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 ? */
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
)
339 /* return depth at which a sched entity is present in the hierarchy */
340 static inline int depth_se(struct sched_entity
*se
)
344 for_each_sched_entity(se
)
351 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
353 int se_depth
, pse_depth
;
356 * preemption test can be made between sibling entities who are in the
357 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
358 * both tasks until we find their ancestors who are siblings of common
362 /* First walk up until both entities are at same depth */
363 se_depth
= depth_se(*se
);
364 pse_depth
= depth_se(*pse
);
366 while (se_depth
> pse_depth
) {
368 *se
= parent_entity(*se
);
371 while (pse_depth
> se_depth
) {
373 *pse
= parent_entity(*pse
);
376 while (!is_same_group(*se
, *pse
)) {
377 *se
= parent_entity(*se
);
378 *pse
= parent_entity(*pse
);
382 #else /* !CONFIG_FAIR_GROUP_SCHED */
384 static inline struct task_struct
*task_of(struct sched_entity
*se
)
386 return container_of(se
, struct task_struct
, se
);
389 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
391 return container_of(cfs_rq
, struct rq
, cfs
);
394 #define entity_is_task(se) 1
396 #define for_each_sched_entity(se) \
397 for (; se; se = NULL)
399 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
401 return &task_rq(p
)->cfs
;
404 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
406 struct task_struct
*p
= task_of(se
);
407 struct rq
*rq
= task_rq(p
);
412 /* runqueue "owned" by this group */
413 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
418 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
422 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
426 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
427 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
430 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
435 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
441 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
445 #endif /* CONFIG_FAIR_GROUP_SCHED */
447 static __always_inline
448 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
450 /**************************************************************
451 * Scheduling class tree data structure manipulation methods:
454 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
456 s64 delta
= (s64
)(vruntime
- max_vruntime
);
458 max_vruntime
= vruntime
;
463 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
465 s64 delta
= (s64
)(vruntime
- min_vruntime
);
467 min_vruntime
= vruntime
;
472 static inline int entity_before(struct sched_entity
*a
,
473 struct sched_entity
*b
)
475 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
478 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
480 u64 vruntime
= cfs_rq
->min_vruntime
;
483 vruntime
= cfs_rq
->curr
->vruntime
;
485 if (cfs_rq
->rb_leftmost
) {
486 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
491 vruntime
= se
->vruntime
;
493 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
496 /* ensure we never gain time by being placed backwards. */
497 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
500 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
505 * Enqueue an entity into the rb-tree:
507 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
509 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
510 struct rb_node
*parent
= NULL
;
511 struct sched_entity
*entry
;
515 * Find the right place in the rbtree:
519 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
521 * We dont care about collisions. Nodes with
522 * the same key stay together.
524 if (entity_before(se
, entry
)) {
525 link
= &parent
->rb_left
;
527 link
= &parent
->rb_right
;
533 * Maintain a cache of leftmost tree entries (it is frequently
537 cfs_rq
->rb_leftmost
= &se
->run_node
;
539 rb_link_node(&se
->run_node
, parent
, link
);
540 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
543 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
545 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
546 struct rb_node
*next_node
;
548 next_node
= rb_next(&se
->run_node
);
549 cfs_rq
->rb_leftmost
= next_node
;
552 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
555 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
557 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
562 return rb_entry(left
, struct sched_entity
, run_node
);
565 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
567 struct rb_node
*next
= rb_next(&se
->run_node
);
572 return rb_entry(next
, struct sched_entity
, run_node
);
575 #ifdef CONFIG_SCHED_DEBUG
576 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
578 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
583 return rb_entry(last
, struct sched_entity
, run_node
);
586 /**************************************************************
587 * Scheduling class statistics methods:
590 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
591 void __user
*buffer
, size_t *lenp
,
594 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
595 int factor
= get_update_sysctl_factor();
600 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
601 sysctl_sched_min_granularity
);
603 #define WRT_SYSCTL(name) \
604 (normalized_sysctl_##name = sysctl_##name / (factor))
605 WRT_SYSCTL(sched_min_granularity
);
606 WRT_SYSCTL(sched_latency
);
607 WRT_SYSCTL(sched_wakeup_granularity
);
617 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
619 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
620 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
626 * The idea is to set a period in which each task runs once.
628 * When there are too many tasks (sched_nr_latency) we have to stretch
629 * this period because otherwise the slices get too small.
631 * p = (nr <= nl) ? l : l*nr/nl
633 static u64
__sched_period(unsigned long nr_running
)
635 u64 period
= sysctl_sched_latency
;
636 unsigned long nr_latency
= sched_nr_latency
;
638 if (unlikely(nr_running
> nr_latency
)) {
639 period
= sysctl_sched_min_granularity
;
640 period
*= nr_running
;
647 * We calculate the wall-time slice from the period by taking a part
648 * proportional to the weight.
652 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
654 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
656 for_each_sched_entity(se
) {
657 struct load_weight
*load
;
658 struct load_weight lw
;
660 cfs_rq
= cfs_rq_of(se
);
661 load
= &cfs_rq
->load
;
663 if (unlikely(!se
->on_rq
)) {
666 update_load_add(&lw
, se
->load
.weight
);
669 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
675 * We calculate the vruntime slice of a to-be-inserted task.
679 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
681 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
685 static unsigned long task_h_load(struct task_struct
*p
);
687 static inline void __update_task_entity_contrib(struct sched_entity
*se
);
689 /* Give new task start runnable values to heavy its load in infant time */
690 void init_task_runnable_average(struct task_struct
*p
)
694 p
->se
.avg
.decay_count
= 0;
695 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
696 p
->se
.avg
.runnable_avg_sum
= slice
;
697 p
->se
.avg
.runnable_avg_period
= slice
;
698 __update_task_entity_contrib(&p
->se
);
701 void init_task_runnable_average(struct task_struct
*p
)
707 * Update the current task's runtime statistics.
709 static void update_curr(struct cfs_rq
*cfs_rq
)
711 struct sched_entity
*curr
= cfs_rq
->curr
;
712 u64 now
= rq_clock_task(rq_of(cfs_rq
));
718 delta_exec
= now
- curr
->exec_start
;
719 if (unlikely((s64
)delta_exec
<= 0))
722 curr
->exec_start
= now
;
724 schedstat_set(curr
->statistics
.exec_max
,
725 max(delta_exec
, curr
->statistics
.exec_max
));
727 curr
->sum_exec_runtime
+= delta_exec
;
728 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
730 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
731 update_min_vruntime(cfs_rq
);
733 if (entity_is_task(curr
)) {
734 struct task_struct
*curtask
= task_of(curr
);
736 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
737 cpuacct_charge(curtask
, delta_exec
);
738 account_group_exec_runtime(curtask
, delta_exec
);
741 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
745 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
747 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
751 * Task is being enqueued - update stats:
753 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
756 * Are we enqueueing a waiting task? (for current tasks
757 * a dequeue/enqueue event is a NOP)
759 if (se
!= cfs_rq
->curr
)
760 update_stats_wait_start(cfs_rq
, se
);
764 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
766 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
767 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
768 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
769 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
770 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
771 #ifdef CONFIG_SCHEDSTATS
772 if (entity_is_task(se
)) {
773 trace_sched_stat_wait(task_of(se
),
774 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
777 schedstat_set(se
->statistics
.wait_start
, 0);
781 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
784 * Mark the end of the wait period if dequeueing a
787 if (se
!= cfs_rq
->curr
)
788 update_stats_wait_end(cfs_rq
, se
);
792 * We are picking a new current task - update its stats:
795 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
798 * We are starting a new run period:
800 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
803 /**************************************************
804 * Scheduling class queueing methods:
807 #ifdef CONFIG_NUMA_BALANCING
809 * Approximate time to scan a full NUMA task in ms. The task scan period is
810 * calculated based on the tasks virtual memory size and
811 * numa_balancing_scan_size.
813 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
814 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
816 /* Portion of address space to scan in MB */
817 unsigned int sysctl_numa_balancing_scan_size
= 256;
819 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
820 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
823 * After skipping a page migration on a shared page, skip N more numa page
824 * migrations unconditionally. This reduces the number of NUMA migrations
825 * in shared memory workloads, and has the effect of pulling tasks towards
826 * where their memory lives, over pulling the memory towards the task.
828 unsigned int sysctl_numa_balancing_migrate_deferred
= 16;
830 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
832 unsigned long rss
= 0;
833 unsigned long nr_scan_pages
;
836 * Calculations based on RSS as non-present and empty pages are skipped
837 * by the PTE scanner and NUMA hinting faults should be trapped based
840 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
841 rss
= get_mm_rss(p
->mm
);
845 rss
= round_up(rss
, nr_scan_pages
);
846 return rss
/ nr_scan_pages
;
849 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
850 #define MAX_SCAN_WINDOW 2560
852 static unsigned int task_scan_min(struct task_struct
*p
)
854 unsigned int scan
, floor
;
855 unsigned int windows
= 1;
857 if (sysctl_numa_balancing_scan_size
< MAX_SCAN_WINDOW
)
858 windows
= MAX_SCAN_WINDOW
/ sysctl_numa_balancing_scan_size
;
859 floor
= 1000 / windows
;
861 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
862 return max_t(unsigned int, floor
, scan
);
865 static unsigned int task_scan_max(struct task_struct
*p
)
867 unsigned int smin
= task_scan_min(p
);
870 /* Watch for min being lower than max due to floor calculations */
871 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
872 return max(smin
, smax
);
875 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
877 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
878 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
881 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
883 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
884 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
890 spinlock_t lock
; /* nr_tasks, tasks */
893 struct list_head task_list
;
896 unsigned long total_faults
;
897 unsigned long faults
[0];
900 pid_t
task_numa_group_id(struct task_struct
*p
)
902 return p
->numa_group
? p
->numa_group
->gid
: 0;
905 static inline int task_faults_idx(int nid
, int priv
)
907 return 2 * nid
+ priv
;
910 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
915 return p
->numa_faults
[task_faults_idx(nid
, 0)] +
916 p
->numa_faults
[task_faults_idx(nid
, 1)];
919 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
924 return p
->numa_group
->faults
[task_faults_idx(nid
, 0)] +
925 p
->numa_group
->faults
[task_faults_idx(nid
, 1)];
929 * These return the fraction of accesses done by a particular task, or
930 * task group, on a particular numa node. The group weight is given a
931 * larger multiplier, in order to group tasks together that are almost
932 * evenly spread out between numa nodes.
934 static inline unsigned long task_weight(struct task_struct
*p
, int nid
)
936 unsigned long total_faults
;
941 total_faults
= p
->total_numa_faults
;
946 return 1000 * task_faults(p
, nid
) / total_faults
;
949 static inline unsigned long group_weight(struct task_struct
*p
, int nid
)
951 if (!p
->numa_group
|| !p
->numa_group
->total_faults
)
954 return 1000 * group_faults(p
, nid
) / p
->numa_group
->total_faults
;
957 static unsigned long weighted_cpuload(const int cpu
);
958 static unsigned long source_load(int cpu
, int type
);
959 static unsigned long target_load(int cpu
, int type
);
960 static unsigned long power_of(int cpu
);
961 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
963 /* Cached statistics for all CPUs within a node */
965 unsigned long nr_running
;
968 /* Total compute capacity of CPUs on a node */
971 /* Approximate capacity in terms of runnable tasks on a node */
972 unsigned long capacity
;
977 * XXX borrowed from update_sg_lb_stats
979 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
983 memset(ns
, 0, sizeof(*ns
));
984 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
985 struct rq
*rq
= cpu_rq(cpu
);
987 ns
->nr_running
+= rq
->nr_running
;
988 ns
->load
+= weighted_cpuload(cpu
);
989 ns
->power
+= power_of(cpu
);
995 * If we raced with hotplug and there are no CPUs left in our mask
996 * the @ns structure is NULL'ed and task_numa_compare() will
997 * not find this node attractive.
999 * We'll either bail at !has_capacity, or we'll detect a huge imbalance
1005 ns
->load
= (ns
->load
* SCHED_POWER_SCALE
) / ns
->power
;
1006 ns
->capacity
= DIV_ROUND_CLOSEST(ns
->power
, SCHED_POWER_SCALE
);
1007 ns
->has_capacity
= (ns
->nr_running
< ns
->capacity
);
1010 struct task_numa_env
{
1011 struct task_struct
*p
;
1013 int src_cpu
, src_nid
;
1014 int dst_cpu
, dst_nid
;
1016 struct numa_stats src_stats
, dst_stats
;
1020 struct task_struct
*best_task
;
1025 static void task_numa_assign(struct task_numa_env
*env
,
1026 struct task_struct
*p
, long imp
)
1029 put_task_struct(env
->best_task
);
1034 env
->best_imp
= imp
;
1035 env
->best_cpu
= env
->dst_cpu
;
1039 * This checks if the overall compute and NUMA accesses of the system would
1040 * be improved if the source tasks was migrated to the target dst_cpu taking
1041 * into account that it might be best if task running on the dst_cpu should
1042 * be exchanged with the source task
1044 static void task_numa_compare(struct task_numa_env
*env
,
1045 long taskimp
, long groupimp
)
1047 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1048 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1049 struct task_struct
*cur
;
1050 long dst_load
, src_load
;
1052 long imp
= (groupimp
> 0) ? groupimp
: taskimp
;
1055 cur
= ACCESS_ONCE(dst_rq
->curr
);
1056 if (cur
->pid
== 0) /* idle */
1060 * "imp" is the fault differential for the source task between the
1061 * source and destination node. Calculate the total differential for
1062 * the source task and potential destination task. The more negative
1063 * the value is, the more rmeote accesses that would be expected to
1064 * be incurred if the tasks were swapped.
1067 /* Skip this swap candidate if cannot move to the source cpu */
1068 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1072 * If dst and source tasks are in the same NUMA group, or not
1073 * in any group then look only at task weights.
1075 if (cur
->numa_group
== env
->p
->numa_group
) {
1076 imp
= taskimp
+ task_weight(cur
, env
->src_nid
) -
1077 task_weight(cur
, env
->dst_nid
);
1079 * Add some hysteresis to prevent swapping the
1080 * tasks within a group over tiny differences.
1082 if (cur
->numa_group
)
1086 * Compare the group weights. If a task is all by
1087 * itself (not part of a group), use the task weight
1090 if (env
->p
->numa_group
)
1095 if (cur
->numa_group
)
1096 imp
+= group_weight(cur
, env
->src_nid
) -
1097 group_weight(cur
, env
->dst_nid
);
1099 imp
+= task_weight(cur
, env
->src_nid
) -
1100 task_weight(cur
, env
->dst_nid
);
1104 if (imp
< env
->best_imp
)
1108 /* Is there capacity at our destination? */
1109 if (env
->src_stats
.has_capacity
&&
1110 !env
->dst_stats
.has_capacity
)
1116 /* Balance doesn't matter much if we're running a task per cpu */
1117 if (src_rq
->nr_running
== 1 && dst_rq
->nr_running
== 1)
1121 * In the overloaded case, try and keep the load balanced.
1124 dst_load
= env
->dst_stats
.load
;
1125 src_load
= env
->src_stats
.load
;
1127 /* XXX missing power terms */
1128 load
= task_h_load(env
->p
);
1133 load
= task_h_load(cur
);
1138 /* make src_load the smaller */
1139 if (dst_load
< src_load
)
1140 swap(dst_load
, src_load
);
1142 if (src_load
* env
->imbalance_pct
< dst_load
* 100)
1146 task_numa_assign(env
, cur
, imp
);
1151 static void task_numa_find_cpu(struct task_numa_env
*env
,
1152 long taskimp
, long groupimp
)
1156 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1157 /* Skip this CPU if the source task cannot migrate */
1158 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1162 task_numa_compare(env
, taskimp
, groupimp
);
1166 static int task_numa_migrate(struct task_struct
*p
)
1168 struct task_numa_env env
= {
1171 .src_cpu
= task_cpu(p
),
1172 .src_nid
= task_node(p
),
1174 .imbalance_pct
= 112,
1180 struct sched_domain
*sd
;
1181 unsigned long taskweight
, groupweight
;
1183 long taskimp
, groupimp
;
1186 * Pick the lowest SD_NUMA domain, as that would have the smallest
1187 * imbalance and would be the first to start moving tasks about.
1189 * And we want to avoid any moving of tasks about, as that would create
1190 * random movement of tasks -- counter the numa conditions we're trying
1194 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1196 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1200 * Cpusets can break the scheduler domain tree into smaller
1201 * balance domains, some of which do not cross NUMA boundaries.
1202 * Tasks that are "trapped" in such domains cannot be migrated
1203 * elsewhere, so there is no point in (re)trying.
1205 if (unlikely(!sd
)) {
1206 p
->numa_preferred_nid
= task_node(p
);
1210 taskweight
= task_weight(p
, env
.src_nid
);
1211 groupweight
= group_weight(p
, env
.src_nid
);
1212 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1213 env
.dst_nid
= p
->numa_preferred_nid
;
1214 taskimp
= task_weight(p
, env
.dst_nid
) - taskweight
;
1215 groupimp
= group_weight(p
, env
.dst_nid
) - groupweight
;
1216 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1218 /* If the preferred nid has capacity, try to use it. */
1219 if (env
.dst_stats
.has_capacity
)
1220 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1222 /* No space available on the preferred nid. Look elsewhere. */
1223 if (env
.best_cpu
== -1) {
1224 for_each_online_node(nid
) {
1225 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1228 /* Only consider nodes where both task and groups benefit */
1229 taskimp
= task_weight(p
, nid
) - taskweight
;
1230 groupimp
= group_weight(p
, nid
) - groupweight
;
1231 if (taskimp
< 0 && groupimp
< 0)
1235 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1236 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1240 /* No better CPU than the current one was found. */
1241 if (env
.best_cpu
== -1)
1244 sched_setnuma(p
, env
.dst_nid
);
1247 * Reset the scan period if the task is being rescheduled on an
1248 * alternative node to recheck if the tasks is now properly placed.
1250 p
->numa_scan_period
= task_scan_min(p
);
1252 if (env
.best_task
== NULL
) {
1253 int ret
= migrate_task_to(p
, env
.best_cpu
);
1257 ret
= migrate_swap(p
, env
.best_task
);
1258 put_task_struct(env
.best_task
);
1262 /* Attempt to migrate a task to a CPU on the preferred node. */
1263 static void numa_migrate_preferred(struct task_struct
*p
)
1265 /* This task has no NUMA fault statistics yet */
1266 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults
))
1269 /* Periodically retry migrating the task to the preferred node */
1270 p
->numa_migrate_retry
= jiffies
+ HZ
;
1272 /* Success if task is already running on preferred CPU */
1273 if (task_node(p
) == p
->numa_preferred_nid
)
1276 /* Otherwise, try migrate to a CPU on the preferred node */
1277 task_numa_migrate(p
);
1281 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1282 * increments. The more local the fault statistics are, the higher the scan
1283 * period will be for the next scan window. If local/remote ratio is below
1284 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1285 * scan period will decrease
1287 #define NUMA_PERIOD_SLOTS 10
1288 #define NUMA_PERIOD_THRESHOLD 3
1291 * Increase the scan period (slow down scanning) if the majority of
1292 * our memory is already on our local node, or if the majority of
1293 * the page accesses are shared with other processes.
1294 * Otherwise, decrease the scan period.
1296 static void update_task_scan_period(struct task_struct
*p
,
1297 unsigned long shared
, unsigned long private)
1299 unsigned int period_slot
;
1303 unsigned long remote
= p
->numa_faults_locality
[0];
1304 unsigned long local
= p
->numa_faults_locality
[1];
1307 * If there were no record hinting faults then either the task is
1308 * completely idle or all activity is areas that are not of interest
1309 * to automatic numa balancing. Scan slower
1311 if (local
+ shared
== 0) {
1312 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1313 p
->numa_scan_period
<< 1);
1315 p
->mm
->numa_next_scan
= jiffies
+
1316 msecs_to_jiffies(p
->numa_scan_period
);
1322 * Prepare to scale scan period relative to the current period.
1323 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1324 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1325 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1327 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1328 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1329 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1330 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1333 diff
= slot
* period_slot
;
1335 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1338 * Scale scan rate increases based on sharing. There is an
1339 * inverse relationship between the degree of sharing and
1340 * the adjustment made to the scanning period. Broadly
1341 * speaking the intent is that there is little point
1342 * scanning faster if shared accesses dominate as it may
1343 * simply bounce migrations uselessly
1345 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
));
1346 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1349 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1350 task_scan_min(p
), task_scan_max(p
));
1351 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1354 static void task_numa_placement(struct task_struct
*p
)
1356 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1357 unsigned long max_faults
= 0, max_group_faults
= 0;
1358 unsigned long fault_types
[2] = { 0, 0 };
1359 spinlock_t
*group_lock
= NULL
;
1361 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1362 if (p
->numa_scan_seq
== seq
)
1364 p
->numa_scan_seq
= seq
;
1365 p
->numa_scan_period_max
= task_scan_max(p
);
1367 /* If the task is part of a group prevent parallel updates to group stats */
1368 if (p
->numa_group
) {
1369 group_lock
= &p
->numa_group
->lock
;
1370 spin_lock(group_lock
);
1373 /* Find the node with the highest number of faults */
1374 for_each_online_node(nid
) {
1375 unsigned long faults
= 0, group_faults
= 0;
1378 for (priv
= 0; priv
< 2; priv
++) {
1381 i
= task_faults_idx(nid
, priv
);
1382 diff
= -p
->numa_faults
[i
];
1384 /* Decay existing window, copy faults since last scan */
1385 p
->numa_faults
[i
] >>= 1;
1386 p
->numa_faults
[i
] += p
->numa_faults_buffer
[i
];
1387 fault_types
[priv
] += p
->numa_faults_buffer
[i
];
1388 p
->numa_faults_buffer
[i
] = 0;
1390 faults
+= p
->numa_faults
[i
];
1391 diff
+= p
->numa_faults
[i
];
1392 p
->total_numa_faults
+= diff
;
1393 if (p
->numa_group
) {
1394 /* safe because we can only change our own group */
1395 p
->numa_group
->faults
[i
] += diff
;
1396 p
->numa_group
->total_faults
+= diff
;
1397 group_faults
+= p
->numa_group
->faults
[i
];
1401 if (faults
> max_faults
) {
1402 max_faults
= faults
;
1406 if (group_faults
> max_group_faults
) {
1407 max_group_faults
= group_faults
;
1408 max_group_nid
= nid
;
1412 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1414 if (p
->numa_group
) {
1416 * If the preferred task and group nids are different,
1417 * iterate over the nodes again to find the best place.
1419 if (max_nid
!= max_group_nid
) {
1420 unsigned long weight
, max_weight
= 0;
1422 for_each_online_node(nid
) {
1423 weight
= task_weight(p
, nid
) + group_weight(p
, nid
);
1424 if (weight
> max_weight
) {
1425 max_weight
= weight
;
1431 spin_unlock(group_lock
);
1434 /* Preferred node as the node with the most faults */
1435 if (max_faults
&& max_nid
!= p
->numa_preferred_nid
) {
1436 /* Update the preferred nid and migrate task if possible */
1437 sched_setnuma(p
, max_nid
);
1438 numa_migrate_preferred(p
);
1442 static inline int get_numa_group(struct numa_group
*grp
)
1444 return atomic_inc_not_zero(&grp
->refcount
);
1447 static inline void put_numa_group(struct numa_group
*grp
)
1449 if (atomic_dec_and_test(&grp
->refcount
))
1450 kfree_rcu(grp
, rcu
);
1453 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1456 struct numa_group
*grp
, *my_grp
;
1457 struct task_struct
*tsk
;
1459 int cpu
= cpupid_to_cpu(cpupid
);
1462 if (unlikely(!p
->numa_group
)) {
1463 unsigned int size
= sizeof(struct numa_group
) +
1464 2*nr_node_ids
*sizeof(unsigned long);
1466 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1470 atomic_set(&grp
->refcount
, 1);
1471 spin_lock_init(&grp
->lock
);
1472 INIT_LIST_HEAD(&grp
->task_list
);
1475 for (i
= 0; i
< 2*nr_node_ids
; i
++)
1476 grp
->faults
[i
] = p
->numa_faults
[i
];
1478 grp
->total_faults
= p
->total_numa_faults
;
1480 list_add(&p
->numa_entry
, &grp
->task_list
);
1482 rcu_assign_pointer(p
->numa_group
, grp
);
1486 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1488 if (!cpupid_match_pid(tsk
, cpupid
))
1491 grp
= rcu_dereference(tsk
->numa_group
);
1495 my_grp
= p
->numa_group
;
1500 * Only join the other group if its bigger; if we're the bigger group,
1501 * the other task will join us.
1503 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1507 * Tie-break on the grp address.
1509 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1512 /* Always join threads in the same process. */
1513 if (tsk
->mm
== current
->mm
)
1516 /* Simple filter to avoid false positives due to PID collisions */
1517 if (flags
& TNF_SHARED
)
1520 /* Update priv based on whether false sharing was detected */
1523 if (join
&& !get_numa_group(grp
))
1531 double_lock(&my_grp
->lock
, &grp
->lock
);
1533 for (i
= 0; i
< 2*nr_node_ids
; i
++) {
1534 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
1535 grp
->faults
[i
] += p
->numa_faults
[i
];
1537 my_grp
->total_faults
-= p
->total_numa_faults
;
1538 grp
->total_faults
+= p
->total_numa_faults
;
1540 list_move(&p
->numa_entry
, &grp
->task_list
);
1544 spin_unlock(&my_grp
->lock
);
1545 spin_unlock(&grp
->lock
);
1547 rcu_assign_pointer(p
->numa_group
, grp
);
1549 put_numa_group(my_grp
);
1557 void task_numa_free(struct task_struct
*p
)
1559 struct numa_group
*grp
= p
->numa_group
;
1561 void *numa_faults
= p
->numa_faults
;
1564 spin_lock(&grp
->lock
);
1565 for (i
= 0; i
< 2*nr_node_ids
; i
++)
1566 grp
->faults
[i
] -= p
->numa_faults
[i
];
1567 grp
->total_faults
-= p
->total_numa_faults
;
1569 list_del(&p
->numa_entry
);
1571 spin_unlock(&grp
->lock
);
1572 rcu_assign_pointer(p
->numa_group
, NULL
);
1573 put_numa_group(grp
);
1576 p
->numa_faults
= NULL
;
1577 p
->numa_faults_buffer
= NULL
;
1582 * Got a PROT_NONE fault for a page on @node.
1584 void task_numa_fault(int last_cpupid
, int node
, int pages
, int flags
)
1586 struct task_struct
*p
= current
;
1587 bool migrated
= flags
& TNF_MIGRATED
;
1590 if (!numabalancing_enabled
)
1593 /* for example, ksmd faulting in a user's mm */
1597 /* Do not worry about placement if exiting */
1598 if (p
->state
== TASK_DEAD
)
1601 /* Allocate buffer to track faults on a per-node basis */
1602 if (unlikely(!p
->numa_faults
)) {
1603 int size
= sizeof(*p
->numa_faults
) * 2 * nr_node_ids
;
1605 /* numa_faults and numa_faults_buffer share the allocation */
1606 p
->numa_faults
= kzalloc(size
* 2, GFP_KERNEL
|__GFP_NOWARN
);
1607 if (!p
->numa_faults
)
1610 BUG_ON(p
->numa_faults_buffer
);
1611 p
->numa_faults_buffer
= p
->numa_faults
+ (2 * nr_node_ids
);
1612 p
->total_numa_faults
= 0;
1613 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1617 * First accesses are treated as private, otherwise consider accesses
1618 * to be private if the accessing pid has not changed
1620 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
1623 priv
= cpupid_match_pid(p
, last_cpupid
);
1624 if (!priv
&& !(flags
& TNF_NO_GROUP
))
1625 task_numa_group(p
, last_cpupid
, flags
, &priv
);
1628 task_numa_placement(p
);
1631 * Retry task to preferred node migration periodically, in case it
1632 * case it previously failed, or the scheduler moved us.
1634 if (time_after(jiffies
, p
->numa_migrate_retry
))
1635 numa_migrate_preferred(p
);
1638 p
->numa_pages_migrated
+= pages
;
1640 p
->numa_faults_buffer
[task_faults_idx(node
, priv
)] += pages
;
1641 p
->numa_faults_locality
[!!(flags
& TNF_FAULT_LOCAL
)] += pages
;
1644 static void reset_ptenuma_scan(struct task_struct
*p
)
1646 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
1647 p
->mm
->numa_scan_offset
= 0;
1651 * The expensive part of numa migration is done from task_work context.
1652 * Triggered from task_tick_numa().
1654 void task_numa_work(struct callback_head
*work
)
1656 unsigned long migrate
, next_scan
, now
= jiffies
;
1657 struct task_struct
*p
= current
;
1658 struct mm_struct
*mm
= p
->mm
;
1659 struct vm_area_struct
*vma
;
1660 unsigned long start
, end
;
1661 unsigned long nr_pte_updates
= 0;
1664 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
1666 work
->next
= work
; /* protect against double add */
1668 * Who cares about NUMA placement when they're dying.
1670 * NOTE: make sure not to dereference p->mm before this check,
1671 * exit_task_work() happens _after_ exit_mm() so we could be called
1672 * without p->mm even though we still had it when we enqueued this
1675 if (p
->flags
& PF_EXITING
)
1678 if (!mm
->numa_next_scan
) {
1679 mm
->numa_next_scan
= now
+
1680 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1684 * Enforce maximal scan/migration frequency..
1686 migrate
= mm
->numa_next_scan
;
1687 if (time_before(now
, migrate
))
1690 if (p
->numa_scan_period
== 0) {
1691 p
->numa_scan_period_max
= task_scan_max(p
);
1692 p
->numa_scan_period
= task_scan_min(p
);
1695 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1696 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
1700 * Delay this task enough that another task of this mm will likely win
1701 * the next time around.
1703 p
->node_stamp
+= 2 * TICK_NSEC
;
1705 start
= mm
->numa_scan_offset
;
1706 pages
= sysctl_numa_balancing_scan_size
;
1707 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
1711 down_read(&mm
->mmap_sem
);
1712 vma
= find_vma(mm
, start
);
1714 reset_ptenuma_scan(p
);
1718 for (; vma
; vma
= vma
->vm_next
) {
1719 if (!vma_migratable(vma
) || !vma_policy_mof(p
, vma
))
1723 * Shared library pages mapped by multiple processes are not
1724 * migrated as it is expected they are cache replicated. Avoid
1725 * hinting faults in read-only file-backed mappings or the vdso
1726 * as migrating the pages will be of marginal benefit.
1729 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
1733 start
= max(start
, vma
->vm_start
);
1734 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
1735 end
= min(end
, vma
->vm_end
);
1736 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
1739 * Scan sysctl_numa_balancing_scan_size but ensure that
1740 * at least one PTE is updated so that unused virtual
1741 * address space is quickly skipped.
1744 pages
-= (end
- start
) >> PAGE_SHIFT
;
1749 } while (end
!= vma
->vm_end
);
1754 * It is possible to reach the end of the VMA list but the last few
1755 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1756 * would find the !migratable VMA on the next scan but not reset the
1757 * scanner to the start so check it now.
1760 mm
->numa_scan_offset
= start
;
1762 reset_ptenuma_scan(p
);
1763 up_read(&mm
->mmap_sem
);
1767 * Drive the periodic memory faults..
1769 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1771 struct callback_head
*work
= &curr
->numa_work
;
1775 * We don't care about NUMA placement if we don't have memory.
1777 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
1781 * Using runtime rather than walltime has the dual advantage that
1782 * we (mostly) drive the selection from busy threads and that the
1783 * task needs to have done some actual work before we bother with
1786 now
= curr
->se
.sum_exec_runtime
;
1787 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
1789 if (now
- curr
->node_stamp
> period
) {
1790 if (!curr
->node_stamp
)
1791 curr
->numa_scan_period
= task_scan_min(curr
);
1792 curr
->node_stamp
+= period
;
1794 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
1795 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
1796 task_work_add(curr
, work
, true);
1801 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1805 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
1809 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
1812 #endif /* CONFIG_NUMA_BALANCING */
1815 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1817 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
1818 if (!parent_entity(se
))
1819 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1821 if (entity_is_task(se
)) {
1822 struct rq
*rq
= rq_of(cfs_rq
);
1824 account_numa_enqueue(rq
, task_of(se
));
1825 list_add(&se
->group_node
, &rq
->cfs_tasks
);
1828 cfs_rq
->nr_running
++;
1832 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1834 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
1835 if (!parent_entity(se
))
1836 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1837 if (entity_is_task(se
)) {
1838 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
1839 list_del_init(&se
->group_node
);
1841 cfs_rq
->nr_running
--;
1844 #ifdef CONFIG_FAIR_GROUP_SCHED
1846 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
1851 * Use this CPU's actual weight instead of the last load_contribution
1852 * to gain a more accurate current total weight. See
1853 * update_cfs_rq_load_contribution().
1855 tg_weight
= atomic_long_read(&tg
->load_avg
);
1856 tg_weight
-= cfs_rq
->tg_load_contrib
;
1857 tg_weight
+= cfs_rq
->load
.weight
;
1862 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1864 long tg_weight
, load
, shares
;
1866 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
1867 load
= cfs_rq
->load
.weight
;
1869 shares
= (tg
->shares
* load
);
1871 shares
/= tg_weight
;
1873 if (shares
< MIN_SHARES
)
1874 shares
= MIN_SHARES
;
1875 if (shares
> tg
->shares
)
1876 shares
= tg
->shares
;
1880 # else /* CONFIG_SMP */
1881 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1885 # endif /* CONFIG_SMP */
1886 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
1887 unsigned long weight
)
1890 /* commit outstanding execution time */
1891 if (cfs_rq
->curr
== se
)
1892 update_curr(cfs_rq
);
1893 account_entity_dequeue(cfs_rq
, se
);
1896 update_load_set(&se
->load
, weight
);
1899 account_entity_enqueue(cfs_rq
, se
);
1902 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
1904 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
1906 struct task_group
*tg
;
1907 struct sched_entity
*se
;
1911 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
1912 if (!se
|| throttled_hierarchy(cfs_rq
))
1915 if (likely(se
->load
.weight
== tg
->shares
))
1918 shares
= calc_cfs_shares(cfs_rq
, tg
);
1920 reweight_entity(cfs_rq_of(se
), se
, shares
);
1922 #else /* CONFIG_FAIR_GROUP_SCHED */
1923 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
1926 #endif /* CONFIG_FAIR_GROUP_SCHED */
1930 * We choose a half-life close to 1 scheduling period.
1931 * Note: The tables below are dependent on this value.
1933 #define LOAD_AVG_PERIOD 32
1934 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1935 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1937 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1938 static const u32 runnable_avg_yN_inv
[] = {
1939 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1940 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1941 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1942 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1943 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1944 0x85aac367, 0x82cd8698,
1948 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1949 * over-estimates when re-combining.
1951 static const u32 runnable_avg_yN_sum
[] = {
1952 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1953 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1954 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1959 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1961 static __always_inline u64
decay_load(u64 val
, u64 n
)
1963 unsigned int local_n
;
1967 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
1970 /* after bounds checking we can collapse to 32-bit */
1974 * As y^PERIOD = 1/2, we can combine
1975 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1976 * With a look-up table which covers k^n (n<PERIOD)
1978 * To achieve constant time decay_load.
1980 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
1981 val
>>= local_n
/ LOAD_AVG_PERIOD
;
1982 local_n
%= LOAD_AVG_PERIOD
;
1985 val
*= runnable_avg_yN_inv
[local_n
];
1986 /* We don't use SRR here since we always want to round down. */
1991 * For updates fully spanning n periods, the contribution to runnable
1992 * average will be: \Sum 1024*y^n
1994 * We can compute this reasonably efficiently by combining:
1995 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1997 static u32
__compute_runnable_contrib(u64 n
)
2001 if (likely(n
<= LOAD_AVG_PERIOD
))
2002 return runnable_avg_yN_sum
[n
];
2003 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2004 return LOAD_AVG_MAX
;
2006 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2008 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2009 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2011 n
-= LOAD_AVG_PERIOD
;
2012 } while (n
> LOAD_AVG_PERIOD
);
2014 contrib
= decay_load(contrib
, n
);
2015 return contrib
+ runnable_avg_yN_sum
[n
];
2019 * We can represent the historical contribution to runnable average as the
2020 * coefficients of a geometric series. To do this we sub-divide our runnable
2021 * history into segments of approximately 1ms (1024us); label the segment that
2022 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2024 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2026 * (now) (~1ms ago) (~2ms ago)
2028 * Let u_i denote the fraction of p_i that the entity was runnable.
2030 * We then designate the fractions u_i as our co-efficients, yielding the
2031 * following representation of historical load:
2032 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2034 * We choose y based on the with of a reasonably scheduling period, fixing:
2037 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2038 * approximately half as much as the contribution to load within the last ms
2041 * When a period "rolls over" and we have new u_0`, multiplying the previous
2042 * sum again by y is sufficient to update:
2043 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2044 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2046 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2047 struct sched_avg
*sa
,
2051 u32 runnable_contrib
;
2052 int delta_w
, decayed
= 0;
2054 delta
= now
- sa
->last_runnable_update
;
2056 * This should only happen when time goes backwards, which it
2057 * unfortunately does during sched clock init when we swap over to TSC.
2059 if ((s64
)delta
< 0) {
2060 sa
->last_runnable_update
= now
;
2065 * Use 1024ns as the unit of measurement since it's a reasonable
2066 * approximation of 1us and fast to compute.
2071 sa
->last_runnable_update
= now
;
2073 /* delta_w is the amount already accumulated against our next period */
2074 delta_w
= sa
->runnable_avg_period
% 1024;
2075 if (delta
+ delta_w
>= 1024) {
2076 /* period roll-over */
2080 * Now that we know we're crossing a period boundary, figure
2081 * out how much from delta we need to complete the current
2082 * period and accrue it.
2084 delta_w
= 1024 - delta_w
;
2086 sa
->runnable_avg_sum
+= delta_w
;
2087 sa
->runnable_avg_period
+= delta_w
;
2091 /* Figure out how many additional periods this update spans */
2092 periods
= delta
/ 1024;
2095 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2097 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2100 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2101 runnable_contrib
= __compute_runnable_contrib(periods
);
2103 sa
->runnable_avg_sum
+= runnable_contrib
;
2104 sa
->runnable_avg_period
+= runnable_contrib
;
2107 /* Remainder of delta accrued against u_0` */
2109 sa
->runnable_avg_sum
+= delta
;
2110 sa
->runnable_avg_period
+= delta
;
2115 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2116 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2118 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2119 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2121 decays
-= se
->avg
.decay_count
;
2125 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2126 se
->avg
.decay_count
= 0;
2131 #ifdef CONFIG_FAIR_GROUP_SCHED
2132 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2135 struct task_group
*tg
= cfs_rq
->tg
;
2138 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2139 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2141 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2142 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2143 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2148 * Aggregate cfs_rq runnable averages into an equivalent task_group
2149 * representation for computing load contributions.
2151 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2152 struct cfs_rq
*cfs_rq
)
2154 struct task_group
*tg
= cfs_rq
->tg
;
2157 /* The fraction of a cpu used by this cfs_rq */
2158 contrib
= div_u64((u64
)sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2159 sa
->runnable_avg_period
+ 1);
2160 contrib
-= cfs_rq
->tg_runnable_contrib
;
2162 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2163 atomic_add(contrib
, &tg
->runnable_avg
);
2164 cfs_rq
->tg_runnable_contrib
+= contrib
;
2168 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2170 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2171 struct task_group
*tg
= cfs_rq
->tg
;
2176 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2177 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2178 atomic_long_read(&tg
->load_avg
) + 1);
2181 * For group entities we need to compute a correction term in the case
2182 * that they are consuming <1 cpu so that we would contribute the same
2183 * load as a task of equal weight.
2185 * Explicitly co-ordinating this measurement would be expensive, but
2186 * fortunately the sum of each cpus contribution forms a usable
2187 * lower-bound on the true value.
2189 * Consider the aggregate of 2 contributions. Either they are disjoint
2190 * (and the sum represents true value) or they are disjoint and we are
2191 * understating by the aggregate of their overlap.
2193 * Extending this to N cpus, for a given overlap, the maximum amount we
2194 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2195 * cpus that overlap for this interval and w_i is the interval width.
2197 * On a small machine; the first term is well-bounded which bounds the
2198 * total error since w_i is a subset of the period. Whereas on a
2199 * larger machine, while this first term can be larger, if w_i is the
2200 * of consequential size guaranteed to see n_i*w_i quickly converge to
2201 * our upper bound of 1-cpu.
2203 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2204 if (runnable_avg
< NICE_0_LOAD
) {
2205 se
->avg
.load_avg_contrib
*= runnable_avg
;
2206 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2210 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2211 int force_update
) {}
2212 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2213 struct cfs_rq
*cfs_rq
) {}
2214 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2217 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2221 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2222 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2223 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2224 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2227 /* Compute the current contribution to load_avg by se, return any delta */
2228 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2230 long old_contrib
= se
->avg
.load_avg_contrib
;
2232 if (entity_is_task(se
)) {
2233 __update_task_entity_contrib(se
);
2235 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2236 __update_group_entity_contrib(se
);
2239 return se
->avg
.load_avg_contrib
- old_contrib
;
2242 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2245 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2246 cfs_rq
->blocked_load_avg
-= load_contrib
;
2248 cfs_rq
->blocked_load_avg
= 0;
2251 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2253 /* Update a sched_entity's runnable average */
2254 static inline void update_entity_load_avg(struct sched_entity
*se
,
2257 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2262 * For a group entity we need to use their owned cfs_rq_clock_task() in
2263 * case they are the parent of a throttled hierarchy.
2265 if (entity_is_task(se
))
2266 now
= cfs_rq_clock_task(cfs_rq
);
2268 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2270 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2273 contrib_delta
= __update_entity_load_avg_contrib(se
);
2279 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2281 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2285 * Decay the load contributed by all blocked children and account this so that
2286 * their contribution may appropriately discounted when they wake up.
2288 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2290 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2293 decays
= now
- cfs_rq
->last_decay
;
2294 if (!decays
&& !force_update
)
2297 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2298 unsigned long removed_load
;
2299 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2300 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2304 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2306 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2307 cfs_rq
->last_decay
= now
;
2310 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2313 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2315 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2316 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2319 /* Add the load generated by se into cfs_rq's child load-average */
2320 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2321 struct sched_entity
*se
,
2325 * We track migrations using entity decay_count <= 0, on a wake-up
2326 * migration we use a negative decay count to track the remote decays
2327 * accumulated while sleeping.
2329 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2330 * are seen by enqueue_entity_load_avg() as a migration with an already
2331 * constructed load_avg_contrib.
2333 if (unlikely(se
->avg
.decay_count
<= 0)) {
2334 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2335 if (se
->avg
.decay_count
) {
2337 * In a wake-up migration we have to approximate the
2338 * time sleeping. This is because we can't synchronize
2339 * clock_task between the two cpus, and it is not
2340 * guaranteed to be read-safe. Instead, we can
2341 * approximate this using our carried decays, which are
2342 * explicitly atomically readable.
2344 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2346 update_entity_load_avg(se
, 0);
2347 /* Indicate that we're now synchronized and on-rq */
2348 se
->avg
.decay_count
= 0;
2353 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2354 * would have made count negative); we must be careful to avoid
2355 * double-accounting blocked time after synchronizing decays.
2357 se
->avg
.last_runnable_update
+= __synchronize_entity_decay(se
)
2361 /* migrated tasks did not contribute to our blocked load */
2363 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2364 update_entity_load_avg(se
, 0);
2367 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2368 /* we force update consideration on load-balancer moves */
2369 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2373 * Remove se's load from this cfs_rq child load-average, if the entity is
2374 * transitioning to a blocked state we track its projected decay using
2377 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2378 struct sched_entity
*se
,
2381 update_entity_load_avg(se
, 1);
2382 /* we force update consideration on load-balancer moves */
2383 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2385 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2387 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2388 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2389 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2393 * Update the rq's load with the elapsed running time before entering
2394 * idle. if the last scheduled task is not a CFS task, idle_enter will
2395 * be the only way to update the runnable statistic.
2397 void idle_enter_fair(struct rq
*this_rq
)
2399 update_rq_runnable_avg(this_rq
, 1);
2403 * Update the rq's load with the elapsed idle time before a task is
2404 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2405 * be the only way to update the runnable statistic.
2407 void idle_exit_fair(struct rq
*this_rq
)
2409 update_rq_runnable_avg(this_rq
, 0);
2413 static inline void update_entity_load_avg(struct sched_entity
*se
,
2414 int update_cfs_rq
) {}
2415 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2416 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2417 struct sched_entity
*se
,
2419 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2420 struct sched_entity
*se
,
2422 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2423 int force_update
) {}
2426 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2428 #ifdef CONFIG_SCHEDSTATS
2429 struct task_struct
*tsk
= NULL
;
2431 if (entity_is_task(se
))
2434 if (se
->statistics
.sleep_start
) {
2435 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2440 if (unlikely(delta
> se
->statistics
.sleep_max
))
2441 se
->statistics
.sleep_max
= delta
;
2443 se
->statistics
.sleep_start
= 0;
2444 se
->statistics
.sum_sleep_runtime
+= delta
;
2447 account_scheduler_latency(tsk
, delta
>> 10, 1);
2448 trace_sched_stat_sleep(tsk
, delta
);
2451 if (se
->statistics
.block_start
) {
2452 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2457 if (unlikely(delta
> se
->statistics
.block_max
))
2458 se
->statistics
.block_max
= delta
;
2460 se
->statistics
.block_start
= 0;
2461 se
->statistics
.sum_sleep_runtime
+= delta
;
2464 if (tsk
->in_iowait
) {
2465 se
->statistics
.iowait_sum
+= delta
;
2466 se
->statistics
.iowait_count
++;
2467 trace_sched_stat_iowait(tsk
, delta
);
2470 trace_sched_stat_blocked(tsk
, delta
);
2473 * Blocking time is in units of nanosecs, so shift by
2474 * 20 to get a milliseconds-range estimation of the
2475 * amount of time that the task spent sleeping:
2477 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2478 profile_hits(SLEEP_PROFILING
,
2479 (void *)get_wchan(tsk
),
2482 account_scheduler_latency(tsk
, delta
>> 10, 0);
2488 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2490 #ifdef CONFIG_SCHED_DEBUG
2491 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2496 if (d
> 3*sysctl_sched_latency
)
2497 schedstat_inc(cfs_rq
, nr_spread_over
);
2502 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2504 u64 vruntime
= cfs_rq
->min_vruntime
;
2507 * The 'current' period is already promised to the current tasks,
2508 * however the extra weight of the new task will slow them down a
2509 * little, place the new task so that it fits in the slot that
2510 * stays open at the end.
2512 if (initial
&& sched_feat(START_DEBIT
))
2513 vruntime
+= sched_vslice(cfs_rq
, se
);
2515 /* sleeps up to a single latency don't count. */
2517 unsigned long thresh
= sysctl_sched_latency
;
2520 * Halve their sleep time's effect, to allow
2521 * for a gentler effect of sleepers:
2523 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2529 /* ensure we never gain time by being placed backwards. */
2530 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
2533 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
2536 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2539 * Update the normalized vruntime before updating min_vruntime
2540 * through calling update_curr().
2542 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
2543 se
->vruntime
+= cfs_rq
->min_vruntime
;
2546 * Update run-time statistics of the 'current'.
2548 update_curr(cfs_rq
);
2549 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
2550 account_entity_enqueue(cfs_rq
, se
);
2551 update_cfs_shares(cfs_rq
);
2553 if (flags
& ENQUEUE_WAKEUP
) {
2554 place_entity(cfs_rq
, se
, 0);
2555 enqueue_sleeper(cfs_rq
, se
);
2558 update_stats_enqueue(cfs_rq
, se
);
2559 check_spread(cfs_rq
, se
);
2560 if (se
!= cfs_rq
->curr
)
2561 __enqueue_entity(cfs_rq
, se
);
2564 if (cfs_rq
->nr_running
== 1) {
2565 list_add_leaf_cfs_rq(cfs_rq
);
2566 check_enqueue_throttle(cfs_rq
);
2570 static void __clear_buddies_last(struct sched_entity
*se
)
2572 for_each_sched_entity(se
) {
2573 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2574 if (cfs_rq
->last
== se
)
2575 cfs_rq
->last
= NULL
;
2581 static void __clear_buddies_next(struct sched_entity
*se
)
2583 for_each_sched_entity(se
) {
2584 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2585 if (cfs_rq
->next
== se
)
2586 cfs_rq
->next
= NULL
;
2592 static void __clear_buddies_skip(struct sched_entity
*se
)
2594 for_each_sched_entity(se
) {
2595 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2596 if (cfs_rq
->skip
== se
)
2597 cfs_rq
->skip
= NULL
;
2603 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2605 if (cfs_rq
->last
== se
)
2606 __clear_buddies_last(se
);
2608 if (cfs_rq
->next
== se
)
2609 __clear_buddies_next(se
);
2611 if (cfs_rq
->skip
== se
)
2612 __clear_buddies_skip(se
);
2615 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2618 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2621 * Update run-time statistics of the 'current'.
2623 update_curr(cfs_rq
);
2624 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
2626 update_stats_dequeue(cfs_rq
, se
);
2627 if (flags
& DEQUEUE_SLEEP
) {
2628 #ifdef CONFIG_SCHEDSTATS
2629 if (entity_is_task(se
)) {
2630 struct task_struct
*tsk
= task_of(se
);
2632 if (tsk
->state
& TASK_INTERRUPTIBLE
)
2633 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
2634 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
2635 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
2640 clear_buddies(cfs_rq
, se
);
2642 if (se
!= cfs_rq
->curr
)
2643 __dequeue_entity(cfs_rq
, se
);
2645 account_entity_dequeue(cfs_rq
, se
);
2648 * Normalize the entity after updating the min_vruntime because the
2649 * update can refer to the ->curr item and we need to reflect this
2650 * movement in our normalized position.
2652 if (!(flags
& DEQUEUE_SLEEP
))
2653 se
->vruntime
-= cfs_rq
->min_vruntime
;
2655 /* return excess runtime on last dequeue */
2656 return_cfs_rq_runtime(cfs_rq
);
2658 update_min_vruntime(cfs_rq
);
2659 update_cfs_shares(cfs_rq
);
2663 * Preempt the current task with a newly woken task if needed:
2666 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2668 unsigned long ideal_runtime
, delta_exec
;
2669 struct sched_entity
*se
;
2672 ideal_runtime
= sched_slice(cfs_rq
, curr
);
2673 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
2674 if (delta_exec
> ideal_runtime
) {
2675 resched_task(rq_of(cfs_rq
)->curr
);
2677 * The current task ran long enough, ensure it doesn't get
2678 * re-elected due to buddy favours.
2680 clear_buddies(cfs_rq
, curr
);
2685 * Ensure that a task that missed wakeup preemption by a
2686 * narrow margin doesn't have to wait for a full slice.
2687 * This also mitigates buddy induced latencies under load.
2689 if (delta_exec
< sysctl_sched_min_granularity
)
2692 se
= __pick_first_entity(cfs_rq
);
2693 delta
= curr
->vruntime
- se
->vruntime
;
2698 if (delta
> ideal_runtime
)
2699 resched_task(rq_of(cfs_rq
)->curr
);
2703 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2705 /* 'current' is not kept within the tree. */
2708 * Any task has to be enqueued before it get to execute on
2709 * a CPU. So account for the time it spent waiting on the
2712 update_stats_wait_end(cfs_rq
, se
);
2713 __dequeue_entity(cfs_rq
, se
);
2716 update_stats_curr_start(cfs_rq
, se
);
2718 #ifdef CONFIG_SCHEDSTATS
2720 * Track our maximum slice length, if the CPU's load is at
2721 * least twice that of our own weight (i.e. dont track it
2722 * when there are only lesser-weight tasks around):
2724 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
2725 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
2726 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
2729 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
2733 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
2736 * Pick the next process, keeping these things in mind, in this order:
2737 * 1) keep things fair between processes/task groups
2738 * 2) pick the "next" process, since someone really wants that to run
2739 * 3) pick the "last" process, for cache locality
2740 * 4) do not run the "skip" process, if something else is available
2742 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
2744 struct sched_entity
*se
= __pick_first_entity(cfs_rq
);
2745 struct sched_entity
*left
= se
;
2748 * Avoid running the skip buddy, if running something else can
2749 * be done without getting too unfair.
2751 if (cfs_rq
->skip
== se
) {
2752 struct sched_entity
*second
= __pick_next_entity(se
);
2753 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
2758 * Prefer last buddy, try to return the CPU to a preempted task.
2760 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
2764 * Someone really wants this to run. If it's not unfair, run it.
2766 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
2769 clear_buddies(cfs_rq
, se
);
2774 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2776 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
2779 * If still on the runqueue then deactivate_task()
2780 * was not called and update_curr() has to be done:
2783 update_curr(cfs_rq
);
2785 /* throttle cfs_rqs exceeding runtime */
2786 check_cfs_rq_runtime(cfs_rq
);
2788 check_spread(cfs_rq
, prev
);
2790 update_stats_wait_start(cfs_rq
, prev
);
2791 /* Put 'current' back into the tree. */
2792 __enqueue_entity(cfs_rq
, prev
);
2793 /* in !on_rq case, update occurred at dequeue */
2794 update_entity_load_avg(prev
, 1);
2796 cfs_rq
->curr
= NULL
;
2800 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
2803 * Update run-time statistics of the 'current'.
2805 update_curr(cfs_rq
);
2808 * Ensure that runnable average is periodically updated.
2810 update_entity_load_avg(curr
, 1);
2811 update_cfs_rq_blocked_load(cfs_rq
, 1);
2812 update_cfs_shares(cfs_rq
);
2814 #ifdef CONFIG_SCHED_HRTICK
2816 * queued ticks are scheduled to match the slice, so don't bother
2817 * validating it and just reschedule.
2820 resched_task(rq_of(cfs_rq
)->curr
);
2824 * don't let the period tick interfere with the hrtick preemption
2826 if (!sched_feat(DOUBLE_TICK
) &&
2827 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
2831 if (cfs_rq
->nr_running
> 1)
2832 check_preempt_tick(cfs_rq
, curr
);
2836 /**************************************************
2837 * CFS bandwidth control machinery
2840 #ifdef CONFIG_CFS_BANDWIDTH
2842 #ifdef HAVE_JUMP_LABEL
2843 static struct static_key __cfs_bandwidth_used
;
2845 static inline bool cfs_bandwidth_used(void)
2847 return static_key_false(&__cfs_bandwidth_used
);
2850 void cfs_bandwidth_usage_inc(void)
2852 static_key_slow_inc(&__cfs_bandwidth_used
);
2855 void cfs_bandwidth_usage_dec(void)
2857 static_key_slow_dec(&__cfs_bandwidth_used
);
2859 #else /* HAVE_JUMP_LABEL */
2860 static bool cfs_bandwidth_used(void)
2865 void cfs_bandwidth_usage_inc(void) {}
2866 void cfs_bandwidth_usage_dec(void) {}
2867 #endif /* HAVE_JUMP_LABEL */
2870 * default period for cfs group bandwidth.
2871 * default: 0.1s, units: nanoseconds
2873 static inline u64
default_cfs_period(void)
2875 return 100000000ULL;
2878 static inline u64
sched_cfs_bandwidth_slice(void)
2880 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
2884 * Replenish runtime according to assigned quota and update expiration time.
2885 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2886 * additional synchronization around rq->lock.
2888 * requires cfs_b->lock
2890 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
2894 if (cfs_b
->quota
== RUNTIME_INF
)
2897 now
= sched_clock_cpu(smp_processor_id());
2898 cfs_b
->runtime
= cfs_b
->quota
;
2899 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
2902 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
2904 return &tg
->cfs_bandwidth
;
2907 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2908 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
2910 if (unlikely(cfs_rq
->throttle_count
))
2911 return cfs_rq
->throttled_clock_task
;
2913 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
2916 /* returns 0 on failure to allocate runtime */
2917 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2919 struct task_group
*tg
= cfs_rq
->tg
;
2920 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
2921 u64 amount
= 0, min_amount
, expires
;
2923 /* note: this is a positive sum as runtime_remaining <= 0 */
2924 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
2926 raw_spin_lock(&cfs_b
->lock
);
2927 if (cfs_b
->quota
== RUNTIME_INF
)
2928 amount
= min_amount
;
2931 * If the bandwidth pool has become inactive, then at least one
2932 * period must have elapsed since the last consumption.
2933 * Refresh the global state and ensure bandwidth timer becomes
2936 if (!cfs_b
->timer_active
) {
2937 __refill_cfs_bandwidth_runtime(cfs_b
);
2938 __start_cfs_bandwidth(cfs_b
);
2941 if (cfs_b
->runtime
> 0) {
2942 amount
= min(cfs_b
->runtime
, min_amount
);
2943 cfs_b
->runtime
-= amount
;
2947 expires
= cfs_b
->runtime_expires
;
2948 raw_spin_unlock(&cfs_b
->lock
);
2950 cfs_rq
->runtime_remaining
+= amount
;
2952 * we may have advanced our local expiration to account for allowed
2953 * spread between our sched_clock and the one on which runtime was
2956 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
2957 cfs_rq
->runtime_expires
= expires
;
2959 return cfs_rq
->runtime_remaining
> 0;
2963 * Note: This depends on the synchronization provided by sched_clock and the
2964 * fact that rq->clock snapshots this value.
2966 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2968 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2970 /* if the deadline is ahead of our clock, nothing to do */
2971 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
2974 if (cfs_rq
->runtime_remaining
< 0)
2978 * If the local deadline has passed we have to consider the
2979 * possibility that our sched_clock is 'fast' and the global deadline
2980 * has not truly expired.
2982 * Fortunately we can check determine whether this the case by checking
2983 * whether the global deadline has advanced.
2986 if ((s64
)(cfs_rq
->runtime_expires
- cfs_b
->runtime_expires
) >= 0) {
2987 /* extend local deadline, drift is bounded above by 2 ticks */
2988 cfs_rq
->runtime_expires
+= TICK_NSEC
;
2990 /* global deadline is ahead, expiration has passed */
2991 cfs_rq
->runtime_remaining
= 0;
2995 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
2997 /* dock delta_exec before expiring quota (as it could span periods) */
2998 cfs_rq
->runtime_remaining
-= delta_exec
;
2999 expire_cfs_rq_runtime(cfs_rq
);
3001 if (likely(cfs_rq
->runtime_remaining
> 0))
3005 * if we're unable to extend our runtime we resched so that the active
3006 * hierarchy can be throttled
3008 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3009 resched_task(rq_of(cfs_rq
)->curr
);
3012 static __always_inline
3013 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3015 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3018 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3021 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3023 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3026 /* check whether cfs_rq, or any parent, is throttled */
3027 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3029 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3033 * Ensure that neither of the group entities corresponding to src_cpu or
3034 * dest_cpu are members of a throttled hierarchy when performing group
3035 * load-balance operations.
3037 static inline int throttled_lb_pair(struct task_group
*tg
,
3038 int src_cpu
, int dest_cpu
)
3040 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3042 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3043 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3045 return throttled_hierarchy(src_cfs_rq
) ||
3046 throttled_hierarchy(dest_cfs_rq
);
3049 /* updated child weight may affect parent so we have to do this bottom up */
3050 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3052 struct rq
*rq
= data
;
3053 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3055 cfs_rq
->throttle_count
--;
3057 if (!cfs_rq
->throttle_count
) {
3058 /* adjust cfs_rq_clock_task() */
3059 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3060 cfs_rq
->throttled_clock_task
;
3067 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3069 struct rq
*rq
= data
;
3070 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3072 /* group is entering throttled state, stop time */
3073 if (!cfs_rq
->throttle_count
)
3074 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3075 cfs_rq
->throttle_count
++;
3080 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3082 struct rq
*rq
= rq_of(cfs_rq
);
3083 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3084 struct sched_entity
*se
;
3085 long task_delta
, dequeue
= 1;
3087 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3089 /* freeze hierarchy runnable averages while throttled */
3091 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3094 task_delta
= cfs_rq
->h_nr_running
;
3095 for_each_sched_entity(se
) {
3096 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3097 /* throttled entity or throttle-on-deactivate */
3102 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3103 qcfs_rq
->h_nr_running
-= task_delta
;
3105 if (qcfs_rq
->load
.weight
)
3110 rq
->nr_running
-= task_delta
;
3112 cfs_rq
->throttled
= 1;
3113 cfs_rq
->throttled_clock
= rq_clock(rq
);
3114 raw_spin_lock(&cfs_b
->lock
);
3115 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3116 if (!cfs_b
->timer_active
)
3117 __start_cfs_bandwidth(cfs_b
);
3118 raw_spin_unlock(&cfs_b
->lock
);
3121 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3123 struct rq
*rq
= rq_of(cfs_rq
);
3124 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3125 struct sched_entity
*se
;
3129 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3131 cfs_rq
->throttled
= 0;
3133 update_rq_clock(rq
);
3135 raw_spin_lock(&cfs_b
->lock
);
3136 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3137 list_del_rcu(&cfs_rq
->throttled_list
);
3138 raw_spin_unlock(&cfs_b
->lock
);
3140 /* update hierarchical throttle state */
3141 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3143 if (!cfs_rq
->load
.weight
)
3146 task_delta
= cfs_rq
->h_nr_running
;
3147 for_each_sched_entity(se
) {
3151 cfs_rq
= cfs_rq_of(se
);
3153 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3154 cfs_rq
->h_nr_running
+= task_delta
;
3156 if (cfs_rq_throttled(cfs_rq
))
3161 rq
->nr_running
+= task_delta
;
3163 /* determine whether we need to wake up potentially idle cpu */
3164 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3165 resched_task(rq
->curr
);
3168 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3169 u64 remaining
, u64 expires
)
3171 struct cfs_rq
*cfs_rq
;
3172 u64 runtime
= remaining
;
3175 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3177 struct rq
*rq
= rq_of(cfs_rq
);
3179 raw_spin_lock(&rq
->lock
);
3180 if (!cfs_rq_throttled(cfs_rq
))
3183 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3184 if (runtime
> remaining
)
3185 runtime
= remaining
;
3186 remaining
-= runtime
;
3188 cfs_rq
->runtime_remaining
+= runtime
;
3189 cfs_rq
->runtime_expires
= expires
;
3191 /* we check whether we're throttled above */
3192 if (cfs_rq
->runtime_remaining
> 0)
3193 unthrottle_cfs_rq(cfs_rq
);
3196 raw_spin_unlock(&rq
->lock
);
3207 * Responsible for refilling a task_group's bandwidth and unthrottling its
3208 * cfs_rqs as appropriate. If there has been no activity within the last
3209 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3210 * used to track this state.
3212 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3214 u64 runtime
, runtime_expires
;
3215 int idle
= 1, throttled
;
3217 raw_spin_lock(&cfs_b
->lock
);
3218 /* no need to continue the timer with no bandwidth constraint */
3219 if (cfs_b
->quota
== RUNTIME_INF
)
3222 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3223 /* idle depends on !throttled (for the case of a large deficit) */
3224 idle
= cfs_b
->idle
&& !throttled
;
3225 cfs_b
->nr_periods
+= overrun
;
3227 /* if we're going inactive then everything else can be deferred */
3232 * if we have relooped after returning idle once, we need to update our
3233 * status as actually running, so that other cpus doing
3234 * __start_cfs_bandwidth will stop trying to cancel us.
3236 cfs_b
->timer_active
= 1;
3238 __refill_cfs_bandwidth_runtime(cfs_b
);
3241 /* mark as potentially idle for the upcoming period */
3246 /* account preceding periods in which throttling occurred */
3247 cfs_b
->nr_throttled
+= overrun
;
3250 * There are throttled entities so we must first use the new bandwidth
3251 * to unthrottle them before making it generally available. This
3252 * ensures that all existing debts will be paid before a new cfs_rq is
3255 runtime
= cfs_b
->runtime
;
3256 runtime_expires
= cfs_b
->runtime_expires
;
3260 * This check is repeated as we are holding onto the new bandwidth
3261 * while we unthrottle. This can potentially race with an unthrottled
3262 * group trying to acquire new bandwidth from the global pool.
3264 while (throttled
&& runtime
> 0) {
3265 raw_spin_unlock(&cfs_b
->lock
);
3266 /* we can't nest cfs_b->lock while distributing bandwidth */
3267 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3269 raw_spin_lock(&cfs_b
->lock
);
3271 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3274 /* return (any) remaining runtime */
3275 cfs_b
->runtime
= runtime
;
3277 * While we are ensured activity in the period following an
3278 * unthrottle, this also covers the case in which the new bandwidth is
3279 * insufficient to cover the existing bandwidth deficit. (Forcing the
3280 * timer to remain active while there are any throttled entities.)
3285 cfs_b
->timer_active
= 0;
3286 raw_spin_unlock(&cfs_b
->lock
);
3291 /* a cfs_rq won't donate quota below this amount */
3292 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3293 /* minimum remaining period time to redistribute slack quota */
3294 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3295 /* how long we wait to gather additional slack before distributing */
3296 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3299 * Are we near the end of the current quota period?
3301 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3302 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3303 * migrate_hrtimers, base is never cleared, so we are fine.
3305 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3307 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3310 /* if the call-back is running a quota refresh is already occurring */
3311 if (hrtimer_callback_running(refresh_timer
))
3314 /* is a quota refresh about to occur? */
3315 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3316 if (remaining
< min_expire
)
3322 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3324 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3326 /* if there's a quota refresh soon don't bother with slack */
3327 if (runtime_refresh_within(cfs_b
, min_left
))
3330 start_bandwidth_timer(&cfs_b
->slack_timer
,
3331 ns_to_ktime(cfs_bandwidth_slack_period
));
3334 /* we know any runtime found here is valid as update_curr() precedes return */
3335 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3337 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3338 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3340 if (slack_runtime
<= 0)
3343 raw_spin_lock(&cfs_b
->lock
);
3344 if (cfs_b
->quota
!= RUNTIME_INF
&&
3345 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3346 cfs_b
->runtime
+= slack_runtime
;
3348 /* we are under rq->lock, defer unthrottling using a timer */
3349 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3350 !list_empty(&cfs_b
->throttled_cfs_rq
))
3351 start_cfs_slack_bandwidth(cfs_b
);
3353 raw_spin_unlock(&cfs_b
->lock
);
3355 /* even if it's not valid for return we don't want to try again */
3356 cfs_rq
->runtime_remaining
-= slack_runtime
;
3359 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3361 if (!cfs_bandwidth_used())
3364 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3367 __return_cfs_rq_runtime(cfs_rq
);
3371 * This is done with a timer (instead of inline with bandwidth return) since
3372 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3374 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3376 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3379 /* confirm we're still not at a refresh boundary */
3380 raw_spin_lock(&cfs_b
->lock
);
3381 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
3382 raw_spin_unlock(&cfs_b
->lock
);
3386 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
) {
3387 runtime
= cfs_b
->runtime
;
3390 expires
= cfs_b
->runtime_expires
;
3391 raw_spin_unlock(&cfs_b
->lock
);
3396 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3398 raw_spin_lock(&cfs_b
->lock
);
3399 if (expires
== cfs_b
->runtime_expires
)
3400 cfs_b
->runtime
= runtime
;
3401 raw_spin_unlock(&cfs_b
->lock
);
3405 * When a group wakes up we want to make sure that its quota is not already
3406 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3407 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3409 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3411 if (!cfs_bandwidth_used())
3414 /* an active group must be handled by the update_curr()->put() path */
3415 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3418 /* ensure the group is not already throttled */
3419 if (cfs_rq_throttled(cfs_rq
))
3422 /* update runtime allocation */
3423 account_cfs_rq_runtime(cfs_rq
, 0);
3424 if (cfs_rq
->runtime_remaining
<= 0)
3425 throttle_cfs_rq(cfs_rq
);
3428 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3429 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3431 if (!cfs_bandwidth_used())
3434 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3438 * it's possible for a throttled entity to be forced into a running
3439 * state (e.g. set_curr_task), in this case we're finished.
3441 if (cfs_rq_throttled(cfs_rq
))
3444 throttle_cfs_rq(cfs_rq
);
3447 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3449 struct cfs_bandwidth
*cfs_b
=
3450 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3451 do_sched_cfs_slack_timer(cfs_b
);
3453 return HRTIMER_NORESTART
;
3456 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3458 struct cfs_bandwidth
*cfs_b
=
3459 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3465 now
= hrtimer_cb_get_time(timer
);
3466 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3471 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3474 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3477 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3479 raw_spin_lock_init(&cfs_b
->lock
);
3481 cfs_b
->quota
= RUNTIME_INF
;
3482 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3484 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3485 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3486 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3487 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3488 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3491 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3493 cfs_rq
->runtime_enabled
= 0;
3494 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3497 /* requires cfs_b->lock, may release to reprogram timer */
3498 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3501 * The timer may be active because we're trying to set a new bandwidth
3502 * period or because we're racing with the tear-down path
3503 * (timer_active==0 becomes visible before the hrtimer call-back
3504 * terminates). In either case we ensure that it's re-programmed
3506 while (unlikely(hrtimer_active(&cfs_b
->period_timer
)) &&
3507 hrtimer_try_to_cancel(&cfs_b
->period_timer
) < 0) {
3508 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3509 raw_spin_unlock(&cfs_b
->lock
);
3511 raw_spin_lock(&cfs_b
->lock
);
3512 /* if someone else restarted the timer then we're done */
3513 if (cfs_b
->timer_active
)
3517 cfs_b
->timer_active
= 1;
3518 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
3521 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3523 hrtimer_cancel(&cfs_b
->period_timer
);
3524 hrtimer_cancel(&cfs_b
->slack_timer
);
3527 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
3529 struct cfs_rq
*cfs_rq
;
3531 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3532 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3534 if (!cfs_rq
->runtime_enabled
)
3538 * clock_task is not advancing so we just need to make sure
3539 * there's some valid quota amount
3541 cfs_rq
->runtime_remaining
= cfs_b
->quota
;
3542 if (cfs_rq_throttled(cfs_rq
))
3543 unthrottle_cfs_rq(cfs_rq
);
3547 #else /* CONFIG_CFS_BANDWIDTH */
3548 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3550 return rq_clock_task(rq_of(cfs_rq
));
3553 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
3554 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3555 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
3556 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3558 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3563 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3568 static inline int throttled_lb_pair(struct task_group
*tg
,
3569 int src_cpu
, int dest_cpu
)
3574 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3576 #ifdef CONFIG_FAIR_GROUP_SCHED
3577 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3580 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3584 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3585 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
3587 #endif /* CONFIG_CFS_BANDWIDTH */
3589 /**************************************************
3590 * CFS operations on tasks:
3593 #ifdef CONFIG_SCHED_HRTICK
3594 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3596 struct sched_entity
*se
= &p
->se
;
3597 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3599 WARN_ON(task_rq(p
) != rq
);
3601 if (cfs_rq
->nr_running
> 1) {
3602 u64 slice
= sched_slice(cfs_rq
, se
);
3603 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
3604 s64 delta
= slice
- ran
;
3613 * Don't schedule slices shorter than 10000ns, that just
3614 * doesn't make sense. Rely on vruntime for fairness.
3617 delta
= max_t(s64
, 10000LL, delta
);
3619 hrtick_start(rq
, delta
);
3624 * called from enqueue/dequeue and updates the hrtick when the
3625 * current task is from our class and nr_running is low enough
3628 static void hrtick_update(struct rq
*rq
)
3630 struct task_struct
*curr
= rq
->curr
;
3632 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
3635 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
3636 hrtick_start_fair(rq
, curr
);
3638 #else /* !CONFIG_SCHED_HRTICK */
3640 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3644 static inline void hrtick_update(struct rq
*rq
)
3650 * The enqueue_task method is called before nr_running is
3651 * increased. Here we update the fair scheduling stats and
3652 * then put the task into the rbtree:
3655 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3657 struct cfs_rq
*cfs_rq
;
3658 struct sched_entity
*se
= &p
->se
;
3660 for_each_sched_entity(se
) {
3663 cfs_rq
= cfs_rq_of(se
);
3664 enqueue_entity(cfs_rq
, se
, flags
);
3667 * end evaluation on encountering a throttled cfs_rq
3669 * note: in the case of encountering a throttled cfs_rq we will
3670 * post the final h_nr_running increment below.
3672 if (cfs_rq_throttled(cfs_rq
))
3674 cfs_rq
->h_nr_running
++;
3676 flags
= ENQUEUE_WAKEUP
;
3679 for_each_sched_entity(se
) {
3680 cfs_rq
= cfs_rq_of(se
);
3681 cfs_rq
->h_nr_running
++;
3683 if (cfs_rq_throttled(cfs_rq
))
3686 update_cfs_shares(cfs_rq
);
3687 update_entity_load_avg(se
, 1);
3691 update_rq_runnable_avg(rq
, rq
->nr_running
);
3697 static void set_next_buddy(struct sched_entity
*se
);
3700 * The dequeue_task method is called before nr_running is
3701 * decreased. We remove the task from the rbtree and
3702 * update the fair scheduling stats:
3704 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3706 struct cfs_rq
*cfs_rq
;
3707 struct sched_entity
*se
= &p
->se
;
3708 int task_sleep
= flags
& DEQUEUE_SLEEP
;
3710 for_each_sched_entity(se
) {
3711 cfs_rq
= cfs_rq_of(se
);
3712 dequeue_entity(cfs_rq
, se
, flags
);
3715 * end evaluation on encountering a throttled cfs_rq
3717 * note: in the case of encountering a throttled cfs_rq we will
3718 * post the final h_nr_running decrement below.
3720 if (cfs_rq_throttled(cfs_rq
))
3722 cfs_rq
->h_nr_running
--;
3724 /* Don't dequeue parent if it has other entities besides us */
3725 if (cfs_rq
->load
.weight
) {
3727 * Bias pick_next to pick a task from this cfs_rq, as
3728 * p is sleeping when it is within its sched_slice.
3730 if (task_sleep
&& parent_entity(se
))
3731 set_next_buddy(parent_entity(se
));
3733 /* avoid re-evaluating load for this entity */
3734 se
= parent_entity(se
);
3737 flags
|= DEQUEUE_SLEEP
;
3740 for_each_sched_entity(se
) {
3741 cfs_rq
= cfs_rq_of(se
);
3742 cfs_rq
->h_nr_running
--;
3744 if (cfs_rq_throttled(cfs_rq
))
3747 update_cfs_shares(cfs_rq
);
3748 update_entity_load_avg(se
, 1);
3753 update_rq_runnable_avg(rq
, 1);
3759 /* Used instead of source_load when we know the type == 0 */
3760 static unsigned long weighted_cpuload(const int cpu
)
3762 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
3766 * Return a low guess at the load of a migration-source cpu weighted
3767 * according to the scheduling class and "nice" value.
3769 * We want to under-estimate the load of migration sources, to
3770 * balance conservatively.
3772 static unsigned long source_load(int cpu
, int type
)
3774 struct rq
*rq
= cpu_rq(cpu
);
3775 unsigned long total
= weighted_cpuload(cpu
);
3777 if (type
== 0 || !sched_feat(LB_BIAS
))
3780 return min(rq
->cpu_load
[type
-1], total
);
3784 * Return a high guess at the load of a migration-target cpu weighted
3785 * according to the scheduling class and "nice" value.
3787 static unsigned long target_load(int cpu
, int type
)
3789 struct rq
*rq
= cpu_rq(cpu
);
3790 unsigned long total
= weighted_cpuload(cpu
);
3792 if (type
== 0 || !sched_feat(LB_BIAS
))
3795 return max(rq
->cpu_load
[type
-1], total
);
3798 static unsigned long power_of(int cpu
)
3800 return cpu_rq(cpu
)->cpu_power
;
3803 static unsigned long cpu_avg_load_per_task(int cpu
)
3805 struct rq
*rq
= cpu_rq(cpu
);
3806 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
3807 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
3810 return load_avg
/ nr_running
;
3815 static void record_wakee(struct task_struct
*p
)
3818 * Rough decay (wiping) for cost saving, don't worry
3819 * about the boundary, really active task won't care
3822 if (jiffies
> current
->wakee_flip_decay_ts
+ HZ
) {
3823 current
->wakee_flips
= 0;
3824 current
->wakee_flip_decay_ts
= jiffies
;
3827 if (current
->last_wakee
!= p
) {
3828 current
->last_wakee
= p
;
3829 current
->wakee_flips
++;
3833 static void task_waking_fair(struct task_struct
*p
)
3835 struct sched_entity
*se
= &p
->se
;
3836 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3839 #ifndef CONFIG_64BIT
3840 u64 min_vruntime_copy
;
3843 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
3845 min_vruntime
= cfs_rq
->min_vruntime
;
3846 } while (min_vruntime
!= min_vruntime_copy
);
3848 min_vruntime
= cfs_rq
->min_vruntime
;
3851 se
->vruntime
-= min_vruntime
;
3855 #ifdef CONFIG_FAIR_GROUP_SCHED
3857 * effective_load() calculates the load change as seen from the root_task_group
3859 * Adding load to a group doesn't make a group heavier, but can cause movement
3860 * of group shares between cpus. Assuming the shares were perfectly aligned one
3861 * can calculate the shift in shares.
3863 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3864 * on this @cpu and results in a total addition (subtraction) of @wg to the
3865 * total group weight.
3867 * Given a runqueue weight distribution (rw_i) we can compute a shares
3868 * distribution (s_i) using:
3870 * s_i = rw_i / \Sum rw_j (1)
3872 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3873 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3874 * shares distribution (s_i):
3876 * rw_i = { 2, 4, 1, 0 }
3877 * s_i = { 2/7, 4/7, 1/7, 0 }
3879 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3880 * task used to run on and the CPU the waker is running on), we need to
3881 * compute the effect of waking a task on either CPU and, in case of a sync
3882 * wakeup, compute the effect of the current task going to sleep.
3884 * So for a change of @wl to the local @cpu with an overall group weight change
3885 * of @wl we can compute the new shares distribution (s'_i) using:
3887 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3889 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3890 * differences in waking a task to CPU 0. The additional task changes the
3891 * weight and shares distributions like:
3893 * rw'_i = { 3, 4, 1, 0 }
3894 * s'_i = { 3/8, 4/8, 1/8, 0 }
3896 * We can then compute the difference in effective weight by using:
3898 * dw_i = S * (s'_i - s_i) (3)
3900 * Where 'S' is the group weight as seen by its parent.
3902 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3903 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3904 * 4/7) times the weight of the group.
3906 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
3908 struct sched_entity
*se
= tg
->se
[cpu
];
3910 if (!tg
->parent
|| !wl
) /* the trivial, non-cgroup case */
3913 for_each_sched_entity(se
) {
3919 * W = @wg + \Sum rw_j
3921 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
3926 w
= se
->my_q
->load
.weight
+ wl
;
3929 * wl = S * s'_i; see (2)
3932 wl
= (w
* tg
->shares
) / W
;
3937 * Per the above, wl is the new se->load.weight value; since
3938 * those are clipped to [MIN_SHARES, ...) do so now. See
3939 * calc_cfs_shares().
3941 if (wl
< MIN_SHARES
)
3945 * wl = dw_i = S * (s'_i - s_i); see (3)
3947 wl
-= se
->load
.weight
;
3950 * Recursively apply this logic to all parent groups to compute
3951 * the final effective load change on the root group. Since
3952 * only the @tg group gets extra weight, all parent groups can
3953 * only redistribute existing shares. @wl is the shift in shares
3954 * resulting from this level per the above.
3963 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
3970 static int wake_wide(struct task_struct
*p
)
3972 int factor
= this_cpu_read(sd_llc_size
);
3975 * Yeah, it's the switching-frequency, could means many wakee or
3976 * rapidly switch, use factor here will just help to automatically
3977 * adjust the loose-degree, so bigger node will lead to more pull.
3979 if (p
->wakee_flips
> factor
) {
3981 * wakee is somewhat hot, it needs certain amount of cpu
3982 * resource, so if waker is far more hot, prefer to leave
3985 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
3992 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
3994 s64 this_load
, load
;
3995 int idx
, this_cpu
, prev_cpu
;
3996 unsigned long tl_per_task
;
3997 struct task_group
*tg
;
3998 unsigned long weight
;
4002 * If we wake multiple tasks be careful to not bounce
4003 * ourselves around too much.
4009 this_cpu
= smp_processor_id();
4010 prev_cpu
= task_cpu(p
);
4011 load
= source_load(prev_cpu
, idx
);
4012 this_load
= target_load(this_cpu
, idx
);
4015 * If sync wakeup then subtract the (maximum possible)
4016 * effect of the currently running task from the load
4017 * of the current CPU:
4020 tg
= task_group(current
);
4021 weight
= current
->se
.load
.weight
;
4023 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4024 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4028 weight
= p
->se
.load
.weight
;
4031 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4032 * due to the sync cause above having dropped this_load to 0, we'll
4033 * always have an imbalance, but there's really nothing you can do
4034 * about that, so that's good too.
4036 * Otherwise check if either cpus are near enough in load to allow this
4037 * task to be woken on this_cpu.
4039 if (this_load
> 0) {
4040 s64 this_eff_load
, prev_eff_load
;
4042 this_eff_load
= 100;
4043 this_eff_load
*= power_of(prev_cpu
);
4044 this_eff_load
*= this_load
+
4045 effective_load(tg
, this_cpu
, weight
, weight
);
4047 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4048 prev_eff_load
*= power_of(this_cpu
);
4049 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4051 balanced
= this_eff_load
<= prev_eff_load
;
4056 * If the currently running task will sleep within
4057 * a reasonable amount of time then attract this newly
4060 if (sync
&& balanced
)
4063 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4064 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
4067 (this_load
<= load
&&
4068 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
4070 * This domain has SD_WAKE_AFFINE and
4071 * p is cache cold in this domain, and
4072 * there is no bad imbalance.
4074 schedstat_inc(sd
, ttwu_move_affine
);
4075 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4083 * find_idlest_group finds and returns the least busy CPU group within the
4086 static struct sched_group
*
4087 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4088 int this_cpu
, int sd_flag
)
4090 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4091 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4092 int load_idx
= sd
->forkexec_idx
;
4093 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4095 if (sd_flag
& SD_BALANCE_WAKE
)
4096 load_idx
= sd
->wake_idx
;
4099 unsigned long load
, avg_load
;
4103 /* Skip over this group if it has no CPUs allowed */
4104 if (!cpumask_intersects(sched_group_cpus(group
),
4105 tsk_cpus_allowed(p
)))
4108 local_group
= cpumask_test_cpu(this_cpu
,
4109 sched_group_cpus(group
));
4111 /* Tally up the load of all CPUs in the group */
4114 for_each_cpu(i
, sched_group_cpus(group
)) {
4115 /* Bias balancing toward cpus of our domain */
4117 load
= source_load(i
, load_idx
);
4119 load
= target_load(i
, load_idx
);
4124 /* Adjust by relative CPU power of the group */
4125 avg_load
= (avg_load
* SCHED_POWER_SCALE
) / group
->sgp
->power
;
4128 this_load
= avg_load
;
4129 } else if (avg_load
< min_load
) {
4130 min_load
= avg_load
;
4133 } while (group
= group
->next
, group
!= sd
->groups
);
4135 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4141 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4144 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4146 unsigned long load
, min_load
= ULONG_MAX
;
4150 /* Traverse only the allowed CPUs */
4151 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4152 load
= weighted_cpuload(i
);
4154 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4164 * Try and locate an idle CPU in the sched_domain.
4166 static int select_idle_sibling(struct task_struct
*p
, int target
)
4168 struct sched_domain
*sd
;
4169 struct sched_group
*sg
;
4170 int i
= task_cpu(p
);
4172 if (idle_cpu(target
))
4176 * If the prevous cpu is cache affine and idle, don't be stupid.
4178 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4182 * Otherwise, iterate the domains and find an elegible idle cpu.
4184 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4185 for_each_lower_domain(sd
) {
4188 if (!cpumask_intersects(sched_group_cpus(sg
),
4189 tsk_cpus_allowed(p
)))
4192 for_each_cpu(i
, sched_group_cpus(sg
)) {
4193 if (i
== target
|| !idle_cpu(i
))
4197 target
= cpumask_first_and(sched_group_cpus(sg
),
4198 tsk_cpus_allowed(p
));
4202 } while (sg
!= sd
->groups
);
4209 * sched_balance_self: balance the current task (running on cpu) in domains
4210 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4213 * Balance, ie. select the least loaded group.
4215 * Returns the target CPU number, or the same CPU if no balancing is needed.
4217 * preempt must be disabled.
4220 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4222 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4223 int cpu
= smp_processor_id();
4225 int want_affine
= 0;
4226 int sync
= wake_flags
& WF_SYNC
;
4228 if (p
->nr_cpus_allowed
== 1)
4231 if (sd_flag
& SD_BALANCE_WAKE
) {
4232 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
4238 for_each_domain(cpu
, tmp
) {
4239 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4243 * If both cpu and prev_cpu are part of this domain,
4244 * cpu is a valid SD_WAKE_AFFINE target.
4246 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4247 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4252 if (tmp
->flags
& sd_flag
)
4257 if (cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4260 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4265 struct sched_group
*group
;
4268 if (!(sd
->flags
& sd_flag
)) {
4273 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
4279 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4280 if (new_cpu
== -1 || new_cpu
== cpu
) {
4281 /* Now try balancing at a lower domain level of cpu */
4286 /* Now try balancing at a lower domain level of new_cpu */
4288 weight
= sd
->span_weight
;
4290 for_each_domain(cpu
, tmp
) {
4291 if (weight
<= tmp
->span_weight
)
4293 if (tmp
->flags
& sd_flag
)
4296 /* while loop will break here if sd == NULL */
4305 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4306 * cfs_rq_of(p) references at time of call are still valid and identify the
4307 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4308 * other assumptions, including the state of rq->lock, should be made.
4311 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4313 struct sched_entity
*se
= &p
->se
;
4314 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4317 * Load tracking: accumulate removed load so that it can be processed
4318 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4319 * to blocked load iff they have a positive decay-count. It can never
4320 * be negative here since on-rq tasks have decay-count == 0.
4322 if (se
->avg
.decay_count
) {
4323 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4324 atomic_long_add(se
->avg
.load_avg_contrib
,
4325 &cfs_rq
->removed_load
);
4328 #endif /* CONFIG_SMP */
4330 static unsigned long
4331 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4333 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4336 * Since its curr running now, convert the gran from real-time
4337 * to virtual-time in his units.
4339 * By using 'se' instead of 'curr' we penalize light tasks, so
4340 * they get preempted easier. That is, if 'se' < 'curr' then
4341 * the resulting gran will be larger, therefore penalizing the
4342 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4343 * be smaller, again penalizing the lighter task.
4345 * This is especially important for buddies when the leftmost
4346 * task is higher priority than the buddy.
4348 return calc_delta_fair(gran
, se
);
4352 * Should 'se' preempt 'curr'.
4366 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4368 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4373 gran
= wakeup_gran(curr
, se
);
4380 static void set_last_buddy(struct sched_entity
*se
)
4382 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4385 for_each_sched_entity(se
)
4386 cfs_rq_of(se
)->last
= se
;
4389 static void set_next_buddy(struct sched_entity
*se
)
4391 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4394 for_each_sched_entity(se
)
4395 cfs_rq_of(se
)->next
= se
;
4398 static void set_skip_buddy(struct sched_entity
*se
)
4400 for_each_sched_entity(se
)
4401 cfs_rq_of(se
)->skip
= se
;
4405 * Preempt the current task with a newly woken task if needed:
4407 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4409 struct task_struct
*curr
= rq
->curr
;
4410 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4411 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4412 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4413 int next_buddy_marked
= 0;
4415 if (unlikely(se
== pse
))
4419 * This is possible from callers such as move_task(), in which we
4420 * unconditionally check_prempt_curr() after an enqueue (which may have
4421 * lead to a throttle). This both saves work and prevents false
4422 * next-buddy nomination below.
4424 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4427 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4428 set_next_buddy(pse
);
4429 next_buddy_marked
= 1;
4433 * We can come here with TIF_NEED_RESCHED already set from new task
4436 * Note: this also catches the edge-case of curr being in a throttled
4437 * group (e.g. via set_curr_task), since update_curr() (in the
4438 * enqueue of curr) will have resulted in resched being set. This
4439 * prevents us from potentially nominating it as a false LAST_BUDDY
4442 if (test_tsk_need_resched(curr
))
4445 /* Idle tasks are by definition preempted by non-idle tasks. */
4446 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4447 likely(p
->policy
!= SCHED_IDLE
))
4451 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4452 * is driven by the tick):
4454 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4457 find_matching_se(&se
, &pse
);
4458 update_curr(cfs_rq_of(se
));
4460 if (wakeup_preempt_entity(se
, pse
) == 1) {
4462 * Bias pick_next to pick the sched entity that is
4463 * triggering this preemption.
4465 if (!next_buddy_marked
)
4466 set_next_buddy(pse
);
4475 * Only set the backward buddy when the current task is still
4476 * on the rq. This can happen when a wakeup gets interleaved
4477 * with schedule on the ->pre_schedule() or idle_balance()
4478 * point, either of which can * drop the rq lock.
4480 * Also, during early boot the idle thread is in the fair class,
4481 * for obvious reasons its a bad idea to schedule back to it.
4483 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4486 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
4490 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
4492 struct task_struct
*p
;
4493 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
4494 struct sched_entity
*se
;
4496 if (!cfs_rq
->nr_running
)
4500 se
= pick_next_entity(cfs_rq
);
4501 set_next_entity(cfs_rq
, se
);
4502 cfs_rq
= group_cfs_rq(se
);
4506 if (hrtick_enabled(rq
))
4507 hrtick_start_fair(rq
, p
);
4513 * Account for a descheduled task:
4515 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4517 struct sched_entity
*se
= &prev
->se
;
4518 struct cfs_rq
*cfs_rq
;
4520 for_each_sched_entity(se
) {
4521 cfs_rq
= cfs_rq_of(se
);
4522 put_prev_entity(cfs_rq
, se
);
4527 * sched_yield() is very simple
4529 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4531 static void yield_task_fair(struct rq
*rq
)
4533 struct task_struct
*curr
= rq
->curr
;
4534 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4535 struct sched_entity
*se
= &curr
->se
;
4538 * Are we the only task in the tree?
4540 if (unlikely(rq
->nr_running
== 1))
4543 clear_buddies(cfs_rq
, se
);
4545 if (curr
->policy
!= SCHED_BATCH
) {
4546 update_rq_clock(rq
);
4548 * Update run-time statistics of the 'current'.
4550 update_curr(cfs_rq
);
4552 * Tell update_rq_clock() that we've just updated,
4553 * so we don't do microscopic update in schedule()
4554 * and double the fastpath cost.
4556 rq
->skip_clock_update
= 1;
4562 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
4564 struct sched_entity
*se
= &p
->se
;
4566 /* throttled hierarchies are not runnable */
4567 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
4570 /* Tell the scheduler that we'd really like pse to run next. */
4573 yield_task_fair(rq
);
4579 /**************************************************
4580 * Fair scheduling class load-balancing methods.
4584 * The purpose of load-balancing is to achieve the same basic fairness the
4585 * per-cpu scheduler provides, namely provide a proportional amount of compute
4586 * time to each task. This is expressed in the following equation:
4588 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4590 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4591 * W_i,0 is defined as:
4593 * W_i,0 = \Sum_j w_i,j (2)
4595 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4596 * is derived from the nice value as per prio_to_weight[].
4598 * The weight average is an exponential decay average of the instantaneous
4601 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4603 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4604 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4605 * can also include other factors [XXX].
4607 * To achieve this balance we define a measure of imbalance which follows
4608 * directly from (1):
4610 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4612 * We them move tasks around to minimize the imbalance. In the continuous
4613 * function space it is obvious this converges, in the discrete case we get
4614 * a few fun cases generally called infeasible weight scenarios.
4617 * - infeasible weights;
4618 * - local vs global optima in the discrete case. ]
4623 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4624 * for all i,j solution, we create a tree of cpus that follows the hardware
4625 * topology where each level pairs two lower groups (or better). This results
4626 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4627 * tree to only the first of the previous level and we decrease the frequency
4628 * of load-balance at each level inv. proportional to the number of cpus in
4634 * \Sum { --- * --- * 2^i } = O(n) (5)
4636 * `- size of each group
4637 * | | `- number of cpus doing load-balance
4639 * `- sum over all levels
4641 * Coupled with a limit on how many tasks we can migrate every balance pass,
4642 * this makes (5) the runtime complexity of the balancer.
4644 * An important property here is that each CPU is still (indirectly) connected
4645 * to every other cpu in at most O(log n) steps:
4647 * The adjacency matrix of the resulting graph is given by:
4650 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4653 * And you'll find that:
4655 * A^(log_2 n)_i,j != 0 for all i,j (7)
4657 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4658 * The task movement gives a factor of O(m), giving a convergence complexity
4661 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4666 * In order to avoid CPUs going idle while there's still work to do, new idle
4667 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4668 * tree itself instead of relying on other CPUs to bring it work.
4670 * This adds some complexity to both (5) and (8) but it reduces the total idle
4678 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4681 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4686 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4688 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4690 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4693 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4694 * rewrite all of this once again.]
4697 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
4699 enum fbq_type
{ regular
, remote
, all
};
4701 #define LBF_ALL_PINNED 0x01
4702 #define LBF_NEED_BREAK 0x02
4703 #define LBF_DST_PINNED 0x04
4704 #define LBF_SOME_PINNED 0x08
4707 struct sched_domain
*sd
;
4715 struct cpumask
*dst_grpmask
;
4717 enum cpu_idle_type idle
;
4719 /* The set of CPUs under consideration for load-balancing */
4720 struct cpumask
*cpus
;
4725 unsigned int loop_break
;
4726 unsigned int loop_max
;
4728 enum fbq_type fbq_type
;
4732 * move_task - move a task from one runqueue to another runqueue.
4733 * Both runqueues must be locked.
4735 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
4737 deactivate_task(env
->src_rq
, p
, 0);
4738 set_task_cpu(p
, env
->dst_cpu
);
4739 activate_task(env
->dst_rq
, p
, 0);
4740 check_preempt_curr(env
->dst_rq
, p
, 0);
4744 * Is this task likely cache-hot:
4747 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
4751 if (p
->sched_class
!= &fair_sched_class
)
4754 if (unlikely(p
->policy
== SCHED_IDLE
))
4758 * Buddy candidates are cache hot:
4760 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
4761 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
4762 &p
->se
== cfs_rq_of(&p
->se
)->last
))
4765 if (sysctl_sched_migration_cost
== -1)
4767 if (sysctl_sched_migration_cost
== 0)
4770 delta
= now
- p
->se
.exec_start
;
4772 return delta
< (s64
)sysctl_sched_migration_cost
;
4775 #ifdef CONFIG_NUMA_BALANCING
4776 /* Returns true if the destination node has incurred more faults */
4777 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
4779 int src_nid
, dst_nid
;
4781 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults
||
4782 !(env
->sd
->flags
& SD_NUMA
)) {
4786 src_nid
= cpu_to_node(env
->src_cpu
);
4787 dst_nid
= cpu_to_node(env
->dst_cpu
);
4789 if (src_nid
== dst_nid
)
4792 /* Always encourage migration to the preferred node. */
4793 if (dst_nid
== p
->numa_preferred_nid
)
4796 /* If both task and group weight improve, this move is a winner. */
4797 if (task_weight(p
, dst_nid
) > task_weight(p
, src_nid
) &&
4798 group_weight(p
, dst_nid
) > group_weight(p
, src_nid
))
4805 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
4807 int src_nid
, dst_nid
;
4809 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
4812 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
4815 src_nid
= cpu_to_node(env
->src_cpu
);
4816 dst_nid
= cpu_to_node(env
->dst_cpu
);
4818 if (src_nid
== dst_nid
)
4821 /* Migrating away from the preferred node is always bad. */
4822 if (src_nid
== p
->numa_preferred_nid
)
4825 /* If either task or group weight get worse, don't do it. */
4826 if (task_weight(p
, dst_nid
) < task_weight(p
, src_nid
) ||
4827 group_weight(p
, dst_nid
) < group_weight(p
, src_nid
))
4834 static inline bool migrate_improves_locality(struct task_struct
*p
,
4840 static inline bool migrate_degrades_locality(struct task_struct
*p
,
4848 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4851 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
4853 int tsk_cache_hot
= 0;
4855 * We do not migrate tasks that are:
4856 * 1) throttled_lb_pair, or
4857 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4858 * 3) running (obviously), or
4859 * 4) are cache-hot on their current CPU.
4861 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
4864 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
4867 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
4869 env
->flags
|= LBF_SOME_PINNED
;
4872 * Remember if this task can be migrated to any other cpu in
4873 * our sched_group. We may want to revisit it if we couldn't
4874 * meet load balance goals by pulling other tasks on src_cpu.
4876 * Also avoid computing new_dst_cpu if we have already computed
4877 * one in current iteration.
4879 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
4882 /* Prevent to re-select dst_cpu via env's cpus */
4883 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
4884 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
4885 env
->flags
|= LBF_DST_PINNED
;
4886 env
->new_dst_cpu
= cpu
;
4894 /* Record that we found atleast one task that could run on dst_cpu */
4895 env
->flags
&= ~LBF_ALL_PINNED
;
4897 if (task_running(env
->src_rq
, p
)) {
4898 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
4903 * Aggressive migration if:
4904 * 1) destination numa is preferred
4905 * 2) task is cache cold, or
4906 * 3) too many balance attempts have failed.
4908 tsk_cache_hot
= task_hot(p
, rq_clock_task(env
->src_rq
), env
->sd
);
4910 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
4912 if (migrate_improves_locality(p
, env
)) {
4913 #ifdef CONFIG_SCHEDSTATS
4914 if (tsk_cache_hot
) {
4915 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
4916 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
4922 if (!tsk_cache_hot
||
4923 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
4925 if (tsk_cache_hot
) {
4926 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
4927 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
4933 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
4938 * move_one_task tries to move exactly one task from busiest to this_rq, as
4939 * part of active balancing operations within "domain".
4940 * Returns 1 if successful and 0 otherwise.
4942 * Called with both runqueues locked.
4944 static int move_one_task(struct lb_env
*env
)
4946 struct task_struct
*p
, *n
;
4948 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
4949 if (!can_migrate_task(p
, env
))
4954 * Right now, this is only the second place move_task()
4955 * is called, so we can safely collect move_task()
4956 * stats here rather than inside move_task().
4958 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
4964 static const unsigned int sched_nr_migrate_break
= 32;
4967 * move_tasks tries to move up to imbalance weighted load from busiest to
4968 * this_rq, as part of a balancing operation within domain "sd".
4969 * Returns 1 if successful and 0 otherwise.
4971 * Called with both runqueues locked.
4973 static int move_tasks(struct lb_env
*env
)
4975 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
4976 struct task_struct
*p
;
4980 if (env
->imbalance
<= 0)
4983 while (!list_empty(tasks
)) {
4984 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
4987 /* We've more or less seen every task there is, call it quits */
4988 if (env
->loop
> env
->loop_max
)
4991 /* take a breather every nr_migrate tasks */
4992 if (env
->loop
> env
->loop_break
) {
4993 env
->loop_break
+= sched_nr_migrate_break
;
4994 env
->flags
|= LBF_NEED_BREAK
;
4998 if (!can_migrate_task(p
, env
))
5001 load
= task_h_load(p
);
5003 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5006 if ((load
/ 2) > env
->imbalance
)
5011 env
->imbalance
-= load
;
5013 #ifdef CONFIG_PREEMPT
5015 * NEWIDLE balancing is a source of latency, so preemptible
5016 * kernels will stop after the first task is pulled to minimize
5017 * the critical section.
5019 if (env
->idle
== CPU_NEWLY_IDLE
)
5024 * We only want to steal up to the prescribed amount of
5027 if (env
->imbalance
<= 0)
5032 list_move_tail(&p
->se
.group_node
, tasks
);
5036 * Right now, this is one of only two places move_task() is called,
5037 * so we can safely collect move_task() stats here rather than
5038 * inside move_task().
5040 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
5045 #ifdef CONFIG_FAIR_GROUP_SCHED
5047 * update tg->load_weight by folding this cpu's load_avg
5049 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5051 struct sched_entity
*se
= tg
->se
[cpu
];
5052 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5054 /* throttled entities do not contribute to load */
5055 if (throttled_hierarchy(cfs_rq
))
5058 update_cfs_rq_blocked_load(cfs_rq
, 1);
5061 update_entity_load_avg(se
, 1);
5063 * We pivot on our runnable average having decayed to zero for
5064 * list removal. This generally implies that all our children
5065 * have also been removed (modulo rounding error or bandwidth
5066 * control); however, such cases are rare and we can fix these
5069 * TODO: fix up out-of-order children on enqueue.
5071 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5072 list_del_leaf_cfs_rq(cfs_rq
);
5074 struct rq
*rq
= rq_of(cfs_rq
);
5075 update_rq_runnable_avg(rq
, rq
->nr_running
);
5079 static void update_blocked_averages(int cpu
)
5081 struct rq
*rq
= cpu_rq(cpu
);
5082 struct cfs_rq
*cfs_rq
;
5083 unsigned long flags
;
5085 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5086 update_rq_clock(rq
);
5088 * Iterates the task_group tree in a bottom up fashion, see
5089 * list_add_leaf_cfs_rq() for details.
5091 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5093 * Note: We may want to consider periodically releasing
5094 * rq->lock about these updates so that creating many task
5095 * groups does not result in continually extending hold time.
5097 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5100 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5104 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5105 * This needs to be done in a top-down fashion because the load of a child
5106 * group is a fraction of its parents load.
5108 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5110 struct rq
*rq
= rq_of(cfs_rq
);
5111 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5112 unsigned long now
= jiffies
;
5115 if (cfs_rq
->last_h_load_update
== now
)
5118 cfs_rq
->h_load_next
= NULL
;
5119 for_each_sched_entity(se
) {
5120 cfs_rq
= cfs_rq_of(se
);
5121 cfs_rq
->h_load_next
= se
;
5122 if (cfs_rq
->last_h_load_update
== now
)
5127 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5128 cfs_rq
->last_h_load_update
= now
;
5131 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5132 load
= cfs_rq
->h_load
;
5133 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5134 cfs_rq
->runnable_load_avg
+ 1);
5135 cfs_rq
= group_cfs_rq(se
);
5136 cfs_rq
->h_load
= load
;
5137 cfs_rq
->last_h_load_update
= now
;
5141 static unsigned long task_h_load(struct task_struct
*p
)
5143 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5145 update_cfs_rq_h_load(cfs_rq
);
5146 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5147 cfs_rq
->runnable_load_avg
+ 1);
5150 static inline void update_blocked_averages(int cpu
)
5154 static unsigned long task_h_load(struct task_struct
*p
)
5156 return p
->se
.avg
.load_avg_contrib
;
5160 /********** Helpers for find_busiest_group ************************/
5162 * sg_lb_stats - stats of a sched_group required for load_balancing
5164 struct sg_lb_stats
{
5165 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5166 unsigned long group_load
; /* Total load over the CPUs of the group */
5167 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5168 unsigned long load_per_task
;
5169 unsigned long group_power
;
5170 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5171 unsigned int group_capacity
;
5172 unsigned int idle_cpus
;
5173 unsigned int group_weight
;
5174 int group_imb
; /* Is there an imbalance in the group ? */
5175 int group_has_capacity
; /* Is there extra capacity in the group? */
5176 #ifdef CONFIG_NUMA_BALANCING
5177 unsigned int nr_numa_running
;
5178 unsigned int nr_preferred_running
;
5183 * sd_lb_stats - Structure to store the statistics of a sched_domain
5184 * during load balancing.
5186 struct sd_lb_stats
{
5187 struct sched_group
*busiest
; /* Busiest group in this sd */
5188 struct sched_group
*local
; /* Local group in this sd */
5189 unsigned long total_load
; /* Total load of all groups in sd */
5190 unsigned long total_pwr
; /* Total power of all groups in sd */
5191 unsigned long avg_load
; /* Average load across all groups in sd */
5193 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5194 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5197 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5200 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5201 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5202 * We must however clear busiest_stat::avg_load because
5203 * update_sd_pick_busiest() reads this before assignment.
5205 *sds
= (struct sd_lb_stats
){
5217 * get_sd_load_idx - Obtain the load index for a given sched domain.
5218 * @sd: The sched_domain whose load_idx is to be obtained.
5219 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5221 * Return: The load index.
5223 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5224 enum cpu_idle_type idle
)
5230 load_idx
= sd
->busy_idx
;
5233 case CPU_NEWLY_IDLE
:
5234 load_idx
= sd
->newidle_idx
;
5237 load_idx
= sd
->idle_idx
;
5244 static unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
5246 return SCHED_POWER_SCALE
;
5249 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
5251 return default_scale_freq_power(sd
, cpu
);
5254 static unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
5256 unsigned long weight
= sd
->span_weight
;
5257 unsigned long smt_gain
= sd
->smt_gain
;
5264 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
5266 return default_scale_smt_power(sd
, cpu
);
5269 static unsigned long scale_rt_power(int cpu
)
5271 struct rq
*rq
= cpu_rq(cpu
);
5272 u64 total
, available
, age_stamp
, avg
;
5275 * Since we're reading these variables without serialization make sure
5276 * we read them once before doing sanity checks on them.
5278 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5279 avg
= ACCESS_ONCE(rq
->rt_avg
);
5281 total
= sched_avg_period() + (rq_clock(rq
) - age_stamp
);
5283 if (unlikely(total
< avg
)) {
5284 /* Ensures that power won't end up being negative */
5287 available
= total
- avg
;
5290 if (unlikely((s64
)total
< SCHED_POWER_SCALE
))
5291 total
= SCHED_POWER_SCALE
;
5293 total
>>= SCHED_POWER_SHIFT
;
5295 return div_u64(available
, total
);
5298 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
5300 unsigned long weight
= sd
->span_weight
;
5301 unsigned long power
= SCHED_POWER_SCALE
;
5302 struct sched_group
*sdg
= sd
->groups
;
5304 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
5305 if (sched_feat(ARCH_POWER
))
5306 power
*= arch_scale_smt_power(sd
, cpu
);
5308 power
*= default_scale_smt_power(sd
, cpu
);
5310 power
>>= SCHED_POWER_SHIFT
;
5313 sdg
->sgp
->power_orig
= power
;
5315 if (sched_feat(ARCH_POWER
))
5316 power
*= arch_scale_freq_power(sd
, cpu
);
5318 power
*= default_scale_freq_power(sd
, cpu
);
5320 power
>>= SCHED_POWER_SHIFT
;
5322 power
*= scale_rt_power(cpu
);
5323 power
>>= SCHED_POWER_SHIFT
;
5328 cpu_rq(cpu
)->cpu_power
= power
;
5329 sdg
->sgp
->power
= power
;
5332 void update_group_power(struct sched_domain
*sd
, int cpu
)
5334 struct sched_domain
*child
= sd
->child
;
5335 struct sched_group
*group
, *sdg
= sd
->groups
;
5336 unsigned long power
, power_orig
;
5337 unsigned long interval
;
5339 interval
= msecs_to_jiffies(sd
->balance_interval
);
5340 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5341 sdg
->sgp
->next_update
= jiffies
+ interval
;
5344 update_cpu_power(sd
, cpu
);
5348 power_orig
= power
= 0;
5350 if (child
->flags
& SD_OVERLAP
) {
5352 * SD_OVERLAP domains cannot assume that child groups
5353 * span the current group.
5356 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
5357 struct sched_group_power
*sgp
;
5358 struct rq
*rq
= cpu_rq(cpu
);
5361 * build_sched_domains() -> init_sched_groups_power()
5362 * gets here before we've attached the domains to the
5365 * Use power_of(), which is set irrespective of domains
5366 * in update_cpu_power().
5368 * This avoids power/power_orig from being 0 and
5369 * causing divide-by-zero issues on boot.
5371 * Runtime updates will correct power_orig.
5373 if (unlikely(!rq
->sd
)) {
5374 power_orig
+= power_of(cpu
);
5375 power
+= power_of(cpu
);
5379 sgp
= rq
->sd
->groups
->sgp
;
5380 power_orig
+= sgp
->power_orig
;
5381 power
+= sgp
->power
;
5385 * !SD_OVERLAP domains can assume that child groups
5386 * span the current group.
5389 group
= child
->groups
;
5391 power_orig
+= group
->sgp
->power_orig
;
5392 power
+= group
->sgp
->power
;
5393 group
= group
->next
;
5394 } while (group
!= child
->groups
);
5397 sdg
->sgp
->power_orig
= power_orig
;
5398 sdg
->sgp
->power
= power
;
5402 * Try and fix up capacity for tiny siblings, this is needed when
5403 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5404 * which on its own isn't powerful enough.
5406 * See update_sd_pick_busiest() and check_asym_packing().
5409 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
5412 * Only siblings can have significantly less than SCHED_POWER_SCALE
5414 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
5418 * If ~90% of the cpu_power is still there, we're good.
5420 if (group
->sgp
->power
* 32 > group
->sgp
->power_orig
* 29)
5427 * Group imbalance indicates (and tries to solve) the problem where balancing
5428 * groups is inadequate due to tsk_cpus_allowed() constraints.
5430 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5431 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5434 * { 0 1 2 3 } { 4 5 6 7 }
5437 * If we were to balance group-wise we'd place two tasks in the first group and
5438 * two tasks in the second group. Clearly this is undesired as it will overload
5439 * cpu 3 and leave one of the cpus in the second group unused.
5441 * The current solution to this issue is detecting the skew in the first group
5442 * by noticing the lower domain failed to reach balance and had difficulty
5443 * moving tasks due to affinity constraints.
5445 * When this is so detected; this group becomes a candidate for busiest; see
5446 * update_sd_pick_busiest(). And calculate_imbalance() and
5447 * find_busiest_group() avoid some of the usual balance conditions to allow it
5448 * to create an effective group imbalance.
5450 * This is a somewhat tricky proposition since the next run might not find the
5451 * group imbalance and decide the groups need to be balanced again. A most
5452 * subtle and fragile situation.
5455 static inline int sg_imbalanced(struct sched_group
*group
)
5457 return group
->sgp
->imbalance
;
5461 * Compute the group capacity.
5463 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5464 * first dividing out the smt factor and computing the actual number of cores
5465 * and limit power unit capacity with that.
5467 static inline int sg_capacity(struct lb_env
*env
, struct sched_group
*group
)
5469 unsigned int capacity
, smt
, cpus
;
5470 unsigned int power
, power_orig
;
5472 power
= group
->sgp
->power
;
5473 power_orig
= group
->sgp
->power_orig
;
5474 cpus
= group
->group_weight
;
5476 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5477 smt
= DIV_ROUND_UP(SCHED_POWER_SCALE
* cpus
, power_orig
);
5478 capacity
= cpus
/ smt
; /* cores */
5480 capacity
= min_t(unsigned, capacity
, DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
));
5482 capacity
= fix_small_capacity(env
->sd
, group
);
5488 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5489 * @env: The load balancing environment.
5490 * @group: sched_group whose statistics are to be updated.
5491 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5492 * @local_group: Does group contain this_cpu.
5493 * @sgs: variable to hold the statistics for this group.
5495 static inline void update_sg_lb_stats(struct lb_env
*env
,
5496 struct sched_group
*group
, int load_idx
,
5497 int local_group
, struct sg_lb_stats
*sgs
)
5502 memset(sgs
, 0, sizeof(*sgs
));
5504 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5505 struct rq
*rq
= cpu_rq(i
);
5507 /* Bias balancing toward cpus of our domain */
5509 load
= target_load(i
, load_idx
);
5511 load
= source_load(i
, load_idx
);
5513 sgs
->group_load
+= load
;
5514 sgs
->sum_nr_running
+= rq
->nr_running
;
5515 #ifdef CONFIG_NUMA_BALANCING
5516 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
5517 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
5519 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
5524 /* Adjust by relative CPU power of the group */
5525 sgs
->group_power
= group
->sgp
->power
;
5526 sgs
->avg_load
= (sgs
->group_load
*SCHED_POWER_SCALE
) / sgs
->group_power
;
5528 if (sgs
->sum_nr_running
)
5529 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
5531 sgs
->group_weight
= group
->group_weight
;
5533 sgs
->group_imb
= sg_imbalanced(group
);
5534 sgs
->group_capacity
= sg_capacity(env
, group
);
5536 if (sgs
->group_capacity
> sgs
->sum_nr_running
)
5537 sgs
->group_has_capacity
= 1;
5541 * update_sd_pick_busiest - return 1 on busiest group
5542 * @env: The load balancing environment.
5543 * @sds: sched_domain statistics
5544 * @sg: sched_group candidate to be checked for being the busiest
5545 * @sgs: sched_group statistics
5547 * Determine if @sg is a busier group than the previously selected
5550 * Return: %true if @sg is a busier group than the previously selected
5551 * busiest group. %false otherwise.
5553 static bool update_sd_pick_busiest(struct lb_env
*env
,
5554 struct sd_lb_stats
*sds
,
5555 struct sched_group
*sg
,
5556 struct sg_lb_stats
*sgs
)
5558 if (sgs
->avg_load
<= sds
->busiest_stat
.avg_load
)
5561 if (sgs
->sum_nr_running
> sgs
->group_capacity
)
5568 * ASYM_PACKING needs to move all the work to the lowest
5569 * numbered CPUs in the group, therefore mark all groups
5570 * higher than ourself as busy.
5572 if ((env
->sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
5573 env
->dst_cpu
< group_first_cpu(sg
)) {
5577 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
5584 #ifdef CONFIG_NUMA_BALANCING
5585 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
5587 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
5589 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
5594 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
5596 if (rq
->nr_running
> rq
->nr_numa_running
)
5598 if (rq
->nr_running
> rq
->nr_preferred_running
)
5603 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
5608 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
5612 #endif /* CONFIG_NUMA_BALANCING */
5615 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5616 * @env: The load balancing environment.
5617 * @sds: variable to hold the statistics for this sched_domain.
5619 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5621 struct sched_domain
*child
= env
->sd
->child
;
5622 struct sched_group
*sg
= env
->sd
->groups
;
5623 struct sg_lb_stats tmp_sgs
;
5624 int load_idx
, prefer_sibling
= 0;
5626 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
5629 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
5632 struct sg_lb_stats
*sgs
= &tmp_sgs
;
5635 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
5638 sgs
= &sds
->local_stat
;
5640 if (env
->idle
!= CPU_NEWLY_IDLE
||
5641 time_after_eq(jiffies
, sg
->sgp
->next_update
))
5642 update_group_power(env
->sd
, env
->dst_cpu
);
5645 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
);
5651 * In case the child domain prefers tasks go to siblings
5652 * first, lower the sg capacity to one so that we'll try
5653 * and move all the excess tasks away. We lower the capacity
5654 * of a group only if the local group has the capacity to fit
5655 * these excess tasks, i.e. nr_running < group_capacity. The
5656 * extra check prevents the case where you always pull from the
5657 * heaviest group when it is already under-utilized (possible
5658 * with a large weight task outweighs the tasks on the system).
5660 if (prefer_sibling
&& sds
->local
&&
5661 sds
->local_stat
.group_has_capacity
)
5662 sgs
->group_capacity
= min(sgs
->group_capacity
, 1U);
5664 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
5666 sds
->busiest_stat
= *sgs
;
5670 /* Now, start updating sd_lb_stats */
5671 sds
->total_load
+= sgs
->group_load
;
5672 sds
->total_pwr
+= sgs
->group_power
;
5675 } while (sg
!= env
->sd
->groups
);
5677 if (env
->sd
->flags
& SD_NUMA
)
5678 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
5682 * check_asym_packing - Check to see if the group is packed into the
5685 * This is primarily intended to used at the sibling level. Some
5686 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5687 * case of POWER7, it can move to lower SMT modes only when higher
5688 * threads are idle. When in lower SMT modes, the threads will
5689 * perform better since they share less core resources. Hence when we
5690 * have idle threads, we want them to be the higher ones.
5692 * This packing function is run on idle threads. It checks to see if
5693 * the busiest CPU in this domain (core in the P7 case) has a higher
5694 * CPU number than the packing function is being run on. Here we are
5695 * assuming lower CPU number will be equivalent to lower a SMT thread
5698 * Return: 1 when packing is required and a task should be moved to
5699 * this CPU. The amount of the imbalance is returned in *imbalance.
5701 * @env: The load balancing environment.
5702 * @sds: Statistics of the sched_domain which is to be packed
5704 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5708 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
5714 busiest_cpu
= group_first_cpu(sds
->busiest
);
5715 if (env
->dst_cpu
> busiest_cpu
)
5718 env
->imbalance
= DIV_ROUND_CLOSEST(
5719 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_power
,
5726 * fix_small_imbalance - Calculate the minor imbalance that exists
5727 * amongst the groups of a sched_domain, during
5729 * @env: The load balancing environment.
5730 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5733 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5735 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
5736 unsigned int imbn
= 2;
5737 unsigned long scaled_busy_load_per_task
;
5738 struct sg_lb_stats
*local
, *busiest
;
5740 local
= &sds
->local_stat
;
5741 busiest
= &sds
->busiest_stat
;
5743 if (!local
->sum_nr_running
)
5744 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
5745 else if (busiest
->load_per_task
> local
->load_per_task
)
5748 scaled_busy_load_per_task
=
5749 (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5750 busiest
->group_power
;
5752 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
5753 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
5754 env
->imbalance
= busiest
->load_per_task
;
5759 * OK, we don't have enough imbalance to justify moving tasks,
5760 * however we may be able to increase total CPU power used by
5764 pwr_now
+= busiest
->group_power
*
5765 min(busiest
->load_per_task
, busiest
->avg_load
);
5766 pwr_now
+= local
->group_power
*
5767 min(local
->load_per_task
, local
->avg_load
);
5768 pwr_now
/= SCHED_POWER_SCALE
;
5770 /* Amount of load we'd subtract */
5771 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5772 busiest
->group_power
;
5773 if (busiest
->avg_load
> tmp
) {
5774 pwr_move
+= busiest
->group_power
*
5775 min(busiest
->load_per_task
,
5776 busiest
->avg_load
- tmp
);
5779 /* Amount of load we'd add */
5780 if (busiest
->avg_load
* busiest
->group_power
<
5781 busiest
->load_per_task
* SCHED_POWER_SCALE
) {
5782 tmp
= (busiest
->avg_load
* busiest
->group_power
) /
5785 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5788 pwr_move
+= local
->group_power
*
5789 min(local
->load_per_task
, local
->avg_load
+ tmp
);
5790 pwr_move
/= SCHED_POWER_SCALE
;
5792 /* Move if we gain throughput */
5793 if (pwr_move
> pwr_now
)
5794 env
->imbalance
= busiest
->load_per_task
;
5798 * calculate_imbalance - Calculate the amount of imbalance present within the
5799 * groups of a given sched_domain during load balance.
5800 * @env: load balance environment
5801 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5803 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5805 unsigned long max_pull
, load_above_capacity
= ~0UL;
5806 struct sg_lb_stats
*local
, *busiest
;
5808 local
= &sds
->local_stat
;
5809 busiest
= &sds
->busiest_stat
;
5811 if (busiest
->group_imb
) {
5813 * In the group_imb case we cannot rely on group-wide averages
5814 * to ensure cpu-load equilibrium, look at wider averages. XXX
5816 busiest
->load_per_task
=
5817 min(busiest
->load_per_task
, sds
->avg_load
);
5821 * In the presence of smp nice balancing, certain scenarios can have
5822 * max load less than avg load(as we skip the groups at or below
5823 * its cpu_power, while calculating max_load..)
5825 if (busiest
->avg_load
<= sds
->avg_load
||
5826 local
->avg_load
>= sds
->avg_load
) {
5828 return fix_small_imbalance(env
, sds
);
5831 if (!busiest
->group_imb
) {
5833 * Don't want to pull so many tasks that a group would go idle.
5834 * Except of course for the group_imb case, since then we might
5835 * have to drop below capacity to reach cpu-load equilibrium.
5837 load_above_capacity
=
5838 (busiest
->sum_nr_running
- busiest
->group_capacity
);
5840 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_POWER_SCALE
);
5841 load_above_capacity
/= busiest
->group_power
;
5845 * We're trying to get all the cpus to the average_load, so we don't
5846 * want to push ourselves above the average load, nor do we wish to
5847 * reduce the max loaded cpu below the average load. At the same time,
5848 * we also don't want to reduce the group load below the group capacity
5849 * (so that we can implement power-savings policies etc). Thus we look
5850 * for the minimum possible imbalance.
5852 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
5854 /* How much load to actually move to equalise the imbalance */
5855 env
->imbalance
= min(
5856 max_pull
* busiest
->group_power
,
5857 (sds
->avg_load
- local
->avg_load
) * local
->group_power
5858 ) / SCHED_POWER_SCALE
;
5861 * if *imbalance is less than the average load per runnable task
5862 * there is no guarantee that any tasks will be moved so we'll have
5863 * a think about bumping its value to force at least one task to be
5866 if (env
->imbalance
< busiest
->load_per_task
)
5867 return fix_small_imbalance(env
, sds
);
5870 /******* find_busiest_group() helpers end here *********************/
5873 * find_busiest_group - Returns the busiest group within the sched_domain
5874 * if there is an imbalance. If there isn't an imbalance, and
5875 * the user has opted for power-savings, it returns a group whose
5876 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5877 * such a group exists.
5879 * Also calculates the amount of weighted load which should be moved
5880 * to restore balance.
5882 * @env: The load balancing environment.
5884 * Return: - The busiest group if imbalance exists.
5885 * - If no imbalance and user has opted for power-savings balance,
5886 * return the least loaded group whose CPUs can be
5887 * put to idle by rebalancing its tasks onto our group.
5889 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
5891 struct sg_lb_stats
*local
, *busiest
;
5892 struct sd_lb_stats sds
;
5894 init_sd_lb_stats(&sds
);
5897 * Compute the various statistics relavent for load balancing at
5900 update_sd_lb_stats(env
, &sds
);
5901 local
= &sds
.local_stat
;
5902 busiest
= &sds
.busiest_stat
;
5904 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
5905 check_asym_packing(env
, &sds
))
5908 /* There is no busy sibling group to pull tasks from */
5909 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
5912 sds
.avg_load
= (SCHED_POWER_SCALE
* sds
.total_load
) / sds
.total_pwr
;
5915 * If the busiest group is imbalanced the below checks don't
5916 * work because they assume all things are equal, which typically
5917 * isn't true due to cpus_allowed constraints and the like.
5919 if (busiest
->group_imb
)
5922 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5923 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_capacity
&&
5924 !busiest
->group_has_capacity
)
5928 * If the local group is more busy than the selected busiest group
5929 * don't try and pull any tasks.
5931 if (local
->avg_load
>= busiest
->avg_load
)
5935 * Don't pull any tasks if this group is already above the domain
5938 if (local
->avg_load
>= sds
.avg_load
)
5941 if (env
->idle
== CPU_IDLE
) {
5943 * This cpu is idle. If the busiest group load doesn't
5944 * have more tasks than the number of available cpu's and
5945 * there is no imbalance between this and busiest group
5946 * wrt to idle cpu's, it is balanced.
5948 if ((local
->idle_cpus
< busiest
->idle_cpus
) &&
5949 busiest
->sum_nr_running
<= busiest
->group_weight
)
5953 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5954 * imbalance_pct to be conservative.
5956 if (100 * busiest
->avg_load
<=
5957 env
->sd
->imbalance_pct
* local
->avg_load
)
5962 /* Looks like there is an imbalance. Compute it */
5963 calculate_imbalance(env
, &sds
);
5972 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5974 static struct rq
*find_busiest_queue(struct lb_env
*env
,
5975 struct sched_group
*group
)
5977 struct rq
*busiest
= NULL
, *rq
;
5978 unsigned long busiest_load
= 0, busiest_power
= 1;
5981 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5982 unsigned long power
, capacity
, wl
;
5986 rt
= fbq_classify_rq(rq
);
5989 * We classify groups/runqueues into three groups:
5990 * - regular: there are !numa tasks
5991 * - remote: there are numa tasks that run on the 'wrong' node
5992 * - all: there is no distinction
5994 * In order to avoid migrating ideally placed numa tasks,
5995 * ignore those when there's better options.
5997 * If we ignore the actual busiest queue to migrate another
5998 * task, the next balance pass can still reduce the busiest
5999 * queue by moving tasks around inside the node.
6001 * If we cannot move enough load due to this classification
6002 * the next pass will adjust the group classification and
6003 * allow migration of more tasks.
6005 * Both cases only affect the total convergence complexity.
6007 if (rt
> env
->fbq_type
)
6010 power
= power_of(i
);
6011 capacity
= DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
);
6013 capacity
= fix_small_capacity(env
->sd
, group
);
6015 wl
= weighted_cpuload(i
);
6018 * When comparing with imbalance, use weighted_cpuload()
6019 * which is not scaled with the cpu power.
6021 if (capacity
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6025 * For the load comparisons with the other cpu's, consider
6026 * the weighted_cpuload() scaled with the cpu power, so that
6027 * the load can be moved away from the cpu that is potentially
6028 * running at a lower capacity.
6030 * Thus we're looking for max(wl_i / power_i), crosswise
6031 * multiplication to rid ourselves of the division works out
6032 * to: wl_i * power_j > wl_j * power_i; where j is our
6035 if (wl
* busiest_power
> busiest_load
* power
) {
6037 busiest_power
= power
;
6046 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6047 * so long as it is large enough.
6049 #define MAX_PINNED_INTERVAL 512
6051 /* Working cpumask for load_balance and load_balance_newidle. */
6052 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6054 static int need_active_balance(struct lb_env
*env
)
6056 struct sched_domain
*sd
= env
->sd
;
6058 if (env
->idle
== CPU_NEWLY_IDLE
) {
6061 * ASYM_PACKING needs to force migrate tasks from busy but
6062 * higher numbered CPUs in order to pack all tasks in the
6063 * lowest numbered CPUs.
6065 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6069 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6072 static int active_load_balance_cpu_stop(void *data
);
6074 static int should_we_balance(struct lb_env
*env
)
6076 struct sched_group
*sg
= env
->sd
->groups
;
6077 struct cpumask
*sg_cpus
, *sg_mask
;
6078 int cpu
, balance_cpu
= -1;
6081 * In the newly idle case, we will allow all the cpu's
6082 * to do the newly idle load balance.
6084 if (env
->idle
== CPU_NEWLY_IDLE
)
6087 sg_cpus
= sched_group_cpus(sg
);
6088 sg_mask
= sched_group_mask(sg
);
6089 /* Try to find first idle cpu */
6090 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6091 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6098 if (balance_cpu
== -1)
6099 balance_cpu
= group_balance_cpu(sg
);
6102 * First idle cpu or the first cpu(busiest) in this sched group
6103 * is eligible for doing load balancing at this and above domains.
6105 return balance_cpu
== env
->dst_cpu
;
6109 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6110 * tasks if there is an imbalance.
6112 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6113 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6114 int *continue_balancing
)
6116 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6117 struct sched_domain
*sd_parent
= sd
->parent
;
6118 struct sched_group
*group
;
6120 unsigned long flags
;
6121 struct cpumask
*cpus
= __get_cpu_var(load_balance_mask
);
6123 struct lb_env env
= {
6125 .dst_cpu
= this_cpu
,
6127 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6129 .loop_break
= sched_nr_migrate_break
,
6135 * For NEWLY_IDLE load_balancing, we don't need to consider
6136 * other cpus in our group
6138 if (idle
== CPU_NEWLY_IDLE
)
6139 env
.dst_grpmask
= NULL
;
6141 cpumask_copy(cpus
, cpu_active_mask
);
6143 schedstat_inc(sd
, lb_count
[idle
]);
6146 if (!should_we_balance(&env
)) {
6147 *continue_balancing
= 0;
6151 group
= find_busiest_group(&env
);
6153 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6157 busiest
= find_busiest_queue(&env
, group
);
6159 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6163 BUG_ON(busiest
== env
.dst_rq
);
6165 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6168 if (busiest
->nr_running
> 1) {
6170 * Attempt to move tasks. If find_busiest_group has found
6171 * an imbalance but busiest->nr_running <= 1, the group is
6172 * still unbalanced. ld_moved simply stays zero, so it is
6173 * correctly treated as an imbalance.
6175 env
.flags
|= LBF_ALL_PINNED
;
6176 env
.src_cpu
= busiest
->cpu
;
6177 env
.src_rq
= busiest
;
6178 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6181 local_irq_save(flags
);
6182 double_rq_lock(env
.dst_rq
, busiest
);
6185 * cur_ld_moved - load moved in current iteration
6186 * ld_moved - cumulative load moved across iterations
6188 cur_ld_moved
= move_tasks(&env
);
6189 ld_moved
+= cur_ld_moved
;
6190 double_rq_unlock(env
.dst_rq
, busiest
);
6191 local_irq_restore(flags
);
6194 * some other cpu did the load balance for us.
6196 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
6197 resched_cpu(env
.dst_cpu
);
6199 if (env
.flags
& LBF_NEED_BREAK
) {
6200 env
.flags
&= ~LBF_NEED_BREAK
;
6205 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6206 * us and move them to an alternate dst_cpu in our sched_group
6207 * where they can run. The upper limit on how many times we
6208 * iterate on same src_cpu is dependent on number of cpus in our
6211 * This changes load balance semantics a bit on who can move
6212 * load to a given_cpu. In addition to the given_cpu itself
6213 * (or a ilb_cpu acting on its behalf where given_cpu is
6214 * nohz-idle), we now have balance_cpu in a position to move
6215 * load to given_cpu. In rare situations, this may cause
6216 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6217 * _independently_ and at _same_ time to move some load to
6218 * given_cpu) causing exceess load to be moved to given_cpu.
6219 * This however should not happen so much in practice and
6220 * moreover subsequent load balance cycles should correct the
6221 * excess load moved.
6223 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6225 /* Prevent to re-select dst_cpu via env's cpus */
6226 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6228 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6229 env
.dst_cpu
= env
.new_dst_cpu
;
6230 env
.flags
&= ~LBF_DST_PINNED
;
6232 env
.loop_break
= sched_nr_migrate_break
;
6235 * Go back to "more_balance" rather than "redo" since we
6236 * need to continue with same src_cpu.
6242 * We failed to reach balance because of affinity.
6245 int *group_imbalance
= &sd_parent
->groups
->sgp
->imbalance
;
6247 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0) {
6248 *group_imbalance
= 1;
6249 } else if (*group_imbalance
)
6250 *group_imbalance
= 0;
6253 /* All tasks on this runqueue were pinned by CPU affinity */
6254 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6255 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6256 if (!cpumask_empty(cpus
)) {
6258 env
.loop_break
= sched_nr_migrate_break
;
6266 schedstat_inc(sd
, lb_failed
[idle
]);
6268 * Increment the failure counter only on periodic balance.
6269 * We do not want newidle balance, which can be very
6270 * frequent, pollute the failure counter causing
6271 * excessive cache_hot migrations and active balances.
6273 if (idle
!= CPU_NEWLY_IDLE
)
6274 sd
->nr_balance_failed
++;
6276 if (need_active_balance(&env
)) {
6277 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6279 /* don't kick the active_load_balance_cpu_stop,
6280 * if the curr task on busiest cpu can't be
6283 if (!cpumask_test_cpu(this_cpu
,
6284 tsk_cpus_allowed(busiest
->curr
))) {
6285 raw_spin_unlock_irqrestore(&busiest
->lock
,
6287 env
.flags
|= LBF_ALL_PINNED
;
6288 goto out_one_pinned
;
6292 * ->active_balance synchronizes accesses to
6293 * ->active_balance_work. Once set, it's cleared
6294 * only after active load balance is finished.
6296 if (!busiest
->active_balance
) {
6297 busiest
->active_balance
= 1;
6298 busiest
->push_cpu
= this_cpu
;
6301 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
6303 if (active_balance
) {
6304 stop_one_cpu_nowait(cpu_of(busiest
),
6305 active_load_balance_cpu_stop
, busiest
,
6306 &busiest
->active_balance_work
);
6310 * We've kicked active balancing, reset the failure
6313 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
6316 sd
->nr_balance_failed
= 0;
6318 if (likely(!active_balance
)) {
6319 /* We were unbalanced, so reset the balancing interval */
6320 sd
->balance_interval
= sd
->min_interval
;
6323 * If we've begun active balancing, start to back off. This
6324 * case may not be covered by the all_pinned logic if there
6325 * is only 1 task on the busy runqueue (because we don't call
6328 if (sd
->balance_interval
< sd
->max_interval
)
6329 sd
->balance_interval
*= 2;
6335 schedstat_inc(sd
, lb_balanced
[idle
]);
6337 sd
->nr_balance_failed
= 0;
6340 /* tune up the balancing interval */
6341 if (((env
.flags
& LBF_ALL_PINNED
) &&
6342 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
6343 (sd
->balance_interval
< sd
->max_interval
))
6344 sd
->balance_interval
*= 2;
6352 * idle_balance is called by schedule() if this_cpu is about to become
6353 * idle. Attempts to pull tasks from other CPUs.
6355 void idle_balance(int this_cpu
, struct rq
*this_rq
)
6357 struct sched_domain
*sd
;
6358 int pulled_task
= 0;
6359 unsigned long next_balance
= jiffies
+ HZ
;
6362 this_rq
->idle_stamp
= rq_clock(this_rq
);
6364 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
6368 * Drop the rq->lock, but keep IRQ/preempt disabled.
6370 raw_spin_unlock(&this_rq
->lock
);
6372 update_blocked_averages(this_cpu
);
6374 for_each_domain(this_cpu
, sd
) {
6375 unsigned long interval
;
6376 int continue_balancing
= 1;
6377 u64 t0
, domain_cost
;
6379 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6382 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
)
6385 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
6386 t0
= sched_clock_cpu(this_cpu
);
6388 /* If we've pulled tasks over stop searching: */
6389 pulled_task
= load_balance(this_cpu
, this_rq
,
6391 &continue_balancing
);
6393 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
6394 if (domain_cost
> sd
->max_newidle_lb_cost
)
6395 sd
->max_newidle_lb_cost
= domain_cost
;
6397 curr_cost
+= domain_cost
;
6400 interval
= msecs_to_jiffies(sd
->balance_interval
);
6401 if (time_after(next_balance
, sd
->last_balance
+ interval
))
6402 next_balance
= sd
->last_balance
+ interval
;
6404 this_rq
->idle_stamp
= 0;
6410 raw_spin_lock(&this_rq
->lock
);
6412 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
6414 * We are going idle. next_balance may be set based on
6415 * a busy processor. So reset next_balance.
6417 this_rq
->next_balance
= next_balance
;
6420 if (curr_cost
> this_rq
->max_idle_balance_cost
)
6421 this_rq
->max_idle_balance_cost
= curr_cost
;
6425 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6426 * running tasks off the busiest CPU onto idle CPUs. It requires at
6427 * least 1 task to be running on each physical CPU where possible, and
6428 * avoids physical / logical imbalances.
6430 static int active_load_balance_cpu_stop(void *data
)
6432 struct rq
*busiest_rq
= data
;
6433 int busiest_cpu
= cpu_of(busiest_rq
);
6434 int target_cpu
= busiest_rq
->push_cpu
;
6435 struct rq
*target_rq
= cpu_rq(target_cpu
);
6436 struct sched_domain
*sd
;
6438 raw_spin_lock_irq(&busiest_rq
->lock
);
6440 /* make sure the requested cpu hasn't gone down in the meantime */
6441 if (unlikely(busiest_cpu
!= smp_processor_id() ||
6442 !busiest_rq
->active_balance
))
6445 /* Is there any task to move? */
6446 if (busiest_rq
->nr_running
<= 1)
6450 * This condition is "impossible", if it occurs
6451 * we need to fix it. Originally reported by
6452 * Bjorn Helgaas on a 128-cpu setup.
6454 BUG_ON(busiest_rq
== target_rq
);
6456 /* move a task from busiest_rq to target_rq */
6457 double_lock_balance(busiest_rq
, target_rq
);
6459 /* Search for an sd spanning us and the target CPU. */
6461 for_each_domain(target_cpu
, sd
) {
6462 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
6463 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
6468 struct lb_env env
= {
6470 .dst_cpu
= target_cpu
,
6471 .dst_rq
= target_rq
,
6472 .src_cpu
= busiest_rq
->cpu
,
6473 .src_rq
= busiest_rq
,
6477 schedstat_inc(sd
, alb_count
);
6479 if (move_one_task(&env
))
6480 schedstat_inc(sd
, alb_pushed
);
6482 schedstat_inc(sd
, alb_failed
);
6485 double_unlock_balance(busiest_rq
, target_rq
);
6487 busiest_rq
->active_balance
= 0;
6488 raw_spin_unlock_irq(&busiest_rq
->lock
);
6492 #ifdef CONFIG_NO_HZ_COMMON
6494 * idle load balancing details
6495 * - When one of the busy CPUs notice that there may be an idle rebalancing
6496 * needed, they will kick the idle load balancer, which then does idle
6497 * load balancing for all the idle CPUs.
6500 cpumask_var_t idle_cpus_mask
;
6502 unsigned long next_balance
; /* in jiffy units */
6503 } nohz ____cacheline_aligned
;
6505 static inline int find_new_ilb(int call_cpu
)
6507 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
6509 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
6516 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6517 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6518 * CPU (if there is one).
6520 static void nohz_balancer_kick(int cpu
)
6524 nohz
.next_balance
++;
6526 ilb_cpu
= find_new_ilb(cpu
);
6528 if (ilb_cpu
>= nr_cpu_ids
)
6531 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
6534 * Use smp_send_reschedule() instead of resched_cpu().
6535 * This way we generate a sched IPI on the target cpu which
6536 * is idle. And the softirq performing nohz idle load balance
6537 * will be run before returning from the IPI.
6539 smp_send_reschedule(ilb_cpu
);
6543 static inline void nohz_balance_exit_idle(int cpu
)
6545 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
6546 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
6547 atomic_dec(&nohz
.nr_cpus
);
6548 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
6552 static inline void set_cpu_sd_state_busy(void)
6554 struct sched_domain
*sd
;
6555 int cpu
= smp_processor_id();
6558 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
6560 if (!sd
|| !sd
->nohz_idle
)
6564 atomic_inc(&sd
->groups
->sgp
->nr_busy_cpus
);
6569 void set_cpu_sd_state_idle(void)
6571 struct sched_domain
*sd
;
6572 int cpu
= smp_processor_id();
6575 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
6577 if (!sd
|| sd
->nohz_idle
)
6581 atomic_dec(&sd
->groups
->sgp
->nr_busy_cpus
);
6587 * This routine will record that the cpu is going idle with tick stopped.
6588 * This info will be used in performing idle load balancing in the future.
6590 void nohz_balance_enter_idle(int cpu
)
6593 * If this cpu is going down, then nothing needs to be done.
6595 if (!cpu_active(cpu
))
6598 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
6601 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
6602 atomic_inc(&nohz
.nr_cpus
);
6603 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
6606 static int sched_ilb_notifier(struct notifier_block
*nfb
,
6607 unsigned long action
, void *hcpu
)
6609 switch (action
& ~CPU_TASKS_FROZEN
) {
6611 nohz_balance_exit_idle(smp_processor_id());
6619 static DEFINE_SPINLOCK(balancing
);
6622 * Scale the max load_balance interval with the number of CPUs in the system.
6623 * This trades load-balance latency on larger machines for less cross talk.
6625 void update_max_interval(void)
6627 max_load_balance_interval
= HZ
*num_online_cpus()/10;
6631 * It checks each scheduling domain to see if it is due to be balanced,
6632 * and initiates a balancing operation if so.
6634 * Balancing parameters are set up in init_sched_domains.
6636 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
6638 int continue_balancing
= 1;
6639 struct rq
*rq
= cpu_rq(cpu
);
6640 unsigned long interval
;
6641 struct sched_domain
*sd
;
6642 /* Earliest time when we have to do rebalance again */
6643 unsigned long next_balance
= jiffies
+ 60*HZ
;
6644 int update_next_balance
= 0;
6645 int need_serialize
, need_decay
= 0;
6648 update_blocked_averages(cpu
);
6651 for_each_domain(cpu
, sd
) {
6653 * Decay the newidle max times here because this is a regular
6654 * visit to all the domains. Decay ~1% per second.
6656 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
6657 sd
->max_newidle_lb_cost
=
6658 (sd
->max_newidle_lb_cost
* 253) / 256;
6659 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
6662 max_cost
+= sd
->max_newidle_lb_cost
;
6664 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6668 * Stop the load balance at this level. There is another
6669 * CPU in our sched group which is doing load balancing more
6672 if (!continue_balancing
) {
6678 interval
= sd
->balance_interval
;
6679 if (idle
!= CPU_IDLE
)
6680 interval
*= sd
->busy_factor
;
6682 /* scale ms to jiffies */
6683 interval
= msecs_to_jiffies(interval
);
6684 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6686 need_serialize
= sd
->flags
& SD_SERIALIZE
;
6688 if (need_serialize
) {
6689 if (!spin_trylock(&balancing
))
6693 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
6694 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
6696 * The LBF_DST_PINNED logic could have changed
6697 * env->dst_cpu, so we can't know our idle
6698 * state even if we migrated tasks. Update it.
6700 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
6702 sd
->last_balance
= jiffies
;
6705 spin_unlock(&balancing
);
6707 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
6708 next_balance
= sd
->last_balance
+ interval
;
6709 update_next_balance
= 1;
6714 * Ensure the rq-wide value also decays but keep it at a
6715 * reasonable floor to avoid funnies with rq->avg_idle.
6717 rq
->max_idle_balance_cost
=
6718 max((u64
)sysctl_sched_migration_cost
, max_cost
);
6723 * next_balance will be updated only when there is a need.
6724 * When the cpu is attached to null domain for ex, it will not be
6727 if (likely(update_next_balance
))
6728 rq
->next_balance
= next_balance
;
6731 #ifdef CONFIG_NO_HZ_COMMON
6733 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6734 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6736 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
)
6738 struct rq
*this_rq
= cpu_rq(this_cpu
);
6742 if (idle
!= CPU_IDLE
||
6743 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
6746 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
6747 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
6751 * If this cpu gets work to do, stop the load balancing
6752 * work being done for other cpus. Next load
6753 * balancing owner will pick it up.
6758 rq
= cpu_rq(balance_cpu
);
6760 raw_spin_lock_irq(&rq
->lock
);
6761 update_rq_clock(rq
);
6762 update_idle_cpu_load(rq
);
6763 raw_spin_unlock_irq(&rq
->lock
);
6765 rebalance_domains(balance_cpu
, CPU_IDLE
);
6767 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
6768 this_rq
->next_balance
= rq
->next_balance
;
6770 nohz
.next_balance
= this_rq
->next_balance
;
6772 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
6776 * Current heuristic for kicking the idle load balancer in the presence
6777 * of an idle cpu is the system.
6778 * - This rq has more than one task.
6779 * - At any scheduler domain level, this cpu's scheduler group has multiple
6780 * busy cpu's exceeding the group's power.
6781 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6782 * domain span are idle.
6784 static inline int nohz_kick_needed(struct rq
*rq
, int cpu
)
6786 unsigned long now
= jiffies
;
6787 struct sched_domain
*sd
;
6788 struct sched_group_power
*sgp
;
6791 if (unlikely(idle_cpu(cpu
)))
6795 * We may be recently in ticked or tickless idle mode. At the first
6796 * busy tick after returning from idle, we will update the busy stats.
6798 set_cpu_sd_state_busy();
6799 nohz_balance_exit_idle(cpu
);
6802 * None are in tickless mode and hence no need for NOHZ idle load
6805 if (likely(!atomic_read(&nohz
.nr_cpus
)))
6808 if (time_before(now
, nohz
.next_balance
))
6811 if (rq
->nr_running
>= 2)
6815 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
6818 sgp
= sd
->groups
->sgp
;
6819 nr_busy
= atomic_read(&sgp
->nr_busy_cpus
);
6822 goto need_kick_unlock
;
6825 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
6827 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
6828 sched_domain_span(sd
)) < cpu
))
6829 goto need_kick_unlock
;
6840 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
) { }
6844 * run_rebalance_domains is triggered when needed from the scheduler tick.
6845 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6847 static void run_rebalance_domains(struct softirq_action
*h
)
6849 int this_cpu
= smp_processor_id();
6850 struct rq
*this_rq
= cpu_rq(this_cpu
);
6851 enum cpu_idle_type idle
= this_rq
->idle_balance
?
6852 CPU_IDLE
: CPU_NOT_IDLE
;
6854 rebalance_domains(this_cpu
, idle
);
6857 * If this cpu has a pending nohz_balance_kick, then do the
6858 * balancing on behalf of the other idle cpus whose ticks are
6861 nohz_idle_balance(this_cpu
, idle
);
6864 static inline int on_null_domain(int cpu
)
6866 return !rcu_dereference_sched(cpu_rq(cpu
)->sd
);
6870 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6872 void trigger_load_balance(struct rq
*rq
, int cpu
)
6874 /* Don't need to rebalance while attached to NULL domain */
6875 if (time_after_eq(jiffies
, rq
->next_balance
) &&
6876 likely(!on_null_domain(cpu
)))
6877 raise_softirq(SCHED_SOFTIRQ
);
6878 #ifdef CONFIG_NO_HZ_COMMON
6879 if (nohz_kick_needed(rq
, cpu
) && likely(!on_null_domain(cpu
)))
6880 nohz_balancer_kick(cpu
);
6884 static void rq_online_fair(struct rq
*rq
)
6889 static void rq_offline_fair(struct rq
*rq
)
6893 /* Ensure any throttled groups are reachable by pick_next_task */
6894 unthrottle_offline_cfs_rqs(rq
);
6897 #endif /* CONFIG_SMP */
6900 * scheduler tick hitting a task of our scheduling class:
6902 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
6904 struct cfs_rq
*cfs_rq
;
6905 struct sched_entity
*se
= &curr
->se
;
6907 for_each_sched_entity(se
) {
6908 cfs_rq
= cfs_rq_of(se
);
6909 entity_tick(cfs_rq
, se
, queued
);
6912 if (numabalancing_enabled
)
6913 task_tick_numa(rq
, curr
);
6915 update_rq_runnable_avg(rq
, 1);
6919 * called on fork with the child task as argument from the parent's context
6920 * - child not yet on the tasklist
6921 * - preemption disabled
6923 static void task_fork_fair(struct task_struct
*p
)
6925 struct cfs_rq
*cfs_rq
;
6926 struct sched_entity
*se
= &p
->se
, *curr
;
6927 int this_cpu
= smp_processor_id();
6928 struct rq
*rq
= this_rq();
6929 unsigned long flags
;
6931 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6933 update_rq_clock(rq
);
6935 cfs_rq
= task_cfs_rq(current
);
6936 curr
= cfs_rq
->curr
;
6939 * Not only the cpu but also the task_group of the parent might have
6940 * been changed after parent->se.parent,cfs_rq were copied to
6941 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6942 * of child point to valid ones.
6945 __set_task_cpu(p
, this_cpu
);
6948 update_curr(cfs_rq
);
6951 se
->vruntime
= curr
->vruntime
;
6952 place_entity(cfs_rq
, se
, 1);
6954 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
6956 * Upon rescheduling, sched_class::put_prev_task() will place
6957 * 'current' within the tree based on its new key value.
6959 swap(curr
->vruntime
, se
->vruntime
);
6960 resched_task(rq
->curr
);
6963 se
->vruntime
-= cfs_rq
->min_vruntime
;
6965 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6969 * Priority of the task has changed. Check to see if we preempt
6973 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
6979 * Reschedule if we are currently running on this runqueue and
6980 * our priority decreased, or if we are not currently running on
6981 * this runqueue and our priority is higher than the current's
6983 if (rq
->curr
== p
) {
6984 if (p
->prio
> oldprio
)
6985 resched_task(rq
->curr
);
6987 check_preempt_curr(rq
, p
, 0);
6990 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
6992 struct sched_entity
*se
= &p
->se
;
6993 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6996 * Ensure the task's vruntime is normalized, so that when its
6997 * switched back to the fair class the enqueue_entity(.flags=0) will
6998 * do the right thing.
7000 * If it was on_rq, then the dequeue_entity(.flags=0) will already
7001 * have normalized the vruntime, if it was !on_rq, then only when
7002 * the task is sleeping will it still have non-normalized vruntime.
7004 if (!se
->on_rq
&& p
->state
!= TASK_RUNNING
) {
7006 * Fix up our vruntime so that the current sleep doesn't
7007 * cause 'unlimited' sleep bonus.
7009 place_entity(cfs_rq
, se
, 0);
7010 se
->vruntime
-= cfs_rq
->min_vruntime
;
7015 * Remove our load from contribution when we leave sched_fair
7016 * and ensure we don't carry in an old decay_count if we
7019 if (se
->avg
.decay_count
) {
7020 __synchronize_entity_decay(se
);
7021 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7027 * We switched to the sched_fair class.
7029 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7035 * We were most likely switched from sched_rt, so
7036 * kick off the schedule if running, otherwise just see
7037 * if we can still preempt the current task.
7040 resched_task(rq
->curr
);
7042 check_preempt_curr(rq
, p
, 0);
7045 /* Account for a task changing its policy or group.
7047 * This routine is mostly called to set cfs_rq->curr field when a task
7048 * migrates between groups/classes.
7050 static void set_curr_task_fair(struct rq
*rq
)
7052 struct sched_entity
*se
= &rq
->curr
->se
;
7054 for_each_sched_entity(se
) {
7055 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7057 set_next_entity(cfs_rq
, se
);
7058 /* ensure bandwidth has been allocated on our new cfs_rq */
7059 account_cfs_rq_runtime(cfs_rq
, 0);
7063 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7065 cfs_rq
->tasks_timeline
= RB_ROOT
;
7066 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7067 #ifndef CONFIG_64BIT
7068 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7071 atomic64_set(&cfs_rq
->decay_counter
, 1);
7072 atomic_long_set(&cfs_rq
->removed_load
, 0);
7076 #ifdef CONFIG_FAIR_GROUP_SCHED
7077 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
7079 struct cfs_rq
*cfs_rq
;
7081 * If the task was not on the rq at the time of this cgroup movement
7082 * it must have been asleep, sleeping tasks keep their ->vruntime
7083 * absolute on their old rq until wakeup (needed for the fair sleeper
7084 * bonus in place_entity()).
7086 * If it was on the rq, we've just 'preempted' it, which does convert
7087 * ->vruntime to a relative base.
7089 * Make sure both cases convert their relative position when migrating
7090 * to another cgroup's rq. This does somewhat interfere with the
7091 * fair sleeper stuff for the first placement, but who cares.
7094 * When !on_rq, vruntime of the task has usually NOT been normalized.
7095 * But there are some cases where it has already been normalized:
7097 * - Moving a forked child which is waiting for being woken up by
7098 * wake_up_new_task().
7099 * - Moving a task which has been woken up by try_to_wake_up() and
7100 * waiting for actually being woken up by sched_ttwu_pending().
7102 * To prevent boost or penalty in the new cfs_rq caused by delta
7103 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7105 if (!on_rq
&& (!p
->se
.sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7109 p
->se
.vruntime
-= cfs_rq_of(&p
->se
)->min_vruntime
;
7110 set_task_rq(p
, task_cpu(p
));
7112 cfs_rq
= cfs_rq_of(&p
->se
);
7113 p
->se
.vruntime
+= cfs_rq
->min_vruntime
;
7116 * migrate_task_rq_fair() will have removed our previous
7117 * contribution, but we must synchronize for ongoing future
7120 p
->se
.avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7121 cfs_rq
->blocked_load_avg
+= p
->se
.avg
.load_avg_contrib
;
7126 void free_fair_sched_group(struct task_group
*tg
)
7130 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7132 for_each_possible_cpu(i
) {
7134 kfree(tg
->cfs_rq
[i
]);
7143 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7145 struct cfs_rq
*cfs_rq
;
7146 struct sched_entity
*se
;
7149 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7152 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7156 tg
->shares
= NICE_0_LOAD
;
7158 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7160 for_each_possible_cpu(i
) {
7161 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7162 GFP_KERNEL
, cpu_to_node(i
));
7166 se
= kzalloc_node(sizeof(struct sched_entity
),
7167 GFP_KERNEL
, cpu_to_node(i
));
7171 init_cfs_rq(cfs_rq
);
7172 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
7183 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7185 struct rq
*rq
= cpu_rq(cpu
);
7186 unsigned long flags
;
7189 * Only empty task groups can be destroyed; so we can speculatively
7190 * check on_list without danger of it being re-added.
7192 if (!tg
->cfs_rq
[cpu
]->on_list
)
7195 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7196 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
7197 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7200 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7201 struct sched_entity
*se
, int cpu
,
7202 struct sched_entity
*parent
)
7204 struct rq
*rq
= cpu_rq(cpu
);
7208 init_cfs_rq_runtime(cfs_rq
);
7210 tg
->cfs_rq
[cpu
] = cfs_rq
;
7213 /* se could be NULL for root_task_group */
7218 se
->cfs_rq
= &rq
->cfs
;
7220 se
->cfs_rq
= parent
->my_q
;
7223 /* guarantee group entities always have weight */
7224 update_load_set(&se
->load
, NICE_0_LOAD
);
7225 se
->parent
= parent
;
7228 static DEFINE_MUTEX(shares_mutex
);
7230 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7233 unsigned long flags
;
7236 * We can't change the weight of the root cgroup.
7241 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
7243 mutex_lock(&shares_mutex
);
7244 if (tg
->shares
== shares
)
7247 tg
->shares
= shares
;
7248 for_each_possible_cpu(i
) {
7249 struct rq
*rq
= cpu_rq(i
);
7250 struct sched_entity
*se
;
7253 /* Propagate contribution to hierarchy */
7254 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7256 /* Possible calls to update_curr() need rq clock */
7257 update_rq_clock(rq
);
7258 for_each_sched_entity(se
)
7259 update_cfs_shares(group_cfs_rq(se
));
7260 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7264 mutex_unlock(&shares_mutex
);
7267 #else /* CONFIG_FAIR_GROUP_SCHED */
7269 void free_fair_sched_group(struct task_group
*tg
) { }
7271 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7276 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
7278 #endif /* CONFIG_FAIR_GROUP_SCHED */
7281 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
7283 struct sched_entity
*se
= &task
->se
;
7284 unsigned int rr_interval
= 0;
7287 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7290 if (rq
->cfs
.load
.weight
)
7291 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
7297 * All the scheduling class methods:
7299 const struct sched_class fair_sched_class
= {
7300 .next
= &idle_sched_class
,
7301 .enqueue_task
= enqueue_task_fair
,
7302 .dequeue_task
= dequeue_task_fair
,
7303 .yield_task
= yield_task_fair
,
7304 .yield_to_task
= yield_to_task_fair
,
7306 .check_preempt_curr
= check_preempt_wakeup
,
7308 .pick_next_task
= pick_next_task_fair
,
7309 .put_prev_task
= put_prev_task_fair
,
7312 .select_task_rq
= select_task_rq_fair
,
7313 .migrate_task_rq
= migrate_task_rq_fair
,
7315 .rq_online
= rq_online_fair
,
7316 .rq_offline
= rq_offline_fair
,
7318 .task_waking
= task_waking_fair
,
7321 .set_curr_task
= set_curr_task_fair
,
7322 .task_tick
= task_tick_fair
,
7323 .task_fork
= task_fork_fair
,
7325 .prio_changed
= prio_changed_fair
,
7326 .switched_from
= switched_from_fair
,
7327 .switched_to
= switched_to_fair
,
7329 .get_rr_interval
= get_rr_interval_fair
,
7331 #ifdef CONFIG_FAIR_GROUP_SCHED
7332 .task_move_group
= task_move_group_fair
,
7336 #ifdef CONFIG_SCHED_DEBUG
7337 void print_cfs_stats(struct seq_file
*m
, int cpu
)
7339 struct cfs_rq
*cfs_rq
;
7342 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
7343 print_cfs_rq(m
, cpu
, cfs_rq
);
7348 __init
void init_sched_fair_class(void)
7351 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7353 #ifdef CONFIG_NO_HZ_COMMON
7354 nohz
.next_balance
= jiffies
;
7355 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7356 cpu_notifier(sched_ilb_notifier
, 0);