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 <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency
= 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG
;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity
= 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency
= 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly
;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
93 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window
= 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
116 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
122 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
128 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling
) {
149 case SCHED_TUNABLESCALING_NONE
:
152 case SCHED_TUNABLESCALING_LINEAR
:
155 case SCHED_TUNABLESCALING_LOG
:
157 factor
= 1 + ilog2(cpus
);
164 static void update_sysctl(void)
166 unsigned int factor
= get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity
);
171 SET_SYSCTL(sched_latency
);
172 SET_SYSCTL(sched_wakeup_granularity
);
176 void sched_init_granularity(void)
181 #define WMULT_CONST (~0U)
182 #define WMULT_SHIFT 32
184 static void __update_inv_weight(struct load_weight
*lw
)
188 if (likely(lw
->inv_weight
))
191 w
= scale_load_down(lw
->weight
);
193 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
195 else if (unlikely(!w
))
196 lw
->inv_weight
= WMULT_CONST
;
198 lw
->inv_weight
= WMULT_CONST
/ w
;
202 * delta_exec * weight / lw.weight
204 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
206 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
207 * we're guaranteed shift stays positive because inv_weight is guaranteed to
208 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
210 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
211 * weight/lw.weight <= 1, and therefore our shift will also be positive.
213 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
215 u64 fact
= scale_load_down(weight
);
216 int shift
= WMULT_SHIFT
;
218 __update_inv_weight(lw
);
220 if (unlikely(fact
>> 32)) {
227 /* hint to use a 32x32->64 mul */
228 fact
= (u64
)(u32
)fact
* lw
->inv_weight
;
235 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
239 const struct sched_class fair_sched_class
;
241 /**************************************************************
242 * CFS operations on generic schedulable entities:
245 #ifdef CONFIG_FAIR_GROUP_SCHED
247 /* cpu runqueue to which this cfs_rq is attached */
248 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
253 /* An entity is a task if it doesn't "own" a runqueue */
254 #define entity_is_task(se) (!se->my_q)
256 static inline struct task_struct
*task_of(struct sched_entity
*se
)
258 #ifdef CONFIG_SCHED_DEBUG
259 WARN_ON_ONCE(!entity_is_task(se
));
261 return container_of(se
, struct task_struct
, se
);
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
268 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
285 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
288 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
290 if (!cfs_rq
->on_list
) {
292 * Ensure we either appear before our parent (if already
293 * enqueued) or force our parent to appear after us when it is
294 * enqueued. The fact that we always enqueue bottom-up
295 * reduces this to two cases.
297 if (cfs_rq
->tg
->parent
&&
298 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
299 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
300 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
302 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
303 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
307 /* We should have no load, but we need to update last_decay. */
308 update_cfs_rq_blocked_load(cfs_rq
, 0);
312 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
314 if (cfs_rq
->on_list
) {
315 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
320 /* Iterate thr' all leaf cfs_rq's on a runqueue */
321 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
322 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
324 /* Do the two (enqueued) entities belong to the same group ? */
325 static inline struct cfs_rq
*
326 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
328 if (se
->cfs_rq
== pse
->cfs_rq
)
334 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
340 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
342 int se_depth
, pse_depth
;
345 * preemption test can be made between sibling entities who are in the
346 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
347 * both tasks until we find their ancestors who are siblings of common
351 /* First walk up until both entities are at same depth */
352 se_depth
= (*se
)->depth
;
353 pse_depth
= (*pse
)->depth
;
355 while (se_depth
> pse_depth
) {
357 *se
= parent_entity(*se
);
360 while (pse_depth
> se_depth
) {
362 *pse
= parent_entity(*pse
);
365 while (!is_same_group(*se
, *pse
)) {
366 *se
= parent_entity(*se
);
367 *pse
= parent_entity(*pse
);
371 #else /* !CONFIG_FAIR_GROUP_SCHED */
373 static inline struct task_struct
*task_of(struct sched_entity
*se
)
375 return container_of(se
, struct task_struct
, se
);
378 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
380 return container_of(cfs_rq
, struct rq
, cfs
);
383 #define entity_is_task(se) 1
385 #define for_each_sched_entity(se) \
386 for (; se; se = NULL)
388 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
390 return &task_rq(p
)->cfs
;
393 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
395 struct task_struct
*p
= task_of(se
);
396 struct rq
*rq
= task_rq(p
);
401 /* runqueue "owned" by this group */
402 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
407 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
411 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
415 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
416 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
418 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
424 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
428 #endif /* CONFIG_FAIR_GROUP_SCHED */
430 static __always_inline
431 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
433 /**************************************************************
434 * Scheduling class tree data structure manipulation methods:
437 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
439 s64 delta
= (s64
)(vruntime
- max_vruntime
);
441 max_vruntime
= vruntime
;
446 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
448 s64 delta
= (s64
)(vruntime
- min_vruntime
);
450 min_vruntime
= vruntime
;
455 static inline int entity_before(struct sched_entity
*a
,
456 struct sched_entity
*b
)
458 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
461 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
463 u64 vruntime
= cfs_rq
->min_vruntime
;
466 vruntime
= cfs_rq
->curr
->vruntime
;
468 if (cfs_rq
->rb_leftmost
) {
469 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
474 vruntime
= se
->vruntime
;
476 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
479 /* ensure we never gain time by being placed backwards. */
480 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
483 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
488 * Enqueue an entity into the rb-tree:
490 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
492 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
493 struct rb_node
*parent
= NULL
;
494 struct sched_entity
*entry
;
498 * Find the right place in the rbtree:
502 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
504 * We dont care about collisions. Nodes with
505 * the same key stay together.
507 if (entity_before(se
, entry
)) {
508 link
= &parent
->rb_left
;
510 link
= &parent
->rb_right
;
516 * Maintain a cache of leftmost tree entries (it is frequently
520 cfs_rq
->rb_leftmost
= &se
->run_node
;
522 rb_link_node(&se
->run_node
, parent
, link
);
523 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
526 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
528 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
529 struct rb_node
*next_node
;
531 next_node
= rb_next(&se
->run_node
);
532 cfs_rq
->rb_leftmost
= next_node
;
535 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
538 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
540 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
545 return rb_entry(left
, struct sched_entity
, run_node
);
548 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
550 struct rb_node
*next
= rb_next(&se
->run_node
);
555 return rb_entry(next
, struct sched_entity
, run_node
);
558 #ifdef CONFIG_SCHED_DEBUG
559 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
561 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
566 return rb_entry(last
, struct sched_entity
, run_node
);
569 /**************************************************************
570 * Scheduling class statistics methods:
573 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
574 void __user
*buffer
, size_t *lenp
,
577 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
578 int factor
= get_update_sysctl_factor();
583 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
584 sysctl_sched_min_granularity
);
586 #define WRT_SYSCTL(name) \
587 (normalized_sysctl_##name = sysctl_##name / (factor))
588 WRT_SYSCTL(sched_min_granularity
);
589 WRT_SYSCTL(sched_latency
);
590 WRT_SYSCTL(sched_wakeup_granularity
);
600 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
602 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
603 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
609 * The idea is to set a period in which each task runs once.
611 * When there are too many tasks (sched_nr_latency) we have to stretch
612 * this period because otherwise the slices get too small.
614 * p = (nr <= nl) ? l : l*nr/nl
616 static u64
__sched_period(unsigned long nr_running
)
618 u64 period
= sysctl_sched_latency
;
619 unsigned long nr_latency
= sched_nr_latency
;
621 if (unlikely(nr_running
> nr_latency
)) {
622 period
= sysctl_sched_min_granularity
;
623 period
*= nr_running
;
630 * We calculate the wall-time slice from the period by taking a part
631 * proportional to the weight.
635 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
637 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
639 for_each_sched_entity(se
) {
640 struct load_weight
*load
;
641 struct load_weight lw
;
643 cfs_rq
= cfs_rq_of(se
);
644 load
= &cfs_rq
->load
;
646 if (unlikely(!se
->on_rq
)) {
649 update_load_add(&lw
, se
->load
.weight
);
652 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
658 * We calculate the vruntime slice of a to-be-inserted task.
662 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
664 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
668 static unsigned long task_h_load(struct task_struct
*p
);
670 static inline void __update_task_entity_contrib(struct sched_entity
*se
);
672 /* Give new task start runnable values to heavy its load in infant time */
673 void init_task_runnable_average(struct task_struct
*p
)
677 p
->se
.avg
.decay_count
= 0;
678 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
679 p
->se
.avg
.runnable_avg_sum
= slice
;
680 p
->se
.avg
.runnable_avg_period
= slice
;
681 __update_task_entity_contrib(&p
->se
);
684 void init_task_runnable_average(struct task_struct
*p
)
690 * Update the current task's runtime statistics.
692 static void update_curr(struct cfs_rq
*cfs_rq
)
694 struct sched_entity
*curr
= cfs_rq
->curr
;
695 u64 now
= rq_clock_task(rq_of(cfs_rq
));
701 delta_exec
= now
- curr
->exec_start
;
702 if (unlikely((s64
)delta_exec
<= 0))
705 curr
->exec_start
= now
;
707 schedstat_set(curr
->statistics
.exec_max
,
708 max(delta_exec
, curr
->statistics
.exec_max
));
710 curr
->sum_exec_runtime
+= delta_exec
;
711 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
713 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
714 update_min_vruntime(cfs_rq
);
716 if (entity_is_task(curr
)) {
717 struct task_struct
*curtask
= task_of(curr
);
719 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
720 cpuacct_charge(curtask
, delta_exec
);
721 account_group_exec_runtime(curtask
, delta_exec
);
724 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
728 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
730 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
734 * Task is being enqueued - update stats:
736 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
739 * Are we enqueueing a waiting task? (for current tasks
740 * a dequeue/enqueue event is a NOP)
742 if (se
!= cfs_rq
->curr
)
743 update_stats_wait_start(cfs_rq
, se
);
747 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
749 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
750 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
751 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
752 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
753 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
754 #ifdef CONFIG_SCHEDSTATS
755 if (entity_is_task(se
)) {
756 trace_sched_stat_wait(task_of(se
),
757 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
760 schedstat_set(se
->statistics
.wait_start
, 0);
764 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
767 * Mark the end of the wait period if dequeueing a
770 if (se
!= cfs_rq
->curr
)
771 update_stats_wait_end(cfs_rq
, se
);
775 * We are picking a new current task - update its stats:
778 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
781 * We are starting a new run period:
783 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
786 /**************************************************
787 * Scheduling class queueing methods:
790 #ifdef CONFIG_NUMA_BALANCING
792 * Approximate time to scan a full NUMA task in ms. The task scan period is
793 * calculated based on the tasks virtual memory size and
794 * numa_balancing_scan_size.
796 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
797 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
799 /* Portion of address space to scan in MB */
800 unsigned int sysctl_numa_balancing_scan_size
= 256;
802 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
803 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
805 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
807 unsigned long rss
= 0;
808 unsigned long nr_scan_pages
;
811 * Calculations based on RSS as non-present and empty pages are skipped
812 * by the PTE scanner and NUMA hinting faults should be trapped based
815 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
816 rss
= get_mm_rss(p
->mm
);
820 rss
= round_up(rss
, nr_scan_pages
);
821 return rss
/ nr_scan_pages
;
824 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
825 #define MAX_SCAN_WINDOW 2560
827 static unsigned int task_scan_min(struct task_struct
*p
)
829 unsigned int scan
, floor
;
830 unsigned int windows
= 1;
832 if (sysctl_numa_balancing_scan_size
< MAX_SCAN_WINDOW
)
833 windows
= MAX_SCAN_WINDOW
/ sysctl_numa_balancing_scan_size
;
834 floor
= 1000 / windows
;
836 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
837 return max_t(unsigned int, floor
, scan
);
840 static unsigned int task_scan_max(struct task_struct
*p
)
842 unsigned int smin
= task_scan_min(p
);
845 /* Watch for min being lower than max due to floor calculations */
846 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
847 return max(smin
, smax
);
850 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
852 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
853 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
856 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
858 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
859 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
865 spinlock_t lock
; /* nr_tasks, tasks */
868 struct list_head task_list
;
871 nodemask_t active_nodes
;
872 unsigned long total_faults
;
874 * Faults_cpu is used to decide whether memory should move
875 * towards the CPU. As a consequence, these stats are weighted
876 * more by CPU use than by memory faults.
878 unsigned long *faults_cpu
;
879 unsigned long faults
[0];
882 /* Shared or private faults. */
883 #define NR_NUMA_HINT_FAULT_TYPES 2
885 /* Memory and CPU locality */
886 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
888 /* Averaged statistics, and temporary buffers. */
889 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
891 pid_t
task_numa_group_id(struct task_struct
*p
)
893 return p
->numa_group
? p
->numa_group
->gid
: 0;
896 static inline int task_faults_idx(int nid
, int priv
)
898 return NR_NUMA_HINT_FAULT_TYPES
* nid
+ priv
;
901 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
903 if (!p
->numa_faults_memory
)
906 return p
->numa_faults_memory
[task_faults_idx(nid
, 0)] +
907 p
->numa_faults_memory
[task_faults_idx(nid
, 1)];
910 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
915 return p
->numa_group
->faults
[task_faults_idx(nid
, 0)] +
916 p
->numa_group
->faults
[task_faults_idx(nid
, 1)];
919 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
921 return group
->faults_cpu
[task_faults_idx(nid
, 0)] +
922 group
->faults_cpu
[task_faults_idx(nid
, 1)];
926 * These return the fraction of accesses done by a particular task, or
927 * task group, on a particular numa node. The group weight is given a
928 * larger multiplier, in order to group tasks together that are almost
929 * evenly spread out between numa nodes.
931 static inline unsigned long task_weight(struct task_struct
*p
, int nid
)
933 unsigned long total_faults
;
935 if (!p
->numa_faults_memory
)
938 total_faults
= p
->total_numa_faults
;
943 return 1000 * task_faults(p
, nid
) / total_faults
;
946 static inline unsigned long group_weight(struct task_struct
*p
, int nid
)
948 if (!p
->numa_group
|| !p
->numa_group
->total_faults
)
951 return 1000 * group_faults(p
, nid
) / p
->numa_group
->total_faults
;
954 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
955 int src_nid
, int dst_cpu
)
957 struct numa_group
*ng
= p
->numa_group
;
958 int dst_nid
= cpu_to_node(dst_cpu
);
959 int last_cpupid
, this_cpupid
;
961 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
964 * Multi-stage node selection is used in conjunction with a periodic
965 * migration fault to build a temporal task<->page relation. By using
966 * a two-stage filter we remove short/unlikely relations.
968 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
969 * a task's usage of a particular page (n_p) per total usage of this
970 * page (n_t) (in a given time-span) to a probability.
972 * Our periodic faults will sample this probability and getting the
973 * same result twice in a row, given these samples are fully
974 * independent, is then given by P(n)^2, provided our sample period
975 * is sufficiently short compared to the usage pattern.
977 * This quadric squishes small probabilities, making it less likely we
978 * act on an unlikely task<->page relation.
980 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
981 if (!cpupid_pid_unset(last_cpupid
) &&
982 cpupid_to_nid(last_cpupid
) != dst_nid
)
985 /* Always allow migrate on private faults */
986 if (cpupid_match_pid(p
, last_cpupid
))
989 /* A shared fault, but p->numa_group has not been set up yet. */
994 * Do not migrate if the destination is not a node that
995 * is actively used by this numa group.
997 if (!node_isset(dst_nid
, ng
->active_nodes
))
1001 * Source is a node that is not actively used by this
1002 * numa group, while the destination is. Migrate.
1004 if (!node_isset(src_nid
, ng
->active_nodes
))
1008 * Both source and destination are nodes in active
1009 * use by this numa group. Maximize memory bandwidth
1010 * by migrating from more heavily used groups, to less
1011 * heavily used ones, spreading the load around.
1012 * Use a 1/4 hysteresis to avoid spurious page movement.
1014 return group_faults(p
, dst_nid
) < (group_faults(p
, src_nid
) * 3 / 4);
1017 static unsigned long weighted_cpuload(const int cpu
);
1018 static unsigned long source_load(int cpu
, int type
);
1019 static unsigned long target_load(int cpu
, int type
);
1020 static unsigned long capacity_of(int cpu
);
1021 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
1023 /* Cached statistics for all CPUs within a node */
1025 unsigned long nr_running
;
1028 /* Total compute capacity of CPUs on a node */
1029 unsigned long compute_capacity
;
1031 /* Approximate capacity in terms of runnable tasks on a node */
1032 unsigned long task_capacity
;
1033 int has_free_capacity
;
1037 * XXX borrowed from update_sg_lb_stats
1039 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1043 memset(ns
, 0, sizeof(*ns
));
1044 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1045 struct rq
*rq
= cpu_rq(cpu
);
1047 ns
->nr_running
+= rq
->nr_running
;
1048 ns
->load
+= weighted_cpuload(cpu
);
1049 ns
->compute_capacity
+= capacity_of(cpu
);
1055 * If we raced with hotplug and there are no CPUs left in our mask
1056 * the @ns structure is NULL'ed and task_numa_compare() will
1057 * not find this node attractive.
1059 * We'll either bail at !has_free_capacity, or we'll detect a huge
1060 * imbalance and bail there.
1066 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
);
1067 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1070 struct task_numa_env
{
1071 struct task_struct
*p
;
1073 int src_cpu
, src_nid
;
1074 int dst_cpu
, dst_nid
;
1076 struct numa_stats src_stats
, dst_stats
;
1080 struct task_struct
*best_task
;
1085 static void task_numa_assign(struct task_numa_env
*env
,
1086 struct task_struct
*p
, long imp
)
1089 put_task_struct(env
->best_task
);
1094 env
->best_imp
= imp
;
1095 env
->best_cpu
= env
->dst_cpu
;
1098 static bool load_too_imbalanced(long src_load
, long dst_load
,
1099 struct task_numa_env
*env
)
1102 long orig_src_load
, orig_dst_load
;
1103 long src_capacity
, dst_capacity
;
1106 * The load is corrected for the CPU capacity available on each node.
1109 * ------------ vs ---------
1110 * src_capacity dst_capacity
1112 src_capacity
= env
->src_stats
.compute_capacity
;
1113 dst_capacity
= env
->dst_stats
.compute_capacity
;
1115 /* We care about the slope of the imbalance, not the direction. */
1116 if (dst_load
< src_load
)
1117 swap(dst_load
, src_load
);
1119 /* Is the difference below the threshold? */
1120 imb
= dst_load
* src_capacity
* 100 -
1121 src_load
* dst_capacity
* env
->imbalance_pct
;
1126 * The imbalance is above the allowed threshold.
1127 * Compare it with the old imbalance.
1129 orig_src_load
= env
->src_stats
.load
;
1130 orig_dst_load
= env
->dst_stats
.load
;
1132 if (orig_dst_load
< orig_src_load
)
1133 swap(orig_dst_load
, orig_src_load
);
1135 old_imb
= orig_dst_load
* src_capacity
* 100 -
1136 orig_src_load
* dst_capacity
* env
->imbalance_pct
;
1138 /* Would this change make things worse? */
1139 return (imb
> old_imb
);
1143 * This checks if the overall compute and NUMA accesses of the system would
1144 * be improved if the source tasks was migrated to the target dst_cpu taking
1145 * into account that it might be best if task running on the dst_cpu should
1146 * be exchanged with the source task
1148 static void task_numa_compare(struct task_numa_env
*env
,
1149 long taskimp
, long groupimp
)
1151 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1152 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1153 struct task_struct
*cur
;
1154 struct task_group
*tg
;
1155 long src_load
, dst_load
;
1157 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1161 cur
= ACCESS_ONCE(dst_rq
->curr
);
1162 if (cur
->pid
== 0) /* idle */
1166 * "imp" is the fault differential for the source task between the
1167 * source and destination node. Calculate the total differential for
1168 * the source task and potential destination task. The more negative
1169 * the value is, the more rmeote accesses that would be expected to
1170 * be incurred if the tasks were swapped.
1173 /* Skip this swap candidate if cannot move to the source cpu */
1174 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1178 * If dst and source tasks are in the same NUMA group, or not
1179 * in any group then look only at task weights.
1181 if (cur
->numa_group
== env
->p
->numa_group
) {
1182 imp
= taskimp
+ task_weight(cur
, env
->src_nid
) -
1183 task_weight(cur
, env
->dst_nid
);
1185 * Add some hysteresis to prevent swapping the
1186 * tasks within a group over tiny differences.
1188 if (cur
->numa_group
)
1192 * Compare the group weights. If a task is all by
1193 * itself (not part of a group), use the task weight
1196 if (cur
->numa_group
)
1197 imp
+= group_weight(cur
, env
->src_nid
) -
1198 group_weight(cur
, env
->dst_nid
);
1200 imp
+= task_weight(cur
, env
->src_nid
) -
1201 task_weight(cur
, env
->dst_nid
);
1205 if (imp
<= env
->best_imp
&& moveimp
<= env
->best_imp
)
1209 /* Is there capacity at our destination? */
1210 if (env
->src_stats
.has_free_capacity
&&
1211 !env
->dst_stats
.has_free_capacity
)
1217 /* Balance doesn't matter much if we're running a task per cpu */
1218 if (imp
> env
->best_imp
&& src_rq
->nr_running
== 1 &&
1219 dst_rq
->nr_running
== 1)
1223 * In the overloaded case, try and keep the load balanced.
1226 src_load
= env
->src_stats
.load
;
1227 dst_load
= env
->dst_stats
.load
;
1229 /* Calculate the effect of moving env->p from src to dst. */
1230 load
= env
->p
->se
.load
.weight
;
1231 tg
= task_group(env
->p
);
1232 src_load
+= effective_load(tg
, env
->src_cpu
, -load
, -load
);
1233 dst_load
+= effective_load(tg
, env
->dst_cpu
, load
, load
);
1235 if (moveimp
> imp
&& moveimp
> env
->best_imp
) {
1237 * If the improvement from just moving env->p direction is
1238 * better than swapping tasks around, check if a move is
1239 * possible. Store a slightly smaller score than moveimp,
1240 * so an actually idle CPU will win.
1242 if (!load_too_imbalanced(src_load
, dst_load
, env
)) {
1249 if (imp
<= env
->best_imp
)
1253 /* Cur moves in the opposite direction. */
1254 load
= cur
->se
.load
.weight
;
1255 tg
= task_group(cur
);
1256 src_load
+= effective_load(tg
, env
->src_cpu
, load
, load
);
1257 dst_load
+= effective_load(tg
, env
->dst_cpu
, -load
, -load
);
1260 if (load_too_imbalanced(src_load
, dst_load
, env
))
1264 task_numa_assign(env
, cur
, imp
);
1269 static void task_numa_find_cpu(struct task_numa_env
*env
,
1270 long taskimp
, long groupimp
)
1274 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1275 /* Skip this CPU if the source task cannot migrate */
1276 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1280 task_numa_compare(env
, taskimp
, groupimp
);
1284 static int task_numa_migrate(struct task_struct
*p
)
1286 struct task_numa_env env
= {
1289 .src_cpu
= task_cpu(p
),
1290 .src_nid
= task_node(p
),
1292 .imbalance_pct
= 112,
1298 struct sched_domain
*sd
;
1299 unsigned long taskweight
, groupweight
;
1301 long taskimp
, groupimp
;
1304 * Pick the lowest SD_NUMA domain, as that would have the smallest
1305 * imbalance and would be the first to start moving tasks about.
1307 * And we want to avoid any moving of tasks about, as that would create
1308 * random movement of tasks -- counter the numa conditions we're trying
1312 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1314 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1318 * Cpusets can break the scheduler domain tree into smaller
1319 * balance domains, some of which do not cross NUMA boundaries.
1320 * Tasks that are "trapped" in such domains cannot be migrated
1321 * elsewhere, so there is no point in (re)trying.
1323 if (unlikely(!sd
)) {
1324 p
->numa_preferred_nid
= task_node(p
);
1328 taskweight
= task_weight(p
, env
.src_nid
);
1329 groupweight
= group_weight(p
, env
.src_nid
);
1330 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1331 env
.dst_nid
= p
->numa_preferred_nid
;
1332 taskimp
= task_weight(p
, env
.dst_nid
) - taskweight
;
1333 groupimp
= group_weight(p
, env
.dst_nid
) - groupweight
;
1334 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1336 /* Try to find a spot on the preferred nid. */
1337 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1339 /* No space available on the preferred nid. Look elsewhere. */
1340 if (env
.best_cpu
== -1) {
1341 for_each_online_node(nid
) {
1342 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1345 /* Only consider nodes where both task and groups benefit */
1346 taskimp
= task_weight(p
, nid
) - taskweight
;
1347 groupimp
= group_weight(p
, nid
) - groupweight
;
1348 if (taskimp
< 0 && groupimp
< 0)
1352 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1353 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1358 * If the task is part of a workload that spans multiple NUMA nodes,
1359 * and is migrating into one of the workload's active nodes, remember
1360 * this node as the task's preferred numa node, so the workload can
1362 * A task that migrated to a second choice node will be better off
1363 * trying for a better one later. Do not set the preferred node here.
1365 if (p
->numa_group
) {
1366 if (env
.best_cpu
== -1)
1371 if (node_isset(nid
, p
->numa_group
->active_nodes
))
1372 sched_setnuma(p
, env
.dst_nid
);
1375 /* No better CPU than the current one was found. */
1376 if (env
.best_cpu
== -1)
1380 * Reset the scan period if the task is being rescheduled on an
1381 * alternative node to recheck if the tasks is now properly placed.
1383 p
->numa_scan_period
= task_scan_min(p
);
1385 if (env
.best_task
== NULL
) {
1386 ret
= migrate_task_to(p
, env
.best_cpu
);
1388 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1392 ret
= migrate_swap(p
, env
.best_task
);
1394 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1395 put_task_struct(env
.best_task
);
1399 /* Attempt to migrate a task to a CPU on the preferred node. */
1400 static void numa_migrate_preferred(struct task_struct
*p
)
1402 unsigned long interval
= HZ
;
1404 /* This task has no NUMA fault statistics yet */
1405 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults_memory
))
1408 /* Periodically retry migrating the task to the preferred node */
1409 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1410 p
->numa_migrate_retry
= jiffies
+ interval
;
1412 /* Success if task is already running on preferred CPU */
1413 if (task_node(p
) == p
->numa_preferred_nid
)
1416 /* Otherwise, try migrate to a CPU on the preferred node */
1417 task_numa_migrate(p
);
1421 * Find the nodes on which the workload is actively running. We do this by
1422 * tracking the nodes from which NUMA hinting faults are triggered. This can
1423 * be different from the set of nodes where the workload's memory is currently
1426 * The bitmask is used to make smarter decisions on when to do NUMA page
1427 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1428 * are added when they cause over 6/16 of the maximum number of faults, but
1429 * only removed when they drop below 3/16.
1431 static void update_numa_active_node_mask(struct numa_group
*numa_group
)
1433 unsigned long faults
, max_faults
= 0;
1436 for_each_online_node(nid
) {
1437 faults
= group_faults_cpu(numa_group
, nid
);
1438 if (faults
> max_faults
)
1439 max_faults
= faults
;
1442 for_each_online_node(nid
) {
1443 faults
= group_faults_cpu(numa_group
, nid
);
1444 if (!node_isset(nid
, numa_group
->active_nodes
)) {
1445 if (faults
> max_faults
* 6 / 16)
1446 node_set(nid
, numa_group
->active_nodes
);
1447 } else if (faults
< max_faults
* 3 / 16)
1448 node_clear(nid
, numa_group
->active_nodes
);
1453 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1454 * increments. The more local the fault statistics are, the higher the scan
1455 * period will be for the next scan window. If local/(local+remote) ratio is
1456 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1457 * the scan period will decrease. Aim for 70% local accesses.
1459 #define NUMA_PERIOD_SLOTS 10
1460 #define NUMA_PERIOD_THRESHOLD 7
1463 * Increase the scan period (slow down scanning) if the majority of
1464 * our memory is already on our local node, or if the majority of
1465 * the page accesses are shared with other processes.
1466 * Otherwise, decrease the scan period.
1468 static void update_task_scan_period(struct task_struct
*p
,
1469 unsigned long shared
, unsigned long private)
1471 unsigned int period_slot
;
1475 unsigned long remote
= p
->numa_faults_locality
[0];
1476 unsigned long local
= p
->numa_faults_locality
[1];
1479 * If there were no record hinting faults then either the task is
1480 * completely idle or all activity is areas that are not of interest
1481 * to automatic numa balancing. Scan slower
1483 if (local
+ shared
== 0) {
1484 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1485 p
->numa_scan_period
<< 1);
1487 p
->mm
->numa_next_scan
= jiffies
+
1488 msecs_to_jiffies(p
->numa_scan_period
);
1494 * Prepare to scale scan period relative to the current period.
1495 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1496 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1497 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1499 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1500 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1501 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1502 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1505 diff
= slot
* period_slot
;
1507 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1510 * Scale scan rate increases based on sharing. There is an
1511 * inverse relationship between the degree of sharing and
1512 * the adjustment made to the scanning period. Broadly
1513 * speaking the intent is that there is little point
1514 * scanning faster if shared accesses dominate as it may
1515 * simply bounce migrations uselessly
1517 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
));
1518 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1521 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1522 task_scan_min(p
), task_scan_max(p
));
1523 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1527 * Get the fraction of time the task has been running since the last
1528 * NUMA placement cycle. The scheduler keeps similar statistics, but
1529 * decays those on a 32ms period, which is orders of magnitude off
1530 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1531 * stats only if the task is so new there are no NUMA statistics yet.
1533 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1535 u64 runtime
, delta
, now
;
1536 /* Use the start of this time slice to avoid calculations. */
1537 now
= p
->se
.exec_start
;
1538 runtime
= p
->se
.sum_exec_runtime
;
1540 if (p
->last_task_numa_placement
) {
1541 delta
= runtime
- p
->last_sum_exec_runtime
;
1542 *period
= now
- p
->last_task_numa_placement
;
1544 delta
= p
->se
.avg
.runnable_avg_sum
;
1545 *period
= p
->se
.avg
.runnable_avg_period
;
1548 p
->last_sum_exec_runtime
= runtime
;
1549 p
->last_task_numa_placement
= now
;
1554 static void task_numa_placement(struct task_struct
*p
)
1556 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1557 unsigned long max_faults
= 0, max_group_faults
= 0;
1558 unsigned long fault_types
[2] = { 0, 0 };
1559 unsigned long total_faults
;
1560 u64 runtime
, period
;
1561 spinlock_t
*group_lock
= NULL
;
1563 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1564 if (p
->numa_scan_seq
== seq
)
1566 p
->numa_scan_seq
= seq
;
1567 p
->numa_scan_period_max
= task_scan_max(p
);
1569 total_faults
= p
->numa_faults_locality
[0] +
1570 p
->numa_faults_locality
[1];
1571 runtime
= numa_get_avg_runtime(p
, &period
);
1573 /* If the task is part of a group prevent parallel updates to group stats */
1574 if (p
->numa_group
) {
1575 group_lock
= &p
->numa_group
->lock
;
1576 spin_lock_irq(group_lock
);
1579 /* Find the node with the highest number of faults */
1580 for_each_online_node(nid
) {
1581 unsigned long faults
= 0, group_faults
= 0;
1584 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
1585 long diff
, f_diff
, f_weight
;
1587 i
= task_faults_idx(nid
, priv
);
1589 /* Decay existing window, copy faults since last scan */
1590 diff
= p
->numa_faults_buffer_memory
[i
] - p
->numa_faults_memory
[i
] / 2;
1591 fault_types
[priv
] += p
->numa_faults_buffer_memory
[i
];
1592 p
->numa_faults_buffer_memory
[i
] = 0;
1595 * Normalize the faults_from, so all tasks in a group
1596 * count according to CPU use, instead of by the raw
1597 * number of faults. Tasks with little runtime have
1598 * little over-all impact on throughput, and thus their
1599 * faults are less important.
1601 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
1602 f_weight
= (f_weight
* p
->numa_faults_buffer_cpu
[i
]) /
1604 f_diff
= f_weight
- p
->numa_faults_cpu
[i
] / 2;
1605 p
->numa_faults_buffer_cpu
[i
] = 0;
1607 p
->numa_faults_memory
[i
] += diff
;
1608 p
->numa_faults_cpu
[i
] += f_diff
;
1609 faults
+= p
->numa_faults_memory
[i
];
1610 p
->total_numa_faults
+= diff
;
1611 if (p
->numa_group
) {
1612 /* safe because we can only change our own group */
1613 p
->numa_group
->faults
[i
] += diff
;
1614 p
->numa_group
->faults_cpu
[i
] += f_diff
;
1615 p
->numa_group
->total_faults
+= diff
;
1616 group_faults
+= p
->numa_group
->faults
[i
];
1620 if (faults
> max_faults
) {
1621 max_faults
= faults
;
1625 if (group_faults
> max_group_faults
) {
1626 max_group_faults
= group_faults
;
1627 max_group_nid
= nid
;
1631 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1633 if (p
->numa_group
) {
1634 update_numa_active_node_mask(p
->numa_group
);
1635 spin_unlock_irq(group_lock
);
1636 max_nid
= max_group_nid
;
1640 /* Set the new preferred node */
1641 if (max_nid
!= p
->numa_preferred_nid
)
1642 sched_setnuma(p
, max_nid
);
1644 if (task_node(p
) != p
->numa_preferred_nid
)
1645 numa_migrate_preferred(p
);
1649 static inline int get_numa_group(struct numa_group
*grp
)
1651 return atomic_inc_not_zero(&grp
->refcount
);
1654 static inline void put_numa_group(struct numa_group
*grp
)
1656 if (atomic_dec_and_test(&grp
->refcount
))
1657 kfree_rcu(grp
, rcu
);
1660 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1663 struct numa_group
*grp
, *my_grp
;
1664 struct task_struct
*tsk
;
1666 int cpu
= cpupid_to_cpu(cpupid
);
1669 if (unlikely(!p
->numa_group
)) {
1670 unsigned int size
= sizeof(struct numa_group
) +
1671 4*nr_node_ids
*sizeof(unsigned long);
1673 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1677 atomic_set(&grp
->refcount
, 1);
1678 spin_lock_init(&grp
->lock
);
1679 INIT_LIST_HEAD(&grp
->task_list
);
1681 /* Second half of the array tracks nids where faults happen */
1682 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
1685 node_set(task_node(current
), grp
->active_nodes
);
1687 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1688 grp
->faults
[i
] = p
->numa_faults_memory
[i
];
1690 grp
->total_faults
= p
->total_numa_faults
;
1692 list_add(&p
->numa_entry
, &grp
->task_list
);
1694 rcu_assign_pointer(p
->numa_group
, grp
);
1698 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1700 if (!cpupid_match_pid(tsk
, cpupid
))
1703 grp
= rcu_dereference(tsk
->numa_group
);
1707 my_grp
= p
->numa_group
;
1712 * Only join the other group if its bigger; if we're the bigger group,
1713 * the other task will join us.
1715 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1719 * Tie-break on the grp address.
1721 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1724 /* Always join threads in the same process. */
1725 if (tsk
->mm
== current
->mm
)
1728 /* Simple filter to avoid false positives due to PID collisions */
1729 if (flags
& TNF_SHARED
)
1732 /* Update priv based on whether false sharing was detected */
1735 if (join
&& !get_numa_group(grp
))
1743 BUG_ON(irqs_disabled());
1744 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
1746 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
1747 my_grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1748 grp
->faults
[i
] += p
->numa_faults_memory
[i
];
1750 my_grp
->total_faults
-= p
->total_numa_faults
;
1751 grp
->total_faults
+= p
->total_numa_faults
;
1753 list_move(&p
->numa_entry
, &grp
->task_list
);
1757 spin_unlock(&my_grp
->lock
);
1758 spin_unlock_irq(&grp
->lock
);
1760 rcu_assign_pointer(p
->numa_group
, grp
);
1762 put_numa_group(my_grp
);
1770 void task_numa_free(struct task_struct
*p
)
1772 struct numa_group
*grp
= p
->numa_group
;
1773 void *numa_faults
= p
->numa_faults_memory
;
1774 unsigned long flags
;
1778 spin_lock_irqsave(&grp
->lock
, flags
);
1779 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1780 grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1781 grp
->total_faults
-= p
->total_numa_faults
;
1783 list_del(&p
->numa_entry
);
1785 spin_unlock_irqrestore(&grp
->lock
, flags
);
1786 rcu_assign_pointer(p
->numa_group
, NULL
);
1787 put_numa_group(grp
);
1790 p
->numa_faults_memory
= NULL
;
1791 p
->numa_faults_buffer_memory
= NULL
;
1792 p
->numa_faults_cpu
= NULL
;
1793 p
->numa_faults_buffer_cpu
= NULL
;
1798 * Got a PROT_NONE fault for a page on @node.
1800 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
1802 struct task_struct
*p
= current
;
1803 bool migrated
= flags
& TNF_MIGRATED
;
1804 int cpu_node
= task_node(current
);
1805 int local
= !!(flags
& TNF_FAULT_LOCAL
);
1808 if (!numabalancing_enabled
)
1811 /* for example, ksmd faulting in a user's mm */
1815 /* Do not worry about placement if exiting */
1816 if (p
->state
== TASK_DEAD
)
1819 /* Allocate buffer to track faults on a per-node basis */
1820 if (unlikely(!p
->numa_faults_memory
)) {
1821 int size
= sizeof(*p
->numa_faults_memory
) *
1822 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
1824 p
->numa_faults_memory
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
1825 if (!p
->numa_faults_memory
)
1828 BUG_ON(p
->numa_faults_buffer_memory
);
1830 * The averaged statistics, shared & private, memory & cpu,
1831 * occupy the first half of the array. The second half of the
1832 * array is for current counters, which are averaged into the
1833 * first set by task_numa_placement.
1835 p
->numa_faults_cpu
= p
->numa_faults_memory
+ (2 * nr_node_ids
);
1836 p
->numa_faults_buffer_memory
= p
->numa_faults_memory
+ (4 * nr_node_ids
);
1837 p
->numa_faults_buffer_cpu
= p
->numa_faults_memory
+ (6 * nr_node_ids
);
1838 p
->total_numa_faults
= 0;
1839 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1843 * First accesses are treated as private, otherwise consider accesses
1844 * to be private if the accessing pid has not changed
1846 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
1849 priv
= cpupid_match_pid(p
, last_cpupid
);
1850 if (!priv
&& !(flags
& TNF_NO_GROUP
))
1851 task_numa_group(p
, last_cpupid
, flags
, &priv
);
1855 * If a workload spans multiple NUMA nodes, a shared fault that
1856 * occurs wholly within the set of nodes that the workload is
1857 * actively using should be counted as local. This allows the
1858 * scan rate to slow down when a workload has settled down.
1860 if (!priv
&& !local
&& p
->numa_group
&&
1861 node_isset(cpu_node
, p
->numa_group
->active_nodes
) &&
1862 node_isset(mem_node
, p
->numa_group
->active_nodes
))
1865 task_numa_placement(p
);
1868 * Retry task to preferred node migration periodically, in case it
1869 * case it previously failed, or the scheduler moved us.
1871 if (time_after(jiffies
, p
->numa_migrate_retry
))
1872 numa_migrate_preferred(p
);
1875 p
->numa_pages_migrated
+= pages
;
1877 p
->numa_faults_buffer_memory
[task_faults_idx(mem_node
, priv
)] += pages
;
1878 p
->numa_faults_buffer_cpu
[task_faults_idx(cpu_node
, priv
)] += pages
;
1879 p
->numa_faults_locality
[local
] += pages
;
1882 static void reset_ptenuma_scan(struct task_struct
*p
)
1884 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
1885 p
->mm
->numa_scan_offset
= 0;
1889 * The expensive part of numa migration is done from task_work context.
1890 * Triggered from task_tick_numa().
1892 void task_numa_work(struct callback_head
*work
)
1894 unsigned long migrate
, next_scan
, now
= jiffies
;
1895 struct task_struct
*p
= current
;
1896 struct mm_struct
*mm
= p
->mm
;
1897 struct vm_area_struct
*vma
;
1898 unsigned long start
, end
;
1899 unsigned long nr_pte_updates
= 0;
1902 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
1904 work
->next
= work
; /* protect against double add */
1906 * Who cares about NUMA placement when they're dying.
1908 * NOTE: make sure not to dereference p->mm before this check,
1909 * exit_task_work() happens _after_ exit_mm() so we could be called
1910 * without p->mm even though we still had it when we enqueued this
1913 if (p
->flags
& PF_EXITING
)
1916 if (!mm
->numa_next_scan
) {
1917 mm
->numa_next_scan
= now
+
1918 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1922 * Enforce maximal scan/migration frequency..
1924 migrate
= mm
->numa_next_scan
;
1925 if (time_before(now
, migrate
))
1928 if (p
->numa_scan_period
== 0) {
1929 p
->numa_scan_period_max
= task_scan_max(p
);
1930 p
->numa_scan_period
= task_scan_min(p
);
1933 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1934 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
1938 * Delay this task enough that another task of this mm will likely win
1939 * the next time around.
1941 p
->node_stamp
+= 2 * TICK_NSEC
;
1943 start
= mm
->numa_scan_offset
;
1944 pages
= sysctl_numa_balancing_scan_size
;
1945 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
1949 down_read(&mm
->mmap_sem
);
1950 vma
= find_vma(mm
, start
);
1952 reset_ptenuma_scan(p
);
1956 for (; vma
; vma
= vma
->vm_next
) {
1957 if (!vma_migratable(vma
) || !vma_policy_mof(p
, vma
))
1961 * Shared library pages mapped by multiple processes are not
1962 * migrated as it is expected they are cache replicated. Avoid
1963 * hinting faults in read-only file-backed mappings or the vdso
1964 * as migrating the pages will be of marginal benefit.
1967 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
1971 * Skip inaccessible VMAs to avoid any confusion between
1972 * PROT_NONE and NUMA hinting ptes
1974 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
1978 start
= max(start
, vma
->vm_start
);
1979 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
1980 end
= min(end
, vma
->vm_end
);
1981 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
1984 * Scan sysctl_numa_balancing_scan_size but ensure that
1985 * at least one PTE is updated so that unused virtual
1986 * address space is quickly skipped.
1989 pages
-= (end
- start
) >> PAGE_SHIFT
;
1996 } while (end
!= vma
->vm_end
);
2001 * It is possible to reach the end of the VMA list but the last few
2002 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2003 * would find the !migratable VMA on the next scan but not reset the
2004 * scanner to the start so check it now.
2007 mm
->numa_scan_offset
= start
;
2009 reset_ptenuma_scan(p
);
2010 up_read(&mm
->mmap_sem
);
2014 * Drive the periodic memory faults..
2016 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2018 struct callback_head
*work
= &curr
->numa_work
;
2022 * We don't care about NUMA placement if we don't have memory.
2024 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2028 * Using runtime rather than walltime has the dual advantage that
2029 * we (mostly) drive the selection from busy threads and that the
2030 * task needs to have done some actual work before we bother with
2033 now
= curr
->se
.sum_exec_runtime
;
2034 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2036 if (now
- curr
->node_stamp
> period
) {
2037 if (!curr
->node_stamp
)
2038 curr
->numa_scan_period
= task_scan_min(curr
);
2039 curr
->node_stamp
+= period
;
2041 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2042 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2043 task_work_add(curr
, work
, true);
2048 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2052 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2056 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2059 #endif /* CONFIG_NUMA_BALANCING */
2062 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2064 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2065 if (!parent_entity(se
))
2066 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2068 if (entity_is_task(se
)) {
2069 struct rq
*rq
= rq_of(cfs_rq
);
2071 account_numa_enqueue(rq
, task_of(se
));
2072 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2075 cfs_rq
->nr_running
++;
2079 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2081 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2082 if (!parent_entity(se
))
2083 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2084 if (entity_is_task(se
)) {
2085 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2086 list_del_init(&se
->group_node
);
2088 cfs_rq
->nr_running
--;
2091 #ifdef CONFIG_FAIR_GROUP_SCHED
2093 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
2098 * Use this CPU's actual weight instead of the last load_contribution
2099 * to gain a more accurate current total weight. See
2100 * update_cfs_rq_load_contribution().
2102 tg_weight
= atomic_long_read(&tg
->load_avg
);
2103 tg_weight
-= cfs_rq
->tg_load_contrib
;
2104 tg_weight
+= cfs_rq
->load
.weight
;
2109 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2111 long tg_weight
, load
, shares
;
2113 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
2114 load
= cfs_rq
->load
.weight
;
2116 shares
= (tg
->shares
* load
);
2118 shares
/= tg_weight
;
2120 if (shares
< MIN_SHARES
)
2121 shares
= MIN_SHARES
;
2122 if (shares
> tg
->shares
)
2123 shares
= tg
->shares
;
2127 # else /* CONFIG_SMP */
2128 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2132 # endif /* CONFIG_SMP */
2133 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2134 unsigned long weight
)
2137 /* commit outstanding execution time */
2138 if (cfs_rq
->curr
== se
)
2139 update_curr(cfs_rq
);
2140 account_entity_dequeue(cfs_rq
, se
);
2143 update_load_set(&se
->load
, weight
);
2146 account_entity_enqueue(cfs_rq
, se
);
2149 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2151 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2153 struct task_group
*tg
;
2154 struct sched_entity
*se
;
2158 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2159 if (!se
|| throttled_hierarchy(cfs_rq
))
2162 if (likely(se
->load
.weight
== tg
->shares
))
2165 shares
= calc_cfs_shares(cfs_rq
, tg
);
2167 reweight_entity(cfs_rq_of(se
), se
, shares
);
2169 #else /* CONFIG_FAIR_GROUP_SCHED */
2170 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2173 #endif /* CONFIG_FAIR_GROUP_SCHED */
2177 * We choose a half-life close to 1 scheduling period.
2178 * Note: The tables below are dependent on this value.
2180 #define LOAD_AVG_PERIOD 32
2181 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2182 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2184 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2185 static const u32 runnable_avg_yN_inv
[] = {
2186 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2187 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2188 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2189 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2190 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2191 0x85aac367, 0x82cd8698,
2195 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2196 * over-estimates when re-combining.
2198 static const u32 runnable_avg_yN_sum
[] = {
2199 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2200 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2201 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2206 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2208 static __always_inline u64
decay_load(u64 val
, u64 n
)
2210 unsigned int local_n
;
2214 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2217 /* after bounds checking we can collapse to 32-bit */
2221 * As y^PERIOD = 1/2, we can combine
2222 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2223 * With a look-up table which covers k^n (n<PERIOD)
2225 * To achieve constant time decay_load.
2227 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2228 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2229 local_n
%= LOAD_AVG_PERIOD
;
2232 val
*= runnable_avg_yN_inv
[local_n
];
2233 /* We don't use SRR here since we always want to round down. */
2238 * For updates fully spanning n periods, the contribution to runnable
2239 * average will be: \Sum 1024*y^n
2241 * We can compute this reasonably efficiently by combining:
2242 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2244 static u32
__compute_runnable_contrib(u64 n
)
2248 if (likely(n
<= LOAD_AVG_PERIOD
))
2249 return runnable_avg_yN_sum
[n
];
2250 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2251 return LOAD_AVG_MAX
;
2253 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2255 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2256 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2258 n
-= LOAD_AVG_PERIOD
;
2259 } while (n
> LOAD_AVG_PERIOD
);
2261 contrib
= decay_load(contrib
, n
);
2262 return contrib
+ runnable_avg_yN_sum
[n
];
2266 * We can represent the historical contribution to runnable average as the
2267 * coefficients of a geometric series. To do this we sub-divide our runnable
2268 * history into segments of approximately 1ms (1024us); label the segment that
2269 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2271 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2273 * (now) (~1ms ago) (~2ms ago)
2275 * Let u_i denote the fraction of p_i that the entity was runnable.
2277 * We then designate the fractions u_i as our co-efficients, yielding the
2278 * following representation of historical load:
2279 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2281 * We choose y based on the with of a reasonably scheduling period, fixing:
2284 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2285 * approximately half as much as the contribution to load within the last ms
2288 * When a period "rolls over" and we have new u_0`, multiplying the previous
2289 * sum again by y is sufficient to update:
2290 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2291 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2293 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2294 struct sched_avg
*sa
,
2298 u32 runnable_contrib
;
2299 int delta_w
, decayed
= 0;
2301 delta
= now
- sa
->last_runnable_update
;
2303 * This should only happen when time goes backwards, which it
2304 * unfortunately does during sched clock init when we swap over to TSC.
2306 if ((s64
)delta
< 0) {
2307 sa
->last_runnable_update
= now
;
2312 * Use 1024ns as the unit of measurement since it's a reasonable
2313 * approximation of 1us and fast to compute.
2318 sa
->last_runnable_update
= now
;
2320 /* delta_w is the amount already accumulated against our next period */
2321 delta_w
= sa
->runnable_avg_period
% 1024;
2322 if (delta
+ delta_w
>= 1024) {
2323 /* period roll-over */
2327 * Now that we know we're crossing a period boundary, figure
2328 * out how much from delta we need to complete the current
2329 * period and accrue it.
2331 delta_w
= 1024 - delta_w
;
2333 sa
->runnable_avg_sum
+= delta_w
;
2334 sa
->runnable_avg_period
+= delta_w
;
2338 /* Figure out how many additional periods this update spans */
2339 periods
= delta
/ 1024;
2342 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2344 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2347 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2348 runnable_contrib
= __compute_runnable_contrib(periods
);
2350 sa
->runnable_avg_sum
+= runnable_contrib
;
2351 sa
->runnable_avg_period
+= runnable_contrib
;
2354 /* Remainder of delta accrued against u_0` */
2356 sa
->runnable_avg_sum
+= delta
;
2357 sa
->runnable_avg_period
+= delta
;
2362 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2363 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2365 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2366 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2368 decays
-= se
->avg
.decay_count
;
2372 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2373 se
->avg
.decay_count
= 0;
2378 #ifdef CONFIG_FAIR_GROUP_SCHED
2379 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2382 struct task_group
*tg
= cfs_rq
->tg
;
2385 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2386 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2388 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2389 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2390 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2395 * Aggregate cfs_rq runnable averages into an equivalent task_group
2396 * representation for computing load contributions.
2398 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2399 struct cfs_rq
*cfs_rq
)
2401 struct task_group
*tg
= cfs_rq
->tg
;
2404 /* The fraction of a cpu used by this cfs_rq */
2405 contrib
= div_u64((u64
)sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2406 sa
->runnable_avg_period
+ 1);
2407 contrib
-= cfs_rq
->tg_runnable_contrib
;
2409 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2410 atomic_add(contrib
, &tg
->runnable_avg
);
2411 cfs_rq
->tg_runnable_contrib
+= contrib
;
2415 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2417 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2418 struct task_group
*tg
= cfs_rq
->tg
;
2423 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2424 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2425 atomic_long_read(&tg
->load_avg
) + 1);
2428 * For group entities we need to compute a correction term in the case
2429 * that they are consuming <1 cpu so that we would contribute the same
2430 * load as a task of equal weight.
2432 * Explicitly co-ordinating this measurement would be expensive, but
2433 * fortunately the sum of each cpus contribution forms a usable
2434 * lower-bound on the true value.
2436 * Consider the aggregate of 2 contributions. Either they are disjoint
2437 * (and the sum represents true value) or they are disjoint and we are
2438 * understating by the aggregate of their overlap.
2440 * Extending this to N cpus, for a given overlap, the maximum amount we
2441 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2442 * cpus that overlap for this interval and w_i is the interval width.
2444 * On a small machine; the first term is well-bounded which bounds the
2445 * total error since w_i is a subset of the period. Whereas on a
2446 * larger machine, while this first term can be larger, if w_i is the
2447 * of consequential size guaranteed to see n_i*w_i quickly converge to
2448 * our upper bound of 1-cpu.
2450 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2451 if (runnable_avg
< NICE_0_LOAD
) {
2452 se
->avg
.load_avg_contrib
*= runnable_avg
;
2453 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2457 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2459 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2460 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2462 #else /* CONFIG_FAIR_GROUP_SCHED */
2463 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2464 int force_update
) {}
2465 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2466 struct cfs_rq
*cfs_rq
) {}
2467 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2468 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2469 #endif /* CONFIG_FAIR_GROUP_SCHED */
2471 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2475 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2476 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2477 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2478 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2481 /* Compute the current contribution to load_avg by se, return any delta */
2482 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2484 long old_contrib
= se
->avg
.load_avg_contrib
;
2486 if (entity_is_task(se
)) {
2487 __update_task_entity_contrib(se
);
2489 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2490 __update_group_entity_contrib(se
);
2493 return se
->avg
.load_avg_contrib
- old_contrib
;
2496 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2499 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2500 cfs_rq
->blocked_load_avg
-= load_contrib
;
2502 cfs_rq
->blocked_load_avg
= 0;
2505 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2507 /* Update a sched_entity's runnable average */
2508 static inline void update_entity_load_avg(struct sched_entity
*se
,
2511 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2516 * For a group entity we need to use their owned cfs_rq_clock_task() in
2517 * case they are the parent of a throttled hierarchy.
2519 if (entity_is_task(se
))
2520 now
= cfs_rq_clock_task(cfs_rq
);
2522 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2524 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2527 contrib_delta
= __update_entity_load_avg_contrib(se
);
2533 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2535 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2539 * Decay the load contributed by all blocked children and account this so that
2540 * their contribution may appropriately discounted when they wake up.
2542 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2544 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2547 decays
= now
- cfs_rq
->last_decay
;
2548 if (!decays
&& !force_update
)
2551 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2552 unsigned long removed_load
;
2553 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2554 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2558 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2560 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2561 cfs_rq
->last_decay
= now
;
2564 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2567 /* Add the load generated by se into cfs_rq's child load-average */
2568 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2569 struct sched_entity
*se
,
2573 * We track migrations using entity decay_count <= 0, on a wake-up
2574 * migration we use a negative decay count to track the remote decays
2575 * accumulated while sleeping.
2577 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2578 * are seen by enqueue_entity_load_avg() as a migration with an already
2579 * constructed load_avg_contrib.
2581 if (unlikely(se
->avg
.decay_count
<= 0)) {
2582 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2583 if (se
->avg
.decay_count
) {
2585 * In a wake-up migration we have to approximate the
2586 * time sleeping. This is because we can't synchronize
2587 * clock_task between the two cpus, and it is not
2588 * guaranteed to be read-safe. Instead, we can
2589 * approximate this using our carried decays, which are
2590 * explicitly atomically readable.
2592 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2594 update_entity_load_avg(se
, 0);
2595 /* Indicate that we're now synchronized and on-rq */
2596 se
->avg
.decay_count
= 0;
2600 __synchronize_entity_decay(se
);
2603 /* migrated tasks did not contribute to our blocked load */
2605 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2606 update_entity_load_avg(se
, 0);
2609 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2610 /* we force update consideration on load-balancer moves */
2611 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2615 * Remove se's load from this cfs_rq child load-average, if the entity is
2616 * transitioning to a blocked state we track its projected decay using
2619 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2620 struct sched_entity
*se
,
2623 update_entity_load_avg(se
, 1);
2624 /* we force update consideration on load-balancer moves */
2625 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2627 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2629 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2630 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2631 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2635 * Update the rq's load with the elapsed running time before entering
2636 * idle. if the last scheduled task is not a CFS task, idle_enter will
2637 * be the only way to update the runnable statistic.
2639 void idle_enter_fair(struct rq
*this_rq
)
2641 update_rq_runnable_avg(this_rq
, 1);
2645 * Update the rq's load with the elapsed idle time before a task is
2646 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2647 * be the only way to update the runnable statistic.
2649 void idle_exit_fair(struct rq
*this_rq
)
2651 update_rq_runnable_avg(this_rq
, 0);
2654 static int idle_balance(struct rq
*this_rq
);
2656 #else /* CONFIG_SMP */
2658 static inline void update_entity_load_avg(struct sched_entity
*se
,
2659 int update_cfs_rq
) {}
2660 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2661 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2662 struct sched_entity
*se
,
2664 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2665 struct sched_entity
*se
,
2667 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2668 int force_update
) {}
2670 static inline int idle_balance(struct rq
*rq
)
2675 #endif /* CONFIG_SMP */
2677 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2679 #ifdef CONFIG_SCHEDSTATS
2680 struct task_struct
*tsk
= NULL
;
2682 if (entity_is_task(se
))
2685 if (se
->statistics
.sleep_start
) {
2686 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2691 if (unlikely(delta
> se
->statistics
.sleep_max
))
2692 se
->statistics
.sleep_max
= delta
;
2694 se
->statistics
.sleep_start
= 0;
2695 se
->statistics
.sum_sleep_runtime
+= delta
;
2698 account_scheduler_latency(tsk
, delta
>> 10, 1);
2699 trace_sched_stat_sleep(tsk
, delta
);
2702 if (se
->statistics
.block_start
) {
2703 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2708 if (unlikely(delta
> se
->statistics
.block_max
))
2709 se
->statistics
.block_max
= delta
;
2711 se
->statistics
.block_start
= 0;
2712 se
->statistics
.sum_sleep_runtime
+= delta
;
2715 if (tsk
->in_iowait
) {
2716 se
->statistics
.iowait_sum
+= delta
;
2717 se
->statistics
.iowait_count
++;
2718 trace_sched_stat_iowait(tsk
, delta
);
2721 trace_sched_stat_blocked(tsk
, delta
);
2724 * Blocking time is in units of nanosecs, so shift by
2725 * 20 to get a milliseconds-range estimation of the
2726 * amount of time that the task spent sleeping:
2728 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2729 profile_hits(SLEEP_PROFILING
,
2730 (void *)get_wchan(tsk
),
2733 account_scheduler_latency(tsk
, delta
>> 10, 0);
2739 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2741 #ifdef CONFIG_SCHED_DEBUG
2742 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2747 if (d
> 3*sysctl_sched_latency
)
2748 schedstat_inc(cfs_rq
, nr_spread_over
);
2753 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2755 u64 vruntime
= cfs_rq
->min_vruntime
;
2758 * The 'current' period is already promised to the current tasks,
2759 * however the extra weight of the new task will slow them down a
2760 * little, place the new task so that it fits in the slot that
2761 * stays open at the end.
2763 if (initial
&& sched_feat(START_DEBIT
))
2764 vruntime
+= sched_vslice(cfs_rq
, se
);
2766 /* sleeps up to a single latency don't count. */
2768 unsigned long thresh
= sysctl_sched_latency
;
2771 * Halve their sleep time's effect, to allow
2772 * for a gentler effect of sleepers:
2774 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2780 /* ensure we never gain time by being placed backwards. */
2781 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
2784 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
2787 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2790 * Update the normalized vruntime before updating min_vruntime
2791 * through calling update_curr().
2793 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
2794 se
->vruntime
+= cfs_rq
->min_vruntime
;
2797 * Update run-time statistics of the 'current'.
2799 update_curr(cfs_rq
);
2800 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
2801 account_entity_enqueue(cfs_rq
, se
);
2802 update_cfs_shares(cfs_rq
);
2804 if (flags
& ENQUEUE_WAKEUP
) {
2805 place_entity(cfs_rq
, se
, 0);
2806 enqueue_sleeper(cfs_rq
, se
);
2809 update_stats_enqueue(cfs_rq
, se
);
2810 check_spread(cfs_rq
, se
);
2811 if (se
!= cfs_rq
->curr
)
2812 __enqueue_entity(cfs_rq
, se
);
2815 if (cfs_rq
->nr_running
== 1) {
2816 list_add_leaf_cfs_rq(cfs_rq
);
2817 check_enqueue_throttle(cfs_rq
);
2821 static void __clear_buddies_last(struct sched_entity
*se
)
2823 for_each_sched_entity(se
) {
2824 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2825 if (cfs_rq
->last
!= se
)
2828 cfs_rq
->last
= NULL
;
2832 static void __clear_buddies_next(struct sched_entity
*se
)
2834 for_each_sched_entity(se
) {
2835 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2836 if (cfs_rq
->next
!= se
)
2839 cfs_rq
->next
= NULL
;
2843 static void __clear_buddies_skip(struct sched_entity
*se
)
2845 for_each_sched_entity(se
) {
2846 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2847 if (cfs_rq
->skip
!= se
)
2850 cfs_rq
->skip
= NULL
;
2854 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2856 if (cfs_rq
->last
== se
)
2857 __clear_buddies_last(se
);
2859 if (cfs_rq
->next
== se
)
2860 __clear_buddies_next(se
);
2862 if (cfs_rq
->skip
== se
)
2863 __clear_buddies_skip(se
);
2866 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2869 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2872 * Update run-time statistics of the 'current'.
2874 update_curr(cfs_rq
);
2875 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
2877 update_stats_dequeue(cfs_rq
, se
);
2878 if (flags
& DEQUEUE_SLEEP
) {
2879 #ifdef CONFIG_SCHEDSTATS
2880 if (entity_is_task(se
)) {
2881 struct task_struct
*tsk
= task_of(se
);
2883 if (tsk
->state
& TASK_INTERRUPTIBLE
)
2884 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
2885 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
2886 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
2891 clear_buddies(cfs_rq
, se
);
2893 if (se
!= cfs_rq
->curr
)
2894 __dequeue_entity(cfs_rq
, se
);
2896 account_entity_dequeue(cfs_rq
, se
);
2899 * Normalize the entity after updating the min_vruntime because the
2900 * update can refer to the ->curr item and we need to reflect this
2901 * movement in our normalized position.
2903 if (!(flags
& DEQUEUE_SLEEP
))
2904 se
->vruntime
-= cfs_rq
->min_vruntime
;
2906 /* return excess runtime on last dequeue */
2907 return_cfs_rq_runtime(cfs_rq
);
2909 update_min_vruntime(cfs_rq
);
2910 update_cfs_shares(cfs_rq
);
2914 * Preempt the current task with a newly woken task if needed:
2917 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2919 unsigned long ideal_runtime
, delta_exec
;
2920 struct sched_entity
*se
;
2923 ideal_runtime
= sched_slice(cfs_rq
, curr
);
2924 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
2925 if (delta_exec
> ideal_runtime
) {
2926 resched_task(rq_of(cfs_rq
)->curr
);
2928 * The current task ran long enough, ensure it doesn't get
2929 * re-elected due to buddy favours.
2931 clear_buddies(cfs_rq
, curr
);
2936 * Ensure that a task that missed wakeup preemption by a
2937 * narrow margin doesn't have to wait for a full slice.
2938 * This also mitigates buddy induced latencies under load.
2940 if (delta_exec
< sysctl_sched_min_granularity
)
2943 se
= __pick_first_entity(cfs_rq
);
2944 delta
= curr
->vruntime
- se
->vruntime
;
2949 if (delta
> ideal_runtime
)
2950 resched_task(rq_of(cfs_rq
)->curr
);
2954 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2956 /* 'current' is not kept within the tree. */
2959 * Any task has to be enqueued before it get to execute on
2960 * a CPU. So account for the time it spent waiting on the
2963 update_stats_wait_end(cfs_rq
, se
);
2964 __dequeue_entity(cfs_rq
, se
);
2967 update_stats_curr_start(cfs_rq
, se
);
2969 #ifdef CONFIG_SCHEDSTATS
2971 * Track our maximum slice length, if the CPU's load is at
2972 * least twice that of our own weight (i.e. dont track it
2973 * when there are only lesser-weight tasks around):
2975 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
2976 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
2977 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
2980 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
2984 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
2987 * Pick the next process, keeping these things in mind, in this order:
2988 * 1) keep things fair between processes/task groups
2989 * 2) pick the "next" process, since someone really wants that to run
2990 * 3) pick the "last" process, for cache locality
2991 * 4) do not run the "skip" process, if something else is available
2993 static struct sched_entity
*
2994 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2996 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
2997 struct sched_entity
*se
;
3000 * If curr is set we have to see if its left of the leftmost entity
3001 * still in the tree, provided there was anything in the tree at all.
3003 if (!left
|| (curr
&& entity_before(curr
, left
)))
3006 se
= left
; /* ideally we run the leftmost entity */
3009 * Avoid running the skip buddy, if running something else can
3010 * be done without getting too unfair.
3012 if (cfs_rq
->skip
== se
) {
3013 struct sched_entity
*second
;
3016 second
= __pick_first_entity(cfs_rq
);
3018 second
= __pick_next_entity(se
);
3019 if (!second
|| (curr
&& entity_before(curr
, second
)))
3023 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3028 * Prefer last buddy, try to return the CPU to a preempted task.
3030 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3034 * Someone really wants this to run. If it's not unfair, run it.
3036 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3039 clear_buddies(cfs_rq
, se
);
3044 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3046 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3049 * If still on the runqueue then deactivate_task()
3050 * was not called and update_curr() has to be done:
3053 update_curr(cfs_rq
);
3055 /* throttle cfs_rqs exceeding runtime */
3056 check_cfs_rq_runtime(cfs_rq
);
3058 check_spread(cfs_rq
, prev
);
3060 update_stats_wait_start(cfs_rq
, prev
);
3061 /* Put 'current' back into the tree. */
3062 __enqueue_entity(cfs_rq
, prev
);
3063 /* in !on_rq case, update occurred at dequeue */
3064 update_entity_load_avg(prev
, 1);
3066 cfs_rq
->curr
= NULL
;
3070 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3073 * Update run-time statistics of the 'current'.
3075 update_curr(cfs_rq
);
3078 * Ensure that runnable average is periodically updated.
3080 update_entity_load_avg(curr
, 1);
3081 update_cfs_rq_blocked_load(cfs_rq
, 1);
3082 update_cfs_shares(cfs_rq
);
3084 #ifdef CONFIG_SCHED_HRTICK
3086 * queued ticks are scheduled to match the slice, so don't bother
3087 * validating it and just reschedule.
3090 resched_task(rq_of(cfs_rq
)->curr
);
3094 * don't let the period tick interfere with the hrtick preemption
3096 if (!sched_feat(DOUBLE_TICK
) &&
3097 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3101 if (cfs_rq
->nr_running
> 1)
3102 check_preempt_tick(cfs_rq
, curr
);
3106 /**************************************************
3107 * CFS bandwidth control machinery
3110 #ifdef CONFIG_CFS_BANDWIDTH
3112 #ifdef HAVE_JUMP_LABEL
3113 static struct static_key __cfs_bandwidth_used
;
3115 static inline bool cfs_bandwidth_used(void)
3117 return static_key_false(&__cfs_bandwidth_used
);
3120 void cfs_bandwidth_usage_inc(void)
3122 static_key_slow_inc(&__cfs_bandwidth_used
);
3125 void cfs_bandwidth_usage_dec(void)
3127 static_key_slow_dec(&__cfs_bandwidth_used
);
3129 #else /* HAVE_JUMP_LABEL */
3130 static bool cfs_bandwidth_used(void)
3135 void cfs_bandwidth_usage_inc(void) {}
3136 void cfs_bandwidth_usage_dec(void) {}
3137 #endif /* HAVE_JUMP_LABEL */
3140 * default period for cfs group bandwidth.
3141 * default: 0.1s, units: nanoseconds
3143 static inline u64
default_cfs_period(void)
3145 return 100000000ULL;
3148 static inline u64
sched_cfs_bandwidth_slice(void)
3150 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3154 * Replenish runtime according to assigned quota and update expiration time.
3155 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3156 * additional synchronization around rq->lock.
3158 * requires cfs_b->lock
3160 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3164 if (cfs_b
->quota
== RUNTIME_INF
)
3167 now
= sched_clock_cpu(smp_processor_id());
3168 cfs_b
->runtime
= cfs_b
->quota
;
3169 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3172 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3174 return &tg
->cfs_bandwidth
;
3177 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3178 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3180 if (unlikely(cfs_rq
->throttle_count
))
3181 return cfs_rq
->throttled_clock_task
;
3183 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3186 /* returns 0 on failure to allocate runtime */
3187 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3189 struct task_group
*tg
= cfs_rq
->tg
;
3190 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3191 u64 amount
= 0, min_amount
, expires
;
3193 /* note: this is a positive sum as runtime_remaining <= 0 */
3194 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3196 raw_spin_lock(&cfs_b
->lock
);
3197 if (cfs_b
->quota
== RUNTIME_INF
)
3198 amount
= min_amount
;
3201 * If the bandwidth pool has become inactive, then at least one
3202 * period must have elapsed since the last consumption.
3203 * Refresh the global state and ensure bandwidth timer becomes
3206 if (!cfs_b
->timer_active
) {
3207 __refill_cfs_bandwidth_runtime(cfs_b
);
3208 __start_cfs_bandwidth(cfs_b
, false);
3211 if (cfs_b
->runtime
> 0) {
3212 amount
= min(cfs_b
->runtime
, min_amount
);
3213 cfs_b
->runtime
-= amount
;
3217 expires
= cfs_b
->runtime_expires
;
3218 raw_spin_unlock(&cfs_b
->lock
);
3220 cfs_rq
->runtime_remaining
+= amount
;
3222 * we may have advanced our local expiration to account for allowed
3223 * spread between our sched_clock and the one on which runtime was
3226 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3227 cfs_rq
->runtime_expires
= expires
;
3229 return cfs_rq
->runtime_remaining
> 0;
3233 * Note: This depends on the synchronization provided by sched_clock and the
3234 * fact that rq->clock snapshots this value.
3236 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3238 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3240 /* if the deadline is ahead of our clock, nothing to do */
3241 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3244 if (cfs_rq
->runtime_remaining
< 0)
3248 * If the local deadline has passed we have to consider the
3249 * possibility that our sched_clock is 'fast' and the global deadline
3250 * has not truly expired.
3252 * Fortunately we can check determine whether this the case by checking
3253 * whether the global deadline has advanced. It is valid to compare
3254 * cfs_b->runtime_expires without any locks since we only care about
3255 * exact equality, so a partial write will still work.
3258 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
3259 /* extend local deadline, drift is bounded above by 2 ticks */
3260 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3262 /* global deadline is ahead, expiration has passed */
3263 cfs_rq
->runtime_remaining
= 0;
3267 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3269 /* dock delta_exec before expiring quota (as it could span periods) */
3270 cfs_rq
->runtime_remaining
-= delta_exec
;
3271 expire_cfs_rq_runtime(cfs_rq
);
3273 if (likely(cfs_rq
->runtime_remaining
> 0))
3277 * if we're unable to extend our runtime we resched so that the active
3278 * hierarchy can be throttled
3280 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3281 resched_task(rq_of(cfs_rq
)->curr
);
3284 static __always_inline
3285 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3287 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3290 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3293 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3295 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3298 /* check whether cfs_rq, or any parent, is throttled */
3299 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3301 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3305 * Ensure that neither of the group entities corresponding to src_cpu or
3306 * dest_cpu are members of a throttled hierarchy when performing group
3307 * load-balance operations.
3309 static inline int throttled_lb_pair(struct task_group
*tg
,
3310 int src_cpu
, int dest_cpu
)
3312 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3314 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3315 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3317 return throttled_hierarchy(src_cfs_rq
) ||
3318 throttled_hierarchy(dest_cfs_rq
);
3321 /* updated child weight may affect parent so we have to do this bottom up */
3322 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3324 struct rq
*rq
= data
;
3325 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3327 cfs_rq
->throttle_count
--;
3329 if (!cfs_rq
->throttle_count
) {
3330 /* adjust cfs_rq_clock_task() */
3331 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3332 cfs_rq
->throttled_clock_task
;
3339 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3341 struct rq
*rq
= data
;
3342 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3344 /* group is entering throttled state, stop time */
3345 if (!cfs_rq
->throttle_count
)
3346 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3347 cfs_rq
->throttle_count
++;
3352 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3354 struct rq
*rq
= rq_of(cfs_rq
);
3355 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3356 struct sched_entity
*se
;
3357 long task_delta
, dequeue
= 1;
3359 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3361 /* freeze hierarchy runnable averages while throttled */
3363 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3366 task_delta
= cfs_rq
->h_nr_running
;
3367 for_each_sched_entity(se
) {
3368 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3369 /* throttled entity or throttle-on-deactivate */
3374 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3375 qcfs_rq
->h_nr_running
-= task_delta
;
3377 if (qcfs_rq
->load
.weight
)
3382 sub_nr_running(rq
, task_delta
);
3384 cfs_rq
->throttled
= 1;
3385 cfs_rq
->throttled_clock
= rq_clock(rq
);
3386 raw_spin_lock(&cfs_b
->lock
);
3388 * Add to the _head_ of the list, so that an already-started
3389 * distribute_cfs_runtime will not see us
3391 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3392 if (!cfs_b
->timer_active
)
3393 __start_cfs_bandwidth(cfs_b
, false);
3394 raw_spin_unlock(&cfs_b
->lock
);
3397 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3399 struct rq
*rq
= rq_of(cfs_rq
);
3400 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3401 struct sched_entity
*se
;
3405 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3407 cfs_rq
->throttled
= 0;
3409 update_rq_clock(rq
);
3411 raw_spin_lock(&cfs_b
->lock
);
3412 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3413 list_del_rcu(&cfs_rq
->throttled_list
);
3414 raw_spin_unlock(&cfs_b
->lock
);
3416 /* update hierarchical throttle state */
3417 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3419 if (!cfs_rq
->load
.weight
)
3422 task_delta
= cfs_rq
->h_nr_running
;
3423 for_each_sched_entity(se
) {
3427 cfs_rq
= cfs_rq_of(se
);
3429 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3430 cfs_rq
->h_nr_running
+= task_delta
;
3432 if (cfs_rq_throttled(cfs_rq
))
3437 add_nr_running(rq
, task_delta
);
3439 /* determine whether we need to wake up potentially idle cpu */
3440 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3441 resched_task(rq
->curr
);
3444 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3445 u64 remaining
, u64 expires
)
3447 struct cfs_rq
*cfs_rq
;
3449 u64 starting_runtime
= remaining
;
3452 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3454 struct rq
*rq
= rq_of(cfs_rq
);
3456 raw_spin_lock(&rq
->lock
);
3457 if (!cfs_rq_throttled(cfs_rq
))
3460 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3461 if (runtime
> remaining
)
3462 runtime
= remaining
;
3463 remaining
-= runtime
;
3465 cfs_rq
->runtime_remaining
+= runtime
;
3466 cfs_rq
->runtime_expires
= expires
;
3468 /* we check whether we're throttled above */
3469 if (cfs_rq
->runtime_remaining
> 0)
3470 unthrottle_cfs_rq(cfs_rq
);
3473 raw_spin_unlock(&rq
->lock
);
3480 return starting_runtime
- remaining
;
3484 * Responsible for refilling a task_group's bandwidth and unthrottling its
3485 * cfs_rqs as appropriate. If there has been no activity within the last
3486 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3487 * used to track this state.
3489 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3491 u64 runtime
, runtime_expires
;
3494 /* no need to continue the timer with no bandwidth constraint */
3495 if (cfs_b
->quota
== RUNTIME_INF
)
3496 goto out_deactivate
;
3498 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3499 cfs_b
->nr_periods
+= overrun
;
3502 * idle depends on !throttled (for the case of a large deficit), and if
3503 * we're going inactive then everything else can be deferred
3505 if (cfs_b
->idle
&& !throttled
)
3506 goto out_deactivate
;
3509 * if we have relooped after returning idle once, we need to update our
3510 * status as actually running, so that other cpus doing
3511 * __start_cfs_bandwidth will stop trying to cancel us.
3513 cfs_b
->timer_active
= 1;
3515 __refill_cfs_bandwidth_runtime(cfs_b
);
3518 /* mark as potentially idle for the upcoming period */
3523 /* account preceding periods in which throttling occurred */
3524 cfs_b
->nr_throttled
+= overrun
;
3526 runtime_expires
= cfs_b
->runtime_expires
;
3529 * This check is repeated as we are holding onto the new bandwidth while
3530 * we unthrottle. This can potentially race with an unthrottled group
3531 * trying to acquire new bandwidth from the global pool. This can result
3532 * in us over-using our runtime if it is all used during this loop, but
3533 * only by limited amounts in that extreme case.
3535 while (throttled
&& cfs_b
->runtime
> 0) {
3536 runtime
= cfs_b
->runtime
;
3537 raw_spin_unlock(&cfs_b
->lock
);
3538 /* we can't nest cfs_b->lock while distributing bandwidth */
3539 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3541 raw_spin_lock(&cfs_b
->lock
);
3543 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3545 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3549 * While we are ensured activity in the period following an
3550 * unthrottle, this also covers the case in which the new bandwidth is
3551 * insufficient to cover the existing bandwidth deficit. (Forcing the
3552 * timer to remain active while there are any throttled entities.)
3559 cfs_b
->timer_active
= 0;
3563 /* a cfs_rq won't donate quota below this amount */
3564 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3565 /* minimum remaining period time to redistribute slack quota */
3566 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3567 /* how long we wait to gather additional slack before distributing */
3568 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3571 * Are we near the end of the current quota period?
3573 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3574 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3575 * migrate_hrtimers, base is never cleared, so we are fine.
3577 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3579 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3582 /* if the call-back is running a quota refresh is already occurring */
3583 if (hrtimer_callback_running(refresh_timer
))
3586 /* is a quota refresh about to occur? */
3587 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3588 if (remaining
< min_expire
)
3594 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3596 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3598 /* if there's a quota refresh soon don't bother with slack */
3599 if (runtime_refresh_within(cfs_b
, min_left
))
3602 start_bandwidth_timer(&cfs_b
->slack_timer
,
3603 ns_to_ktime(cfs_bandwidth_slack_period
));
3606 /* we know any runtime found here is valid as update_curr() precedes return */
3607 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3609 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3610 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3612 if (slack_runtime
<= 0)
3615 raw_spin_lock(&cfs_b
->lock
);
3616 if (cfs_b
->quota
!= RUNTIME_INF
&&
3617 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3618 cfs_b
->runtime
+= slack_runtime
;
3620 /* we are under rq->lock, defer unthrottling using a timer */
3621 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3622 !list_empty(&cfs_b
->throttled_cfs_rq
))
3623 start_cfs_slack_bandwidth(cfs_b
);
3625 raw_spin_unlock(&cfs_b
->lock
);
3627 /* even if it's not valid for return we don't want to try again */
3628 cfs_rq
->runtime_remaining
-= slack_runtime
;
3631 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3633 if (!cfs_bandwidth_used())
3636 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3639 __return_cfs_rq_runtime(cfs_rq
);
3643 * This is done with a timer (instead of inline with bandwidth return) since
3644 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3646 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3648 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3651 /* confirm we're still not at a refresh boundary */
3652 raw_spin_lock(&cfs_b
->lock
);
3653 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
3654 raw_spin_unlock(&cfs_b
->lock
);
3658 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
3659 runtime
= cfs_b
->runtime
;
3661 expires
= cfs_b
->runtime_expires
;
3662 raw_spin_unlock(&cfs_b
->lock
);
3667 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3669 raw_spin_lock(&cfs_b
->lock
);
3670 if (expires
== cfs_b
->runtime_expires
)
3671 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3672 raw_spin_unlock(&cfs_b
->lock
);
3676 * When a group wakes up we want to make sure that its quota is not already
3677 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3678 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3680 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3682 if (!cfs_bandwidth_used())
3685 /* an active group must be handled by the update_curr()->put() path */
3686 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3689 /* ensure the group is not already throttled */
3690 if (cfs_rq_throttled(cfs_rq
))
3693 /* update runtime allocation */
3694 account_cfs_rq_runtime(cfs_rq
, 0);
3695 if (cfs_rq
->runtime_remaining
<= 0)
3696 throttle_cfs_rq(cfs_rq
);
3699 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3700 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3702 if (!cfs_bandwidth_used())
3705 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3709 * it's possible for a throttled entity to be forced into a running
3710 * state (e.g. set_curr_task), in this case we're finished.
3712 if (cfs_rq_throttled(cfs_rq
))
3715 throttle_cfs_rq(cfs_rq
);
3719 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3721 struct cfs_bandwidth
*cfs_b
=
3722 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3723 do_sched_cfs_slack_timer(cfs_b
);
3725 return HRTIMER_NORESTART
;
3728 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3730 struct cfs_bandwidth
*cfs_b
=
3731 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3736 raw_spin_lock(&cfs_b
->lock
);
3738 now
= hrtimer_cb_get_time(timer
);
3739 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3744 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3746 raw_spin_unlock(&cfs_b
->lock
);
3748 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3751 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3753 raw_spin_lock_init(&cfs_b
->lock
);
3755 cfs_b
->quota
= RUNTIME_INF
;
3756 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3758 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3759 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3760 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3761 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3762 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3765 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3767 cfs_rq
->runtime_enabled
= 0;
3768 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3771 /* requires cfs_b->lock, may release to reprogram timer */
3772 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
, bool force
)
3775 * The timer may be active because we're trying to set a new bandwidth
3776 * period or because we're racing with the tear-down path
3777 * (timer_active==0 becomes visible before the hrtimer call-back
3778 * terminates). In either case we ensure that it's re-programmed
3780 while (unlikely(hrtimer_active(&cfs_b
->period_timer
)) &&
3781 hrtimer_try_to_cancel(&cfs_b
->period_timer
) < 0) {
3782 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3783 raw_spin_unlock(&cfs_b
->lock
);
3785 raw_spin_lock(&cfs_b
->lock
);
3786 /* if someone else restarted the timer then we're done */
3787 if (!force
&& cfs_b
->timer_active
)
3791 cfs_b
->timer_active
= 1;
3792 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
3795 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3797 hrtimer_cancel(&cfs_b
->period_timer
);
3798 hrtimer_cancel(&cfs_b
->slack_timer
);
3801 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
3803 struct cfs_rq
*cfs_rq
;
3805 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3806 struct cfs_bandwidth
*cfs_b
= &cfs_rq
->tg
->cfs_bandwidth
;
3808 raw_spin_lock(&cfs_b
->lock
);
3809 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
3810 raw_spin_unlock(&cfs_b
->lock
);
3814 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
3816 struct cfs_rq
*cfs_rq
;
3818 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3819 if (!cfs_rq
->runtime_enabled
)
3823 * clock_task is not advancing so we just need to make sure
3824 * there's some valid quota amount
3826 cfs_rq
->runtime_remaining
= 1;
3828 * Offline rq is schedulable till cpu is completely disabled
3829 * in take_cpu_down(), so we prevent new cfs throttling here.
3831 cfs_rq
->runtime_enabled
= 0;
3833 if (cfs_rq_throttled(cfs_rq
))
3834 unthrottle_cfs_rq(cfs_rq
);
3838 #else /* CONFIG_CFS_BANDWIDTH */
3839 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3841 return rq_clock_task(rq_of(cfs_rq
));
3844 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
3845 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
3846 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
3847 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3849 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3854 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3859 static inline int throttled_lb_pair(struct task_group
*tg
,
3860 int src_cpu
, int dest_cpu
)
3865 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3867 #ifdef CONFIG_FAIR_GROUP_SCHED
3868 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3871 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3875 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3876 static inline void update_runtime_enabled(struct rq
*rq
) {}
3877 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
3879 #endif /* CONFIG_CFS_BANDWIDTH */
3881 /**************************************************
3882 * CFS operations on tasks:
3885 #ifdef CONFIG_SCHED_HRTICK
3886 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3888 struct sched_entity
*se
= &p
->se
;
3889 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3891 WARN_ON(task_rq(p
) != rq
);
3893 if (cfs_rq
->nr_running
> 1) {
3894 u64 slice
= sched_slice(cfs_rq
, se
);
3895 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
3896 s64 delta
= slice
- ran
;
3905 * Don't schedule slices shorter than 10000ns, that just
3906 * doesn't make sense. Rely on vruntime for fairness.
3909 delta
= max_t(s64
, 10000LL, delta
);
3911 hrtick_start(rq
, delta
);
3916 * called from enqueue/dequeue and updates the hrtick when the
3917 * current task is from our class and nr_running is low enough
3920 static void hrtick_update(struct rq
*rq
)
3922 struct task_struct
*curr
= rq
->curr
;
3924 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
3927 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
3928 hrtick_start_fair(rq
, curr
);
3930 #else /* !CONFIG_SCHED_HRTICK */
3932 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3936 static inline void hrtick_update(struct rq
*rq
)
3942 * The enqueue_task method is called before nr_running is
3943 * increased. Here we update the fair scheduling stats and
3944 * then put the task into the rbtree:
3947 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3949 struct cfs_rq
*cfs_rq
;
3950 struct sched_entity
*se
= &p
->se
;
3952 for_each_sched_entity(se
) {
3955 cfs_rq
= cfs_rq_of(se
);
3956 enqueue_entity(cfs_rq
, se
, flags
);
3959 * end evaluation on encountering a throttled cfs_rq
3961 * note: in the case of encountering a throttled cfs_rq we will
3962 * post the final h_nr_running increment below.
3964 if (cfs_rq_throttled(cfs_rq
))
3966 cfs_rq
->h_nr_running
++;
3968 flags
= ENQUEUE_WAKEUP
;
3971 for_each_sched_entity(se
) {
3972 cfs_rq
= cfs_rq_of(se
);
3973 cfs_rq
->h_nr_running
++;
3975 if (cfs_rq_throttled(cfs_rq
))
3978 update_cfs_shares(cfs_rq
);
3979 update_entity_load_avg(se
, 1);
3983 update_rq_runnable_avg(rq
, rq
->nr_running
);
3984 add_nr_running(rq
, 1);
3989 static void set_next_buddy(struct sched_entity
*se
);
3992 * The dequeue_task method is called before nr_running is
3993 * decreased. We remove the task from the rbtree and
3994 * update the fair scheduling stats:
3996 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3998 struct cfs_rq
*cfs_rq
;
3999 struct sched_entity
*se
= &p
->se
;
4000 int task_sleep
= flags
& DEQUEUE_SLEEP
;
4002 for_each_sched_entity(se
) {
4003 cfs_rq
= cfs_rq_of(se
);
4004 dequeue_entity(cfs_rq
, se
, flags
);
4007 * end evaluation on encountering a throttled cfs_rq
4009 * note: in the case of encountering a throttled cfs_rq we will
4010 * post the final h_nr_running decrement below.
4012 if (cfs_rq_throttled(cfs_rq
))
4014 cfs_rq
->h_nr_running
--;
4016 /* Don't dequeue parent if it has other entities besides us */
4017 if (cfs_rq
->load
.weight
) {
4019 * Bias pick_next to pick a task from this cfs_rq, as
4020 * p is sleeping when it is within its sched_slice.
4022 if (task_sleep
&& parent_entity(se
))
4023 set_next_buddy(parent_entity(se
));
4025 /* avoid re-evaluating load for this entity */
4026 se
= parent_entity(se
);
4029 flags
|= DEQUEUE_SLEEP
;
4032 for_each_sched_entity(se
) {
4033 cfs_rq
= cfs_rq_of(se
);
4034 cfs_rq
->h_nr_running
--;
4036 if (cfs_rq_throttled(cfs_rq
))
4039 update_cfs_shares(cfs_rq
);
4040 update_entity_load_avg(se
, 1);
4044 sub_nr_running(rq
, 1);
4045 update_rq_runnable_avg(rq
, 1);
4051 /* Used instead of source_load when we know the type == 0 */
4052 static unsigned long weighted_cpuload(const int cpu
)
4054 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
4058 * Return a low guess at the load of a migration-source cpu weighted
4059 * according to the scheduling class and "nice" value.
4061 * We want to under-estimate the load of migration sources, to
4062 * balance conservatively.
4064 static unsigned long source_load(int cpu
, int type
)
4066 struct rq
*rq
= cpu_rq(cpu
);
4067 unsigned long total
= weighted_cpuload(cpu
);
4069 if (type
== 0 || !sched_feat(LB_BIAS
))
4072 return min(rq
->cpu_load
[type
-1], total
);
4076 * Return a high guess at the load of a migration-target cpu weighted
4077 * according to the scheduling class and "nice" value.
4079 static unsigned long target_load(int cpu
, int type
)
4081 struct rq
*rq
= cpu_rq(cpu
);
4082 unsigned long total
= weighted_cpuload(cpu
);
4084 if (type
== 0 || !sched_feat(LB_BIAS
))
4087 return max(rq
->cpu_load
[type
-1], total
);
4090 static unsigned long capacity_of(int cpu
)
4092 return cpu_rq(cpu
)->cpu_capacity
;
4095 static unsigned long cpu_avg_load_per_task(int cpu
)
4097 struct rq
*rq
= cpu_rq(cpu
);
4098 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
4099 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
4102 return load_avg
/ nr_running
;
4107 static void record_wakee(struct task_struct
*p
)
4110 * Rough decay (wiping) for cost saving, don't worry
4111 * about the boundary, really active task won't care
4114 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
4115 current
->wakee_flips
>>= 1;
4116 current
->wakee_flip_decay_ts
= jiffies
;
4119 if (current
->last_wakee
!= p
) {
4120 current
->last_wakee
= p
;
4121 current
->wakee_flips
++;
4125 static void task_waking_fair(struct task_struct
*p
)
4127 struct sched_entity
*se
= &p
->se
;
4128 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4131 #ifndef CONFIG_64BIT
4132 u64 min_vruntime_copy
;
4135 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
4137 min_vruntime
= cfs_rq
->min_vruntime
;
4138 } while (min_vruntime
!= min_vruntime_copy
);
4140 min_vruntime
= cfs_rq
->min_vruntime
;
4143 se
->vruntime
-= min_vruntime
;
4147 #ifdef CONFIG_FAIR_GROUP_SCHED
4149 * effective_load() calculates the load change as seen from the root_task_group
4151 * Adding load to a group doesn't make a group heavier, but can cause movement
4152 * of group shares between cpus. Assuming the shares were perfectly aligned one
4153 * can calculate the shift in shares.
4155 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4156 * on this @cpu and results in a total addition (subtraction) of @wg to the
4157 * total group weight.
4159 * Given a runqueue weight distribution (rw_i) we can compute a shares
4160 * distribution (s_i) using:
4162 * s_i = rw_i / \Sum rw_j (1)
4164 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4165 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4166 * shares distribution (s_i):
4168 * rw_i = { 2, 4, 1, 0 }
4169 * s_i = { 2/7, 4/7, 1/7, 0 }
4171 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4172 * task used to run on and the CPU the waker is running on), we need to
4173 * compute the effect of waking a task on either CPU and, in case of a sync
4174 * wakeup, compute the effect of the current task going to sleep.
4176 * So for a change of @wl to the local @cpu with an overall group weight change
4177 * of @wl we can compute the new shares distribution (s'_i) using:
4179 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4181 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4182 * differences in waking a task to CPU 0. The additional task changes the
4183 * weight and shares distributions like:
4185 * rw'_i = { 3, 4, 1, 0 }
4186 * s'_i = { 3/8, 4/8, 1/8, 0 }
4188 * We can then compute the difference in effective weight by using:
4190 * dw_i = S * (s'_i - s_i) (3)
4192 * Where 'S' is the group weight as seen by its parent.
4194 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4195 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4196 * 4/7) times the weight of the group.
4198 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4200 struct sched_entity
*se
= tg
->se
[cpu
];
4202 if (!tg
->parent
) /* the trivial, non-cgroup case */
4205 for_each_sched_entity(se
) {
4211 * W = @wg + \Sum rw_j
4213 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4218 w
= se
->my_q
->load
.weight
+ wl
;
4221 * wl = S * s'_i; see (2)
4224 wl
= (w
* tg
->shares
) / W
;
4229 * Per the above, wl is the new se->load.weight value; since
4230 * those are clipped to [MIN_SHARES, ...) do so now. See
4231 * calc_cfs_shares().
4233 if (wl
< MIN_SHARES
)
4237 * wl = dw_i = S * (s'_i - s_i); see (3)
4239 wl
-= se
->load
.weight
;
4242 * Recursively apply this logic to all parent groups to compute
4243 * the final effective load change on the root group. Since
4244 * only the @tg group gets extra weight, all parent groups can
4245 * only redistribute existing shares. @wl is the shift in shares
4246 * resulting from this level per the above.
4255 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4262 static int wake_wide(struct task_struct
*p
)
4264 int factor
= this_cpu_read(sd_llc_size
);
4267 * Yeah, it's the switching-frequency, could means many wakee or
4268 * rapidly switch, use factor here will just help to automatically
4269 * adjust the loose-degree, so bigger node will lead to more pull.
4271 if (p
->wakee_flips
> factor
) {
4273 * wakee is somewhat hot, it needs certain amount of cpu
4274 * resource, so if waker is far more hot, prefer to leave
4277 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
4284 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4286 s64 this_load
, load
;
4287 int idx
, this_cpu
, prev_cpu
;
4288 unsigned long tl_per_task
;
4289 struct task_group
*tg
;
4290 unsigned long weight
;
4294 * If we wake multiple tasks be careful to not bounce
4295 * ourselves around too much.
4301 this_cpu
= smp_processor_id();
4302 prev_cpu
= task_cpu(p
);
4303 load
= source_load(prev_cpu
, idx
);
4304 this_load
= target_load(this_cpu
, idx
);
4307 * If sync wakeup then subtract the (maximum possible)
4308 * effect of the currently running task from the load
4309 * of the current CPU:
4312 tg
= task_group(current
);
4313 weight
= current
->se
.load
.weight
;
4315 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4316 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4320 weight
= p
->se
.load
.weight
;
4323 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4324 * due to the sync cause above having dropped this_load to 0, we'll
4325 * always have an imbalance, but there's really nothing you can do
4326 * about that, so that's good too.
4328 * Otherwise check if either cpus are near enough in load to allow this
4329 * task to be woken on this_cpu.
4331 if (this_load
> 0) {
4332 s64 this_eff_load
, prev_eff_load
;
4334 this_eff_load
= 100;
4335 this_eff_load
*= capacity_of(prev_cpu
);
4336 this_eff_load
*= this_load
+
4337 effective_load(tg
, this_cpu
, weight
, weight
);
4339 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4340 prev_eff_load
*= capacity_of(this_cpu
);
4341 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4343 balanced
= this_eff_load
<= prev_eff_load
;
4348 * If the currently running task will sleep within
4349 * a reasonable amount of time then attract this newly
4352 if (sync
&& balanced
)
4355 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4356 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
4359 (this_load
<= load
&&
4360 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
4362 * This domain has SD_WAKE_AFFINE and
4363 * p is cache cold in this domain, and
4364 * there is no bad imbalance.
4366 schedstat_inc(sd
, ttwu_move_affine
);
4367 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4375 * find_idlest_group finds and returns the least busy CPU group within the
4378 static struct sched_group
*
4379 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4380 int this_cpu
, int sd_flag
)
4382 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4383 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4384 int load_idx
= sd
->forkexec_idx
;
4385 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4387 if (sd_flag
& SD_BALANCE_WAKE
)
4388 load_idx
= sd
->wake_idx
;
4391 unsigned long load
, avg_load
;
4395 /* Skip over this group if it has no CPUs allowed */
4396 if (!cpumask_intersects(sched_group_cpus(group
),
4397 tsk_cpus_allowed(p
)))
4400 local_group
= cpumask_test_cpu(this_cpu
,
4401 sched_group_cpus(group
));
4403 /* Tally up the load of all CPUs in the group */
4406 for_each_cpu(i
, sched_group_cpus(group
)) {
4407 /* Bias balancing toward cpus of our domain */
4409 load
= source_load(i
, load_idx
);
4411 load
= target_load(i
, load_idx
);
4416 /* Adjust by relative CPU capacity of the group */
4417 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) / group
->sgc
->capacity
;
4420 this_load
= avg_load
;
4421 } else if (avg_load
< min_load
) {
4422 min_load
= avg_load
;
4425 } while (group
= group
->next
, group
!= sd
->groups
);
4427 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4433 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4436 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4438 unsigned long load
, min_load
= ULONG_MAX
;
4442 /* Traverse only the allowed CPUs */
4443 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4444 load
= weighted_cpuload(i
);
4446 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4456 * Try and locate an idle CPU in the sched_domain.
4458 static int select_idle_sibling(struct task_struct
*p
, int target
)
4460 struct sched_domain
*sd
;
4461 struct sched_group
*sg
;
4462 int i
= task_cpu(p
);
4464 if (idle_cpu(target
))
4468 * If the prevous cpu is cache affine and idle, don't be stupid.
4470 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4474 * Otherwise, iterate the domains and find an elegible idle cpu.
4476 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4477 for_each_lower_domain(sd
) {
4480 if (!cpumask_intersects(sched_group_cpus(sg
),
4481 tsk_cpus_allowed(p
)))
4484 for_each_cpu(i
, sched_group_cpus(sg
)) {
4485 if (i
== target
|| !idle_cpu(i
))
4489 target
= cpumask_first_and(sched_group_cpus(sg
),
4490 tsk_cpus_allowed(p
));
4494 } while (sg
!= sd
->groups
);
4501 * select_task_rq_fair: Select target runqueue for the waking task in domains
4502 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4503 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4505 * Balances load by selecting the idlest cpu in the idlest group, or under
4506 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4508 * Returns the target cpu number.
4510 * preempt must be disabled.
4513 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4515 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4516 int cpu
= smp_processor_id();
4518 int want_affine
= 0;
4519 int sync
= wake_flags
& WF_SYNC
;
4521 if (p
->nr_cpus_allowed
== 1)
4524 if (sd_flag
& SD_BALANCE_WAKE
) {
4525 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
4531 for_each_domain(cpu
, tmp
) {
4532 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4536 * If both cpu and prev_cpu are part of this domain,
4537 * cpu is a valid SD_WAKE_AFFINE target.
4539 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4540 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4545 if (tmp
->flags
& sd_flag
)
4549 if (affine_sd
&& cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4552 if (sd_flag
& SD_BALANCE_WAKE
) {
4553 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4558 struct sched_group
*group
;
4561 if (!(sd
->flags
& sd_flag
)) {
4566 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
4572 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4573 if (new_cpu
== -1 || new_cpu
== cpu
) {
4574 /* Now try balancing at a lower domain level of cpu */
4579 /* Now try balancing at a lower domain level of new_cpu */
4581 weight
= sd
->span_weight
;
4583 for_each_domain(cpu
, tmp
) {
4584 if (weight
<= tmp
->span_weight
)
4586 if (tmp
->flags
& sd_flag
)
4589 /* while loop will break here if sd == NULL */
4598 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4599 * cfs_rq_of(p) references at time of call are still valid and identify the
4600 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4601 * other assumptions, including the state of rq->lock, should be made.
4604 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4606 struct sched_entity
*se
= &p
->se
;
4607 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4610 * Load tracking: accumulate removed load so that it can be processed
4611 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4612 * to blocked load iff they have a positive decay-count. It can never
4613 * be negative here since on-rq tasks have decay-count == 0.
4615 if (se
->avg
.decay_count
) {
4616 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4617 atomic_long_add(se
->avg
.load_avg_contrib
,
4618 &cfs_rq
->removed_load
);
4621 /* We have migrated, no longer consider this task hot */
4624 #endif /* CONFIG_SMP */
4626 static unsigned long
4627 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4629 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4632 * Since its curr running now, convert the gran from real-time
4633 * to virtual-time in his units.
4635 * By using 'se' instead of 'curr' we penalize light tasks, so
4636 * they get preempted easier. That is, if 'se' < 'curr' then
4637 * the resulting gran will be larger, therefore penalizing the
4638 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4639 * be smaller, again penalizing the lighter task.
4641 * This is especially important for buddies when the leftmost
4642 * task is higher priority than the buddy.
4644 return calc_delta_fair(gran
, se
);
4648 * Should 'se' preempt 'curr'.
4662 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4664 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4669 gran
= wakeup_gran(curr
, se
);
4676 static void set_last_buddy(struct sched_entity
*se
)
4678 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4681 for_each_sched_entity(se
)
4682 cfs_rq_of(se
)->last
= se
;
4685 static void set_next_buddy(struct sched_entity
*se
)
4687 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4690 for_each_sched_entity(se
)
4691 cfs_rq_of(se
)->next
= se
;
4694 static void set_skip_buddy(struct sched_entity
*se
)
4696 for_each_sched_entity(se
)
4697 cfs_rq_of(se
)->skip
= se
;
4701 * Preempt the current task with a newly woken task if needed:
4703 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4705 struct task_struct
*curr
= rq
->curr
;
4706 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4707 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4708 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4709 int next_buddy_marked
= 0;
4711 if (unlikely(se
== pse
))
4715 * This is possible from callers such as move_task(), in which we
4716 * unconditionally check_prempt_curr() after an enqueue (which may have
4717 * lead to a throttle). This both saves work and prevents false
4718 * next-buddy nomination below.
4720 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4723 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4724 set_next_buddy(pse
);
4725 next_buddy_marked
= 1;
4729 * We can come here with TIF_NEED_RESCHED already set from new task
4732 * Note: this also catches the edge-case of curr being in a throttled
4733 * group (e.g. via set_curr_task), since update_curr() (in the
4734 * enqueue of curr) will have resulted in resched being set. This
4735 * prevents us from potentially nominating it as a false LAST_BUDDY
4738 if (test_tsk_need_resched(curr
))
4741 /* Idle tasks are by definition preempted by non-idle tasks. */
4742 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4743 likely(p
->policy
!= SCHED_IDLE
))
4747 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4748 * is driven by the tick):
4750 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4753 find_matching_se(&se
, &pse
);
4754 update_curr(cfs_rq_of(se
));
4756 if (wakeup_preempt_entity(se
, pse
) == 1) {
4758 * Bias pick_next to pick the sched entity that is
4759 * triggering this preemption.
4761 if (!next_buddy_marked
)
4762 set_next_buddy(pse
);
4771 * Only set the backward buddy when the current task is still
4772 * on the rq. This can happen when a wakeup gets interleaved
4773 * with schedule on the ->pre_schedule() or idle_balance()
4774 * point, either of which can * drop the rq lock.
4776 * Also, during early boot the idle thread is in the fair class,
4777 * for obvious reasons its a bad idea to schedule back to it.
4779 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4782 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
4786 static struct task_struct
*
4787 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4789 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
4790 struct sched_entity
*se
;
4791 struct task_struct
*p
;
4795 #ifdef CONFIG_FAIR_GROUP_SCHED
4796 if (!cfs_rq
->nr_running
)
4799 if (prev
->sched_class
!= &fair_sched_class
)
4803 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4804 * likely that a next task is from the same cgroup as the current.
4806 * Therefore attempt to avoid putting and setting the entire cgroup
4807 * hierarchy, only change the part that actually changes.
4811 struct sched_entity
*curr
= cfs_rq
->curr
;
4814 * Since we got here without doing put_prev_entity() we also
4815 * have to consider cfs_rq->curr. If it is still a runnable
4816 * entity, update_curr() will update its vruntime, otherwise
4817 * forget we've ever seen it.
4819 if (curr
&& curr
->on_rq
)
4820 update_curr(cfs_rq
);
4825 * This call to check_cfs_rq_runtime() will do the throttle and
4826 * dequeue its entity in the parent(s). Therefore the 'simple'
4827 * nr_running test will indeed be correct.
4829 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
4832 se
= pick_next_entity(cfs_rq
, curr
);
4833 cfs_rq
= group_cfs_rq(se
);
4839 * Since we haven't yet done put_prev_entity and if the selected task
4840 * is a different task than we started out with, try and touch the
4841 * least amount of cfs_rqs.
4844 struct sched_entity
*pse
= &prev
->se
;
4846 while (!(cfs_rq
= is_same_group(se
, pse
))) {
4847 int se_depth
= se
->depth
;
4848 int pse_depth
= pse
->depth
;
4850 if (se_depth
<= pse_depth
) {
4851 put_prev_entity(cfs_rq_of(pse
), pse
);
4852 pse
= parent_entity(pse
);
4854 if (se_depth
>= pse_depth
) {
4855 set_next_entity(cfs_rq_of(se
), se
);
4856 se
= parent_entity(se
);
4860 put_prev_entity(cfs_rq
, pse
);
4861 set_next_entity(cfs_rq
, se
);
4864 if (hrtick_enabled(rq
))
4865 hrtick_start_fair(rq
, p
);
4872 if (!cfs_rq
->nr_running
)
4875 put_prev_task(rq
, prev
);
4878 se
= pick_next_entity(cfs_rq
, NULL
);
4879 set_next_entity(cfs_rq
, se
);
4880 cfs_rq
= group_cfs_rq(se
);
4885 if (hrtick_enabled(rq
))
4886 hrtick_start_fair(rq
, p
);
4891 new_tasks
= idle_balance(rq
);
4893 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4894 * possible for any higher priority task to appear. In that case we
4895 * must re-start the pick_next_entity() loop.
4907 * Account for a descheduled task:
4909 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4911 struct sched_entity
*se
= &prev
->se
;
4912 struct cfs_rq
*cfs_rq
;
4914 for_each_sched_entity(se
) {
4915 cfs_rq
= cfs_rq_of(se
);
4916 put_prev_entity(cfs_rq
, se
);
4921 * sched_yield() is very simple
4923 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4925 static void yield_task_fair(struct rq
*rq
)
4927 struct task_struct
*curr
= rq
->curr
;
4928 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4929 struct sched_entity
*se
= &curr
->se
;
4932 * Are we the only task in the tree?
4934 if (unlikely(rq
->nr_running
== 1))
4937 clear_buddies(cfs_rq
, se
);
4939 if (curr
->policy
!= SCHED_BATCH
) {
4940 update_rq_clock(rq
);
4942 * Update run-time statistics of the 'current'.
4944 update_curr(cfs_rq
);
4946 * Tell update_rq_clock() that we've just updated,
4947 * so we don't do microscopic update in schedule()
4948 * and double the fastpath cost.
4950 rq
->skip_clock_update
= 1;
4956 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
4958 struct sched_entity
*se
= &p
->se
;
4960 /* throttled hierarchies are not runnable */
4961 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
4964 /* Tell the scheduler that we'd really like pse to run next. */
4967 yield_task_fair(rq
);
4973 /**************************************************
4974 * Fair scheduling class load-balancing methods.
4978 * The purpose of load-balancing is to achieve the same basic fairness the
4979 * per-cpu scheduler provides, namely provide a proportional amount of compute
4980 * time to each task. This is expressed in the following equation:
4982 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4984 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4985 * W_i,0 is defined as:
4987 * W_i,0 = \Sum_j w_i,j (2)
4989 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4990 * is derived from the nice value as per prio_to_weight[].
4992 * The weight average is an exponential decay average of the instantaneous
4995 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4997 * C_i is the compute capacity of cpu i, typically it is the
4998 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4999 * can also include other factors [XXX].
5001 * To achieve this balance we define a measure of imbalance which follows
5002 * directly from (1):
5004 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5006 * We them move tasks around to minimize the imbalance. In the continuous
5007 * function space it is obvious this converges, in the discrete case we get
5008 * a few fun cases generally called infeasible weight scenarios.
5011 * - infeasible weights;
5012 * - local vs global optima in the discrete case. ]
5017 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5018 * for all i,j solution, we create a tree of cpus that follows the hardware
5019 * topology where each level pairs two lower groups (or better). This results
5020 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5021 * tree to only the first of the previous level and we decrease the frequency
5022 * of load-balance at each level inv. proportional to the number of cpus in
5028 * \Sum { --- * --- * 2^i } = O(n) (5)
5030 * `- size of each group
5031 * | | `- number of cpus doing load-balance
5033 * `- sum over all levels
5035 * Coupled with a limit on how many tasks we can migrate every balance pass,
5036 * this makes (5) the runtime complexity of the balancer.
5038 * An important property here is that each CPU is still (indirectly) connected
5039 * to every other cpu in at most O(log n) steps:
5041 * The adjacency matrix of the resulting graph is given by:
5044 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5047 * And you'll find that:
5049 * A^(log_2 n)_i,j != 0 for all i,j (7)
5051 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5052 * The task movement gives a factor of O(m), giving a convergence complexity
5055 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5060 * In order to avoid CPUs going idle while there's still work to do, new idle
5061 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5062 * tree itself instead of relying on other CPUs to bring it work.
5064 * This adds some complexity to both (5) and (8) but it reduces the total idle
5072 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5075 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5080 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5082 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5084 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5087 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5088 * rewrite all of this once again.]
5091 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
5093 enum fbq_type
{ regular
, remote
, all
};
5095 #define LBF_ALL_PINNED 0x01
5096 #define LBF_NEED_BREAK 0x02
5097 #define LBF_DST_PINNED 0x04
5098 #define LBF_SOME_PINNED 0x08
5101 struct sched_domain
*sd
;
5109 struct cpumask
*dst_grpmask
;
5111 enum cpu_idle_type idle
;
5113 /* The set of CPUs under consideration for load-balancing */
5114 struct cpumask
*cpus
;
5119 unsigned int loop_break
;
5120 unsigned int loop_max
;
5122 enum fbq_type fbq_type
;
5126 * move_task - move a task from one runqueue to another runqueue.
5127 * Both runqueues must be locked.
5129 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
5131 deactivate_task(env
->src_rq
, p
, 0);
5132 set_task_cpu(p
, env
->dst_cpu
);
5133 activate_task(env
->dst_rq
, p
, 0);
5134 check_preempt_curr(env
->dst_rq
, p
, 0);
5138 * Is this task likely cache-hot:
5140 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
5144 if (p
->sched_class
!= &fair_sched_class
)
5147 if (unlikely(p
->policy
== SCHED_IDLE
))
5151 * Buddy candidates are cache hot:
5153 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
5154 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
5155 &p
->se
== cfs_rq_of(&p
->se
)->last
))
5158 if (sysctl_sched_migration_cost
== -1)
5160 if (sysctl_sched_migration_cost
== 0)
5163 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
5165 return delta
< (s64
)sysctl_sched_migration_cost
;
5168 #ifdef CONFIG_NUMA_BALANCING
5169 /* Returns true if the destination node has incurred more faults */
5170 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
5172 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5173 int src_nid
, dst_nid
;
5175 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults_memory
||
5176 !(env
->sd
->flags
& SD_NUMA
)) {
5180 src_nid
= cpu_to_node(env
->src_cpu
);
5181 dst_nid
= cpu_to_node(env
->dst_cpu
);
5183 if (src_nid
== dst_nid
)
5187 /* Task is already in the group's interleave set. */
5188 if (node_isset(src_nid
, numa_group
->active_nodes
))
5191 /* Task is moving into the group's interleave set. */
5192 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5195 return group_faults(p
, dst_nid
) > group_faults(p
, src_nid
);
5198 /* Encourage migration to the preferred node. */
5199 if (dst_nid
== p
->numa_preferred_nid
)
5202 return task_faults(p
, dst_nid
) > task_faults(p
, src_nid
);
5206 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
5208 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5209 int src_nid
, dst_nid
;
5211 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
5214 if (!p
->numa_faults_memory
|| !(env
->sd
->flags
& SD_NUMA
))
5217 src_nid
= cpu_to_node(env
->src_cpu
);
5218 dst_nid
= cpu_to_node(env
->dst_cpu
);
5220 if (src_nid
== dst_nid
)
5224 /* Task is moving within/into the group's interleave set. */
5225 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5228 /* Task is moving out of the group's interleave set. */
5229 if (node_isset(src_nid
, numa_group
->active_nodes
))
5232 return group_faults(p
, dst_nid
) < group_faults(p
, src_nid
);
5235 /* Migrating away from the preferred node is always bad. */
5236 if (src_nid
== p
->numa_preferred_nid
)
5239 return task_faults(p
, dst_nid
) < task_faults(p
, src_nid
);
5243 static inline bool migrate_improves_locality(struct task_struct
*p
,
5249 static inline bool migrate_degrades_locality(struct task_struct
*p
,
5257 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5260 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
5262 int tsk_cache_hot
= 0;
5264 * We do not migrate tasks that are:
5265 * 1) throttled_lb_pair, or
5266 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5267 * 3) running (obviously), or
5268 * 4) are cache-hot on their current CPU.
5270 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
5273 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
5276 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
5278 env
->flags
|= LBF_SOME_PINNED
;
5281 * Remember if this task can be migrated to any other cpu in
5282 * our sched_group. We may want to revisit it if we couldn't
5283 * meet load balance goals by pulling other tasks on src_cpu.
5285 * Also avoid computing new_dst_cpu if we have already computed
5286 * one in current iteration.
5288 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
5291 /* Prevent to re-select dst_cpu via env's cpus */
5292 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
5293 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
5294 env
->flags
|= LBF_DST_PINNED
;
5295 env
->new_dst_cpu
= cpu
;
5303 /* Record that we found atleast one task that could run on dst_cpu */
5304 env
->flags
&= ~LBF_ALL_PINNED
;
5306 if (task_running(env
->src_rq
, p
)) {
5307 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
5312 * Aggressive migration if:
5313 * 1) destination numa is preferred
5314 * 2) task is cache cold, or
5315 * 3) too many balance attempts have failed.
5317 tsk_cache_hot
= task_hot(p
, env
);
5319 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
5321 if (migrate_improves_locality(p
, env
)) {
5322 #ifdef CONFIG_SCHEDSTATS
5323 if (tsk_cache_hot
) {
5324 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5325 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5331 if (!tsk_cache_hot
||
5332 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
5334 if (tsk_cache_hot
) {
5335 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5336 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5342 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
5347 * move_one_task tries to move exactly one task from busiest to this_rq, as
5348 * part of active balancing operations within "domain".
5349 * Returns 1 if successful and 0 otherwise.
5351 * Called with both runqueues locked.
5353 static int move_one_task(struct lb_env
*env
)
5355 struct task_struct
*p
, *n
;
5357 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
5358 if (!can_migrate_task(p
, env
))
5363 * Right now, this is only the second place move_task()
5364 * is called, so we can safely collect move_task()
5365 * stats here rather than inside move_task().
5367 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
5373 static const unsigned int sched_nr_migrate_break
= 32;
5376 * move_tasks tries to move up to imbalance weighted load from busiest to
5377 * this_rq, as part of a balancing operation within domain "sd".
5378 * Returns 1 if successful and 0 otherwise.
5380 * Called with both runqueues locked.
5382 static int move_tasks(struct lb_env
*env
)
5384 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
5385 struct task_struct
*p
;
5389 if (env
->imbalance
<= 0)
5392 while (!list_empty(tasks
)) {
5393 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5396 /* We've more or less seen every task there is, call it quits */
5397 if (env
->loop
> env
->loop_max
)
5400 /* take a breather every nr_migrate tasks */
5401 if (env
->loop
> env
->loop_break
) {
5402 env
->loop_break
+= sched_nr_migrate_break
;
5403 env
->flags
|= LBF_NEED_BREAK
;
5407 if (!can_migrate_task(p
, env
))
5410 load
= task_h_load(p
);
5412 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5415 if ((load
/ 2) > env
->imbalance
)
5420 env
->imbalance
-= load
;
5422 #ifdef CONFIG_PREEMPT
5424 * NEWIDLE balancing is a source of latency, so preemptible
5425 * kernels will stop after the first task is pulled to minimize
5426 * the critical section.
5428 if (env
->idle
== CPU_NEWLY_IDLE
)
5433 * We only want to steal up to the prescribed amount of
5436 if (env
->imbalance
<= 0)
5441 list_move_tail(&p
->se
.group_node
, tasks
);
5445 * Right now, this is one of only two places move_task() is called,
5446 * so we can safely collect move_task() stats here rather than
5447 * inside move_task().
5449 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
5454 #ifdef CONFIG_FAIR_GROUP_SCHED
5456 * update tg->load_weight by folding this cpu's load_avg
5458 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5460 struct sched_entity
*se
= tg
->se
[cpu
];
5461 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5463 /* throttled entities do not contribute to load */
5464 if (throttled_hierarchy(cfs_rq
))
5467 update_cfs_rq_blocked_load(cfs_rq
, 1);
5470 update_entity_load_avg(se
, 1);
5472 * We pivot on our runnable average having decayed to zero for
5473 * list removal. This generally implies that all our children
5474 * have also been removed (modulo rounding error or bandwidth
5475 * control); however, such cases are rare and we can fix these
5478 * TODO: fix up out-of-order children on enqueue.
5480 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5481 list_del_leaf_cfs_rq(cfs_rq
);
5483 struct rq
*rq
= rq_of(cfs_rq
);
5484 update_rq_runnable_avg(rq
, rq
->nr_running
);
5488 static void update_blocked_averages(int cpu
)
5490 struct rq
*rq
= cpu_rq(cpu
);
5491 struct cfs_rq
*cfs_rq
;
5492 unsigned long flags
;
5494 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5495 update_rq_clock(rq
);
5497 * Iterates the task_group tree in a bottom up fashion, see
5498 * list_add_leaf_cfs_rq() for details.
5500 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5502 * Note: We may want to consider periodically releasing
5503 * rq->lock about these updates so that creating many task
5504 * groups does not result in continually extending hold time.
5506 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5509 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5513 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5514 * This needs to be done in a top-down fashion because the load of a child
5515 * group is a fraction of its parents load.
5517 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5519 struct rq
*rq
= rq_of(cfs_rq
);
5520 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5521 unsigned long now
= jiffies
;
5524 if (cfs_rq
->last_h_load_update
== now
)
5527 cfs_rq
->h_load_next
= NULL
;
5528 for_each_sched_entity(se
) {
5529 cfs_rq
= cfs_rq_of(se
);
5530 cfs_rq
->h_load_next
= se
;
5531 if (cfs_rq
->last_h_load_update
== now
)
5536 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5537 cfs_rq
->last_h_load_update
= now
;
5540 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5541 load
= cfs_rq
->h_load
;
5542 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5543 cfs_rq
->runnable_load_avg
+ 1);
5544 cfs_rq
= group_cfs_rq(se
);
5545 cfs_rq
->h_load
= load
;
5546 cfs_rq
->last_h_load_update
= now
;
5550 static unsigned long task_h_load(struct task_struct
*p
)
5552 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5554 update_cfs_rq_h_load(cfs_rq
);
5555 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5556 cfs_rq
->runnable_load_avg
+ 1);
5559 static inline void update_blocked_averages(int cpu
)
5563 static unsigned long task_h_load(struct task_struct
*p
)
5565 return p
->se
.avg
.load_avg_contrib
;
5569 /********** Helpers for find_busiest_group ************************/
5571 * sg_lb_stats - stats of a sched_group required for load_balancing
5573 struct sg_lb_stats
{
5574 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5575 unsigned long group_load
; /* Total load over the CPUs of the group */
5576 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5577 unsigned long load_per_task
;
5578 unsigned long group_capacity
;
5579 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5580 unsigned int group_capacity_factor
;
5581 unsigned int idle_cpus
;
5582 unsigned int group_weight
;
5583 int group_imb
; /* Is there an imbalance in the group ? */
5584 int group_has_free_capacity
;
5585 #ifdef CONFIG_NUMA_BALANCING
5586 unsigned int nr_numa_running
;
5587 unsigned int nr_preferred_running
;
5592 * sd_lb_stats - Structure to store the statistics of a sched_domain
5593 * during load balancing.
5595 struct sd_lb_stats
{
5596 struct sched_group
*busiest
; /* Busiest group in this sd */
5597 struct sched_group
*local
; /* Local group in this sd */
5598 unsigned long total_load
; /* Total load of all groups in sd */
5599 unsigned long total_capacity
; /* Total capacity of all groups in sd */
5600 unsigned long avg_load
; /* Average load across all groups in sd */
5602 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5603 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5606 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5609 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5610 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5611 * We must however clear busiest_stat::avg_load because
5612 * update_sd_pick_busiest() reads this before assignment.
5614 *sds
= (struct sd_lb_stats
){
5618 .total_capacity
= 0UL,
5626 * get_sd_load_idx - Obtain the load index for a given sched domain.
5627 * @sd: The sched_domain whose load_idx is to be obtained.
5628 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5630 * Return: The load index.
5632 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5633 enum cpu_idle_type idle
)
5639 load_idx
= sd
->busy_idx
;
5642 case CPU_NEWLY_IDLE
:
5643 load_idx
= sd
->newidle_idx
;
5646 load_idx
= sd
->idle_idx
;
5653 static unsigned long default_scale_capacity(struct sched_domain
*sd
, int cpu
)
5655 return SCHED_CAPACITY_SCALE
;
5658 unsigned long __weak
arch_scale_freq_capacity(struct sched_domain
*sd
, int cpu
)
5660 return default_scale_capacity(sd
, cpu
);
5663 static unsigned long default_scale_smt_capacity(struct sched_domain
*sd
, int cpu
)
5665 unsigned long weight
= sd
->span_weight
;
5666 unsigned long smt_gain
= sd
->smt_gain
;
5673 unsigned long __weak
arch_scale_smt_capacity(struct sched_domain
*sd
, int cpu
)
5675 return default_scale_smt_capacity(sd
, cpu
);
5678 static unsigned long scale_rt_capacity(int cpu
)
5680 struct rq
*rq
= cpu_rq(cpu
);
5681 u64 total
, available
, age_stamp
, avg
;
5685 * Since we're reading these variables without serialization make sure
5686 * we read them once before doing sanity checks on them.
5688 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5689 avg
= ACCESS_ONCE(rq
->rt_avg
);
5691 delta
= rq_clock(rq
) - age_stamp
;
5692 if (unlikely(delta
< 0))
5695 total
= sched_avg_period() + delta
;
5697 if (unlikely(total
< avg
)) {
5698 /* Ensures that capacity won't end up being negative */
5701 available
= total
- avg
;
5704 if (unlikely((s64
)total
< SCHED_CAPACITY_SCALE
))
5705 total
= SCHED_CAPACITY_SCALE
;
5707 total
>>= SCHED_CAPACITY_SHIFT
;
5709 return div_u64(available
, total
);
5712 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5714 unsigned long weight
= sd
->span_weight
;
5715 unsigned long capacity
= SCHED_CAPACITY_SCALE
;
5716 struct sched_group
*sdg
= sd
->groups
;
5718 if ((sd
->flags
& SD_SHARE_CPUCAPACITY
) && weight
> 1) {
5719 if (sched_feat(ARCH_CAPACITY
))
5720 capacity
*= arch_scale_smt_capacity(sd
, cpu
);
5722 capacity
*= default_scale_smt_capacity(sd
, cpu
);
5724 capacity
>>= SCHED_CAPACITY_SHIFT
;
5727 sdg
->sgc
->capacity_orig
= capacity
;
5729 if (sched_feat(ARCH_CAPACITY
))
5730 capacity
*= arch_scale_freq_capacity(sd
, cpu
);
5732 capacity
*= default_scale_capacity(sd
, cpu
);
5734 capacity
>>= SCHED_CAPACITY_SHIFT
;
5736 capacity
*= scale_rt_capacity(cpu
);
5737 capacity
>>= SCHED_CAPACITY_SHIFT
;
5742 cpu_rq(cpu
)->cpu_capacity
= capacity
;
5743 sdg
->sgc
->capacity
= capacity
;
5746 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
5748 struct sched_domain
*child
= sd
->child
;
5749 struct sched_group
*group
, *sdg
= sd
->groups
;
5750 unsigned long capacity
, capacity_orig
;
5751 unsigned long interval
;
5753 interval
= msecs_to_jiffies(sd
->balance_interval
);
5754 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5755 sdg
->sgc
->next_update
= jiffies
+ interval
;
5758 update_cpu_capacity(sd
, cpu
);
5762 capacity_orig
= capacity
= 0;
5764 if (child
->flags
& SD_OVERLAP
) {
5766 * SD_OVERLAP domains cannot assume that child groups
5767 * span the current group.
5770 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
5771 struct sched_group_capacity
*sgc
;
5772 struct rq
*rq
= cpu_rq(cpu
);
5775 * build_sched_domains() -> init_sched_groups_capacity()
5776 * gets here before we've attached the domains to the
5779 * Use capacity_of(), which is set irrespective of domains
5780 * in update_cpu_capacity().
5782 * This avoids capacity/capacity_orig from being 0 and
5783 * causing divide-by-zero issues on boot.
5785 * Runtime updates will correct capacity_orig.
5787 if (unlikely(!rq
->sd
)) {
5788 capacity_orig
+= capacity_of(cpu
);
5789 capacity
+= capacity_of(cpu
);
5793 sgc
= rq
->sd
->groups
->sgc
;
5794 capacity_orig
+= sgc
->capacity_orig
;
5795 capacity
+= sgc
->capacity
;
5799 * !SD_OVERLAP domains can assume that child groups
5800 * span the current group.
5803 group
= child
->groups
;
5805 capacity_orig
+= group
->sgc
->capacity_orig
;
5806 capacity
+= group
->sgc
->capacity
;
5807 group
= group
->next
;
5808 } while (group
!= child
->groups
);
5811 sdg
->sgc
->capacity_orig
= capacity_orig
;
5812 sdg
->sgc
->capacity
= capacity
;
5816 * Try and fix up capacity for tiny siblings, this is needed when
5817 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5818 * which on its own isn't powerful enough.
5820 * See update_sd_pick_busiest() and check_asym_packing().
5823 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
5826 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
5828 if (!(sd
->flags
& SD_SHARE_CPUCAPACITY
))
5832 * If ~90% of the cpu_capacity is still there, we're good.
5834 if (group
->sgc
->capacity
* 32 > group
->sgc
->capacity_orig
* 29)
5841 * Group imbalance indicates (and tries to solve) the problem where balancing
5842 * groups is inadequate due to tsk_cpus_allowed() constraints.
5844 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5845 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5848 * { 0 1 2 3 } { 4 5 6 7 }
5851 * If we were to balance group-wise we'd place two tasks in the first group and
5852 * two tasks in the second group. Clearly this is undesired as it will overload
5853 * cpu 3 and leave one of the cpus in the second group unused.
5855 * The current solution to this issue is detecting the skew in the first group
5856 * by noticing the lower domain failed to reach balance and had difficulty
5857 * moving tasks due to affinity constraints.
5859 * When this is so detected; this group becomes a candidate for busiest; see
5860 * update_sd_pick_busiest(). And calculate_imbalance() and
5861 * find_busiest_group() avoid some of the usual balance conditions to allow it
5862 * to create an effective group imbalance.
5864 * This is a somewhat tricky proposition since the next run might not find the
5865 * group imbalance and decide the groups need to be balanced again. A most
5866 * subtle and fragile situation.
5869 static inline int sg_imbalanced(struct sched_group
*group
)
5871 return group
->sgc
->imbalance
;
5875 * Compute the group capacity factor.
5877 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
5878 * first dividing out the smt factor and computing the actual number of cores
5879 * and limit unit capacity with that.
5881 static inline int sg_capacity_factor(struct lb_env
*env
, struct sched_group
*group
)
5883 unsigned int capacity_factor
, smt
, cpus
;
5884 unsigned int capacity
, capacity_orig
;
5886 capacity
= group
->sgc
->capacity
;
5887 capacity_orig
= group
->sgc
->capacity_orig
;
5888 cpus
= group
->group_weight
;
5890 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
5891 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, capacity_orig
);
5892 capacity_factor
= cpus
/ smt
; /* cores */
5894 capacity_factor
= min_t(unsigned,
5895 capacity_factor
, DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
));
5896 if (!capacity_factor
)
5897 capacity_factor
= fix_small_capacity(env
->sd
, group
);
5899 return capacity_factor
;
5903 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5904 * @env: The load balancing environment.
5905 * @group: sched_group whose statistics are to be updated.
5906 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5907 * @local_group: Does group contain this_cpu.
5908 * @sgs: variable to hold the statistics for this group.
5910 static inline void update_sg_lb_stats(struct lb_env
*env
,
5911 struct sched_group
*group
, int load_idx
,
5912 int local_group
, struct sg_lb_stats
*sgs
,
5918 memset(sgs
, 0, sizeof(*sgs
));
5920 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5921 struct rq
*rq
= cpu_rq(i
);
5923 /* Bias balancing toward cpus of our domain */
5925 load
= target_load(i
, load_idx
);
5927 load
= source_load(i
, load_idx
);
5929 sgs
->group_load
+= load
;
5930 sgs
->sum_nr_running
+= rq
->nr_running
;
5932 if (rq
->nr_running
> 1)
5935 #ifdef CONFIG_NUMA_BALANCING
5936 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
5937 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
5939 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
5944 /* Adjust by relative CPU capacity of the group */
5945 sgs
->group_capacity
= group
->sgc
->capacity
;
5946 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
5948 if (sgs
->sum_nr_running
)
5949 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
5951 sgs
->group_weight
= group
->group_weight
;
5953 sgs
->group_imb
= sg_imbalanced(group
);
5954 sgs
->group_capacity_factor
= sg_capacity_factor(env
, group
);
5956 if (sgs
->group_capacity_factor
> sgs
->sum_nr_running
)
5957 sgs
->group_has_free_capacity
= 1;
5961 * update_sd_pick_busiest - return 1 on busiest group
5962 * @env: The load balancing environment.
5963 * @sds: sched_domain statistics
5964 * @sg: sched_group candidate to be checked for being the busiest
5965 * @sgs: sched_group statistics
5967 * Determine if @sg is a busier group than the previously selected
5970 * Return: %true if @sg is a busier group than the previously selected
5971 * busiest group. %false otherwise.
5973 static bool update_sd_pick_busiest(struct lb_env
*env
,
5974 struct sd_lb_stats
*sds
,
5975 struct sched_group
*sg
,
5976 struct sg_lb_stats
*sgs
)
5978 if (sgs
->avg_load
<= sds
->busiest_stat
.avg_load
)
5981 if (sgs
->sum_nr_running
> sgs
->group_capacity_factor
)
5988 * ASYM_PACKING needs to move all the work to the lowest
5989 * numbered CPUs in the group, therefore mark all groups
5990 * higher than ourself as busy.
5992 if ((env
->sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
5993 env
->dst_cpu
< group_first_cpu(sg
)) {
5997 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
6004 #ifdef CONFIG_NUMA_BALANCING
6005 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6007 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
6009 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
6014 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6016 if (rq
->nr_running
> rq
->nr_numa_running
)
6018 if (rq
->nr_running
> rq
->nr_preferred_running
)
6023 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6028 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6032 #endif /* CONFIG_NUMA_BALANCING */
6035 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6036 * @env: The load balancing environment.
6037 * @sds: variable to hold the statistics for this sched_domain.
6039 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6041 struct sched_domain
*child
= env
->sd
->child
;
6042 struct sched_group
*sg
= env
->sd
->groups
;
6043 struct sg_lb_stats tmp_sgs
;
6044 int load_idx
, prefer_sibling
= 0;
6045 bool overload
= false;
6047 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
6050 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
6053 struct sg_lb_stats
*sgs
= &tmp_sgs
;
6056 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
6059 sgs
= &sds
->local_stat
;
6061 if (env
->idle
!= CPU_NEWLY_IDLE
||
6062 time_after_eq(jiffies
, sg
->sgc
->next_update
))
6063 update_group_capacity(env
->sd
, env
->dst_cpu
);
6066 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
,
6073 * In case the child domain prefers tasks go to siblings
6074 * first, lower the sg capacity factor to one so that we'll try
6075 * and move all the excess tasks away. We lower the capacity
6076 * of a group only if the local group has the capacity to fit
6077 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6078 * extra check prevents the case where you always pull from the
6079 * heaviest group when it is already under-utilized (possible
6080 * with a large weight task outweighs the tasks on the system).
6082 if (prefer_sibling
&& sds
->local
&&
6083 sds
->local_stat
.group_has_free_capacity
)
6084 sgs
->group_capacity_factor
= min(sgs
->group_capacity_factor
, 1U);
6086 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
6088 sds
->busiest_stat
= *sgs
;
6092 /* Now, start updating sd_lb_stats */
6093 sds
->total_load
+= sgs
->group_load
;
6094 sds
->total_capacity
+= sgs
->group_capacity
;
6097 } while (sg
!= env
->sd
->groups
);
6099 if (env
->sd
->flags
& SD_NUMA
)
6100 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
6102 if (!env
->sd
->parent
) {
6103 /* update overload indicator if we are at root domain */
6104 if (env
->dst_rq
->rd
->overload
!= overload
)
6105 env
->dst_rq
->rd
->overload
= overload
;
6111 * check_asym_packing - Check to see if the group is packed into the
6114 * This is primarily intended to used at the sibling level. Some
6115 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6116 * case of POWER7, it can move to lower SMT modes only when higher
6117 * threads are idle. When in lower SMT modes, the threads will
6118 * perform better since they share less core resources. Hence when we
6119 * have idle threads, we want them to be the higher ones.
6121 * This packing function is run on idle threads. It checks to see if
6122 * the busiest CPU in this domain (core in the P7 case) has a higher
6123 * CPU number than the packing function is being run on. Here we are
6124 * assuming lower CPU number will be equivalent to lower a SMT thread
6127 * Return: 1 when packing is required and a task should be moved to
6128 * this CPU. The amount of the imbalance is returned in *imbalance.
6130 * @env: The load balancing environment.
6131 * @sds: Statistics of the sched_domain which is to be packed
6133 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6137 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6143 busiest_cpu
= group_first_cpu(sds
->busiest
);
6144 if (env
->dst_cpu
> busiest_cpu
)
6147 env
->imbalance
= DIV_ROUND_CLOSEST(
6148 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
6149 SCHED_CAPACITY_SCALE
);
6155 * fix_small_imbalance - Calculate the minor imbalance that exists
6156 * amongst the groups of a sched_domain, during
6158 * @env: The load balancing environment.
6159 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6162 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6164 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
6165 unsigned int imbn
= 2;
6166 unsigned long scaled_busy_load_per_task
;
6167 struct sg_lb_stats
*local
, *busiest
;
6169 local
= &sds
->local_stat
;
6170 busiest
= &sds
->busiest_stat
;
6172 if (!local
->sum_nr_running
)
6173 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
6174 else if (busiest
->load_per_task
> local
->load_per_task
)
6177 scaled_busy_load_per_task
=
6178 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6179 busiest
->group_capacity
;
6181 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
6182 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
6183 env
->imbalance
= busiest
->load_per_task
;
6188 * OK, we don't have enough imbalance to justify moving tasks,
6189 * however we may be able to increase total CPU capacity used by
6193 capa_now
+= busiest
->group_capacity
*
6194 min(busiest
->load_per_task
, busiest
->avg_load
);
6195 capa_now
+= local
->group_capacity
*
6196 min(local
->load_per_task
, local
->avg_load
);
6197 capa_now
/= SCHED_CAPACITY_SCALE
;
6199 /* Amount of load we'd subtract */
6200 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
6201 capa_move
+= busiest
->group_capacity
*
6202 min(busiest
->load_per_task
,
6203 busiest
->avg_load
- scaled_busy_load_per_task
);
6206 /* Amount of load we'd add */
6207 if (busiest
->avg_load
* busiest
->group_capacity
<
6208 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
6209 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
6210 local
->group_capacity
;
6212 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6213 local
->group_capacity
;
6215 capa_move
+= local
->group_capacity
*
6216 min(local
->load_per_task
, local
->avg_load
+ tmp
);
6217 capa_move
/= SCHED_CAPACITY_SCALE
;
6219 /* Move if we gain throughput */
6220 if (capa_move
> capa_now
)
6221 env
->imbalance
= busiest
->load_per_task
;
6225 * calculate_imbalance - Calculate the amount of imbalance present within the
6226 * groups of a given sched_domain during load balance.
6227 * @env: load balance environment
6228 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6230 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6232 unsigned long max_pull
, load_above_capacity
= ~0UL;
6233 struct sg_lb_stats
*local
, *busiest
;
6235 local
= &sds
->local_stat
;
6236 busiest
= &sds
->busiest_stat
;
6238 if (busiest
->group_imb
) {
6240 * In the group_imb case we cannot rely on group-wide averages
6241 * to ensure cpu-load equilibrium, look at wider averages. XXX
6243 busiest
->load_per_task
=
6244 min(busiest
->load_per_task
, sds
->avg_load
);
6248 * In the presence of smp nice balancing, certain scenarios can have
6249 * max load less than avg load(as we skip the groups at or below
6250 * its cpu_capacity, while calculating max_load..)
6252 if (busiest
->avg_load
<= sds
->avg_load
||
6253 local
->avg_load
>= sds
->avg_load
) {
6255 return fix_small_imbalance(env
, sds
);
6258 if (!busiest
->group_imb
) {
6260 * Don't want to pull so many tasks that a group would go idle.
6261 * Except of course for the group_imb case, since then we might
6262 * have to drop below capacity to reach cpu-load equilibrium.
6264 load_above_capacity
=
6265 (busiest
->sum_nr_running
- busiest
->group_capacity_factor
);
6267 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_CAPACITY_SCALE
);
6268 load_above_capacity
/= busiest
->group_capacity
;
6272 * We're trying to get all the cpus to the average_load, so we don't
6273 * want to push ourselves above the average load, nor do we wish to
6274 * reduce the max loaded cpu below the average load. At the same time,
6275 * we also don't want to reduce the group load below the group capacity
6276 * (so that we can implement power-savings policies etc). Thus we look
6277 * for the minimum possible imbalance.
6279 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
6281 /* How much load to actually move to equalise the imbalance */
6282 env
->imbalance
= min(
6283 max_pull
* busiest
->group_capacity
,
6284 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
6285 ) / SCHED_CAPACITY_SCALE
;
6288 * if *imbalance is less than the average load per runnable task
6289 * there is no guarantee that any tasks will be moved so we'll have
6290 * a think about bumping its value to force at least one task to be
6293 if (env
->imbalance
< busiest
->load_per_task
)
6294 return fix_small_imbalance(env
, sds
);
6297 /******* find_busiest_group() helpers end here *********************/
6300 * find_busiest_group - Returns the busiest group within the sched_domain
6301 * if there is an imbalance. If there isn't an imbalance, and
6302 * the user has opted for power-savings, it returns a group whose
6303 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6304 * such a group exists.
6306 * Also calculates the amount of weighted load which should be moved
6307 * to restore balance.
6309 * @env: The load balancing environment.
6311 * Return: - The busiest group if imbalance exists.
6312 * - If no imbalance and user has opted for power-savings balance,
6313 * return the least loaded group whose CPUs can be
6314 * put to idle by rebalancing its tasks onto our group.
6316 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
6318 struct sg_lb_stats
*local
, *busiest
;
6319 struct sd_lb_stats sds
;
6321 init_sd_lb_stats(&sds
);
6324 * Compute the various statistics relavent for load balancing at
6327 update_sd_lb_stats(env
, &sds
);
6328 local
= &sds
.local_stat
;
6329 busiest
= &sds
.busiest_stat
;
6331 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
6332 check_asym_packing(env
, &sds
))
6335 /* There is no busy sibling group to pull tasks from */
6336 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
6339 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
6340 / sds
.total_capacity
;
6343 * If the busiest group is imbalanced the below checks don't
6344 * work because they assume all things are equal, which typically
6345 * isn't true due to cpus_allowed constraints and the like.
6347 if (busiest
->group_imb
)
6350 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6351 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_free_capacity
&&
6352 !busiest
->group_has_free_capacity
)
6356 * If the local group is more busy than the selected busiest group
6357 * don't try and pull any tasks.
6359 if (local
->avg_load
>= busiest
->avg_load
)
6363 * Don't pull any tasks if this group is already above the domain
6366 if (local
->avg_load
>= sds
.avg_load
)
6369 if (env
->idle
== CPU_IDLE
) {
6371 * This cpu is idle. If the busiest group load doesn't
6372 * have more tasks than the number of available cpu's and
6373 * there is no imbalance between this and busiest group
6374 * wrt to idle cpu's, it is balanced.
6376 if ((local
->idle_cpus
< busiest
->idle_cpus
) &&
6377 busiest
->sum_nr_running
<= busiest
->group_weight
)
6381 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6382 * imbalance_pct to be conservative.
6384 if (100 * busiest
->avg_load
<=
6385 env
->sd
->imbalance_pct
* local
->avg_load
)
6390 /* Looks like there is an imbalance. Compute it */
6391 calculate_imbalance(env
, &sds
);
6400 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6402 static struct rq
*find_busiest_queue(struct lb_env
*env
,
6403 struct sched_group
*group
)
6405 struct rq
*busiest
= NULL
, *rq
;
6406 unsigned long busiest_load
= 0, busiest_capacity
= 1;
6409 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6410 unsigned long capacity
, capacity_factor
, wl
;
6414 rt
= fbq_classify_rq(rq
);
6417 * We classify groups/runqueues into three groups:
6418 * - regular: there are !numa tasks
6419 * - remote: there are numa tasks that run on the 'wrong' node
6420 * - all: there is no distinction
6422 * In order to avoid migrating ideally placed numa tasks,
6423 * ignore those when there's better options.
6425 * If we ignore the actual busiest queue to migrate another
6426 * task, the next balance pass can still reduce the busiest
6427 * queue by moving tasks around inside the node.
6429 * If we cannot move enough load due to this classification
6430 * the next pass will adjust the group classification and
6431 * allow migration of more tasks.
6433 * Both cases only affect the total convergence complexity.
6435 if (rt
> env
->fbq_type
)
6438 capacity
= capacity_of(i
);
6439 capacity_factor
= DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
);
6440 if (!capacity_factor
)
6441 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6443 wl
= weighted_cpuload(i
);
6446 * When comparing with imbalance, use weighted_cpuload()
6447 * which is not scaled with the cpu capacity.
6449 if (capacity_factor
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6453 * For the load comparisons with the other cpu's, consider
6454 * the weighted_cpuload() scaled with the cpu capacity, so
6455 * that the load can be moved away from the cpu that is
6456 * potentially running at a lower capacity.
6458 * Thus we're looking for max(wl_i / capacity_i), crosswise
6459 * multiplication to rid ourselves of the division works out
6460 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6461 * our previous maximum.
6463 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
6465 busiest_capacity
= capacity
;
6474 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6475 * so long as it is large enough.
6477 #define MAX_PINNED_INTERVAL 512
6479 /* Working cpumask for load_balance and load_balance_newidle. */
6480 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6482 static int need_active_balance(struct lb_env
*env
)
6484 struct sched_domain
*sd
= env
->sd
;
6486 if (env
->idle
== CPU_NEWLY_IDLE
) {
6489 * ASYM_PACKING needs to force migrate tasks from busy but
6490 * higher numbered CPUs in order to pack all tasks in the
6491 * lowest numbered CPUs.
6493 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6497 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6500 static int active_load_balance_cpu_stop(void *data
);
6502 static int should_we_balance(struct lb_env
*env
)
6504 struct sched_group
*sg
= env
->sd
->groups
;
6505 struct cpumask
*sg_cpus
, *sg_mask
;
6506 int cpu
, balance_cpu
= -1;
6509 * In the newly idle case, we will allow all the cpu's
6510 * to do the newly idle load balance.
6512 if (env
->idle
== CPU_NEWLY_IDLE
)
6515 sg_cpus
= sched_group_cpus(sg
);
6516 sg_mask
= sched_group_mask(sg
);
6517 /* Try to find first idle cpu */
6518 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6519 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6526 if (balance_cpu
== -1)
6527 balance_cpu
= group_balance_cpu(sg
);
6530 * First idle cpu or the first cpu(busiest) in this sched group
6531 * is eligible for doing load balancing at this and above domains.
6533 return balance_cpu
== env
->dst_cpu
;
6537 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6538 * tasks if there is an imbalance.
6540 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6541 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6542 int *continue_balancing
)
6544 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6545 struct sched_domain
*sd_parent
= sd
->parent
;
6546 struct sched_group
*group
;
6548 unsigned long flags
;
6549 struct cpumask
*cpus
= __get_cpu_var(load_balance_mask
);
6551 struct lb_env env
= {
6553 .dst_cpu
= this_cpu
,
6555 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6557 .loop_break
= sched_nr_migrate_break
,
6563 * For NEWLY_IDLE load_balancing, we don't need to consider
6564 * other cpus in our group
6566 if (idle
== CPU_NEWLY_IDLE
)
6567 env
.dst_grpmask
= NULL
;
6569 cpumask_copy(cpus
, cpu_active_mask
);
6571 schedstat_inc(sd
, lb_count
[idle
]);
6574 if (!should_we_balance(&env
)) {
6575 *continue_balancing
= 0;
6579 group
= find_busiest_group(&env
);
6581 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6585 busiest
= find_busiest_queue(&env
, group
);
6587 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6591 BUG_ON(busiest
== env
.dst_rq
);
6593 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6596 if (busiest
->nr_running
> 1) {
6598 * Attempt to move tasks. If find_busiest_group has found
6599 * an imbalance but busiest->nr_running <= 1, the group is
6600 * still unbalanced. ld_moved simply stays zero, so it is
6601 * correctly treated as an imbalance.
6603 env
.flags
|= LBF_ALL_PINNED
;
6604 env
.src_cpu
= busiest
->cpu
;
6605 env
.src_rq
= busiest
;
6606 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6609 local_irq_save(flags
);
6610 double_rq_lock(env
.dst_rq
, busiest
);
6613 * cur_ld_moved - load moved in current iteration
6614 * ld_moved - cumulative load moved across iterations
6616 cur_ld_moved
= move_tasks(&env
);
6617 ld_moved
+= cur_ld_moved
;
6618 double_rq_unlock(env
.dst_rq
, busiest
);
6619 local_irq_restore(flags
);
6622 * some other cpu did the load balance for us.
6624 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
6625 resched_cpu(env
.dst_cpu
);
6627 if (env
.flags
& LBF_NEED_BREAK
) {
6628 env
.flags
&= ~LBF_NEED_BREAK
;
6633 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6634 * us and move them to an alternate dst_cpu in our sched_group
6635 * where they can run. The upper limit on how many times we
6636 * iterate on same src_cpu is dependent on number of cpus in our
6639 * This changes load balance semantics a bit on who can move
6640 * load to a given_cpu. In addition to the given_cpu itself
6641 * (or a ilb_cpu acting on its behalf where given_cpu is
6642 * nohz-idle), we now have balance_cpu in a position to move
6643 * load to given_cpu. In rare situations, this may cause
6644 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6645 * _independently_ and at _same_ time to move some load to
6646 * given_cpu) causing exceess load to be moved to given_cpu.
6647 * This however should not happen so much in practice and
6648 * moreover subsequent load balance cycles should correct the
6649 * excess load moved.
6651 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6653 /* Prevent to re-select dst_cpu via env's cpus */
6654 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6656 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6657 env
.dst_cpu
= env
.new_dst_cpu
;
6658 env
.flags
&= ~LBF_DST_PINNED
;
6660 env
.loop_break
= sched_nr_migrate_break
;
6663 * Go back to "more_balance" rather than "redo" since we
6664 * need to continue with same src_cpu.
6670 * We failed to reach balance because of affinity.
6673 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
6675 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0) {
6676 *group_imbalance
= 1;
6677 } else if (*group_imbalance
)
6678 *group_imbalance
= 0;
6681 /* All tasks on this runqueue were pinned by CPU affinity */
6682 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6683 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6684 if (!cpumask_empty(cpus
)) {
6686 env
.loop_break
= sched_nr_migrate_break
;
6694 schedstat_inc(sd
, lb_failed
[idle
]);
6696 * Increment the failure counter only on periodic balance.
6697 * We do not want newidle balance, which can be very
6698 * frequent, pollute the failure counter causing
6699 * excessive cache_hot migrations and active balances.
6701 if (idle
!= CPU_NEWLY_IDLE
)
6702 sd
->nr_balance_failed
++;
6704 if (need_active_balance(&env
)) {
6705 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6707 /* don't kick the active_load_balance_cpu_stop,
6708 * if the curr task on busiest cpu can't be
6711 if (!cpumask_test_cpu(this_cpu
,
6712 tsk_cpus_allowed(busiest
->curr
))) {
6713 raw_spin_unlock_irqrestore(&busiest
->lock
,
6715 env
.flags
|= LBF_ALL_PINNED
;
6716 goto out_one_pinned
;
6720 * ->active_balance synchronizes accesses to
6721 * ->active_balance_work. Once set, it's cleared
6722 * only after active load balance is finished.
6724 if (!busiest
->active_balance
) {
6725 busiest
->active_balance
= 1;
6726 busiest
->push_cpu
= this_cpu
;
6729 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
6731 if (active_balance
) {
6732 stop_one_cpu_nowait(cpu_of(busiest
),
6733 active_load_balance_cpu_stop
, busiest
,
6734 &busiest
->active_balance_work
);
6738 * We've kicked active balancing, reset the failure
6741 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
6744 sd
->nr_balance_failed
= 0;
6746 if (likely(!active_balance
)) {
6747 /* We were unbalanced, so reset the balancing interval */
6748 sd
->balance_interval
= sd
->min_interval
;
6751 * If we've begun active balancing, start to back off. This
6752 * case may not be covered by the all_pinned logic if there
6753 * is only 1 task on the busy runqueue (because we don't call
6756 if (sd
->balance_interval
< sd
->max_interval
)
6757 sd
->balance_interval
*= 2;
6763 schedstat_inc(sd
, lb_balanced
[idle
]);
6765 sd
->nr_balance_failed
= 0;
6768 /* tune up the balancing interval */
6769 if (((env
.flags
& LBF_ALL_PINNED
) &&
6770 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
6771 (sd
->balance_interval
< sd
->max_interval
))
6772 sd
->balance_interval
*= 2;
6779 static inline unsigned long
6780 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
6782 unsigned long interval
= sd
->balance_interval
;
6785 interval
*= sd
->busy_factor
;
6787 /* scale ms to jiffies */
6788 interval
= msecs_to_jiffies(interval
);
6789 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6795 update_next_balance(struct sched_domain
*sd
, int cpu_busy
, unsigned long *next_balance
)
6797 unsigned long interval
, next
;
6799 interval
= get_sd_balance_interval(sd
, cpu_busy
);
6800 next
= sd
->last_balance
+ interval
;
6802 if (time_after(*next_balance
, next
))
6803 *next_balance
= next
;
6807 * idle_balance is called by schedule() if this_cpu is about to become
6808 * idle. Attempts to pull tasks from other CPUs.
6810 static int idle_balance(struct rq
*this_rq
)
6812 unsigned long next_balance
= jiffies
+ HZ
;
6813 int this_cpu
= this_rq
->cpu
;
6814 struct sched_domain
*sd
;
6815 int pulled_task
= 0;
6818 idle_enter_fair(this_rq
);
6821 * We must set idle_stamp _before_ calling idle_balance(), such that we
6822 * measure the duration of idle_balance() as idle time.
6824 this_rq
->idle_stamp
= rq_clock(this_rq
);
6826 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
6827 !this_rq
->rd
->overload
) {
6829 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
6831 update_next_balance(sd
, 0, &next_balance
);
6838 * Drop the rq->lock, but keep IRQ/preempt disabled.
6840 raw_spin_unlock(&this_rq
->lock
);
6842 update_blocked_averages(this_cpu
);
6844 for_each_domain(this_cpu
, sd
) {
6845 int continue_balancing
= 1;
6846 u64 t0
, domain_cost
;
6848 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6851 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
6852 update_next_balance(sd
, 0, &next_balance
);
6856 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
6857 t0
= sched_clock_cpu(this_cpu
);
6859 pulled_task
= load_balance(this_cpu
, this_rq
,
6861 &continue_balancing
);
6863 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
6864 if (domain_cost
> sd
->max_newidle_lb_cost
)
6865 sd
->max_newidle_lb_cost
= domain_cost
;
6867 curr_cost
+= domain_cost
;
6870 update_next_balance(sd
, 0, &next_balance
);
6873 * Stop searching for tasks to pull if there are
6874 * now runnable tasks on this rq.
6876 if (pulled_task
|| this_rq
->nr_running
> 0)
6881 raw_spin_lock(&this_rq
->lock
);
6883 if (curr_cost
> this_rq
->max_idle_balance_cost
)
6884 this_rq
->max_idle_balance_cost
= curr_cost
;
6887 * While browsing the domains, we released the rq lock, a task could
6888 * have been enqueued in the meantime. Since we're not going idle,
6889 * pretend we pulled a task.
6891 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
6895 /* Move the next balance forward */
6896 if (time_after(this_rq
->next_balance
, next_balance
))
6897 this_rq
->next_balance
= next_balance
;
6899 /* Is there a task of a high priority class? */
6900 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
6904 idle_exit_fair(this_rq
);
6905 this_rq
->idle_stamp
= 0;
6912 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6913 * running tasks off the busiest CPU onto idle CPUs. It requires at
6914 * least 1 task to be running on each physical CPU where possible, and
6915 * avoids physical / logical imbalances.
6917 static int active_load_balance_cpu_stop(void *data
)
6919 struct rq
*busiest_rq
= data
;
6920 int busiest_cpu
= cpu_of(busiest_rq
);
6921 int target_cpu
= busiest_rq
->push_cpu
;
6922 struct rq
*target_rq
= cpu_rq(target_cpu
);
6923 struct sched_domain
*sd
;
6925 raw_spin_lock_irq(&busiest_rq
->lock
);
6927 /* make sure the requested cpu hasn't gone down in the meantime */
6928 if (unlikely(busiest_cpu
!= smp_processor_id() ||
6929 !busiest_rq
->active_balance
))
6932 /* Is there any task to move? */
6933 if (busiest_rq
->nr_running
<= 1)
6937 * This condition is "impossible", if it occurs
6938 * we need to fix it. Originally reported by
6939 * Bjorn Helgaas on a 128-cpu setup.
6941 BUG_ON(busiest_rq
== target_rq
);
6943 /* move a task from busiest_rq to target_rq */
6944 double_lock_balance(busiest_rq
, target_rq
);
6946 /* Search for an sd spanning us and the target CPU. */
6948 for_each_domain(target_cpu
, sd
) {
6949 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
6950 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
6955 struct lb_env env
= {
6957 .dst_cpu
= target_cpu
,
6958 .dst_rq
= target_rq
,
6959 .src_cpu
= busiest_rq
->cpu
,
6960 .src_rq
= busiest_rq
,
6964 schedstat_inc(sd
, alb_count
);
6966 if (move_one_task(&env
))
6967 schedstat_inc(sd
, alb_pushed
);
6969 schedstat_inc(sd
, alb_failed
);
6972 double_unlock_balance(busiest_rq
, target_rq
);
6974 busiest_rq
->active_balance
= 0;
6975 raw_spin_unlock_irq(&busiest_rq
->lock
);
6979 static inline int on_null_domain(struct rq
*rq
)
6981 return unlikely(!rcu_dereference_sched(rq
->sd
));
6984 #ifdef CONFIG_NO_HZ_COMMON
6986 * idle load balancing details
6987 * - When one of the busy CPUs notice that there may be an idle rebalancing
6988 * needed, they will kick the idle load balancer, which then does idle
6989 * load balancing for all the idle CPUs.
6992 cpumask_var_t idle_cpus_mask
;
6994 unsigned long next_balance
; /* in jiffy units */
6995 } nohz ____cacheline_aligned
;
6997 static inline int find_new_ilb(void)
6999 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
7001 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
7008 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7009 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7010 * CPU (if there is one).
7012 static void nohz_balancer_kick(void)
7016 nohz
.next_balance
++;
7018 ilb_cpu
= find_new_ilb();
7020 if (ilb_cpu
>= nr_cpu_ids
)
7023 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
7026 * Use smp_send_reschedule() instead of resched_cpu().
7027 * This way we generate a sched IPI on the target cpu which
7028 * is idle. And the softirq performing nohz idle load balance
7029 * will be run before returning from the IPI.
7031 smp_send_reschedule(ilb_cpu
);
7035 static inline void nohz_balance_exit_idle(int cpu
)
7037 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
7039 * Completely isolated CPUs don't ever set, so we must test.
7041 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
7042 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
7043 atomic_dec(&nohz
.nr_cpus
);
7045 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7049 static inline void set_cpu_sd_state_busy(void)
7051 struct sched_domain
*sd
;
7052 int cpu
= smp_processor_id();
7055 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7057 if (!sd
|| !sd
->nohz_idle
)
7061 atomic_inc(&sd
->groups
->sgc
->nr_busy_cpus
);
7066 void set_cpu_sd_state_idle(void)
7068 struct sched_domain
*sd
;
7069 int cpu
= smp_processor_id();
7072 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7074 if (!sd
|| sd
->nohz_idle
)
7078 atomic_dec(&sd
->groups
->sgc
->nr_busy_cpus
);
7084 * This routine will record that the cpu is going idle with tick stopped.
7085 * This info will be used in performing idle load balancing in the future.
7087 void nohz_balance_enter_idle(int cpu
)
7090 * If this cpu is going down, then nothing needs to be done.
7092 if (!cpu_active(cpu
))
7095 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
7099 * If we're a completely isolated CPU, we don't play.
7101 if (on_null_domain(cpu_rq(cpu
)))
7104 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
7105 atomic_inc(&nohz
.nr_cpus
);
7106 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7109 static int sched_ilb_notifier(struct notifier_block
*nfb
,
7110 unsigned long action
, void *hcpu
)
7112 switch (action
& ~CPU_TASKS_FROZEN
) {
7114 nohz_balance_exit_idle(smp_processor_id());
7122 static DEFINE_SPINLOCK(balancing
);
7125 * Scale the max load_balance interval with the number of CPUs in the system.
7126 * This trades load-balance latency on larger machines for less cross talk.
7128 void update_max_interval(void)
7130 max_load_balance_interval
= HZ
*num_online_cpus()/10;
7134 * It checks each scheduling domain to see if it is due to be balanced,
7135 * and initiates a balancing operation if so.
7137 * Balancing parameters are set up in init_sched_domains.
7139 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
7141 int continue_balancing
= 1;
7143 unsigned long interval
;
7144 struct sched_domain
*sd
;
7145 /* Earliest time when we have to do rebalance again */
7146 unsigned long next_balance
= jiffies
+ 60*HZ
;
7147 int update_next_balance
= 0;
7148 int need_serialize
, need_decay
= 0;
7151 update_blocked_averages(cpu
);
7154 for_each_domain(cpu
, sd
) {
7156 * Decay the newidle max times here because this is a regular
7157 * visit to all the domains. Decay ~1% per second.
7159 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
7160 sd
->max_newidle_lb_cost
=
7161 (sd
->max_newidle_lb_cost
* 253) / 256;
7162 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
7165 max_cost
+= sd
->max_newidle_lb_cost
;
7167 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7171 * Stop the load balance at this level. There is another
7172 * CPU in our sched group which is doing load balancing more
7175 if (!continue_balancing
) {
7181 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7183 need_serialize
= sd
->flags
& SD_SERIALIZE
;
7184 if (need_serialize
) {
7185 if (!spin_trylock(&balancing
))
7189 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
7190 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
7192 * The LBF_DST_PINNED logic could have changed
7193 * env->dst_cpu, so we can't know our idle
7194 * state even if we migrated tasks. Update it.
7196 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
7198 sd
->last_balance
= jiffies
;
7199 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7202 spin_unlock(&balancing
);
7204 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
7205 next_balance
= sd
->last_balance
+ interval
;
7206 update_next_balance
= 1;
7211 * Ensure the rq-wide value also decays but keep it at a
7212 * reasonable floor to avoid funnies with rq->avg_idle.
7214 rq
->max_idle_balance_cost
=
7215 max((u64
)sysctl_sched_migration_cost
, max_cost
);
7220 * next_balance will be updated only when there is a need.
7221 * When the cpu is attached to null domain for ex, it will not be
7224 if (likely(update_next_balance
))
7225 rq
->next_balance
= next_balance
;
7228 #ifdef CONFIG_NO_HZ_COMMON
7230 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7231 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7233 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
7235 int this_cpu
= this_rq
->cpu
;
7239 if (idle
!= CPU_IDLE
||
7240 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
7243 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
7244 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
7248 * If this cpu gets work to do, stop the load balancing
7249 * work being done for other cpus. Next load
7250 * balancing owner will pick it up.
7255 rq
= cpu_rq(balance_cpu
);
7258 * If time for next balance is due,
7261 if (time_after_eq(jiffies
, rq
->next_balance
)) {
7262 raw_spin_lock_irq(&rq
->lock
);
7263 update_rq_clock(rq
);
7264 update_idle_cpu_load(rq
);
7265 raw_spin_unlock_irq(&rq
->lock
);
7266 rebalance_domains(rq
, CPU_IDLE
);
7269 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
7270 this_rq
->next_balance
= rq
->next_balance
;
7272 nohz
.next_balance
= this_rq
->next_balance
;
7274 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
7278 * Current heuristic for kicking the idle load balancer in the presence
7279 * of an idle cpu is the system.
7280 * - This rq has more than one task.
7281 * - At any scheduler domain level, this cpu's scheduler group has multiple
7282 * busy cpu's exceeding the group's capacity.
7283 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7284 * domain span are idle.
7286 static inline int nohz_kick_needed(struct rq
*rq
)
7288 unsigned long now
= jiffies
;
7289 struct sched_domain
*sd
;
7290 struct sched_group_capacity
*sgc
;
7291 int nr_busy
, cpu
= rq
->cpu
;
7293 if (unlikely(rq
->idle_balance
))
7297 * We may be recently in ticked or tickless idle mode. At the first
7298 * busy tick after returning from idle, we will update the busy stats.
7300 set_cpu_sd_state_busy();
7301 nohz_balance_exit_idle(cpu
);
7304 * None are in tickless mode and hence no need for NOHZ idle load
7307 if (likely(!atomic_read(&nohz
.nr_cpus
)))
7310 if (time_before(now
, nohz
.next_balance
))
7313 if (rq
->nr_running
>= 2)
7317 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7320 sgc
= sd
->groups
->sgc
;
7321 nr_busy
= atomic_read(&sgc
->nr_busy_cpus
);
7324 goto need_kick_unlock
;
7327 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
7329 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
7330 sched_domain_span(sd
)) < cpu
))
7331 goto need_kick_unlock
;
7342 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
7346 * run_rebalance_domains is triggered when needed from the scheduler tick.
7347 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7349 static void run_rebalance_domains(struct softirq_action
*h
)
7351 struct rq
*this_rq
= this_rq();
7352 enum cpu_idle_type idle
= this_rq
->idle_balance
?
7353 CPU_IDLE
: CPU_NOT_IDLE
;
7355 rebalance_domains(this_rq
, idle
);
7358 * If this cpu has a pending nohz_balance_kick, then do the
7359 * balancing on behalf of the other idle cpus whose ticks are
7362 nohz_idle_balance(this_rq
, idle
);
7366 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7368 void trigger_load_balance(struct rq
*rq
)
7370 /* Don't need to rebalance while attached to NULL domain */
7371 if (unlikely(on_null_domain(rq
)))
7374 if (time_after_eq(jiffies
, rq
->next_balance
))
7375 raise_softirq(SCHED_SOFTIRQ
);
7376 #ifdef CONFIG_NO_HZ_COMMON
7377 if (nohz_kick_needed(rq
))
7378 nohz_balancer_kick();
7382 static void rq_online_fair(struct rq
*rq
)
7386 update_runtime_enabled(rq
);
7389 static void rq_offline_fair(struct rq
*rq
)
7393 /* Ensure any throttled groups are reachable by pick_next_task */
7394 unthrottle_offline_cfs_rqs(rq
);
7397 #endif /* CONFIG_SMP */
7400 * scheduler tick hitting a task of our scheduling class:
7402 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
7404 struct cfs_rq
*cfs_rq
;
7405 struct sched_entity
*se
= &curr
->se
;
7407 for_each_sched_entity(se
) {
7408 cfs_rq
= cfs_rq_of(se
);
7409 entity_tick(cfs_rq
, se
, queued
);
7412 if (numabalancing_enabled
)
7413 task_tick_numa(rq
, curr
);
7415 update_rq_runnable_avg(rq
, 1);
7419 * called on fork with the child task as argument from the parent's context
7420 * - child not yet on the tasklist
7421 * - preemption disabled
7423 static void task_fork_fair(struct task_struct
*p
)
7425 struct cfs_rq
*cfs_rq
;
7426 struct sched_entity
*se
= &p
->se
, *curr
;
7427 int this_cpu
= smp_processor_id();
7428 struct rq
*rq
= this_rq();
7429 unsigned long flags
;
7431 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7433 update_rq_clock(rq
);
7435 cfs_rq
= task_cfs_rq(current
);
7436 curr
= cfs_rq
->curr
;
7439 * Not only the cpu but also the task_group of the parent might have
7440 * been changed after parent->se.parent,cfs_rq were copied to
7441 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7442 * of child point to valid ones.
7445 __set_task_cpu(p
, this_cpu
);
7448 update_curr(cfs_rq
);
7451 se
->vruntime
= curr
->vruntime
;
7452 place_entity(cfs_rq
, se
, 1);
7454 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
7456 * Upon rescheduling, sched_class::put_prev_task() will place
7457 * 'current' within the tree based on its new key value.
7459 swap(curr
->vruntime
, se
->vruntime
);
7460 resched_task(rq
->curr
);
7463 se
->vruntime
-= cfs_rq
->min_vruntime
;
7465 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7469 * Priority of the task has changed. Check to see if we preempt
7473 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
7479 * Reschedule if we are currently running on this runqueue and
7480 * our priority decreased, or if we are not currently running on
7481 * this runqueue and our priority is higher than the current's
7483 if (rq
->curr
== p
) {
7484 if (p
->prio
> oldprio
)
7485 resched_task(rq
->curr
);
7487 check_preempt_curr(rq
, p
, 0);
7490 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
7492 struct sched_entity
*se
= &p
->se
;
7493 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7496 * Ensure the task's vruntime is normalized, so that when it's
7497 * switched back to the fair class the enqueue_entity(.flags=0) will
7498 * do the right thing.
7500 * If it's on_rq, then the dequeue_entity(.flags=0) will already
7501 * have normalized the vruntime, if it's !on_rq, then only when
7502 * the task is sleeping will it still have non-normalized vruntime.
7504 if (!p
->on_rq
&& p
->state
!= TASK_RUNNING
) {
7506 * Fix up our vruntime so that the current sleep doesn't
7507 * cause 'unlimited' sleep bonus.
7509 place_entity(cfs_rq
, se
, 0);
7510 se
->vruntime
-= cfs_rq
->min_vruntime
;
7515 * Remove our load from contribution when we leave sched_fair
7516 * and ensure we don't carry in an old decay_count if we
7519 if (se
->avg
.decay_count
) {
7520 __synchronize_entity_decay(se
);
7521 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7527 * We switched to the sched_fair class.
7529 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7531 struct sched_entity
*se
= &p
->se
;
7532 #ifdef CONFIG_FAIR_GROUP_SCHED
7534 * Since the real-depth could have been changed (only FAIR
7535 * class maintain depth value), reset depth properly.
7537 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7543 * We were most likely switched from sched_rt, so
7544 * kick off the schedule if running, otherwise just see
7545 * if we can still preempt the current task.
7548 resched_task(rq
->curr
);
7550 check_preempt_curr(rq
, p
, 0);
7553 /* Account for a task changing its policy or group.
7555 * This routine is mostly called to set cfs_rq->curr field when a task
7556 * migrates between groups/classes.
7558 static void set_curr_task_fair(struct rq
*rq
)
7560 struct sched_entity
*se
= &rq
->curr
->se
;
7562 for_each_sched_entity(se
) {
7563 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7565 set_next_entity(cfs_rq
, se
);
7566 /* ensure bandwidth has been allocated on our new cfs_rq */
7567 account_cfs_rq_runtime(cfs_rq
, 0);
7571 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7573 cfs_rq
->tasks_timeline
= RB_ROOT
;
7574 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7575 #ifndef CONFIG_64BIT
7576 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7579 atomic64_set(&cfs_rq
->decay_counter
, 1);
7580 atomic_long_set(&cfs_rq
->removed_load
, 0);
7584 #ifdef CONFIG_FAIR_GROUP_SCHED
7585 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
7587 struct sched_entity
*se
= &p
->se
;
7588 struct cfs_rq
*cfs_rq
;
7591 * If the task was not on the rq at the time of this cgroup movement
7592 * it must have been asleep, sleeping tasks keep their ->vruntime
7593 * absolute on their old rq until wakeup (needed for the fair sleeper
7594 * bonus in place_entity()).
7596 * If it was on the rq, we've just 'preempted' it, which does convert
7597 * ->vruntime to a relative base.
7599 * Make sure both cases convert their relative position when migrating
7600 * to another cgroup's rq. This does somewhat interfere with the
7601 * fair sleeper stuff for the first placement, but who cares.
7604 * When !on_rq, vruntime of the task has usually NOT been normalized.
7605 * But there are some cases where it has already been normalized:
7607 * - Moving a forked child which is waiting for being woken up by
7608 * wake_up_new_task().
7609 * - Moving a task which has been woken up by try_to_wake_up() and
7610 * waiting for actually being woken up by sched_ttwu_pending().
7612 * To prevent boost or penalty in the new cfs_rq caused by delta
7613 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7615 if (!on_rq
&& (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7619 se
->vruntime
-= cfs_rq_of(se
)->min_vruntime
;
7620 set_task_rq(p
, task_cpu(p
));
7621 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7623 cfs_rq
= cfs_rq_of(se
);
7624 se
->vruntime
+= cfs_rq
->min_vruntime
;
7627 * migrate_task_rq_fair() will have removed our previous
7628 * contribution, but we must synchronize for ongoing future
7631 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7632 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
7637 void free_fair_sched_group(struct task_group
*tg
)
7641 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7643 for_each_possible_cpu(i
) {
7645 kfree(tg
->cfs_rq
[i
]);
7654 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7656 struct cfs_rq
*cfs_rq
;
7657 struct sched_entity
*se
;
7660 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7663 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7667 tg
->shares
= NICE_0_LOAD
;
7669 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7671 for_each_possible_cpu(i
) {
7672 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7673 GFP_KERNEL
, cpu_to_node(i
));
7677 se
= kzalloc_node(sizeof(struct sched_entity
),
7678 GFP_KERNEL
, cpu_to_node(i
));
7682 init_cfs_rq(cfs_rq
);
7683 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
7694 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7696 struct rq
*rq
= cpu_rq(cpu
);
7697 unsigned long flags
;
7700 * Only empty task groups can be destroyed; so we can speculatively
7701 * check on_list without danger of it being re-added.
7703 if (!tg
->cfs_rq
[cpu
]->on_list
)
7706 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7707 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
7708 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7711 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7712 struct sched_entity
*se
, int cpu
,
7713 struct sched_entity
*parent
)
7715 struct rq
*rq
= cpu_rq(cpu
);
7719 init_cfs_rq_runtime(cfs_rq
);
7721 tg
->cfs_rq
[cpu
] = cfs_rq
;
7724 /* se could be NULL for root_task_group */
7729 se
->cfs_rq
= &rq
->cfs
;
7732 se
->cfs_rq
= parent
->my_q
;
7733 se
->depth
= parent
->depth
+ 1;
7737 /* guarantee group entities always have weight */
7738 update_load_set(&se
->load
, NICE_0_LOAD
);
7739 se
->parent
= parent
;
7742 static DEFINE_MUTEX(shares_mutex
);
7744 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7747 unsigned long flags
;
7750 * We can't change the weight of the root cgroup.
7755 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
7757 mutex_lock(&shares_mutex
);
7758 if (tg
->shares
== shares
)
7761 tg
->shares
= shares
;
7762 for_each_possible_cpu(i
) {
7763 struct rq
*rq
= cpu_rq(i
);
7764 struct sched_entity
*se
;
7767 /* Propagate contribution to hierarchy */
7768 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7770 /* Possible calls to update_curr() need rq clock */
7771 update_rq_clock(rq
);
7772 for_each_sched_entity(se
)
7773 update_cfs_shares(group_cfs_rq(se
));
7774 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7778 mutex_unlock(&shares_mutex
);
7781 #else /* CONFIG_FAIR_GROUP_SCHED */
7783 void free_fair_sched_group(struct task_group
*tg
) { }
7785 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7790 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
7792 #endif /* CONFIG_FAIR_GROUP_SCHED */
7795 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
7797 struct sched_entity
*se
= &task
->se
;
7798 unsigned int rr_interval
= 0;
7801 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7804 if (rq
->cfs
.load
.weight
)
7805 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
7811 * All the scheduling class methods:
7813 const struct sched_class fair_sched_class
= {
7814 .next
= &idle_sched_class
,
7815 .enqueue_task
= enqueue_task_fair
,
7816 .dequeue_task
= dequeue_task_fair
,
7817 .yield_task
= yield_task_fair
,
7818 .yield_to_task
= yield_to_task_fair
,
7820 .check_preempt_curr
= check_preempt_wakeup
,
7822 .pick_next_task
= pick_next_task_fair
,
7823 .put_prev_task
= put_prev_task_fair
,
7826 .select_task_rq
= select_task_rq_fair
,
7827 .migrate_task_rq
= migrate_task_rq_fair
,
7829 .rq_online
= rq_online_fair
,
7830 .rq_offline
= rq_offline_fair
,
7832 .task_waking
= task_waking_fair
,
7835 .set_curr_task
= set_curr_task_fair
,
7836 .task_tick
= task_tick_fair
,
7837 .task_fork
= task_fork_fair
,
7839 .prio_changed
= prio_changed_fair
,
7840 .switched_from
= switched_from_fair
,
7841 .switched_to
= switched_to_fair
,
7843 .get_rr_interval
= get_rr_interval_fair
,
7845 #ifdef CONFIG_FAIR_GROUP_SCHED
7846 .task_move_group
= task_move_group_fair
,
7850 #ifdef CONFIG_SCHED_DEBUG
7851 void print_cfs_stats(struct seq_file
*m
, int cpu
)
7853 struct cfs_rq
*cfs_rq
;
7856 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
7857 print_cfs_rq(m
, cpu
, cfs_rq
);
7862 __init
void init_sched_fair_class(void)
7865 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7867 #ifdef CONFIG_NO_HZ_COMMON
7868 nohz
.next_balance
= jiffies
;
7869 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7870 cpu_notifier(sched_ilb_notifier
, 0);