2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency
= 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG
;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity
= 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency
= 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly
;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
93 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window
= 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
116 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
122 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
128 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling
) {
149 case SCHED_TUNABLESCALING_NONE
:
152 case SCHED_TUNABLESCALING_LINEAR
:
155 case SCHED_TUNABLESCALING_LOG
:
157 factor
= 1 + ilog2(cpus
);
164 static void update_sysctl(void)
166 unsigned int factor
= get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity
);
171 SET_SYSCTL(sched_latency
);
172 SET_SYSCTL(sched_wakeup_granularity
);
176 void sched_init_granularity(void)
181 #define WMULT_CONST (~0U)
182 #define WMULT_SHIFT 32
184 static void __update_inv_weight(struct load_weight
*lw
)
188 if (likely(lw
->inv_weight
))
191 w
= scale_load_down(lw
->weight
);
193 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
195 else if (unlikely(!w
))
196 lw
->inv_weight
= WMULT_CONST
;
198 lw
->inv_weight
= WMULT_CONST
/ w
;
202 * delta_exec * weight / lw.weight
204 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
206 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
207 * we're guaranteed shift stays positive because inv_weight is guaranteed to
208 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
210 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
211 * weight/lw.weight <= 1, and therefore our shift will also be positive.
213 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
215 u64 fact
= scale_load_down(weight
);
216 int shift
= WMULT_SHIFT
;
218 __update_inv_weight(lw
);
220 if (unlikely(fact
>> 32)) {
227 /* hint to use a 32x32->64 mul */
228 fact
= (u64
)(u32
)fact
* lw
->inv_weight
;
235 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
239 const struct sched_class fair_sched_class
;
241 /**************************************************************
242 * CFS operations on generic schedulable entities:
245 #ifdef CONFIG_FAIR_GROUP_SCHED
247 /* cpu runqueue to which this cfs_rq is attached */
248 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
253 /* An entity is a task if it doesn't "own" a runqueue */
254 #define entity_is_task(se) (!se->my_q)
256 static inline struct task_struct
*task_of(struct sched_entity
*se
)
258 #ifdef CONFIG_SCHED_DEBUG
259 WARN_ON_ONCE(!entity_is_task(se
));
261 return container_of(se
, struct task_struct
, se
);
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
268 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
285 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
288 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
290 if (!cfs_rq
->on_list
) {
292 * Ensure we either appear before our parent (if already
293 * enqueued) or force our parent to appear after us when it is
294 * enqueued. The fact that we always enqueue bottom-up
295 * reduces this to two cases.
297 if (cfs_rq
->tg
->parent
&&
298 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
299 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
300 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
302 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
303 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
307 /* We should have no load, but we need to update last_decay. */
308 update_cfs_rq_blocked_load(cfs_rq
, 0);
312 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
314 if (cfs_rq
->on_list
) {
315 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
320 /* Iterate thr' all leaf cfs_rq's on a runqueue */
321 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
322 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
324 /* Do the two (enqueued) entities belong to the same group ? */
325 static inline struct cfs_rq
*
326 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
328 if (se
->cfs_rq
== pse
->cfs_rq
)
334 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
340 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
342 int se_depth
, pse_depth
;
345 * preemption test can be made between sibling entities who are in the
346 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
347 * both tasks until we find their ancestors who are siblings of common
351 /* First walk up until both entities are at same depth */
352 se_depth
= (*se
)->depth
;
353 pse_depth
= (*pse
)->depth
;
355 while (se_depth
> pse_depth
) {
357 *se
= parent_entity(*se
);
360 while (pse_depth
> se_depth
) {
362 *pse
= parent_entity(*pse
);
365 while (!is_same_group(*se
, *pse
)) {
366 *se
= parent_entity(*se
);
367 *pse
= parent_entity(*pse
);
371 #else /* !CONFIG_FAIR_GROUP_SCHED */
373 static inline struct task_struct
*task_of(struct sched_entity
*se
)
375 return container_of(se
, struct task_struct
, se
);
378 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
380 return container_of(cfs_rq
, struct rq
, cfs
);
383 #define entity_is_task(se) 1
385 #define for_each_sched_entity(se) \
386 for (; se; se = NULL)
388 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
390 return &task_rq(p
)->cfs
;
393 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
395 struct task_struct
*p
= task_of(se
);
396 struct rq
*rq
= task_rq(p
);
401 /* runqueue "owned" by this group */
402 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
407 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
411 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
415 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
416 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
418 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
424 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
428 #endif /* CONFIG_FAIR_GROUP_SCHED */
430 static __always_inline
431 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
433 /**************************************************************
434 * Scheduling class tree data structure manipulation methods:
437 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
439 s64 delta
= (s64
)(vruntime
- max_vruntime
);
441 max_vruntime
= vruntime
;
446 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
448 s64 delta
= (s64
)(vruntime
- min_vruntime
);
450 min_vruntime
= vruntime
;
455 static inline int entity_before(struct sched_entity
*a
,
456 struct sched_entity
*b
)
458 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
461 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
463 u64 vruntime
= cfs_rq
->min_vruntime
;
466 vruntime
= cfs_rq
->curr
->vruntime
;
468 if (cfs_rq
->rb_leftmost
) {
469 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
474 vruntime
= se
->vruntime
;
476 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
479 /* ensure we never gain time by being placed backwards. */
480 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
483 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
488 * Enqueue an entity into the rb-tree:
490 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
492 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
493 struct rb_node
*parent
= NULL
;
494 struct sched_entity
*entry
;
498 * Find the right place in the rbtree:
502 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
504 * We dont care about collisions. Nodes with
505 * the same key stay together.
507 if (entity_before(se
, entry
)) {
508 link
= &parent
->rb_left
;
510 link
= &parent
->rb_right
;
516 * Maintain a cache of leftmost tree entries (it is frequently
520 cfs_rq
->rb_leftmost
= &se
->run_node
;
522 rb_link_node(&se
->run_node
, parent
, link
);
523 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
526 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
528 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
529 struct rb_node
*next_node
;
531 next_node
= rb_next(&se
->run_node
);
532 cfs_rq
->rb_leftmost
= next_node
;
535 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
538 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
540 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
545 return rb_entry(left
, struct sched_entity
, run_node
);
548 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
550 struct rb_node
*next
= rb_next(&se
->run_node
);
555 return rb_entry(next
, struct sched_entity
, run_node
);
558 #ifdef CONFIG_SCHED_DEBUG
559 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
561 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
566 return rb_entry(last
, struct sched_entity
, run_node
);
569 /**************************************************************
570 * Scheduling class statistics methods:
573 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
574 void __user
*buffer
, size_t *lenp
,
577 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
578 int factor
= get_update_sysctl_factor();
583 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
584 sysctl_sched_min_granularity
);
586 #define WRT_SYSCTL(name) \
587 (normalized_sysctl_##name = sysctl_##name / (factor))
588 WRT_SYSCTL(sched_min_granularity
);
589 WRT_SYSCTL(sched_latency
);
590 WRT_SYSCTL(sched_wakeup_granularity
);
600 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
602 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
603 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
609 * The idea is to set a period in which each task runs once.
611 * When there are too many tasks (sched_nr_latency) we have to stretch
612 * this period because otherwise the slices get too small.
614 * p = (nr <= nl) ? l : l*nr/nl
616 static u64
__sched_period(unsigned long nr_running
)
618 u64 period
= sysctl_sched_latency
;
619 unsigned long nr_latency
= sched_nr_latency
;
621 if (unlikely(nr_running
> nr_latency
)) {
622 period
= sysctl_sched_min_granularity
;
623 period
*= nr_running
;
630 * We calculate the wall-time slice from the period by taking a part
631 * proportional to the weight.
635 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
637 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
639 for_each_sched_entity(se
) {
640 struct load_weight
*load
;
641 struct load_weight lw
;
643 cfs_rq
= cfs_rq_of(se
);
644 load
= &cfs_rq
->load
;
646 if (unlikely(!se
->on_rq
)) {
649 update_load_add(&lw
, se
->load
.weight
);
652 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
658 * We calculate the vruntime slice of a to-be-inserted task.
662 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
664 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
668 static int select_idle_sibling(struct task_struct
*p
, int cpu
);
669 static unsigned long task_h_load(struct task_struct
*p
);
671 static inline void __update_task_entity_contrib(struct sched_entity
*se
);
673 /* Give new task start runnable values to heavy its load in infant time */
674 void init_task_runnable_average(struct task_struct
*p
)
678 p
->se
.avg
.decay_count
= 0;
679 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
680 p
->se
.avg
.runnable_avg_sum
= slice
;
681 p
->se
.avg
.runnable_avg_period
= slice
;
682 __update_task_entity_contrib(&p
->se
);
685 void init_task_runnable_average(struct task_struct
*p
)
691 * Update the current task's runtime statistics.
693 static void update_curr(struct cfs_rq
*cfs_rq
)
695 struct sched_entity
*curr
= cfs_rq
->curr
;
696 u64 now
= rq_clock_task(rq_of(cfs_rq
));
702 delta_exec
= now
- curr
->exec_start
;
703 if (unlikely((s64
)delta_exec
<= 0))
706 curr
->exec_start
= now
;
708 schedstat_set(curr
->statistics
.exec_max
,
709 max(delta_exec
, curr
->statistics
.exec_max
));
711 curr
->sum_exec_runtime
+= delta_exec
;
712 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
714 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
715 update_min_vruntime(cfs_rq
);
717 if (entity_is_task(curr
)) {
718 struct task_struct
*curtask
= task_of(curr
);
720 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
721 cpuacct_charge(curtask
, delta_exec
);
722 account_group_exec_runtime(curtask
, delta_exec
);
725 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
729 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
731 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
735 * Task is being enqueued - update stats:
737 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
740 * Are we enqueueing a waiting task? (for current tasks
741 * a dequeue/enqueue event is a NOP)
743 if (se
!= cfs_rq
->curr
)
744 update_stats_wait_start(cfs_rq
, se
);
748 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
750 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
751 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
752 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
753 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
754 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
755 #ifdef CONFIG_SCHEDSTATS
756 if (entity_is_task(se
)) {
757 trace_sched_stat_wait(task_of(se
),
758 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
761 schedstat_set(se
->statistics
.wait_start
, 0);
765 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
768 * Mark the end of the wait period if dequeueing a
771 if (se
!= cfs_rq
->curr
)
772 update_stats_wait_end(cfs_rq
, se
);
776 * We are picking a new current task - update its stats:
779 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
782 * We are starting a new run period:
784 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
787 /**************************************************
788 * Scheduling class queueing methods:
791 #ifdef CONFIG_NUMA_BALANCING
793 * Approximate time to scan a full NUMA task in ms. The task scan period is
794 * calculated based on the tasks virtual memory size and
795 * numa_balancing_scan_size.
797 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
798 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
800 /* Portion of address space to scan in MB */
801 unsigned int sysctl_numa_balancing_scan_size
= 256;
803 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
804 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
806 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
808 unsigned long rss
= 0;
809 unsigned long nr_scan_pages
;
812 * Calculations based on RSS as non-present and empty pages are skipped
813 * by the PTE scanner and NUMA hinting faults should be trapped based
816 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
817 rss
= get_mm_rss(p
->mm
);
821 rss
= round_up(rss
, nr_scan_pages
);
822 return rss
/ nr_scan_pages
;
825 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
826 #define MAX_SCAN_WINDOW 2560
828 static unsigned int task_scan_min(struct task_struct
*p
)
830 unsigned int scan
, floor
;
831 unsigned int windows
= 1;
833 if (sysctl_numa_balancing_scan_size
< MAX_SCAN_WINDOW
)
834 windows
= MAX_SCAN_WINDOW
/ sysctl_numa_balancing_scan_size
;
835 floor
= 1000 / windows
;
837 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
838 return max_t(unsigned int, floor
, scan
);
841 static unsigned int task_scan_max(struct task_struct
*p
)
843 unsigned int smin
= task_scan_min(p
);
846 /* Watch for min being lower than max due to floor calculations */
847 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
848 return max(smin
, smax
);
851 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
853 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
854 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
857 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
859 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
860 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
866 spinlock_t lock
; /* nr_tasks, tasks */
869 struct list_head task_list
;
872 nodemask_t active_nodes
;
873 unsigned long total_faults
;
875 * Faults_cpu is used to decide whether memory should move
876 * towards the CPU. As a consequence, these stats are weighted
877 * more by CPU use than by memory faults.
879 unsigned long *faults_cpu
;
880 unsigned long faults
[0];
883 /* Shared or private faults. */
884 #define NR_NUMA_HINT_FAULT_TYPES 2
886 /* Memory and CPU locality */
887 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
889 /* Averaged statistics, and temporary buffers. */
890 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
892 pid_t
task_numa_group_id(struct task_struct
*p
)
894 return p
->numa_group
? p
->numa_group
->gid
: 0;
897 static inline int task_faults_idx(int nid
, int priv
)
899 return NR_NUMA_HINT_FAULT_TYPES
* nid
+ priv
;
902 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
904 if (!p
->numa_faults_memory
)
907 return p
->numa_faults_memory
[task_faults_idx(nid
, 0)] +
908 p
->numa_faults_memory
[task_faults_idx(nid
, 1)];
911 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
916 return p
->numa_group
->faults
[task_faults_idx(nid
, 0)] +
917 p
->numa_group
->faults
[task_faults_idx(nid
, 1)];
920 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
922 return group
->faults_cpu
[task_faults_idx(nid
, 0)] +
923 group
->faults_cpu
[task_faults_idx(nid
, 1)];
927 * These return the fraction of accesses done by a particular task, or
928 * task group, on a particular numa node. The group weight is given a
929 * larger multiplier, in order to group tasks together that are almost
930 * evenly spread out between numa nodes.
932 static inline unsigned long task_weight(struct task_struct
*p
, int nid
)
934 unsigned long total_faults
;
936 if (!p
->numa_faults_memory
)
939 total_faults
= p
->total_numa_faults
;
944 return 1000 * task_faults(p
, nid
) / total_faults
;
947 static inline unsigned long group_weight(struct task_struct
*p
, int nid
)
949 if (!p
->numa_group
|| !p
->numa_group
->total_faults
)
952 return 1000 * group_faults(p
, nid
) / p
->numa_group
->total_faults
;
955 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
956 int src_nid
, int dst_cpu
)
958 struct numa_group
*ng
= p
->numa_group
;
959 int dst_nid
= cpu_to_node(dst_cpu
);
960 int last_cpupid
, this_cpupid
;
962 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
965 * Multi-stage node selection is used in conjunction with a periodic
966 * migration fault to build a temporal task<->page relation. By using
967 * a two-stage filter we remove short/unlikely relations.
969 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
970 * a task's usage of a particular page (n_p) per total usage of this
971 * page (n_t) (in a given time-span) to a probability.
973 * Our periodic faults will sample this probability and getting the
974 * same result twice in a row, given these samples are fully
975 * independent, is then given by P(n)^2, provided our sample period
976 * is sufficiently short compared to the usage pattern.
978 * This quadric squishes small probabilities, making it less likely we
979 * act on an unlikely task<->page relation.
981 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
982 if (!cpupid_pid_unset(last_cpupid
) &&
983 cpupid_to_nid(last_cpupid
) != dst_nid
)
986 /* Always allow migrate on private faults */
987 if (cpupid_match_pid(p
, last_cpupid
))
990 /* A shared fault, but p->numa_group has not been set up yet. */
995 * Do not migrate if the destination is not a node that
996 * is actively used by this numa group.
998 if (!node_isset(dst_nid
, ng
->active_nodes
))
1002 * Source is a node that is not actively used by this
1003 * numa group, while the destination is. Migrate.
1005 if (!node_isset(src_nid
, ng
->active_nodes
))
1009 * Both source and destination are nodes in active
1010 * use by this numa group. Maximize memory bandwidth
1011 * by migrating from more heavily used groups, to less
1012 * heavily used ones, spreading the load around.
1013 * Use a 1/4 hysteresis to avoid spurious page movement.
1015 return group_faults(p
, dst_nid
) < (group_faults(p
, src_nid
) * 3 / 4);
1018 static unsigned long weighted_cpuload(const int cpu
);
1019 static unsigned long source_load(int cpu
, int type
);
1020 static unsigned long target_load(int cpu
, int type
);
1021 static unsigned long capacity_of(int cpu
);
1022 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
1024 /* Cached statistics for all CPUs within a node */
1026 unsigned long nr_running
;
1029 /* Total compute capacity of CPUs on a node */
1030 unsigned long compute_capacity
;
1032 /* Approximate capacity in terms of runnable tasks on a node */
1033 unsigned long task_capacity
;
1034 int has_free_capacity
;
1038 * XXX borrowed from update_sg_lb_stats
1040 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1042 int smt
, cpu
, cpus
= 0;
1043 unsigned long capacity
;
1045 memset(ns
, 0, sizeof(*ns
));
1046 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1047 struct rq
*rq
= cpu_rq(cpu
);
1049 ns
->nr_running
+= rq
->nr_running
;
1050 ns
->load
+= weighted_cpuload(cpu
);
1051 ns
->compute_capacity
+= capacity_of(cpu
);
1057 * If we raced with hotplug and there are no CPUs left in our mask
1058 * the @ns structure is NULL'ed and task_numa_compare() will
1059 * not find this node attractive.
1061 * We'll either bail at !has_free_capacity, or we'll detect a huge
1062 * imbalance and bail there.
1067 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1068 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, ns
->compute_capacity
);
1069 capacity
= cpus
/ smt
; /* cores */
1071 ns
->task_capacity
= min_t(unsigned, capacity
,
1072 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
));
1073 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1076 struct task_numa_env
{
1077 struct task_struct
*p
;
1079 int src_cpu
, src_nid
;
1080 int dst_cpu
, dst_nid
;
1082 struct numa_stats src_stats
, dst_stats
;
1086 struct task_struct
*best_task
;
1091 static void task_numa_assign(struct task_numa_env
*env
,
1092 struct task_struct
*p
, long imp
)
1095 put_task_struct(env
->best_task
);
1100 env
->best_imp
= imp
;
1101 env
->best_cpu
= env
->dst_cpu
;
1104 static bool load_too_imbalanced(long src_load
, long dst_load
,
1105 struct task_numa_env
*env
)
1108 long orig_src_load
, orig_dst_load
;
1109 long src_capacity
, dst_capacity
;
1112 * The load is corrected for the CPU capacity available on each node.
1115 * ------------ vs ---------
1116 * src_capacity dst_capacity
1118 src_capacity
= env
->src_stats
.compute_capacity
;
1119 dst_capacity
= env
->dst_stats
.compute_capacity
;
1121 /* We care about the slope of the imbalance, not the direction. */
1122 if (dst_load
< src_load
)
1123 swap(dst_load
, src_load
);
1125 /* Is the difference below the threshold? */
1126 imb
= dst_load
* src_capacity
* 100 -
1127 src_load
* dst_capacity
* env
->imbalance_pct
;
1132 * The imbalance is above the allowed threshold.
1133 * Compare it with the old imbalance.
1135 orig_src_load
= env
->src_stats
.load
;
1136 orig_dst_load
= env
->dst_stats
.load
;
1138 if (orig_dst_load
< orig_src_load
)
1139 swap(orig_dst_load
, orig_src_load
);
1141 old_imb
= orig_dst_load
* src_capacity
* 100 -
1142 orig_src_load
* dst_capacity
* env
->imbalance_pct
;
1144 /* Would this change make things worse? */
1145 return (imb
> old_imb
);
1149 * This checks if the overall compute and NUMA accesses of the system would
1150 * be improved if the source tasks was migrated to the target dst_cpu taking
1151 * into account that it might be best if task running on the dst_cpu should
1152 * be exchanged with the source task
1154 static void task_numa_compare(struct task_numa_env
*env
,
1155 long taskimp
, long groupimp
)
1157 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1158 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1159 struct task_struct
*cur
;
1160 long src_load
, dst_load
;
1162 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1166 cur
= ACCESS_ONCE(dst_rq
->curr
);
1167 if (cur
->pid
== 0) /* idle */
1171 * "imp" is the fault differential for the source task between the
1172 * source and destination node. Calculate the total differential for
1173 * the source task and potential destination task. The more negative
1174 * the value is, the more rmeote accesses that would be expected to
1175 * be incurred if the tasks were swapped.
1178 /* Skip this swap candidate if cannot move to the source cpu */
1179 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1183 * If dst and source tasks are in the same NUMA group, or not
1184 * in any group then look only at task weights.
1186 if (cur
->numa_group
== env
->p
->numa_group
) {
1187 imp
= taskimp
+ task_weight(cur
, env
->src_nid
) -
1188 task_weight(cur
, env
->dst_nid
);
1190 * Add some hysteresis to prevent swapping the
1191 * tasks within a group over tiny differences.
1193 if (cur
->numa_group
)
1197 * Compare the group weights. If a task is all by
1198 * itself (not part of a group), use the task weight
1201 if (cur
->numa_group
)
1202 imp
+= group_weight(cur
, env
->src_nid
) -
1203 group_weight(cur
, env
->dst_nid
);
1205 imp
+= task_weight(cur
, env
->src_nid
) -
1206 task_weight(cur
, env
->dst_nid
);
1210 if (imp
<= env
->best_imp
&& moveimp
<= env
->best_imp
)
1214 /* Is there capacity at our destination? */
1215 if (env
->src_stats
.nr_running
<= env
->src_stats
.task_capacity
&&
1216 !env
->dst_stats
.has_free_capacity
)
1222 /* Balance doesn't matter much if we're running a task per cpu */
1223 if (imp
> env
->best_imp
&& src_rq
->nr_running
== 1 &&
1224 dst_rq
->nr_running
== 1)
1228 * In the overloaded case, try and keep the load balanced.
1231 load
= task_h_load(env
->p
);
1232 dst_load
= env
->dst_stats
.load
+ load
;
1233 src_load
= env
->src_stats
.load
- load
;
1235 if (moveimp
> imp
&& moveimp
> env
->best_imp
) {
1237 * If the improvement from just moving env->p direction is
1238 * better than swapping tasks around, check if a move is
1239 * possible. Store a slightly smaller score than moveimp,
1240 * so an actually idle CPU will win.
1242 if (!load_too_imbalanced(src_load
, dst_load
, env
)) {
1249 if (imp
<= env
->best_imp
)
1253 load
= task_h_load(cur
);
1258 if (load_too_imbalanced(src_load
, dst_load
, env
))
1262 * One idle CPU per node is evaluated for a task numa move.
1263 * Call select_idle_sibling to maybe find a better one.
1266 env
->dst_cpu
= select_idle_sibling(env
->p
, env
->dst_cpu
);
1269 task_numa_assign(env
, cur
, imp
);
1274 static void task_numa_find_cpu(struct task_numa_env
*env
,
1275 long taskimp
, long groupimp
)
1279 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1280 /* Skip this CPU if the source task cannot migrate */
1281 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1285 task_numa_compare(env
, taskimp
, groupimp
);
1289 static int task_numa_migrate(struct task_struct
*p
)
1291 struct task_numa_env env
= {
1294 .src_cpu
= task_cpu(p
),
1295 .src_nid
= task_node(p
),
1297 .imbalance_pct
= 112,
1303 struct sched_domain
*sd
;
1304 unsigned long taskweight
, groupweight
;
1306 long taskimp
, groupimp
;
1309 * Pick the lowest SD_NUMA domain, as that would have the smallest
1310 * imbalance and would be the first to start moving tasks about.
1312 * And we want to avoid any moving of tasks about, as that would create
1313 * random movement of tasks -- counter the numa conditions we're trying
1317 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1319 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1323 * Cpusets can break the scheduler domain tree into smaller
1324 * balance domains, some of which do not cross NUMA boundaries.
1325 * Tasks that are "trapped" in such domains cannot be migrated
1326 * elsewhere, so there is no point in (re)trying.
1328 if (unlikely(!sd
)) {
1329 p
->numa_preferred_nid
= task_node(p
);
1333 taskweight
= task_weight(p
, env
.src_nid
);
1334 groupweight
= group_weight(p
, env
.src_nid
);
1335 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1336 env
.dst_nid
= p
->numa_preferred_nid
;
1337 taskimp
= task_weight(p
, env
.dst_nid
) - taskweight
;
1338 groupimp
= group_weight(p
, env
.dst_nid
) - groupweight
;
1339 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1341 /* Try to find a spot on the preferred nid. */
1342 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1344 /* No space available on the preferred nid. Look elsewhere. */
1345 if (env
.best_cpu
== -1) {
1346 for_each_online_node(nid
) {
1347 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1350 /* Only consider nodes where both task and groups benefit */
1351 taskimp
= task_weight(p
, nid
) - taskweight
;
1352 groupimp
= group_weight(p
, nid
) - groupweight
;
1353 if (taskimp
< 0 && groupimp
< 0)
1357 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1358 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1363 * If the task is part of a workload that spans multiple NUMA nodes,
1364 * and is migrating into one of the workload's active nodes, remember
1365 * this node as the task's preferred numa node, so the workload can
1367 * A task that migrated to a second choice node will be better off
1368 * trying for a better one later. Do not set the preferred node here.
1370 if (p
->numa_group
) {
1371 if (env
.best_cpu
== -1)
1376 if (node_isset(nid
, p
->numa_group
->active_nodes
))
1377 sched_setnuma(p
, env
.dst_nid
);
1380 /* No better CPU than the current one was found. */
1381 if (env
.best_cpu
== -1)
1385 * Reset the scan period if the task is being rescheduled on an
1386 * alternative node to recheck if the tasks is now properly placed.
1388 p
->numa_scan_period
= task_scan_min(p
);
1390 if (env
.best_task
== NULL
) {
1391 ret
= migrate_task_to(p
, env
.best_cpu
);
1393 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1397 ret
= migrate_swap(p
, env
.best_task
);
1399 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1400 put_task_struct(env
.best_task
);
1404 /* Attempt to migrate a task to a CPU on the preferred node. */
1405 static void numa_migrate_preferred(struct task_struct
*p
)
1407 unsigned long interval
= HZ
;
1409 /* This task has no NUMA fault statistics yet */
1410 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults_memory
))
1413 /* Periodically retry migrating the task to the preferred node */
1414 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1415 p
->numa_migrate_retry
= jiffies
+ interval
;
1417 /* Success if task is already running on preferred CPU */
1418 if (task_node(p
) == p
->numa_preferred_nid
)
1421 /* Otherwise, try migrate to a CPU on the preferred node */
1422 task_numa_migrate(p
);
1426 * Find the nodes on which the workload is actively running. We do this by
1427 * tracking the nodes from which NUMA hinting faults are triggered. This can
1428 * be different from the set of nodes where the workload's memory is currently
1431 * The bitmask is used to make smarter decisions on when to do NUMA page
1432 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1433 * are added when they cause over 6/16 of the maximum number of faults, but
1434 * only removed when they drop below 3/16.
1436 static void update_numa_active_node_mask(struct numa_group
*numa_group
)
1438 unsigned long faults
, max_faults
= 0;
1441 for_each_online_node(nid
) {
1442 faults
= group_faults_cpu(numa_group
, nid
);
1443 if (faults
> max_faults
)
1444 max_faults
= faults
;
1447 for_each_online_node(nid
) {
1448 faults
= group_faults_cpu(numa_group
, nid
);
1449 if (!node_isset(nid
, numa_group
->active_nodes
)) {
1450 if (faults
> max_faults
* 6 / 16)
1451 node_set(nid
, numa_group
->active_nodes
);
1452 } else if (faults
< max_faults
* 3 / 16)
1453 node_clear(nid
, numa_group
->active_nodes
);
1458 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1459 * increments. The more local the fault statistics are, the higher the scan
1460 * period will be for the next scan window. If local/(local+remote) ratio is
1461 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1462 * the scan period will decrease. Aim for 70% local accesses.
1464 #define NUMA_PERIOD_SLOTS 10
1465 #define NUMA_PERIOD_THRESHOLD 7
1468 * Increase the scan period (slow down scanning) if the majority of
1469 * our memory is already on our local node, or if the majority of
1470 * the page accesses are shared with other processes.
1471 * Otherwise, decrease the scan period.
1473 static void update_task_scan_period(struct task_struct
*p
,
1474 unsigned long shared
, unsigned long private)
1476 unsigned int period_slot
;
1480 unsigned long remote
= p
->numa_faults_locality
[0];
1481 unsigned long local
= p
->numa_faults_locality
[1];
1484 * If there were no record hinting faults then either the task is
1485 * completely idle or all activity is areas that are not of interest
1486 * to automatic numa balancing. Scan slower
1488 if (local
+ shared
== 0) {
1489 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1490 p
->numa_scan_period
<< 1);
1492 p
->mm
->numa_next_scan
= jiffies
+
1493 msecs_to_jiffies(p
->numa_scan_period
);
1499 * Prepare to scale scan period relative to the current period.
1500 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1501 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1502 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1504 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1505 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1506 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1507 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1510 diff
= slot
* period_slot
;
1512 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1515 * Scale scan rate increases based on sharing. There is an
1516 * inverse relationship between the degree of sharing and
1517 * the adjustment made to the scanning period. Broadly
1518 * speaking the intent is that there is little point
1519 * scanning faster if shared accesses dominate as it may
1520 * simply bounce migrations uselessly
1522 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
));
1523 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1526 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1527 task_scan_min(p
), task_scan_max(p
));
1528 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1532 * Get the fraction of time the task has been running since the last
1533 * NUMA placement cycle. The scheduler keeps similar statistics, but
1534 * decays those on a 32ms period, which is orders of magnitude off
1535 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1536 * stats only if the task is so new there are no NUMA statistics yet.
1538 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1540 u64 runtime
, delta
, now
;
1541 /* Use the start of this time slice to avoid calculations. */
1542 now
= p
->se
.exec_start
;
1543 runtime
= p
->se
.sum_exec_runtime
;
1545 if (p
->last_task_numa_placement
) {
1546 delta
= runtime
- p
->last_sum_exec_runtime
;
1547 *period
= now
- p
->last_task_numa_placement
;
1549 delta
= p
->se
.avg
.runnable_avg_sum
;
1550 *period
= p
->se
.avg
.runnable_avg_period
;
1553 p
->last_sum_exec_runtime
= runtime
;
1554 p
->last_task_numa_placement
= now
;
1559 static void task_numa_placement(struct task_struct
*p
)
1561 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1562 unsigned long max_faults
= 0, max_group_faults
= 0;
1563 unsigned long fault_types
[2] = { 0, 0 };
1564 unsigned long total_faults
;
1565 u64 runtime
, period
;
1566 spinlock_t
*group_lock
= NULL
;
1568 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1569 if (p
->numa_scan_seq
== seq
)
1571 p
->numa_scan_seq
= seq
;
1572 p
->numa_scan_period_max
= task_scan_max(p
);
1574 total_faults
= p
->numa_faults_locality
[0] +
1575 p
->numa_faults_locality
[1];
1576 runtime
= numa_get_avg_runtime(p
, &period
);
1578 /* If the task is part of a group prevent parallel updates to group stats */
1579 if (p
->numa_group
) {
1580 group_lock
= &p
->numa_group
->lock
;
1581 spin_lock_irq(group_lock
);
1584 /* Find the node with the highest number of faults */
1585 for_each_online_node(nid
) {
1586 unsigned long faults
= 0, group_faults
= 0;
1589 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
1590 long diff
, f_diff
, f_weight
;
1592 i
= task_faults_idx(nid
, priv
);
1594 /* Decay existing window, copy faults since last scan */
1595 diff
= p
->numa_faults_buffer_memory
[i
] - p
->numa_faults_memory
[i
] / 2;
1596 fault_types
[priv
] += p
->numa_faults_buffer_memory
[i
];
1597 p
->numa_faults_buffer_memory
[i
] = 0;
1600 * Normalize the faults_from, so all tasks in a group
1601 * count according to CPU use, instead of by the raw
1602 * number of faults. Tasks with little runtime have
1603 * little over-all impact on throughput, and thus their
1604 * faults are less important.
1606 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
1607 f_weight
= (f_weight
* p
->numa_faults_buffer_cpu
[i
]) /
1609 f_diff
= f_weight
- p
->numa_faults_cpu
[i
] / 2;
1610 p
->numa_faults_buffer_cpu
[i
] = 0;
1612 p
->numa_faults_memory
[i
] += diff
;
1613 p
->numa_faults_cpu
[i
] += f_diff
;
1614 faults
+= p
->numa_faults_memory
[i
];
1615 p
->total_numa_faults
+= diff
;
1616 if (p
->numa_group
) {
1617 /* safe because we can only change our own group */
1618 p
->numa_group
->faults
[i
] += diff
;
1619 p
->numa_group
->faults_cpu
[i
] += f_diff
;
1620 p
->numa_group
->total_faults
+= diff
;
1621 group_faults
+= p
->numa_group
->faults
[i
];
1625 if (faults
> max_faults
) {
1626 max_faults
= faults
;
1630 if (group_faults
> max_group_faults
) {
1631 max_group_faults
= group_faults
;
1632 max_group_nid
= nid
;
1636 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1638 if (p
->numa_group
) {
1639 update_numa_active_node_mask(p
->numa_group
);
1640 spin_unlock_irq(group_lock
);
1641 max_nid
= max_group_nid
;
1645 /* Set the new preferred node */
1646 if (max_nid
!= p
->numa_preferred_nid
)
1647 sched_setnuma(p
, max_nid
);
1649 if (task_node(p
) != p
->numa_preferred_nid
)
1650 numa_migrate_preferred(p
);
1654 static inline int get_numa_group(struct numa_group
*grp
)
1656 return atomic_inc_not_zero(&grp
->refcount
);
1659 static inline void put_numa_group(struct numa_group
*grp
)
1661 if (atomic_dec_and_test(&grp
->refcount
))
1662 kfree_rcu(grp
, rcu
);
1665 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1668 struct numa_group
*grp
, *my_grp
;
1669 struct task_struct
*tsk
;
1671 int cpu
= cpupid_to_cpu(cpupid
);
1674 if (unlikely(!p
->numa_group
)) {
1675 unsigned int size
= sizeof(struct numa_group
) +
1676 4*nr_node_ids
*sizeof(unsigned long);
1678 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1682 atomic_set(&grp
->refcount
, 1);
1683 spin_lock_init(&grp
->lock
);
1684 INIT_LIST_HEAD(&grp
->task_list
);
1686 /* Second half of the array tracks nids where faults happen */
1687 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
1690 node_set(task_node(current
), grp
->active_nodes
);
1692 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1693 grp
->faults
[i
] = p
->numa_faults_memory
[i
];
1695 grp
->total_faults
= p
->total_numa_faults
;
1697 list_add(&p
->numa_entry
, &grp
->task_list
);
1699 rcu_assign_pointer(p
->numa_group
, grp
);
1703 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1705 if (!cpupid_match_pid(tsk
, cpupid
))
1708 grp
= rcu_dereference(tsk
->numa_group
);
1712 my_grp
= p
->numa_group
;
1717 * Only join the other group if its bigger; if we're the bigger group,
1718 * the other task will join us.
1720 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1724 * Tie-break on the grp address.
1726 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1729 /* Always join threads in the same process. */
1730 if (tsk
->mm
== current
->mm
)
1733 /* Simple filter to avoid false positives due to PID collisions */
1734 if (flags
& TNF_SHARED
)
1737 /* Update priv based on whether false sharing was detected */
1740 if (join
&& !get_numa_group(grp
))
1748 BUG_ON(irqs_disabled());
1749 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
1751 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
1752 my_grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1753 grp
->faults
[i
] += p
->numa_faults_memory
[i
];
1755 my_grp
->total_faults
-= p
->total_numa_faults
;
1756 grp
->total_faults
+= p
->total_numa_faults
;
1758 list_move(&p
->numa_entry
, &grp
->task_list
);
1762 spin_unlock(&my_grp
->lock
);
1763 spin_unlock_irq(&grp
->lock
);
1765 rcu_assign_pointer(p
->numa_group
, grp
);
1767 put_numa_group(my_grp
);
1775 void task_numa_free(struct task_struct
*p
)
1777 struct numa_group
*grp
= p
->numa_group
;
1778 void *numa_faults
= p
->numa_faults_memory
;
1779 unsigned long flags
;
1783 spin_lock_irqsave(&grp
->lock
, flags
);
1784 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1785 grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1786 grp
->total_faults
-= p
->total_numa_faults
;
1788 list_del(&p
->numa_entry
);
1790 spin_unlock_irqrestore(&grp
->lock
, flags
);
1791 RCU_INIT_POINTER(p
->numa_group
, NULL
);
1792 put_numa_group(grp
);
1795 p
->numa_faults_memory
= NULL
;
1796 p
->numa_faults_buffer_memory
= NULL
;
1797 p
->numa_faults_cpu
= NULL
;
1798 p
->numa_faults_buffer_cpu
= NULL
;
1803 * Got a PROT_NONE fault for a page on @node.
1805 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
1807 struct task_struct
*p
= current
;
1808 bool migrated
= flags
& TNF_MIGRATED
;
1809 int cpu_node
= task_node(current
);
1810 int local
= !!(flags
& TNF_FAULT_LOCAL
);
1813 if (!numabalancing_enabled
)
1816 /* for example, ksmd faulting in a user's mm */
1820 /* Allocate buffer to track faults on a per-node basis */
1821 if (unlikely(!p
->numa_faults_memory
)) {
1822 int size
= sizeof(*p
->numa_faults_memory
) *
1823 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
1825 p
->numa_faults_memory
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
1826 if (!p
->numa_faults_memory
)
1829 BUG_ON(p
->numa_faults_buffer_memory
);
1831 * The averaged statistics, shared & private, memory & cpu,
1832 * occupy the first half of the array. The second half of the
1833 * array is for current counters, which are averaged into the
1834 * first set by task_numa_placement.
1836 p
->numa_faults_cpu
= p
->numa_faults_memory
+ (2 * nr_node_ids
);
1837 p
->numa_faults_buffer_memory
= p
->numa_faults_memory
+ (4 * nr_node_ids
);
1838 p
->numa_faults_buffer_cpu
= p
->numa_faults_memory
+ (6 * nr_node_ids
);
1839 p
->total_numa_faults
= 0;
1840 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1844 * First accesses are treated as private, otherwise consider accesses
1845 * to be private if the accessing pid has not changed
1847 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
1850 priv
= cpupid_match_pid(p
, last_cpupid
);
1851 if (!priv
&& !(flags
& TNF_NO_GROUP
))
1852 task_numa_group(p
, last_cpupid
, flags
, &priv
);
1856 * If a workload spans multiple NUMA nodes, a shared fault that
1857 * occurs wholly within the set of nodes that the workload is
1858 * actively using should be counted as local. This allows the
1859 * scan rate to slow down when a workload has settled down.
1861 if (!priv
&& !local
&& p
->numa_group
&&
1862 node_isset(cpu_node
, p
->numa_group
->active_nodes
) &&
1863 node_isset(mem_node
, p
->numa_group
->active_nodes
))
1866 task_numa_placement(p
);
1869 * Retry task to preferred node migration periodically, in case it
1870 * case it previously failed, or the scheduler moved us.
1872 if (time_after(jiffies
, p
->numa_migrate_retry
))
1873 numa_migrate_preferred(p
);
1876 p
->numa_pages_migrated
+= pages
;
1878 p
->numa_faults_buffer_memory
[task_faults_idx(mem_node
, priv
)] += pages
;
1879 p
->numa_faults_buffer_cpu
[task_faults_idx(cpu_node
, priv
)] += pages
;
1880 p
->numa_faults_locality
[local
] += pages
;
1883 static void reset_ptenuma_scan(struct task_struct
*p
)
1885 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
1886 p
->mm
->numa_scan_offset
= 0;
1890 * The expensive part of numa migration is done from task_work context.
1891 * Triggered from task_tick_numa().
1893 void task_numa_work(struct callback_head
*work
)
1895 unsigned long migrate
, next_scan
, now
= jiffies
;
1896 struct task_struct
*p
= current
;
1897 struct mm_struct
*mm
= p
->mm
;
1898 struct vm_area_struct
*vma
;
1899 unsigned long start
, end
;
1900 unsigned long nr_pte_updates
= 0;
1903 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
1905 work
->next
= work
; /* protect against double add */
1907 * Who cares about NUMA placement when they're dying.
1909 * NOTE: make sure not to dereference p->mm before this check,
1910 * exit_task_work() happens _after_ exit_mm() so we could be called
1911 * without p->mm even though we still had it when we enqueued this
1914 if (p
->flags
& PF_EXITING
)
1917 if (!mm
->numa_next_scan
) {
1918 mm
->numa_next_scan
= now
+
1919 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1923 * Enforce maximal scan/migration frequency..
1925 migrate
= mm
->numa_next_scan
;
1926 if (time_before(now
, migrate
))
1929 if (p
->numa_scan_period
== 0) {
1930 p
->numa_scan_period_max
= task_scan_max(p
);
1931 p
->numa_scan_period
= task_scan_min(p
);
1934 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1935 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
1939 * Delay this task enough that another task of this mm will likely win
1940 * the next time around.
1942 p
->node_stamp
+= 2 * TICK_NSEC
;
1944 start
= mm
->numa_scan_offset
;
1945 pages
= sysctl_numa_balancing_scan_size
;
1946 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
1950 down_read(&mm
->mmap_sem
);
1951 vma
= find_vma(mm
, start
);
1953 reset_ptenuma_scan(p
);
1957 for (; vma
; vma
= vma
->vm_next
) {
1958 if (!vma_migratable(vma
) || !vma_policy_mof(p
, vma
))
1962 * Shared library pages mapped by multiple processes are not
1963 * migrated as it is expected they are cache replicated. Avoid
1964 * hinting faults in read-only file-backed mappings or the vdso
1965 * as migrating the pages will be of marginal benefit.
1968 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
1972 * Skip inaccessible VMAs to avoid any confusion between
1973 * PROT_NONE and NUMA hinting ptes
1975 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
1979 start
= max(start
, vma
->vm_start
);
1980 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
1981 end
= min(end
, vma
->vm_end
);
1982 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
1985 * Scan sysctl_numa_balancing_scan_size but ensure that
1986 * at least one PTE is updated so that unused virtual
1987 * address space is quickly skipped.
1990 pages
-= (end
- start
) >> PAGE_SHIFT
;
1997 } while (end
!= vma
->vm_end
);
2002 * It is possible to reach the end of the VMA list but the last few
2003 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2004 * would find the !migratable VMA on the next scan but not reset the
2005 * scanner to the start so check it now.
2008 mm
->numa_scan_offset
= start
;
2010 reset_ptenuma_scan(p
);
2011 up_read(&mm
->mmap_sem
);
2015 * Drive the periodic memory faults..
2017 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2019 struct callback_head
*work
= &curr
->numa_work
;
2023 * We don't care about NUMA placement if we don't have memory.
2025 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2029 * Using runtime rather than walltime has the dual advantage that
2030 * we (mostly) drive the selection from busy threads and that the
2031 * task needs to have done some actual work before we bother with
2034 now
= curr
->se
.sum_exec_runtime
;
2035 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2037 if (now
- curr
->node_stamp
> period
) {
2038 if (!curr
->node_stamp
)
2039 curr
->numa_scan_period
= task_scan_min(curr
);
2040 curr
->node_stamp
+= period
;
2042 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2043 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2044 task_work_add(curr
, work
, true);
2049 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2053 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2057 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2060 #endif /* CONFIG_NUMA_BALANCING */
2063 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2065 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2066 if (!parent_entity(se
))
2067 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2069 if (entity_is_task(se
)) {
2070 struct rq
*rq
= rq_of(cfs_rq
);
2072 account_numa_enqueue(rq
, task_of(se
));
2073 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2076 cfs_rq
->nr_running
++;
2080 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2082 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2083 if (!parent_entity(se
))
2084 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2085 if (entity_is_task(se
)) {
2086 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2087 list_del_init(&se
->group_node
);
2089 cfs_rq
->nr_running
--;
2092 #ifdef CONFIG_FAIR_GROUP_SCHED
2094 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
2099 * Use this CPU's actual weight instead of the last load_contribution
2100 * to gain a more accurate current total weight. See
2101 * update_cfs_rq_load_contribution().
2103 tg_weight
= atomic_long_read(&tg
->load_avg
);
2104 tg_weight
-= cfs_rq
->tg_load_contrib
;
2105 tg_weight
+= cfs_rq
->load
.weight
;
2110 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2112 long tg_weight
, load
, shares
;
2114 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
2115 load
= cfs_rq
->load
.weight
;
2117 shares
= (tg
->shares
* load
);
2119 shares
/= tg_weight
;
2121 if (shares
< MIN_SHARES
)
2122 shares
= MIN_SHARES
;
2123 if (shares
> tg
->shares
)
2124 shares
= tg
->shares
;
2128 # else /* CONFIG_SMP */
2129 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2133 # endif /* CONFIG_SMP */
2134 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2135 unsigned long weight
)
2138 /* commit outstanding execution time */
2139 if (cfs_rq
->curr
== se
)
2140 update_curr(cfs_rq
);
2141 account_entity_dequeue(cfs_rq
, se
);
2144 update_load_set(&se
->load
, weight
);
2147 account_entity_enqueue(cfs_rq
, se
);
2150 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2152 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2154 struct task_group
*tg
;
2155 struct sched_entity
*se
;
2159 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2160 if (!se
|| throttled_hierarchy(cfs_rq
))
2163 if (likely(se
->load
.weight
== tg
->shares
))
2166 shares
= calc_cfs_shares(cfs_rq
, tg
);
2168 reweight_entity(cfs_rq_of(se
), se
, shares
);
2170 #else /* CONFIG_FAIR_GROUP_SCHED */
2171 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2174 #endif /* CONFIG_FAIR_GROUP_SCHED */
2178 * We choose a half-life close to 1 scheduling period.
2179 * Note: The tables below are dependent on this value.
2181 #define LOAD_AVG_PERIOD 32
2182 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2183 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2185 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2186 static const u32 runnable_avg_yN_inv
[] = {
2187 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2188 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2189 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2190 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2191 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2192 0x85aac367, 0x82cd8698,
2196 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2197 * over-estimates when re-combining.
2199 static const u32 runnable_avg_yN_sum
[] = {
2200 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2201 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2202 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2207 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2209 static __always_inline u64
decay_load(u64 val
, u64 n
)
2211 unsigned int local_n
;
2215 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2218 /* after bounds checking we can collapse to 32-bit */
2222 * As y^PERIOD = 1/2, we can combine
2223 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2224 * With a look-up table which covers y^n (n<PERIOD)
2226 * To achieve constant time decay_load.
2228 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2229 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2230 local_n
%= LOAD_AVG_PERIOD
;
2233 val
*= runnable_avg_yN_inv
[local_n
];
2234 /* We don't use SRR here since we always want to round down. */
2239 * For updates fully spanning n periods, the contribution to runnable
2240 * average will be: \Sum 1024*y^n
2242 * We can compute this reasonably efficiently by combining:
2243 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2245 static u32
__compute_runnable_contrib(u64 n
)
2249 if (likely(n
<= LOAD_AVG_PERIOD
))
2250 return runnable_avg_yN_sum
[n
];
2251 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2252 return LOAD_AVG_MAX
;
2254 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2256 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2257 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2259 n
-= LOAD_AVG_PERIOD
;
2260 } while (n
> LOAD_AVG_PERIOD
);
2262 contrib
= decay_load(contrib
, n
);
2263 return contrib
+ runnable_avg_yN_sum
[n
];
2267 * We can represent the historical contribution to runnable average as the
2268 * coefficients of a geometric series. To do this we sub-divide our runnable
2269 * history into segments of approximately 1ms (1024us); label the segment that
2270 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2272 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2274 * (now) (~1ms ago) (~2ms ago)
2276 * Let u_i denote the fraction of p_i that the entity was runnable.
2278 * We then designate the fractions u_i as our co-efficients, yielding the
2279 * following representation of historical load:
2280 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2282 * We choose y based on the with of a reasonably scheduling period, fixing:
2285 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2286 * approximately half as much as the contribution to load within the last ms
2289 * When a period "rolls over" and we have new u_0`, multiplying the previous
2290 * sum again by y is sufficient to update:
2291 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2292 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2294 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2295 struct sched_avg
*sa
,
2299 u32 runnable_contrib
;
2300 int delta_w
, decayed
= 0;
2302 delta
= now
- sa
->last_runnable_update
;
2304 * This should only happen when time goes backwards, which it
2305 * unfortunately does during sched clock init when we swap over to TSC.
2307 if ((s64
)delta
< 0) {
2308 sa
->last_runnable_update
= now
;
2313 * Use 1024ns as the unit of measurement since it's a reasonable
2314 * approximation of 1us and fast to compute.
2319 sa
->last_runnable_update
= now
;
2321 /* delta_w is the amount already accumulated against our next period */
2322 delta_w
= sa
->runnable_avg_period
% 1024;
2323 if (delta
+ delta_w
>= 1024) {
2324 /* period roll-over */
2328 * Now that we know we're crossing a period boundary, figure
2329 * out how much from delta we need to complete the current
2330 * period and accrue it.
2332 delta_w
= 1024 - delta_w
;
2334 sa
->runnable_avg_sum
+= delta_w
;
2335 sa
->runnable_avg_period
+= delta_w
;
2339 /* Figure out how many additional periods this update spans */
2340 periods
= delta
/ 1024;
2343 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2345 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2348 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2349 runnable_contrib
= __compute_runnable_contrib(periods
);
2351 sa
->runnable_avg_sum
+= runnable_contrib
;
2352 sa
->runnable_avg_period
+= runnable_contrib
;
2355 /* Remainder of delta accrued against u_0` */
2357 sa
->runnable_avg_sum
+= delta
;
2358 sa
->runnable_avg_period
+= delta
;
2363 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2364 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2366 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2367 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2369 decays
-= se
->avg
.decay_count
;
2373 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2374 se
->avg
.decay_count
= 0;
2379 #ifdef CONFIG_FAIR_GROUP_SCHED
2380 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2383 struct task_group
*tg
= cfs_rq
->tg
;
2386 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2387 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2392 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2393 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2394 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2399 * Aggregate cfs_rq runnable averages into an equivalent task_group
2400 * representation for computing load contributions.
2402 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2403 struct cfs_rq
*cfs_rq
)
2405 struct task_group
*tg
= cfs_rq
->tg
;
2408 /* The fraction of a cpu used by this cfs_rq */
2409 contrib
= div_u64((u64
)sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2410 sa
->runnable_avg_period
+ 1);
2411 contrib
-= cfs_rq
->tg_runnable_contrib
;
2413 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2414 atomic_add(contrib
, &tg
->runnable_avg
);
2415 cfs_rq
->tg_runnable_contrib
+= contrib
;
2419 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2421 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2422 struct task_group
*tg
= cfs_rq
->tg
;
2427 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2428 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2429 atomic_long_read(&tg
->load_avg
) + 1);
2432 * For group entities we need to compute a correction term in the case
2433 * that they are consuming <1 cpu so that we would contribute the same
2434 * load as a task of equal weight.
2436 * Explicitly co-ordinating this measurement would be expensive, but
2437 * fortunately the sum of each cpus contribution forms a usable
2438 * lower-bound on the true value.
2440 * Consider the aggregate of 2 contributions. Either they are disjoint
2441 * (and the sum represents true value) or they are disjoint and we are
2442 * understating by the aggregate of their overlap.
2444 * Extending this to N cpus, for a given overlap, the maximum amount we
2445 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2446 * cpus that overlap for this interval and w_i is the interval width.
2448 * On a small machine; the first term is well-bounded which bounds the
2449 * total error since w_i is a subset of the period. Whereas on a
2450 * larger machine, while this first term can be larger, if w_i is the
2451 * of consequential size guaranteed to see n_i*w_i quickly converge to
2452 * our upper bound of 1-cpu.
2454 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2455 if (runnable_avg
< NICE_0_LOAD
) {
2456 se
->avg
.load_avg_contrib
*= runnable_avg
;
2457 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2461 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2463 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2464 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2466 #else /* CONFIG_FAIR_GROUP_SCHED */
2467 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2468 int force_update
) {}
2469 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2470 struct cfs_rq
*cfs_rq
) {}
2471 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2472 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2473 #endif /* CONFIG_FAIR_GROUP_SCHED */
2475 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2479 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2480 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2481 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2482 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2485 /* Compute the current contribution to load_avg by se, return any delta */
2486 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2488 long old_contrib
= se
->avg
.load_avg_contrib
;
2490 if (entity_is_task(se
)) {
2491 __update_task_entity_contrib(se
);
2493 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2494 __update_group_entity_contrib(se
);
2497 return se
->avg
.load_avg_contrib
- old_contrib
;
2500 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2503 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2504 cfs_rq
->blocked_load_avg
-= load_contrib
;
2506 cfs_rq
->blocked_load_avg
= 0;
2509 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2511 /* Update a sched_entity's runnable average */
2512 static inline void update_entity_load_avg(struct sched_entity
*se
,
2515 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2520 * For a group entity we need to use their owned cfs_rq_clock_task() in
2521 * case they are the parent of a throttled hierarchy.
2523 if (entity_is_task(se
))
2524 now
= cfs_rq_clock_task(cfs_rq
);
2526 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2528 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2531 contrib_delta
= __update_entity_load_avg_contrib(se
);
2537 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2539 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2543 * Decay the load contributed by all blocked children and account this so that
2544 * their contribution may appropriately discounted when they wake up.
2546 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2548 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2551 decays
= now
- cfs_rq
->last_decay
;
2552 if (!decays
&& !force_update
)
2555 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2556 unsigned long removed_load
;
2557 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2558 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2562 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2564 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2565 cfs_rq
->last_decay
= now
;
2568 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2571 /* Add the load generated by se into cfs_rq's child load-average */
2572 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2573 struct sched_entity
*se
,
2577 * We track migrations using entity decay_count <= 0, on a wake-up
2578 * migration we use a negative decay count to track the remote decays
2579 * accumulated while sleeping.
2581 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2582 * are seen by enqueue_entity_load_avg() as a migration with an already
2583 * constructed load_avg_contrib.
2585 if (unlikely(se
->avg
.decay_count
<= 0)) {
2586 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2587 if (se
->avg
.decay_count
) {
2589 * In a wake-up migration we have to approximate the
2590 * time sleeping. This is because we can't synchronize
2591 * clock_task between the two cpus, and it is not
2592 * guaranteed to be read-safe. Instead, we can
2593 * approximate this using our carried decays, which are
2594 * explicitly atomically readable.
2596 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2598 update_entity_load_avg(se
, 0);
2599 /* Indicate that we're now synchronized and on-rq */
2600 se
->avg
.decay_count
= 0;
2604 __synchronize_entity_decay(se
);
2607 /* migrated tasks did not contribute to our blocked load */
2609 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2610 update_entity_load_avg(se
, 0);
2613 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2614 /* we force update consideration on load-balancer moves */
2615 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2619 * Remove se's load from this cfs_rq child load-average, if the entity is
2620 * transitioning to a blocked state we track its projected decay using
2623 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2624 struct sched_entity
*se
,
2627 update_entity_load_avg(se
, 1);
2628 /* we force update consideration on load-balancer moves */
2629 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2631 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2633 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2634 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2635 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2639 * Update the rq's load with the elapsed running time before entering
2640 * idle. if the last scheduled task is not a CFS task, idle_enter will
2641 * be the only way to update the runnable statistic.
2643 void idle_enter_fair(struct rq
*this_rq
)
2645 update_rq_runnable_avg(this_rq
, 1);
2649 * Update the rq's load with the elapsed idle time before a task is
2650 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2651 * be the only way to update the runnable statistic.
2653 void idle_exit_fair(struct rq
*this_rq
)
2655 update_rq_runnable_avg(this_rq
, 0);
2658 static int idle_balance(struct rq
*this_rq
);
2660 #else /* CONFIG_SMP */
2662 static inline void update_entity_load_avg(struct sched_entity
*se
,
2663 int update_cfs_rq
) {}
2664 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2665 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2666 struct sched_entity
*se
,
2668 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2669 struct sched_entity
*se
,
2671 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2672 int force_update
) {}
2674 static inline int idle_balance(struct rq
*rq
)
2679 #endif /* CONFIG_SMP */
2681 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2683 #ifdef CONFIG_SCHEDSTATS
2684 struct task_struct
*tsk
= NULL
;
2686 if (entity_is_task(se
))
2689 if (se
->statistics
.sleep_start
) {
2690 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2695 if (unlikely(delta
> se
->statistics
.sleep_max
))
2696 se
->statistics
.sleep_max
= delta
;
2698 se
->statistics
.sleep_start
= 0;
2699 se
->statistics
.sum_sleep_runtime
+= delta
;
2702 account_scheduler_latency(tsk
, delta
>> 10, 1);
2703 trace_sched_stat_sleep(tsk
, delta
);
2706 if (se
->statistics
.block_start
) {
2707 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2712 if (unlikely(delta
> se
->statistics
.block_max
))
2713 se
->statistics
.block_max
= delta
;
2715 se
->statistics
.block_start
= 0;
2716 se
->statistics
.sum_sleep_runtime
+= delta
;
2719 if (tsk
->in_iowait
) {
2720 se
->statistics
.iowait_sum
+= delta
;
2721 se
->statistics
.iowait_count
++;
2722 trace_sched_stat_iowait(tsk
, delta
);
2725 trace_sched_stat_blocked(tsk
, delta
);
2728 * Blocking time is in units of nanosecs, so shift by
2729 * 20 to get a milliseconds-range estimation of the
2730 * amount of time that the task spent sleeping:
2732 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2733 profile_hits(SLEEP_PROFILING
,
2734 (void *)get_wchan(tsk
),
2737 account_scheduler_latency(tsk
, delta
>> 10, 0);
2743 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2745 #ifdef CONFIG_SCHED_DEBUG
2746 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2751 if (d
> 3*sysctl_sched_latency
)
2752 schedstat_inc(cfs_rq
, nr_spread_over
);
2757 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2759 u64 vruntime
= cfs_rq
->min_vruntime
;
2762 * The 'current' period is already promised to the current tasks,
2763 * however the extra weight of the new task will slow them down a
2764 * little, place the new task so that it fits in the slot that
2765 * stays open at the end.
2767 if (initial
&& sched_feat(START_DEBIT
))
2768 vruntime
+= sched_vslice(cfs_rq
, se
);
2770 /* sleeps up to a single latency don't count. */
2772 unsigned long thresh
= sysctl_sched_latency
;
2775 * Halve their sleep time's effect, to allow
2776 * for a gentler effect of sleepers:
2778 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2784 /* ensure we never gain time by being placed backwards. */
2785 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
2788 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
2791 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2794 * Update the normalized vruntime before updating min_vruntime
2795 * through calling update_curr().
2797 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
2798 se
->vruntime
+= cfs_rq
->min_vruntime
;
2801 * Update run-time statistics of the 'current'.
2803 update_curr(cfs_rq
);
2804 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
2805 account_entity_enqueue(cfs_rq
, se
);
2806 update_cfs_shares(cfs_rq
);
2808 if (flags
& ENQUEUE_WAKEUP
) {
2809 place_entity(cfs_rq
, se
, 0);
2810 enqueue_sleeper(cfs_rq
, se
);
2813 update_stats_enqueue(cfs_rq
, se
);
2814 check_spread(cfs_rq
, se
);
2815 if (se
!= cfs_rq
->curr
)
2816 __enqueue_entity(cfs_rq
, se
);
2819 if (cfs_rq
->nr_running
== 1) {
2820 list_add_leaf_cfs_rq(cfs_rq
);
2821 check_enqueue_throttle(cfs_rq
);
2825 static void __clear_buddies_last(struct sched_entity
*se
)
2827 for_each_sched_entity(se
) {
2828 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2829 if (cfs_rq
->last
!= se
)
2832 cfs_rq
->last
= NULL
;
2836 static void __clear_buddies_next(struct sched_entity
*se
)
2838 for_each_sched_entity(se
) {
2839 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2840 if (cfs_rq
->next
!= se
)
2843 cfs_rq
->next
= NULL
;
2847 static void __clear_buddies_skip(struct sched_entity
*se
)
2849 for_each_sched_entity(se
) {
2850 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2851 if (cfs_rq
->skip
!= se
)
2854 cfs_rq
->skip
= NULL
;
2858 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2860 if (cfs_rq
->last
== se
)
2861 __clear_buddies_last(se
);
2863 if (cfs_rq
->next
== se
)
2864 __clear_buddies_next(se
);
2866 if (cfs_rq
->skip
== se
)
2867 __clear_buddies_skip(se
);
2870 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2873 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2876 * Update run-time statistics of the 'current'.
2878 update_curr(cfs_rq
);
2879 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
2881 update_stats_dequeue(cfs_rq
, se
);
2882 if (flags
& DEQUEUE_SLEEP
) {
2883 #ifdef CONFIG_SCHEDSTATS
2884 if (entity_is_task(se
)) {
2885 struct task_struct
*tsk
= task_of(se
);
2887 if (tsk
->state
& TASK_INTERRUPTIBLE
)
2888 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
2889 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
2890 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
2895 clear_buddies(cfs_rq
, se
);
2897 if (se
!= cfs_rq
->curr
)
2898 __dequeue_entity(cfs_rq
, se
);
2900 account_entity_dequeue(cfs_rq
, se
);
2903 * Normalize the entity after updating the min_vruntime because the
2904 * update can refer to the ->curr item and we need to reflect this
2905 * movement in our normalized position.
2907 if (!(flags
& DEQUEUE_SLEEP
))
2908 se
->vruntime
-= cfs_rq
->min_vruntime
;
2910 /* return excess runtime on last dequeue */
2911 return_cfs_rq_runtime(cfs_rq
);
2913 update_min_vruntime(cfs_rq
);
2914 update_cfs_shares(cfs_rq
);
2918 * Preempt the current task with a newly woken task if needed:
2921 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2923 unsigned long ideal_runtime
, delta_exec
;
2924 struct sched_entity
*se
;
2927 ideal_runtime
= sched_slice(cfs_rq
, curr
);
2928 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
2929 if (delta_exec
> ideal_runtime
) {
2930 resched_curr(rq_of(cfs_rq
));
2932 * The current task ran long enough, ensure it doesn't get
2933 * re-elected due to buddy favours.
2935 clear_buddies(cfs_rq
, curr
);
2940 * Ensure that a task that missed wakeup preemption by a
2941 * narrow margin doesn't have to wait for a full slice.
2942 * This also mitigates buddy induced latencies under load.
2944 if (delta_exec
< sysctl_sched_min_granularity
)
2947 se
= __pick_first_entity(cfs_rq
);
2948 delta
= curr
->vruntime
- se
->vruntime
;
2953 if (delta
> ideal_runtime
)
2954 resched_curr(rq_of(cfs_rq
));
2958 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2960 /* 'current' is not kept within the tree. */
2963 * Any task has to be enqueued before it get to execute on
2964 * a CPU. So account for the time it spent waiting on the
2967 update_stats_wait_end(cfs_rq
, se
);
2968 __dequeue_entity(cfs_rq
, se
);
2971 update_stats_curr_start(cfs_rq
, se
);
2973 #ifdef CONFIG_SCHEDSTATS
2975 * Track our maximum slice length, if the CPU's load is at
2976 * least twice that of our own weight (i.e. dont track it
2977 * when there are only lesser-weight tasks around):
2979 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
2980 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
2981 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
2984 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
2988 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
2991 * Pick the next process, keeping these things in mind, in this order:
2992 * 1) keep things fair between processes/task groups
2993 * 2) pick the "next" process, since someone really wants that to run
2994 * 3) pick the "last" process, for cache locality
2995 * 4) do not run the "skip" process, if something else is available
2997 static struct sched_entity
*
2998 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3000 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
3001 struct sched_entity
*se
;
3004 * If curr is set we have to see if its left of the leftmost entity
3005 * still in the tree, provided there was anything in the tree at all.
3007 if (!left
|| (curr
&& entity_before(curr
, left
)))
3010 se
= left
; /* ideally we run the leftmost entity */
3013 * Avoid running the skip buddy, if running something else can
3014 * be done without getting too unfair.
3016 if (cfs_rq
->skip
== se
) {
3017 struct sched_entity
*second
;
3020 second
= __pick_first_entity(cfs_rq
);
3022 second
= __pick_next_entity(se
);
3023 if (!second
|| (curr
&& entity_before(curr
, second
)))
3027 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3032 * Prefer last buddy, try to return the CPU to a preempted task.
3034 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3038 * Someone really wants this to run. If it's not unfair, run it.
3040 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3043 clear_buddies(cfs_rq
, se
);
3048 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3050 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3053 * If still on the runqueue then deactivate_task()
3054 * was not called and update_curr() has to be done:
3057 update_curr(cfs_rq
);
3059 /* throttle cfs_rqs exceeding runtime */
3060 check_cfs_rq_runtime(cfs_rq
);
3062 check_spread(cfs_rq
, prev
);
3064 update_stats_wait_start(cfs_rq
, prev
);
3065 /* Put 'current' back into the tree. */
3066 __enqueue_entity(cfs_rq
, prev
);
3067 /* in !on_rq case, update occurred at dequeue */
3068 update_entity_load_avg(prev
, 1);
3070 cfs_rq
->curr
= NULL
;
3074 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3077 * Update run-time statistics of the 'current'.
3079 update_curr(cfs_rq
);
3082 * Ensure that runnable average is periodically updated.
3084 update_entity_load_avg(curr
, 1);
3085 update_cfs_rq_blocked_load(cfs_rq
, 1);
3086 update_cfs_shares(cfs_rq
);
3088 #ifdef CONFIG_SCHED_HRTICK
3090 * queued ticks are scheduled to match the slice, so don't bother
3091 * validating it and just reschedule.
3094 resched_curr(rq_of(cfs_rq
));
3098 * don't let the period tick interfere with the hrtick preemption
3100 if (!sched_feat(DOUBLE_TICK
) &&
3101 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3105 if (cfs_rq
->nr_running
> 1)
3106 check_preempt_tick(cfs_rq
, curr
);
3110 /**************************************************
3111 * CFS bandwidth control machinery
3114 #ifdef CONFIG_CFS_BANDWIDTH
3116 #ifdef HAVE_JUMP_LABEL
3117 static struct static_key __cfs_bandwidth_used
;
3119 static inline bool cfs_bandwidth_used(void)
3121 return static_key_false(&__cfs_bandwidth_used
);
3124 void cfs_bandwidth_usage_inc(void)
3126 static_key_slow_inc(&__cfs_bandwidth_used
);
3129 void cfs_bandwidth_usage_dec(void)
3131 static_key_slow_dec(&__cfs_bandwidth_used
);
3133 #else /* HAVE_JUMP_LABEL */
3134 static bool cfs_bandwidth_used(void)
3139 void cfs_bandwidth_usage_inc(void) {}
3140 void cfs_bandwidth_usage_dec(void) {}
3141 #endif /* HAVE_JUMP_LABEL */
3144 * default period for cfs group bandwidth.
3145 * default: 0.1s, units: nanoseconds
3147 static inline u64
default_cfs_period(void)
3149 return 100000000ULL;
3152 static inline u64
sched_cfs_bandwidth_slice(void)
3154 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3158 * Replenish runtime according to assigned quota and update expiration time.
3159 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3160 * additional synchronization around rq->lock.
3162 * requires cfs_b->lock
3164 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3168 if (cfs_b
->quota
== RUNTIME_INF
)
3171 now
= sched_clock_cpu(smp_processor_id());
3172 cfs_b
->runtime
= cfs_b
->quota
;
3173 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3176 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3178 return &tg
->cfs_bandwidth
;
3181 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3182 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3184 if (unlikely(cfs_rq
->throttle_count
))
3185 return cfs_rq
->throttled_clock_task
;
3187 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3190 /* returns 0 on failure to allocate runtime */
3191 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3193 struct task_group
*tg
= cfs_rq
->tg
;
3194 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3195 u64 amount
= 0, min_amount
, expires
;
3197 /* note: this is a positive sum as runtime_remaining <= 0 */
3198 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3200 raw_spin_lock(&cfs_b
->lock
);
3201 if (cfs_b
->quota
== RUNTIME_INF
)
3202 amount
= min_amount
;
3205 * If the bandwidth pool has become inactive, then at least one
3206 * period must have elapsed since the last consumption.
3207 * Refresh the global state and ensure bandwidth timer becomes
3210 if (!cfs_b
->timer_active
) {
3211 __refill_cfs_bandwidth_runtime(cfs_b
);
3212 __start_cfs_bandwidth(cfs_b
, false);
3215 if (cfs_b
->runtime
> 0) {
3216 amount
= min(cfs_b
->runtime
, min_amount
);
3217 cfs_b
->runtime
-= amount
;
3221 expires
= cfs_b
->runtime_expires
;
3222 raw_spin_unlock(&cfs_b
->lock
);
3224 cfs_rq
->runtime_remaining
+= amount
;
3226 * we may have advanced our local expiration to account for allowed
3227 * spread between our sched_clock and the one on which runtime was
3230 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3231 cfs_rq
->runtime_expires
= expires
;
3233 return cfs_rq
->runtime_remaining
> 0;
3237 * Note: This depends on the synchronization provided by sched_clock and the
3238 * fact that rq->clock snapshots this value.
3240 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3242 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3244 /* if the deadline is ahead of our clock, nothing to do */
3245 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3248 if (cfs_rq
->runtime_remaining
< 0)
3252 * If the local deadline has passed we have to consider the
3253 * possibility that our sched_clock is 'fast' and the global deadline
3254 * has not truly expired.
3256 * Fortunately we can check determine whether this the case by checking
3257 * whether the global deadline has advanced. It is valid to compare
3258 * cfs_b->runtime_expires without any locks since we only care about
3259 * exact equality, so a partial write will still work.
3262 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
3263 /* extend local deadline, drift is bounded above by 2 ticks */
3264 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3266 /* global deadline is ahead, expiration has passed */
3267 cfs_rq
->runtime_remaining
= 0;
3271 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3273 /* dock delta_exec before expiring quota (as it could span periods) */
3274 cfs_rq
->runtime_remaining
-= delta_exec
;
3275 expire_cfs_rq_runtime(cfs_rq
);
3277 if (likely(cfs_rq
->runtime_remaining
> 0))
3281 * if we're unable to extend our runtime we resched so that the active
3282 * hierarchy can be throttled
3284 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3285 resched_curr(rq_of(cfs_rq
));
3288 static __always_inline
3289 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3291 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3294 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3297 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3299 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3302 /* check whether cfs_rq, or any parent, is throttled */
3303 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3305 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3309 * Ensure that neither of the group entities corresponding to src_cpu or
3310 * dest_cpu are members of a throttled hierarchy when performing group
3311 * load-balance operations.
3313 static inline int throttled_lb_pair(struct task_group
*tg
,
3314 int src_cpu
, int dest_cpu
)
3316 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3318 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3319 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3321 return throttled_hierarchy(src_cfs_rq
) ||
3322 throttled_hierarchy(dest_cfs_rq
);
3325 /* updated child weight may affect parent so we have to do this bottom up */
3326 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3328 struct rq
*rq
= data
;
3329 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3331 cfs_rq
->throttle_count
--;
3333 if (!cfs_rq
->throttle_count
) {
3334 /* adjust cfs_rq_clock_task() */
3335 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3336 cfs_rq
->throttled_clock_task
;
3343 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3345 struct rq
*rq
= data
;
3346 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3348 /* group is entering throttled state, stop time */
3349 if (!cfs_rq
->throttle_count
)
3350 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3351 cfs_rq
->throttle_count
++;
3356 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3358 struct rq
*rq
= rq_of(cfs_rq
);
3359 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3360 struct sched_entity
*se
;
3361 long task_delta
, dequeue
= 1;
3363 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3365 /* freeze hierarchy runnable averages while throttled */
3367 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3370 task_delta
= cfs_rq
->h_nr_running
;
3371 for_each_sched_entity(se
) {
3372 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3373 /* throttled entity or throttle-on-deactivate */
3378 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3379 qcfs_rq
->h_nr_running
-= task_delta
;
3381 if (qcfs_rq
->load
.weight
)
3386 sub_nr_running(rq
, task_delta
);
3388 cfs_rq
->throttled
= 1;
3389 cfs_rq
->throttled_clock
= rq_clock(rq
);
3390 raw_spin_lock(&cfs_b
->lock
);
3392 * Add to the _head_ of the list, so that an already-started
3393 * distribute_cfs_runtime will not see us
3395 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3396 if (!cfs_b
->timer_active
)
3397 __start_cfs_bandwidth(cfs_b
, false);
3398 raw_spin_unlock(&cfs_b
->lock
);
3401 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3403 struct rq
*rq
= rq_of(cfs_rq
);
3404 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3405 struct sched_entity
*se
;
3409 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3411 cfs_rq
->throttled
= 0;
3413 update_rq_clock(rq
);
3415 raw_spin_lock(&cfs_b
->lock
);
3416 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3417 list_del_rcu(&cfs_rq
->throttled_list
);
3418 raw_spin_unlock(&cfs_b
->lock
);
3420 /* update hierarchical throttle state */
3421 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3423 if (!cfs_rq
->load
.weight
)
3426 task_delta
= cfs_rq
->h_nr_running
;
3427 for_each_sched_entity(se
) {
3431 cfs_rq
= cfs_rq_of(se
);
3433 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3434 cfs_rq
->h_nr_running
+= task_delta
;
3436 if (cfs_rq_throttled(cfs_rq
))
3441 add_nr_running(rq
, task_delta
);
3443 /* determine whether we need to wake up potentially idle cpu */
3444 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3448 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3449 u64 remaining
, u64 expires
)
3451 struct cfs_rq
*cfs_rq
;
3453 u64 starting_runtime
= remaining
;
3456 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3458 struct rq
*rq
= rq_of(cfs_rq
);
3460 raw_spin_lock(&rq
->lock
);
3461 if (!cfs_rq_throttled(cfs_rq
))
3464 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3465 if (runtime
> remaining
)
3466 runtime
= remaining
;
3467 remaining
-= runtime
;
3469 cfs_rq
->runtime_remaining
+= runtime
;
3470 cfs_rq
->runtime_expires
= expires
;
3472 /* we check whether we're throttled above */
3473 if (cfs_rq
->runtime_remaining
> 0)
3474 unthrottle_cfs_rq(cfs_rq
);
3477 raw_spin_unlock(&rq
->lock
);
3484 return starting_runtime
- remaining
;
3488 * Responsible for refilling a task_group's bandwidth and unthrottling its
3489 * cfs_rqs as appropriate. If there has been no activity within the last
3490 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3491 * used to track this state.
3493 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3495 u64 runtime
, runtime_expires
;
3498 /* no need to continue the timer with no bandwidth constraint */
3499 if (cfs_b
->quota
== RUNTIME_INF
)
3500 goto out_deactivate
;
3502 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3503 cfs_b
->nr_periods
+= overrun
;
3506 * idle depends on !throttled (for the case of a large deficit), and if
3507 * we're going inactive then everything else can be deferred
3509 if (cfs_b
->idle
&& !throttled
)
3510 goto out_deactivate
;
3513 * if we have relooped after returning idle once, we need to update our
3514 * status as actually running, so that other cpus doing
3515 * __start_cfs_bandwidth will stop trying to cancel us.
3517 cfs_b
->timer_active
= 1;
3519 __refill_cfs_bandwidth_runtime(cfs_b
);
3522 /* mark as potentially idle for the upcoming period */
3527 /* account preceding periods in which throttling occurred */
3528 cfs_b
->nr_throttled
+= overrun
;
3530 runtime_expires
= cfs_b
->runtime_expires
;
3533 * This check is repeated as we are holding onto the new bandwidth while
3534 * we unthrottle. This can potentially race with an unthrottled group
3535 * trying to acquire new bandwidth from the global pool. This can result
3536 * in us over-using our runtime if it is all used during this loop, but
3537 * only by limited amounts in that extreme case.
3539 while (throttled
&& cfs_b
->runtime
> 0) {
3540 runtime
= cfs_b
->runtime
;
3541 raw_spin_unlock(&cfs_b
->lock
);
3542 /* we can't nest cfs_b->lock while distributing bandwidth */
3543 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3545 raw_spin_lock(&cfs_b
->lock
);
3547 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3549 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3553 * While we are ensured activity in the period following an
3554 * unthrottle, this also covers the case in which the new bandwidth is
3555 * insufficient to cover the existing bandwidth deficit. (Forcing the
3556 * timer to remain active while there are any throttled entities.)
3563 cfs_b
->timer_active
= 0;
3567 /* a cfs_rq won't donate quota below this amount */
3568 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3569 /* minimum remaining period time to redistribute slack quota */
3570 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3571 /* how long we wait to gather additional slack before distributing */
3572 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3575 * Are we near the end of the current quota period?
3577 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3578 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3579 * migrate_hrtimers, base is never cleared, so we are fine.
3581 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3583 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3586 /* if the call-back is running a quota refresh is already occurring */
3587 if (hrtimer_callback_running(refresh_timer
))
3590 /* is a quota refresh about to occur? */
3591 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3592 if (remaining
< min_expire
)
3598 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3600 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3602 /* if there's a quota refresh soon don't bother with slack */
3603 if (runtime_refresh_within(cfs_b
, min_left
))
3606 start_bandwidth_timer(&cfs_b
->slack_timer
,
3607 ns_to_ktime(cfs_bandwidth_slack_period
));
3610 /* we know any runtime found here is valid as update_curr() precedes return */
3611 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3613 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3614 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3616 if (slack_runtime
<= 0)
3619 raw_spin_lock(&cfs_b
->lock
);
3620 if (cfs_b
->quota
!= RUNTIME_INF
&&
3621 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3622 cfs_b
->runtime
+= slack_runtime
;
3624 /* we are under rq->lock, defer unthrottling using a timer */
3625 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3626 !list_empty(&cfs_b
->throttled_cfs_rq
))
3627 start_cfs_slack_bandwidth(cfs_b
);
3629 raw_spin_unlock(&cfs_b
->lock
);
3631 /* even if it's not valid for return we don't want to try again */
3632 cfs_rq
->runtime_remaining
-= slack_runtime
;
3635 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3637 if (!cfs_bandwidth_used())
3640 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3643 __return_cfs_rq_runtime(cfs_rq
);
3647 * This is done with a timer (instead of inline with bandwidth return) since
3648 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3650 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3652 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3655 /* confirm we're still not at a refresh boundary */
3656 raw_spin_lock(&cfs_b
->lock
);
3657 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
3658 raw_spin_unlock(&cfs_b
->lock
);
3662 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
3663 runtime
= cfs_b
->runtime
;
3665 expires
= cfs_b
->runtime_expires
;
3666 raw_spin_unlock(&cfs_b
->lock
);
3671 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3673 raw_spin_lock(&cfs_b
->lock
);
3674 if (expires
== cfs_b
->runtime_expires
)
3675 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3676 raw_spin_unlock(&cfs_b
->lock
);
3680 * When a group wakes up we want to make sure that its quota is not already
3681 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3682 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3684 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3686 if (!cfs_bandwidth_used())
3689 /* an active group must be handled by the update_curr()->put() path */
3690 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3693 /* ensure the group is not already throttled */
3694 if (cfs_rq_throttled(cfs_rq
))
3697 /* update runtime allocation */
3698 account_cfs_rq_runtime(cfs_rq
, 0);
3699 if (cfs_rq
->runtime_remaining
<= 0)
3700 throttle_cfs_rq(cfs_rq
);
3703 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3704 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3706 if (!cfs_bandwidth_used())
3709 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3713 * it's possible for a throttled entity to be forced into a running
3714 * state (e.g. set_curr_task), in this case we're finished.
3716 if (cfs_rq_throttled(cfs_rq
))
3719 throttle_cfs_rq(cfs_rq
);
3723 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3725 struct cfs_bandwidth
*cfs_b
=
3726 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3727 do_sched_cfs_slack_timer(cfs_b
);
3729 return HRTIMER_NORESTART
;
3732 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3734 struct cfs_bandwidth
*cfs_b
=
3735 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3740 raw_spin_lock(&cfs_b
->lock
);
3742 now
= hrtimer_cb_get_time(timer
);
3743 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3748 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3750 raw_spin_unlock(&cfs_b
->lock
);
3752 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3755 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3757 raw_spin_lock_init(&cfs_b
->lock
);
3759 cfs_b
->quota
= RUNTIME_INF
;
3760 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3762 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3763 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3764 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3765 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3766 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3769 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3771 cfs_rq
->runtime_enabled
= 0;
3772 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3775 /* requires cfs_b->lock, may release to reprogram timer */
3776 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
, bool force
)
3779 * The timer may be active because we're trying to set a new bandwidth
3780 * period or because we're racing with the tear-down path
3781 * (timer_active==0 becomes visible before the hrtimer call-back
3782 * terminates). In either case we ensure that it's re-programmed
3784 while (unlikely(hrtimer_active(&cfs_b
->period_timer
)) &&
3785 hrtimer_try_to_cancel(&cfs_b
->period_timer
) < 0) {
3786 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3787 raw_spin_unlock(&cfs_b
->lock
);
3789 raw_spin_lock(&cfs_b
->lock
);
3790 /* if someone else restarted the timer then we're done */
3791 if (!force
&& cfs_b
->timer_active
)
3795 cfs_b
->timer_active
= 1;
3796 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
3799 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3801 hrtimer_cancel(&cfs_b
->period_timer
);
3802 hrtimer_cancel(&cfs_b
->slack_timer
);
3805 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
3807 struct cfs_rq
*cfs_rq
;
3809 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3810 struct cfs_bandwidth
*cfs_b
= &cfs_rq
->tg
->cfs_bandwidth
;
3812 raw_spin_lock(&cfs_b
->lock
);
3813 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
3814 raw_spin_unlock(&cfs_b
->lock
);
3818 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
3820 struct cfs_rq
*cfs_rq
;
3822 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3823 if (!cfs_rq
->runtime_enabled
)
3827 * clock_task is not advancing so we just need to make sure
3828 * there's some valid quota amount
3830 cfs_rq
->runtime_remaining
= 1;
3832 * Offline rq is schedulable till cpu is completely disabled
3833 * in take_cpu_down(), so we prevent new cfs throttling here.
3835 cfs_rq
->runtime_enabled
= 0;
3837 if (cfs_rq_throttled(cfs_rq
))
3838 unthrottle_cfs_rq(cfs_rq
);
3842 #else /* CONFIG_CFS_BANDWIDTH */
3843 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3845 return rq_clock_task(rq_of(cfs_rq
));
3848 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
3849 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
3850 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
3851 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3853 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3858 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3863 static inline int throttled_lb_pair(struct task_group
*tg
,
3864 int src_cpu
, int dest_cpu
)
3869 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3871 #ifdef CONFIG_FAIR_GROUP_SCHED
3872 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3875 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3879 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3880 static inline void update_runtime_enabled(struct rq
*rq
) {}
3881 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
3883 #endif /* CONFIG_CFS_BANDWIDTH */
3885 /**************************************************
3886 * CFS operations on tasks:
3889 #ifdef CONFIG_SCHED_HRTICK
3890 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3892 struct sched_entity
*se
= &p
->se
;
3893 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3895 WARN_ON(task_rq(p
) != rq
);
3897 if (cfs_rq
->nr_running
> 1) {
3898 u64 slice
= sched_slice(cfs_rq
, se
);
3899 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
3900 s64 delta
= slice
- ran
;
3907 hrtick_start(rq
, delta
);
3912 * called from enqueue/dequeue and updates the hrtick when the
3913 * current task is from our class and nr_running is low enough
3916 static void hrtick_update(struct rq
*rq
)
3918 struct task_struct
*curr
= rq
->curr
;
3920 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
3923 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
3924 hrtick_start_fair(rq
, curr
);
3926 #else /* !CONFIG_SCHED_HRTICK */
3928 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3932 static inline void hrtick_update(struct rq
*rq
)
3938 * The enqueue_task method is called before nr_running is
3939 * increased. Here we update the fair scheduling stats and
3940 * then put the task into the rbtree:
3943 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3945 struct cfs_rq
*cfs_rq
;
3946 struct sched_entity
*se
= &p
->se
;
3948 for_each_sched_entity(se
) {
3951 cfs_rq
= cfs_rq_of(se
);
3952 enqueue_entity(cfs_rq
, se
, flags
);
3955 * end evaluation on encountering a throttled cfs_rq
3957 * note: in the case of encountering a throttled cfs_rq we will
3958 * post the final h_nr_running increment below.
3960 if (cfs_rq_throttled(cfs_rq
))
3962 cfs_rq
->h_nr_running
++;
3964 flags
= ENQUEUE_WAKEUP
;
3967 for_each_sched_entity(se
) {
3968 cfs_rq
= cfs_rq_of(se
);
3969 cfs_rq
->h_nr_running
++;
3971 if (cfs_rq_throttled(cfs_rq
))
3974 update_cfs_shares(cfs_rq
);
3975 update_entity_load_avg(se
, 1);
3979 update_rq_runnable_avg(rq
, rq
->nr_running
);
3980 add_nr_running(rq
, 1);
3985 static void set_next_buddy(struct sched_entity
*se
);
3988 * The dequeue_task method is called before nr_running is
3989 * decreased. We remove the task from the rbtree and
3990 * update the fair scheduling stats:
3992 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3994 struct cfs_rq
*cfs_rq
;
3995 struct sched_entity
*se
= &p
->se
;
3996 int task_sleep
= flags
& DEQUEUE_SLEEP
;
3998 for_each_sched_entity(se
) {
3999 cfs_rq
= cfs_rq_of(se
);
4000 dequeue_entity(cfs_rq
, se
, flags
);
4003 * end evaluation on encountering a throttled cfs_rq
4005 * note: in the case of encountering a throttled cfs_rq we will
4006 * post the final h_nr_running decrement below.
4008 if (cfs_rq_throttled(cfs_rq
))
4010 cfs_rq
->h_nr_running
--;
4012 /* Don't dequeue parent if it has other entities besides us */
4013 if (cfs_rq
->load
.weight
) {
4015 * Bias pick_next to pick a task from this cfs_rq, as
4016 * p is sleeping when it is within its sched_slice.
4018 if (task_sleep
&& parent_entity(se
))
4019 set_next_buddy(parent_entity(se
));
4021 /* avoid re-evaluating load for this entity */
4022 se
= parent_entity(se
);
4025 flags
|= DEQUEUE_SLEEP
;
4028 for_each_sched_entity(se
) {
4029 cfs_rq
= cfs_rq_of(se
);
4030 cfs_rq
->h_nr_running
--;
4032 if (cfs_rq_throttled(cfs_rq
))
4035 update_cfs_shares(cfs_rq
);
4036 update_entity_load_avg(se
, 1);
4040 sub_nr_running(rq
, 1);
4041 update_rq_runnable_avg(rq
, 1);
4047 /* Used instead of source_load when we know the type == 0 */
4048 static unsigned long weighted_cpuload(const int cpu
)
4050 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
4054 * Return a low guess at the load of a migration-source cpu weighted
4055 * according to the scheduling class and "nice" value.
4057 * We want to under-estimate the load of migration sources, to
4058 * balance conservatively.
4060 static unsigned long source_load(int cpu
, int type
)
4062 struct rq
*rq
= cpu_rq(cpu
);
4063 unsigned long total
= weighted_cpuload(cpu
);
4065 if (type
== 0 || !sched_feat(LB_BIAS
))
4068 return min(rq
->cpu_load
[type
-1], total
);
4072 * Return a high guess at the load of a migration-target cpu weighted
4073 * according to the scheduling class and "nice" value.
4075 static unsigned long target_load(int cpu
, int type
)
4077 struct rq
*rq
= cpu_rq(cpu
);
4078 unsigned long total
= weighted_cpuload(cpu
);
4080 if (type
== 0 || !sched_feat(LB_BIAS
))
4083 return max(rq
->cpu_load
[type
-1], total
);
4086 static unsigned long capacity_of(int cpu
)
4088 return cpu_rq(cpu
)->cpu_capacity
;
4091 static unsigned long cpu_avg_load_per_task(int cpu
)
4093 struct rq
*rq
= cpu_rq(cpu
);
4094 unsigned long nr_running
= ACCESS_ONCE(rq
->cfs
.h_nr_running
);
4095 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
4098 return load_avg
/ nr_running
;
4103 static void record_wakee(struct task_struct
*p
)
4106 * Rough decay (wiping) for cost saving, don't worry
4107 * about the boundary, really active task won't care
4110 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
4111 current
->wakee_flips
>>= 1;
4112 current
->wakee_flip_decay_ts
= jiffies
;
4115 if (current
->last_wakee
!= p
) {
4116 current
->last_wakee
= p
;
4117 current
->wakee_flips
++;
4121 static void task_waking_fair(struct task_struct
*p
)
4123 struct sched_entity
*se
= &p
->se
;
4124 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4127 #ifndef CONFIG_64BIT
4128 u64 min_vruntime_copy
;
4131 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
4133 min_vruntime
= cfs_rq
->min_vruntime
;
4134 } while (min_vruntime
!= min_vruntime_copy
);
4136 min_vruntime
= cfs_rq
->min_vruntime
;
4139 se
->vruntime
-= min_vruntime
;
4143 #ifdef CONFIG_FAIR_GROUP_SCHED
4145 * effective_load() calculates the load change as seen from the root_task_group
4147 * Adding load to a group doesn't make a group heavier, but can cause movement
4148 * of group shares between cpus. Assuming the shares were perfectly aligned one
4149 * can calculate the shift in shares.
4151 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4152 * on this @cpu and results in a total addition (subtraction) of @wg to the
4153 * total group weight.
4155 * Given a runqueue weight distribution (rw_i) we can compute a shares
4156 * distribution (s_i) using:
4158 * s_i = rw_i / \Sum rw_j (1)
4160 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4161 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4162 * shares distribution (s_i):
4164 * rw_i = { 2, 4, 1, 0 }
4165 * s_i = { 2/7, 4/7, 1/7, 0 }
4167 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4168 * task used to run on and the CPU the waker is running on), we need to
4169 * compute the effect of waking a task on either CPU and, in case of a sync
4170 * wakeup, compute the effect of the current task going to sleep.
4172 * So for a change of @wl to the local @cpu with an overall group weight change
4173 * of @wl we can compute the new shares distribution (s'_i) using:
4175 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4177 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4178 * differences in waking a task to CPU 0. The additional task changes the
4179 * weight and shares distributions like:
4181 * rw'_i = { 3, 4, 1, 0 }
4182 * s'_i = { 3/8, 4/8, 1/8, 0 }
4184 * We can then compute the difference in effective weight by using:
4186 * dw_i = S * (s'_i - s_i) (3)
4188 * Where 'S' is the group weight as seen by its parent.
4190 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4191 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4192 * 4/7) times the weight of the group.
4194 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4196 struct sched_entity
*se
= tg
->se
[cpu
];
4198 if (!tg
->parent
) /* the trivial, non-cgroup case */
4201 for_each_sched_entity(se
) {
4207 * W = @wg + \Sum rw_j
4209 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4214 w
= se
->my_q
->load
.weight
+ wl
;
4217 * wl = S * s'_i; see (2)
4220 wl
= (w
* tg
->shares
) / W
;
4225 * Per the above, wl is the new se->load.weight value; since
4226 * those are clipped to [MIN_SHARES, ...) do so now. See
4227 * calc_cfs_shares().
4229 if (wl
< MIN_SHARES
)
4233 * wl = dw_i = S * (s'_i - s_i); see (3)
4235 wl
-= se
->load
.weight
;
4238 * Recursively apply this logic to all parent groups to compute
4239 * the final effective load change on the root group. Since
4240 * only the @tg group gets extra weight, all parent groups can
4241 * only redistribute existing shares. @wl is the shift in shares
4242 * resulting from this level per the above.
4251 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4258 static int wake_wide(struct task_struct
*p
)
4260 int factor
= this_cpu_read(sd_llc_size
);
4263 * Yeah, it's the switching-frequency, could means many wakee or
4264 * rapidly switch, use factor here will just help to automatically
4265 * adjust the loose-degree, so bigger node will lead to more pull.
4267 if (p
->wakee_flips
> factor
) {
4269 * wakee is somewhat hot, it needs certain amount of cpu
4270 * resource, so if waker is far more hot, prefer to leave
4273 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
4280 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4282 s64 this_load
, load
;
4283 s64 this_eff_load
, prev_eff_load
;
4284 int idx
, this_cpu
, prev_cpu
;
4285 struct task_group
*tg
;
4286 unsigned long weight
;
4290 * If we wake multiple tasks be careful to not bounce
4291 * ourselves around too much.
4297 this_cpu
= smp_processor_id();
4298 prev_cpu
= task_cpu(p
);
4299 load
= source_load(prev_cpu
, idx
);
4300 this_load
= target_load(this_cpu
, idx
);
4303 * If sync wakeup then subtract the (maximum possible)
4304 * effect of the currently running task from the load
4305 * of the current CPU:
4308 tg
= task_group(current
);
4309 weight
= current
->se
.load
.weight
;
4311 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4312 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4316 weight
= p
->se
.load
.weight
;
4319 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4320 * due to the sync cause above having dropped this_load to 0, we'll
4321 * always have an imbalance, but there's really nothing you can do
4322 * about that, so that's good too.
4324 * Otherwise check if either cpus are near enough in load to allow this
4325 * task to be woken on this_cpu.
4327 this_eff_load
= 100;
4328 this_eff_load
*= capacity_of(prev_cpu
);
4330 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4331 prev_eff_load
*= capacity_of(this_cpu
);
4333 if (this_load
> 0) {
4334 this_eff_load
*= this_load
+
4335 effective_load(tg
, this_cpu
, weight
, weight
);
4337 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4340 balanced
= this_eff_load
<= prev_eff_load
;
4342 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4347 schedstat_inc(sd
, ttwu_move_affine
);
4348 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4354 * find_idlest_group finds and returns the least busy CPU group within the
4357 static struct sched_group
*
4358 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4359 int this_cpu
, int sd_flag
)
4361 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4362 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4363 int load_idx
= sd
->forkexec_idx
;
4364 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4366 if (sd_flag
& SD_BALANCE_WAKE
)
4367 load_idx
= sd
->wake_idx
;
4370 unsigned long load
, avg_load
;
4374 /* Skip over this group if it has no CPUs allowed */
4375 if (!cpumask_intersects(sched_group_cpus(group
),
4376 tsk_cpus_allowed(p
)))
4379 local_group
= cpumask_test_cpu(this_cpu
,
4380 sched_group_cpus(group
));
4382 /* Tally up the load of all CPUs in the group */
4385 for_each_cpu(i
, sched_group_cpus(group
)) {
4386 /* Bias balancing toward cpus of our domain */
4388 load
= source_load(i
, load_idx
);
4390 load
= target_load(i
, load_idx
);
4395 /* Adjust by relative CPU capacity of the group */
4396 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) / group
->sgc
->capacity
;
4399 this_load
= avg_load
;
4400 } else if (avg_load
< min_load
) {
4401 min_load
= avg_load
;
4404 } while (group
= group
->next
, group
!= sd
->groups
);
4406 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4412 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4415 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4417 unsigned long load
, min_load
= ULONG_MAX
;
4421 /* Traverse only the allowed CPUs */
4422 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4423 load
= weighted_cpuload(i
);
4425 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4435 * Try and locate an idle CPU in the sched_domain.
4437 static int select_idle_sibling(struct task_struct
*p
, int target
)
4439 struct sched_domain
*sd
;
4440 struct sched_group
*sg
;
4441 int i
= task_cpu(p
);
4443 if (idle_cpu(target
))
4447 * If the prevous cpu is cache affine and idle, don't be stupid.
4449 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4453 * Otherwise, iterate the domains and find an elegible idle cpu.
4455 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4456 for_each_lower_domain(sd
) {
4459 if (!cpumask_intersects(sched_group_cpus(sg
),
4460 tsk_cpus_allowed(p
)))
4463 for_each_cpu(i
, sched_group_cpus(sg
)) {
4464 if (i
== target
|| !idle_cpu(i
))
4468 target
= cpumask_first_and(sched_group_cpus(sg
),
4469 tsk_cpus_allowed(p
));
4473 } while (sg
!= sd
->groups
);
4480 * select_task_rq_fair: Select target runqueue for the waking task in domains
4481 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4482 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4484 * Balances load by selecting the idlest cpu in the idlest group, or under
4485 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4487 * Returns the target cpu number.
4489 * preempt must be disabled.
4492 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4494 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4495 int cpu
= smp_processor_id();
4497 int want_affine
= 0;
4498 int sync
= wake_flags
& WF_SYNC
;
4500 if (p
->nr_cpus_allowed
== 1)
4503 if (sd_flag
& SD_BALANCE_WAKE
)
4504 want_affine
= cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
));
4507 for_each_domain(cpu
, tmp
) {
4508 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4512 * If both cpu and prev_cpu are part of this domain,
4513 * cpu is a valid SD_WAKE_AFFINE target.
4515 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4516 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4521 if (tmp
->flags
& sd_flag
)
4525 if (affine_sd
&& cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4528 if (sd_flag
& SD_BALANCE_WAKE
) {
4529 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4534 struct sched_group
*group
;
4537 if (!(sd
->flags
& sd_flag
)) {
4542 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
4548 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4549 if (new_cpu
== -1 || new_cpu
== cpu
) {
4550 /* Now try balancing at a lower domain level of cpu */
4555 /* Now try balancing at a lower domain level of new_cpu */
4557 weight
= sd
->span_weight
;
4559 for_each_domain(cpu
, tmp
) {
4560 if (weight
<= tmp
->span_weight
)
4562 if (tmp
->flags
& sd_flag
)
4565 /* while loop will break here if sd == NULL */
4574 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4575 * cfs_rq_of(p) references at time of call are still valid and identify the
4576 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4577 * other assumptions, including the state of rq->lock, should be made.
4580 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4582 struct sched_entity
*se
= &p
->se
;
4583 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4586 * Load tracking: accumulate removed load so that it can be processed
4587 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4588 * to blocked load iff they have a positive decay-count. It can never
4589 * be negative here since on-rq tasks have decay-count == 0.
4591 if (se
->avg
.decay_count
) {
4592 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4593 atomic_long_add(se
->avg
.load_avg_contrib
,
4594 &cfs_rq
->removed_load
);
4597 /* We have migrated, no longer consider this task hot */
4600 #endif /* CONFIG_SMP */
4602 static unsigned long
4603 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4605 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4608 * Since its curr running now, convert the gran from real-time
4609 * to virtual-time in his units.
4611 * By using 'se' instead of 'curr' we penalize light tasks, so
4612 * they get preempted easier. That is, if 'se' < 'curr' then
4613 * the resulting gran will be larger, therefore penalizing the
4614 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4615 * be smaller, again penalizing the lighter task.
4617 * This is especially important for buddies when the leftmost
4618 * task is higher priority than the buddy.
4620 return calc_delta_fair(gran
, se
);
4624 * Should 'se' preempt 'curr'.
4638 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4640 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4645 gran
= wakeup_gran(curr
, se
);
4652 static void set_last_buddy(struct sched_entity
*se
)
4654 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4657 for_each_sched_entity(se
)
4658 cfs_rq_of(se
)->last
= se
;
4661 static void set_next_buddy(struct sched_entity
*se
)
4663 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4666 for_each_sched_entity(se
)
4667 cfs_rq_of(se
)->next
= se
;
4670 static void set_skip_buddy(struct sched_entity
*se
)
4672 for_each_sched_entity(se
)
4673 cfs_rq_of(se
)->skip
= se
;
4677 * Preempt the current task with a newly woken task if needed:
4679 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4681 struct task_struct
*curr
= rq
->curr
;
4682 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4683 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4684 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4685 int next_buddy_marked
= 0;
4687 if (unlikely(se
== pse
))
4691 * This is possible from callers such as attach_tasks(), in which we
4692 * unconditionally check_prempt_curr() after an enqueue (which may have
4693 * lead to a throttle). This both saves work and prevents false
4694 * next-buddy nomination below.
4696 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4699 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4700 set_next_buddy(pse
);
4701 next_buddy_marked
= 1;
4705 * We can come here with TIF_NEED_RESCHED already set from new task
4708 * Note: this also catches the edge-case of curr being in a throttled
4709 * group (e.g. via set_curr_task), since update_curr() (in the
4710 * enqueue of curr) will have resulted in resched being set. This
4711 * prevents us from potentially nominating it as a false LAST_BUDDY
4714 if (test_tsk_need_resched(curr
))
4717 /* Idle tasks are by definition preempted by non-idle tasks. */
4718 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4719 likely(p
->policy
!= SCHED_IDLE
))
4723 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4724 * is driven by the tick):
4726 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4729 find_matching_se(&se
, &pse
);
4730 update_curr(cfs_rq_of(se
));
4732 if (wakeup_preempt_entity(se
, pse
) == 1) {
4734 * Bias pick_next to pick the sched entity that is
4735 * triggering this preemption.
4737 if (!next_buddy_marked
)
4738 set_next_buddy(pse
);
4747 * Only set the backward buddy when the current task is still
4748 * on the rq. This can happen when a wakeup gets interleaved
4749 * with schedule on the ->pre_schedule() or idle_balance()
4750 * point, either of which can * drop the rq lock.
4752 * Also, during early boot the idle thread is in the fair class,
4753 * for obvious reasons its a bad idea to schedule back to it.
4755 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4758 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
4762 static struct task_struct
*
4763 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4765 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
4766 struct sched_entity
*se
;
4767 struct task_struct
*p
;
4771 #ifdef CONFIG_FAIR_GROUP_SCHED
4772 if (!cfs_rq
->nr_running
)
4775 if (prev
->sched_class
!= &fair_sched_class
)
4779 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4780 * likely that a next task is from the same cgroup as the current.
4782 * Therefore attempt to avoid putting and setting the entire cgroup
4783 * hierarchy, only change the part that actually changes.
4787 struct sched_entity
*curr
= cfs_rq
->curr
;
4790 * Since we got here without doing put_prev_entity() we also
4791 * have to consider cfs_rq->curr. If it is still a runnable
4792 * entity, update_curr() will update its vruntime, otherwise
4793 * forget we've ever seen it.
4795 if (curr
&& curr
->on_rq
)
4796 update_curr(cfs_rq
);
4801 * This call to check_cfs_rq_runtime() will do the throttle and
4802 * dequeue its entity in the parent(s). Therefore the 'simple'
4803 * nr_running test will indeed be correct.
4805 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
4808 se
= pick_next_entity(cfs_rq
, curr
);
4809 cfs_rq
= group_cfs_rq(se
);
4815 * Since we haven't yet done put_prev_entity and if the selected task
4816 * is a different task than we started out with, try and touch the
4817 * least amount of cfs_rqs.
4820 struct sched_entity
*pse
= &prev
->se
;
4822 while (!(cfs_rq
= is_same_group(se
, pse
))) {
4823 int se_depth
= se
->depth
;
4824 int pse_depth
= pse
->depth
;
4826 if (se_depth
<= pse_depth
) {
4827 put_prev_entity(cfs_rq_of(pse
), pse
);
4828 pse
= parent_entity(pse
);
4830 if (se_depth
>= pse_depth
) {
4831 set_next_entity(cfs_rq_of(se
), se
);
4832 se
= parent_entity(se
);
4836 put_prev_entity(cfs_rq
, pse
);
4837 set_next_entity(cfs_rq
, se
);
4840 if (hrtick_enabled(rq
))
4841 hrtick_start_fair(rq
, p
);
4848 if (!cfs_rq
->nr_running
)
4851 put_prev_task(rq
, prev
);
4854 se
= pick_next_entity(cfs_rq
, NULL
);
4855 set_next_entity(cfs_rq
, se
);
4856 cfs_rq
= group_cfs_rq(se
);
4861 if (hrtick_enabled(rq
))
4862 hrtick_start_fair(rq
, p
);
4867 new_tasks
= idle_balance(rq
);
4869 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4870 * possible for any higher priority task to appear. In that case we
4871 * must re-start the pick_next_entity() loop.
4883 * Account for a descheduled task:
4885 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4887 struct sched_entity
*se
= &prev
->se
;
4888 struct cfs_rq
*cfs_rq
;
4890 for_each_sched_entity(se
) {
4891 cfs_rq
= cfs_rq_of(se
);
4892 put_prev_entity(cfs_rq
, se
);
4897 * sched_yield() is very simple
4899 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4901 static void yield_task_fair(struct rq
*rq
)
4903 struct task_struct
*curr
= rq
->curr
;
4904 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4905 struct sched_entity
*se
= &curr
->se
;
4908 * Are we the only task in the tree?
4910 if (unlikely(rq
->nr_running
== 1))
4913 clear_buddies(cfs_rq
, se
);
4915 if (curr
->policy
!= SCHED_BATCH
) {
4916 update_rq_clock(rq
);
4918 * Update run-time statistics of the 'current'.
4920 update_curr(cfs_rq
);
4922 * Tell update_rq_clock() that we've just updated,
4923 * so we don't do microscopic update in schedule()
4924 * and double the fastpath cost.
4926 rq
->skip_clock_update
= 1;
4932 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
4934 struct sched_entity
*se
= &p
->se
;
4936 /* throttled hierarchies are not runnable */
4937 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
4940 /* Tell the scheduler that we'd really like pse to run next. */
4943 yield_task_fair(rq
);
4949 /**************************************************
4950 * Fair scheduling class load-balancing methods.
4954 * The purpose of load-balancing is to achieve the same basic fairness the
4955 * per-cpu scheduler provides, namely provide a proportional amount of compute
4956 * time to each task. This is expressed in the following equation:
4958 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4960 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4961 * W_i,0 is defined as:
4963 * W_i,0 = \Sum_j w_i,j (2)
4965 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4966 * is derived from the nice value as per prio_to_weight[].
4968 * The weight average is an exponential decay average of the instantaneous
4971 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4973 * C_i is the compute capacity of cpu i, typically it is the
4974 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4975 * can also include other factors [XXX].
4977 * To achieve this balance we define a measure of imbalance which follows
4978 * directly from (1):
4980 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
4982 * We them move tasks around to minimize the imbalance. In the continuous
4983 * function space it is obvious this converges, in the discrete case we get
4984 * a few fun cases generally called infeasible weight scenarios.
4987 * - infeasible weights;
4988 * - local vs global optima in the discrete case. ]
4993 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4994 * for all i,j solution, we create a tree of cpus that follows the hardware
4995 * topology where each level pairs two lower groups (or better). This results
4996 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4997 * tree to only the first of the previous level and we decrease the frequency
4998 * of load-balance at each level inv. proportional to the number of cpus in
5004 * \Sum { --- * --- * 2^i } = O(n) (5)
5006 * `- size of each group
5007 * | | `- number of cpus doing load-balance
5009 * `- sum over all levels
5011 * Coupled with a limit on how many tasks we can migrate every balance pass,
5012 * this makes (5) the runtime complexity of the balancer.
5014 * An important property here is that each CPU is still (indirectly) connected
5015 * to every other cpu in at most O(log n) steps:
5017 * The adjacency matrix of the resulting graph is given by:
5020 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5023 * And you'll find that:
5025 * A^(log_2 n)_i,j != 0 for all i,j (7)
5027 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5028 * The task movement gives a factor of O(m), giving a convergence complexity
5031 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5036 * In order to avoid CPUs going idle while there's still work to do, new idle
5037 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5038 * tree itself instead of relying on other CPUs to bring it work.
5040 * This adds some complexity to both (5) and (8) but it reduces the total idle
5048 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5051 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5056 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5058 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5060 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5063 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5064 * rewrite all of this once again.]
5067 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
5069 enum fbq_type
{ regular
, remote
, all
};
5071 #define LBF_ALL_PINNED 0x01
5072 #define LBF_NEED_BREAK 0x02
5073 #define LBF_DST_PINNED 0x04
5074 #define LBF_SOME_PINNED 0x08
5077 struct sched_domain
*sd
;
5085 struct cpumask
*dst_grpmask
;
5087 enum cpu_idle_type idle
;
5089 /* The set of CPUs under consideration for load-balancing */
5090 struct cpumask
*cpus
;
5095 unsigned int loop_break
;
5096 unsigned int loop_max
;
5098 enum fbq_type fbq_type
;
5099 struct list_head tasks
;
5103 * Is this task likely cache-hot:
5105 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
5109 lockdep_assert_held(&env
->src_rq
->lock
);
5111 if (p
->sched_class
!= &fair_sched_class
)
5114 if (unlikely(p
->policy
== SCHED_IDLE
))
5118 * Buddy candidates are cache hot:
5120 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
5121 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
5122 &p
->se
== cfs_rq_of(&p
->se
)->last
))
5125 if (sysctl_sched_migration_cost
== -1)
5127 if (sysctl_sched_migration_cost
== 0)
5130 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
5132 return delta
< (s64
)sysctl_sched_migration_cost
;
5135 #ifdef CONFIG_NUMA_BALANCING
5136 /* Returns true if the destination node has incurred more faults */
5137 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
5139 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5140 int src_nid
, dst_nid
;
5142 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults_memory
||
5143 !(env
->sd
->flags
& SD_NUMA
)) {
5147 src_nid
= cpu_to_node(env
->src_cpu
);
5148 dst_nid
= cpu_to_node(env
->dst_cpu
);
5150 if (src_nid
== dst_nid
)
5154 /* Task is already in the group's interleave set. */
5155 if (node_isset(src_nid
, numa_group
->active_nodes
))
5158 /* Task is moving into the group's interleave set. */
5159 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5162 return group_faults(p
, dst_nid
) > group_faults(p
, src_nid
);
5165 /* Encourage migration to the preferred node. */
5166 if (dst_nid
== p
->numa_preferred_nid
)
5169 return task_faults(p
, dst_nid
) > task_faults(p
, src_nid
);
5173 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
5175 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5176 int src_nid
, dst_nid
;
5178 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
5181 if (!p
->numa_faults_memory
|| !(env
->sd
->flags
& SD_NUMA
))
5184 src_nid
= cpu_to_node(env
->src_cpu
);
5185 dst_nid
= cpu_to_node(env
->dst_cpu
);
5187 if (src_nid
== dst_nid
)
5191 /* Task is moving within/into the group's interleave set. */
5192 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5195 /* Task is moving out of the group's interleave set. */
5196 if (node_isset(src_nid
, numa_group
->active_nodes
))
5199 return group_faults(p
, dst_nid
) < group_faults(p
, src_nid
);
5202 /* Migrating away from the preferred node is always bad. */
5203 if (src_nid
== p
->numa_preferred_nid
)
5206 return task_faults(p
, dst_nid
) < task_faults(p
, src_nid
);
5210 static inline bool migrate_improves_locality(struct task_struct
*p
,
5216 static inline bool migrate_degrades_locality(struct task_struct
*p
,
5224 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5227 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
5229 int tsk_cache_hot
= 0;
5231 lockdep_assert_held(&env
->src_rq
->lock
);
5234 * We do not migrate tasks that are:
5235 * 1) throttled_lb_pair, or
5236 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5237 * 3) running (obviously), or
5238 * 4) are cache-hot on their current CPU.
5240 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
5243 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
5246 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
5248 env
->flags
|= LBF_SOME_PINNED
;
5251 * Remember if this task can be migrated to any other cpu in
5252 * our sched_group. We may want to revisit it if we couldn't
5253 * meet load balance goals by pulling other tasks on src_cpu.
5255 * Also avoid computing new_dst_cpu if we have already computed
5256 * one in current iteration.
5258 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
5261 /* Prevent to re-select dst_cpu via env's cpus */
5262 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
5263 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
5264 env
->flags
|= LBF_DST_PINNED
;
5265 env
->new_dst_cpu
= cpu
;
5273 /* Record that we found atleast one task that could run on dst_cpu */
5274 env
->flags
&= ~LBF_ALL_PINNED
;
5276 if (task_running(env
->src_rq
, p
)) {
5277 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
5282 * Aggressive migration if:
5283 * 1) destination numa is preferred
5284 * 2) task is cache cold, or
5285 * 3) too many balance attempts have failed.
5287 tsk_cache_hot
= task_hot(p
, env
);
5289 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
5291 if (migrate_improves_locality(p
, env
)) {
5292 #ifdef CONFIG_SCHEDSTATS
5293 if (tsk_cache_hot
) {
5294 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5295 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5301 if (!tsk_cache_hot
||
5302 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
5304 if (tsk_cache_hot
) {
5305 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5306 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5312 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
5317 * detach_task() -- detach the task for the migration specified in env
5319 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
5321 lockdep_assert_held(&env
->src_rq
->lock
);
5323 deactivate_task(env
->src_rq
, p
, 0);
5324 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
5325 set_task_cpu(p
, env
->dst_cpu
);
5329 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5330 * part of active balancing operations within "domain".
5332 * Returns a task if successful and NULL otherwise.
5334 static struct task_struct
*detach_one_task(struct lb_env
*env
)
5336 struct task_struct
*p
, *n
;
5338 lockdep_assert_held(&env
->src_rq
->lock
);
5340 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
5341 if (!can_migrate_task(p
, env
))
5344 detach_task(p
, env
);
5347 * Right now, this is only the second place where
5348 * lb_gained[env->idle] is updated (other is detach_tasks)
5349 * so we can safely collect stats here rather than
5350 * inside detach_tasks().
5352 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
5358 static const unsigned int sched_nr_migrate_break
= 32;
5361 * detach_tasks() -- tries to detach up to imbalance weighted load from
5362 * busiest_rq, as part of a balancing operation within domain "sd".
5364 * Returns number of detached tasks if successful and 0 otherwise.
5366 static int detach_tasks(struct lb_env
*env
)
5368 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
5369 struct task_struct
*p
;
5373 lockdep_assert_held(&env
->src_rq
->lock
);
5375 if (env
->imbalance
<= 0)
5378 while (!list_empty(tasks
)) {
5379 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5382 /* We've more or less seen every task there is, call it quits */
5383 if (env
->loop
> env
->loop_max
)
5386 /* take a breather every nr_migrate tasks */
5387 if (env
->loop
> env
->loop_break
) {
5388 env
->loop_break
+= sched_nr_migrate_break
;
5389 env
->flags
|= LBF_NEED_BREAK
;
5393 if (!can_migrate_task(p
, env
))
5396 load
= task_h_load(p
);
5398 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5401 if ((load
/ 2) > env
->imbalance
)
5404 detach_task(p
, env
);
5405 list_add(&p
->se
.group_node
, &env
->tasks
);
5408 env
->imbalance
-= load
;
5410 #ifdef CONFIG_PREEMPT
5412 * NEWIDLE balancing is a source of latency, so preemptible
5413 * kernels will stop after the first task is detached to minimize
5414 * the critical section.
5416 if (env
->idle
== CPU_NEWLY_IDLE
)
5421 * We only want to steal up to the prescribed amount of
5424 if (env
->imbalance
<= 0)
5429 list_move_tail(&p
->se
.group_node
, tasks
);
5433 * Right now, this is one of only two places we collect this stat
5434 * so we can safely collect detach_one_task() stats here rather
5435 * than inside detach_one_task().
5437 schedstat_add(env
->sd
, lb_gained
[env
->idle
], detached
);
5443 * attach_task() -- attach the task detached by detach_task() to its new rq.
5445 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
5447 lockdep_assert_held(&rq
->lock
);
5449 BUG_ON(task_rq(p
) != rq
);
5450 p
->on_rq
= TASK_ON_RQ_QUEUED
;
5451 activate_task(rq
, p
, 0);
5452 check_preempt_curr(rq
, p
, 0);
5456 * attach_one_task() -- attaches the task returned from detach_one_task() to
5459 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
5461 raw_spin_lock(&rq
->lock
);
5463 raw_spin_unlock(&rq
->lock
);
5467 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5470 static void attach_tasks(struct lb_env
*env
)
5472 struct list_head
*tasks
= &env
->tasks
;
5473 struct task_struct
*p
;
5475 raw_spin_lock(&env
->dst_rq
->lock
);
5477 while (!list_empty(tasks
)) {
5478 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5479 list_del_init(&p
->se
.group_node
);
5481 attach_task(env
->dst_rq
, p
);
5484 raw_spin_unlock(&env
->dst_rq
->lock
);
5487 #ifdef CONFIG_FAIR_GROUP_SCHED
5489 * update tg->load_weight by folding this cpu's load_avg
5491 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5493 struct sched_entity
*se
= tg
->se
[cpu
];
5494 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5496 /* throttled entities do not contribute to load */
5497 if (throttled_hierarchy(cfs_rq
))
5500 update_cfs_rq_blocked_load(cfs_rq
, 1);
5503 update_entity_load_avg(se
, 1);
5505 * We pivot on our runnable average having decayed to zero for
5506 * list removal. This generally implies that all our children
5507 * have also been removed (modulo rounding error or bandwidth
5508 * control); however, such cases are rare and we can fix these
5511 * TODO: fix up out-of-order children on enqueue.
5513 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5514 list_del_leaf_cfs_rq(cfs_rq
);
5516 struct rq
*rq
= rq_of(cfs_rq
);
5517 update_rq_runnable_avg(rq
, rq
->nr_running
);
5521 static void update_blocked_averages(int cpu
)
5523 struct rq
*rq
= cpu_rq(cpu
);
5524 struct cfs_rq
*cfs_rq
;
5525 unsigned long flags
;
5527 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5528 update_rq_clock(rq
);
5530 * Iterates the task_group tree in a bottom up fashion, see
5531 * list_add_leaf_cfs_rq() for details.
5533 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5535 * Note: We may want to consider periodically releasing
5536 * rq->lock about these updates so that creating many task
5537 * groups does not result in continually extending hold time.
5539 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5542 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5546 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5547 * This needs to be done in a top-down fashion because the load of a child
5548 * group is a fraction of its parents load.
5550 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5552 struct rq
*rq
= rq_of(cfs_rq
);
5553 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5554 unsigned long now
= jiffies
;
5557 if (cfs_rq
->last_h_load_update
== now
)
5560 cfs_rq
->h_load_next
= NULL
;
5561 for_each_sched_entity(se
) {
5562 cfs_rq
= cfs_rq_of(se
);
5563 cfs_rq
->h_load_next
= se
;
5564 if (cfs_rq
->last_h_load_update
== now
)
5569 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5570 cfs_rq
->last_h_load_update
= now
;
5573 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5574 load
= cfs_rq
->h_load
;
5575 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5576 cfs_rq
->runnable_load_avg
+ 1);
5577 cfs_rq
= group_cfs_rq(se
);
5578 cfs_rq
->h_load
= load
;
5579 cfs_rq
->last_h_load_update
= now
;
5583 static unsigned long task_h_load(struct task_struct
*p
)
5585 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5587 update_cfs_rq_h_load(cfs_rq
);
5588 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5589 cfs_rq
->runnable_load_avg
+ 1);
5592 static inline void update_blocked_averages(int cpu
)
5596 static unsigned long task_h_load(struct task_struct
*p
)
5598 return p
->se
.avg
.load_avg_contrib
;
5602 /********** Helpers for find_busiest_group ************************/
5611 * sg_lb_stats - stats of a sched_group required for load_balancing
5613 struct sg_lb_stats
{
5614 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5615 unsigned long group_load
; /* Total load over the CPUs of the group */
5616 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5617 unsigned long load_per_task
;
5618 unsigned long group_capacity
;
5619 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5620 unsigned int group_capacity_factor
;
5621 unsigned int idle_cpus
;
5622 unsigned int group_weight
;
5623 enum group_type group_type
;
5624 int group_has_free_capacity
;
5625 #ifdef CONFIG_NUMA_BALANCING
5626 unsigned int nr_numa_running
;
5627 unsigned int nr_preferred_running
;
5632 * sd_lb_stats - Structure to store the statistics of a sched_domain
5633 * during load balancing.
5635 struct sd_lb_stats
{
5636 struct sched_group
*busiest
; /* Busiest group in this sd */
5637 struct sched_group
*local
; /* Local group in this sd */
5638 unsigned long total_load
; /* Total load of all groups in sd */
5639 unsigned long total_capacity
; /* Total capacity of all groups in sd */
5640 unsigned long avg_load
; /* Average load across all groups in sd */
5642 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5643 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5646 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5649 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5650 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5651 * We must however clear busiest_stat::avg_load because
5652 * update_sd_pick_busiest() reads this before assignment.
5654 *sds
= (struct sd_lb_stats
){
5658 .total_capacity
= 0UL,
5661 .sum_nr_running
= 0,
5662 .group_type
= group_other
,
5668 * get_sd_load_idx - Obtain the load index for a given sched domain.
5669 * @sd: The sched_domain whose load_idx is to be obtained.
5670 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5672 * Return: The load index.
5674 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5675 enum cpu_idle_type idle
)
5681 load_idx
= sd
->busy_idx
;
5684 case CPU_NEWLY_IDLE
:
5685 load_idx
= sd
->newidle_idx
;
5688 load_idx
= sd
->idle_idx
;
5695 static unsigned long default_scale_capacity(struct sched_domain
*sd
, int cpu
)
5697 return SCHED_CAPACITY_SCALE
;
5700 unsigned long __weak
arch_scale_freq_capacity(struct sched_domain
*sd
, int cpu
)
5702 return default_scale_capacity(sd
, cpu
);
5705 static unsigned long default_scale_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5707 if ((sd
->flags
& SD_SHARE_CPUCAPACITY
) && (sd
->span_weight
> 1))
5708 return sd
->smt_gain
/ sd
->span_weight
;
5710 return SCHED_CAPACITY_SCALE
;
5713 unsigned long __weak
arch_scale_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5715 return default_scale_cpu_capacity(sd
, cpu
);
5718 static unsigned long scale_rt_capacity(int cpu
)
5720 struct rq
*rq
= cpu_rq(cpu
);
5721 u64 total
, available
, age_stamp
, avg
;
5725 * Since we're reading these variables without serialization make sure
5726 * we read them once before doing sanity checks on them.
5728 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5729 avg
= ACCESS_ONCE(rq
->rt_avg
);
5731 delta
= rq_clock(rq
) - age_stamp
;
5732 if (unlikely(delta
< 0))
5735 total
= sched_avg_period() + delta
;
5737 if (unlikely(total
< avg
)) {
5738 /* Ensures that capacity won't end up being negative */
5741 available
= total
- avg
;
5744 if (unlikely((s64
)total
< SCHED_CAPACITY_SCALE
))
5745 total
= SCHED_CAPACITY_SCALE
;
5747 total
>>= SCHED_CAPACITY_SHIFT
;
5749 return div_u64(available
, total
);
5752 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5754 unsigned long capacity
= SCHED_CAPACITY_SCALE
;
5755 struct sched_group
*sdg
= sd
->groups
;
5757 if (sched_feat(ARCH_CAPACITY
))
5758 capacity
*= arch_scale_cpu_capacity(sd
, cpu
);
5760 capacity
*= default_scale_cpu_capacity(sd
, cpu
);
5762 capacity
>>= SCHED_CAPACITY_SHIFT
;
5764 sdg
->sgc
->capacity_orig
= capacity
;
5766 if (sched_feat(ARCH_CAPACITY
))
5767 capacity
*= arch_scale_freq_capacity(sd
, cpu
);
5769 capacity
*= default_scale_capacity(sd
, cpu
);
5771 capacity
>>= SCHED_CAPACITY_SHIFT
;
5773 capacity
*= scale_rt_capacity(cpu
);
5774 capacity
>>= SCHED_CAPACITY_SHIFT
;
5779 cpu_rq(cpu
)->cpu_capacity
= capacity
;
5780 sdg
->sgc
->capacity
= capacity
;
5783 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
5785 struct sched_domain
*child
= sd
->child
;
5786 struct sched_group
*group
, *sdg
= sd
->groups
;
5787 unsigned long capacity
, capacity_orig
;
5788 unsigned long interval
;
5790 interval
= msecs_to_jiffies(sd
->balance_interval
);
5791 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5792 sdg
->sgc
->next_update
= jiffies
+ interval
;
5795 update_cpu_capacity(sd
, cpu
);
5799 capacity_orig
= capacity
= 0;
5801 if (child
->flags
& SD_OVERLAP
) {
5803 * SD_OVERLAP domains cannot assume that child groups
5804 * span the current group.
5807 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
5808 struct sched_group_capacity
*sgc
;
5809 struct rq
*rq
= cpu_rq(cpu
);
5812 * build_sched_domains() -> init_sched_groups_capacity()
5813 * gets here before we've attached the domains to the
5816 * Use capacity_of(), which is set irrespective of domains
5817 * in update_cpu_capacity().
5819 * This avoids capacity/capacity_orig from being 0 and
5820 * causing divide-by-zero issues on boot.
5822 * Runtime updates will correct capacity_orig.
5824 if (unlikely(!rq
->sd
)) {
5825 capacity_orig
+= capacity_of(cpu
);
5826 capacity
+= capacity_of(cpu
);
5830 sgc
= rq
->sd
->groups
->sgc
;
5831 capacity_orig
+= sgc
->capacity_orig
;
5832 capacity
+= sgc
->capacity
;
5836 * !SD_OVERLAP domains can assume that child groups
5837 * span the current group.
5840 group
= child
->groups
;
5842 capacity_orig
+= group
->sgc
->capacity_orig
;
5843 capacity
+= group
->sgc
->capacity
;
5844 group
= group
->next
;
5845 } while (group
!= child
->groups
);
5848 sdg
->sgc
->capacity_orig
= capacity_orig
;
5849 sdg
->sgc
->capacity
= capacity
;
5853 * Try and fix up capacity for tiny siblings, this is needed when
5854 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5855 * which on its own isn't powerful enough.
5857 * See update_sd_pick_busiest() and check_asym_packing().
5860 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
5863 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
5865 if (!(sd
->flags
& SD_SHARE_CPUCAPACITY
))
5869 * If ~90% of the cpu_capacity is still there, we're good.
5871 if (group
->sgc
->capacity
* 32 > group
->sgc
->capacity_orig
* 29)
5878 * Group imbalance indicates (and tries to solve) the problem where balancing
5879 * groups is inadequate due to tsk_cpus_allowed() constraints.
5881 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5882 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5885 * { 0 1 2 3 } { 4 5 6 7 }
5888 * If we were to balance group-wise we'd place two tasks in the first group and
5889 * two tasks in the second group. Clearly this is undesired as it will overload
5890 * cpu 3 and leave one of the cpus in the second group unused.
5892 * The current solution to this issue is detecting the skew in the first group
5893 * by noticing the lower domain failed to reach balance and had difficulty
5894 * moving tasks due to affinity constraints.
5896 * When this is so detected; this group becomes a candidate for busiest; see
5897 * update_sd_pick_busiest(). And calculate_imbalance() and
5898 * find_busiest_group() avoid some of the usual balance conditions to allow it
5899 * to create an effective group imbalance.
5901 * This is a somewhat tricky proposition since the next run might not find the
5902 * group imbalance and decide the groups need to be balanced again. A most
5903 * subtle and fragile situation.
5906 static inline int sg_imbalanced(struct sched_group
*group
)
5908 return group
->sgc
->imbalance
;
5912 * Compute the group capacity factor.
5914 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
5915 * first dividing out the smt factor and computing the actual number of cores
5916 * and limit unit capacity with that.
5918 static inline int sg_capacity_factor(struct lb_env
*env
, struct sched_group
*group
)
5920 unsigned int capacity_factor
, smt
, cpus
;
5921 unsigned int capacity
, capacity_orig
;
5923 capacity
= group
->sgc
->capacity
;
5924 capacity_orig
= group
->sgc
->capacity_orig
;
5925 cpus
= group
->group_weight
;
5927 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
5928 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, capacity_orig
);
5929 capacity_factor
= cpus
/ smt
; /* cores */
5931 capacity_factor
= min_t(unsigned,
5932 capacity_factor
, DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
));
5933 if (!capacity_factor
)
5934 capacity_factor
= fix_small_capacity(env
->sd
, group
);
5936 return capacity_factor
;
5939 static enum group_type
5940 group_classify(struct sched_group
*group
, struct sg_lb_stats
*sgs
)
5942 if (sgs
->sum_nr_running
> sgs
->group_capacity_factor
)
5943 return group_overloaded
;
5945 if (sg_imbalanced(group
))
5946 return group_imbalanced
;
5952 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5953 * @env: The load balancing environment.
5954 * @group: sched_group whose statistics are to be updated.
5955 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5956 * @local_group: Does group contain this_cpu.
5957 * @sgs: variable to hold the statistics for this group.
5958 * @overload: Indicate more than one runnable task for any CPU.
5960 static inline void update_sg_lb_stats(struct lb_env
*env
,
5961 struct sched_group
*group
, int load_idx
,
5962 int local_group
, struct sg_lb_stats
*sgs
,
5968 memset(sgs
, 0, sizeof(*sgs
));
5970 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5971 struct rq
*rq
= cpu_rq(i
);
5973 /* Bias balancing toward cpus of our domain */
5975 load
= target_load(i
, load_idx
);
5977 load
= source_load(i
, load_idx
);
5979 sgs
->group_load
+= load
;
5980 sgs
->sum_nr_running
+= rq
->cfs
.h_nr_running
;
5982 if (rq
->nr_running
> 1)
5985 #ifdef CONFIG_NUMA_BALANCING
5986 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
5987 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
5989 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
5994 /* Adjust by relative CPU capacity of the group */
5995 sgs
->group_capacity
= group
->sgc
->capacity
;
5996 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
5998 if (sgs
->sum_nr_running
)
5999 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
6001 sgs
->group_weight
= group
->group_weight
;
6002 sgs
->group_capacity_factor
= sg_capacity_factor(env
, group
);
6003 sgs
->group_type
= group_classify(group
, sgs
);
6005 if (sgs
->group_capacity_factor
> sgs
->sum_nr_running
)
6006 sgs
->group_has_free_capacity
= 1;
6010 * update_sd_pick_busiest - return 1 on busiest group
6011 * @env: The load balancing environment.
6012 * @sds: sched_domain statistics
6013 * @sg: sched_group candidate to be checked for being the busiest
6014 * @sgs: sched_group statistics
6016 * Determine if @sg is a busier group than the previously selected
6019 * Return: %true if @sg is a busier group than the previously selected
6020 * busiest group. %false otherwise.
6022 static bool update_sd_pick_busiest(struct lb_env
*env
,
6023 struct sd_lb_stats
*sds
,
6024 struct sched_group
*sg
,
6025 struct sg_lb_stats
*sgs
)
6027 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
6029 if (sgs
->group_type
> busiest
->group_type
)
6032 if (sgs
->group_type
< busiest
->group_type
)
6035 if (sgs
->avg_load
<= busiest
->avg_load
)
6038 /* This is the busiest node in its class. */
6039 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6043 * ASYM_PACKING needs to move all the work to the lowest
6044 * numbered CPUs in the group, therefore mark all groups
6045 * higher than ourself as busy.
6047 if (sgs
->sum_nr_running
&& env
->dst_cpu
< group_first_cpu(sg
)) {
6051 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
6058 #ifdef CONFIG_NUMA_BALANCING
6059 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6061 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
6063 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
6068 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6070 if (rq
->nr_running
> rq
->nr_numa_running
)
6072 if (rq
->nr_running
> rq
->nr_preferred_running
)
6077 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6082 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6086 #endif /* CONFIG_NUMA_BALANCING */
6089 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6090 * @env: The load balancing environment.
6091 * @sds: variable to hold the statistics for this sched_domain.
6093 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6095 struct sched_domain
*child
= env
->sd
->child
;
6096 struct sched_group
*sg
= env
->sd
->groups
;
6097 struct sg_lb_stats tmp_sgs
;
6098 int load_idx
, prefer_sibling
= 0;
6099 bool overload
= false;
6101 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
6104 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
6107 struct sg_lb_stats
*sgs
= &tmp_sgs
;
6110 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
6113 sgs
= &sds
->local_stat
;
6115 if (env
->idle
!= CPU_NEWLY_IDLE
||
6116 time_after_eq(jiffies
, sg
->sgc
->next_update
))
6117 update_group_capacity(env
->sd
, env
->dst_cpu
);
6120 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
,
6127 * In case the child domain prefers tasks go to siblings
6128 * first, lower the sg capacity factor to one so that we'll try
6129 * and move all the excess tasks away. We lower the capacity
6130 * of a group only if the local group has the capacity to fit
6131 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6132 * extra check prevents the case where you always pull from the
6133 * heaviest group when it is already under-utilized (possible
6134 * with a large weight task outweighs the tasks on the system).
6136 if (prefer_sibling
&& sds
->local
&&
6137 sds
->local_stat
.group_has_free_capacity
)
6138 sgs
->group_capacity_factor
= min(sgs
->group_capacity_factor
, 1U);
6140 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
6142 sds
->busiest_stat
= *sgs
;
6146 /* Now, start updating sd_lb_stats */
6147 sds
->total_load
+= sgs
->group_load
;
6148 sds
->total_capacity
+= sgs
->group_capacity
;
6151 } while (sg
!= env
->sd
->groups
);
6153 if (env
->sd
->flags
& SD_NUMA
)
6154 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
6156 if (!env
->sd
->parent
) {
6157 /* update overload indicator if we are at root domain */
6158 if (env
->dst_rq
->rd
->overload
!= overload
)
6159 env
->dst_rq
->rd
->overload
= overload
;
6165 * check_asym_packing - Check to see if the group is packed into the
6168 * This is primarily intended to used at the sibling level. Some
6169 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6170 * case of POWER7, it can move to lower SMT modes only when higher
6171 * threads are idle. When in lower SMT modes, the threads will
6172 * perform better since they share less core resources. Hence when we
6173 * have idle threads, we want them to be the higher ones.
6175 * This packing function is run on idle threads. It checks to see if
6176 * the busiest CPU in this domain (core in the P7 case) has a higher
6177 * CPU number than the packing function is being run on. Here we are
6178 * assuming lower CPU number will be equivalent to lower a SMT thread
6181 * Return: 1 when packing is required and a task should be moved to
6182 * this CPU. The amount of the imbalance is returned in *imbalance.
6184 * @env: The load balancing environment.
6185 * @sds: Statistics of the sched_domain which is to be packed
6187 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6191 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6197 busiest_cpu
= group_first_cpu(sds
->busiest
);
6198 if (env
->dst_cpu
> busiest_cpu
)
6201 env
->imbalance
= DIV_ROUND_CLOSEST(
6202 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
6203 SCHED_CAPACITY_SCALE
);
6209 * fix_small_imbalance - Calculate the minor imbalance that exists
6210 * amongst the groups of a sched_domain, during
6212 * @env: The load balancing environment.
6213 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6216 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6218 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
6219 unsigned int imbn
= 2;
6220 unsigned long scaled_busy_load_per_task
;
6221 struct sg_lb_stats
*local
, *busiest
;
6223 local
= &sds
->local_stat
;
6224 busiest
= &sds
->busiest_stat
;
6226 if (!local
->sum_nr_running
)
6227 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
6228 else if (busiest
->load_per_task
> local
->load_per_task
)
6231 scaled_busy_load_per_task
=
6232 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6233 busiest
->group_capacity
;
6235 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
6236 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
6237 env
->imbalance
= busiest
->load_per_task
;
6242 * OK, we don't have enough imbalance to justify moving tasks,
6243 * however we may be able to increase total CPU capacity used by
6247 capa_now
+= busiest
->group_capacity
*
6248 min(busiest
->load_per_task
, busiest
->avg_load
);
6249 capa_now
+= local
->group_capacity
*
6250 min(local
->load_per_task
, local
->avg_load
);
6251 capa_now
/= SCHED_CAPACITY_SCALE
;
6253 /* Amount of load we'd subtract */
6254 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
6255 capa_move
+= busiest
->group_capacity
*
6256 min(busiest
->load_per_task
,
6257 busiest
->avg_load
- scaled_busy_load_per_task
);
6260 /* Amount of load we'd add */
6261 if (busiest
->avg_load
* busiest
->group_capacity
<
6262 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
6263 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
6264 local
->group_capacity
;
6266 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6267 local
->group_capacity
;
6269 capa_move
+= local
->group_capacity
*
6270 min(local
->load_per_task
, local
->avg_load
+ tmp
);
6271 capa_move
/= SCHED_CAPACITY_SCALE
;
6273 /* Move if we gain throughput */
6274 if (capa_move
> capa_now
)
6275 env
->imbalance
= busiest
->load_per_task
;
6279 * calculate_imbalance - Calculate the amount of imbalance present within the
6280 * groups of a given sched_domain during load balance.
6281 * @env: load balance environment
6282 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6284 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6286 unsigned long max_pull
, load_above_capacity
= ~0UL;
6287 struct sg_lb_stats
*local
, *busiest
;
6289 local
= &sds
->local_stat
;
6290 busiest
= &sds
->busiest_stat
;
6292 if (busiest
->group_type
== group_imbalanced
) {
6294 * In the group_imb case we cannot rely on group-wide averages
6295 * to ensure cpu-load equilibrium, look at wider averages. XXX
6297 busiest
->load_per_task
=
6298 min(busiest
->load_per_task
, sds
->avg_load
);
6302 * In the presence of smp nice balancing, certain scenarios can have
6303 * max load less than avg load(as we skip the groups at or below
6304 * its cpu_capacity, while calculating max_load..)
6306 if (busiest
->avg_load
<= sds
->avg_load
||
6307 local
->avg_load
>= sds
->avg_load
) {
6309 return fix_small_imbalance(env
, sds
);
6313 * If there aren't any idle cpus, avoid creating some.
6315 if (busiest
->group_type
== group_overloaded
&&
6316 local
->group_type
== group_overloaded
) {
6317 load_above_capacity
=
6318 (busiest
->sum_nr_running
- busiest
->group_capacity_factor
);
6320 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_CAPACITY_SCALE
);
6321 load_above_capacity
/= busiest
->group_capacity
;
6325 * We're trying to get all the cpus to the average_load, so we don't
6326 * want to push ourselves above the average load, nor do we wish to
6327 * reduce the max loaded cpu below the average load. At the same time,
6328 * we also don't want to reduce the group load below the group capacity
6329 * (so that we can implement power-savings policies etc). Thus we look
6330 * for the minimum possible imbalance.
6332 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
6334 /* How much load to actually move to equalise the imbalance */
6335 env
->imbalance
= min(
6336 max_pull
* busiest
->group_capacity
,
6337 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
6338 ) / SCHED_CAPACITY_SCALE
;
6341 * if *imbalance is less than the average load per runnable task
6342 * there is no guarantee that any tasks will be moved so we'll have
6343 * a think about bumping its value to force at least one task to be
6346 if (env
->imbalance
< busiest
->load_per_task
)
6347 return fix_small_imbalance(env
, sds
);
6350 /******* find_busiest_group() helpers end here *********************/
6353 * find_busiest_group - Returns the busiest group within the sched_domain
6354 * if there is an imbalance. If there isn't an imbalance, and
6355 * the user has opted for power-savings, it returns a group whose
6356 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6357 * such a group exists.
6359 * Also calculates the amount of weighted load which should be moved
6360 * to restore balance.
6362 * @env: The load balancing environment.
6364 * Return: - The busiest group if imbalance exists.
6365 * - If no imbalance and user has opted for power-savings balance,
6366 * return the least loaded group whose CPUs can be
6367 * put to idle by rebalancing its tasks onto our group.
6369 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
6371 struct sg_lb_stats
*local
, *busiest
;
6372 struct sd_lb_stats sds
;
6374 init_sd_lb_stats(&sds
);
6377 * Compute the various statistics relavent for load balancing at
6380 update_sd_lb_stats(env
, &sds
);
6381 local
= &sds
.local_stat
;
6382 busiest
= &sds
.busiest_stat
;
6384 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
6385 check_asym_packing(env
, &sds
))
6388 /* There is no busy sibling group to pull tasks from */
6389 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
6392 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
6393 / sds
.total_capacity
;
6396 * If the busiest group is imbalanced the below checks don't
6397 * work because they assume all things are equal, which typically
6398 * isn't true due to cpus_allowed constraints and the like.
6400 if (busiest
->group_type
== group_imbalanced
)
6403 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6404 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_free_capacity
&&
6405 !busiest
->group_has_free_capacity
)
6409 * If the local group is busier than the selected busiest group
6410 * don't try and pull any tasks.
6412 if (local
->avg_load
>= busiest
->avg_load
)
6416 * Don't pull any tasks if this group is already above the domain
6419 if (local
->avg_load
>= sds
.avg_load
)
6422 if (env
->idle
== CPU_IDLE
) {
6424 * This cpu is idle. If the busiest group load doesn't
6425 * have more tasks than the number of available cpu's and
6426 * there is no imbalance between this and busiest group
6427 * wrt to idle cpu's, it is balanced.
6429 if ((local
->idle_cpus
< busiest
->idle_cpus
) &&
6430 busiest
->sum_nr_running
<= busiest
->group_weight
)
6434 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6435 * imbalance_pct to be conservative.
6437 if (100 * busiest
->avg_load
<=
6438 env
->sd
->imbalance_pct
* local
->avg_load
)
6443 /* Looks like there is an imbalance. Compute it */
6444 calculate_imbalance(env
, &sds
);
6453 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6455 static struct rq
*find_busiest_queue(struct lb_env
*env
,
6456 struct sched_group
*group
)
6458 struct rq
*busiest
= NULL
, *rq
;
6459 unsigned long busiest_load
= 0, busiest_capacity
= 1;
6462 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6463 unsigned long capacity
, capacity_factor
, wl
;
6467 rt
= fbq_classify_rq(rq
);
6470 * We classify groups/runqueues into three groups:
6471 * - regular: there are !numa tasks
6472 * - remote: there are numa tasks that run on the 'wrong' node
6473 * - all: there is no distinction
6475 * In order to avoid migrating ideally placed numa tasks,
6476 * ignore those when there's better options.
6478 * If we ignore the actual busiest queue to migrate another
6479 * task, the next balance pass can still reduce the busiest
6480 * queue by moving tasks around inside the node.
6482 * If we cannot move enough load due to this classification
6483 * the next pass will adjust the group classification and
6484 * allow migration of more tasks.
6486 * Both cases only affect the total convergence complexity.
6488 if (rt
> env
->fbq_type
)
6491 capacity
= capacity_of(i
);
6492 capacity_factor
= DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
);
6493 if (!capacity_factor
)
6494 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6496 wl
= weighted_cpuload(i
);
6499 * When comparing with imbalance, use weighted_cpuload()
6500 * which is not scaled with the cpu capacity.
6502 if (capacity_factor
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6506 * For the load comparisons with the other cpu's, consider
6507 * the weighted_cpuload() scaled with the cpu capacity, so
6508 * that the load can be moved away from the cpu that is
6509 * potentially running at a lower capacity.
6511 * Thus we're looking for max(wl_i / capacity_i), crosswise
6512 * multiplication to rid ourselves of the division works out
6513 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6514 * our previous maximum.
6516 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
6518 busiest_capacity
= capacity
;
6527 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6528 * so long as it is large enough.
6530 #define MAX_PINNED_INTERVAL 512
6532 /* Working cpumask for load_balance and load_balance_newidle. */
6533 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6535 static int need_active_balance(struct lb_env
*env
)
6537 struct sched_domain
*sd
= env
->sd
;
6539 if (env
->idle
== CPU_NEWLY_IDLE
) {
6542 * ASYM_PACKING needs to force migrate tasks from busy but
6543 * higher numbered CPUs in order to pack all tasks in the
6544 * lowest numbered CPUs.
6546 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6550 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6553 static int active_load_balance_cpu_stop(void *data
);
6555 static int should_we_balance(struct lb_env
*env
)
6557 struct sched_group
*sg
= env
->sd
->groups
;
6558 struct cpumask
*sg_cpus
, *sg_mask
;
6559 int cpu
, balance_cpu
= -1;
6562 * In the newly idle case, we will allow all the cpu's
6563 * to do the newly idle load balance.
6565 if (env
->idle
== CPU_NEWLY_IDLE
)
6568 sg_cpus
= sched_group_cpus(sg
);
6569 sg_mask
= sched_group_mask(sg
);
6570 /* Try to find first idle cpu */
6571 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6572 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6579 if (balance_cpu
== -1)
6580 balance_cpu
= group_balance_cpu(sg
);
6583 * First idle cpu or the first cpu(busiest) in this sched group
6584 * is eligible for doing load balancing at this and above domains.
6586 return balance_cpu
== env
->dst_cpu
;
6590 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6591 * tasks if there is an imbalance.
6593 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6594 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6595 int *continue_balancing
)
6597 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6598 struct sched_domain
*sd_parent
= sd
->parent
;
6599 struct sched_group
*group
;
6601 unsigned long flags
;
6602 struct cpumask
*cpus
= __get_cpu_var(load_balance_mask
);
6604 struct lb_env env
= {
6606 .dst_cpu
= this_cpu
,
6608 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6610 .loop_break
= sched_nr_migrate_break
,
6613 .tasks
= LIST_HEAD_INIT(env
.tasks
),
6617 * For NEWLY_IDLE load_balancing, we don't need to consider
6618 * other cpus in our group
6620 if (idle
== CPU_NEWLY_IDLE
)
6621 env
.dst_grpmask
= NULL
;
6623 cpumask_copy(cpus
, cpu_active_mask
);
6625 schedstat_inc(sd
, lb_count
[idle
]);
6628 if (!should_we_balance(&env
)) {
6629 *continue_balancing
= 0;
6633 group
= find_busiest_group(&env
);
6635 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6639 busiest
= find_busiest_queue(&env
, group
);
6641 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6645 BUG_ON(busiest
== env
.dst_rq
);
6647 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6650 if (busiest
->nr_running
> 1) {
6652 * Attempt to move tasks. If find_busiest_group has found
6653 * an imbalance but busiest->nr_running <= 1, the group is
6654 * still unbalanced. ld_moved simply stays zero, so it is
6655 * correctly treated as an imbalance.
6657 env
.flags
|= LBF_ALL_PINNED
;
6658 env
.src_cpu
= busiest
->cpu
;
6659 env
.src_rq
= busiest
;
6660 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6663 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6666 * cur_ld_moved - load moved in current iteration
6667 * ld_moved - cumulative load moved across iterations
6669 cur_ld_moved
= detach_tasks(&env
);
6672 * We've detached some tasks from busiest_rq. Every
6673 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6674 * unlock busiest->lock, and we are able to be sure
6675 * that nobody can manipulate the tasks in parallel.
6676 * See task_rq_lock() family for the details.
6679 raw_spin_unlock(&busiest
->lock
);
6683 ld_moved
+= cur_ld_moved
;
6686 local_irq_restore(flags
);
6689 * some other cpu did the load balance for us.
6691 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
6692 resched_cpu(env
.dst_cpu
);
6694 if (env
.flags
& LBF_NEED_BREAK
) {
6695 env
.flags
&= ~LBF_NEED_BREAK
;
6700 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6701 * us and move them to an alternate dst_cpu in our sched_group
6702 * where they can run. The upper limit on how many times we
6703 * iterate on same src_cpu is dependent on number of cpus in our
6706 * This changes load balance semantics a bit on who can move
6707 * load to a given_cpu. In addition to the given_cpu itself
6708 * (or a ilb_cpu acting on its behalf where given_cpu is
6709 * nohz-idle), we now have balance_cpu in a position to move
6710 * load to given_cpu. In rare situations, this may cause
6711 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6712 * _independently_ and at _same_ time to move some load to
6713 * given_cpu) causing exceess load to be moved to given_cpu.
6714 * This however should not happen so much in practice and
6715 * moreover subsequent load balance cycles should correct the
6716 * excess load moved.
6718 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6720 /* Prevent to re-select dst_cpu via env's cpus */
6721 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6723 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6724 env
.dst_cpu
= env
.new_dst_cpu
;
6725 env
.flags
&= ~LBF_DST_PINNED
;
6727 env
.loop_break
= sched_nr_migrate_break
;
6730 * Go back to "more_balance" rather than "redo" since we
6731 * need to continue with same src_cpu.
6737 * We failed to reach balance because of affinity.
6740 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
6742 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
6743 *group_imbalance
= 1;
6746 /* All tasks on this runqueue were pinned by CPU affinity */
6747 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6748 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6749 if (!cpumask_empty(cpus
)) {
6751 env
.loop_break
= sched_nr_migrate_break
;
6754 goto out_all_pinned
;
6759 schedstat_inc(sd
, lb_failed
[idle
]);
6761 * Increment the failure counter only on periodic balance.
6762 * We do not want newidle balance, which can be very
6763 * frequent, pollute the failure counter causing
6764 * excessive cache_hot migrations and active balances.
6766 if (idle
!= CPU_NEWLY_IDLE
)
6767 sd
->nr_balance_failed
++;
6769 if (need_active_balance(&env
)) {
6770 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6772 /* don't kick the active_load_balance_cpu_stop,
6773 * if the curr task on busiest cpu can't be
6776 if (!cpumask_test_cpu(this_cpu
,
6777 tsk_cpus_allowed(busiest
->curr
))) {
6778 raw_spin_unlock_irqrestore(&busiest
->lock
,
6780 env
.flags
|= LBF_ALL_PINNED
;
6781 goto out_one_pinned
;
6785 * ->active_balance synchronizes accesses to
6786 * ->active_balance_work. Once set, it's cleared
6787 * only after active load balance is finished.
6789 if (!busiest
->active_balance
) {
6790 busiest
->active_balance
= 1;
6791 busiest
->push_cpu
= this_cpu
;
6794 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
6796 if (active_balance
) {
6797 stop_one_cpu_nowait(cpu_of(busiest
),
6798 active_load_balance_cpu_stop
, busiest
,
6799 &busiest
->active_balance_work
);
6803 * We've kicked active balancing, reset the failure
6806 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
6809 sd
->nr_balance_failed
= 0;
6811 if (likely(!active_balance
)) {
6812 /* We were unbalanced, so reset the balancing interval */
6813 sd
->balance_interval
= sd
->min_interval
;
6816 * If we've begun active balancing, start to back off. This
6817 * case may not be covered by the all_pinned logic if there
6818 * is only 1 task on the busy runqueue (because we don't call
6821 if (sd
->balance_interval
< sd
->max_interval
)
6822 sd
->balance_interval
*= 2;
6829 * We reach balance although we may have faced some affinity
6830 * constraints. Clear the imbalance flag if it was set.
6833 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
6835 if (*group_imbalance
)
6836 *group_imbalance
= 0;
6841 * We reach balance because all tasks are pinned at this level so
6842 * we can't migrate them. Let the imbalance flag set so parent level
6843 * can try to migrate them.
6845 schedstat_inc(sd
, lb_balanced
[idle
]);
6847 sd
->nr_balance_failed
= 0;
6850 /* tune up the balancing interval */
6851 if (((env
.flags
& LBF_ALL_PINNED
) &&
6852 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
6853 (sd
->balance_interval
< sd
->max_interval
))
6854 sd
->balance_interval
*= 2;
6861 static inline unsigned long
6862 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
6864 unsigned long interval
= sd
->balance_interval
;
6867 interval
*= sd
->busy_factor
;
6869 /* scale ms to jiffies */
6870 interval
= msecs_to_jiffies(interval
);
6871 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6877 update_next_balance(struct sched_domain
*sd
, int cpu_busy
, unsigned long *next_balance
)
6879 unsigned long interval
, next
;
6881 interval
= get_sd_balance_interval(sd
, cpu_busy
);
6882 next
= sd
->last_balance
+ interval
;
6884 if (time_after(*next_balance
, next
))
6885 *next_balance
= next
;
6889 * idle_balance is called by schedule() if this_cpu is about to become
6890 * idle. Attempts to pull tasks from other CPUs.
6892 static int idle_balance(struct rq
*this_rq
)
6894 unsigned long next_balance
= jiffies
+ HZ
;
6895 int this_cpu
= this_rq
->cpu
;
6896 struct sched_domain
*sd
;
6897 int pulled_task
= 0;
6900 idle_enter_fair(this_rq
);
6903 * We must set idle_stamp _before_ calling idle_balance(), such that we
6904 * measure the duration of idle_balance() as idle time.
6906 this_rq
->idle_stamp
= rq_clock(this_rq
);
6908 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
6909 !this_rq
->rd
->overload
) {
6911 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
6913 update_next_balance(sd
, 0, &next_balance
);
6920 * Drop the rq->lock, but keep IRQ/preempt disabled.
6922 raw_spin_unlock(&this_rq
->lock
);
6924 update_blocked_averages(this_cpu
);
6926 for_each_domain(this_cpu
, sd
) {
6927 int continue_balancing
= 1;
6928 u64 t0
, domain_cost
;
6930 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6933 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
6934 update_next_balance(sd
, 0, &next_balance
);
6938 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
6939 t0
= sched_clock_cpu(this_cpu
);
6941 pulled_task
= load_balance(this_cpu
, this_rq
,
6943 &continue_balancing
);
6945 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
6946 if (domain_cost
> sd
->max_newidle_lb_cost
)
6947 sd
->max_newidle_lb_cost
= domain_cost
;
6949 curr_cost
+= domain_cost
;
6952 update_next_balance(sd
, 0, &next_balance
);
6955 * Stop searching for tasks to pull if there are
6956 * now runnable tasks on this rq.
6958 if (pulled_task
|| this_rq
->nr_running
> 0)
6963 raw_spin_lock(&this_rq
->lock
);
6965 if (curr_cost
> this_rq
->max_idle_balance_cost
)
6966 this_rq
->max_idle_balance_cost
= curr_cost
;
6969 * While browsing the domains, we released the rq lock, a task could
6970 * have been enqueued in the meantime. Since we're not going idle,
6971 * pretend we pulled a task.
6973 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
6977 /* Move the next balance forward */
6978 if (time_after(this_rq
->next_balance
, next_balance
))
6979 this_rq
->next_balance
= next_balance
;
6981 /* Is there a task of a high priority class? */
6982 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
6986 idle_exit_fair(this_rq
);
6987 this_rq
->idle_stamp
= 0;
6994 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6995 * running tasks off the busiest CPU onto idle CPUs. It requires at
6996 * least 1 task to be running on each physical CPU where possible, and
6997 * avoids physical / logical imbalances.
6999 static int active_load_balance_cpu_stop(void *data
)
7001 struct rq
*busiest_rq
= data
;
7002 int busiest_cpu
= cpu_of(busiest_rq
);
7003 int target_cpu
= busiest_rq
->push_cpu
;
7004 struct rq
*target_rq
= cpu_rq(target_cpu
);
7005 struct sched_domain
*sd
;
7006 struct task_struct
*p
= NULL
;
7008 raw_spin_lock_irq(&busiest_rq
->lock
);
7010 /* make sure the requested cpu hasn't gone down in the meantime */
7011 if (unlikely(busiest_cpu
!= smp_processor_id() ||
7012 !busiest_rq
->active_balance
))
7015 /* Is there any task to move? */
7016 if (busiest_rq
->nr_running
<= 1)
7020 * This condition is "impossible", if it occurs
7021 * we need to fix it. Originally reported by
7022 * Bjorn Helgaas on a 128-cpu setup.
7024 BUG_ON(busiest_rq
== target_rq
);
7026 /* Search for an sd spanning us and the target CPU. */
7028 for_each_domain(target_cpu
, sd
) {
7029 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
7030 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
7035 struct lb_env env
= {
7037 .dst_cpu
= target_cpu
,
7038 .dst_rq
= target_rq
,
7039 .src_cpu
= busiest_rq
->cpu
,
7040 .src_rq
= busiest_rq
,
7044 schedstat_inc(sd
, alb_count
);
7046 p
= detach_one_task(&env
);
7048 schedstat_inc(sd
, alb_pushed
);
7050 schedstat_inc(sd
, alb_failed
);
7054 busiest_rq
->active_balance
= 0;
7055 raw_spin_unlock(&busiest_rq
->lock
);
7058 attach_one_task(target_rq
, p
);
7065 static inline int on_null_domain(struct rq
*rq
)
7067 return unlikely(!rcu_dereference_sched(rq
->sd
));
7070 #ifdef CONFIG_NO_HZ_COMMON
7072 * idle load balancing details
7073 * - When one of the busy CPUs notice that there may be an idle rebalancing
7074 * needed, they will kick the idle load balancer, which then does idle
7075 * load balancing for all the idle CPUs.
7078 cpumask_var_t idle_cpus_mask
;
7080 unsigned long next_balance
; /* in jiffy units */
7081 } nohz ____cacheline_aligned
;
7083 static inline int find_new_ilb(void)
7085 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
7087 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
7094 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7095 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7096 * CPU (if there is one).
7098 static void nohz_balancer_kick(void)
7102 nohz
.next_balance
++;
7104 ilb_cpu
= find_new_ilb();
7106 if (ilb_cpu
>= nr_cpu_ids
)
7109 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
7112 * Use smp_send_reschedule() instead of resched_cpu().
7113 * This way we generate a sched IPI on the target cpu which
7114 * is idle. And the softirq performing nohz idle load balance
7115 * will be run before returning from the IPI.
7117 smp_send_reschedule(ilb_cpu
);
7121 static inline void nohz_balance_exit_idle(int cpu
)
7123 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
7125 * Completely isolated CPUs don't ever set, so we must test.
7127 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
7128 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
7129 atomic_dec(&nohz
.nr_cpus
);
7131 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7135 static inline void set_cpu_sd_state_busy(void)
7137 struct sched_domain
*sd
;
7138 int cpu
= smp_processor_id();
7141 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7143 if (!sd
|| !sd
->nohz_idle
)
7147 atomic_inc(&sd
->groups
->sgc
->nr_busy_cpus
);
7152 void set_cpu_sd_state_idle(void)
7154 struct sched_domain
*sd
;
7155 int cpu
= smp_processor_id();
7158 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7160 if (!sd
|| sd
->nohz_idle
)
7164 atomic_dec(&sd
->groups
->sgc
->nr_busy_cpus
);
7170 * This routine will record that the cpu is going idle with tick stopped.
7171 * This info will be used in performing idle load balancing in the future.
7173 void nohz_balance_enter_idle(int cpu
)
7176 * If this cpu is going down, then nothing needs to be done.
7178 if (!cpu_active(cpu
))
7181 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
7185 * If we're a completely isolated CPU, we don't play.
7187 if (on_null_domain(cpu_rq(cpu
)))
7190 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
7191 atomic_inc(&nohz
.nr_cpus
);
7192 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7195 static int sched_ilb_notifier(struct notifier_block
*nfb
,
7196 unsigned long action
, void *hcpu
)
7198 switch (action
& ~CPU_TASKS_FROZEN
) {
7200 nohz_balance_exit_idle(smp_processor_id());
7208 static DEFINE_SPINLOCK(balancing
);
7211 * Scale the max load_balance interval with the number of CPUs in the system.
7212 * This trades load-balance latency on larger machines for less cross talk.
7214 void update_max_interval(void)
7216 max_load_balance_interval
= HZ
*num_online_cpus()/10;
7220 * It checks each scheduling domain to see if it is due to be balanced,
7221 * and initiates a balancing operation if so.
7223 * Balancing parameters are set up in init_sched_domains.
7225 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
7227 int continue_balancing
= 1;
7229 unsigned long interval
;
7230 struct sched_domain
*sd
;
7231 /* Earliest time when we have to do rebalance again */
7232 unsigned long next_balance
= jiffies
+ 60*HZ
;
7233 int update_next_balance
= 0;
7234 int need_serialize
, need_decay
= 0;
7237 update_blocked_averages(cpu
);
7240 for_each_domain(cpu
, sd
) {
7242 * Decay the newidle max times here because this is a regular
7243 * visit to all the domains. Decay ~1% per second.
7245 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
7246 sd
->max_newidle_lb_cost
=
7247 (sd
->max_newidle_lb_cost
* 253) / 256;
7248 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
7251 max_cost
+= sd
->max_newidle_lb_cost
;
7253 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7257 * Stop the load balance at this level. There is another
7258 * CPU in our sched group which is doing load balancing more
7261 if (!continue_balancing
) {
7267 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7269 need_serialize
= sd
->flags
& SD_SERIALIZE
;
7270 if (need_serialize
) {
7271 if (!spin_trylock(&balancing
))
7275 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
7276 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
7278 * The LBF_DST_PINNED logic could have changed
7279 * env->dst_cpu, so we can't know our idle
7280 * state even if we migrated tasks. Update it.
7282 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
7284 sd
->last_balance
= jiffies
;
7285 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7288 spin_unlock(&balancing
);
7290 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
7291 next_balance
= sd
->last_balance
+ interval
;
7292 update_next_balance
= 1;
7297 * Ensure the rq-wide value also decays but keep it at a
7298 * reasonable floor to avoid funnies with rq->avg_idle.
7300 rq
->max_idle_balance_cost
=
7301 max((u64
)sysctl_sched_migration_cost
, max_cost
);
7306 * next_balance will be updated only when there is a need.
7307 * When the cpu is attached to null domain for ex, it will not be
7310 if (likely(update_next_balance
))
7311 rq
->next_balance
= next_balance
;
7314 #ifdef CONFIG_NO_HZ_COMMON
7316 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7317 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7319 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
7321 int this_cpu
= this_rq
->cpu
;
7325 if (idle
!= CPU_IDLE
||
7326 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
7329 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
7330 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
7334 * If this cpu gets work to do, stop the load balancing
7335 * work being done for other cpus. Next load
7336 * balancing owner will pick it up.
7341 rq
= cpu_rq(balance_cpu
);
7344 * If time for next balance is due,
7347 if (time_after_eq(jiffies
, rq
->next_balance
)) {
7348 raw_spin_lock_irq(&rq
->lock
);
7349 update_rq_clock(rq
);
7350 update_idle_cpu_load(rq
);
7351 raw_spin_unlock_irq(&rq
->lock
);
7352 rebalance_domains(rq
, CPU_IDLE
);
7355 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
7356 this_rq
->next_balance
= rq
->next_balance
;
7358 nohz
.next_balance
= this_rq
->next_balance
;
7360 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
7364 * Current heuristic for kicking the idle load balancer in the presence
7365 * of an idle cpu is the system.
7366 * - This rq has more than one task.
7367 * - At any scheduler domain level, this cpu's scheduler group has multiple
7368 * busy cpu's exceeding the group's capacity.
7369 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7370 * domain span are idle.
7372 static inline int nohz_kick_needed(struct rq
*rq
)
7374 unsigned long now
= jiffies
;
7375 struct sched_domain
*sd
;
7376 struct sched_group_capacity
*sgc
;
7377 int nr_busy
, cpu
= rq
->cpu
;
7379 if (unlikely(rq
->idle_balance
))
7383 * We may be recently in ticked or tickless idle mode. At the first
7384 * busy tick after returning from idle, we will update the busy stats.
7386 set_cpu_sd_state_busy();
7387 nohz_balance_exit_idle(cpu
);
7390 * None are in tickless mode and hence no need for NOHZ idle load
7393 if (likely(!atomic_read(&nohz
.nr_cpus
)))
7396 if (time_before(now
, nohz
.next_balance
))
7399 if (rq
->nr_running
>= 2)
7403 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7406 sgc
= sd
->groups
->sgc
;
7407 nr_busy
= atomic_read(&sgc
->nr_busy_cpus
);
7410 goto need_kick_unlock
;
7413 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
7415 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
7416 sched_domain_span(sd
)) < cpu
))
7417 goto need_kick_unlock
;
7428 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
7432 * run_rebalance_domains is triggered when needed from the scheduler tick.
7433 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7435 static void run_rebalance_domains(struct softirq_action
*h
)
7437 struct rq
*this_rq
= this_rq();
7438 enum cpu_idle_type idle
= this_rq
->idle_balance
?
7439 CPU_IDLE
: CPU_NOT_IDLE
;
7441 rebalance_domains(this_rq
, idle
);
7444 * If this cpu has a pending nohz_balance_kick, then do the
7445 * balancing on behalf of the other idle cpus whose ticks are
7448 nohz_idle_balance(this_rq
, idle
);
7452 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7454 void trigger_load_balance(struct rq
*rq
)
7456 /* Don't need to rebalance while attached to NULL domain */
7457 if (unlikely(on_null_domain(rq
)))
7460 if (time_after_eq(jiffies
, rq
->next_balance
))
7461 raise_softirq(SCHED_SOFTIRQ
);
7462 #ifdef CONFIG_NO_HZ_COMMON
7463 if (nohz_kick_needed(rq
))
7464 nohz_balancer_kick();
7468 static void rq_online_fair(struct rq
*rq
)
7472 update_runtime_enabled(rq
);
7475 static void rq_offline_fair(struct rq
*rq
)
7479 /* Ensure any throttled groups are reachable by pick_next_task */
7480 unthrottle_offline_cfs_rqs(rq
);
7483 #endif /* CONFIG_SMP */
7486 * scheduler tick hitting a task of our scheduling class:
7488 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
7490 struct cfs_rq
*cfs_rq
;
7491 struct sched_entity
*se
= &curr
->se
;
7493 for_each_sched_entity(se
) {
7494 cfs_rq
= cfs_rq_of(se
);
7495 entity_tick(cfs_rq
, se
, queued
);
7498 if (numabalancing_enabled
)
7499 task_tick_numa(rq
, curr
);
7501 update_rq_runnable_avg(rq
, 1);
7505 * called on fork with the child task as argument from the parent's context
7506 * - child not yet on the tasklist
7507 * - preemption disabled
7509 static void task_fork_fair(struct task_struct
*p
)
7511 struct cfs_rq
*cfs_rq
;
7512 struct sched_entity
*se
= &p
->se
, *curr
;
7513 int this_cpu
= smp_processor_id();
7514 struct rq
*rq
= this_rq();
7515 unsigned long flags
;
7517 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7519 update_rq_clock(rq
);
7521 cfs_rq
= task_cfs_rq(current
);
7522 curr
= cfs_rq
->curr
;
7525 * Not only the cpu but also the task_group of the parent might have
7526 * been changed after parent->se.parent,cfs_rq were copied to
7527 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7528 * of child point to valid ones.
7531 __set_task_cpu(p
, this_cpu
);
7534 update_curr(cfs_rq
);
7537 se
->vruntime
= curr
->vruntime
;
7538 place_entity(cfs_rq
, se
, 1);
7540 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
7542 * Upon rescheduling, sched_class::put_prev_task() will place
7543 * 'current' within the tree based on its new key value.
7545 swap(curr
->vruntime
, se
->vruntime
);
7549 se
->vruntime
-= cfs_rq
->min_vruntime
;
7551 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7555 * Priority of the task has changed. Check to see if we preempt
7559 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
7561 if (!task_on_rq_queued(p
))
7565 * Reschedule if we are currently running on this runqueue and
7566 * our priority decreased, or if we are not currently running on
7567 * this runqueue and our priority is higher than the current's
7569 if (rq
->curr
== p
) {
7570 if (p
->prio
> oldprio
)
7573 check_preempt_curr(rq
, p
, 0);
7576 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
7578 struct sched_entity
*se
= &p
->se
;
7579 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7582 * Ensure the task's vruntime is normalized, so that when it's
7583 * switched back to the fair class the enqueue_entity(.flags=0) will
7584 * do the right thing.
7586 * If it's queued, then the dequeue_entity(.flags=0) will already
7587 * have normalized the vruntime, if it's !queued, then only when
7588 * the task is sleeping will it still have non-normalized vruntime.
7590 if (!task_on_rq_queued(p
) && p
->state
!= TASK_RUNNING
) {
7592 * Fix up our vruntime so that the current sleep doesn't
7593 * cause 'unlimited' sleep bonus.
7595 place_entity(cfs_rq
, se
, 0);
7596 se
->vruntime
-= cfs_rq
->min_vruntime
;
7601 * Remove our load from contribution when we leave sched_fair
7602 * and ensure we don't carry in an old decay_count if we
7605 if (se
->avg
.decay_count
) {
7606 __synchronize_entity_decay(se
);
7607 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7613 * We switched to the sched_fair class.
7615 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7617 #ifdef CONFIG_FAIR_GROUP_SCHED
7618 struct sched_entity
*se
= &p
->se
;
7620 * Since the real-depth could have been changed (only FAIR
7621 * class maintain depth value), reset depth properly.
7623 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7625 if (!task_on_rq_queued(p
))
7629 * We were most likely switched from sched_rt, so
7630 * kick off the schedule if running, otherwise just see
7631 * if we can still preempt the current task.
7636 check_preempt_curr(rq
, p
, 0);
7639 /* Account for a task changing its policy or group.
7641 * This routine is mostly called to set cfs_rq->curr field when a task
7642 * migrates between groups/classes.
7644 static void set_curr_task_fair(struct rq
*rq
)
7646 struct sched_entity
*se
= &rq
->curr
->se
;
7648 for_each_sched_entity(se
) {
7649 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7651 set_next_entity(cfs_rq
, se
);
7652 /* ensure bandwidth has been allocated on our new cfs_rq */
7653 account_cfs_rq_runtime(cfs_rq
, 0);
7657 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7659 cfs_rq
->tasks_timeline
= RB_ROOT
;
7660 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7661 #ifndef CONFIG_64BIT
7662 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7665 atomic64_set(&cfs_rq
->decay_counter
, 1);
7666 atomic_long_set(&cfs_rq
->removed_load
, 0);
7670 #ifdef CONFIG_FAIR_GROUP_SCHED
7671 static void task_move_group_fair(struct task_struct
*p
, int queued
)
7673 struct sched_entity
*se
= &p
->se
;
7674 struct cfs_rq
*cfs_rq
;
7677 * If the task was not on the rq at the time of this cgroup movement
7678 * it must have been asleep, sleeping tasks keep their ->vruntime
7679 * absolute on their old rq until wakeup (needed for the fair sleeper
7680 * bonus in place_entity()).
7682 * If it was on the rq, we've just 'preempted' it, which does convert
7683 * ->vruntime to a relative base.
7685 * Make sure both cases convert their relative position when migrating
7686 * to another cgroup's rq. This does somewhat interfere with the
7687 * fair sleeper stuff for the first placement, but who cares.
7690 * When !queued, vruntime of the task has usually NOT been normalized.
7691 * But there are some cases where it has already been normalized:
7693 * - Moving a forked child which is waiting for being woken up by
7694 * wake_up_new_task().
7695 * - Moving a task which has been woken up by try_to_wake_up() and
7696 * waiting for actually being woken up by sched_ttwu_pending().
7698 * To prevent boost or penalty in the new cfs_rq caused by delta
7699 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7701 if (!queued
&& (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7705 se
->vruntime
-= cfs_rq_of(se
)->min_vruntime
;
7706 set_task_rq(p
, task_cpu(p
));
7707 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7709 cfs_rq
= cfs_rq_of(se
);
7710 se
->vruntime
+= cfs_rq
->min_vruntime
;
7713 * migrate_task_rq_fair() will have removed our previous
7714 * contribution, but we must synchronize for ongoing future
7717 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7718 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
7723 void free_fair_sched_group(struct task_group
*tg
)
7727 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7729 for_each_possible_cpu(i
) {
7731 kfree(tg
->cfs_rq
[i
]);
7740 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7742 struct cfs_rq
*cfs_rq
;
7743 struct sched_entity
*se
;
7746 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7749 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7753 tg
->shares
= NICE_0_LOAD
;
7755 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7757 for_each_possible_cpu(i
) {
7758 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7759 GFP_KERNEL
, cpu_to_node(i
));
7763 se
= kzalloc_node(sizeof(struct sched_entity
),
7764 GFP_KERNEL
, cpu_to_node(i
));
7768 init_cfs_rq(cfs_rq
);
7769 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
7780 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7782 struct rq
*rq
= cpu_rq(cpu
);
7783 unsigned long flags
;
7786 * Only empty task groups can be destroyed; so we can speculatively
7787 * check on_list without danger of it being re-added.
7789 if (!tg
->cfs_rq
[cpu
]->on_list
)
7792 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7793 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
7794 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7797 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7798 struct sched_entity
*se
, int cpu
,
7799 struct sched_entity
*parent
)
7801 struct rq
*rq
= cpu_rq(cpu
);
7805 init_cfs_rq_runtime(cfs_rq
);
7807 tg
->cfs_rq
[cpu
] = cfs_rq
;
7810 /* se could be NULL for root_task_group */
7815 se
->cfs_rq
= &rq
->cfs
;
7818 se
->cfs_rq
= parent
->my_q
;
7819 se
->depth
= parent
->depth
+ 1;
7823 /* guarantee group entities always have weight */
7824 update_load_set(&se
->load
, NICE_0_LOAD
);
7825 se
->parent
= parent
;
7828 static DEFINE_MUTEX(shares_mutex
);
7830 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7833 unsigned long flags
;
7836 * We can't change the weight of the root cgroup.
7841 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
7843 mutex_lock(&shares_mutex
);
7844 if (tg
->shares
== shares
)
7847 tg
->shares
= shares
;
7848 for_each_possible_cpu(i
) {
7849 struct rq
*rq
= cpu_rq(i
);
7850 struct sched_entity
*se
;
7853 /* Propagate contribution to hierarchy */
7854 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7856 /* Possible calls to update_curr() need rq clock */
7857 update_rq_clock(rq
);
7858 for_each_sched_entity(se
)
7859 update_cfs_shares(group_cfs_rq(se
));
7860 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7864 mutex_unlock(&shares_mutex
);
7867 #else /* CONFIG_FAIR_GROUP_SCHED */
7869 void free_fair_sched_group(struct task_group
*tg
) { }
7871 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7876 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
7878 #endif /* CONFIG_FAIR_GROUP_SCHED */
7881 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
7883 struct sched_entity
*se
= &task
->se
;
7884 unsigned int rr_interval
= 0;
7887 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7890 if (rq
->cfs
.load
.weight
)
7891 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
7897 * All the scheduling class methods:
7899 const struct sched_class fair_sched_class
= {
7900 .next
= &idle_sched_class
,
7901 .enqueue_task
= enqueue_task_fair
,
7902 .dequeue_task
= dequeue_task_fair
,
7903 .yield_task
= yield_task_fair
,
7904 .yield_to_task
= yield_to_task_fair
,
7906 .check_preempt_curr
= check_preempt_wakeup
,
7908 .pick_next_task
= pick_next_task_fair
,
7909 .put_prev_task
= put_prev_task_fair
,
7912 .select_task_rq
= select_task_rq_fair
,
7913 .migrate_task_rq
= migrate_task_rq_fair
,
7915 .rq_online
= rq_online_fair
,
7916 .rq_offline
= rq_offline_fair
,
7918 .task_waking
= task_waking_fair
,
7921 .set_curr_task
= set_curr_task_fair
,
7922 .task_tick
= task_tick_fair
,
7923 .task_fork
= task_fork_fair
,
7925 .prio_changed
= prio_changed_fair
,
7926 .switched_from
= switched_from_fair
,
7927 .switched_to
= switched_to_fair
,
7929 .get_rr_interval
= get_rr_interval_fair
,
7931 #ifdef CONFIG_FAIR_GROUP_SCHED
7932 .task_move_group
= task_move_group_fair
,
7936 #ifdef CONFIG_SCHED_DEBUG
7937 void print_cfs_stats(struct seq_file
*m
, int cpu
)
7939 struct cfs_rq
*cfs_rq
;
7942 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
7943 print_cfs_rq(m
, cpu
, cfs_rq
);
7948 __init
void init_sched_fair_class(void)
7951 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7953 #ifdef CONFIG_NO_HZ_COMMON
7954 nohz
.next_balance
= jiffies
;
7955 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7956 cpu_notifier(sched_ilb_notifier
, 0);