa_group
;
2080 * Only join the other group if its bigger; if we're the bigger group,
2081 * the other task will join us.
2083 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
2087 * Tie-break on the grp address.
2089 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
2092 /* Always join threads in the same process. */
2093 if (tsk
->mm
== current
->mm
)
2096 /* Simple filter to avoid false positives due to PID collisions */
2097 if (flags
& TNF_SHARED
)
2100 /* Update priv based on whether false sharing was detected */
2103 if (join
&& !get_numa_group(grp
))
2111 BUG_ON(irqs_disabled());
2112 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
2114 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
2115 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
2116 grp
->faults
[i
] += p
->numa_faults
[i
];
2118 my_grp
->total_faults
-= p
->total_numa_faults
;
2119 grp
->total_faults
+= p
->total_numa_faults
;
2124 spin_unlock(&my_grp
->lock
);
2125 spin_unlock_irq(&grp
->lock
);
2127 rcu_assign_pointer(p
->numa_group
, grp
);
2129 put_numa_group(my_grp
);
2137 void task_numa_free(struct task_struct
*p
)
2139 struct numa_group
*grp
= p
->numa_group
;
2140 void *numa_faults
= p
->numa_faults
;
2141 unsigned long flags
;
2145 spin_lock_irqsave(&grp
->lock
, flags
);
2146 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2147 grp
->faults
[i
] -= p
->numa_faults
[i
];
2148 grp
->total_faults
-= p
->total_numa_faults
;
2151 spin_unlock_irqrestore(&grp
->lock
, flags
);
2152 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2153 put_numa_group(grp
);
2156 p
->numa_faults
= NULL
;
2161 * Got a PROT_NONE fault for a page on @node.
2163 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2165 struct task_struct
*p
= current
;
2166 bool migrated
= flags
& TNF_MIGRATED
;
2167 int cpu_node
= task_node(current
);
2168 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2169 struct numa_group
*ng
;
2172 if (!static_branch_likely(&sched_numa_balancing
))
2175 /* for example, ksmd faulting in a user's mm */
2179 /* Allocate buffer to track faults on a per-node basis */
2180 if (unlikely(!p
->numa_faults
)) {
2181 int size
= sizeof(*p
->numa_faults
) *
2182 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2184 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2185 if (!p
->numa_faults
)
2188 p
->total_numa_faults
= 0;
2189 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2193 * First accesses are treated as private, otherwise consider accesses
2194 * to be private if the accessing pid has not changed
2196 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2199 priv
= cpupid_match_pid(p
, last_cpupid
);
2200 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2201 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2205 * If a workload spans multiple NUMA nodes, a shared fault that
2206 * occurs wholly within the set of nodes that the workload is
2207 * actively using should be counted as local. This allows the
2208 * scan rate to slow down when a workload has settled down.
2211 if (!priv
&& !local
&& ng
&& ng
->active_nodes
> 1 &&
2212 numa_is_active_node(cpu_node
, ng
) &&
2213 numa_is_active_node(mem_node
, ng
))
2216 task_numa_placement(p
);
2219 * Retry task to preferred node migration periodically, in case it
2220 * case it previously failed, or the scheduler moved us.
2222 if (time_after(jiffies
, p
->numa_migrate_retry
))
2223 numa_migrate_preferred(p
);
2226 p
->numa_pages_migrated
+= pages
;
2227 if (flags
& TNF_MIGRATE_FAIL
)
2228 p
->numa_faults_locality
[2] += pages
;
2230 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2231 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2232 p
->numa_faults_locality
[local
] += pages
;
2235 static void reset_ptenuma_scan(struct task_struct
*p
)
2238 * We only did a read acquisition of the mmap sem, so
2239 * p->mm->numa_scan_seq is written to without exclusive access
2240 * and the update is not guaranteed to be atomic. That's not
2241 * much of an issue though, since this is just used for
2242 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2243 * expensive, to avoid any form of compiler optimizations:
2245 WRITE_ONCE(p
->mm
->numa_scan_seq
, READ_ONCE(p
->mm
->numa_scan_seq
) + 1);
2246 p
->mm
->numa_scan_offset
= 0;
2250 * The expensive part of numa migration is done from task_work context.
2251 * Triggered from task_tick_numa().
2253 void task_numa_work(struct callback_head
*work
)
2255 unsigned long migrate
, next_scan
, now
= jiffies
;
2256 struct task_struct
*p
= current
;
2257 struct mm_struct
*mm
= p
->mm
;
2258 u64 runtime
= p
->se
.sum_exec_runtime
;
2259 struct vm_area_struct
*vma
;
2260 unsigned long start
, end
;
2261 unsigned long nr_pte_updates
= 0;
2262 long pages
, virtpages
;
2264 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
2266 work
->next
= work
; /* protect against double add */
2268 * Who cares about NUMA placement when they're dying.
2270 * NOTE: make sure not to dereference p->mm before this check,
2271 * exit_task_work() happens _after_ exit_mm() so we could be called
2272 * without p->mm even though we still had it when we enqueued this
2275 if (p
->flags
& PF_EXITING
)
2278 if (!mm
->numa_next_scan
) {
2279 mm
->numa_next_scan
= now
+
2280 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2284 * Enforce maximal scan/migration frequency..
2286 migrate
= mm
->numa_next_scan
;
2287 if (time_before(now
, migrate
))
2290 if (p
->numa_scan_period
== 0) {
2291 p
->numa_scan_period_max
= task_scan_max(p
);
2292 p
->numa_scan_period
= task_scan_min(p
);
2295 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2296 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2300 * Delay this task enough that another task of this mm will likely win
2301 * the next time around.
2303 p
->node_stamp
+= 2 * TICK_NSEC
;
2305 start
= mm
->numa_scan_offset
;
2306 pages
= sysctl_numa_balancing_scan_size
;
2307 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2308 virtpages
= pages
* 8; /* Scan up to this much virtual space */
2313 down_read(&mm
->mmap_sem
);
2314 vma
= find_vma(mm
, start
);
2316 reset_ptenuma_scan(p
);
2320 for (; vma
; vma
= vma
->vm_next
) {
2321 if (!vma_migratable(vma
) || !vma_policy_mof(vma
) ||
2322 is_vm_hugetlb_page(vma
) || (vma
->vm_flags
& VM_MIXEDMAP
)) {
2327 * Shared library pages mapped by multiple processes are not
2328 * migrated as it is expected they are cache replicated. Avoid
2329 * hinting faults in read-only file-backed mappings or the vdso
2330 * as migrating the pages will be of marginal benefit.
2333 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2337 * Skip inaccessible VMAs to avoid any confusion between
2338 * PROT_NONE and NUMA hinting ptes
2340 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
2344 start
= max(start
, vma
->vm_start
);
2345 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2346 end
= min(end
, vma
->vm_end
);
2347 nr_pte_updates
= change_prot_numa(vma
, start
, end
);
2350 * Try to scan sysctl_numa_balancing_size worth of
2351 * hpages that have at least one present PTE that
2352 * is not already pte-numa. If the VMA contains
2353 * areas that are unused or already full of prot_numa
2354 * PTEs, scan up to virtpages, to skip through those
2358 pages
-= (end
- start
) >> PAGE_SHIFT
;
2359 virtpages
-= (end
- start
) >> PAGE_SHIFT
;
2362 if (pages
<= 0 || virtpages
<= 0)
2366 } while (end
!= vma
->vm_end
);
2371 * It is possible to reach the end of the VMA list but the last few
2372 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2373 * would find the !migratable VMA on the next scan but not reset the
2374 * scanner to the start so check it now.
2377 mm
->numa_scan_offset
= start
;
2379 reset_ptenuma_scan(p
);
2380 up_read(&mm
->mmap_sem
);
2383 * Make sure tasks use at least 32x as much time to run other code
2384 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2385 * Usually update_task_scan_period slows down scanning enough; on an
2386 * overloaded system we need to limit overhead on a per task basis.
2388 if (unlikely(p
->se
.sum_exec_runtime
!= runtime
)) {
2389 u64 diff
= p
->se
.sum_exec_runtime
- runtime
;
2390 p
->node_stamp
+= 32 * diff
;
2395 * Drive the periodic memory faults..
2397 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2399 struct callback_head
*work
= &curr
->numa_work
;
2403 * We don't care about NUMA placement if we don't have memory.
2405 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2409 * Using runtime rather than walltime has the dual advantage that
2410 * we (mostly) drive the selection from busy threads and that the
2411 * task needs to have done some actual work before we bother with
2414 now
= curr
->se
.sum_exec_runtime
;
2415 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2417 if (now
> curr
->node_stamp
+ period
) {
2418 if (!curr
->node_stamp
)
2419 curr
->numa_scan_period
= task_scan_min(curr
);
2420 curr
->node_stamp
+= period
;
2422 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2423 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2424 task_work_add(curr
, work
, true);
2429 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2433 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2437 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2440 #endif /* CONFIG_NUMA_BALANCING */
2443 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2445 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2446 if (!parent_entity(se
))
2447 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2449 if (entity_is_task(se
)) {
2450 struct rq
*rq
= rq_of(cfs_rq
);
2452 account_numa_enqueue(rq
, task_of(se
));
2453 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2456 cfs_rq
->nr_running
++;
2460 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2462 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2463 if (!parent_entity(se
))
2464 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2466 if (entity_is_task(se
)) {
2467 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2468 list_del_init(&se
->group_node
);
2471 cfs_rq
->nr_running
--;
2474 #ifdef CONFIG_FAIR_GROUP_SCHED
2476 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
2481 * Use this CPU's real-time load instead of the last load contribution
2482 * as the updating of the contribution is delayed, and we will use the
2483 * the real-time load to calc the share. See update_tg_load_avg().
2485 tg_weight
= atomic_long_read(&tg
->load_avg
);
2486 tg_weight
-= cfs_rq
->tg_load_avg_contrib
;
2487 tg_weight
+= cfs_rq
->load
.weight
;
2492 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2494 long tg_weight
, load
, shares
;
2496 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
2497 load
= cfs_rq
->load
.weight
;
2499 shares
= (tg
->shares
* load
);
2501 shares
/= tg_weight
;
2503 if (shares
< MIN_SHARES
)
2504 shares
= MIN_SHARES
;
2505 if (shares
> tg
->shares
)
2506 shares
= tg
->shares
;
2510 # else /* CONFIG_SMP */
2511 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2515 # endif /* CONFIG_SMP */
2516 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2517 unsigned long weight
)
2520 /* commit outstanding execution time */
2521 if (cfs_rq
->curr
== se
)
2522 update_curr(cfs_rq
);
2523 account_entity_dequeue(cfs_rq
, se
);
2526 update_load_set(&se
->load
, weight
);
2529 account_entity_enqueue(cfs_rq
, se
);
2532 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2534 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2536 struct task_group
*tg
;
2537 struct sched_entity
*se
;
2541 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2542 if (!se
|| throttled_hierarchy(cfs_rq
))
2545 if (likely(se
->load
.weight
== tg
->shares
))
2548 shares
= calc_cfs_shares(cfs_rq
, tg
);
2550 reweight_entity(cfs_rq_of(se
), se
, shares
);
2552 #else /* CONFIG_FAIR_GROUP_SCHED */
2553 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2556 #endif /* CONFIG_FAIR_GROUP_SCHED */
2559 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2560 static const u32 runnable_avg_yN_inv
[] = {
2561 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2562 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2563 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2564 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2565 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2566 0x85aac367, 0x82cd8698,
2570 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2571 * over-estimates when re-combining.
2573 static const u32 runnable_avg_yN_sum
[] = {
2574 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2575 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2576 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2580 * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to
2581 * lower integers. See Documentation/scheduler/sched-avg.txt how these
2584 static const u32 __accumulated_sum_N32
[] = {
2585 0, 23371, 35056, 40899, 43820, 45281,
2586 46011, 46376, 46559, 46650, 46696, 46719,
2591 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2593 static __always_inline u64
decay_load(u64 val
, u64 n
)
2595 unsigned int local_n
;
2599 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2602 /* after bounds checking we can collapse to 32-bit */
2606 * As y^PERIOD = 1/2, we can combine
2607 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2608 * With a look-up table which covers y^n (n<PERIOD)
2610 * To achieve constant time decay_load.
2612 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2613 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2614 local_n
%= LOAD_AVG_PERIOD
;
2617 val
= mul_u64_u32_shr(val
, runnable_avg_yN_inv
[local_n
], 32);
2622 * For updates fully spanning n periods, the contribution to runnable
2623 * average will be: \Sum 1024*y^n
2625 * We can compute this reasonably efficiently by combining:
2626 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2628 static u32
__compute_runnable_contrib(u64 n
)
2632 if (likely(n
<= LOAD_AVG_PERIOD
))
2633 return runnable_avg_yN_sum
[n
];
2634 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2635 return LOAD_AVG_MAX
;
2637 /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */
2638 contrib
= __accumulated_sum_N32
[n
/LOAD_AVG_PERIOD
];
2639 n
%= LOAD_AVG_PERIOD
;
2640 contrib
= decay_load(contrib
, n
);
2641 return contrib
+ runnable_avg_yN_sum
[n
];
2644 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2647 * We can represent the historical contribution to runnable average as the
2648 * coefficients of a geometric series. To do this we sub-divide our runnable
2649 * history into segments of approximately 1ms (1024us); label the segment that
2650 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2652 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2654 * (now) (~1ms ago) (~2ms ago)
2656 * Let u_i denote the fraction of p_i that the entity was runnable.
2658 * We then designate the fractions u_i as our co-efficients, yielding the
2659 * following representation of historical load:
2660 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2662 * We choose y based on the with of a reasonably scheduling period, fixing:
2665 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2666 * approximately half as much as the contribution to load within the last ms
2669 * When a period "rolls over" and we have new u_0`, multiplying the previous
2670 * sum again by y is sufficient to update:
2671 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2672 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2674 static __always_inline
int
2675 __update_load_avg(u64 now
, int cpu
, struct sched_avg
*sa
,
2676 unsigned long weight
, int running
, struct cfs_rq
*cfs_rq
)
2678 u64 delta
, scaled_delta
, periods
;
2680 unsigned int delta_w
, scaled_delta_w
, decayed
= 0;
2681 unsigned long scale_freq
, scale_cpu
;
2683 delta
= now
- sa
->last_update_time
;
2685 * This should only happen when time goes backwards, which it
2686 * unfortunately does during sched clock init when we swap over to TSC.
2688 if ((s64
)delta
< 0) {
2689 sa
->last_update_time
= now
;
2694 * Use 1024ns as the unit of measurement since it's a reasonable
2695 * approximation of 1us and fast to compute.
2700 sa
->last_update_time
= now
;
2702 scale_freq
= arch_scale_freq_capacity(NULL
, cpu
);
2703 scale_cpu
= arch_scale_cpu_capacity(NULL
, cpu
);
2705 /* delta_w is the amount already accumulated against our next period */
2706 delta_w
= sa
->period_contrib
;
2707 if (delta
+ delta_w
>= 1024) {
2710 /* how much left for next period will start over, we don't know yet */
2711 sa
->period_contrib
= 0;
2714 * Now that we know we're crossing a period boundary, figure
2715 * out how much from delta we need to complete the current
2716 * period and accrue it.
2718 delta_w
= 1024 - delta_w
;
2719 scaled_delta_w
= cap_scale(delta_w
, scale_freq
);
2721 sa
->load_sum
+= weight
* scaled_delta_w
;
2723 cfs_rq
->runnable_load_sum
+=
2724 weight
* scaled_delta_w
;
2728 sa
->util_sum
+= scaled_delta_w
* scale_cpu
;
2732 /* Figure out how many additional periods this update spans */
2733 periods
= delta
/ 1024;
2736 sa
->load_sum
= decay_load(sa
->load_sum
, periods
+ 1);
2738 cfs_rq
->runnable_load_sum
=
2739 decay_load(cfs_rq
->runnable_load_sum
, periods
+ 1);
2741 sa
->util_sum
= decay_load((u64
)(sa
->util_sum
), periods
+ 1);
2743 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2744 contrib
= __compute_runnable_contrib(periods
);
2745 contrib
= cap_scale(contrib
, scale_freq
);
2747 sa
->load_sum
+= weight
* contrib
;
2749 cfs_rq
->runnable_load_sum
+= weight
* contrib
;
2752 sa
->util_sum
+= contrib
* scale_cpu
;
2755 /* Remainder of delta accrued against u_0` */
2756 scaled_delta
= cap_scale(delta
, scale_freq
);
2758 sa
->load_sum
+= weight
* scaled_delta
;
2760 cfs_rq
->runnable_load_sum
+= weight
* scaled_delta
;
2763 sa
->util_sum
+= scaled_delta
* scale_cpu
;
2765 sa
->period_contrib
+= delta
;
2768 sa
->load_avg
= div_u64(sa
->load_sum
, LOAD_AVG_MAX
);
2770 cfs_rq
->runnable_load_avg
=
2771 div_u64(cfs_rq
->runnable_load_sum
, LOAD_AVG_MAX
);
2773 sa
->util_avg
= sa
->util_sum
/ LOAD_AVG_MAX
;
2779 #ifdef CONFIG_FAIR_GROUP_SCHED
2781 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2782 * and effective_load (which is not done because it is too costly).
2784 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
)
2786 long delta
= cfs_rq
->avg
.load_avg
- cfs_rq
->tg_load_avg_contrib
;
2789 * No need to update load_avg for root_task_group as it is not used.
2791 if (cfs_rq
->tg
== &root_task_group
)
2794 if (force
|| abs(delta
) > cfs_rq
->tg_load_avg_contrib
/ 64) {
2795 atomic_long_add(delta
, &cfs_rq
->tg
->load_avg
);
2796 cfs_rq
->tg_load_avg_contrib
= cfs_rq
->avg
.load_avg
;
2801 * Called within set_task_rq() right before setting a task's cpu. The
2802 * caller only guarantees p->pi_lock is held; no other assumptions,
2803 * including the state of rq->lock, should be made.
2805 void set_task_rq_fair(struct sched_entity
*se
,
2806 struct cfs_rq
*prev
, struct cfs_rq
*next
)
2808 if (!sched_feat(ATTACH_AGE_LOAD
))
2812 * We are supposed to update the task to "current" time, then its up to
2813 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2814 * getting what current time is, so simply throw away the out-of-date
2815 * time. This will result in the wakee task is less decayed, but giving
2816 * the wakee more load sounds not bad.
2818 if (se
->avg
.last_update_time
&& prev
) {
2819 u64 p_last_update_time
;
2820 u64 n_last_update_time
;
2822 #ifndef CONFIG_64BIT
2823 u64 p_last_update_time_copy
;
2824 u64 n_last_update_time_copy
;
2827 p_last_update_time_copy
= prev
->load_last_update_time_copy
;
2828 n_last_update_time_copy
= next
->load_last_update_time_copy
;
2832 p_last_update_time
= prev
->avg
.last_update_time
;
2833 n_last_update_time
= next
->avg
.last_update_time
;
2835 } while (p_last_update_time
!= p_last_update_time_copy
||
2836 n_last_update_time
!= n_last_update_time_copy
);
2838 p_last_update_time
= prev
->avg
.last_update_time
;
2839 n_last_update_time
= next
->avg
.last_update_time
;
2841 __update_load_avg(p_last_update_time
, cpu_of(rq_of(prev
)),
2842 &se
->avg
, 0, 0, NULL
);
2843 se
->avg
.last_update_time
= n_last_update_time
;
2846 #else /* CONFIG_FAIR_GROUP_SCHED */
2847 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
) {}
2848 #endif /* CONFIG_FAIR_GROUP_SCHED */
2850 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2852 static inline void cfs_rq_util_change(struct cfs_rq
*cfs_rq
)
2854 struct rq
*rq
= rq_of(cfs_rq
);
2855 int cpu
= cpu_of(rq
);
2857 if (cpu
== smp_processor_id() && &rq
->cfs
== cfs_rq
) {
2858 unsigned long max
= rq
->cpu_capacity_orig
;
2861 * There are a few boundary cases this might miss but it should
2862 * get called often enough that that should (hopefully) not be
2863 * a real problem -- added to that it only calls on the local
2864 * CPU, so if we enqueue remotely we'll miss an update, but
2865 * the next tick/schedule should update.
2867 * It will not get called when we go idle, because the idle
2868 * thread is a different class (!fair), nor will the utilization
2869 * number include things like RT tasks.
2871 * As is, the util number is not freq-invariant (we'd have to
2872 * implement arch_scale_freq_capacity() for that).
2876 cpufreq_update_util(rq_clock(rq
),
2877 min(cfs_rq
->avg
.util_avg
, max
), max
);
2881 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2883 update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
, bool update_freq
)
2885 struct sched_avg
*sa
= &cfs_rq
->avg
;
2886 int decayed
, removed_load
= 0, removed_util
= 0;
2888 if (atomic_long_read(&cfs_rq
->removed_load_avg
)) {
2889 s64 r
= atomic_long_xchg(&cfs_rq
->removed_load_avg
, 0);
2890 sa
->load_avg
= max_t(long, sa
->load_avg
- r
, 0);
2891 sa
->load_sum
= max_t(s64
, sa
->load_sum
- r
* LOAD_AVG_MAX
, 0);
2895 if (atomic_long_read(&cfs_rq
->removed_util_avg
)) {
2896 long r
= atomic_long_xchg(&cfs_rq
->removed_util_avg
, 0);
2897 sa
->util_avg
= max_t(long, sa
->util_avg
- r
, 0);
2898 sa
->util_sum
= max_t(s32
, sa
->util_sum
- r
* LOAD_AVG_MAX
, 0);
2902 decayed
= __update_load_avg(now
, cpu_of(rq_of(cfs_rq
)), sa
,
2903 scale_load_down(cfs_rq
->load
.weight
), cfs_rq
->curr
!= NULL
, cfs_rq
);
2905 #ifndef CONFIG_64BIT
2907 cfs_rq
->load_last_update_time_copy
= sa
->last_update_time
;
2910 if (update_freq
&& (decayed
|| removed_util
))
2911 cfs_rq_util_change(cfs_rq
);
2913 return decayed
|| removed_load
;
2916 /* Update task and its cfs_rq load average */
2917 static inline void update_load_avg(struct sched_entity
*se
, int update_tg
)
2919 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2920 u64 now
= cfs_rq_clock_task(cfs_rq
);
2921 struct rq
*rq
= rq_of(cfs_rq
);
2922 int cpu
= cpu_of(rq
);
2925 * Track task load average for carrying it to new CPU after migrated, and
2926 * track group sched_entity load average for task_h_load calc in migration
2928 __update_load_avg(now
, cpu
, &se
->avg
,
2929 se
->on_rq
* scale_load_down(se
->load
.weight
),
2930 cfs_rq
->curr
== se
, NULL
);
2932 if (update_cfs_rq_load_avg(now
, cfs_rq
, true) && update_tg
)
2933 update_tg_load_avg(cfs_rq
, 0);
2936 static void attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2938 if (!sched_feat(ATTACH_AGE_LOAD
))
2942 * If we got migrated (either between CPUs or between cgroups) we'll
2943 * have aged the average right before clearing @last_update_time.
2945 if (se
->avg
.last_update_time
) {
2946 __update_load_avg(cfs_rq
->avg
.last_update_time
, cpu_of(rq_of(cfs_rq
)),
2947 &se
->avg
, 0, 0, NULL
);
2950 * XXX: we could have just aged the entire load away if we've been
2951 * absent from the fair class for too long.
2956 se
->avg
.last_update_time
= cfs_rq
->avg
.last_update_time
;
2957 cfs_rq
->avg
.load_avg
+= se
->avg
.load_avg
;
2958 cfs_rq
->avg
.load_sum
+= se
->avg
.load_sum
;
2959 cfs_rq
->avg
.util_avg
+= se
->avg
.util_avg
;
2960 cfs_rq
->avg
.util_sum
+= se
->avg
.util_sum
;
2962 cfs_rq_util_change(cfs_rq
);
2965 static void detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2967 __update_load_avg(cfs_rq
->avg
.last_update_time
, cpu_of(rq_of(cfs_rq
)),
2968 &se
->avg
, se
->on_rq
* scale_load_down(se
->load
.weight
),
2969 cfs_rq
->curr
== se
, NULL
);
2971 cfs_rq
->avg
.load_avg
= max_t(long, cfs_rq
->avg
.load_avg
- se
->avg
.load_avg
, 0);
2972 cfs_rq
->avg
.load_sum
= max_t(s64
, cfs_rq
->avg
.load_sum
- se
->avg
.load_sum
, 0);
2973 cfs_rq
->avg
.util_avg
= max_t(long, cfs_rq
->avg
.util_avg
- se
->avg
.util_avg
, 0);
2974 cfs_rq
->avg
.util_sum
= max_t(s32
, cfs_rq
->avg
.util_sum
- se
->avg
.util_sum
, 0);
2976 cfs_rq_util_change(cfs_rq
);
2979 /* Add the load generated by se into cfs_rq's load average */
2981 enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2983 struct sched_avg
*sa
= &se
->avg
;
2984 u64 now
= cfs_rq_clock_task(cfs_rq
);
2985 int migrated
, decayed
;
2987 migrated
= !sa
->last_update_time
;
2989 __update_load_avg(now
, cpu_of(rq_of(cfs_rq
)), sa
,
2990 se
->on_rq
* scale_load_down(se
->load
.weight
),
2991 cfs_rq
->curr
== se
, NULL
);
2994 decayed
= update_cfs_rq_load_avg(now
, cfs_rq
, !migrated
);
2996 cfs_rq
->runnable_load_avg
+= sa
->load_avg
;
2997 cfs_rq
->runnable_load_sum
+= sa
->load_sum
;
3000 attach_entity_load_avg(cfs_rq
, se
);
3002 if (decayed
|| migrated
)
3003 update_tg_load_avg(cfs_rq
, 0);
3006 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3008 dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3010 update_load_avg(se
, 1);
3012 cfs_rq
->runnable_load_avg
=
3013 max_t(long, cfs_rq
->runnable_load_avg
- se
->avg
.load_avg
, 0);
3014 cfs_rq
->runnable_load_sum
=
3015 max_t(s64
, cfs_rq
->runnable_load_sum
- se
->avg
.load_sum
, 0);
3018 #ifndef CONFIG_64BIT
3019 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3021 u64 last_update_time_copy
;
3022 u64 last_update_time
;
3025 last_update_time_copy
= cfs_rq
->load_last_update_time_copy
;
3027 last_update_time
= cfs_rq
->avg
.last_update_time
;
3028 } while (last_update_time
!= last_update_time_copy
);
3030 return last_update_time
;
3033 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
3035 return cfs_rq
->avg
.last_update_time
;
3040 * Task first catches up with cfs_rq, and then subtract
3041 * itself from the cfs_rq (task must be off the queue now).
3043 void remove_entity_load_avg(struct sched_entity
*se
)
3045 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3046 u64 last_update_time
;
3049 * Newly created task or never used group entity should not be removed
3050 * from its (source) cfs_rq
3052 if (se
->avg
.last_update_time
== 0)
3055 last_update_time
= cfs_rq_last_update_time(cfs_rq
);
3057 __update_load_avg(last_update_time
, cpu_of(rq_of(cfs_rq
)), &se
->avg
, 0, 0, NULL
);
3058 atomic_long_add(se
->avg
.load_avg
, &cfs_rq
->removed_load_avg
);
3059 atomic_long_add(se
->avg
.util_avg
, &cfs_rq
->removed_util_avg
);
3062 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq
*cfs_rq
)
3064 return cfs_rq
->runnable_load_avg
;
3067 static inline unsigned long cfs_rq_load_avg(struct cfs_rq
*cfs_rq
)
3069 return cfs_rq
->avg
.load_avg
;
3072 static int idle_balance(struct rq
*this_rq
);
3074 #else /* CONFIG_SMP */
3076 static inline void update_load_avg(struct sched_entity
*se
, int not_used
)
3078 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3079 struct rq
*rq
= rq_of(cfs_rq
);
3081 cpufreq_trigger_update(rq_clock(rq
));
3085 enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3087 dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3088 static inline void remove_entity_load_avg(struct sched_entity
*se
) {}
3091 attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3093 detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3095 static inline int idle_balance(struct rq
*rq
)
3100 #endif /* CONFIG_SMP */
3102 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3104 #ifdef CONFIG_SCHEDSTATS
3105 struct task_struct
*tsk
= NULL
;
3107 if (entity_is_task(se
))
3110 if (se
->statistics
.sleep_start
) {
3111 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
3116 if (unlikely(delta
> se
->statistics
.sleep_max
))
3117 se
->statistics
.sleep_max
= delta
;
3119 se
->statistics
.sleep_start
= 0;
3120 se
->statistics
.sum_sleep_runtime
+= delta
;
3123 account_scheduler_latency(tsk
, delta
>> 10, 1);
3124 trace_sched_stat_sleep(tsk
, delta
);
3127 if (se
->statistics
.block_start
) {
3128 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
3133 if (unlikely(delta
> se
->statistics
.block_max
))
3134 se
->statistics
.block_max
= delta
;
3136 se
->statistics
.block_start
= 0;
3137 se
->statistics
.sum_sleep_runtime
+= delta
;
3140 if (tsk
->in_iowait
) {
3141 se
->statistics
.iowait_sum
+= delta
;
3142 se
->statistics
.iowait_count
++;
3143 trace_sched_stat_iowait(tsk
, delta
);
3146 trace_sched_stat_blocked(tsk
, delta
);
3149 * Blocking time is in units of nanosecs, so shift by
3150 * 20 to get a milliseconds-range estimation of the
3151 * amount of time that the task spent sleeping:
3153 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
3154 profile_hits(SLEEP_PROFILING
,
3155 (void *)get_wchan(tsk
),
3158 account_scheduler_latency(tsk
, delta
>> 10, 0);
3164 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3166 #ifdef CONFIG_SCHED_DEBUG
3167 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
3172 if (d
> 3*sysctl_sched_latency
)
3173 schedstat_inc(cfs_rq
, nr_spread_over
);
3178 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
3180 u64 vruntime
= cfs_rq
->min_vruntime
;
3183 * The 'current' period is already promised to the current tasks,
3184 * however the extra weight of the new task will slow them down a
3185 * little, place the new task so that it fits in the slot that
3186 * stays open at the end.
3188 if (initial
&& sched_feat(START_DEBIT
))
3189 vruntime
+= sched_vslice(cfs_rq
, se
);
3191 /* sleeps up to a single latency don't count. */
3193 unsigned long thresh
= sysctl_sched_latency
;
3196 * Halve their sleep time's effect, to allow
3197 * for a gentler effect of sleepers:
3199 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
3205 /* ensure we never gain time by being placed backwards. */
3206 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
3209 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
3211 static inline void check_schedstat_required(void)
3213 #ifdef CONFIG_SCHEDSTATS
3214 if (schedstat_enabled())
3217 /* Force schedstat enabled if a dependent tracepoint is active */
3218 if (trace_sched_stat_wait_enabled() ||
3219 trace_sched_stat_sleep_enabled() ||
3220 trace_sched_stat_iowait_enabled() ||
3221 trace_sched_stat_blocked_enabled() ||
3222 trace_sched_stat_runtime_enabled()) {
3223 pr_warn_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3224 "stat_blocked and stat_runtime require the "
3225 "kernel parameter schedstats=enabled or "
3226 "kernel.sched_schedstats=1\n");
3237 * update_min_vruntime()
3238 * vruntime -= min_vruntime
3242 * update_min_vruntime()
3243 * vruntime += min_vruntime
3245 * this way the vruntime transition between RQs is done when both
3246 * min_vruntime are up-to-date.
3250 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3251 * vruntime -= min_vruntime
3255 * update_min_vruntime()
3256 * vruntime += min_vruntime
3258 * this way we don't have the most up-to-date min_vruntime on the originating
3259 * CPU and an up-to-date min_vruntime on the destination CPU.
3263 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3265 bool renorm
= !(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_MIGRATED
);
3266 bool curr
= cfs_rq
->curr
== se
;
3269 * If we're the current task, we must renormalise before calling
3273 se
->vruntime
+= cfs_rq
->min_vruntime
;
3275 update_curr(cfs_rq
);
3278 * Otherwise, renormalise after, such that we're placed at the current
3279 * moment in time, instead of some random moment in the past. Being
3280 * placed in the past could significantly boost this task to the
3281 * fairness detriment of existing tasks.
3283 if (renorm
&& !curr
)
3284 se
->vruntime
+= cfs_rq
->min_vruntime
;
3286 enqueue_entity_load_avg(cfs_rq
, se
);
3287 account_entity_enqueue(cfs_rq
, se
);
3288 update_cfs_shares(cfs_rq
);
3290 if (flags
& ENQUEUE_WAKEUP
) {
3291 place_entity(cfs_rq
, se
, 0);
3292 if (schedstat_enabled())
3293 enqueue_sleeper(cfs_rq
, se
);
3296 check_schedstat_required();
3297 if (schedstat_enabled()) {
3298 update_stats_enqueue(cfs_rq
, se
);
3299 check_spread(cfs_rq
, se
);
3302 __enqueue_entity(cfs_rq
, se
);
3305 if (cfs_rq
->nr_running
== 1) {
3306 list_add_leaf_cfs_rq(cfs_rq
);
3307 check_enqueue_throttle(cfs_rq
);
3311 static void __clear_buddies_last(struct sched_entity
*se
)
3313 for_each_sched_entity(se
) {
3314 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3315 if (cfs_rq
->last
!= se
)
3318 cfs_rq
->last
= NULL
;
3322 static void __clear_buddies_next(struct sched_entity
*se
)
3324 for_each_sched_entity(se
) {
3325 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3326 if (cfs_rq
->next
!= se
)
3329 cfs_rq
->next
= NULL
;
3333 static void __clear_buddies_skip(struct sched_entity
*se
)
3335 for_each_sched_entity(se
) {
3336 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3337 if (cfs_rq
->skip
!= se
)
3340 cfs_rq
->skip
= NULL
;
3344 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3346 if (cfs_rq
->last
== se
)
3347 __clear_buddies_last(se
);
3349 if (cfs_rq
->next
== se
)
3350 __clear_buddies_next(se
);
3352 if (cfs_rq
->skip
== se
)
3353 __clear_buddies_skip(se
);
3356 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3359 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3362 * Update run-time statistics of the 'current'.
3364 update_curr(cfs_rq
);
3365 dequeue_entity_load_avg(cfs_rq
, se
);
3367 if (schedstat_enabled())
3368 update_stats_dequeue(cfs_rq
, se
, flags
);
3370 clear_buddies(cfs_rq
, se
);
3372 if (se
!= cfs_rq
->curr
)
3373 __dequeue_entity(cfs_rq
, se
);
3375 account_entity_dequeue(cfs_rq
, se
);
3378 * Normalize the entity after updating the min_vruntime because the
3379 * update can refer to the ->curr item and we need to reflect this
3380 * movement in our normalized position.
3382 if (!(flags
& DEQUEUE_SLEEP
))
3383 se
->vruntime
-= cfs_rq
->min_vruntime
;
3385 /* return excess runtime on last dequeue */
3386 return_cfs_rq_runtime(cfs_rq
);
3388 update_min_vruntime(cfs_rq
);
3389 update_cfs_shares(cfs_rq
);
3393 * Preempt the current task with a newly woken task if needed:
3396 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3398 unsigned long ideal_runtime
, delta_exec
;
3399 struct sched_entity
*se
;
3402 ideal_runtime
= sched_slice(cfs_rq
, curr
);
3403 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
3404 if (delta_exec
> ideal_runtime
) {
3405 resched_curr(rq_of(cfs_rq
));
3407 * The current task ran long enough, ensure it doesn't get
3408 * re-elected due to buddy favours.
3410 clear_buddies(cfs_rq
, curr
);
3415 * Ensure that a task that missed wakeup preemption by a
3416 * narrow margin doesn't have to wait for a full slice.
3417 * This also mitigates buddy induced latencies under load.
3419 if (delta_exec
< sysctl_sched_min_granularity
)
3422 se
= __pick_first_entity(cfs_rq
);
3423 delta
= curr
->vruntime
- se
->vruntime
;
3428 if (delta
> ideal_runtime
)
3429 resched_curr(rq_of(cfs_rq
));
3433 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3435 /* 'current' is not kept within the tree. */
3438 * Any task has to be enqueued before it get to execute on
3439 * a CPU. So account for the time it spent waiting on the
3442 if (schedstat_enabled())
3443 update_stats_wait_end(cfs_rq
, se
);
3444 __dequeue_entity(cfs_rq
, se
);
3445 update_load_avg(se
, 1);
3448 update_stats_curr_start(cfs_rq
, se
);
3450 #ifdef CONFIG_SCHEDSTATS
3452 * Track our maximum slice length, if the CPU's load is at
3453 * least twice that of our own weight (i.e. dont track it
3454 * when there are only lesser-weight tasks around):
3456 if (schedstat_enabled() && rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
3457 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
3458 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
3461 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
3465 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
3468 * Pick the next process, keeping these things in mind, in this order:
3469 * 1) keep things fair between processes/task groups
3470 * 2) pick the "next" process, since someone really wants that to run
3471 * 3) pick the "last" process, for cache locality
3472 * 4) do not run the "skip" process, if something else is available
3474 static struct sched_entity
*
3475 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3477 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
3478 struct sched_entity
*se
;
3481 * If curr is set we have to see if its left of the leftmost entity
3482 * still in the tree, provided there was anything in the tree at all.
3484 if (!left
|| (curr
&& entity_before(curr
, left
)))
3487 se
= left
; /* ideally we run the leftmost entity */
3490 * Avoid running the skip buddy, if running something else can
3491 * be done without getting too unfair.
3493 if (cfs_rq
->skip
== se
) {
3494 struct sched_entity
*second
;
3497 second
= __pick_first_entity(cfs_rq
);
3499 second
= __pick_next_entity(se
);
3500 if (!second
|| (curr
&& entity_before(curr
, second
)))
3504 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3509 * Prefer last buddy, try to return the CPU to a preempted task.
3511 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3515 * Someone really wants this to run. If it's not unfair, run it.
3517 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3520 clear_buddies(cfs_rq
, se
);
3525 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3527 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3530 * If still on the runqueue then deactivate_task()
3531 * was not called and update_curr() has to be done:
3534 update_curr(cfs_rq
);
3536 /* throttle cfs_rqs exceeding runtime */
3537 check_cfs_rq_runtime(cfs_rq
);
3539 if (schedstat_enabled()) {
3540 check_spread(cfs_rq
, prev
);
3542 update_stats_wait_start(cfs_rq
, prev
);
3546 /* Put 'current' back into the tree. */
3547 __enqueue_entity(cfs_rq
, prev
);
3548 /* in !on_rq case, update occurred at dequeue */
3549 update_load_avg(prev
, 0);
3551 cfs_rq
->curr
= NULL
;
3555 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3558 * Update run-time statistics of the 'current'.
3560 update_curr(cfs_rq
);
3563 * Ensure that runnable average is periodically updated.
3565 update_load_avg(curr
, 1);
3566 update_cfs_shares(cfs_rq
);
3568 #ifdef CONFIG_SCHED_HRTICK
3570 * queued ticks are scheduled to match the slice, so don't bother
3571 * validating it and just reschedule.
3574 resched_curr(rq_of(cfs_rq
));
3578 * don't let the period tick interfere with the hrtick preemption
3580 if (!sched_feat(DOUBLE_TICK
) &&
3581 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3585 if (cfs_rq
->nr_running
> 1)
3586 check_preempt_tick(cfs_rq
, curr
);
3590 /**************************************************
3591 * CFS bandwidth control machinery
3594 #ifdef CONFIG_CFS_BANDWIDTH
3596 #ifdef HAVE_JUMP_LABEL
3597 static struct static_key __cfs_bandwidth_used
;
3599 static inline bool cfs_bandwidth_used(void)
3601 return static_key_false(&__cfs_bandwidth_used
);
3604 void cfs_bandwidth_usage_inc(void)
3606 static_key_slow_inc(&__cfs_bandwidth_used
);
3609 void cfs_bandwidth_usage_dec(void)
3611 static_key_slow_dec(&__cfs_bandwidth_used
);
3613 #else /* HAVE_JUMP_LABEL */
3614 static bool cfs_bandwidth_used(void)
3619 void cfs_bandwidth_usage_inc(void) {}
3620 void cfs_bandwidth_usage_dec(void) {}
3621 #endif /* HAVE_JUMP_LABEL */
3624 * default period for cfs group bandwidth.
3625 * default: 0.1s, units: nanoseconds
3627 static inline u64
default_cfs_period(void)
3629 return 100000000ULL;
3632 static inline u64
sched_cfs_bandwidth_slice(void)
3634 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3638 * Replenish runtime according to assigned quota and update expiration time.
3639 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3640 * additional synchronization around rq->lock.
3642 * requires cfs_b->lock
3644 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3648 if (cfs_b
->quota
== RUNTIME_INF
)
3651 now
= sched_clock_cpu(smp_processor_id());
3652 cfs_b
->runtime
= cfs_b
->quota
;
3653 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3656 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3658 return &tg
->cfs_bandwidth
;
3661 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3662 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3664 if (unlikely(cfs_rq
->throttle_count
))
3665 return cfs_rq
->throttled_clock_task
- cfs_rq
->throttled_clock_task_time
;
3667 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3670 /* returns 0 on failure to allocate runtime */
3671 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3673 struct task_group
*tg
= cfs_rq
->tg
;
3674 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3675 u64 amount
= 0, min_amount
, expires
;
3677 /* note: this is a positive sum as runtime_remaining <= 0 */
3678 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3680 raw_spin_lock(&cfs_b
->lock
);
3681 if (cfs_b
->quota
== RUNTIME_INF
)
3682 amount
= min_amount
;
3684 start_cfs_bandwidth(cfs_b
);
3686 if (cfs_b
->runtime
> 0) {
3687 amount
= min(cfs_b
->runtime
, min_amount
);
3688 cfs_b
->runtime
-= amount
;
3692 expires
= cfs_b
->runtime_expires
;
3693 raw_spin_unlock(&cfs_b
->lock
);
3695 cfs_rq
->runtime_remaining
+= amount
;
3697 * we may have advanced our local expiration to account for allowed
3698 * spread between our sched_clock and the one on which runtime was
3701 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3702 cfs_rq
->runtime_expires
= expires
;
3704 return cfs_rq
->runtime_remaining
> 0;
3708 * Note: This depends on the synchronization provided by sched_clock and the
3709 * fact that rq->clock snapshots this value.
3711 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3713 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3715 /* if the deadline is ahead of our clock, nothing to do */
3716 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3719 if (cfs_rq
->runtime_remaining
< 0)
3723 * If the local deadline has passed we have to consider the
3724 * possibility that our sched_clock is 'fast' and the global deadline
3725 * has not truly expired.
3727 * Fortunately we can check determine whether this the case by checking
3728 * whether the global deadline has advanced. It is valid to compare
3729 * cfs_b->runtime_expires without any locks since we only care about
3730 * exact equality, so a partial write will still work.
3733 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
3734 /* extend local deadline, drift is bounded above by 2 ticks */
3735 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3737 /* global deadline is ahead, expiration has passed */
3738 cfs_rq
->runtime_remaining
= 0;
3742 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3744 /* dock delta_exec before expiring quota (as it could span periods) */
3745 cfs_rq
->runtime_remaining
-= delta_exec
;
3746 expire_cfs_rq_runtime(cfs_rq
);
3748 if (likely(cfs_rq
->runtime_remaining
> 0))
3752 * if we're unable to extend our runtime we resched so that the active
3753 * hierarchy can be throttled
3755 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3756 resched_curr(rq_of(cfs_rq
));
3759 static __always_inline
3760 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3762 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3765 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3768 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3770 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3773 /* check whether cfs_rq, or any parent, is throttled */
3774 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3776 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3780 * Ensure that neither of the group entities corresponding to src_cpu or
3781 * dest_cpu are members of a throttled hierarchy when performing group
3782 * load-balance operations.
3784 static inline int throttled_lb_pair(struct task_group
*tg
,
3785 int src_cpu
, int dest_cpu
)
3787 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3789 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3790 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3792 return throttled_hierarchy(src_cfs_rq
) ||
3793 throttled_hierarchy(dest_cfs_rq
);
3796 /* updated child weight may affect parent so we have to do this bottom up */
3797 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3799 struct rq
*rq
= data
;
3800 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3802 cfs_rq
->throttle_count
--;
3803 if (!cfs_rq
->throttle_count
) {
3804 /* adjust cfs_rq_clock_task() */
3805 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3806 cfs_rq
->throttled_clock_task
;
3812 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3814 struct rq
*rq
= data
;
3815 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3817 /* group is entering throttled state, stop time */
3818 if (!cfs_rq
->throttle_count
)
3819 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3820 cfs_rq
->throttle_count
++;
3825 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3827 struct rq
*rq
= rq_of(cfs_rq
);
3828 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3829 struct sched_entity
*se
;
3830 long task_delta
, dequeue
= 1;
3833 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3835 /* freeze hierarchy runnable averages while throttled */
3837 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3840 task_delta
= cfs_rq
->h_nr_running
;
3841 for_each_sched_entity(se
) {
3842 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3843 /* throttled entity or throttle-on-deactivate */
3848 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3849 qcfs_rq
->h_nr_running
-= task_delta
;
3851 if (qcfs_rq
->load
.weight
)
3856 sub_nr_running(rq
, task_delta
);
3858 cfs_rq
->throttled
= 1;
3859 cfs_rq
->throttled_clock
= rq_clock(rq
);
3860 raw_spin_lock(&cfs_b
->lock
);
3861 empty
= list_empty(&cfs_b
->throttled_cfs_rq
);
3864 * Add to the _head_ of the list, so that an already-started
3865 * distribute_cfs_runtime will not see us
3867 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3870 * If we're the first throttled task, make sure the bandwidth
3874 start_cfs_bandwidth(cfs_b
);
3876 raw_spin_unlock(&cfs_b
->lock
);
3879 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3881 struct rq
*rq
= rq_of(cfs_rq
);
3882 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3883 struct sched_entity
*se
;
3887 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3889 cfs_rq
->throttled
= 0;
3891 update_rq_clock(rq
);
3893 raw_spin_lock(&cfs_b
->lock
);
3894 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3895 list_del_rcu(&cfs_rq
->throttled_list
);
3896 raw_spin_unlock(&cfs_b
->lock
);
3898 /* update hierarchical throttle state */
3899 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3901 if (!cfs_rq
->load
.weight
)
3904 task_delta
= cfs_rq
->h_nr_running
;
3905 for_each_sched_entity(se
) {
3909 cfs_rq
= cfs_rq_of(se
);
3911 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3912 cfs_rq
->h_nr_running
+= task_delta
;
3914 if (cfs_rq_throttled(cfs_rq
))
3919 add_nr_running(rq
, task_delta
);
3921 /* determine whether we need to wake up potentially idle cpu */
3922 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3926 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3927 u64 remaining
, u64 expires
)
3929 struct cfs_rq
*cfs_rq
;
3931 u64 starting_runtime
= remaining
;
3934 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3936 struct rq
*rq
= rq_of(cfs_rq
);
3938 raw_spin_lock(&rq
->lock
);
3939 if (!cfs_rq_throttled(cfs_rq
))
3942 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3943 if (runtime
> remaining
)
3944 runtime
= remaining
;
3945 remaining
-= runtime
;
3947 cfs_rq
->runtime_remaining
+= runtime
;
3948 cfs_rq
->runtime_expires
= expires
;
3950 /* we check whether we're throttled above */
3951 if (cfs_rq
->runtime_remaining
> 0)
3952 unthrottle_cfs_rq(cfs_rq
);
3955 raw_spin_unlock(&rq
->lock
);
3962 return starting_runtime
- remaining
;
3966 * Responsible for refilling a task_group's bandwidth and unthrottling its
3967 * cfs_rqs as appropriate. If there has been no activity within the last
3968 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3969 * used to track this state.
3971 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3973 u64 runtime
, runtime_expires
;
3976 /* no need to continue the timer with no bandwidth constraint */
3977 if (cfs_b
->quota
== RUNTIME_INF
)
3978 goto out_deactivate
;
3980 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3981 cfs_b
->nr_periods
+= overrun
;
3984 * idle depends on !throttled (for the case of a large deficit), and if
3985 * we're going inactive then everything else can be deferred
3987 if (cfs_b
->idle
&& !throttled
)
3988 goto out_deactivate
;
3990 __refill_cfs_bandwidth_runtime(cfs_b
);
3993 /* mark as potentially idle for the upcoming period */
3998 /* account preceding periods in which throttling occurred */
3999 cfs_b
->nr_throttled
+= overrun
;
4001 runtime_expires
= cfs_b
->runtime_expires
;
4004 * This check is repeated as we are holding onto the new bandwidth while
4005 * we unthrottle. This can potentially race with an unthrottled group
4006 * trying to acquire new bandwidth from the global pool. This can result
4007 * in us over-using our runtime if it is all used during this loop, but
4008 * only by limited amounts in that extreme case.
4010 while (throttled
&& cfs_b
->runtime
> 0) {
4011 runtime
= cfs_b
->runtime
;
4012 raw_spin_unlock(&cfs_b
->lock
);
4013 /* we can't nest cfs_b->lock while distributing bandwidth */
4014 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
4016 raw_spin_lock(&cfs_b
->lock
);
4018 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
4020 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
4024 * While we are ensured activity in the period following an
4025 * unthrottle, this also covers the case in which the new bandwidth is
4026 * insufficient to cover the existing bandwidth deficit. (Forcing the
4027 * timer to remain active while there are any throttled entities.)
4037 /* a cfs_rq won't donate quota below this amount */
4038 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
4039 /* minimum remaining period time to redistribute slack quota */
4040 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
4041 /* how long we wait to gather additional slack before distributing */
4042 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
4045 * Are we near the end of the current quota period?
4047 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4048 * hrtimer base being cleared by hrtimer_start. In the case of
4049 * migrate_hrtimers, base is never cleared, so we are fine.
4051 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
4053 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
4056 /* if the call-back is running a quota refresh is already occurring */
4057 if (hrtimer_callback_running(refresh_timer
))
4060 /* is a quota refresh about to occur? */
4061 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
4062 if (remaining
< min_expire
)
4068 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
4070 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
4072 /* if there's a quota refresh soon don't bother with slack */
4073 if (runtime_refresh_within(cfs_b
, min_left
))
4076 hrtimer_start(&cfs_b
->slack_timer
,
4077 ns_to_ktime(cfs_bandwidth_slack_period
),
4081 /* we know any runtime found here is valid as update_curr() precedes return */
4082 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4084 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
4085 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
4087 if (slack_runtime
<= 0)
4090 raw_spin_lock(&cfs_b
->lock
);
4091 if (cfs_b
->quota
!= RUNTIME_INF
&&
4092 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
4093 cfs_b
->runtime
+= slack_runtime
;
4095 /* we are under rq->lock, defer unthrottling using a timer */
4096 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
4097 !list_empty(&cfs_b
->throttled_cfs_rq
))
4098 start_cfs_slack_bandwidth(cfs_b
);
4100 raw_spin_unlock(&cfs_b
->lock
);
4102 /* even if it's not valid for return we don't want to try again */
4103 cfs_rq
->runtime_remaining
-= slack_runtime
;
4106 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4108 if (!cfs_bandwidth_used())
4111 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
4114 __return_cfs_rq_runtime(cfs_rq
);
4118 * This is done with a timer (instead of inline with bandwidth return) since
4119 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4121 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
4123 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
4126 /* confirm we're still not at a refresh boundary */
4127 raw_spin_lock(&cfs_b
->lock
);
4128 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
4129 raw_spin_unlock(&cfs_b
->lock
);
4133 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
4134 runtime
= cfs_b
->runtime
;
4136 expires
= cfs_b
->runtime_expires
;
4137 raw_spin_unlock(&cfs_b
->lock
);
4142 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
4144 raw_spin_lock(&cfs_b
->lock
);
4145 if (expires
== cfs_b
->runtime_expires
)
4146 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
4147 raw_spin_unlock(&cfs_b
->lock
);
4151 * When a group wakes up we want to make sure that its quota is not already
4152 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4153 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4155 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
4157 if (!cfs_bandwidth_used())
4160 /* an active group must be handled by the update_curr()->put() path */
4161 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
4164 /* ensure the group is not already throttled */
4165 if (cfs_rq_throttled(cfs_rq
))
4168 /* update runtime allocation */
4169 account_cfs_rq_runtime(cfs_rq
, 0);
4170 if (cfs_rq
->runtime_remaining
<= 0)
4171 throttle_cfs_rq(cfs_rq
);
4174 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4175 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4177 if (!cfs_bandwidth_used())
4180 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
4184 * it's possible for a throttled entity to be forced into a running
4185 * state (e.g. set_curr_task), in this case we're finished.
4187 if (cfs_rq_throttled(cfs_rq
))
4190 throttle_cfs_rq(cfs_rq
);
4194 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
4196 struct cfs_bandwidth
*cfs_b
=
4197 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
4199 do_sched_cfs_slack_timer(cfs_b
);
4201 return HRTIMER_NORESTART
;
4204 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
4206 struct cfs_bandwidth
*cfs_b
=
4207 container_of(timer
, struct cfs_bandwidth
, period_timer
);
4211 raw_spin_lock(&cfs_b
->lock
);
4213 overrun
= hrtimer_forward_now(timer
, cfs_b
->period
);
4217 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
4220 cfs_b
->period_active
= 0;
4221 raw_spin_unlock(&cfs_b
->lock
);
4223 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
4226 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4228 raw_spin_lock_init(&cfs_b
->lock
);
4230 cfs_b
->quota
= RUNTIME_INF
;
4231 cfs_b
->period
= ns_to_ktime(default_cfs_period());
4233 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
4234 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_PINNED
);
4235 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
4236 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4237 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
4240 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4242 cfs_rq
->runtime_enabled
= 0;
4243 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
4246 void start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4248 lockdep_assert_held(&cfs_b
->lock
);
4250 if (!cfs_b
->period_active
) {
4251 cfs_b
->period_active
= 1;
4252 hrtimer_forward_now(&cfs_b
->period_timer
, cfs_b
->period
);
4253 hrtimer_start_expires(&cfs_b
->period_timer
, HRTIMER_MODE_ABS_PINNED
);
4257 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4259 /* init_cfs_bandwidth() was not called */
4260 if (!cfs_b
->throttled_cfs_rq
.next
)
4263 hrtimer_cancel(&cfs_b
->period_timer
);
4264 hrtimer_cancel(&cfs_b
->slack_timer
);
4267 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
4269 struct cfs_rq
*cfs_rq
;
4271 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4272 struct cfs_bandwidth
*cfs_b
= &cfs_rq
->tg
->cfs_bandwidth
;
4274 raw_spin_lock(&cfs_b
->lock
);
4275 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
4276 raw_spin_unlock(&cfs_b
->lock
);
4280 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
4282 struct cfs_rq
*cfs_rq
;
4284 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4285 if (!cfs_rq
->runtime_enabled
)
4289 * clock_task is not advancing so we just need to make sure
4290 * there's some valid quota amount
4292 cfs_rq
->runtime_remaining
= 1;
4294 * Offline rq is schedulable till cpu is completely disabled
4295 * in take_cpu_down(), so we prevent new cfs throttling here.
4297 cfs_rq
->runtime_enabled
= 0;
4299 if (cfs_rq_throttled(cfs_rq
))
4300 unthrottle_cfs_rq(cfs_rq
);
4304 #else /* CONFIG_CFS_BANDWIDTH */
4305 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
4307 return rq_clock_task(rq_of(cfs_rq
));
4310 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
4311 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
4312 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
4313 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4315 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4320 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4325 static inline int throttled_lb_pair(struct task_group
*tg
,
4326 int src_cpu
, int dest_cpu
)
4331 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4333 #ifdef CONFIG_FAIR_GROUP_SCHED
4334 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4337 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4341 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4342 static inline void update_runtime_enabled(struct rq
*rq
) {}
4343 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
4345 #endif /* CONFIG_CFS_BANDWIDTH */
4347 /**************************************************
4348 * CFS operations on tasks:
4351 #ifdef CONFIG_SCHED_HRTICK
4352 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4354 struct sched_entity
*se
= &p
->se
;
4355 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4357 WARN_ON(task_rq(p
) != rq
);
4359 if (cfs_rq
->nr_running
> 1) {
4360 u64 slice
= sched_slice(cfs_rq
, se
);
4361 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
4362 s64 delta
= slice
- ran
;
4369 hrtick_start(rq
, delta
);
4374 * called from enqueue/dequeue and updates the hrtick when the
4375 * current task is from our class and nr_running is low enough
4378 static void hrtick_update(struct rq
*rq
)
4380 struct task_struct
*curr
= rq
->curr
;
4382 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
4385 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
4386 hrtick_start_fair(rq
, curr
);
4388 #else /* !CONFIG_SCHED_HRTICK */
4390 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4394 static inline void hrtick_update(struct rq
*rq
)
4400 * The enqueue_task method is called before nr_running is
4401 * increased. Here we update the fair scheduling stats and
4402 * then put the task into the rbtree:
4405 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4407 struct cfs_rq
*cfs_rq
;
4408 struct sched_entity
*se
= &p
->se
;
4410 for_each_sched_entity(se
) {
4413 cfs_rq
= cfs_rq_of(se
);
4414 enqueue_entity(cfs_rq
, se
, flags
);
4417 * end evaluation on encountering a throttled cfs_rq
4419 * note: in the case of encountering a throttled cfs_rq we will
4420 * post the final h_nr_running increment below.
4422 if (cfs_rq_throttled(cfs_rq
))
4424 cfs_rq
->h_nr_running
++;
4426 flags
= ENQUEUE_WAKEUP
;
4429 for_each_sched_entity(se
) {
4430 cfs_rq
= cfs_rq_of(se
);
4431 cfs_rq
->h_nr_running
++;
4433 if (cfs_rq_throttled(cfs_rq
))
4436 update_load_avg(se
, 1);
4437 update_cfs_shares(cfs_rq
);
4441 add_nr_running(rq
, 1);
4446 static void set_next_buddy(struct sched_entity
*se
);
4449 * The dequeue_task method is called before nr_running is
4450 * decreased. We remove the task from the rbtree and
4451 * update the fair scheduling stats:
4453 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4455 struct cfs_rq
*cfs_rq
;
4456 struct sched_entity
*se
= &p
->se
;
4457 int task_sleep
= flags
& DEQUEUE_SLEEP
;
4459 for_each_sched_entity(se
) {
4460 cfs_rq
= cfs_rq_of(se
);
4461 dequeue_entity(cfs_rq
, se
, flags
);
4464 * end evaluation on encountering a throttled cfs_rq
4466 * note: in the case of encountering a throttled cfs_rq we will
4467 * post the final h_nr_running decrement below.
4469 if (cfs_rq_throttled(cfs_rq
))
4471 cfs_rq
->h_nr_running
--;
4473 /* Don't dequeue parent if it has other entities besides us */
4474 if (cfs_rq
->load
.weight
) {
4476 * Bias pick_next to pick a task from this cfs_rq, as
4477 * p is sleeping when it is within its sched_slice.
4479 if (task_sleep
&& parent_entity(se
))
4480 set_next_buddy(parent_entity(se
));
4482 /* avoid re-evaluating load for this entity */
4483 se
= parent_entity(se
);
4486 flags
|= DEQUEUE_SLEEP
;
4489 for_each_sched_entity(se
) {
4490 cfs_rq
= cfs_rq_of(se
);
4491 cfs_rq
->h_nr_running
--;
4493 if (cfs_rq_throttled(cfs_rq
))
4496 update_load_avg(se
, 1);
4497 update_cfs_shares(cfs_rq
);
4501 sub_nr_running(rq
, 1);
4507 #ifdef CONFIG_NO_HZ_COMMON
4509 * per rq 'load' arrray crap; XXX kill this.
4513 * The exact cpuload calculated at every tick would be:
4515 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4517 * If a cpu misses updates for n ticks (as it was idle) and update gets
4518 * called on the n+1-th tick when cpu may be busy, then we have:
4520 * load_n = (1 - 1/2^i)^n * load_0
4521 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4523 * decay_load_missed() below does efficient calculation of
4525 * load' = (1 - 1/2^i)^n * load
4527 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4528 * This allows us to precompute the above in said factors, thereby allowing the
4529 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4530 * fixed_power_int())
4532 * The calculation is approximated on a 128 point scale.
4534 #define DEGRADE_SHIFT 7
4536 static const u8 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
4537 static const u8 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
4538 { 0, 0, 0, 0, 0, 0, 0, 0 },
4539 { 64, 32, 8, 0, 0, 0, 0, 0 },
4540 { 96, 72, 40, 12, 1, 0, 0, 0 },
4541 { 112, 98, 75, 43, 15, 1, 0, 0 },
4542 { 120, 112, 98, 76, 45, 16, 2, 0 }
4546 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4547 * would be when CPU is idle and so we just decay the old load without
4548 * adding any new load.
4550 static unsigned long
4551 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
4555 if (!missed_updates
)
4558 if (missed_updates
>= degrade_zero_ticks
[idx
])
4562 return load
>> missed_updates
;
4564 while (missed_updates
) {
4565 if (missed_updates
% 2)
4566 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
4568 missed_updates
>>= 1;
4573 #endif /* CONFIG_NO_HZ_COMMON */
4576 * __cpu_load_update - update the rq->cpu_load[] statistics
4577 * @this_rq: The rq to update statistics for
4578 * @this_load: The current load
4579 * @pending_updates: The number of missed updates
4581 * Update rq->cpu_load[] statistics. This function is usually called every
4582 * scheduler tick (TICK_NSEC).
4584 * This function computes a decaying average:
4586 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4588 * Because of NOHZ it might not get called on every tick which gives need for
4589 * the @pending_updates argument.
4591 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4592 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4593 * = A * (A * load[i]_n-2 + B) + B
4594 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4595 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4596 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4597 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4598 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4600 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4601 * any change in load would have resulted in the tick being turned back on.
4603 * For regular NOHZ, this reduces to:
4605 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4607 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4610 static void cpu_load_update(struct rq
*this_rq
, unsigned long this_load
,
4611 unsigned long pending_updates
)
4613 unsigned long __maybe_unused tickless_load
= this_rq
->cpu_load
[0];
4616 this_rq
->nr_load_updates
++;
4618 /* Update our load: */
4619 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
4620 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
4621 unsigned long old_load
, new_load
;
4623 /* scale is effectively 1 << i now, and >> i divides by scale */
4625 old_load
= this_rq
->cpu_load
[i
];
4626 #ifdef CONFIG_NO_HZ_COMMON
4627 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
4628 if (tickless_load
) {
4629 old_load
-= decay_load_missed(tickless_load
, pending_updates
- 1, i
);
4631 * old_load can never be a negative value because a
4632 * decayed tickless_load cannot be greater than the
4633 * original tickless_load.
4635 old_load
+= tickless_load
;
4638 new_load
= this_load
;
4640 * Round up the averaging division if load is increasing. This
4641 * prevents us from getting stuck on 9 if the load is 10, for
4644 if (new_load
> old_load
)
4645 new_load
+= scale
- 1;
4647 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
4650 sched_avg_update(this_rq
);
4653 /* Used instead of source_load when we know the type == 0 */
4654 static unsigned long weighted_cpuload(const int cpu
)
4656 return cfs_rq_runnable_load_avg(&cpu_rq(cpu
)->cfs
);
4659 #ifdef CONFIG_NO_HZ_COMMON
4661 * There is no sane way to deal with nohz on smp when using jiffies because the
4662 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4663 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4665 * Therefore we need to avoid the delta approach from the regular tick when
4666 * possible since that would seriously skew the load calculation. This is why we
4667 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
4668 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
4669 * loop exit, nohz_idle_balance, nohz full exit...)
4671 * This means we might still be one tick off for nohz periods.
4674 static void cpu_load_update_nohz(struct rq
*this_rq
,
4675 unsigned long curr_jiffies
,
4678 unsigned long pending_updates
;
4680 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
4681 if (pending_updates
) {
4682 this_rq
->last_load_update_tick
= curr_jiffies
;
4684 * In the regular NOHZ case, we were idle, this means load 0.
4685 * In the NOHZ_FULL case, we were non-idle, we should consider
4686 * its weighted load.
4688 cpu_load_update(this_rq
, load
, pending_updates
);
4693 * Called from nohz_idle_balance() to update the load ratings before doing the
4696 static void cpu_load_update_idle(struct rq
*this_rq
)
4699 * bail if there's load or we're actually up-to-date.
4701 if (weighted_cpuload(cpu_of(this_rq
)))
4704 cpu_load_update_nohz(this_rq
, READ_ONCE(jiffies
), 0);
4708 * Record CPU load on nohz entry so we know the tickless load to account
4709 * on nohz exit. cpu_load[0] happens then to be updated more frequently
4710 * than other cpu_load[idx] but it should be fine as cpu_load readers
4711 * shouldn't rely into synchronized cpu_load[*] updates.
4713 void cpu_load_update_nohz_start(void)
4715 struct rq
*this_rq
= this_rq();
4718 * This is all lockless but should be fine. If weighted_cpuload changes
4719 * concurrently we'll exit nohz. And cpu_load write can race with
4720 * cpu_load_update_idle() but both updater would be writing the same.
4722 this_rq
->cpu_load
[0] = weighted_cpuload(cpu_of(this_rq
));
4726 * Account the tickless load in the end of a nohz frame.
4728 void cpu_load_update_nohz_stop(void)
4730 unsigned long curr_jiffies
= READ_ONCE(jiffies
);
4731 struct rq
*this_rq
= this_rq();
4734 if (curr_jiffies
== this_rq
->last_load_update_tick
)
4737 load
= weighted_cpuload(cpu_of(this_rq
));
4738 raw_spin_lock(&this_rq
->lock
);
4739 update_rq_clock(this_rq
);
4740 cpu_load_update_nohz(this_rq
, curr_jiffies
, load
);
4741 raw_spin_unlock(&this_rq
->lock
);
4743 #else /* !CONFIG_NO_HZ_COMMON */
4744 static inline void cpu_load_update_nohz(struct rq
*this_rq
,
4745 unsigned long curr_jiffies
,
4746 unsigned long load
) { }
4747 #endif /* CONFIG_NO_HZ_COMMON */
4749 static void cpu_load_update_periodic(struct rq
*this_rq
, unsigned long load
)
4751 #ifdef CONFIG_NO_HZ_COMMON
4752 /* See the mess around cpu_load_update_nohz(). */
4753 this_rq
->last_load_update_tick
= READ_ONCE(jiffies
);
4755 cpu_load_update(this_rq
, load
, 1);
4759 * Called from scheduler_tick()
4761 void cpu_load_update_active(struct rq
*this_rq
)
4763 unsigned long load
= weighted_cpuload(cpu_of(this_rq
));
4765 if (tick_nohz_tick_stopped())
4766 cpu_load_update_nohz(this_rq
, READ_ONCE(jiffies
), load
);
4768 cpu_load_update_periodic(this_rq
, load
);
4772 * Return a low guess at the load of a migration-source cpu weighted
4773 * according to the scheduling class and "nice" value.
4775 * We want to under-estimate the load of migration sources, to
4776 * balance conservatively.
4778 static unsigned long source_load(int cpu
, int type
)
4780 struct rq
*rq
= cpu_rq(cpu
);
4781 unsigned long total
= weighted_cpuload(cpu
);
4783 if (type
== 0 || !sched_feat(LB_BIAS
))
4786 return min(rq
->cpu_load
[type
-1], total
);
4790 * Return a high guess at the load of a migration-target cpu weighted
4791 * according to the scheduling class and "nice" value.
4793 static unsigned long target_load(int cpu
, int type
)
4795 struct rq
*rq
= cpu_rq(cpu
);
4796 unsigned long total
= weighted_cpuload(cpu
);
4798 if (type
== 0 || !sched_feat(LB_BIAS
))
4801 return max(rq
->cpu_load
[type
-1], total
);
4804 static unsigned long capacity_of(int cpu
)
4806 return cpu_rq(cpu
)->cpu_capacity
;
4809 static unsigned long capacity_orig_of(int cpu
)
4811 return cpu_rq(cpu
)->cpu_capacity_orig
;
4814 static unsigned long cpu_avg_load_per_task(int cpu
)
4816 struct rq
*rq
= cpu_rq(cpu
);
4817 unsigned long nr_running
= READ_ONCE(rq
->cfs
.h_nr_running
);
4818 unsigned long load_avg
= weighted_cpuload(cpu
);
4821 return load_avg
/ nr_running
;
4826 #ifdef CONFIG_FAIR_GROUP_SCHED
4828 * effective_load() calculates the load change as seen from the root_task_group
4830 * Adding load to a group doesn't make a group heavier, but can cause movement
4831 * of group shares between cpus. Assuming the shares were perfectly aligned one
4832 * can calculate the shift in shares.
4834 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4835 * on this @cpu and results in a total addition (subtraction) of @wg to the
4836 * total group weight.
4838 * Given a runqueue weight distribution (rw_i) we can compute a shares
4839 * distribution (s_i) using:
4841 * s_i = rw_i / \Sum rw_j (1)
4843 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4844 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4845 * shares distribution (s_i):
4847 * rw_i = { 2, 4, 1, 0 }
4848 * s_i = { 2/7, 4/7, 1/7, 0 }
4850 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4851 * task used to run on and the CPU the waker is running on), we need to
4852 * compute the effect of waking a task on either CPU and, in case of a sync
4853 * wakeup, compute the effect of the current task going to sleep.
4855 * So for a change of @wl to the local @cpu with an overall group weight change
4856 * of @wl we can compute the new shares distribution (s'_i) using:
4858 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4860 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4861 * differences in waking a task to CPU 0. The additional task changes the
4862 * weight and shares distributions like:
4864 * rw'_i = { 3, 4, 1, 0 }
4865 * s'_i = { 3/8, 4/8, 1/8, 0 }
4867 * We can then compute the difference in effective weight by using:
4869 * dw_i = S * (s'_i - s_i) (3)
4871 * Where 'S' is the group weight as seen by its parent.
4873 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4874 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4875 * 4/7) times the weight of the group.
4877 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4879 struct sched_entity
*se
= tg
->se
[cpu
];
4881 if (!tg
->parent
) /* the trivial, non-cgroup case */
4884 for_each_sched_entity(se
) {
4890 * W = @wg + \Sum rw_j
4892 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4897 w
= cfs_rq_load_avg(se
->my_q
) + wl
;
4900 * wl = S * s'_i; see (2)
4903 wl
= (w
* (long)tg
->shares
) / W
;
4908 * Per the above, wl is the new se->load.weight value; since
4909 * those are clipped to [MIN_SHARES, ...) do so now. See
4910 * calc_cfs_shares().
4912 if (wl
< MIN_SHARES
)
4916 * wl = dw_i = S * (s'_i - s_i); see (3)
4918 wl
-= se
->avg
.load_avg
;
4921 * Recursively apply this logic to all parent groups to compute
4922 * the final effective load change on the root group. Since
4923 * only the @tg group gets extra weight, all parent groups can
4924 * only redistribute existing shares. @wl is the shift in shares
4925 * resulting from this level per the above.
4934 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4941 static void record_wakee(struct task_struct
*p
)
4944 * Only decay a single time; tasks that have less then 1 wakeup per
4945 * jiffy will not have built up many flips.
4947 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
4948 current
->wakee_flips
>>= 1;
4949 current
->wakee_flip_decay_ts
= jiffies
;
4952 if (current
->last_wakee
!= p
) {
4953 current
->last_wakee
= p
;
4954 current
->wakee_flips
++;
4959 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4961 * A waker of many should wake a different task than the one last awakened
4962 * at a frequency roughly N times higher than one of its wakees.
4964 * In order to determine whether we should let the load spread vs consolidating
4965 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
4966 * partner, and a factor of lls_size higher frequency in the other.
4968 * With both conditions met, we can be relatively sure that the relationship is
4969 * non-monogamous, with partner count exceeding socket size.
4971 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
4972 * whatever is irrelevant, spread criteria is apparent partner count exceeds
4975 static int wake_wide(struct task_struct
*p
)
4977 unsigned int master
= current
->wakee_flips
;
4978 unsigned int slave
= p
->wakee_flips
;
4979 int factor
= this_cpu_read(sd_llc_size
);
4982 swap(master
, slave
);
4983 if (slave
< factor
|| master
< slave
* factor
)
4988 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4990 s64 this_load
, load
;
4991 s64 this_eff_load
, prev_eff_load
;
4992 int idx
, this_cpu
, prev_cpu
;
4993 struct task_group
*tg
;
4994 unsigned long weight
;
4998 this_cpu
= smp_processor_id();
4999 prev_cpu
= task_cpu(p
);
5000 load
= source_load(prev_cpu
, idx
);
5001 this_load
= target_load(this_cpu
, idx
);
5004 * If sync wakeup then subtract the (maximum possible)
5005 * effect of the currently running task from the load
5006 * of the current CPU:
5009 tg
= task_group(current
);
5010 weight
= current
->se
.avg
.load_avg
;
5012 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
5013 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
5017 weight
= p
->se
.avg
.load_avg
;
5020 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5021 * due to the sync cause above having dropped this_load to 0, we'll
5022 * always have an imbalance, but there's really nothing you can do
5023 * about that, so that's good too.
5025 * Otherwise check if either cpus are near enough in load to allow this
5026 * task to be woken on this_cpu.
5028 this_eff_load
= 100;
5029 this_eff_load
*= capacity_of(prev_cpu
);
5031 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
5032 prev_eff_load
*= capacity_of(this_cpu
);
5034 if (this_load
> 0) {
5035 this_eff_load
*= this_load
+
5036 effective_load(tg
, this_cpu
, weight
, weight
);
5038 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
5041 balanced
= this_eff_load
<= prev_eff_load
;
5043 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
5048 schedstat_inc(sd
, ttwu_move_affine
);
5049 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
5055 * find_idlest_group finds and returns the least busy CPU group within the
5058 static struct sched_group
*
5059 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
5060 int this_cpu
, int sd_flag
)
5062 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
5063 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
5064 int load_idx
= sd
->forkexec_idx
;
5065 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
5067 if (sd_flag
& SD_BALANCE_WAKE
)
5068 load_idx
= sd
->wake_idx
;
5071 unsigned long load
, avg_load
;
5075 /* Skip over this group if it has no CPUs allowed */
5076 if (!cpumask_intersects(sched_group_cpus(group
),
5077 tsk_cpus_allowed(p
)))
5080 local_group
= cpumask_test_cpu(this_cpu
,
5081 sched_group_cpus(group
));
5083 /* Tally up the load of all CPUs in the group */
5086 for_each_cpu(i
, sched_group_cpus(group
)) {
5087 /* Bias balancing toward cpus of our domain */
5089 load
= source_load(i
, load_idx
);
5091 load
= target_load(i
, load_idx
);
5096 /* Adjust by relative CPU capacity of the group */
5097 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) / group
->sgc
->capacity
;
5100 this_load
= avg_load
;
5101 } else if (avg_load
< min_load
) {
5102 min_load
= avg_load
;
5105 } while (group
= group
->next
, group
!= sd
->groups
);
5107 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
5113 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5116 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
5118 unsigned long load
, min_load
= ULONG_MAX
;
5119 unsigned int min_exit_latency
= UINT_MAX
;
5120 u64 latest_idle_timestamp
= 0;
5121 int least_loaded_cpu
= this_cpu
;
5122 int shallowest_idle_cpu
= -1;
5125 /* Traverse only the allowed CPUs */
5126 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
5128 struct rq
*rq
= cpu_rq(i
);
5129 struct cpuidle_state
*idle
= idle_get_state(rq
);
5130 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
5132 * We give priority to a CPU whose idle state
5133 * has the smallest exit latency irrespective
5134 * of any idle timestamp.
5136 min_exit_latency
= idle
->exit_latency
;
5137 latest_idle_timestamp
= rq
->idle_stamp
;
5138 shallowest_idle_cpu
= i
;
5139 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
5140 rq
->idle_stamp
> latest_idle_timestamp
) {
5142 * If equal or no active idle state, then
5143 * the most recently idled CPU might have
5146 latest_idle_timestamp
= rq
->idle_stamp
;
5147 shallowest_idle_cpu
= i
;
5149 } else if (shallowest_idle_cpu
== -1) {
5150 load
= weighted_cpuload(i
);
5151 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
5153 least_loaded_cpu
= i
;
5158 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
5162 * Try and locate an idle CPU in the sched_domain.
5164 static int select_idle_sibling(struct task_struct
*p
, int target
)
5166 struct sched_domain
*sd
;
5167 struct sched_group
*sg
;
5168 int i
= task_cpu(p
);
5170 if (idle_cpu(target
))
5174 * If the prevous cpu is cache affine and idle, don't be stupid.
5176 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
5180 * Otherwise, iterate the domains and find an eligible idle cpu.
5182 * A completely idle sched group at higher domains is more
5183 * desirable than an idle group at a lower level, because lower
5184 * domains have smaller groups and usually share hardware
5185 * resources which causes tasks to contend on them, e.g. x86
5186 * hyperthread siblings in the lowest domain (SMT) can contend
5187 * on the shared cpu pipeline.
5189 * However, while we prefer idle groups at higher domains
5190 * finding an idle cpu at the lowest domain is still better than
5191 * returning 'target', which we've already established, isn't
5194 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
5195 for_each_lower_domain(sd
) {
5198 if (!cpumask_intersects(sched_group_cpus(sg
),
5199 tsk_cpus_allowed(p
)))
5202 /* Ensure the entire group is idle */
5203 for_each_cpu(i
, sched_group_cpus(sg
)) {
5204 if (i
== target
|| !idle_cpu(i
))
5209 * It doesn't matter which cpu we pick, the
5210 * whole group is idle.
5212 target
= cpumask_first_and(sched_group_cpus(sg
),
5213 tsk_cpus_allowed(p
));
5217 } while (sg
!= sd
->groups
);
5224 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5225 * tasks. The unit of the return value must be the one of capacity so we can
5226 * compare the utilization with the capacity of the CPU that is available for
5227 * CFS task (ie cpu_capacity).
5229 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5230 * recent utilization of currently non-runnable tasks on a CPU. It represents
5231 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5232 * capacity_orig is the cpu_capacity available at the highest frequency
5233 * (arch_scale_freq_capacity()).
5234 * The utilization of a CPU converges towards a sum equal to or less than the
5235 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5236 * the running time on this CPU scaled by capacity_curr.
5238 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5239 * higher than capacity_orig because of unfortunate rounding in
5240 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5241 * the average stabilizes with the new running time. We need to check that the
5242 * utilization stays within the range of [0..capacity_orig] and cap it if
5243 * necessary. Without utilization capping, a group could be seen as overloaded
5244 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5245 * available capacity. We allow utilization to overshoot capacity_curr (but not
5246 * capacity_orig) as it useful for predicting the capacity required after task
5247 * migrations (scheduler-driven DVFS).
5249 static int cpu_util(int cpu
)
5251 unsigned long util
= cpu_rq(cpu
)->cfs
.avg
.util_avg
;
5252 unsigned long capacity
= capacity_orig_of(cpu
);
5254 return (util
>= capacity
) ? capacity
: util
;
5258 * select_task_rq_fair: Select target runqueue for the waking task in domains
5259 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5260 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5262 * Balances load by selecting the idlest cpu in the idlest group, or under
5263 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5265 * Returns the target cpu number.
5267 * preempt must be disabled.
5270 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
5272 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
5273 int cpu
= smp_processor_id();
5274 int new_cpu
= prev_cpu
;
5275 int want_affine
= 0;
5276 int sync
= wake_flags
& WF_SYNC
;
5278 if (sd_flag
& SD_BALANCE_WAKE
) {
5280 want_affine
= !wake_wide(p
) && cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
));
5284 for_each_domain(cpu
, tmp
) {
5285 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
5289 * If both cpu and prev_cpu are part of this domain,
5290 * cpu is a valid SD_WAKE_AFFINE target.
5292 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
5293 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
5298 if (tmp
->flags
& sd_flag
)
5300 else if (!want_affine
)
5305 sd
= NULL
; /* Prefer wake_affine over balance flags */
5306 if (cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
5311 if (sd_flag
& SD_BALANCE_WAKE
) /* XXX always ? */
5312 new_cpu
= select_idle_sibling(p
, new_cpu
);
5315 struct sched_group
*group
;
5318 if (!(sd
->flags
& sd_flag
)) {
5323 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
5329 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
5330 if (new_cpu
== -1 || new_cpu
== cpu
) {
5331 /* Now try balancing at a lower domain level of cpu */
5336 /* Now try balancing at a lower domain level of new_cpu */
5338 weight
= sd
->span_weight
;
5340 for_each_domain(cpu
, tmp
) {
5341 if (weight
<= tmp
->span_weight
)
5343 if (tmp
->flags
& sd_flag
)
5346 /* while loop will break here if sd == NULL */
5354 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5355 * cfs_rq_of(p) references at time of call are still valid and identify the
5356 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5358 static void migrate_task_rq_fair(struct task_struct
*p
)
5361 * As blocked tasks retain absolute vruntime the migration needs to
5362 * deal with this by subtracting the old and adding the new
5363 * min_vruntime -- the latter is done by enqueue_entity() when placing
5364 * the task on the new runqueue.
5366 if (p
->state
== TASK_WAKING
) {
5367 struct sched_entity
*se
= &p
->se
;
5368 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5371 #ifndef CONFIG_64BIT
5372 u64 min_vruntime_copy
;
5375 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
5377 min_vruntime
= cfs_rq
->min_vruntime
;
5378 } while (min_vruntime
!= min_vruntime_copy
);
5380 min_vruntime
= cfs_rq
->min_vruntime
;
5383 se
->vruntime
-= min_vruntime
;
5387 * We are supposed to update the task to "current" time, then its up to date
5388 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5389 * what current time is, so simply throw away the out-of-date time. This
5390 * will result in the wakee task is less decayed, but giving the wakee more
5391 * load sounds not bad.
5393 remove_entity_load_avg(&p
->se
);
5395 /* Tell new CPU we are migrated */
5396 p
->se
.avg
.last_update_time
= 0;
5398 /* We have migrated, no longer consider this task hot */
5399 p
->se
.exec_start
= 0;
5402 static void task_dead_fair(struct task_struct
*p
)
5404 remove_entity_load_avg(&p
->se
);
5406 #endif /* CONFIG_SMP */
5408 static unsigned long
5409 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
5411 unsigned long gran
= sysctl_sched_wakeup_granularity
;
5414 * Since its curr running now, convert the gran from real-time
5415 * to virtual-time in his units.
5417 * By using 'se' instead of 'curr' we penalize light tasks, so
5418 * they get preempted easier. That is, if 'se' < 'curr' then
5419 * the resulting gran will be larger, therefore penalizing the
5420 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5421 * be smaller, again penalizing the lighter task.
5423 * This is especially important for buddies when the leftmost
5424 * task is higher priority than the buddy.
5426 return calc_delta_fair(gran
, se
);
5430 * Should 'se' preempt 'curr'.
5444 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
5446 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
5451 gran
= wakeup_gran(curr
, se
);
5458 static void set_last_buddy(struct sched_entity
*se
)
5460 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
5463 for_each_sched_entity(se
)
5464 cfs_rq_of(se
)->last
= se
;
5467 static void set_next_buddy(struct sched_entity
*se
)
5469 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
5472 for_each_sched_entity(se
)
5473 cfs_rq_of(se
)->next
= se
;
5476 static void set_skip_buddy(struct sched_entity
*se
)
5478 for_each_sched_entity(se
)
5479 cfs_rq_of(se
)->skip
= se
;
5483 * Preempt the current task with a newly woken task if needed:
5485 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
5487 struct task_struct
*curr
= rq
->curr
;
5488 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
5489 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
5490 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
5491 int next_buddy_marked
= 0;
5493 if (unlikely(se
== pse
))
5497 * This is possible from callers such as attach_tasks(), in which we
5498 * unconditionally check_prempt_curr() after an enqueue (which may have
5499 * lead to a throttle). This both saves work and prevents false
5500 * next-buddy nomination below.
5502 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
5505 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
5506 set_next_buddy(pse
);
5507 next_buddy_marked
= 1;
5511 * We can come here with TIF_NEED_RESCHED already set from new task
5514 * Note: this also catches the edge-case of curr being in a throttled
5515 * group (e.g. via set_curr_task), since update_curr() (in the
5516 * enqueue of curr) will have resulted in resched being set. This
5517 * prevents us from potentially nominating it as a false LAST_BUDDY
5520 if (test_tsk_need_resched(curr
))
5523 /* Idle tasks are by definition preempted by non-idle tasks. */
5524 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
5525 likely(p
->policy
!= SCHED_IDLE
))
5529 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5530 * is driven by the tick):
5532 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
5535 find_matching_se(&se
, &pse
);
5536 update_curr(cfs_rq_of(se
));
5538 if (wakeup_preempt_entity(se
, pse
) == 1) {
5540 * Bias pick_next to pick the sched entity that is
5541 * triggering this preemption.
5543 if (!next_buddy_marked
)
5544 set_next_buddy(pse
);
5553 * Only set the backward buddy when the current task is still
5554 * on the rq. This can happen when a wakeup gets interleaved
5555 * with schedule on the ->pre_schedule() or idle_balance()
5556 * point, either of which can * drop the rq lock.
5558 * Also, during early boot the idle thread is in the fair class,
5559 * for obvious reasons its a bad idea to schedule back to it.
5561 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
5564 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
5568 static struct task_struct
*
5569 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
, struct pin_cookie cookie
)
5571 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
5572 struct sched_entity
*se
;
5573 struct task_struct
*p
;
5577 #ifdef CONFIG_FAIR_GROUP_SCHED
5578 if (!cfs_rq
->nr_running
)
5581 if (prev
->sched_class
!= &fair_sched_class
)
5585 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5586 * likely that a next task is from the same cgroup as the current.
5588 * Therefore attempt to avoid putting and setting the entire cgroup
5589 * hierarchy, only change the part that actually changes.
5593 struct sched_entity
*curr
= cfs_rq
->curr
;
5596 * Since we got here without doing put_prev_entity() we also
5597 * have to consider cfs_rq->curr. If it is still a runnable
5598 * entity, update_curr() will update its vruntime, otherwise
5599 * forget we've ever seen it.
5603 update_curr(cfs_rq
);
5608 * This call to check_cfs_rq_runtime() will do the
5609 * throttle and dequeue its entity in the parent(s).
5610 * Therefore the 'simple' nr_running test will indeed
5613 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
5617 se
= pick_next_entity(cfs_rq
, curr
);
5618 cfs_rq
= group_cfs_rq(se
);
5624 * Since we haven't yet done put_prev_entity and if the selected task
5625 * is a different task than we started out with, try and touch the
5626 * least amount of cfs_rqs.
5629 struct sched_entity
*pse
= &prev
->se
;
5631 while (!(cfs_rq
= is_same_group(se
, pse
))) {
5632 int se_depth
= se
->depth
;
5633 int pse_depth
= pse
->depth
;
5635 if (se_depth
<= pse_depth
) {
5636 put_prev_entity(cfs_rq_of(pse
), pse
);
5637 pse
= parent_entity(pse
);
5639 if (se_depth
>= pse_depth
) {
5640 set_next_entity(cfs_rq_of(se
), se
);
5641 se
= parent_entity(se
);
5645 put_prev_entity(cfs_rq
, pse
);
5646 set_next_entity(cfs_rq
, se
);
5649 if (hrtick_enabled(rq
))
5650 hrtick_start_fair(rq
, p
);
5657 if (!cfs_rq
->nr_running
)
5660 put_prev_task(rq
, prev
);
5663 se
= pick_next_entity(cfs_rq
, NULL
);
5664 set_next_entity(cfs_rq
, se
);
5665 cfs_rq
= group_cfs_rq(se
);
5670 if (hrtick_enabled(rq
))
5671 hrtick_start_fair(rq
, p
);
5677 * This is OK, because current is on_cpu, which avoids it being picked
5678 * for load-balance and preemption/IRQs are still disabled avoiding
5679 * further scheduler activity on it and we're being very careful to
5680 * re-start the picking loop.
5682 lockdep_unpin_lock(&rq
->lock
, cookie
);
5683 new_tasks
= idle_balance(rq
);
5684 lockdep_repin_lock(&rq
->lock
, cookie
);
5686 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5687 * possible for any higher priority task to appear. In that case we
5688 * must re-start the pick_next_entity() loop.
5700 * Account for a descheduled task:
5702 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
5704 struct sched_entity
*se
= &prev
->se
;
5705 struct cfs_rq
*cfs_rq
;
5707 for_each_sched_entity(se
) {
5708 cfs_rq
= cfs_rq_of(se
);
5709 put_prev_entity(cfs_rq
, se
);
5714 * sched_yield() is very simple
5716 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5718 static void yield_task_fair(struct rq
*rq
)
5720 struct task_struct
*curr
= rq
->curr
;
5721 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
5722 struct sched_entity
*se
= &curr
->se
;
5725 * Are we the only task in the tree?
5727 if (unlikely(rq
->nr_running
== 1))
5730 clear_buddies(cfs_rq
, se
);
5732 if (curr
->policy
!= SCHED_BATCH
) {
5733 update_rq_clock(rq
);
5735 * Update run-time statistics of the 'current'.
5737 update_curr(cfs_rq
);
5739 * Tell update_rq_clock() that we've just updated,
5740 * so we don't do microscopic update in schedule()
5741 * and double the fastpath cost.
5743 rq_clock_skip_update(rq
, true);
5749 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
5751 struct sched_entity
*se
= &p
->se
;
5753 /* throttled hierarchies are not runnable */
5754 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
5757 /* Tell the scheduler that we'd really like pse to run next. */
5760 yield_task_fair(rq
);
5766 /**************************************************
5767 * Fair scheduling class load-balancing methods.
5771 * The purpose of load-balancing is to achieve the same basic fairness the
5772 * per-cpu scheduler provides, namely provide a proportional amount of compute
5773 * time to each task. This is expressed in the following equation:
5775 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5777 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5778 * W_i,0 is defined as:
5780 * W_i,0 = \Sum_j w_i,j (2)
5782 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5783 * is derived from the nice value as per sched_prio_to_weight[].
5785 * The weight average is an exponential decay average of the instantaneous
5788 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5790 * C_i is the compute capacity of cpu i, typically it is the
5791 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5792 * can also include other factors [XXX].
5794 * To achieve this balance we define a measure of imbalance which follows
5795 * directly from (1):
5797 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5799 * We them move tasks around to minimize the imbalance. In the continuous
5800 * function space it is obvious this converges, in the discrete case we get
5801 * a few fun cases generally called infeasible weight scenarios.
5804 * - infeasible weights;
5805 * - local vs global optima in the discrete case. ]
5810 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5811 * for all i,j solution, we create a tree of cpus that follows the hardware
5812 * topology where each level pairs two lower groups (or better). This results
5813 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5814 * tree to only the first of the previous level and we decrease the frequency
5815 * of load-balance at each level inv. proportional to the number of cpus in
5821 * \Sum { --- * --- * 2^i } = O(n) (5)
5823 * `- size of each group
5824 * | | `- number of cpus doing load-balance
5826 * `- sum over all levels
5828 * Coupled with a limit on how many tasks we can migrate every balance pass,
5829 * this makes (5) the runtime complexity of the balancer.
5831 * An important property here is that each CPU is still (indirectly) connected
5832 * to every other cpu in at most O(log n) steps:
5834 * The adjacency matrix of the resulting graph is given by:
5837 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5840 * And you'll find that:
5842 * A^(log_2 n)_i,j != 0 for all i,j (7)
5844 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5845 * The task movement gives a factor of O(m), giving a convergence complexity
5848 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5853 * In order to avoid CPUs going idle while there's still work to do, new idle
5854 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5855 * tree itself instead of relying on other CPUs to bring it work.
5857 * This adds some complexity to both (5) and (8) but it reduces the total idle
5865 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5868 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5873 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5875 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5877 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5880 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5881 * rewrite all of this once again.]
5884 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
5886 enum fbq_type
{ regular
, remote
, all
};
5888 #define LBF_ALL_PINNED 0x01
5889 #define LBF_NEED_BREAK 0x02
5890 #define LBF_DST_PINNED 0x04
5891 #define LBF_SOME_PINNED 0x08
5894 struct sched_domain
*sd
;
5902 struct cpumask
*dst_grpmask
;
5904 enum cpu_idle_type idle
;
5906 /* The set of CPUs under consideration for load-balancing */
5907 struct cpumask
*cpus
;
5912 unsigned int loop_break
;
5913 unsigned int loop_max
;
5915 enum fbq_type fbq_type
;
5916 struct list_head tasks
;
5920 * Is this task likely cache-hot:
5922 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
5926 lockdep_assert_held(&env
->src_rq
->lock
);
5928 if (p
->sched_class
!= &fair_sched_class
)
5931 if (unlikely(p
->policy
== SCHED_IDLE
))
5935 * Buddy candidates are cache hot:
5937 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
5938 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
5939 &p
->se
== cfs_rq_of(&p
->se
)->last
))
5942 if (sysctl_sched_migration_cost
== -1)
5944 if (sysctl_sched_migration_cost
== 0)
5947 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
5949 return delta
< (s64
)sysctl_sched_migration_cost
;
5952 #ifdef CONFIG_NUMA_BALANCING
5954 * Returns 1, if task migration degrades locality
5955 * Returns 0, if task migration improves locality i.e migration preferred.
5956 * Returns -1, if task migration is not affected by locality.
5958 static int migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
5960 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5961 unsigned long src_faults
, dst_faults
;
5962 int src_nid
, dst_nid
;
5964 if (!static_branch_likely(&sched_numa_balancing
))
5967 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
5970 src_nid
= cpu_to_node(env
->src_cpu
);
5971 dst_nid
= cpu_to_node(env
->dst_cpu
);
5973 if (src_nid
== dst_nid
)
5976 /* Migrating away from the preferred node is always bad. */
5977 if (src_nid
== p
->numa_preferred_nid
) {
5978 if (env
->src_rq
->nr_running
> env
->src_rq
->nr_preferred_running
)
5984 /* Encourage migration to the preferred node. */
5985 if (dst_nid
== p
->numa_preferred_nid
)
5989 src_faults
= group_faults(p
, src_nid
);
5990 dst_faults
= group_faults(p
, dst_nid
);
5992 src_faults
= task_faults(p
, src_nid
);
5993 dst_faults
= task_faults(p
, dst_nid
);
5996 return dst_faults
< src_faults
;
6000 static inline int migrate_degrades_locality(struct task_struct
*p
,
6008 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6011 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
6015 lockdep_assert_held(&env
->src_rq
->lock
);
6018 * We do not migrate tasks that are:
6019 * 1) throttled_lb_pair, or
6020 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6021 * 3) running (obviously), or
6022 * 4) are cache-hot on their current CPU.
6024 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
6027 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
6030 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
6032 env
->flags
|= LBF_SOME_PINNED
;
6035 * Remember if this task can be migrated to any other cpu in
6036 * our sched_group. We may want to revisit it if we couldn't
6037 * meet load balance goals by pulling other tasks on src_cpu.
6039 * Also avoid computing new_dst_cpu if we have already computed
6040 * one in current iteration.
6042 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
6045 /* Prevent to re-select dst_cpu via env's cpus */
6046 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
6047 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
6048 env
->flags
|= LBF_DST_PINNED
;
6049 env
->new_dst_cpu
= cpu
;
6057 /* Record that we found atleast one task that could run on dst_cpu */
6058 env
->flags
&= ~LBF_ALL_PINNED
;
6060 if (task_running(env
->src_rq
, p
)) {
6061 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
6066 * Aggressive migration if:
6067 * 1) destination numa is preferred
6068 * 2) task is cache cold, or
6069 * 3) too many balance attempts have failed.
6071 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
6072 if (tsk_cache_hot
== -1)
6073 tsk_cache_hot
= task_hot(p
, env
);
6075 if (tsk_cache_hot
<= 0 ||
6076 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
6077 if (tsk_cache_hot
== 1) {
6078 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
6079 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
6084 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
6089 * detach_task() -- detach the task for the migration specified in env
6091 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
6093 lockdep_assert_held(&env
->src_rq
->lock
);
6095 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
6096 deactivate_task(env
->src_rq
, p
, 0);
6097 set_task_cpu(p
, env
->dst_cpu
);
6101 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6102 * part of active balancing operations within "domain".
6104 * Returns a task if successful and NULL otherwise.
6106 static struct task_struct
*detach_one_task(struct lb_env
*env
)
6108 struct task_struct
*p
, *n
;
6110 lockdep_assert_held(&env
->src_rq
->lock
);
6112 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
6113 if (!can_migrate_task(p
, env
))
6116 detach_task(p
, env
);
6119 * Right now, this is only the second place where
6120 * lb_gained[env->idle] is updated (other is detach_tasks)
6121 * so we can safely collect stats here rather than
6122 * inside detach_tasks().
6124 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
6130 static const unsigned int sched_nr_migrate_break
= 32;
6133 * detach_tasks() -- tries to detach up to imbalance weighted load from
6134 * busiest_rq, as part of a balancing operation within domain "sd".
6136 * Returns number of detached tasks if successful and 0 otherwise.
6138 static int detach_tasks(struct lb_env
*env
)
6140 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
6141 struct task_struct
*p
;
6145 lockdep_assert_held(&env
->src_rq
->lock
);
6147 if (env
->imbalance
<= 0)
6150 while (!list_empty(tasks
)) {
6152 * We don't want to steal all, otherwise we may be treated likewise,
6153 * which could at worst lead to a livelock crash.
6155 if (env
->idle
!= CPU_NOT_IDLE
&& env
->src_rq
->nr_running
<= 1)
6158 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
6161 /* We've more or less seen every task there is, call it quits */
6162 if (env
->loop
> env
->loop_max
)
6165 /* take a breather every nr_migrate tasks */
6166 if (env
->loop
> env
->loop_break
) {
6167 env
->loop_break
+= sched_nr_migrate_break
;
6168 env
->flags
|= LBF_NEED_BREAK
;
6172 if (!can_migrate_task(p
, env
))
6175 load
= task_h_load(p
);
6177 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
6180 if ((load
/ 2) > env
->imbalance
)
6183 detach_task(p
, env
);
6184 list_add(&p
->se
.group_node
, &env
->tasks
);
6187 env
->imbalance
-= load
;
6189 #ifdef CONFIG_PREEMPT
6191 * NEWIDLE balancing is a source of latency, so preemptible
6192 * kernels will stop after the first task is detached to minimize
6193 * the critical section.
6195 if (env
->idle
== CPU_NEWLY_IDLE
)
6200 * We only want to steal up to the prescribed amount of
6203 if (env
->imbalance
<= 0)
6208 list_move_tail(&p
->se
.group_node
, tasks
);
6212 * Right now, this is one of only two places we collect this stat
6213 * so we can safely collect detach_one_task() stats here rather
6214 * than inside detach_one_task().
6216 schedstat_add(env
->sd
, lb_gained
[env
->idle
], detached
);
6222 * attach_task() -- attach the task detached by detach_task() to its new rq.
6224 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
6226 lockdep_assert_held(&rq
->lock
);
6228 BUG_ON(task_rq(p
) != rq
);
6229 activate_task(rq
, p
, 0);
6230 p
->on_rq
= TASK_ON_RQ_QUEUED
;
6231 check_preempt_curr(rq
, p
, 0);
6235 * attach_one_task() -- attaches the task returned from detach_one_task() to
6238 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
6240 raw_spin_lock(&rq
->lock
);
6242 raw_spin_unlock(&rq
->lock
);
6246 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6249 static void attach_tasks(struct lb_env
*env
)
6251 struct list_head
*tasks
= &env
->tasks
;
6252 struct task_struct
*p
;
6254 raw_spin_lock(&env
->dst_rq
->lock
);
6256 while (!list_empty(tasks
)) {
6257 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
6258 list_del_init(&p
->se
.group_node
);
6260 attach_task(env
->dst_rq
, p
);
6263 raw_spin_unlock(&env
->dst_rq
->lock
);
6266 #ifdef CONFIG_FAIR_GROUP_SCHED
6267 static void update_blocked_averages(int cpu
)
6269 struct rq
*rq
= cpu_rq(cpu
);
6270 struct cfs_rq
*cfs_rq
;
6271 unsigned long flags
;
6273 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6274 update_rq_clock(rq
);
6277 * Iterates the task_group tree in a bottom up fashion, see
6278 * list_add_leaf_cfs_rq() for details.
6280 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
6281 /* throttled entities do not contribute to load */
6282 if (throttled_hierarchy(cfs_rq
))
6285 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq
), cfs_rq
, true))
6286 update_tg_load_avg(cfs_rq
, 0);
6288 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6292 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6293 * This needs to be done in a top-down fashion because the load of a child
6294 * group is a fraction of its parents load.
6296 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
6298 struct rq
*rq
= rq_of(cfs_rq
);
6299 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
6300 unsigned long now
= jiffies
;
6303 if (cfs_rq
->last_h_load_update
== now
)
6306 cfs_rq
->h_load_next
= NULL
;
6307 for_each_sched_entity(se
) {
6308 cfs_rq
= cfs_rq_of(se
);
6309 cfs_rq
->h_load_next
= se
;
6310 if (cfs_rq
->last_h_load_update
== now
)
6315 cfs_rq
->h_load
= cfs_rq_load_avg(cfs_rq
);
6316 cfs_rq
->last_h_load_update
= now
;
6319 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
6320 load
= cfs_rq
->h_load
;
6321 load
= div64_ul(load
* se
->avg
.load_avg
,
6322 cfs_rq_load_avg(cfs_rq
) + 1);
6323 cfs_rq
= group_cfs_rq(se
);
6324 cfs_rq
->h_load
= load
;
6325 cfs_rq
->last_h_load_update
= now
;
6329 static unsigned long task_h_load(struct task_struct
*p
)
6331 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
6333 update_cfs_rq_h_load(cfs_rq
);
6334 return div64_ul(p
->se
.avg
.load_avg
* cfs_rq
->h_load
,
6335 cfs_rq_load_avg(cfs_rq
) + 1);
6338 static inline void update_blocked_averages(int cpu
)
6340 struct rq
*rq
= cpu_rq(cpu
);
6341 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
6342 unsigned long flags
;
6344 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6345 update_rq_clock(rq
);
6346 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq
), cfs_rq
, true);
6347 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6350 static unsigned long task_h_load(struct task_struct
*p
)
6352 return p
->se
.avg
.load_avg
;
6356 /********** Helpers for find_busiest_group ************************/
6365 * sg_lb_stats - stats of a sched_group required for load_balancing
6367 struct sg_lb_stats
{
6368 unsigned long avg_load
; /*Avg load across the CPUs of the group */
6369 unsigned long group_load
; /* Total load over the CPUs of the group */
6370 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
6371 unsigned long load_per_task
;
6372 unsigned long group_capacity
;
6373 unsigned long group_util
; /* Total utilization of the group */
6374 unsigned int sum_nr_running
; /* Nr tasks running in the group */
6375 unsigned int idle_cpus
;
6376 unsigned int group_weight
;
6377 enum group_type group_type
;
6378 int group_no_capacity
;
6379 #ifdef CONFIG_NUMA_BALANCING
6380 unsigned int nr_numa_running
;
6381 unsigned int nr_preferred_running
;
6386 * sd_lb_stats - Structure to store the statistics of a sched_domain
6387 * during load balancing.
6389 struct sd_lb_stats
{
6390 struct sched_group
*busiest
; /* Busiest group in this sd */
6391 struct sched_group
*local
; /* Local group in this sd */
6392 unsigned long total_load
; /* Total load of all groups in sd */
6393 unsigned long total_capacity
; /* Total capacity of all groups in sd */
6394 unsigned long avg_load
; /* Average load across all groups in sd */
6396 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
6397 struct sg_lb_stats local_stat
; /* Statistics of the local group */
6400 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
6403 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6404 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6405 * We must however clear busiest_stat::avg_load because
6406 * update_sd_pick_busiest() reads this before assignment.
6408 *sds
= (struct sd_lb_stats
){
6412 .total_capacity
= 0UL,
6415 .sum_nr_running
= 0,
6416 .group_type
= group_other
,
6422 * get_sd_load_idx - Obtain the load index for a given sched domain.
6423 * @sd: The sched_domain whose load_idx is to be obtained.
6424 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6426 * Return: The load index.
6428 static inline int get_sd_load_idx(struct sched_domain
*sd
,
6429 enum cpu_idle_type idle
)
6435 load_idx
= sd
->busy_idx
;
6438 case CPU_NEWLY_IDLE
:
6439 load_idx
= sd
->newidle_idx
;
6442 load_idx
= sd
->idle_idx
;
6449 static unsigned long scale_rt_capacity(int cpu
)
6451 struct rq
*rq
= cpu_rq(cpu
);
6452 u64 total
, used
, age_stamp
, avg
;
6456 * Since we're reading these variables without serialization make sure
6457 * we read them once before doing sanity checks on them.
6459 age_stamp
= READ_ONCE(rq
->age_stamp
);
6460 avg
= READ_ONCE(rq
->rt_avg
);
6461 delta
= __rq_clock_broken(rq
) - age_stamp
;
6463 if (unlikely(delta
< 0))
6466 total
= sched_avg_period() + delta
;
6468 used
= div_u64(avg
, total
);
6470 if (likely(used
< SCHED_CAPACITY_SCALE
))
6471 return SCHED_CAPACITY_SCALE
- used
;
6476 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
6478 unsigned long capacity
= arch_scale_cpu_capacity(sd
, cpu
);
6479 struct sched_group
*sdg
= sd
->groups
;
6481 cpu_rq(cpu
)->cpu_capacity_orig
= capacity
;
6483 capacity
*= scale_rt_capacity(cpu
);
6484 capacity
>>= SCHED_CAPACITY_SHIFT
;
6489 cpu_rq(cpu
)->cpu_capacity
= capacity
;
6490 sdg
->sgc
->capacity
= capacity
;
6493 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
6495 struct sched_domain
*child
= sd
->child
;
6496 struct sched_group
*group
, *sdg
= sd
->groups
;
6497 unsigned long capacity
;
6498 unsigned long interval
;
6500 interval
= msecs_to_jiffies(sd
->balance_interval
);
6501 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6502 sdg
->sgc
->next_update
= jiffies
+ interval
;
6505 update_cpu_capacity(sd
, cpu
);
6511 if (child
->flags
& SD_OVERLAP
) {
6513 * SD_OVERLAP domains cannot assume that child groups
6514 * span the current group.
6517 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
6518 struct sched_group_capacity
*sgc
;
6519 struct rq
*rq
= cpu_rq(cpu
);
6522 * build_sched_domains() -> init_sched_groups_capacity()
6523 * gets here before we've attached the domains to the
6526 * Use capacity_of(), which is set irrespective of domains
6527 * in update_cpu_capacity().
6529 * This avoids capacity from being 0 and
6530 * causing divide-by-zero issues on boot.
6532 if (unlikely(!rq
->sd
)) {
6533 capacity
+= capacity_of(cpu
);
6537 sgc
= rq
->sd
->groups
->sgc
;
6538 capacity
+= sgc
->capacity
;
6542 * !SD_OVERLAP domains can assume that child groups
6543 * span the current group.
6546 group
= child
->groups
;
6548 capacity
+= group
->sgc
->capacity
;
6549 group
= group
->next
;
6550 } while (group
!= child
->groups
);
6553 sdg
->sgc
->capacity
= capacity
;
6557 * Check whether the capacity of the rq has been noticeably reduced by side
6558 * activity. The imbalance_pct is used for the threshold.
6559 * Return true is the capacity is reduced
6562 check_cpu_capacity(struct rq
*rq
, struct sched_domain
*sd
)
6564 return ((rq
->cpu_capacity
* sd
->imbalance_pct
) <
6565 (rq
->cpu_capacity_orig
* 100));
6569 * Group imbalance indicates (and tries to solve) the problem where balancing
6570 * groups is inadequate due to tsk_cpus_allowed() constraints.
6572 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6573 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6576 * { 0 1 2 3 } { 4 5 6 7 }
6579 * If we were to balance group-wise we'd place two tasks in the first group and
6580 * two tasks in the second group. Clearly this is undesired as it will overload
6581 * cpu 3 and leave one of the cpus in the second group unused.
6583 * The current solution to this issue is detecting the skew in the first group
6584 * by noticing the lower domain failed to reach balance and had difficulty
6585 * moving tasks due to affinity constraints.
6587 * When this is so detected; this group becomes a candidate for busiest; see
6588 * update_sd_pick_busiest(). And calculate_imbalance() and
6589 * find_busiest_group() avoid some of the usual balance conditions to allow it
6590 * to create an effective group imbalance.
6592 * This is a somewhat tricky proposition since the next run might not find the
6593 * group imbalance and decide the groups need to be balanced again. A most
6594 * subtle and fragile situation.
6597 static inline int sg_imbalanced(struct sched_group
*group
)
6599 return group
->sgc
->imbalance
;
6603 * group_has_capacity returns true if the group has spare capacity that could
6604 * be used by some tasks.
6605 * We consider that a group has spare capacity if the * number of task is
6606 * smaller than the number of CPUs or if the utilization is lower than the
6607 * available capacity for CFS tasks.
6608 * For the latter, we use a threshold to stabilize the state, to take into
6609 * account the variance of the tasks' load and to return true if the available
6610 * capacity in meaningful for the load balancer.
6611 * As an example, an available capacity of 1% can appear but it doesn't make
6612 * any benefit for the load balance.
6615 group_has_capacity(struct lb_env
*env
, struct sg_lb_stats
*sgs
)
6617 if (sgs
->sum_nr_running
< sgs
->group_weight
)
6620 if ((sgs
->group_capacity
* 100) >
6621 (sgs
->group_util
* env
->sd
->imbalance_pct
))
6628 * group_is_overloaded returns true if the group has more tasks than it can
6630 * group_is_overloaded is not equals to !group_has_capacity because a group
6631 * with the exact right number of tasks, has no more spare capacity but is not
6632 * overloaded so both group_has_capacity and group_is_overloaded return
6636 group_is_overloaded(struct lb_env
*env
, struct sg_lb_stats
*sgs
)
6638 if (sgs
->sum_nr_running
<= sgs
->group_weight
)
6641 if ((sgs
->group_capacity
* 100) <
6642 (sgs
->group_util
* env
->sd
->imbalance_pct
))
6649 group_type
group_classify(struct sched_group
*group
,
6650 struct sg_lb_stats
*sgs
)
6652 if (sgs
->group_no_capacity
)
6653 return group_overloaded
;
6655 if (sg_imbalanced(group
))
6656 return group_imbalanced
;
6662 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6663 * @env: The load balancing environment.
6664 * @group: sched_group whose statistics are to be updated.
6665 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6666 * @local_group: Does group contain this_cpu.
6667 * @sgs: variable to hold the statistics for this group.
6668 * @overload: Indicate more than one runnable task for any CPU.
6670 static inline void update_sg_lb_stats(struct lb_env
*env
,
6671 struct sched_group
*group
, int load_idx
,
6672 int local_group
, struct sg_lb_stats
*sgs
,
6678 memset(sgs
, 0, sizeof(*sgs
));
6680 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6681 struct rq
*rq
= cpu_rq(i
);
6683 /* Bias balancing toward cpus of our domain */
6685 load
= target_load(i
, load_idx
);
6687 load
= source_load(i
, load_idx
);
6689 sgs
->group_load
+= load
;
6690 sgs
->group_util
+= cpu_util(i
);
6691 sgs
->sum_nr_running
+= rq
->cfs
.h_nr_running
;
6693 nr_running
= rq
->nr_running
;
6697 #ifdef CONFIG_NUMA_BALANCING
6698 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
6699 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
6701 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
6703 * No need to call idle_cpu() if nr_running is not 0
6705 if (!nr_running
&& idle_cpu(i
))
6709 /* Adjust by relative CPU capacity of the group */
6710 sgs
->group_capacity
= group
->sgc
->capacity
;
6711 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
6713 if (sgs
->sum_nr_running
)
6714 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
6716 sgs
->group_weight
= group
->group_weight
;
6718 sgs
->group_no_capacity
= group_is_overloaded(env
, sgs
);
6719 sgs
->group_type
= group_classify(group
, sgs
);
6723 * update_sd_pick_busiest - return 1 on busiest group
6724 * @env: The load balancing environment.
6725 * @sds: sched_domain statistics
6726 * @sg: sched_group candidate to be checked for being the busiest
6727 * @sgs: sched_group statistics
6729 * Determine if @sg is a busier group than the previously selected
6732 * Return: %true if @sg is a busier group than the previously selected
6733 * busiest group. %false otherwise.
6735 static bool update_sd_pick_busiest(struct lb_env
*env
,
6736 struct sd_lb_stats
*sds
,
6737 struct sched_group
*sg
,
6738 struct sg_lb_stats
*sgs
)
6740 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
6742 if (sgs
->group_type
> busiest
->group_type
)
6745 if (sgs
->group_type
< busiest
->group_type
)
6748 if (sgs
->avg_load
<= busiest
->avg_load
)
6751 /* This is the busiest node in its class. */
6752 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6755 /* No ASYM_PACKING if target cpu is already busy */
6756 if (env
->idle
== CPU_NOT_IDLE
)
6759 * ASYM_PACKING needs to move all the work to the lowest
6760 * numbered CPUs in the group, therefore mark all groups
6761 * higher than ourself as busy.
6763 if (sgs
->sum_nr_running
&& env
->dst_cpu
< group_first_cpu(sg
)) {
6767 /* Prefer to move from highest possible cpu's work */
6768 if (group_first_cpu(sds
->busiest
) < group_first_cpu(sg
))
6775 #ifdef CONFIG_NUMA_BALANCING
6776 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6778 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
6780 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
6785 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6787 if (rq
->nr_running
> rq
->nr_numa_running
)
6789 if (rq
->nr_running
> rq
->nr_preferred_running
)
6794 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6799 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6803 #endif /* CONFIG_NUMA_BALANCING */
6806 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6807 * @env: The load balancing environment.
6808 * @sds: variable to hold the statistics for this sched_domain.
6810 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6812 struct sched_domain
*child
= env
->sd
->child
;
6813 struct sched_group
*sg
= env
->sd
->groups
;
6814 struct sg_lb_stats tmp_sgs
;
6815 int load_idx
, prefer_sibling
= 0;
6816 bool overload
= false;
6818 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
6821 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
6824 struct sg_lb_stats
*sgs
= &tmp_sgs
;
6827 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
6830 sgs
= &sds
->local_stat
;
6832 if (env
->idle
!= CPU_NEWLY_IDLE
||
6833 time_after_eq(jiffies
, sg
->sgc
->next_update
))
6834 update_group_capacity(env
->sd
, env
->dst_cpu
);
6837 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
,
6844 * In case the child domain prefers tasks go to siblings
6845 * first, lower the sg capacity so that we'll try
6846 * and move all the excess tasks away. We lower the capacity
6847 * of a group only if the local group has the capacity to fit
6848 * these excess tasks. The extra check prevents the case where
6849 * you always pull from the heaviest group when it is already
6850 * under-utilized (possible with a large weight task outweighs
6851 * the tasks on the system).
6853 if (prefer_sibling
&& sds
->local
&&
6854 group_has_capacity(env
, &sds
->local_stat
) &&
6855 (sgs
->sum_nr_running
> 1)) {
6856 sgs
->group_no_capacity
= 1;
6857 sgs
->group_type
= group_classify(sg
, sgs
);
6860 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
6862 sds
->busiest_stat
= *sgs
;
6866 /* Now, start updating sd_lb_stats */
6867 sds
->total_load
+= sgs
->group_load
;
6868 sds
->total_capacity
+= sgs
->group_capacity
;
6871 } while (sg
!= env
->sd
->groups
);
6873 if (env
->sd
->flags
& SD_NUMA
)
6874 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
6876 if (!env
->sd
->parent
) {
6877 /* update overload indicator if we are at root domain */
6878 if (env
->dst_rq
->rd
->overload
!= overload
)
6879 env
->dst_rq
->rd
->overload
= overload
;
6885 * check_asym_packing - Check to see if the group is packed into the
6888 * This is primarily intended to used at the sibling level. Some
6889 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6890 * case of POWER7, it can move to lower SMT modes only when higher
6891 * threads are idle. When in lower SMT modes, the threads will
6892 * perform better since they share less core resources. Hence when we
6893 * have idle threads, we want them to be the higher ones.
6895 * This packing function is run on idle threads. It checks to see if
6896 * the busiest CPU in this domain (core in the P7 case) has a higher
6897 * CPU number than the packing function is being run on. Here we are
6898 * assuming lower CPU number will be equivalent to lower a SMT thread
6901 * Return: 1 when packing is required and a task should be moved to
6902 * this CPU. The amount of the imbalance is returned in *imbalance.
6904 * @env: The load balancing environment.
6905 * @sds: Statistics of the sched_domain which is to be packed
6907 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6911 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6914 if (env
->idle
== CPU_NOT_IDLE
)
6920 busiest_cpu
= group_first_cpu(sds
->busiest
);
6921 if (env
->dst_cpu
> busiest_cpu
)
6924 env
->imbalance
= DIV_ROUND_CLOSEST(
6925 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
6926 SCHED_CAPACITY_SCALE
);
6932 * fix_small_imbalance - Calculate the minor imbalance that exists
6933 * amongst the groups of a sched_domain, during
6935 * @env: The load balancing environment.
6936 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6939 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6941 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
6942 unsigned int imbn
= 2;
6943 unsigned long scaled_busy_load_per_task
;
6944 struct sg_lb_stats
*local
, *busiest
;
6946 local
= &sds
->local_stat
;
6947 busiest
= &sds
->busiest_stat
;
6949 if (!local
->sum_nr_running
)
6950 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
6951 else if (busiest
->load_per_task
> local
->load_per_task
)
6954 scaled_busy_load_per_task
=
6955 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6956 busiest
->group_capacity
;
6958 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
6959 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
6960 env
->imbalance
= busiest
->load_per_task
;
6965 * OK, we don't have enough imbalance to justify moving tasks,
6966 * however we may be able to increase total CPU capacity used by
6970 capa_now
+= busiest
->group_capacity
*
6971 min(busiest
->load_per_task
, busiest
->avg_load
);
6972 capa_now
+= local
->group_capacity
*
6973 min(local
->load_per_task
, local
->avg_load
);
6974 capa_now
/= SCHED_CAPACITY_SCALE
;
6976 /* Amount of load we'd subtract */
6977 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
6978 capa_move
+= busiest
->group_capacity
*
6979 min(busiest
->load_per_task
,
6980 busiest
->avg_load
- scaled_busy_load_per_task
);
6983 /* Amount of load we'd add */
6984 if (busiest
->avg_load
* busiest
->group_capacity
<
6985 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
6986 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
6987 local
->group_capacity
;
6989 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6990 local
->group_capacity
;
6992 capa_move
+= local
->group_capacity
*
6993 min(local
->load_per_task
, local
->avg_load
+ tmp
);
6994 capa_move
/= SCHED_CAPACITY_SCALE
;
6996 /* Move if we gain throughput */
6997 if (capa_move
> capa_now
)
6998 env
->imbalance
= busiest
->load_per_task
;
7002 * calculate_imbalance - Calculate the amount of imbalance present within the
7003 * groups of a given sched_domain during load balance.
7004 * @env: load balance environment
7005 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7007 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
7009 unsigned long max_pull
, load_above_capacity
= ~0UL;
7010 struct sg_lb_stats
*local
, *busiest
;
7012 local
= &sds
->local_stat
;
7013 busiest
= &sds
->busiest_stat
;
7015 if (busiest
->group_type
== group_imbalanced
) {
7017 * In the group_imb case we cannot rely on group-wide averages
7018 * to ensure cpu-load equilibrium, look at wider averages. XXX
7020 busiest
->load_per_task
=
7021 min(busiest
->load_per_task
, sds
->avg_load
);
7025 * Avg load of busiest sg can be less and avg load of local sg can
7026 * be greater than avg load across all sgs of sd because avg load
7027 * factors in sg capacity and sgs with smaller group_type are
7028 * skipped when updating the busiest sg:
7030 if (busiest
->avg_load
<= sds
->avg_load
||
7031 local
->avg_load
>= sds
->avg_load
) {
7033 return fix_small_imbalance(env
, sds
);
7037 * If there aren't any idle cpus, avoid creating some.
7039 if (busiest
->group_type
== group_overloaded
&&
7040 local
->group_type
== group_overloaded
) {
7041 load_above_capacity
= busiest
->sum_nr_running
* SCHED_CAPACITY_SCALE
;
7042 if (load_above_capacity
> busiest
->group_capacity
) {
7043 load_above_capacity
-= busiest
->group_capacity
;
7044 load_above_capacity
*= NICE_0_LOAD
;
7045 load_above_capacity
/= busiest
->group_capacity
;
7047 load_above_capacity
= ~0UL;
7051 * We're trying to get all the cpus to the average_load, so we don't
7052 * want to push ourselves above the average load, nor do we wish to
7053 * reduce the max loaded cpu below the average load. At the same time,
7054 * we also don't want to reduce the group load below the group
7055 * capacity. Thus we look for the minimum possible imbalance.
7057 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
7059 /* How much load to actually move to equalise the imbalance */
7060 env
->imbalance
= min(
7061 max_pull
* busiest
->group_capacity
,
7062 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
7063 ) / SCHED_CAPACITY_SCALE
;
7066 * if *imbalance is less than the average load per runnable task
7067 * there is no guarantee that any tasks will be moved so we'll have
7068 * a think about bumping its value to force at least one task to be
7071 if (env
->imbalance
< busiest
->load_per_task
)
7072 return fix_small_imbalance(env
, sds
);
7075 /******* find_busiest_group() helpers end here *********************/
7078 * find_busiest_group - Returns the busiest group within the sched_domain
7079 * if there is an imbalance.
7081 * Also calculates the amount of weighted load which should be moved
7082 * to restore balance.
7084 * @env: The load balancing environment.
7086 * Return: - The busiest group if imbalance exists.
7088 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
7090 struct sg_lb_stats
*local
, *busiest
;
7091 struct sd_lb_stats sds
;
7093 init_sd_lb_stats(&sds
);
7096 * Compute the various statistics relavent for load balancing at
7099 update_sd_lb_stats(env
, &sds
);
7100 local
= &sds
.local_stat
;
7101 busiest
= &sds
.busiest_stat
;
7103 /* ASYM feature bypasses nice load balance check */
7104 if (check_asym_packing(env
, &sds
))
7107 /* There is no busy sibling group to pull tasks from */
7108 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
7111 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
7112 / sds
.total_capacity
;
7115 * If the busiest group is imbalanced the below checks don't
7116 * work because they assume all things are equal, which typically
7117 * isn't true due to cpus_allowed constraints and the like.
7119 if (busiest
->group_type
== group_imbalanced
)
7122 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7123 if (env
->idle
== CPU_NEWLY_IDLE
&& group_has_capacity(env
, local
) &&
7124 busiest
->group_no_capacity
)
7128 * If the local group is busier than the selected busiest group
7129 * don't try and pull any tasks.
7131 if (local
->avg_load
>= busiest
->avg_load
)
7135 * Don't pull any tasks if this group is already above the domain
7138 if (local
->avg_load
>= sds
.avg_load
)
7141 if (env
->idle
== CPU_IDLE
) {
7143 * This cpu is idle. If the busiest group is not overloaded
7144 * and there is no imbalance between this and busiest group
7145 * wrt idle cpus, it is balanced. The imbalance becomes
7146 * significant if the diff is greater than 1 otherwise we
7147 * might end up to just move the imbalance on another group
7149 if ((busiest
->group_type
!= group_overloaded
) &&
7150 (local
->idle_cpus
<= (busiest
->idle_cpus
+ 1)))
7154 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7155 * imbalance_pct to be conservative.
7157 if (100 * busiest
->avg_load
<=
7158 env
->sd
->imbalance_pct
* local
->avg_load
)
7163 /* Looks like there is an imbalance. Compute it */
7164 calculate_imbalance(env
, &sds
);
7173 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7175 static struct rq
*find_busiest_queue(struct lb_env
*env
,
7176 struct sched_group
*group
)
7178 struct rq
*busiest
= NULL
, *rq
;
7179 unsigned long busiest_load
= 0, busiest_capacity
= 1;
7182 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
7183 unsigned long capacity
, wl
;
7187 rt
= fbq_classify_rq(rq
);
7190 * We classify groups/runqueues into three groups:
7191 * - regular: there are !numa tasks
7192 * - remote: there are numa tasks that run on the 'wrong' node
7193 * - all: there is no distinction
7195 * In order to avoid migrating ideally placed numa tasks,
7196 * ignore those when there's better options.
7198 * If we ignore the actual busiest queue to migrate another
7199 * task, the next balance pass can still reduce the busiest
7200 * queue by moving tasks around inside the node.
7202 * If we cannot move enough load due to this classification
7203 * the next pass will adjust the group classification and
7204 * allow migration of more tasks.
7206 * Both cases only affect the total convergence complexity.
7208 if (rt
> env
->fbq_type
)
7211 capacity
= capacity_of(i
);
7213 wl
= weighted_cpuload(i
);
7216 * When comparing with imbalance, use weighted_cpuload()
7217 * which is not scaled with the cpu capacity.
7220 if (rq
->nr_running
== 1 && wl
> env
->imbalance
&&
7221 !check_cpu_capacity(rq
, env
->sd
))
7225 * For the load comparisons with the other cpu's, consider
7226 * the weighted_cpuload() scaled with the cpu capacity, so
7227 * that the load can be moved away from the cpu that is
7228 * potentially running at a lower capacity.
7230 * Thus we're looking for max(wl_i / capacity_i), crosswise
7231 * multiplication to rid ourselves of the division works out
7232 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7233 * our previous maximum.
7235 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
7237 busiest_capacity
= capacity
;
7246 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7247 * so long as it is large enough.
7249 #define MAX_PINNED_INTERVAL 512
7251 /* Working cpumask for load_balance and load_balance_newidle. */
7252 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7254 static int need_active_balance(struct lb_env
*env
)
7256 struct sched_domain
*sd
= env
->sd
;
7258 if (env
->idle
== CPU_NEWLY_IDLE
) {
7261 * ASYM_PACKING needs to force migrate tasks from busy but
7262 * higher numbered CPUs in order to pack all tasks in the
7263 * lowest numbered CPUs.
7265 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
7270 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7271 * It's worth migrating the task if the src_cpu's capacity is reduced
7272 * because of other sched_class or IRQs if more capacity stays
7273 * available on dst_cpu.
7275 if ((env
->idle
!= CPU_NOT_IDLE
) &&
7276 (env
->src_rq
->cfs
.h_nr_running
== 1)) {
7277 if ((check_cpu_capacity(env
->src_rq
, sd
)) &&
7278 (capacity_of(env
->src_cpu
)*sd
->imbalance_pct
< capacity_of(env
->dst_cpu
)*100))
7282 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
7285 static int active_load_balance_cpu_stop(void *data
);
7287 static int should_we_balance(struct lb_env
*env
)
7289 struct sched_group
*sg
= env
->sd
->groups
;
7290 struct cpumask
*sg_cpus
, *sg_mask
;
7291 int cpu
, balance_cpu
= -1;
7294 * In the newly idle case, we will allow all the cpu's
7295 * to do the newly idle load balance.
7297 if (env
->idle
== CPU_NEWLY_IDLE
)
7300 sg_cpus
= sched_group_cpus(sg
);
7301 sg_mask
= sched_group_mask(sg
);
7302 /* Try to find first idle cpu */
7303 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
7304 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
7311 if (balance_cpu
== -1)
7312 balance_cpu
= group_balance_cpu(sg
);
7315 * First idle cpu or the first cpu(busiest) in this sched group
7316 * is eligible for doing load balancing at this and above domains.
7318 return balance_cpu
== env
->dst_cpu
;
7322 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7323 * tasks if there is an imbalance.
7325 static int load_balance(int this_cpu
, struct rq
*this_rq
,
7326 struct sched_domain
*sd
, enum cpu_idle_type idle
,
7327 int *continue_balancing
)
7329 int ld_moved
, cur_ld_moved
, active_balance
= 0;
7330 struct sched_domain
*sd_parent
= sd
->parent
;
7331 struct sched_group
*group
;
7333 unsigned long flags
;
7334 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
7336 struct lb_env env
= {
7338 .dst_cpu
= this_cpu
,
7340 .dst_grpmask
= sched_group_cpus(sd
->groups
),
7342 .loop_break
= sched_nr_migrate_break
,
7345 .tasks
= LIST_HEAD_INIT(env
.tasks
),
7349 * For NEWLY_IDLE load_balancing, we don't need to consider
7350 * other cpus in our group
7352 if (idle
== CPU_NEWLY_IDLE
)
7353 env
.dst_grpmask
= NULL
;
7355 cpumask_copy(cpus
, cpu_active_mask
);
7357 schedstat_inc(sd
, lb_count
[idle
]);
7360 if (!should_we_balance(&env
)) {
7361 *continue_balancing
= 0;
7365 group
= find_busiest_group(&env
);
7367 schedstat_inc(sd
, lb_nobusyg
[idle
]);
7371 busiest
= find_busiest_queue(&env
, group
);
7373 schedstat_inc(sd
, lb_nobusyq
[idle
]);
7377 BUG_ON(busiest
== env
.dst_rq
);
7379 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
7381 env
.src_cpu
= busiest
->cpu
;
7382 env
.src_rq
= busiest
;
7385 if (busiest
->nr_running
> 1) {
7387 * Attempt to move tasks. If find_busiest_group has found
7388 * an imbalance but busiest->nr_running <= 1, the group is
7389 * still unbalanced. ld_moved simply stays zero, so it is
7390 * correctly treated as an imbalance.
7392 env
.flags
|= LBF_ALL_PINNED
;
7393 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
7396 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
7399 * cur_ld_moved - load moved in current iteration
7400 * ld_moved - cumulative load moved across iterations
7402 cur_ld_moved
= detach_tasks(&env
);
7405 * We've detached some tasks from busiest_rq. Every
7406 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7407 * unlock busiest->lock, and we are able to be sure
7408 * that nobody can manipulate the tasks in parallel.
7409 * See task_rq_lock() family for the details.
7412 raw_spin_unlock(&busiest
->lock
);
7416 ld_moved
+= cur_ld_moved
;
7419 local_irq_restore(flags
);
7421 if (env
.flags
& LBF_NEED_BREAK
) {
7422 env
.flags
&= ~LBF_NEED_BREAK
;
7427 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7428 * us and move them to an alternate dst_cpu in our sched_group
7429 * where they can run. The upper limit on how many times we
7430 * iterate on same src_cpu is dependent on number of cpus in our
7433 * This changes load balance semantics a bit on who can move
7434 * load to a given_cpu. In addition to the given_cpu itself
7435 * (or a ilb_cpu acting on its behalf where given_cpu is
7436 * nohz-idle), we now have balance_cpu in a position to move
7437 * load to given_cpu. In rare situations, this may cause
7438 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7439 * _independently_ and at _same_ time to move some load to
7440 * given_cpu) causing exceess load to be moved to given_cpu.
7441 * This however should not happen so much in practice and
7442 * moreover subsequent load balance cycles should correct the
7443 * excess load moved.
7445 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
7447 /* Prevent to re-select dst_cpu via env's cpus */
7448 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
7450 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
7451 env
.dst_cpu
= env
.new_dst_cpu
;
7452 env
.flags
&= ~LBF_DST_PINNED
;
7454 env
.loop_break
= sched_nr_migrate_break
;
7457 * Go back to "more_balance" rather than "redo" since we
7458 * need to continue with same src_cpu.
7464 * We failed to reach balance because of affinity.
7467 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
7469 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
7470 *group_imbalance
= 1;
7473 /* All tasks on this runqueue were pinned by CPU affinity */
7474 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
7475 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
7476 if (!cpumask_empty(cpus
)) {
7478 env
.loop_break
= sched_nr_migrate_break
;
7481 goto out_all_pinned
;
7486 schedstat_inc(sd
, lb_failed
[idle
]);
7488 * Increment the failure counter only on periodic balance.
7489 * We do not want newidle balance, which can be very
7490 * frequent, pollute the failure counter causing
7491 * excessive cache_hot migrations and active balances.
7493 if (idle
!= CPU_NEWLY_IDLE
)
7494 sd
->nr_balance_failed
++;
7496 if (need_active_balance(&env
)) {
7497 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
7499 /* don't kick the active_load_balance_cpu_stop,
7500 * if the curr task on busiest cpu can't be
7503 if (!cpumask_test_cpu(this_cpu
,
7504 tsk_cpus_allowed(busiest
->curr
))) {
7505 raw_spin_unlock_irqrestore(&busiest
->lock
,
7507 env
.flags
|= LBF_ALL_PINNED
;
7508 goto out_one_pinned
;
7512 * ->active_balance synchronizes accesses to
7513 * ->active_balance_work. Once set, it's cleared
7514 * only after active load balance is finished.
7516 if (!busiest
->active_balance
) {
7517 busiest
->active_balance
= 1;
7518 busiest
->push_cpu
= this_cpu
;
7521 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
7523 if (active_balance
) {
7524 stop_one_cpu_nowait(cpu_of(busiest
),
7525 active_load_balance_cpu_stop
, busiest
,
7526 &busiest
->active_balance_work
);
7529 /* We've kicked active balancing, force task migration. */
7530 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
7533 sd
->nr_balance_failed
= 0;
7535 if (likely(!active_balance
)) {
7536 /* We were unbalanced, so reset the balancing interval */
7537 sd
->balance_interval
= sd
->min_interval
;
7540 * If we've begun active balancing, start to back off. This
7541 * case may not be covered by the all_pinned logic if there
7542 * is only 1 task on the busy runqueue (because we don't call
7545 if (sd
->balance_interval
< sd
->max_interval
)
7546 sd
->balance_interval
*= 2;
7553 * We reach balance although we may have faced some affinity
7554 * constraints. Clear the imbalance flag if it was set.
7557 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
7559 if (*group_imbalance
)
7560 *group_imbalance
= 0;
7565 * We reach balance because all tasks are pinned at this level so
7566 * we can't migrate them. Let the imbalance flag set so parent level
7567 * can try to migrate them.
7569 schedstat_inc(sd
, lb_balanced
[idle
]);
7571 sd
->nr_balance_failed
= 0;
7574 /* tune up the balancing interval */
7575 if (((env
.flags
& LBF_ALL_PINNED
) &&
7576 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
7577 (sd
->balance_interval
< sd
->max_interval
))
7578 sd
->balance_interval
*= 2;
7585 static inline unsigned long
7586 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
7588 unsigned long interval
= sd
->balance_interval
;
7591 interval
*= sd
->busy_factor
;
7593 /* scale ms to jiffies */
7594 interval
= msecs_to_jiffies(interval
);
7595 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
7601 update_next_balance(struct sched_domain
*sd
, int cpu_busy
, unsigned long *next_balance
)
7603 unsigned long interval
, next
;
7605 interval
= get_sd_balance_interval(sd
, cpu_busy
);
7606 next
= sd
->last_balance
+ interval
;
7608 if (time_after(*next_balance
, next
))
7609 *next_balance
= next
;
7613 * idle_balance is called by schedule() if this_cpu is about to become
7614 * idle. Attempts to pull tasks from other CPUs.
7616 static int idle_balance(struct rq
*this_rq
)
7618 unsigned long next_balance
= jiffies
+ HZ
;
7619 int this_cpu
= this_rq
->cpu
;
7620 struct sched_domain
*sd
;
7621 int pulled_task
= 0;
7625 * We must set idle_stamp _before_ calling idle_balance(), such that we
7626 * measure the duration of idle_balance() as idle time.
7628 this_rq
->idle_stamp
= rq_clock(this_rq
);
7630 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
7631 !this_rq
->rd
->overload
) {
7633 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
7635 update_next_balance(sd
, 0, &next_balance
);
7641 raw_spin_unlock(&this_rq
->lock
);
7643 update_blocked_averages(this_cpu
);
7645 for_each_domain(this_cpu
, sd
) {
7646 int continue_balancing
= 1;
7647 u64 t0
, domain_cost
;
7649 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7652 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
7653 update_next_balance(sd
, 0, &next_balance
);
7657 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
7658 t0
= sched_clock_cpu(this_cpu
);
7660 pulled_task
= load_balance(this_cpu
, this_rq
,
7662 &continue_balancing
);
7664 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
7665 if (domain_cost
> sd
->max_newidle_lb_cost
)
7666 sd
->max_newidle_lb_cost
= domain_cost
;
7668 curr_cost
+= domain_cost
;
7671 update_next_balance(sd
, 0, &next_balance
);
7674 * Stop searching for tasks to pull if there are
7675 * now runnable tasks on this rq.
7677 if (pulled_task
|| this_rq
->nr_running
> 0)
7682 raw_spin_lock(&this_rq
->lock
);
7684 if (curr_cost
> this_rq
->max_idle_balance_cost
)
7685 this_rq
->max_idle_balance_cost
= curr_cost
;
7688 * While browsing the domains, we released the rq lock, a task could
7689 * have been enqueued in the meantime. Since we're not going idle,
7690 * pretend we pulled a task.
7692 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
7696 /* Move the next balance forward */
7697 if (time_after(this_rq
->next_balance
, next_balance
))
7698 this_rq
->next_balance
= next_balance
;
7700 /* Is there a task of a high priority class? */
7701 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
7705 this_rq
->idle_stamp
= 0;
7711 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7712 * running tasks off the busiest CPU onto idle CPUs. It requires at
7713 * least 1 task to be running on each physical CPU where possible, and
7714 * avoids physical / logical imbalances.
7716 static int active_load_balance_cpu_stop(void *data
)
7718 struct rq
*busiest_rq
= data
;
7719 int busiest_cpu
= cpu_of(busiest_rq
);
7720 int target_cpu
= busiest_rq
->push_cpu
;
7721 struct rq
*target_rq
= cpu_rq(target_cpu
);
7722 struct sched_domain
*sd
;
7723 struct task_struct
*p
= NULL
;
7725 raw_spin_lock_irq(&busiest_rq
->lock
);
7727 /* make sure the requested cpu hasn't gone down in the meantime */
7728 if (unlikely(busiest_cpu
!= smp_processor_id() ||
7729 !busiest_rq
->active_balance
))
7732 /* Is there any task to move? */
7733 if (busiest_rq
->nr_running
<= 1)
7737 * This condition is "impossible", if it occurs
7738 * we need to fix it. Originally reported by
7739 * Bjorn Helgaas on a 128-cpu setup.
7741 BUG_ON(busiest_rq
== target_rq
);
7743 /* Search for an sd spanning us and the target CPU. */
7745 for_each_domain(target_cpu
, sd
) {
7746 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
7747 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
7752 struct lb_env env
= {
7754 .dst_cpu
= target_cpu
,
7755 .dst_rq
= target_rq
,
7756 .src_cpu
= busiest_rq
->cpu
,
7757 .src_rq
= busiest_rq
,
7761 schedstat_inc(sd
, alb_count
);
7763 p
= detach_one_task(&env
);
7765 schedstat_inc(sd
, alb_pushed
);
7766 /* Active balancing done, reset the failure counter. */
7767 sd
->nr_balance_failed
= 0;
7769 schedstat_inc(sd
, alb_failed
);
7774 busiest_rq
->active_balance
= 0;
7775 raw_spin_unlock(&busiest_rq
->lock
);
7778 attach_one_task(target_rq
, p
);
7785 static inline int on_null_domain(struct rq
*rq
)
7787 return unlikely(!rcu_dereference_sched(rq
->sd
));
7790 #ifdef CONFIG_NO_HZ_COMMON
7792 * idle load balancing details
7793 * - When one of the busy CPUs notice that there may be an idle rebalancing
7794 * needed, they will kick the idle load balancer, which then does idle
7795 * load balancing for all the idle CPUs.
7798 cpumask_var_t idle_cpus_mask
;
7800 unsigned long next_balance
; /* in jiffy units */
7801 } nohz ____cacheline_aligned
;
7803 static inline int find_new_ilb(void)
7805 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
7807 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
7814 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7815 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7816 * CPU (if there is one).
7818 static void nohz_balancer_kick(void)
7822 nohz
.next_balance
++;
7824 ilb_cpu
= find_new_ilb();
7826 if (ilb_cpu
>= nr_cpu_ids
)
7829 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
7832 * Use smp_send_reschedule() instead of resched_cpu().
7833 * This way we generate a sched IPI on the target cpu which
7834 * is idle. And the softirq performing nohz idle load balance
7835 * will be run before returning from the IPI.
7837 smp_send_reschedule(ilb_cpu
);
7841 void nohz_balance_exit_idle(unsigned int cpu
)
7843 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
7845 * Completely isolated CPUs don't ever set, so we must test.
7847 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
7848 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
7849 atomic_dec(&nohz
.nr_cpus
);
7851 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7855 static inline void set_cpu_sd_state_busy(void)
7857 struct sched_domain
*sd
;
7858 int cpu
= smp_processor_id();
7861 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7863 if (!sd
|| !sd
->nohz_idle
)
7867 atomic_inc(&sd
->groups
->sgc
->nr_busy_cpus
);
7872 void set_cpu_sd_state_idle(void)
7874 struct sched_domain
*sd
;
7875 int cpu
= smp_processor_id();
7878 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7880 if (!sd
|| sd
->nohz_idle
)
7884 atomic_dec(&sd
->groups
->sgc
->nr_busy_cpus
);
7890 * This routine will record that the cpu is going idle with tick stopped.
7891 * This info will be used in performing idle load balancing in the future.
7893 void nohz_balance_enter_idle(int cpu
)
7896 * If this cpu is going down, then nothing needs to be done.
7898 if (!cpu_active(cpu
))
7901 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
7905 * If we're a completely isolated CPU, we don't play.
7907 if (on_null_domain(cpu_rq(cpu
)))
7910 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
7911 atomic_inc(&nohz
.nr_cpus
);
7912 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7916 static DEFINE_SPINLOCK(balancing
);
7919 * Scale the max load_balance interval with the number of CPUs in the system.
7920 * This trades load-balance latency on larger machines for less cross talk.
7922 void update_max_interval(void)
7924 max_load_balance_interval
= HZ
*num_online_cpus()/10;
7928 * It checks each scheduling domain to see if it is due to be balanced,
7929 * and initiates a balancing operation if so.
7931 * Balancing parameters are set up in init_sched_domains.
7933 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
7935 int continue_balancing
= 1;
7937 unsigned long interval
;
7938 struct sched_domain
*sd
;
7939 /* Earliest time when we have to do rebalance again */
7940 unsigned long next_balance
= jiffies
+ 60*HZ
;
7941 int update_next_balance
= 0;
7942 int need_serialize
, need_decay
= 0;
7945 update_blocked_averages(cpu
);
7948 for_each_domain(cpu
, sd
) {
7950 * Decay the newidle max times here because this is a regular
7951 * visit to all the domains. Decay ~1% per second.
7953 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
7954 sd
->max_newidle_lb_cost
=
7955 (sd
->max_newidle_lb_cost
* 253) / 256;
7956 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
7959 max_cost
+= sd
->max_newidle_lb_cost
;
7961 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7965 * Stop the load balance at this level. There is another
7966 * CPU in our sched group which is doing load balancing more
7969 if (!continue_balancing
) {
7975 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7977 need_serialize
= sd
->flags
& SD_SERIALIZE
;
7978 if (need_serialize
) {
7979 if (!spin_trylock(&balancing
))
7983 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
7984 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
7986 * The LBF_DST_PINNED logic could have changed
7987 * env->dst_cpu, so we can't know our idle
7988 * state even if we migrated tasks. Update it.
7990 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
7992 sd
->last_balance
= jiffies
;
7993 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7996 spin_unlock(&balancing
);
7998 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
7999 next_balance
= sd
->last_balance
+ interval
;
8000 update_next_balance
= 1;
8005 * Ensure the rq-wide value also decays but keep it at a
8006 * reasonable floor to avoid funnies with rq->avg_idle.
8008 rq
->max_idle_balance_cost
=
8009 max((u64
)sysctl_sched_migration_cost
, max_cost
);
8014 * next_balance will be updated only when there is a need.
8015 * When the cpu is attached to null domain for ex, it will not be
8018 if (likely(update_next_balance
)) {
8019 rq
->next_balance
= next_balance
;
8021 #ifdef CONFIG_NO_HZ_COMMON
8023 * If this CPU has been elected to perform the nohz idle
8024 * balance. Other idle CPUs have already rebalanced with
8025 * nohz_idle_balance() and nohz.next_balance has been
8026 * updated accordingly. This CPU is now running the idle load
8027 * balance for itself and we need to update the
8028 * nohz.next_balance accordingly.
8030 if ((idle
== CPU_IDLE
) && time_after(nohz
.next_balance
, rq
->next_balance
))
8031 nohz
.next_balance
= rq
->next_balance
;
8036 #ifdef CONFIG_NO_HZ_COMMON
8038 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8039 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8041 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
8043 int this_cpu
= this_rq
->cpu
;
8046 /* Earliest time when we have to do rebalance again */
8047 unsigned long next_balance
= jiffies
+ 60*HZ
;
8048 int update_next_balance
= 0;
8050 if (idle
!= CPU_IDLE
||
8051 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
8054 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
8055 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
8059 * If this cpu gets work to do, stop the load balancing
8060 * work being done for other cpus. Next load
8061 * balancing owner will pick it up.
8066 rq
= cpu_rq(balance_cpu
);
8069 * If time for next balance is due,
8072 if (time_after_eq(jiffies
, rq
->next_balance
)) {
8073 raw_spin_lock_irq(&rq
->lock
);
8074 update_rq_clock(rq
);
8075 cpu_load_update_idle(rq
);
8076 raw_spin_unlock_irq(&rq
->lock
);
8077 rebalance_domains(rq
, CPU_IDLE
);
8080 if (time_after(next_balance
, rq
->next_balance
)) {
8081 next_balance
= rq
->next_balance
;
8082 update_next_balance
= 1;
8087 * next_balance will be updated only when there is a need.
8088 * When the CPU is attached to null domain for ex, it will not be
8091 if (likely(update_next_balance
))
8092 nohz
.next_balance
= next_balance
;
8094 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
8098 * Current heuristic for kicking the idle load balancer in the presence
8099 * of an idle cpu in the system.
8100 * - This rq has more than one task.
8101 * - This rq has at least one CFS task and the capacity of the CPU is
8102 * significantly reduced because of RT tasks or IRQs.
8103 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8104 * multiple busy cpu.
8105 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8106 * domain span are idle.
8108 static inline bool nohz_kick_needed(struct rq
*rq
)
8110 unsigned long now
= jiffies
;
8111 struct sched_domain
*sd
;
8112 struct sched_group_capacity
*sgc
;
8113 int nr_busy
, cpu
= rq
->cpu
;
8116 if (unlikely(rq
->idle_balance
))
8120 * We may be recently in ticked or tickless idle mode. At the first
8121 * busy tick after returning from idle, we will update the busy stats.
8123 set_cpu_sd_state_busy();
8124 nohz_balance_exit_idle(cpu
);
8127 * None are in tickless mode and hence no need for NOHZ idle load
8130 if (likely(!atomic_read(&nohz
.nr_cpus
)))
8133 if (time_before(now
, nohz
.next_balance
))
8136 if (rq
->nr_running
>= 2)
8140 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
8142 sgc
= sd
->groups
->sgc
;
8143 nr_busy
= atomic_read(&sgc
->nr_busy_cpus
);
8152 sd
= rcu_dereference(rq
->sd
);
8154 if ((rq
->cfs
.h_nr_running
>= 1) &&
8155 check_cpu_capacity(rq
, sd
)) {
8161 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
8162 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
8163 sched_domain_span(sd
)) < cpu
)) {
8173 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
8177 * run_rebalance_domains is triggered when needed from the scheduler tick.
8178 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8180 static void run_rebalance_domains(struct softirq_action
*h
)
8182 struct rq
*this_rq
= this_rq();
8183 enum cpu_idle_type idle
= this_rq
->idle_balance
?
8184 CPU_IDLE
: CPU_NOT_IDLE
;
8187 * If this cpu has a pending nohz_balance_kick, then do the
8188 * balancing on behalf of the other idle cpus whose ticks are
8189 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8190 * give the idle cpus a chance to load balance. Else we may
8191 * load balance only within the local sched_domain hierarchy
8192 * and abort nohz_idle_balance altogether if we pull some load.
8194 nohz_idle_balance(this_rq
, idle
);
8195 rebalance_domains(this_rq
, idle
);
8199 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8201 void trigger_load_balance(struct rq
*rq
)
8203 /* Don't need to rebalance while attached to NULL domain */
8204 if (unlikely(on_null_domain(rq
)))
8207 if (time_after_eq(jiffies
, rq
->next_balance
))
8208 raise_softirq(SCHED_SOFTIRQ
);
8209 #ifdef CONFIG_NO_HZ_COMMON
8210 if (nohz_kick_needed(rq
))
8211 nohz_balancer_kick();
8215 static void rq_online_fair(struct rq
*rq
)
8219 update_runtime_enabled(rq
);
8222 static void rq_offline_fair(struct rq
*rq
)
8226 /* Ensure any throttled groups are reachable by pick_next_task */
8227 unthrottle_offline_cfs_rqs(rq
);
8230 #endif /* CONFIG_SMP */
8233 * scheduler tick hitting a task of our scheduling class:
8235 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
8237 struct cfs_rq
*cfs_rq
;
8238 struct sched_entity
*se
= &curr
->se
;
8240 for_each_sched_entity(se
) {
8241 cfs_rq
= cfs_rq_of(se
);
8242 entity_tick(cfs_rq
, se
, queued
);
8245 if (static_branch_unlikely(&sched_numa_balancing
))
8246 task_tick_numa(rq
, curr
);
8250 * called on fork with the child task as argument from the parent's context
8251 * - child not yet on the tasklist
8252 * - preemption disabled
8254 static void task_fork_fair(struct task_struct
*p
)
8256 struct cfs_rq
*cfs_rq
;
8257 struct sched_entity
*se
= &p
->se
, *curr
;
8258 int this_cpu
= smp_processor_id();
8259 struct rq
*rq
= this_rq();
8260 unsigned long flags
;
8262 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8264 update_rq_clock(rq
);
8266 cfs_rq
= task_cfs_rq(current
);
8267 curr
= cfs_rq
->curr
;
8270 * Not only the cpu but also the task_group of the parent might have
8271 * been changed after parent->se.parent,cfs_rq were copied to
8272 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8273 * of child point to valid ones.
8276 __set_task_cpu(p
, this_cpu
);
8279 update_curr(cfs_rq
);
8282 se
->vruntime
= curr
->vruntime
;
8283 place_entity(cfs_rq
, se
, 1);
8285 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
8287 * Upon rescheduling, sched_class::put_prev_task() will place
8288 * 'current' within the tree based on its new key value.
8290 swap(curr
->vruntime
, se
->vruntime
);
8294 se
->vruntime
-= cfs_rq
->min_vruntime
;
8296 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8300 * Priority of the task has changed. Check to see if we preempt
8304 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
8306 if (!task_on_rq_queued(p
))
8310 * Reschedule if we are currently running on this runqueue and
8311 * our priority decreased, or if we are not currently running on
8312 * this runqueue and our priority is higher than the current's
8314 if (rq
->curr
== p
) {
8315 if (p
->prio
> oldprio
)
8318 check_preempt_curr(rq
, p
, 0);
8321 static inline bool vruntime_normalized(struct task_struct
*p
)
8323 struct sched_entity
*se
= &p
->se
;
8326 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8327 * the dequeue_entity(.flags=0) will already have normalized the
8334 * When !on_rq, vruntime of the task has usually NOT been normalized.
8335 * But there are some cases where it has already been normalized:
8337 * - A forked child which is waiting for being woken up by
8338 * wake_up_new_task().
8339 * - A task which has been woken up by try_to_wake_up() and
8340 * waiting for actually being woken up by sched_ttwu_pending().
8342 if (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
)
8348 static void detach_task_cfs_rq(struct task_struct
*p
)
8350 struct sched_entity
*se
= &p
->se
;
8351 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
8353 if (!vruntime_normalized(p
)) {
8355 * Fix up our vruntime so that the current sleep doesn't
8356 * cause 'unlimited' sleep bonus.
8358 place_entity(cfs_rq
, se
, 0);
8359 se
->vruntime
-= cfs_rq
->min_vruntime
;
8362 /* Catch up with the cfs_rq and remove our load when we leave */
8363 detach_entity_load_avg(cfs_rq
, se
);
8366 static void attach_task_cfs_rq(struct task_struct
*p
)
8368 struct sched_entity
*se
= &p
->se
;
8369 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
8371 #ifdef CONFIG_FAIR_GROUP_SCHED
8373 * Since the real-depth could have been changed (only FAIR
8374 * class maintain depth value), reset depth properly.
8376 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
8379 /* Synchronize task with its cfs_rq */
8380 attach_entity_load_avg(cfs_rq
, se
);
8382 if (!vruntime_normalized(p
))
8383 se
->vruntime
+= cfs_rq
->min_vruntime
;
8386 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
8388 detach_task_cfs_rq(p
);
8391 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
8393 attach_task_cfs_rq(p
);
8395 if (task_on_rq_queued(p
)) {
8397 * We were most likely switched from sched_rt, so
8398 * kick off the schedule if running, otherwise just see
8399 * if we can still preempt the current task.
8404 check_preempt_curr(rq
, p
, 0);
8408 /* Account for a task changing its policy or group.
8410 * This routine is mostly called to set cfs_rq->curr field when a task
8411 * migrates between groups/classes.
8413 static void set_curr_task_fair(struct rq
*rq
)
8415 struct sched_entity
*se
= &rq
->curr
->se
;
8417 for_each_sched_entity(se
) {
8418 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
8420 set_next_entity(cfs_rq
, se
);
8421 /* ensure bandwidth has been allocated on our new cfs_rq */
8422 account_cfs_rq_runtime(cfs_rq
, 0);
8426 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
8428 cfs_rq
->tasks_timeline
= RB_ROOT
;
8429 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8430 #ifndef CONFIG_64BIT
8431 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
8434 atomic_long_set(&cfs_rq
->removed_load_avg
, 0);
8435 atomic_long_set(&cfs_rq
->removed_util_avg
, 0);
8439 #ifdef CONFIG_FAIR_GROUP_SCHED
8440 static void task_move_group_fair(struct task_struct
*p
)
8442 detach_task_cfs_rq(p
);
8443 set_task_rq(p
, task_cpu(p
));
8446 /* Tell se's cfs_rq has been changed -- migrated */
8447 p
->se
.avg
.last_update_time
= 0;
8449 attach_task_cfs_rq(p
);
8452 void free_fair_sched_group(struct task_group
*tg
)
8456 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
8458 for_each_possible_cpu(i
) {
8460 kfree(tg
->cfs_rq
[i
]);
8469 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8471 struct cfs_rq
*cfs_rq
;
8472 struct sched_entity
*se
;
8475 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8478 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8482 tg
->shares
= NICE_0_LOAD
;
8484 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
8486 for_each_possible_cpu(i
) {
8487 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8488 GFP_KERNEL
, cpu_to_node(i
));
8492 se
= kzalloc_node(sizeof(struct sched_entity
),
8493 GFP_KERNEL
, cpu_to_node(i
));
8497 init_cfs_rq(cfs_rq
);
8498 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
8499 init_entity_runnable_average(se
);
8500 post_init_entity_util_avg(se
);
8511 void unregister_fair_sched_group(struct task_group
*tg
)
8513 unsigned long flags
;
8517 for_each_possible_cpu(cpu
) {
8519 remove_entity_load_avg(tg
->se
[cpu
]);
8522 * Only empty task groups can be destroyed; so we can speculatively
8523 * check on_list without danger of it being re-added.
8525 if (!tg
->cfs_rq
[cpu
]->on_list
)
8530 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8531 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
8532 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8536 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8537 struct sched_entity
*se
, int cpu
,
8538 struct sched_entity
*parent
)
8540 struct rq
*rq
= cpu_rq(cpu
);
8544 init_cfs_rq_runtime(cfs_rq
);
8546 tg
->cfs_rq
[cpu
] = cfs_rq
;
8549 /* se could be NULL for root_task_group */
8554 se
->cfs_rq
= &rq
->cfs
;
8557 se
->cfs_rq
= parent
->my_q
;
8558 se
->depth
= parent
->depth
+ 1;
8562 /* guarantee group entities always have weight */
8563 update_load_set(&se
->load
, NICE_0_LOAD
);
8564 se
->parent
= parent
;
8567 static DEFINE_MUTEX(shares_mutex
);
8569 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8572 unsigned long flags
;
8575 * We can't change the weight of the root cgroup.
8580 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
8582 mutex_lock(&shares_mutex
);
8583 if (tg
->shares
== shares
)
8586 tg
->shares
= shares
;
8587 for_each_possible_cpu(i
) {
8588 struct rq
*rq
= cpu_rq(i
);
8589 struct sched_entity
*se
;
8592 /* Propagate contribution to hierarchy */
8593 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8595 /* Possible calls to update_curr() need rq clock */
8596 update_rq_clock(rq
);
8597 for_each_sched_entity(se
)
8598 update_cfs_shares(group_cfs_rq(se
));
8599 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8603 mutex_unlock(&shares_mutex
);
8606 #else /* CONFIG_FAIR_GROUP_SCHED */
8608 void free_fair_sched_group(struct task_group
*tg
) { }
8610 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8615 void unregister_fair_sched_group(struct task_group
*tg
) { }
8617 #endif /* CONFIG_FAIR_GROUP_SCHED */
8620 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
8622 struct sched_entity
*se
= &task
->se
;
8623 unsigned int rr_interval
= 0;
8626 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8629 if (rq
->cfs
.load
.weight
)
8630 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
8636 * All the scheduling class methods:
8638 const struct sched_class fair_sched_class
= {
8639 .next
= &idle_sched_class
,
8640 .enqueue_task
= enqueue_task_fair
,
8641 .dequeue_task
= dequeue_task_fair
,
8642 .yield_task
= yield_task_fair
,
8643 .yield_to_task
= yield_to_task_fair
,
8645 .check_preempt_curr
= check_preempt_wakeup
,
8647 .pick_next_task
= pick_next_task_fair
,
8648 .put_prev_task
= put_prev_task_fair
,
8651 .select_task_rq
= select_task_rq_fair
,
8652 .migrate_task_rq
= migrate_task_rq_fair
,
8654 .rq_online
= rq_online_fair
,
8655 .rq_offline
= rq_offline_fair
,
8657 .task_dead
= task_dead_fair
,
8658 .set_cpus_allowed
= set_cpus_allowed_common
,
8661 .set_curr_task
= set_curr_task_fair
,
8662 .task_tick
= task_tick_fair
,
8663 .task_fork
= task_fork_fair
,
8665 .prio_changed
= prio_changed_fair
,
8666 .switched_from
= switched_from_fair
,
8667 .switched_to
= switched_to_fair
,
8669 .get_rr_interval
= get_rr_interval_fair
,
8671 .update_curr
= update_curr_fair
,
8673 #ifdef CONFIG_FAIR_GROUP_SCHED
8674 .task_move_group
= task_move_group_fair
,
8678 #ifdef CONFIG_SCHED_DEBUG
8679 void print_cfs_stats(struct seq_file
*m
, int cpu
)
8681 struct cfs_rq
*cfs_rq
;
8684 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
8685 print_cfs_rq(m
, cpu
, cfs_rq
);
8689 #ifdef CONFIG_NUMA_BALANCING
8690 void show_numa_stats(struct task_struct
*p
, struct seq_file
*m
)
8693 unsigned long tsf
= 0, tpf
= 0, gsf
= 0, gpf
= 0;
8695 for_each_online_node(node
) {
8696 if (p
->numa_faults
) {
8697 tsf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
8698 tpf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
8700 if (p
->numa_group
) {
8701 gsf
= p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)],
8702 gpf
= p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
8704 print_numa_stats(m
, node
, tsf
, tpf
, gsf
, gpf
);
8707 #endif /* CONFIG_NUMA_BALANCING */
8708 #endif /* CONFIG_SCHED_DEBUG */
8710 __init
void init_sched_fair_class(void)
8713 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8715 #ifdef CONFIG_NO_HZ_COMMON
8716 nohz
.next_balance
= jiffies
;
8717 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);