2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency
= 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG
;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity
= 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency
= 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly
;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
93 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window
= 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
116 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
122 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
128 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling
) {
149 case SCHED_TUNABLESCALING_NONE
:
152 case SCHED_TUNABLESCALING_LINEAR
:
155 case SCHED_TUNABLESCALING_LOG
:
157 factor
= 1 + ilog2(cpus
);
164 static void update_sysctl(void)
166 unsigned int factor
= get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity
);
171 SET_SYSCTL(sched_latency
);
172 SET_SYSCTL(sched_wakeup_granularity
);
176 void sched_init_granularity(void)
181 #define WMULT_CONST (~0U)
182 #define WMULT_SHIFT 32
184 static void __update_inv_weight(struct load_weight
*lw
)
188 if (likely(lw
->inv_weight
))
191 w
= scale_load_down(lw
->weight
);
193 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
195 else if (unlikely(!w
))
196 lw
->inv_weight
= WMULT_CONST
;
198 lw
->inv_weight
= WMULT_CONST
/ w
;
202 * delta_exec * weight / lw.weight
204 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
206 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
207 * we're guaranteed shift stays positive because inv_weight is guaranteed to
208 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
210 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
211 * weight/lw.weight <= 1, and therefore our shift will also be positive.
213 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
215 u64 fact
= scale_load_down(weight
);
216 int shift
= WMULT_SHIFT
;
218 __update_inv_weight(lw
);
220 if (unlikely(fact
>> 32)) {
227 /* hint to use a 32x32->64 mul */
228 fact
= (u64
)(u32
)fact
* lw
->inv_weight
;
235 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
239 const struct sched_class fair_sched_class
;
241 /**************************************************************
242 * CFS operations on generic schedulable entities:
245 #ifdef CONFIG_FAIR_GROUP_SCHED
247 /* cpu runqueue to which this cfs_rq is attached */
248 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
253 /* An entity is a task if it doesn't "own" a runqueue */
254 #define entity_is_task(se) (!se->my_q)
256 static inline struct task_struct
*task_of(struct sched_entity
*se
)
258 #ifdef CONFIG_SCHED_DEBUG
259 WARN_ON_ONCE(!entity_is_task(se
));
261 return container_of(se
, struct task_struct
, se
);
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
268 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
285 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
288 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
290 if (!cfs_rq
->on_list
) {
292 * Ensure we either appear before our parent (if already
293 * enqueued) or force our parent to appear after us when it is
294 * enqueued. The fact that we always enqueue bottom-up
295 * reduces this to two cases.
297 if (cfs_rq
->tg
->parent
&&
298 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
299 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
300 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
302 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
303 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
307 /* We should have no load, but we need to update last_decay. */
308 update_cfs_rq_blocked_load(cfs_rq
, 0);
312 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
314 if (cfs_rq
->on_list
) {
315 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
320 /* Iterate thr' all leaf cfs_rq's on a runqueue */
321 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
322 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
324 /* Do the two (enqueued) entities belong to the same group ? */
325 static inline struct cfs_rq
*
326 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
328 if (se
->cfs_rq
== pse
->cfs_rq
)
334 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
340 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
342 int se_depth
, pse_depth
;
345 * preemption test can be made between sibling entities who are in the
346 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
347 * both tasks until we find their ancestors who are siblings of common
351 /* First walk up until both entities are at same depth */
352 se_depth
= (*se
)->depth
;
353 pse_depth
= (*pse
)->depth
;
355 while (se_depth
> pse_depth
) {
357 *se
= parent_entity(*se
);
360 while (pse_depth
> se_depth
) {
362 *pse
= parent_entity(*pse
);
365 while (!is_same_group(*se
, *pse
)) {
366 *se
= parent_entity(*se
);
367 *pse
= parent_entity(*pse
);
371 #else /* !CONFIG_FAIR_GROUP_SCHED */
373 static inline struct task_struct
*task_of(struct sched_entity
*se
)
375 return container_of(se
, struct task_struct
, se
);
378 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
380 return container_of(cfs_rq
, struct rq
, cfs
);
383 #define entity_is_task(se) 1
385 #define for_each_sched_entity(se) \
386 for (; se; se = NULL)
388 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
390 return &task_rq(p
)->cfs
;
393 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
395 struct task_struct
*p
= task_of(se
);
396 struct rq
*rq
= task_rq(p
);
401 /* runqueue "owned" by this group */
402 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
407 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
411 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
415 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
416 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
418 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
424 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
428 #endif /* CONFIG_FAIR_GROUP_SCHED */
430 static __always_inline
431 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
433 /**************************************************************
434 * Scheduling class tree data structure manipulation methods:
437 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
439 s64 delta
= (s64
)(vruntime
- max_vruntime
);
441 max_vruntime
= vruntime
;
446 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
448 s64 delta
= (s64
)(vruntime
- min_vruntime
);
450 min_vruntime
= vruntime
;
455 static inline int entity_before(struct sched_entity
*a
,
456 struct sched_entity
*b
)
458 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
461 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
463 u64 vruntime
= cfs_rq
->min_vruntime
;
466 vruntime
= cfs_rq
->curr
->vruntime
;
468 if (cfs_rq
->rb_leftmost
) {
469 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
474 vruntime
= se
->vruntime
;
476 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
479 /* ensure we never gain time by being placed backwards. */
480 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
483 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
488 * Enqueue an entity into the rb-tree:
490 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
492 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
493 struct rb_node
*parent
= NULL
;
494 struct sched_entity
*entry
;
498 * Find the right place in the rbtree:
502 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
504 * We dont care about collisions. Nodes with
505 * the same key stay together.
507 if (entity_before(se
, entry
)) {
508 link
= &parent
->rb_left
;
510 link
= &parent
->rb_right
;
516 * Maintain a cache of leftmost tree entries (it is frequently
520 cfs_rq
->rb_leftmost
= &se
->run_node
;
522 rb_link_node(&se
->run_node
, parent
, link
);
523 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
526 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
528 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
529 struct rb_node
*next_node
;
531 next_node
= rb_next(&se
->run_node
);
532 cfs_rq
->rb_leftmost
= next_node
;
535 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
538 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
540 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
545 return rb_entry(left
, struct sched_entity
, run_node
);
548 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
550 struct rb_node
*next
= rb_next(&se
->run_node
);
555 return rb_entry(next
, struct sched_entity
, run_node
);
558 #ifdef CONFIG_SCHED_DEBUG
559 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
561 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
566 return rb_entry(last
, struct sched_entity
, run_node
);
569 /**************************************************************
570 * Scheduling class statistics methods:
573 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
574 void __user
*buffer
, size_t *lenp
,
577 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
578 int factor
= get_update_sysctl_factor();
583 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
584 sysctl_sched_min_granularity
);
586 #define WRT_SYSCTL(name) \
587 (normalized_sysctl_##name = sysctl_##name / (factor))
588 WRT_SYSCTL(sched_min_granularity
);
589 WRT_SYSCTL(sched_latency
);
590 WRT_SYSCTL(sched_wakeup_granularity
);
600 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
602 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
603 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
609 * The idea is to set a period in which each task runs once.
611 * When there are too many tasks (sched_nr_latency) we have to stretch
612 * this period because otherwise the slices get too small.
614 * p = (nr <= nl) ? l : l*nr/nl
616 static u64
__sched_period(unsigned long nr_running
)
618 u64 period
= sysctl_sched_latency
;
619 unsigned long nr_latency
= sched_nr_latency
;
621 if (unlikely(nr_running
> nr_latency
)) {
622 period
= sysctl_sched_min_granularity
;
623 period
*= nr_running
;
630 * We calculate the wall-time slice from the period by taking a part
631 * proportional to the weight.
635 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
637 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
639 for_each_sched_entity(se
) {
640 struct load_weight
*load
;
641 struct load_weight lw
;
643 cfs_rq
= cfs_rq_of(se
);
644 load
= &cfs_rq
->load
;
646 if (unlikely(!se
->on_rq
)) {
649 update_load_add(&lw
, se
->load
.weight
);
652 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
658 * We calculate the vruntime slice of a to-be-inserted task.
662 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
664 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
668 static unsigned long task_h_load(struct task_struct
*p
);
670 static inline void __update_task_entity_contrib(struct sched_entity
*se
);
672 /* Give new task start runnable values to heavy its load in infant time */
673 void init_task_runnable_average(struct task_struct
*p
)
677 p
->se
.avg
.decay_count
= 0;
678 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
679 p
->se
.avg
.runnable_avg_sum
= slice
;
680 p
->se
.avg
.runnable_avg_period
= slice
;
681 __update_task_entity_contrib(&p
->se
);
684 void init_task_runnable_average(struct task_struct
*p
)
690 * Update the current task's runtime statistics.
692 static void update_curr(struct cfs_rq
*cfs_rq
)
694 struct sched_entity
*curr
= cfs_rq
->curr
;
695 u64 now
= rq_clock_task(rq_of(cfs_rq
));
701 delta_exec
= now
- curr
->exec_start
;
702 if (unlikely((s64
)delta_exec
<= 0))
705 curr
->exec_start
= now
;
707 schedstat_set(curr
->statistics
.exec_max
,
708 max(delta_exec
, curr
->statistics
.exec_max
));
710 curr
->sum_exec_runtime
+= delta_exec
;
711 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
713 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
714 update_min_vruntime(cfs_rq
);
716 if (entity_is_task(curr
)) {
717 struct task_struct
*curtask
= task_of(curr
);
719 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
720 cpuacct_charge(curtask
, delta_exec
);
721 account_group_exec_runtime(curtask
, delta_exec
);
724 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
728 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
730 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
734 * Task is being enqueued - update stats:
736 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
739 * Are we enqueueing a waiting task? (for current tasks
740 * a dequeue/enqueue event is a NOP)
742 if (se
!= cfs_rq
->curr
)
743 update_stats_wait_start(cfs_rq
, se
);
747 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
749 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
750 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
751 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
752 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
753 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
754 #ifdef CONFIG_SCHEDSTATS
755 if (entity_is_task(se
)) {
756 trace_sched_stat_wait(task_of(se
),
757 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
760 schedstat_set(se
->statistics
.wait_start
, 0);
764 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
767 * Mark the end of the wait period if dequeueing a
770 if (se
!= cfs_rq
->curr
)
771 update_stats_wait_end(cfs_rq
, se
);
775 * We are picking a new current task - update its stats:
778 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
781 * We are starting a new run period:
783 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
786 /**************************************************
787 * Scheduling class queueing methods:
790 #ifdef CONFIG_NUMA_BALANCING
792 * Approximate time to scan a full NUMA task in ms. The task scan period is
793 * calculated based on the tasks virtual memory size and
794 * numa_balancing_scan_size.
796 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
797 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
799 /* Portion of address space to scan in MB */
800 unsigned int sysctl_numa_balancing_scan_size
= 256;
802 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
803 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
805 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
807 unsigned long rss
= 0;
808 unsigned long nr_scan_pages
;
811 * Calculations based on RSS as non-present and empty pages are skipped
812 * by the PTE scanner and NUMA hinting faults should be trapped based
815 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
816 rss
= get_mm_rss(p
->mm
);
820 rss
= round_up(rss
, nr_scan_pages
);
821 return rss
/ nr_scan_pages
;
824 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
825 #define MAX_SCAN_WINDOW 2560
827 static unsigned int task_scan_min(struct task_struct
*p
)
829 unsigned int scan
, floor
;
830 unsigned int windows
= 1;
832 if (sysctl_numa_balancing_scan_size
< MAX_SCAN_WINDOW
)
833 windows
= MAX_SCAN_WINDOW
/ sysctl_numa_balancing_scan_size
;
834 floor
= 1000 / windows
;
836 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
837 return max_t(unsigned int, floor
, scan
);
840 static unsigned int task_scan_max(struct task_struct
*p
)
842 unsigned int smin
= task_scan_min(p
);
845 /* Watch for min being lower than max due to floor calculations */
846 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
847 return max(smin
, smax
);
850 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
852 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
853 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
856 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
858 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
859 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
865 spinlock_t lock
; /* nr_tasks, tasks */
868 struct list_head task_list
;
871 nodemask_t active_nodes
;
872 unsigned long total_faults
;
874 * Faults_cpu is used to decide whether memory should move
875 * towards the CPU. As a consequence, these stats are weighted
876 * more by CPU use than by memory faults.
878 unsigned long *faults_cpu
;
879 unsigned long faults
[0];
882 /* Shared or private faults. */
883 #define NR_NUMA_HINT_FAULT_TYPES 2
885 /* Memory and CPU locality */
886 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
888 /* Averaged statistics, and temporary buffers. */
889 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
891 pid_t
task_numa_group_id(struct task_struct
*p
)
893 return p
->numa_group
? p
->numa_group
->gid
: 0;
896 static inline int task_faults_idx(int nid
, int priv
)
898 return NR_NUMA_HINT_FAULT_TYPES
* nid
+ priv
;
901 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
903 if (!p
->numa_faults_memory
)
906 return p
->numa_faults_memory
[task_faults_idx(nid
, 0)] +
907 p
->numa_faults_memory
[task_faults_idx(nid
, 1)];
910 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
915 return p
->numa_group
->faults
[task_faults_idx(nid
, 0)] +
916 p
->numa_group
->faults
[task_faults_idx(nid
, 1)];
919 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
921 return group
->faults_cpu
[task_faults_idx(nid
, 0)] +
922 group
->faults_cpu
[task_faults_idx(nid
, 1)];
926 * These return the fraction of accesses done by a particular task, or
927 * task group, on a particular numa node. The group weight is given a
928 * larger multiplier, in order to group tasks together that are almost
929 * evenly spread out between numa nodes.
931 static inline unsigned long task_weight(struct task_struct
*p
, int nid
)
933 unsigned long total_faults
;
935 if (!p
->numa_faults_memory
)
938 total_faults
= p
->total_numa_faults
;
943 return 1000 * task_faults(p
, nid
) / total_faults
;
946 static inline unsigned long group_weight(struct task_struct
*p
, int nid
)
948 if (!p
->numa_group
|| !p
->numa_group
->total_faults
)
951 return 1000 * group_faults(p
, nid
) / p
->numa_group
->total_faults
;
954 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
955 int src_nid
, int dst_cpu
)
957 struct numa_group
*ng
= p
->numa_group
;
958 int dst_nid
= cpu_to_node(dst_cpu
);
959 int last_cpupid
, this_cpupid
;
961 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
964 * Multi-stage node selection is used in conjunction with a periodic
965 * migration fault to build a temporal task<->page relation. By using
966 * a two-stage filter we remove short/unlikely relations.
968 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
969 * a task's usage of a particular page (n_p) per total usage of this
970 * page (n_t) (in a given time-span) to a probability.
972 * Our periodic faults will sample this probability and getting the
973 * same result twice in a row, given these samples are fully
974 * independent, is then given by P(n)^2, provided our sample period
975 * is sufficiently short compared to the usage pattern.
977 * This quadric squishes small probabilities, making it less likely we
978 * act on an unlikely task<->page relation.
980 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
981 if (!cpupid_pid_unset(last_cpupid
) &&
982 cpupid_to_nid(last_cpupid
) != dst_nid
)
985 /* Always allow migrate on private faults */
986 if (cpupid_match_pid(p
, last_cpupid
))
989 /* A shared fault, but p->numa_group has not been set up yet. */
994 * Do not migrate if the destination is not a node that
995 * is actively used by this numa group.
997 if (!node_isset(dst_nid
, ng
->active_nodes
))
1001 * Source is a node that is not actively used by this
1002 * numa group, while the destination is. Migrate.
1004 if (!node_isset(src_nid
, ng
->active_nodes
))
1008 * Both source and destination are nodes in active
1009 * use by this numa group. Maximize memory bandwidth
1010 * by migrating from more heavily used groups, to less
1011 * heavily used ones, spreading the load around.
1012 * Use a 1/4 hysteresis to avoid spurious page movement.
1014 return group_faults(p
, dst_nid
) < (group_faults(p
, src_nid
) * 3 / 4);
1017 static unsigned long weighted_cpuload(const int cpu
);
1018 static unsigned long source_load(int cpu
, int type
);
1019 static unsigned long target_load(int cpu
, int type
);
1020 static unsigned long capacity_of(int cpu
);
1021 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
1023 /* Cached statistics for all CPUs within a node */
1025 unsigned long nr_running
;
1028 /* Total compute capacity of CPUs on a node */
1029 unsigned long compute_capacity
;
1031 /* Approximate capacity in terms of runnable tasks on a node */
1032 unsigned long task_capacity
;
1033 int has_free_capacity
;
1037 * XXX borrowed from update_sg_lb_stats
1039 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1043 memset(ns
, 0, sizeof(*ns
));
1044 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1045 struct rq
*rq
= cpu_rq(cpu
);
1047 ns
->nr_running
+= rq
->nr_running
;
1048 ns
->load
+= weighted_cpuload(cpu
);
1049 ns
->compute_capacity
+= capacity_of(cpu
);
1055 * If we raced with hotplug and there are no CPUs left in our mask
1056 * the @ns structure is NULL'ed and task_numa_compare() will
1057 * not find this node attractive.
1059 * We'll either bail at !has_free_capacity, or we'll detect a huge
1060 * imbalance and bail there.
1065 ns
->load
= (ns
->load
* SCHED_CAPACITY_SCALE
) / ns
->compute_capacity
;
1067 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
);
1068 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1071 struct task_numa_env
{
1072 struct task_struct
*p
;
1074 int src_cpu
, src_nid
;
1075 int dst_cpu
, dst_nid
;
1077 struct numa_stats src_stats
, dst_stats
;
1081 struct task_struct
*best_task
;
1086 static void task_numa_assign(struct task_numa_env
*env
,
1087 struct task_struct
*p
, long imp
)
1090 put_task_struct(env
->best_task
);
1095 env
->best_imp
= imp
;
1096 env
->best_cpu
= env
->dst_cpu
;
1099 static bool load_too_imbalanced(long orig_src_load
, long orig_dst_load
,
1100 long src_load
, long dst_load
,
1101 struct task_numa_env
*env
)
1105 /* We care about the slope of the imbalance, not the direction. */
1106 if (dst_load
< src_load
)
1107 swap(dst_load
, src_load
);
1109 /* Is the difference below the threshold? */
1110 imb
= dst_load
* 100 - src_load
* env
->imbalance_pct
;
1115 * The imbalance is above the allowed threshold.
1116 * Compare it with the old imbalance.
1118 if (orig_dst_load
< orig_src_load
)
1119 swap(orig_dst_load
, orig_src_load
);
1121 old_imb
= orig_dst_load
* 100 - orig_src_load
* env
->imbalance_pct
;
1123 /* Would this change make things worse? */
1124 return (imb
> old_imb
);
1128 * This checks if the overall compute and NUMA accesses of the system would
1129 * be improved if the source tasks was migrated to the target dst_cpu taking
1130 * into account that it might be best if task running on the dst_cpu should
1131 * be exchanged with the source task
1133 static void task_numa_compare(struct task_numa_env
*env
,
1134 long taskimp
, long groupimp
)
1136 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1137 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1138 struct task_struct
*cur
;
1139 long orig_src_load
, src_load
;
1140 long orig_dst_load
, dst_load
;
1142 long imp
= (groupimp
> 0) ? groupimp
: taskimp
;
1145 cur
= ACCESS_ONCE(dst_rq
->curr
);
1146 if (cur
->pid
== 0) /* idle */
1150 * "imp" is the fault differential for the source task between the
1151 * source and destination node. Calculate the total differential for
1152 * the source task and potential destination task. The more negative
1153 * the value is, the more rmeote accesses that would be expected to
1154 * be incurred if the tasks were swapped.
1157 /* Skip this swap candidate if cannot move to the source cpu */
1158 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1162 * If dst and source tasks are in the same NUMA group, or not
1163 * in any group then look only at task weights.
1165 if (cur
->numa_group
== env
->p
->numa_group
) {
1166 imp
= taskimp
+ task_weight(cur
, env
->src_nid
) -
1167 task_weight(cur
, env
->dst_nid
);
1169 * Add some hysteresis to prevent swapping the
1170 * tasks within a group over tiny differences.
1172 if (cur
->numa_group
)
1176 * Compare the group weights. If a task is all by
1177 * itself (not part of a group), use the task weight
1180 if (env
->p
->numa_group
)
1185 if (cur
->numa_group
)
1186 imp
+= group_weight(cur
, env
->src_nid
) -
1187 group_weight(cur
, env
->dst_nid
);
1189 imp
+= task_weight(cur
, env
->src_nid
) -
1190 task_weight(cur
, env
->dst_nid
);
1194 if (imp
< env
->best_imp
)
1198 /* Is there capacity at our destination? */
1199 if (env
->src_stats
.has_free_capacity
&&
1200 !env
->dst_stats
.has_free_capacity
)
1206 /* Balance doesn't matter much if we're running a task per cpu */
1207 if (src_rq
->nr_running
== 1 && dst_rq
->nr_running
== 1)
1211 * In the overloaded case, try and keep the load balanced.
1214 orig_dst_load
= env
->dst_stats
.load
;
1215 orig_src_load
= env
->src_stats
.load
;
1217 /* XXX missing capacity terms */
1218 load
= task_h_load(env
->p
);
1219 dst_load
= orig_dst_load
+ load
;
1220 src_load
= orig_src_load
- load
;
1223 load
= task_h_load(cur
);
1228 if (load_too_imbalanced(orig_src_load
, orig_dst_load
,
1229 src_load
, dst_load
, env
))
1233 task_numa_assign(env
, cur
, imp
);
1238 static void task_numa_find_cpu(struct task_numa_env
*env
,
1239 long taskimp
, long groupimp
)
1243 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1244 /* Skip this CPU if the source task cannot migrate */
1245 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1249 task_numa_compare(env
, taskimp
, groupimp
);
1253 static int task_numa_migrate(struct task_struct
*p
)
1255 struct task_numa_env env
= {
1258 .src_cpu
= task_cpu(p
),
1259 .src_nid
= task_node(p
),
1261 .imbalance_pct
= 112,
1267 struct sched_domain
*sd
;
1268 unsigned long taskweight
, groupweight
;
1270 long taskimp
, groupimp
;
1273 * Pick the lowest SD_NUMA domain, as that would have the smallest
1274 * imbalance and would be the first to start moving tasks about.
1276 * And we want to avoid any moving of tasks about, as that would create
1277 * random movement of tasks -- counter the numa conditions we're trying
1281 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1283 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1287 * Cpusets can break the scheduler domain tree into smaller
1288 * balance domains, some of which do not cross NUMA boundaries.
1289 * Tasks that are "trapped" in such domains cannot be migrated
1290 * elsewhere, so there is no point in (re)trying.
1292 if (unlikely(!sd
)) {
1293 p
->numa_preferred_nid
= task_node(p
);
1297 taskweight
= task_weight(p
, env
.src_nid
);
1298 groupweight
= group_weight(p
, env
.src_nid
);
1299 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1300 env
.dst_nid
= p
->numa_preferred_nid
;
1301 taskimp
= task_weight(p
, env
.dst_nid
) - taskweight
;
1302 groupimp
= group_weight(p
, env
.dst_nid
) - groupweight
;
1303 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1305 /* If the preferred nid has free capacity, try to use it. */
1306 if (env
.dst_stats
.has_free_capacity
)
1307 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1309 /* No space available on the preferred nid. Look elsewhere. */
1310 if (env
.best_cpu
== -1) {
1311 for_each_online_node(nid
) {
1312 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1315 /* Only consider nodes where both task and groups benefit */
1316 taskimp
= task_weight(p
, nid
) - taskweight
;
1317 groupimp
= group_weight(p
, nid
) - groupweight
;
1318 if (taskimp
< 0 && groupimp
< 0)
1322 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1323 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1327 /* No better CPU than the current one was found. */
1328 if (env
.best_cpu
== -1)
1332 * If the task is part of a workload that spans multiple NUMA nodes,
1333 * and is migrating into one of the workload's active nodes, remember
1334 * this node as the task's preferred numa node, so the workload can
1336 * A task that migrated to a second choice node will be better off
1337 * trying for a better one later. Do not set the preferred node here.
1339 if (p
->numa_group
&& node_isset(env
.dst_nid
, p
->numa_group
->active_nodes
))
1340 sched_setnuma(p
, env
.dst_nid
);
1343 * Reset the scan period if the task is being rescheduled on an
1344 * alternative node to recheck if the tasks is now properly placed.
1346 p
->numa_scan_period
= task_scan_min(p
);
1348 if (env
.best_task
== NULL
) {
1349 ret
= migrate_task_to(p
, env
.best_cpu
);
1351 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1355 ret
= migrate_swap(p
, env
.best_task
);
1357 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1358 put_task_struct(env
.best_task
);
1362 /* Attempt to migrate a task to a CPU on the preferred node. */
1363 static void numa_migrate_preferred(struct task_struct
*p
)
1365 unsigned long interval
= HZ
;
1367 /* This task has no NUMA fault statistics yet */
1368 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults_memory
))
1371 /* Periodically retry migrating the task to the preferred node */
1372 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1373 p
->numa_migrate_retry
= jiffies
+ interval
;
1375 /* Success if task is already running on preferred CPU */
1376 if (task_node(p
) == p
->numa_preferred_nid
)
1379 /* Otherwise, try migrate to a CPU on the preferred node */
1380 task_numa_migrate(p
);
1384 * Find the nodes on which the workload is actively running. We do this by
1385 * tracking the nodes from which NUMA hinting faults are triggered. This can
1386 * be different from the set of nodes where the workload's memory is currently
1389 * The bitmask is used to make smarter decisions on when to do NUMA page
1390 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1391 * are added when they cause over 6/16 of the maximum number of faults, but
1392 * only removed when they drop below 3/16.
1394 static void update_numa_active_node_mask(struct numa_group
*numa_group
)
1396 unsigned long faults
, max_faults
= 0;
1399 for_each_online_node(nid
) {
1400 faults
= group_faults_cpu(numa_group
, nid
);
1401 if (faults
> max_faults
)
1402 max_faults
= faults
;
1405 for_each_online_node(nid
) {
1406 faults
= group_faults_cpu(numa_group
, nid
);
1407 if (!node_isset(nid
, numa_group
->active_nodes
)) {
1408 if (faults
> max_faults
* 6 / 16)
1409 node_set(nid
, numa_group
->active_nodes
);
1410 } else if (faults
< max_faults
* 3 / 16)
1411 node_clear(nid
, numa_group
->active_nodes
);
1416 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1417 * increments. The more local the fault statistics are, the higher the scan
1418 * period will be for the next scan window. If local/remote ratio is below
1419 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1420 * scan period will decrease
1422 #define NUMA_PERIOD_SLOTS 10
1423 #define NUMA_PERIOD_THRESHOLD 3
1426 * Increase the scan period (slow down scanning) if the majority of
1427 * our memory is already on our local node, or if the majority of
1428 * the page accesses are shared with other processes.
1429 * Otherwise, decrease the scan period.
1431 static void update_task_scan_period(struct task_struct
*p
,
1432 unsigned long shared
, unsigned long private)
1434 unsigned int period_slot
;
1438 unsigned long remote
= p
->numa_faults_locality
[0];
1439 unsigned long local
= p
->numa_faults_locality
[1];
1442 * If there were no record hinting faults then either the task is
1443 * completely idle or all activity is areas that are not of interest
1444 * to automatic numa balancing. Scan slower
1446 if (local
+ shared
== 0) {
1447 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1448 p
->numa_scan_period
<< 1);
1450 p
->mm
->numa_next_scan
= jiffies
+
1451 msecs_to_jiffies(p
->numa_scan_period
);
1457 * Prepare to scale scan period relative to the current period.
1458 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1459 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1460 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1462 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1463 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1464 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1465 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1468 diff
= slot
* period_slot
;
1470 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1473 * Scale scan rate increases based on sharing. There is an
1474 * inverse relationship between the degree of sharing and
1475 * the adjustment made to the scanning period. Broadly
1476 * speaking the intent is that there is little point
1477 * scanning faster if shared accesses dominate as it may
1478 * simply bounce migrations uselessly
1480 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
));
1481 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1484 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1485 task_scan_min(p
), task_scan_max(p
));
1486 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1490 * Get the fraction of time the task has been running since the last
1491 * NUMA placement cycle. The scheduler keeps similar statistics, but
1492 * decays those on a 32ms period, which is orders of magnitude off
1493 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1494 * stats only if the task is so new there are no NUMA statistics yet.
1496 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1498 u64 runtime
, delta
, now
;
1499 /* Use the start of this time slice to avoid calculations. */
1500 now
= p
->se
.exec_start
;
1501 runtime
= p
->se
.sum_exec_runtime
;
1503 if (p
->last_task_numa_placement
) {
1504 delta
= runtime
- p
->last_sum_exec_runtime
;
1505 *period
= now
- p
->last_task_numa_placement
;
1507 delta
= p
->se
.avg
.runnable_avg_sum
;
1508 *period
= p
->se
.avg
.runnable_avg_period
;
1511 p
->last_sum_exec_runtime
= runtime
;
1512 p
->last_task_numa_placement
= now
;
1517 static void task_numa_placement(struct task_struct
*p
)
1519 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1520 unsigned long max_faults
= 0, max_group_faults
= 0;
1521 unsigned long fault_types
[2] = { 0, 0 };
1522 unsigned long total_faults
;
1523 u64 runtime
, period
;
1524 spinlock_t
*group_lock
= NULL
;
1526 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1527 if (p
->numa_scan_seq
== seq
)
1529 p
->numa_scan_seq
= seq
;
1530 p
->numa_scan_period_max
= task_scan_max(p
);
1532 total_faults
= p
->numa_faults_locality
[0] +
1533 p
->numa_faults_locality
[1];
1534 runtime
= numa_get_avg_runtime(p
, &period
);
1536 /* If the task is part of a group prevent parallel updates to group stats */
1537 if (p
->numa_group
) {
1538 group_lock
= &p
->numa_group
->lock
;
1539 spin_lock_irq(group_lock
);
1542 /* Find the node with the highest number of faults */
1543 for_each_online_node(nid
) {
1544 unsigned long faults
= 0, group_faults
= 0;
1547 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
1548 long diff
, f_diff
, f_weight
;
1550 i
= task_faults_idx(nid
, priv
);
1552 /* Decay existing window, copy faults since last scan */
1553 diff
= p
->numa_faults_buffer_memory
[i
] - p
->numa_faults_memory
[i
] / 2;
1554 fault_types
[priv
] += p
->numa_faults_buffer_memory
[i
];
1555 p
->numa_faults_buffer_memory
[i
] = 0;
1558 * Normalize the faults_from, so all tasks in a group
1559 * count according to CPU use, instead of by the raw
1560 * number of faults. Tasks with little runtime have
1561 * little over-all impact on throughput, and thus their
1562 * faults are less important.
1564 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
1565 f_weight
= (f_weight
* p
->numa_faults_buffer_cpu
[i
]) /
1567 f_diff
= f_weight
- p
->numa_faults_cpu
[i
] / 2;
1568 p
->numa_faults_buffer_cpu
[i
] = 0;
1570 p
->numa_faults_memory
[i
] += diff
;
1571 p
->numa_faults_cpu
[i
] += f_diff
;
1572 faults
+= p
->numa_faults_memory
[i
];
1573 p
->total_numa_faults
+= diff
;
1574 if (p
->numa_group
) {
1575 /* safe because we can only change our own group */
1576 p
->numa_group
->faults
[i
] += diff
;
1577 p
->numa_group
->faults_cpu
[i
] += f_diff
;
1578 p
->numa_group
->total_faults
+= diff
;
1579 group_faults
+= p
->numa_group
->faults
[i
];
1583 if (faults
> max_faults
) {
1584 max_faults
= faults
;
1588 if (group_faults
> max_group_faults
) {
1589 max_group_faults
= group_faults
;
1590 max_group_nid
= nid
;
1594 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1596 if (p
->numa_group
) {
1597 update_numa_active_node_mask(p
->numa_group
);
1599 * If the preferred task and group nids are different,
1600 * iterate over the nodes again to find the best place.
1602 if (max_nid
!= max_group_nid
) {
1603 unsigned long weight
, max_weight
= 0;
1605 for_each_online_node(nid
) {
1606 weight
= task_weight(p
, nid
) + group_weight(p
, nid
);
1607 if (weight
> max_weight
) {
1608 max_weight
= weight
;
1614 spin_unlock_irq(group_lock
);
1617 /* Preferred node as the node with the most faults */
1618 if (max_faults
&& max_nid
!= p
->numa_preferred_nid
) {
1619 /* Update the preferred nid and migrate task if possible */
1620 sched_setnuma(p
, max_nid
);
1621 numa_migrate_preferred(p
);
1625 static inline int get_numa_group(struct numa_group
*grp
)
1627 return atomic_inc_not_zero(&grp
->refcount
);
1630 static inline void put_numa_group(struct numa_group
*grp
)
1632 if (atomic_dec_and_test(&grp
->refcount
))
1633 kfree_rcu(grp
, rcu
);
1636 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1639 struct numa_group
*grp
, *my_grp
;
1640 struct task_struct
*tsk
;
1642 int cpu
= cpupid_to_cpu(cpupid
);
1645 if (unlikely(!p
->numa_group
)) {
1646 unsigned int size
= sizeof(struct numa_group
) +
1647 4*nr_node_ids
*sizeof(unsigned long);
1649 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1653 atomic_set(&grp
->refcount
, 1);
1654 spin_lock_init(&grp
->lock
);
1655 INIT_LIST_HEAD(&grp
->task_list
);
1657 /* Second half of the array tracks nids where faults happen */
1658 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
1661 node_set(task_node(current
), grp
->active_nodes
);
1663 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1664 grp
->faults
[i
] = p
->numa_faults_memory
[i
];
1666 grp
->total_faults
= p
->total_numa_faults
;
1668 list_add(&p
->numa_entry
, &grp
->task_list
);
1670 rcu_assign_pointer(p
->numa_group
, grp
);
1674 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1676 if (!cpupid_match_pid(tsk
, cpupid
))
1679 grp
= rcu_dereference(tsk
->numa_group
);
1683 my_grp
= p
->numa_group
;
1688 * Only join the other group if its bigger; if we're the bigger group,
1689 * the other task will join us.
1691 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1695 * Tie-break on the grp address.
1697 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1700 /* Always join threads in the same process. */
1701 if (tsk
->mm
== current
->mm
)
1704 /* Simple filter to avoid false positives due to PID collisions */
1705 if (flags
& TNF_SHARED
)
1708 /* Update priv based on whether false sharing was detected */
1711 if (join
&& !get_numa_group(grp
))
1719 BUG_ON(irqs_disabled());
1720 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
1722 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
1723 my_grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1724 grp
->faults
[i
] += p
->numa_faults_memory
[i
];
1726 my_grp
->total_faults
-= p
->total_numa_faults
;
1727 grp
->total_faults
+= p
->total_numa_faults
;
1729 list_move(&p
->numa_entry
, &grp
->task_list
);
1733 spin_unlock(&my_grp
->lock
);
1734 spin_unlock_irq(&grp
->lock
);
1736 rcu_assign_pointer(p
->numa_group
, grp
);
1738 put_numa_group(my_grp
);
1746 void task_numa_free(struct task_struct
*p
)
1748 struct numa_group
*grp
= p
->numa_group
;
1749 void *numa_faults
= p
->numa_faults_memory
;
1750 unsigned long flags
;
1754 spin_lock_irqsave(&grp
->lock
, flags
);
1755 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1756 grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1757 grp
->total_faults
-= p
->total_numa_faults
;
1759 list_del(&p
->numa_entry
);
1761 spin_unlock_irqrestore(&grp
->lock
, flags
);
1762 rcu_assign_pointer(p
->numa_group
, NULL
);
1763 put_numa_group(grp
);
1766 p
->numa_faults_memory
= NULL
;
1767 p
->numa_faults_buffer_memory
= NULL
;
1768 p
->numa_faults_cpu
= NULL
;
1769 p
->numa_faults_buffer_cpu
= NULL
;
1774 * Got a PROT_NONE fault for a page on @node.
1776 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
1778 struct task_struct
*p
= current
;
1779 bool migrated
= flags
& TNF_MIGRATED
;
1780 int cpu_node
= task_node(current
);
1781 int local
= !!(flags
& TNF_FAULT_LOCAL
);
1784 if (!numabalancing_enabled
)
1787 /* for example, ksmd faulting in a user's mm */
1791 /* Do not worry about placement if exiting */
1792 if (p
->state
== TASK_DEAD
)
1795 /* Allocate buffer to track faults on a per-node basis */
1796 if (unlikely(!p
->numa_faults_memory
)) {
1797 int size
= sizeof(*p
->numa_faults_memory
) *
1798 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
1800 p
->numa_faults_memory
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
1801 if (!p
->numa_faults_memory
)
1804 BUG_ON(p
->numa_faults_buffer_memory
);
1806 * The averaged statistics, shared & private, memory & cpu,
1807 * occupy the first half of the array. The second half of the
1808 * array is for current counters, which are averaged into the
1809 * first set by task_numa_placement.
1811 p
->numa_faults_cpu
= p
->numa_faults_memory
+ (2 * nr_node_ids
);
1812 p
->numa_faults_buffer_memory
= p
->numa_faults_memory
+ (4 * nr_node_ids
);
1813 p
->numa_faults_buffer_cpu
= p
->numa_faults_memory
+ (6 * nr_node_ids
);
1814 p
->total_numa_faults
= 0;
1815 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1819 * First accesses are treated as private, otherwise consider accesses
1820 * to be private if the accessing pid has not changed
1822 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
1825 priv
= cpupid_match_pid(p
, last_cpupid
);
1826 if (!priv
&& !(flags
& TNF_NO_GROUP
))
1827 task_numa_group(p
, last_cpupid
, flags
, &priv
);
1831 * If a workload spans multiple NUMA nodes, a shared fault that
1832 * occurs wholly within the set of nodes that the workload is
1833 * actively using should be counted as local. This allows the
1834 * scan rate to slow down when a workload has settled down.
1836 if (!priv
&& !local
&& p
->numa_group
&&
1837 node_isset(cpu_node
, p
->numa_group
->active_nodes
) &&
1838 node_isset(mem_node
, p
->numa_group
->active_nodes
))
1841 task_numa_placement(p
);
1844 * Retry task to preferred node migration periodically, in case it
1845 * case it previously failed, or the scheduler moved us.
1847 if (time_after(jiffies
, p
->numa_migrate_retry
))
1848 numa_migrate_preferred(p
);
1851 p
->numa_pages_migrated
+= pages
;
1853 p
->numa_faults_buffer_memory
[task_faults_idx(mem_node
, priv
)] += pages
;
1854 p
->numa_faults_buffer_cpu
[task_faults_idx(cpu_node
, priv
)] += pages
;
1855 p
->numa_faults_locality
[local
] += pages
;
1858 static void reset_ptenuma_scan(struct task_struct
*p
)
1860 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
1861 p
->mm
->numa_scan_offset
= 0;
1865 * The expensive part of numa migration is done from task_work context.
1866 * Triggered from task_tick_numa().
1868 void task_numa_work(struct callback_head
*work
)
1870 unsigned long migrate
, next_scan
, now
= jiffies
;
1871 struct task_struct
*p
= current
;
1872 struct mm_struct
*mm
= p
->mm
;
1873 struct vm_area_struct
*vma
;
1874 unsigned long start
, end
;
1875 unsigned long nr_pte_updates
= 0;
1878 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
1880 work
->next
= work
; /* protect against double add */
1882 * Who cares about NUMA placement when they're dying.
1884 * NOTE: make sure not to dereference p->mm before this check,
1885 * exit_task_work() happens _after_ exit_mm() so we could be called
1886 * without p->mm even though we still had it when we enqueued this
1889 if (p
->flags
& PF_EXITING
)
1892 if (!mm
->numa_next_scan
) {
1893 mm
->numa_next_scan
= now
+
1894 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1898 * Enforce maximal scan/migration frequency..
1900 migrate
= mm
->numa_next_scan
;
1901 if (time_before(now
, migrate
))
1904 if (p
->numa_scan_period
== 0) {
1905 p
->numa_scan_period_max
= task_scan_max(p
);
1906 p
->numa_scan_period
= task_scan_min(p
);
1909 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1910 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
1914 * Delay this task enough that another task of this mm will likely win
1915 * the next time around.
1917 p
->node_stamp
+= 2 * TICK_NSEC
;
1919 start
= mm
->numa_scan_offset
;
1920 pages
= sysctl_numa_balancing_scan_size
;
1921 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
1925 down_read(&mm
->mmap_sem
);
1926 vma
= find_vma(mm
, start
);
1928 reset_ptenuma_scan(p
);
1932 for (; vma
; vma
= vma
->vm_next
) {
1933 if (!vma_migratable(vma
) || !vma_policy_mof(p
, vma
))
1937 * Shared library pages mapped by multiple processes are not
1938 * migrated as it is expected they are cache replicated. Avoid
1939 * hinting faults in read-only file-backed mappings or the vdso
1940 * as migrating the pages will be of marginal benefit.
1943 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
1947 * Skip inaccessible VMAs to avoid any confusion between
1948 * PROT_NONE and NUMA hinting ptes
1950 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
1954 start
= max(start
, vma
->vm_start
);
1955 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
1956 end
= min(end
, vma
->vm_end
);
1957 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
1960 * Scan sysctl_numa_balancing_scan_size but ensure that
1961 * at least one PTE is updated so that unused virtual
1962 * address space is quickly skipped.
1965 pages
-= (end
- start
) >> PAGE_SHIFT
;
1972 } while (end
!= vma
->vm_end
);
1977 * It is possible to reach the end of the VMA list but the last few
1978 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1979 * would find the !migratable VMA on the next scan but not reset the
1980 * scanner to the start so check it now.
1983 mm
->numa_scan_offset
= start
;
1985 reset_ptenuma_scan(p
);
1986 up_read(&mm
->mmap_sem
);
1990 * Drive the periodic memory faults..
1992 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1994 struct callback_head
*work
= &curr
->numa_work
;
1998 * We don't care about NUMA placement if we don't have memory.
2000 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2004 * Using runtime rather than walltime has the dual advantage that
2005 * we (mostly) drive the selection from busy threads and that the
2006 * task needs to have done some actual work before we bother with
2009 now
= curr
->se
.sum_exec_runtime
;
2010 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2012 if (now
- curr
->node_stamp
> period
) {
2013 if (!curr
->node_stamp
)
2014 curr
->numa_scan_period
= task_scan_min(curr
);
2015 curr
->node_stamp
+= period
;
2017 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2018 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2019 task_work_add(curr
, work
, true);
2024 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2028 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2032 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2035 #endif /* CONFIG_NUMA_BALANCING */
2038 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2040 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2041 if (!parent_entity(se
))
2042 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2044 if (entity_is_task(se
)) {
2045 struct rq
*rq
= rq_of(cfs_rq
);
2047 account_numa_enqueue(rq
, task_of(se
));
2048 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2051 cfs_rq
->nr_running
++;
2055 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2057 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2058 if (!parent_entity(se
))
2059 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2060 if (entity_is_task(se
)) {
2061 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2062 list_del_init(&se
->group_node
);
2064 cfs_rq
->nr_running
--;
2067 #ifdef CONFIG_FAIR_GROUP_SCHED
2069 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
2074 * Use this CPU's actual weight instead of the last load_contribution
2075 * to gain a more accurate current total weight. See
2076 * update_cfs_rq_load_contribution().
2078 tg_weight
= atomic_long_read(&tg
->load_avg
);
2079 tg_weight
-= cfs_rq
->tg_load_contrib
;
2080 tg_weight
+= cfs_rq
->load
.weight
;
2085 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2087 long tg_weight
, load
, shares
;
2089 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
2090 load
= cfs_rq
->load
.weight
;
2092 shares
= (tg
->shares
* load
);
2094 shares
/= tg_weight
;
2096 if (shares
< MIN_SHARES
)
2097 shares
= MIN_SHARES
;
2098 if (shares
> tg
->shares
)
2099 shares
= tg
->shares
;
2103 # else /* CONFIG_SMP */
2104 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2108 # endif /* CONFIG_SMP */
2109 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2110 unsigned long weight
)
2113 /* commit outstanding execution time */
2114 if (cfs_rq
->curr
== se
)
2115 update_curr(cfs_rq
);
2116 account_entity_dequeue(cfs_rq
, se
);
2119 update_load_set(&se
->load
, weight
);
2122 account_entity_enqueue(cfs_rq
, se
);
2125 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2127 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2129 struct task_group
*tg
;
2130 struct sched_entity
*se
;
2134 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2135 if (!se
|| throttled_hierarchy(cfs_rq
))
2138 if (likely(se
->load
.weight
== tg
->shares
))
2141 shares
= calc_cfs_shares(cfs_rq
, tg
);
2143 reweight_entity(cfs_rq_of(se
), se
, shares
);
2145 #else /* CONFIG_FAIR_GROUP_SCHED */
2146 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2149 #endif /* CONFIG_FAIR_GROUP_SCHED */
2153 * We choose a half-life close to 1 scheduling period.
2154 * Note: The tables below are dependent on this value.
2156 #define LOAD_AVG_PERIOD 32
2157 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2158 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2160 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2161 static const u32 runnable_avg_yN_inv
[] = {
2162 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2163 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2164 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2165 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2166 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2167 0x85aac367, 0x82cd8698,
2171 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2172 * over-estimates when re-combining.
2174 static const u32 runnable_avg_yN_sum
[] = {
2175 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2176 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2177 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2182 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2184 static __always_inline u64
decay_load(u64 val
, u64 n
)
2186 unsigned int local_n
;
2190 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2193 /* after bounds checking we can collapse to 32-bit */
2197 * As y^PERIOD = 1/2, we can combine
2198 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2199 * With a look-up table which covers k^n (n<PERIOD)
2201 * To achieve constant time decay_load.
2203 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2204 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2205 local_n
%= LOAD_AVG_PERIOD
;
2208 val
*= runnable_avg_yN_inv
[local_n
];
2209 /* We don't use SRR here since we always want to round down. */
2214 * For updates fully spanning n periods, the contribution to runnable
2215 * average will be: \Sum 1024*y^n
2217 * We can compute this reasonably efficiently by combining:
2218 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2220 static u32
__compute_runnable_contrib(u64 n
)
2224 if (likely(n
<= LOAD_AVG_PERIOD
))
2225 return runnable_avg_yN_sum
[n
];
2226 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2227 return LOAD_AVG_MAX
;
2229 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2231 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2232 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2234 n
-= LOAD_AVG_PERIOD
;
2235 } while (n
> LOAD_AVG_PERIOD
);
2237 contrib
= decay_load(contrib
, n
);
2238 return contrib
+ runnable_avg_yN_sum
[n
];
2242 * We can represent the historical contribution to runnable average as the
2243 * coefficients of a geometric series. To do this we sub-divide our runnable
2244 * history into segments of approximately 1ms (1024us); label the segment that
2245 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2247 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2249 * (now) (~1ms ago) (~2ms ago)
2251 * Let u_i denote the fraction of p_i that the entity was runnable.
2253 * We then designate the fractions u_i as our co-efficients, yielding the
2254 * following representation of historical load:
2255 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2257 * We choose y based on the with of a reasonably scheduling period, fixing:
2260 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2261 * approximately half as much as the contribution to load within the last ms
2264 * When a period "rolls over" and we have new u_0`, multiplying the previous
2265 * sum again by y is sufficient to update:
2266 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2267 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2269 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2270 struct sched_avg
*sa
,
2274 u32 runnable_contrib
;
2275 int delta_w
, decayed
= 0;
2277 delta
= now
- sa
->last_runnable_update
;
2279 * This should only happen when time goes backwards, which it
2280 * unfortunately does during sched clock init when we swap over to TSC.
2282 if ((s64
)delta
< 0) {
2283 sa
->last_runnable_update
= now
;
2288 * Use 1024ns as the unit of measurement since it's a reasonable
2289 * approximation of 1us and fast to compute.
2294 sa
->last_runnable_update
= now
;
2296 /* delta_w is the amount already accumulated against our next period */
2297 delta_w
= sa
->runnable_avg_period
% 1024;
2298 if (delta
+ delta_w
>= 1024) {
2299 /* period roll-over */
2303 * Now that we know we're crossing a period boundary, figure
2304 * out how much from delta we need to complete the current
2305 * period and accrue it.
2307 delta_w
= 1024 - delta_w
;
2309 sa
->runnable_avg_sum
+= delta_w
;
2310 sa
->runnable_avg_period
+= delta_w
;
2314 /* Figure out how many additional periods this update spans */
2315 periods
= delta
/ 1024;
2318 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2320 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2323 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2324 runnable_contrib
= __compute_runnable_contrib(periods
);
2326 sa
->runnable_avg_sum
+= runnable_contrib
;
2327 sa
->runnable_avg_period
+= runnable_contrib
;
2330 /* Remainder of delta accrued against u_0` */
2332 sa
->runnable_avg_sum
+= delta
;
2333 sa
->runnable_avg_period
+= delta
;
2338 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2339 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2341 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2342 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2344 decays
-= se
->avg
.decay_count
;
2348 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2349 se
->avg
.decay_count
= 0;
2354 #ifdef CONFIG_FAIR_GROUP_SCHED
2355 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2358 struct task_group
*tg
= cfs_rq
->tg
;
2361 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2362 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2364 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2365 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2366 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2371 * Aggregate cfs_rq runnable averages into an equivalent task_group
2372 * representation for computing load contributions.
2374 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2375 struct cfs_rq
*cfs_rq
)
2377 struct task_group
*tg
= cfs_rq
->tg
;
2380 /* The fraction of a cpu used by this cfs_rq */
2381 contrib
= div_u64((u64
)sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2382 sa
->runnable_avg_period
+ 1);
2383 contrib
-= cfs_rq
->tg_runnable_contrib
;
2385 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2386 atomic_add(contrib
, &tg
->runnable_avg
);
2387 cfs_rq
->tg_runnable_contrib
+= contrib
;
2391 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2393 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2394 struct task_group
*tg
= cfs_rq
->tg
;
2399 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2400 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2401 atomic_long_read(&tg
->load_avg
) + 1);
2404 * For group entities we need to compute a correction term in the case
2405 * that they are consuming <1 cpu so that we would contribute the same
2406 * load as a task of equal weight.
2408 * Explicitly co-ordinating this measurement would be expensive, but
2409 * fortunately the sum of each cpus contribution forms a usable
2410 * lower-bound on the true value.
2412 * Consider the aggregate of 2 contributions. Either they are disjoint
2413 * (and the sum represents true value) or they are disjoint and we are
2414 * understating by the aggregate of their overlap.
2416 * Extending this to N cpus, for a given overlap, the maximum amount we
2417 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2418 * cpus that overlap for this interval and w_i is the interval width.
2420 * On a small machine; the first term is well-bounded which bounds the
2421 * total error since w_i is a subset of the period. Whereas on a
2422 * larger machine, while this first term can be larger, if w_i is the
2423 * of consequential size guaranteed to see n_i*w_i quickly converge to
2424 * our upper bound of 1-cpu.
2426 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2427 if (runnable_avg
< NICE_0_LOAD
) {
2428 se
->avg
.load_avg_contrib
*= runnable_avg
;
2429 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2433 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2435 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2436 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2438 #else /* CONFIG_FAIR_GROUP_SCHED */
2439 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2440 int force_update
) {}
2441 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2442 struct cfs_rq
*cfs_rq
) {}
2443 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2444 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2445 #endif /* CONFIG_FAIR_GROUP_SCHED */
2447 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2451 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2452 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2453 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2454 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2457 /* Compute the current contribution to load_avg by se, return any delta */
2458 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2460 long old_contrib
= se
->avg
.load_avg_contrib
;
2462 if (entity_is_task(se
)) {
2463 __update_task_entity_contrib(se
);
2465 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2466 __update_group_entity_contrib(se
);
2469 return se
->avg
.load_avg_contrib
- old_contrib
;
2472 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2475 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2476 cfs_rq
->blocked_load_avg
-= load_contrib
;
2478 cfs_rq
->blocked_load_avg
= 0;
2481 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2483 /* Update a sched_entity's runnable average */
2484 static inline void update_entity_load_avg(struct sched_entity
*se
,
2487 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2492 * For a group entity we need to use their owned cfs_rq_clock_task() in
2493 * case they are the parent of a throttled hierarchy.
2495 if (entity_is_task(se
))
2496 now
= cfs_rq_clock_task(cfs_rq
);
2498 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2500 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2503 contrib_delta
= __update_entity_load_avg_contrib(se
);
2509 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2511 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2515 * Decay the load contributed by all blocked children and account this so that
2516 * their contribution may appropriately discounted when they wake up.
2518 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2520 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2523 decays
= now
- cfs_rq
->last_decay
;
2524 if (!decays
&& !force_update
)
2527 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2528 unsigned long removed_load
;
2529 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2530 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2534 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2536 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2537 cfs_rq
->last_decay
= now
;
2540 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2543 /* Add the load generated by se into cfs_rq's child load-average */
2544 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2545 struct sched_entity
*se
,
2549 * We track migrations using entity decay_count <= 0, on a wake-up
2550 * migration we use a negative decay count to track the remote decays
2551 * accumulated while sleeping.
2553 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2554 * are seen by enqueue_entity_load_avg() as a migration with an already
2555 * constructed load_avg_contrib.
2557 if (unlikely(se
->avg
.decay_count
<= 0)) {
2558 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2559 if (se
->avg
.decay_count
) {
2561 * In a wake-up migration we have to approximate the
2562 * time sleeping. This is because we can't synchronize
2563 * clock_task between the two cpus, and it is not
2564 * guaranteed to be read-safe. Instead, we can
2565 * approximate this using our carried decays, which are
2566 * explicitly atomically readable.
2568 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2570 update_entity_load_avg(se
, 0);
2571 /* Indicate that we're now synchronized and on-rq */
2572 se
->avg
.decay_count
= 0;
2576 __synchronize_entity_decay(se
);
2579 /* migrated tasks did not contribute to our blocked load */
2581 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2582 update_entity_load_avg(se
, 0);
2585 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2586 /* we force update consideration on load-balancer moves */
2587 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2591 * Remove se's load from this cfs_rq child load-average, if the entity is
2592 * transitioning to a blocked state we track its projected decay using
2595 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2596 struct sched_entity
*se
,
2599 update_entity_load_avg(se
, 1);
2600 /* we force update consideration on load-balancer moves */
2601 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2603 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2605 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2606 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2607 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2611 * Update the rq's load with the elapsed running time before entering
2612 * idle. if the last scheduled task is not a CFS task, idle_enter will
2613 * be the only way to update the runnable statistic.
2615 void idle_enter_fair(struct rq
*this_rq
)
2617 update_rq_runnable_avg(this_rq
, 1);
2621 * Update the rq's load with the elapsed idle time before a task is
2622 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2623 * be the only way to update the runnable statistic.
2625 void idle_exit_fair(struct rq
*this_rq
)
2627 update_rq_runnable_avg(this_rq
, 0);
2630 static int idle_balance(struct rq
*this_rq
);
2632 #else /* CONFIG_SMP */
2634 static inline void update_entity_load_avg(struct sched_entity
*se
,
2635 int update_cfs_rq
) {}
2636 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2637 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2638 struct sched_entity
*se
,
2640 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2641 struct sched_entity
*se
,
2643 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2644 int force_update
) {}
2646 static inline int idle_balance(struct rq
*rq
)
2651 #endif /* CONFIG_SMP */
2653 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2655 #ifdef CONFIG_SCHEDSTATS
2656 struct task_struct
*tsk
= NULL
;
2658 if (entity_is_task(se
))
2661 if (se
->statistics
.sleep_start
) {
2662 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2667 if (unlikely(delta
> se
->statistics
.sleep_max
))
2668 se
->statistics
.sleep_max
= delta
;
2670 se
->statistics
.sleep_start
= 0;
2671 se
->statistics
.sum_sleep_runtime
+= delta
;
2674 account_scheduler_latency(tsk
, delta
>> 10, 1);
2675 trace_sched_stat_sleep(tsk
, delta
);
2678 if (se
->statistics
.block_start
) {
2679 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2684 if (unlikely(delta
> se
->statistics
.block_max
))
2685 se
->statistics
.block_max
= delta
;
2687 se
->statistics
.block_start
= 0;
2688 se
->statistics
.sum_sleep_runtime
+= delta
;
2691 if (tsk
->in_iowait
) {
2692 se
->statistics
.iowait_sum
+= delta
;
2693 se
->statistics
.iowait_count
++;
2694 trace_sched_stat_iowait(tsk
, delta
);
2697 trace_sched_stat_blocked(tsk
, delta
);
2700 * Blocking time is in units of nanosecs, so shift by
2701 * 20 to get a milliseconds-range estimation of the
2702 * amount of time that the task spent sleeping:
2704 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2705 profile_hits(SLEEP_PROFILING
,
2706 (void *)get_wchan(tsk
),
2709 account_scheduler_latency(tsk
, delta
>> 10, 0);
2715 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2717 #ifdef CONFIG_SCHED_DEBUG
2718 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2723 if (d
> 3*sysctl_sched_latency
)
2724 schedstat_inc(cfs_rq
, nr_spread_over
);
2729 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2731 u64 vruntime
= cfs_rq
->min_vruntime
;
2734 * The 'current' period is already promised to the current tasks,
2735 * however the extra weight of the new task will slow them down a
2736 * little, place the new task so that it fits in the slot that
2737 * stays open at the end.
2739 if (initial
&& sched_feat(START_DEBIT
))
2740 vruntime
+= sched_vslice(cfs_rq
, se
);
2742 /* sleeps up to a single latency don't count. */
2744 unsigned long thresh
= sysctl_sched_latency
;
2747 * Halve their sleep time's effect, to allow
2748 * for a gentler effect of sleepers:
2750 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2756 /* ensure we never gain time by being placed backwards. */
2757 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
2760 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
2763 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2766 * Update the normalized vruntime before updating min_vruntime
2767 * through calling update_curr().
2769 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
2770 se
->vruntime
+= cfs_rq
->min_vruntime
;
2773 * Update run-time statistics of the 'current'.
2775 update_curr(cfs_rq
);
2776 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
2777 account_entity_enqueue(cfs_rq
, se
);
2778 update_cfs_shares(cfs_rq
);
2780 if (flags
& ENQUEUE_WAKEUP
) {
2781 place_entity(cfs_rq
, se
, 0);
2782 enqueue_sleeper(cfs_rq
, se
);
2785 update_stats_enqueue(cfs_rq
, se
);
2786 check_spread(cfs_rq
, se
);
2787 if (se
!= cfs_rq
->curr
)
2788 __enqueue_entity(cfs_rq
, se
);
2791 if (cfs_rq
->nr_running
== 1) {
2792 list_add_leaf_cfs_rq(cfs_rq
);
2793 check_enqueue_throttle(cfs_rq
);
2797 static void __clear_buddies_last(struct sched_entity
*se
)
2799 for_each_sched_entity(se
) {
2800 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2801 if (cfs_rq
->last
!= se
)
2804 cfs_rq
->last
= NULL
;
2808 static void __clear_buddies_next(struct sched_entity
*se
)
2810 for_each_sched_entity(se
) {
2811 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2812 if (cfs_rq
->next
!= se
)
2815 cfs_rq
->next
= NULL
;
2819 static void __clear_buddies_skip(struct sched_entity
*se
)
2821 for_each_sched_entity(se
) {
2822 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2823 if (cfs_rq
->skip
!= se
)
2826 cfs_rq
->skip
= NULL
;
2830 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2832 if (cfs_rq
->last
== se
)
2833 __clear_buddies_last(se
);
2835 if (cfs_rq
->next
== se
)
2836 __clear_buddies_next(se
);
2838 if (cfs_rq
->skip
== se
)
2839 __clear_buddies_skip(se
);
2842 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2845 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2848 * Update run-time statistics of the 'current'.
2850 update_curr(cfs_rq
);
2851 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
2853 update_stats_dequeue(cfs_rq
, se
);
2854 if (flags
& DEQUEUE_SLEEP
) {
2855 #ifdef CONFIG_SCHEDSTATS
2856 if (entity_is_task(se
)) {
2857 struct task_struct
*tsk
= task_of(se
);
2859 if (tsk
->state
& TASK_INTERRUPTIBLE
)
2860 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
2861 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
2862 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
2867 clear_buddies(cfs_rq
, se
);
2869 if (se
!= cfs_rq
->curr
)
2870 __dequeue_entity(cfs_rq
, se
);
2872 account_entity_dequeue(cfs_rq
, se
);
2875 * Normalize the entity after updating the min_vruntime because the
2876 * update can refer to the ->curr item and we need to reflect this
2877 * movement in our normalized position.
2879 if (!(flags
& DEQUEUE_SLEEP
))
2880 se
->vruntime
-= cfs_rq
->min_vruntime
;
2882 /* return excess runtime on last dequeue */
2883 return_cfs_rq_runtime(cfs_rq
);
2885 update_min_vruntime(cfs_rq
);
2886 update_cfs_shares(cfs_rq
);
2890 * Preempt the current task with a newly woken task if needed:
2893 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2895 unsigned long ideal_runtime
, delta_exec
;
2896 struct sched_entity
*se
;
2899 ideal_runtime
= sched_slice(cfs_rq
, curr
);
2900 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
2901 if (delta_exec
> ideal_runtime
) {
2902 resched_task(rq_of(cfs_rq
)->curr
);
2904 * The current task ran long enough, ensure it doesn't get
2905 * re-elected due to buddy favours.
2907 clear_buddies(cfs_rq
, curr
);
2912 * Ensure that a task that missed wakeup preemption by a
2913 * narrow margin doesn't have to wait for a full slice.
2914 * This also mitigates buddy induced latencies under load.
2916 if (delta_exec
< sysctl_sched_min_granularity
)
2919 se
= __pick_first_entity(cfs_rq
);
2920 delta
= curr
->vruntime
- se
->vruntime
;
2925 if (delta
> ideal_runtime
)
2926 resched_task(rq_of(cfs_rq
)->curr
);
2930 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2932 /* 'current' is not kept within the tree. */
2935 * Any task has to be enqueued before it get to execute on
2936 * a CPU. So account for the time it spent waiting on the
2939 update_stats_wait_end(cfs_rq
, se
);
2940 __dequeue_entity(cfs_rq
, se
);
2943 update_stats_curr_start(cfs_rq
, se
);
2945 #ifdef CONFIG_SCHEDSTATS
2947 * Track our maximum slice length, if the CPU's load is at
2948 * least twice that of our own weight (i.e. dont track it
2949 * when there are only lesser-weight tasks around):
2951 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
2952 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
2953 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
2956 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
2960 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
2963 * Pick the next process, keeping these things in mind, in this order:
2964 * 1) keep things fair between processes/task groups
2965 * 2) pick the "next" process, since someone really wants that to run
2966 * 3) pick the "last" process, for cache locality
2967 * 4) do not run the "skip" process, if something else is available
2969 static struct sched_entity
*
2970 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2972 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
2973 struct sched_entity
*se
;
2976 * If curr is set we have to see if its left of the leftmost entity
2977 * still in the tree, provided there was anything in the tree at all.
2979 if (!left
|| (curr
&& entity_before(curr
, left
)))
2982 se
= left
; /* ideally we run the leftmost entity */
2985 * Avoid running the skip buddy, if running something else can
2986 * be done without getting too unfair.
2988 if (cfs_rq
->skip
== se
) {
2989 struct sched_entity
*second
;
2992 second
= __pick_first_entity(cfs_rq
);
2994 second
= __pick_next_entity(se
);
2995 if (!second
|| (curr
&& entity_before(curr
, second
)))
2999 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3004 * Prefer last buddy, try to return the CPU to a preempted task.
3006 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3010 * Someone really wants this to run. If it's not unfair, run it.
3012 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3015 clear_buddies(cfs_rq
, se
);
3020 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3022 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3025 * If still on the runqueue then deactivate_task()
3026 * was not called and update_curr() has to be done:
3029 update_curr(cfs_rq
);
3031 /* throttle cfs_rqs exceeding runtime */
3032 check_cfs_rq_runtime(cfs_rq
);
3034 check_spread(cfs_rq
, prev
);
3036 update_stats_wait_start(cfs_rq
, prev
);
3037 /* Put 'current' back into the tree. */
3038 __enqueue_entity(cfs_rq
, prev
);
3039 /* in !on_rq case, update occurred at dequeue */
3040 update_entity_load_avg(prev
, 1);
3042 cfs_rq
->curr
= NULL
;
3046 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3049 * Update run-time statistics of the 'current'.
3051 update_curr(cfs_rq
);
3054 * Ensure that runnable average is periodically updated.
3056 update_entity_load_avg(curr
, 1);
3057 update_cfs_rq_blocked_load(cfs_rq
, 1);
3058 update_cfs_shares(cfs_rq
);
3060 #ifdef CONFIG_SCHED_HRTICK
3062 * queued ticks are scheduled to match the slice, so don't bother
3063 * validating it and just reschedule.
3066 resched_task(rq_of(cfs_rq
)->curr
);
3070 * don't let the period tick interfere with the hrtick preemption
3072 if (!sched_feat(DOUBLE_TICK
) &&
3073 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3077 if (cfs_rq
->nr_running
> 1)
3078 check_preempt_tick(cfs_rq
, curr
);
3082 /**************************************************
3083 * CFS bandwidth control machinery
3086 #ifdef CONFIG_CFS_BANDWIDTH
3088 #ifdef HAVE_JUMP_LABEL
3089 static struct static_key __cfs_bandwidth_used
;
3091 static inline bool cfs_bandwidth_used(void)
3093 return static_key_false(&__cfs_bandwidth_used
);
3096 void cfs_bandwidth_usage_inc(void)
3098 static_key_slow_inc(&__cfs_bandwidth_used
);
3101 void cfs_bandwidth_usage_dec(void)
3103 static_key_slow_dec(&__cfs_bandwidth_used
);
3105 #else /* HAVE_JUMP_LABEL */
3106 static bool cfs_bandwidth_used(void)
3111 void cfs_bandwidth_usage_inc(void) {}
3112 void cfs_bandwidth_usage_dec(void) {}
3113 #endif /* HAVE_JUMP_LABEL */
3116 * default period for cfs group bandwidth.
3117 * default: 0.1s, units: nanoseconds
3119 static inline u64
default_cfs_period(void)
3121 return 100000000ULL;
3124 static inline u64
sched_cfs_bandwidth_slice(void)
3126 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3130 * Replenish runtime according to assigned quota and update expiration time.
3131 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3132 * additional synchronization around rq->lock.
3134 * requires cfs_b->lock
3136 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3140 if (cfs_b
->quota
== RUNTIME_INF
)
3143 now
= sched_clock_cpu(smp_processor_id());
3144 cfs_b
->runtime
= cfs_b
->quota
;
3145 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3148 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3150 return &tg
->cfs_bandwidth
;
3153 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3154 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3156 if (unlikely(cfs_rq
->throttle_count
))
3157 return cfs_rq
->throttled_clock_task
;
3159 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3162 /* returns 0 on failure to allocate runtime */
3163 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3165 struct task_group
*tg
= cfs_rq
->tg
;
3166 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3167 u64 amount
= 0, min_amount
, expires
;
3169 /* note: this is a positive sum as runtime_remaining <= 0 */
3170 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3172 raw_spin_lock(&cfs_b
->lock
);
3173 if (cfs_b
->quota
== RUNTIME_INF
)
3174 amount
= min_amount
;
3177 * If the bandwidth pool has become inactive, then at least one
3178 * period must have elapsed since the last consumption.
3179 * Refresh the global state and ensure bandwidth timer becomes
3182 if (!cfs_b
->timer_active
) {
3183 __refill_cfs_bandwidth_runtime(cfs_b
);
3184 __start_cfs_bandwidth(cfs_b
, false);
3187 if (cfs_b
->runtime
> 0) {
3188 amount
= min(cfs_b
->runtime
, min_amount
);
3189 cfs_b
->runtime
-= amount
;
3193 expires
= cfs_b
->runtime_expires
;
3194 raw_spin_unlock(&cfs_b
->lock
);
3196 cfs_rq
->runtime_remaining
+= amount
;
3198 * we may have advanced our local expiration to account for allowed
3199 * spread between our sched_clock and the one on which runtime was
3202 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3203 cfs_rq
->runtime_expires
= expires
;
3205 return cfs_rq
->runtime_remaining
> 0;
3209 * Note: This depends on the synchronization provided by sched_clock and the
3210 * fact that rq->clock snapshots this value.
3212 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3214 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3216 /* if the deadline is ahead of our clock, nothing to do */
3217 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3220 if (cfs_rq
->runtime_remaining
< 0)
3224 * If the local deadline has passed we have to consider the
3225 * possibility that our sched_clock is 'fast' and the global deadline
3226 * has not truly expired.
3228 * Fortunately we can check determine whether this the case by checking
3229 * whether the global deadline has advanced. It is valid to compare
3230 * cfs_b->runtime_expires without any locks since we only care about
3231 * exact equality, so a partial write will still work.
3234 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
3235 /* extend local deadline, drift is bounded above by 2 ticks */
3236 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3238 /* global deadline is ahead, expiration has passed */
3239 cfs_rq
->runtime_remaining
= 0;
3243 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3245 /* dock delta_exec before expiring quota (as it could span periods) */
3246 cfs_rq
->runtime_remaining
-= delta_exec
;
3247 expire_cfs_rq_runtime(cfs_rq
);
3249 if (likely(cfs_rq
->runtime_remaining
> 0))
3253 * if we're unable to extend our runtime we resched so that the active
3254 * hierarchy can be throttled
3256 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3257 resched_task(rq_of(cfs_rq
)->curr
);
3260 static __always_inline
3261 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3263 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3266 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3269 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3271 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3274 /* check whether cfs_rq, or any parent, is throttled */
3275 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3277 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3281 * Ensure that neither of the group entities corresponding to src_cpu or
3282 * dest_cpu are members of a throttled hierarchy when performing group
3283 * load-balance operations.
3285 static inline int throttled_lb_pair(struct task_group
*tg
,
3286 int src_cpu
, int dest_cpu
)
3288 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3290 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3291 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3293 return throttled_hierarchy(src_cfs_rq
) ||
3294 throttled_hierarchy(dest_cfs_rq
);
3297 /* updated child weight may affect parent so we have to do this bottom up */
3298 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3300 struct rq
*rq
= data
;
3301 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3303 cfs_rq
->throttle_count
--;
3305 if (!cfs_rq
->throttle_count
) {
3306 /* adjust cfs_rq_clock_task() */
3307 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3308 cfs_rq
->throttled_clock_task
;
3315 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3317 struct rq
*rq
= data
;
3318 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3320 /* group is entering throttled state, stop time */
3321 if (!cfs_rq
->throttle_count
)
3322 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3323 cfs_rq
->throttle_count
++;
3328 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3330 struct rq
*rq
= rq_of(cfs_rq
);
3331 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3332 struct sched_entity
*se
;
3333 long task_delta
, dequeue
= 1;
3335 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3337 /* freeze hierarchy runnable averages while throttled */
3339 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3342 task_delta
= cfs_rq
->h_nr_running
;
3343 for_each_sched_entity(se
) {
3344 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3345 /* throttled entity or throttle-on-deactivate */
3350 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3351 qcfs_rq
->h_nr_running
-= task_delta
;
3353 if (qcfs_rq
->load
.weight
)
3358 sub_nr_running(rq
, task_delta
);
3360 cfs_rq
->throttled
= 1;
3361 cfs_rq
->throttled_clock
= rq_clock(rq
);
3362 raw_spin_lock(&cfs_b
->lock
);
3363 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3364 if (!cfs_b
->timer_active
)
3365 __start_cfs_bandwidth(cfs_b
, false);
3366 raw_spin_unlock(&cfs_b
->lock
);
3369 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3371 struct rq
*rq
= rq_of(cfs_rq
);
3372 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3373 struct sched_entity
*se
;
3377 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3379 cfs_rq
->throttled
= 0;
3381 update_rq_clock(rq
);
3383 raw_spin_lock(&cfs_b
->lock
);
3384 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3385 list_del_rcu(&cfs_rq
->throttled_list
);
3386 raw_spin_unlock(&cfs_b
->lock
);
3388 /* update hierarchical throttle state */
3389 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3391 if (!cfs_rq
->load
.weight
)
3394 task_delta
= cfs_rq
->h_nr_running
;
3395 for_each_sched_entity(se
) {
3399 cfs_rq
= cfs_rq_of(se
);
3401 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3402 cfs_rq
->h_nr_running
+= task_delta
;
3404 if (cfs_rq_throttled(cfs_rq
))
3409 add_nr_running(rq
, task_delta
);
3411 /* determine whether we need to wake up potentially idle cpu */
3412 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3413 resched_task(rq
->curr
);
3416 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3417 u64 remaining
, u64 expires
)
3419 struct cfs_rq
*cfs_rq
;
3420 u64 runtime
= remaining
;
3423 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3425 struct rq
*rq
= rq_of(cfs_rq
);
3427 raw_spin_lock(&rq
->lock
);
3428 if (!cfs_rq_throttled(cfs_rq
))
3431 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3432 if (runtime
> remaining
)
3433 runtime
= remaining
;
3434 remaining
-= runtime
;
3436 cfs_rq
->runtime_remaining
+= runtime
;
3437 cfs_rq
->runtime_expires
= expires
;
3439 /* we check whether we're throttled above */
3440 if (cfs_rq
->runtime_remaining
> 0)
3441 unthrottle_cfs_rq(cfs_rq
);
3444 raw_spin_unlock(&rq
->lock
);
3455 * Responsible for refilling a task_group's bandwidth and unthrottling its
3456 * cfs_rqs as appropriate. If there has been no activity within the last
3457 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3458 * used to track this state.
3460 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3462 u64 runtime
, runtime_expires
;
3465 /* no need to continue the timer with no bandwidth constraint */
3466 if (cfs_b
->quota
== RUNTIME_INF
)
3467 goto out_deactivate
;
3469 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3470 cfs_b
->nr_periods
+= overrun
;
3473 * idle depends on !throttled (for the case of a large deficit), and if
3474 * we're going inactive then everything else can be deferred
3476 if (cfs_b
->idle
&& !throttled
)
3477 goto out_deactivate
;
3480 * if we have relooped after returning idle once, we need to update our
3481 * status as actually running, so that other cpus doing
3482 * __start_cfs_bandwidth will stop trying to cancel us.
3484 cfs_b
->timer_active
= 1;
3486 __refill_cfs_bandwidth_runtime(cfs_b
);
3489 /* mark as potentially idle for the upcoming period */
3494 /* account preceding periods in which throttling occurred */
3495 cfs_b
->nr_throttled
+= overrun
;
3498 * There are throttled entities so we must first use the new bandwidth
3499 * to unthrottle them before making it generally available. This
3500 * ensures that all existing debts will be paid before a new cfs_rq is
3503 runtime
= cfs_b
->runtime
;
3504 runtime_expires
= cfs_b
->runtime_expires
;
3508 * This check is repeated as we are holding onto the new bandwidth
3509 * while we unthrottle. This can potentially race with an unthrottled
3510 * group trying to acquire new bandwidth from the global pool.
3512 while (throttled
&& runtime
> 0) {
3513 raw_spin_unlock(&cfs_b
->lock
);
3514 /* we can't nest cfs_b->lock while distributing bandwidth */
3515 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3517 raw_spin_lock(&cfs_b
->lock
);
3519 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3522 /* return (any) remaining runtime */
3523 cfs_b
->runtime
= runtime
;
3525 * While we are ensured activity in the period following an
3526 * unthrottle, this also covers the case in which the new bandwidth is
3527 * insufficient to cover the existing bandwidth deficit. (Forcing the
3528 * timer to remain active while there are any throttled entities.)
3535 cfs_b
->timer_active
= 0;
3539 /* a cfs_rq won't donate quota below this amount */
3540 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3541 /* minimum remaining period time to redistribute slack quota */
3542 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3543 /* how long we wait to gather additional slack before distributing */
3544 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3547 * Are we near the end of the current quota period?
3549 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3550 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3551 * migrate_hrtimers, base is never cleared, so we are fine.
3553 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3555 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3558 /* if the call-back is running a quota refresh is already occurring */
3559 if (hrtimer_callback_running(refresh_timer
))
3562 /* is a quota refresh about to occur? */
3563 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3564 if (remaining
< min_expire
)
3570 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3572 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3574 /* if there's a quota refresh soon don't bother with slack */
3575 if (runtime_refresh_within(cfs_b
, min_left
))
3578 start_bandwidth_timer(&cfs_b
->slack_timer
,
3579 ns_to_ktime(cfs_bandwidth_slack_period
));
3582 /* we know any runtime found here is valid as update_curr() precedes return */
3583 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3585 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3586 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3588 if (slack_runtime
<= 0)
3591 raw_spin_lock(&cfs_b
->lock
);
3592 if (cfs_b
->quota
!= RUNTIME_INF
&&
3593 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3594 cfs_b
->runtime
+= slack_runtime
;
3596 /* we are under rq->lock, defer unthrottling using a timer */
3597 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3598 !list_empty(&cfs_b
->throttled_cfs_rq
))
3599 start_cfs_slack_bandwidth(cfs_b
);
3601 raw_spin_unlock(&cfs_b
->lock
);
3603 /* even if it's not valid for return we don't want to try again */
3604 cfs_rq
->runtime_remaining
-= slack_runtime
;
3607 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3609 if (!cfs_bandwidth_used())
3612 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3615 __return_cfs_rq_runtime(cfs_rq
);
3619 * This is done with a timer (instead of inline with bandwidth return) since
3620 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3622 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3624 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3627 /* confirm we're still not at a refresh boundary */
3628 raw_spin_lock(&cfs_b
->lock
);
3629 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
3630 raw_spin_unlock(&cfs_b
->lock
);
3634 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
) {
3635 runtime
= cfs_b
->runtime
;
3638 expires
= cfs_b
->runtime_expires
;
3639 raw_spin_unlock(&cfs_b
->lock
);
3644 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3646 raw_spin_lock(&cfs_b
->lock
);
3647 if (expires
== cfs_b
->runtime_expires
)
3648 cfs_b
->runtime
= runtime
;
3649 raw_spin_unlock(&cfs_b
->lock
);
3653 * When a group wakes up we want to make sure that its quota is not already
3654 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3655 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3657 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3659 if (!cfs_bandwidth_used())
3662 /* an active group must be handled by the update_curr()->put() path */
3663 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3666 /* ensure the group is not already throttled */
3667 if (cfs_rq_throttled(cfs_rq
))
3670 /* update runtime allocation */
3671 account_cfs_rq_runtime(cfs_rq
, 0);
3672 if (cfs_rq
->runtime_remaining
<= 0)
3673 throttle_cfs_rq(cfs_rq
);
3676 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3677 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3679 if (!cfs_bandwidth_used())
3682 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3686 * it's possible for a throttled entity to be forced into a running
3687 * state (e.g. set_curr_task), in this case we're finished.
3689 if (cfs_rq_throttled(cfs_rq
))
3692 throttle_cfs_rq(cfs_rq
);
3696 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3698 struct cfs_bandwidth
*cfs_b
=
3699 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3700 do_sched_cfs_slack_timer(cfs_b
);
3702 return HRTIMER_NORESTART
;
3705 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3707 struct cfs_bandwidth
*cfs_b
=
3708 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3713 raw_spin_lock(&cfs_b
->lock
);
3715 now
= hrtimer_cb_get_time(timer
);
3716 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3721 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3723 raw_spin_unlock(&cfs_b
->lock
);
3725 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3728 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3730 raw_spin_lock_init(&cfs_b
->lock
);
3732 cfs_b
->quota
= RUNTIME_INF
;
3733 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3735 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3736 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3737 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3738 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3739 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3742 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3744 cfs_rq
->runtime_enabled
= 0;
3745 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3748 /* requires cfs_b->lock, may release to reprogram timer */
3749 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
, bool force
)
3752 * The timer may be active because we're trying to set a new bandwidth
3753 * period or because we're racing with the tear-down path
3754 * (timer_active==0 becomes visible before the hrtimer call-back
3755 * terminates). In either case we ensure that it's re-programmed
3757 while (unlikely(hrtimer_active(&cfs_b
->period_timer
)) &&
3758 hrtimer_try_to_cancel(&cfs_b
->period_timer
) < 0) {
3759 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3760 raw_spin_unlock(&cfs_b
->lock
);
3762 raw_spin_lock(&cfs_b
->lock
);
3763 /* if someone else restarted the timer then we're done */
3764 if (!force
&& cfs_b
->timer_active
)
3768 cfs_b
->timer_active
= 1;
3769 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
3772 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3774 hrtimer_cancel(&cfs_b
->period_timer
);
3775 hrtimer_cancel(&cfs_b
->slack_timer
);
3778 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
3780 struct cfs_rq
*cfs_rq
;
3782 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3783 if (!cfs_rq
->runtime_enabled
)
3787 * clock_task is not advancing so we just need to make sure
3788 * there's some valid quota amount
3790 cfs_rq
->runtime_remaining
= 1;
3791 if (cfs_rq_throttled(cfs_rq
))
3792 unthrottle_cfs_rq(cfs_rq
);
3796 #else /* CONFIG_CFS_BANDWIDTH */
3797 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3799 return rq_clock_task(rq_of(cfs_rq
));
3802 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
3803 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
3804 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
3805 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3807 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3812 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3817 static inline int throttled_lb_pair(struct task_group
*tg
,
3818 int src_cpu
, int dest_cpu
)
3823 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3825 #ifdef CONFIG_FAIR_GROUP_SCHED
3826 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3829 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3833 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3834 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
3836 #endif /* CONFIG_CFS_BANDWIDTH */
3838 /**************************************************
3839 * CFS operations on tasks:
3842 #ifdef CONFIG_SCHED_HRTICK
3843 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3845 struct sched_entity
*se
= &p
->se
;
3846 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3848 WARN_ON(task_rq(p
) != rq
);
3850 if (cfs_rq
->nr_running
> 1) {
3851 u64 slice
= sched_slice(cfs_rq
, se
);
3852 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
3853 s64 delta
= slice
- ran
;
3862 * Don't schedule slices shorter than 10000ns, that just
3863 * doesn't make sense. Rely on vruntime for fairness.
3866 delta
= max_t(s64
, 10000LL, delta
);
3868 hrtick_start(rq
, delta
);
3873 * called from enqueue/dequeue and updates the hrtick when the
3874 * current task is from our class and nr_running is low enough
3877 static void hrtick_update(struct rq
*rq
)
3879 struct task_struct
*curr
= rq
->curr
;
3881 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
3884 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
3885 hrtick_start_fair(rq
, curr
);
3887 #else /* !CONFIG_SCHED_HRTICK */
3889 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3893 static inline void hrtick_update(struct rq
*rq
)
3899 * The enqueue_task method is called before nr_running is
3900 * increased. Here we update the fair scheduling stats and
3901 * then put the task into the rbtree:
3904 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3906 struct cfs_rq
*cfs_rq
;
3907 struct sched_entity
*se
= &p
->se
;
3909 for_each_sched_entity(se
) {
3912 cfs_rq
= cfs_rq_of(se
);
3913 enqueue_entity(cfs_rq
, se
, flags
);
3916 * end evaluation on encountering a throttled cfs_rq
3918 * note: in the case of encountering a throttled cfs_rq we will
3919 * post the final h_nr_running increment below.
3921 if (cfs_rq_throttled(cfs_rq
))
3923 cfs_rq
->h_nr_running
++;
3925 flags
= ENQUEUE_WAKEUP
;
3928 for_each_sched_entity(se
) {
3929 cfs_rq
= cfs_rq_of(se
);
3930 cfs_rq
->h_nr_running
++;
3932 if (cfs_rq_throttled(cfs_rq
))
3935 update_cfs_shares(cfs_rq
);
3936 update_entity_load_avg(se
, 1);
3940 update_rq_runnable_avg(rq
, rq
->nr_running
);
3941 add_nr_running(rq
, 1);
3946 static void set_next_buddy(struct sched_entity
*se
);
3949 * The dequeue_task method is called before nr_running is
3950 * decreased. We remove the task from the rbtree and
3951 * update the fair scheduling stats:
3953 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3955 struct cfs_rq
*cfs_rq
;
3956 struct sched_entity
*se
= &p
->se
;
3957 int task_sleep
= flags
& DEQUEUE_SLEEP
;
3959 for_each_sched_entity(se
) {
3960 cfs_rq
= cfs_rq_of(se
);
3961 dequeue_entity(cfs_rq
, se
, flags
);
3964 * end evaluation on encountering a throttled cfs_rq
3966 * note: in the case of encountering a throttled cfs_rq we will
3967 * post the final h_nr_running decrement below.
3969 if (cfs_rq_throttled(cfs_rq
))
3971 cfs_rq
->h_nr_running
--;
3973 /* Don't dequeue parent if it has other entities besides us */
3974 if (cfs_rq
->load
.weight
) {
3976 * Bias pick_next to pick a task from this cfs_rq, as
3977 * p is sleeping when it is within its sched_slice.
3979 if (task_sleep
&& parent_entity(se
))
3980 set_next_buddy(parent_entity(se
));
3982 /* avoid re-evaluating load for this entity */
3983 se
= parent_entity(se
);
3986 flags
|= DEQUEUE_SLEEP
;
3989 for_each_sched_entity(se
) {
3990 cfs_rq
= cfs_rq_of(se
);
3991 cfs_rq
->h_nr_running
--;
3993 if (cfs_rq_throttled(cfs_rq
))
3996 update_cfs_shares(cfs_rq
);
3997 update_entity_load_avg(se
, 1);
4001 sub_nr_running(rq
, 1);
4002 update_rq_runnable_avg(rq
, 1);
4008 /* Used instead of source_load when we know the type == 0 */
4009 static unsigned long weighted_cpuload(const int cpu
)
4011 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
4015 * Return a low guess at the load of a migration-source cpu weighted
4016 * according to the scheduling class and "nice" value.
4018 * We want to under-estimate the load of migration sources, to
4019 * balance conservatively.
4021 static unsigned long source_load(int cpu
, int type
)
4023 struct rq
*rq
= cpu_rq(cpu
);
4024 unsigned long total
= weighted_cpuload(cpu
);
4026 if (type
== 0 || !sched_feat(LB_BIAS
))
4029 return min(rq
->cpu_load
[type
-1], total
);
4033 * Return a high guess at the load of a migration-target cpu weighted
4034 * according to the scheduling class and "nice" value.
4036 static unsigned long target_load(int cpu
, int type
)
4038 struct rq
*rq
= cpu_rq(cpu
);
4039 unsigned long total
= weighted_cpuload(cpu
);
4041 if (type
== 0 || !sched_feat(LB_BIAS
))
4044 return max(rq
->cpu_load
[type
-1], total
);
4047 static unsigned long capacity_of(int cpu
)
4049 return cpu_rq(cpu
)->cpu_capacity
;
4052 static unsigned long cpu_avg_load_per_task(int cpu
)
4054 struct rq
*rq
= cpu_rq(cpu
);
4055 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
4056 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
4059 return load_avg
/ nr_running
;
4064 static void record_wakee(struct task_struct
*p
)
4067 * Rough decay (wiping) for cost saving, don't worry
4068 * about the boundary, really active task won't care
4071 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
4072 current
->wakee_flips
>>= 1;
4073 current
->wakee_flip_decay_ts
= jiffies
;
4076 if (current
->last_wakee
!= p
) {
4077 current
->last_wakee
= p
;
4078 current
->wakee_flips
++;
4082 static void task_waking_fair(struct task_struct
*p
)
4084 struct sched_entity
*se
= &p
->se
;
4085 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4088 #ifndef CONFIG_64BIT
4089 u64 min_vruntime_copy
;
4092 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
4094 min_vruntime
= cfs_rq
->min_vruntime
;
4095 } while (min_vruntime
!= min_vruntime_copy
);
4097 min_vruntime
= cfs_rq
->min_vruntime
;
4100 se
->vruntime
-= min_vruntime
;
4104 #ifdef CONFIG_FAIR_GROUP_SCHED
4106 * effective_load() calculates the load change as seen from the root_task_group
4108 * Adding load to a group doesn't make a group heavier, but can cause movement
4109 * of group shares between cpus. Assuming the shares were perfectly aligned one
4110 * can calculate the shift in shares.
4112 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4113 * on this @cpu and results in a total addition (subtraction) of @wg to the
4114 * total group weight.
4116 * Given a runqueue weight distribution (rw_i) we can compute a shares
4117 * distribution (s_i) using:
4119 * s_i = rw_i / \Sum rw_j (1)
4121 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4122 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4123 * shares distribution (s_i):
4125 * rw_i = { 2, 4, 1, 0 }
4126 * s_i = { 2/7, 4/7, 1/7, 0 }
4128 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4129 * task used to run on and the CPU the waker is running on), we need to
4130 * compute the effect of waking a task on either CPU and, in case of a sync
4131 * wakeup, compute the effect of the current task going to sleep.
4133 * So for a change of @wl to the local @cpu with an overall group weight change
4134 * of @wl we can compute the new shares distribution (s'_i) using:
4136 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4138 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4139 * differences in waking a task to CPU 0. The additional task changes the
4140 * weight and shares distributions like:
4142 * rw'_i = { 3, 4, 1, 0 }
4143 * s'_i = { 3/8, 4/8, 1/8, 0 }
4145 * We can then compute the difference in effective weight by using:
4147 * dw_i = S * (s'_i - s_i) (3)
4149 * Where 'S' is the group weight as seen by its parent.
4151 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4152 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4153 * 4/7) times the weight of the group.
4155 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4157 struct sched_entity
*se
= tg
->se
[cpu
];
4159 if (!tg
->parent
) /* the trivial, non-cgroup case */
4162 for_each_sched_entity(se
) {
4168 * W = @wg + \Sum rw_j
4170 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4175 w
= se
->my_q
->load
.weight
+ wl
;
4178 * wl = S * s'_i; see (2)
4181 wl
= (w
* tg
->shares
) / W
;
4186 * Per the above, wl is the new se->load.weight value; since
4187 * those are clipped to [MIN_SHARES, ...) do so now. See
4188 * calc_cfs_shares().
4190 if (wl
< MIN_SHARES
)
4194 * wl = dw_i = S * (s'_i - s_i); see (3)
4196 wl
-= se
->load
.weight
;
4199 * Recursively apply this logic to all parent groups to compute
4200 * the final effective load change on the root group. Since
4201 * only the @tg group gets extra weight, all parent groups can
4202 * only redistribute existing shares. @wl is the shift in shares
4203 * resulting from this level per the above.
4212 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4219 static int wake_wide(struct task_struct
*p
)
4221 int factor
= this_cpu_read(sd_llc_size
);
4224 * Yeah, it's the switching-frequency, could means many wakee or
4225 * rapidly switch, use factor here will just help to automatically
4226 * adjust the loose-degree, so bigger node will lead to more pull.
4228 if (p
->wakee_flips
> factor
) {
4230 * wakee is somewhat hot, it needs certain amount of cpu
4231 * resource, so if waker is far more hot, prefer to leave
4234 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
4241 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4243 s64 this_load
, load
;
4244 int idx
, this_cpu
, prev_cpu
;
4245 unsigned long tl_per_task
;
4246 struct task_group
*tg
;
4247 unsigned long weight
;
4251 * If we wake multiple tasks be careful to not bounce
4252 * ourselves around too much.
4258 this_cpu
= smp_processor_id();
4259 prev_cpu
= task_cpu(p
);
4260 load
= source_load(prev_cpu
, idx
);
4261 this_load
= target_load(this_cpu
, idx
);
4264 * If sync wakeup then subtract the (maximum possible)
4265 * effect of the currently running task from the load
4266 * of the current CPU:
4269 tg
= task_group(current
);
4270 weight
= current
->se
.load
.weight
;
4272 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4273 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4277 weight
= p
->se
.load
.weight
;
4280 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4281 * due to the sync cause above having dropped this_load to 0, we'll
4282 * always have an imbalance, but there's really nothing you can do
4283 * about that, so that's good too.
4285 * Otherwise check if either cpus are near enough in load to allow this
4286 * task to be woken on this_cpu.
4288 if (this_load
> 0) {
4289 s64 this_eff_load
, prev_eff_load
;
4291 this_eff_load
= 100;
4292 this_eff_load
*= capacity_of(prev_cpu
);
4293 this_eff_load
*= this_load
+
4294 effective_load(tg
, this_cpu
, weight
, weight
);
4296 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4297 prev_eff_load
*= capacity_of(this_cpu
);
4298 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4300 balanced
= this_eff_load
<= prev_eff_load
;
4305 * If the currently running task will sleep within
4306 * a reasonable amount of time then attract this newly
4309 if (sync
&& balanced
)
4312 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4313 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
4316 (this_load
<= load
&&
4317 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
4319 * This domain has SD_WAKE_AFFINE and
4320 * p is cache cold in this domain, and
4321 * there is no bad imbalance.
4323 schedstat_inc(sd
, ttwu_move_affine
);
4324 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4332 * find_idlest_group finds and returns the least busy CPU group within the
4335 static struct sched_group
*
4336 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4337 int this_cpu
, int sd_flag
)
4339 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4340 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4341 int load_idx
= sd
->forkexec_idx
;
4342 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4344 if (sd_flag
& SD_BALANCE_WAKE
)
4345 load_idx
= sd
->wake_idx
;
4348 unsigned long load
, avg_load
;
4352 /* Skip over this group if it has no CPUs allowed */
4353 if (!cpumask_intersects(sched_group_cpus(group
),
4354 tsk_cpus_allowed(p
)))
4357 local_group
= cpumask_test_cpu(this_cpu
,
4358 sched_group_cpus(group
));
4360 /* Tally up the load of all CPUs in the group */
4363 for_each_cpu(i
, sched_group_cpus(group
)) {
4364 /* Bias balancing toward cpus of our domain */
4366 load
= source_load(i
, load_idx
);
4368 load
= target_load(i
, load_idx
);
4373 /* Adjust by relative CPU capacity of the group */
4374 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) / group
->sgc
->capacity
;
4377 this_load
= avg_load
;
4378 } else if (avg_load
< min_load
) {
4379 min_load
= avg_load
;
4382 } while (group
= group
->next
, group
!= sd
->groups
);
4384 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4390 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4393 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4395 unsigned long load
, min_load
= ULONG_MAX
;
4399 /* Traverse only the allowed CPUs */
4400 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4401 load
= weighted_cpuload(i
);
4403 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4413 * Try and locate an idle CPU in the sched_domain.
4415 static int select_idle_sibling(struct task_struct
*p
, int target
)
4417 struct sched_domain
*sd
;
4418 struct sched_group
*sg
;
4419 int i
= task_cpu(p
);
4421 if (idle_cpu(target
))
4425 * If the prevous cpu is cache affine and idle, don't be stupid.
4427 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4431 * Otherwise, iterate the domains and find an elegible idle cpu.
4433 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4434 for_each_lower_domain(sd
) {
4437 if (!cpumask_intersects(sched_group_cpus(sg
),
4438 tsk_cpus_allowed(p
)))
4441 for_each_cpu(i
, sched_group_cpus(sg
)) {
4442 if (i
== target
|| !idle_cpu(i
))
4446 target
= cpumask_first_and(sched_group_cpus(sg
),
4447 tsk_cpus_allowed(p
));
4451 } while (sg
!= sd
->groups
);
4458 * select_task_rq_fair: Select target runqueue for the waking task in domains
4459 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4460 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4462 * Balances load by selecting the idlest cpu in the idlest group, or under
4463 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4465 * Returns the target cpu number.
4467 * preempt must be disabled.
4470 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4472 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4473 int cpu
= smp_processor_id();
4475 int want_affine
= 0;
4476 int sync
= wake_flags
& WF_SYNC
;
4478 if (p
->nr_cpus_allowed
== 1)
4481 if (sd_flag
& SD_BALANCE_WAKE
) {
4482 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
4488 for_each_domain(cpu
, tmp
) {
4489 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4493 * If both cpu and prev_cpu are part of this domain,
4494 * cpu is a valid SD_WAKE_AFFINE target.
4496 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4497 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4502 if (tmp
->flags
& sd_flag
)
4506 if (affine_sd
&& cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4509 if (sd_flag
& SD_BALANCE_WAKE
) {
4510 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4515 struct sched_group
*group
;
4518 if (!(sd
->flags
& sd_flag
)) {
4523 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
4529 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4530 if (new_cpu
== -1 || new_cpu
== cpu
) {
4531 /* Now try balancing at a lower domain level of cpu */
4536 /* Now try balancing at a lower domain level of new_cpu */
4538 weight
= sd
->span_weight
;
4540 for_each_domain(cpu
, tmp
) {
4541 if (weight
<= tmp
->span_weight
)
4543 if (tmp
->flags
& sd_flag
)
4546 /* while loop will break here if sd == NULL */
4555 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4556 * cfs_rq_of(p) references at time of call are still valid and identify the
4557 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4558 * other assumptions, including the state of rq->lock, should be made.
4561 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4563 struct sched_entity
*se
= &p
->se
;
4564 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4567 * Load tracking: accumulate removed load so that it can be processed
4568 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4569 * to blocked load iff they have a positive decay-count. It can never
4570 * be negative here since on-rq tasks have decay-count == 0.
4572 if (se
->avg
.decay_count
) {
4573 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4574 atomic_long_add(se
->avg
.load_avg_contrib
,
4575 &cfs_rq
->removed_load
);
4578 /* We have migrated, no longer consider this task hot */
4581 #endif /* CONFIG_SMP */
4583 static unsigned long
4584 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4586 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4589 * Since its curr running now, convert the gran from real-time
4590 * to virtual-time in his units.
4592 * By using 'se' instead of 'curr' we penalize light tasks, so
4593 * they get preempted easier. That is, if 'se' < 'curr' then
4594 * the resulting gran will be larger, therefore penalizing the
4595 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4596 * be smaller, again penalizing the lighter task.
4598 * This is especially important for buddies when the leftmost
4599 * task is higher priority than the buddy.
4601 return calc_delta_fair(gran
, se
);
4605 * Should 'se' preempt 'curr'.
4619 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4621 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4626 gran
= wakeup_gran(curr
, se
);
4633 static void set_last_buddy(struct sched_entity
*se
)
4635 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4638 for_each_sched_entity(se
)
4639 cfs_rq_of(se
)->last
= se
;
4642 static void set_next_buddy(struct sched_entity
*se
)
4644 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4647 for_each_sched_entity(se
)
4648 cfs_rq_of(se
)->next
= se
;
4651 static void set_skip_buddy(struct sched_entity
*se
)
4653 for_each_sched_entity(se
)
4654 cfs_rq_of(se
)->skip
= se
;
4658 * Preempt the current task with a newly woken task if needed:
4660 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4662 struct task_struct
*curr
= rq
->curr
;
4663 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4664 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4665 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4666 int next_buddy_marked
= 0;
4668 if (unlikely(se
== pse
))
4672 * This is possible from callers such as move_task(), in which we
4673 * unconditionally check_prempt_curr() after an enqueue (which may have
4674 * lead to a throttle). This both saves work and prevents false
4675 * next-buddy nomination below.
4677 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4680 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4681 set_next_buddy(pse
);
4682 next_buddy_marked
= 1;
4686 * We can come here with TIF_NEED_RESCHED already set from new task
4689 * Note: this also catches the edge-case of curr being in a throttled
4690 * group (e.g. via set_curr_task), since update_curr() (in the
4691 * enqueue of curr) will have resulted in resched being set. This
4692 * prevents us from potentially nominating it as a false LAST_BUDDY
4695 if (test_tsk_need_resched(curr
))
4698 /* Idle tasks are by definition preempted by non-idle tasks. */
4699 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4700 likely(p
->policy
!= SCHED_IDLE
))
4704 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4705 * is driven by the tick):
4707 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4710 find_matching_se(&se
, &pse
);
4711 update_curr(cfs_rq_of(se
));
4713 if (wakeup_preempt_entity(se
, pse
) == 1) {
4715 * Bias pick_next to pick the sched entity that is
4716 * triggering this preemption.
4718 if (!next_buddy_marked
)
4719 set_next_buddy(pse
);
4728 * Only set the backward buddy when the current task is still
4729 * on the rq. This can happen when a wakeup gets interleaved
4730 * with schedule on the ->pre_schedule() or idle_balance()
4731 * point, either of which can * drop the rq lock.
4733 * Also, during early boot the idle thread is in the fair class,
4734 * for obvious reasons its a bad idea to schedule back to it.
4736 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4739 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
4743 static struct task_struct
*
4744 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4746 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
4747 struct sched_entity
*se
;
4748 struct task_struct
*p
;
4752 #ifdef CONFIG_FAIR_GROUP_SCHED
4753 if (!cfs_rq
->nr_running
)
4756 if (prev
->sched_class
!= &fair_sched_class
)
4760 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4761 * likely that a next task is from the same cgroup as the current.
4763 * Therefore attempt to avoid putting and setting the entire cgroup
4764 * hierarchy, only change the part that actually changes.
4768 struct sched_entity
*curr
= cfs_rq
->curr
;
4771 * Since we got here without doing put_prev_entity() we also
4772 * have to consider cfs_rq->curr. If it is still a runnable
4773 * entity, update_curr() will update its vruntime, otherwise
4774 * forget we've ever seen it.
4776 if (curr
&& curr
->on_rq
)
4777 update_curr(cfs_rq
);
4782 * This call to check_cfs_rq_runtime() will do the throttle and
4783 * dequeue its entity in the parent(s). Therefore the 'simple'
4784 * nr_running test will indeed be correct.
4786 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
4789 se
= pick_next_entity(cfs_rq
, curr
);
4790 cfs_rq
= group_cfs_rq(se
);
4796 * Since we haven't yet done put_prev_entity and if the selected task
4797 * is a different task than we started out with, try and touch the
4798 * least amount of cfs_rqs.
4801 struct sched_entity
*pse
= &prev
->se
;
4803 while (!(cfs_rq
= is_same_group(se
, pse
))) {
4804 int se_depth
= se
->depth
;
4805 int pse_depth
= pse
->depth
;
4807 if (se_depth
<= pse_depth
) {
4808 put_prev_entity(cfs_rq_of(pse
), pse
);
4809 pse
= parent_entity(pse
);
4811 if (se_depth
>= pse_depth
) {
4812 set_next_entity(cfs_rq_of(se
), se
);
4813 se
= parent_entity(se
);
4817 put_prev_entity(cfs_rq
, pse
);
4818 set_next_entity(cfs_rq
, se
);
4821 if (hrtick_enabled(rq
))
4822 hrtick_start_fair(rq
, p
);
4829 if (!cfs_rq
->nr_running
)
4832 put_prev_task(rq
, prev
);
4835 se
= pick_next_entity(cfs_rq
, NULL
);
4836 set_next_entity(cfs_rq
, se
);
4837 cfs_rq
= group_cfs_rq(se
);
4842 if (hrtick_enabled(rq
))
4843 hrtick_start_fair(rq
, p
);
4848 new_tasks
= idle_balance(rq
);
4850 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4851 * possible for any higher priority task to appear. In that case we
4852 * must re-start the pick_next_entity() loop.
4864 * Account for a descheduled task:
4866 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4868 struct sched_entity
*se
= &prev
->se
;
4869 struct cfs_rq
*cfs_rq
;
4871 for_each_sched_entity(se
) {
4872 cfs_rq
= cfs_rq_of(se
);
4873 put_prev_entity(cfs_rq
, se
);
4878 * sched_yield() is very simple
4880 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4882 static void yield_task_fair(struct rq
*rq
)
4884 struct task_struct
*curr
= rq
->curr
;
4885 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4886 struct sched_entity
*se
= &curr
->se
;
4889 * Are we the only task in the tree?
4891 if (unlikely(rq
->nr_running
== 1))
4894 clear_buddies(cfs_rq
, se
);
4896 if (curr
->policy
!= SCHED_BATCH
) {
4897 update_rq_clock(rq
);
4899 * Update run-time statistics of the 'current'.
4901 update_curr(cfs_rq
);
4903 * Tell update_rq_clock() that we've just updated,
4904 * so we don't do microscopic update in schedule()
4905 * and double the fastpath cost.
4907 rq
->skip_clock_update
= 1;
4913 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
4915 struct sched_entity
*se
= &p
->se
;
4917 /* throttled hierarchies are not runnable */
4918 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
4921 /* Tell the scheduler that we'd really like pse to run next. */
4924 yield_task_fair(rq
);
4930 /**************************************************
4931 * Fair scheduling class load-balancing methods.
4935 * The purpose of load-balancing is to achieve the same basic fairness the
4936 * per-cpu scheduler provides, namely provide a proportional amount of compute
4937 * time to each task. This is expressed in the following equation:
4939 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4941 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4942 * W_i,0 is defined as:
4944 * W_i,0 = \Sum_j w_i,j (2)
4946 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4947 * is derived from the nice value as per prio_to_weight[].
4949 * The weight average is an exponential decay average of the instantaneous
4952 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4954 * C_i is the compute capacity of cpu i, typically it is the
4955 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4956 * can also include other factors [XXX].
4958 * To achieve this balance we define a measure of imbalance which follows
4959 * directly from (1):
4961 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
4963 * We them move tasks around to minimize the imbalance. In the continuous
4964 * function space it is obvious this converges, in the discrete case we get
4965 * a few fun cases generally called infeasible weight scenarios.
4968 * - infeasible weights;
4969 * - local vs global optima in the discrete case. ]
4974 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4975 * for all i,j solution, we create a tree of cpus that follows the hardware
4976 * topology where each level pairs two lower groups (or better). This results
4977 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4978 * tree to only the first of the previous level and we decrease the frequency
4979 * of load-balance at each level inv. proportional to the number of cpus in
4985 * \Sum { --- * --- * 2^i } = O(n) (5)
4987 * `- size of each group
4988 * | | `- number of cpus doing load-balance
4990 * `- sum over all levels
4992 * Coupled with a limit on how many tasks we can migrate every balance pass,
4993 * this makes (5) the runtime complexity of the balancer.
4995 * An important property here is that each CPU is still (indirectly) connected
4996 * to every other cpu in at most O(log n) steps:
4998 * The adjacency matrix of the resulting graph is given by:
5001 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5004 * And you'll find that:
5006 * A^(log_2 n)_i,j != 0 for all i,j (7)
5008 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5009 * The task movement gives a factor of O(m), giving a convergence complexity
5012 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5017 * In order to avoid CPUs going idle while there's still work to do, new idle
5018 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5019 * tree itself instead of relying on other CPUs to bring it work.
5021 * This adds some complexity to both (5) and (8) but it reduces the total idle
5029 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5032 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5037 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5039 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5041 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5044 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5045 * rewrite all of this once again.]
5048 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
5050 enum fbq_type
{ regular
, remote
, all
};
5052 #define LBF_ALL_PINNED 0x01
5053 #define LBF_NEED_BREAK 0x02
5054 #define LBF_DST_PINNED 0x04
5055 #define LBF_SOME_PINNED 0x08
5058 struct sched_domain
*sd
;
5066 struct cpumask
*dst_grpmask
;
5068 enum cpu_idle_type idle
;
5070 /* The set of CPUs under consideration for load-balancing */
5071 struct cpumask
*cpus
;
5076 unsigned int loop_break
;
5077 unsigned int loop_max
;
5079 enum fbq_type fbq_type
;
5083 * move_task - move a task from one runqueue to another runqueue.
5084 * Both runqueues must be locked.
5086 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
5088 deactivate_task(env
->src_rq
, p
, 0);
5089 set_task_cpu(p
, env
->dst_cpu
);
5090 activate_task(env
->dst_rq
, p
, 0);
5091 check_preempt_curr(env
->dst_rq
, p
, 0);
5095 * Is this task likely cache-hot:
5098 task_hot(struct task_struct
*p
, u64 now
)
5102 if (p
->sched_class
!= &fair_sched_class
)
5105 if (unlikely(p
->policy
== SCHED_IDLE
))
5109 * Buddy candidates are cache hot:
5111 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
5112 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
5113 &p
->se
== cfs_rq_of(&p
->se
)->last
))
5116 if (sysctl_sched_migration_cost
== -1)
5118 if (sysctl_sched_migration_cost
== 0)
5121 delta
= now
- p
->se
.exec_start
;
5123 return delta
< (s64
)sysctl_sched_migration_cost
;
5126 #ifdef CONFIG_NUMA_BALANCING
5127 /* Returns true if the destination node has incurred more faults */
5128 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
5130 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5131 int src_nid
, dst_nid
;
5133 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults_memory
||
5134 !(env
->sd
->flags
& SD_NUMA
)) {
5138 src_nid
= cpu_to_node(env
->src_cpu
);
5139 dst_nid
= cpu_to_node(env
->dst_cpu
);
5141 if (src_nid
== dst_nid
)
5145 /* Task is already in the group's interleave set. */
5146 if (node_isset(src_nid
, numa_group
->active_nodes
))
5149 /* Task is moving into the group's interleave set. */
5150 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5153 return group_faults(p
, dst_nid
) > group_faults(p
, src_nid
);
5156 /* Encourage migration to the preferred node. */
5157 if (dst_nid
== p
->numa_preferred_nid
)
5160 return task_faults(p
, dst_nid
) > task_faults(p
, src_nid
);
5164 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
5166 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5167 int src_nid
, dst_nid
;
5169 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
5172 if (!p
->numa_faults_memory
|| !(env
->sd
->flags
& SD_NUMA
))
5175 src_nid
= cpu_to_node(env
->src_cpu
);
5176 dst_nid
= cpu_to_node(env
->dst_cpu
);
5178 if (src_nid
== dst_nid
)
5182 /* Task is moving within/into the group's interleave set. */
5183 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5186 /* Task is moving out of the group's interleave set. */
5187 if (node_isset(src_nid
, numa_group
->active_nodes
))
5190 return group_faults(p
, dst_nid
) < group_faults(p
, src_nid
);
5193 /* Migrating away from the preferred node is always bad. */
5194 if (src_nid
== p
->numa_preferred_nid
)
5197 return task_faults(p
, dst_nid
) < task_faults(p
, src_nid
);
5201 static inline bool migrate_improves_locality(struct task_struct
*p
,
5207 static inline bool migrate_degrades_locality(struct task_struct
*p
,
5215 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5218 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
5220 int tsk_cache_hot
= 0;
5222 * We do not migrate tasks that are:
5223 * 1) throttled_lb_pair, or
5224 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5225 * 3) running (obviously), or
5226 * 4) are cache-hot on their current CPU.
5228 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
5231 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
5234 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
5236 env
->flags
|= LBF_SOME_PINNED
;
5239 * Remember if this task can be migrated to any other cpu in
5240 * our sched_group. We may want to revisit it if we couldn't
5241 * meet load balance goals by pulling other tasks on src_cpu.
5243 * Also avoid computing new_dst_cpu if we have already computed
5244 * one in current iteration.
5246 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
5249 /* Prevent to re-select dst_cpu via env's cpus */
5250 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
5251 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
5252 env
->flags
|= LBF_DST_PINNED
;
5253 env
->new_dst_cpu
= cpu
;
5261 /* Record that we found atleast one task that could run on dst_cpu */
5262 env
->flags
&= ~LBF_ALL_PINNED
;
5264 if (task_running(env
->src_rq
, p
)) {
5265 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
5270 * Aggressive migration if:
5271 * 1) destination numa is preferred
5272 * 2) task is cache cold, or
5273 * 3) too many balance attempts have failed.
5275 tsk_cache_hot
= task_hot(p
, rq_clock_task(env
->src_rq
));
5277 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
5279 if (migrate_improves_locality(p
, env
)) {
5280 #ifdef CONFIG_SCHEDSTATS
5281 if (tsk_cache_hot
) {
5282 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5283 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5289 if (!tsk_cache_hot
||
5290 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
5292 if (tsk_cache_hot
) {
5293 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5294 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5300 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
5305 * move_one_task tries to move exactly one task from busiest to this_rq, as
5306 * part of active balancing operations within "domain".
5307 * Returns 1 if successful and 0 otherwise.
5309 * Called with both runqueues locked.
5311 static int move_one_task(struct lb_env
*env
)
5313 struct task_struct
*p
, *n
;
5315 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
5316 if (!can_migrate_task(p
, env
))
5321 * Right now, this is only the second place move_task()
5322 * is called, so we can safely collect move_task()
5323 * stats here rather than inside move_task().
5325 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
5331 static const unsigned int sched_nr_migrate_break
= 32;
5334 * move_tasks tries to move up to imbalance weighted load from busiest to
5335 * this_rq, as part of a balancing operation within domain "sd".
5336 * Returns 1 if successful and 0 otherwise.
5338 * Called with both runqueues locked.
5340 static int move_tasks(struct lb_env
*env
)
5342 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
5343 struct task_struct
*p
;
5347 if (env
->imbalance
<= 0)
5350 while (!list_empty(tasks
)) {
5351 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5354 /* We've more or less seen every task there is, call it quits */
5355 if (env
->loop
> env
->loop_max
)
5358 /* take a breather every nr_migrate tasks */
5359 if (env
->loop
> env
->loop_break
) {
5360 env
->loop_break
+= sched_nr_migrate_break
;
5361 env
->flags
|= LBF_NEED_BREAK
;
5365 if (!can_migrate_task(p
, env
))
5368 load
= task_h_load(p
);
5370 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5373 if ((load
/ 2) > env
->imbalance
)
5378 env
->imbalance
-= load
;
5380 #ifdef CONFIG_PREEMPT
5382 * NEWIDLE balancing is a source of latency, so preemptible
5383 * kernels will stop after the first task is pulled to minimize
5384 * the critical section.
5386 if (env
->idle
== CPU_NEWLY_IDLE
)
5391 * We only want to steal up to the prescribed amount of
5394 if (env
->imbalance
<= 0)
5399 list_move_tail(&p
->se
.group_node
, tasks
);
5403 * Right now, this is one of only two places move_task() is called,
5404 * so we can safely collect move_task() stats here rather than
5405 * inside move_task().
5407 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
5412 #ifdef CONFIG_FAIR_GROUP_SCHED
5414 * update tg->load_weight by folding this cpu's load_avg
5416 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5418 struct sched_entity
*se
= tg
->se
[cpu
];
5419 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5421 /* throttled entities do not contribute to load */
5422 if (throttled_hierarchy(cfs_rq
))
5425 update_cfs_rq_blocked_load(cfs_rq
, 1);
5428 update_entity_load_avg(se
, 1);
5430 * We pivot on our runnable average having decayed to zero for
5431 * list removal. This generally implies that all our children
5432 * have also been removed (modulo rounding error or bandwidth
5433 * control); however, such cases are rare and we can fix these
5436 * TODO: fix up out-of-order children on enqueue.
5438 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5439 list_del_leaf_cfs_rq(cfs_rq
);
5441 struct rq
*rq
= rq_of(cfs_rq
);
5442 update_rq_runnable_avg(rq
, rq
->nr_running
);
5446 static void update_blocked_averages(int cpu
)
5448 struct rq
*rq
= cpu_rq(cpu
);
5449 struct cfs_rq
*cfs_rq
;
5450 unsigned long flags
;
5452 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5453 update_rq_clock(rq
);
5455 * Iterates the task_group tree in a bottom up fashion, see
5456 * list_add_leaf_cfs_rq() for details.
5458 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5460 * Note: We may want to consider periodically releasing
5461 * rq->lock about these updates so that creating many task
5462 * groups does not result in continually extending hold time.
5464 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5467 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5471 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5472 * This needs to be done in a top-down fashion because the load of a child
5473 * group is a fraction of its parents load.
5475 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5477 struct rq
*rq
= rq_of(cfs_rq
);
5478 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5479 unsigned long now
= jiffies
;
5482 if (cfs_rq
->last_h_load_update
== now
)
5485 cfs_rq
->h_load_next
= NULL
;
5486 for_each_sched_entity(se
) {
5487 cfs_rq
= cfs_rq_of(se
);
5488 cfs_rq
->h_load_next
= se
;
5489 if (cfs_rq
->last_h_load_update
== now
)
5494 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5495 cfs_rq
->last_h_load_update
= now
;
5498 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5499 load
= cfs_rq
->h_load
;
5500 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5501 cfs_rq
->runnable_load_avg
+ 1);
5502 cfs_rq
= group_cfs_rq(se
);
5503 cfs_rq
->h_load
= load
;
5504 cfs_rq
->last_h_load_update
= now
;
5508 static unsigned long task_h_load(struct task_struct
*p
)
5510 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5512 update_cfs_rq_h_load(cfs_rq
);
5513 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5514 cfs_rq
->runnable_load_avg
+ 1);
5517 static inline void update_blocked_averages(int cpu
)
5521 static unsigned long task_h_load(struct task_struct
*p
)
5523 return p
->se
.avg
.load_avg_contrib
;
5527 /********** Helpers for find_busiest_group ************************/
5529 * sg_lb_stats - stats of a sched_group required for load_balancing
5531 struct sg_lb_stats
{
5532 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5533 unsigned long group_load
; /* Total load over the CPUs of the group */
5534 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5535 unsigned long load_per_task
;
5536 unsigned long group_capacity
;
5537 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5538 unsigned int group_capacity_factor
;
5539 unsigned int idle_cpus
;
5540 unsigned int group_weight
;
5541 int group_imb
; /* Is there an imbalance in the group ? */
5542 int group_has_free_capacity
;
5543 #ifdef CONFIG_NUMA_BALANCING
5544 unsigned int nr_numa_running
;
5545 unsigned int nr_preferred_running
;
5550 * sd_lb_stats - Structure to store the statistics of a sched_domain
5551 * during load balancing.
5553 struct sd_lb_stats
{
5554 struct sched_group
*busiest
; /* Busiest group in this sd */
5555 struct sched_group
*local
; /* Local group in this sd */
5556 unsigned long total_load
; /* Total load of all groups in sd */
5557 unsigned long total_capacity
; /* Total capacity of all groups in sd */
5558 unsigned long avg_load
; /* Average load across all groups in sd */
5560 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5561 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5564 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5567 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5568 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5569 * We must however clear busiest_stat::avg_load because
5570 * update_sd_pick_busiest() reads this before assignment.
5572 *sds
= (struct sd_lb_stats
){
5576 .total_capacity
= 0UL,
5584 * get_sd_load_idx - Obtain the load index for a given sched domain.
5585 * @sd: The sched_domain whose load_idx is to be obtained.
5586 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5588 * Return: The load index.
5590 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5591 enum cpu_idle_type idle
)
5597 load_idx
= sd
->busy_idx
;
5600 case CPU_NEWLY_IDLE
:
5601 load_idx
= sd
->newidle_idx
;
5604 load_idx
= sd
->idle_idx
;
5611 static unsigned long default_scale_capacity(struct sched_domain
*sd
, int cpu
)
5613 return SCHED_CAPACITY_SCALE
;
5616 unsigned long __weak
arch_scale_freq_capacity(struct sched_domain
*sd
, int cpu
)
5618 return default_scale_capacity(sd
, cpu
);
5621 static unsigned long default_scale_smt_capacity(struct sched_domain
*sd
, int cpu
)
5623 unsigned long weight
= sd
->span_weight
;
5624 unsigned long smt_gain
= sd
->smt_gain
;
5631 unsigned long __weak
arch_scale_smt_capacity(struct sched_domain
*sd
, int cpu
)
5633 return default_scale_smt_capacity(sd
, cpu
);
5636 static unsigned long scale_rt_capacity(int cpu
)
5638 struct rq
*rq
= cpu_rq(cpu
);
5639 u64 total
, available
, age_stamp
, avg
;
5643 * Since we're reading these variables without serialization make sure
5644 * we read them once before doing sanity checks on them.
5646 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5647 avg
= ACCESS_ONCE(rq
->rt_avg
);
5649 delta
= rq_clock(rq
) - age_stamp
;
5650 if (unlikely(delta
< 0))
5653 total
= sched_avg_period() + delta
;
5655 if (unlikely(total
< avg
)) {
5656 /* Ensures that capacity won't end up being negative */
5659 available
= total
- avg
;
5662 if (unlikely((s64
)total
< SCHED_CAPACITY_SCALE
))
5663 total
= SCHED_CAPACITY_SCALE
;
5665 total
>>= SCHED_CAPACITY_SHIFT
;
5667 return div_u64(available
, total
);
5670 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5672 unsigned long weight
= sd
->span_weight
;
5673 unsigned long capacity
= SCHED_CAPACITY_SCALE
;
5674 struct sched_group
*sdg
= sd
->groups
;
5676 if ((sd
->flags
& SD_SHARE_CPUCAPACITY
) && weight
> 1) {
5677 if (sched_feat(ARCH_CAPACITY
))
5678 capacity
*= arch_scale_smt_capacity(sd
, cpu
);
5680 capacity
*= default_scale_smt_capacity(sd
, cpu
);
5682 capacity
>>= SCHED_CAPACITY_SHIFT
;
5685 sdg
->sgc
->capacity_orig
= capacity
;
5687 if (sched_feat(ARCH_CAPACITY
))
5688 capacity
*= arch_scale_freq_capacity(sd
, cpu
);
5690 capacity
*= default_scale_capacity(sd
, cpu
);
5692 capacity
>>= SCHED_CAPACITY_SHIFT
;
5694 capacity
*= scale_rt_capacity(cpu
);
5695 capacity
>>= SCHED_CAPACITY_SHIFT
;
5700 cpu_rq(cpu
)->cpu_capacity
= capacity
;
5701 sdg
->sgc
->capacity
= capacity
;
5704 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
5706 struct sched_domain
*child
= sd
->child
;
5707 struct sched_group
*group
, *sdg
= sd
->groups
;
5708 unsigned long capacity
, capacity_orig
;
5709 unsigned long interval
;
5711 interval
= msecs_to_jiffies(sd
->balance_interval
);
5712 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5713 sdg
->sgc
->next_update
= jiffies
+ interval
;
5716 update_cpu_capacity(sd
, cpu
);
5720 capacity_orig
= capacity
= 0;
5722 if (child
->flags
& SD_OVERLAP
) {
5724 * SD_OVERLAP domains cannot assume that child groups
5725 * span the current group.
5728 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
5729 struct sched_group_capacity
*sgc
;
5730 struct rq
*rq
= cpu_rq(cpu
);
5733 * build_sched_domains() -> init_sched_groups_capacity()
5734 * gets here before we've attached the domains to the
5737 * Use capacity_of(), which is set irrespective of domains
5738 * in update_cpu_capacity().
5740 * This avoids capacity/capacity_orig from being 0 and
5741 * causing divide-by-zero issues on boot.
5743 * Runtime updates will correct capacity_orig.
5745 if (unlikely(!rq
->sd
)) {
5746 capacity_orig
+= capacity_of(cpu
);
5747 capacity
+= capacity_of(cpu
);
5751 sgc
= rq
->sd
->groups
->sgc
;
5752 capacity_orig
+= sgc
->capacity_orig
;
5753 capacity
+= sgc
->capacity
;
5757 * !SD_OVERLAP domains can assume that child groups
5758 * span the current group.
5761 group
= child
->groups
;
5763 capacity_orig
+= group
->sgc
->capacity_orig
;
5764 capacity
+= group
->sgc
->capacity
;
5765 group
= group
->next
;
5766 } while (group
!= child
->groups
);
5769 sdg
->sgc
->capacity_orig
= capacity_orig
;
5770 sdg
->sgc
->capacity
= capacity
;
5774 * Try and fix up capacity for tiny siblings, this is needed when
5775 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5776 * which on its own isn't powerful enough.
5778 * See update_sd_pick_busiest() and check_asym_packing().
5781 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
5784 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
5786 if (!(sd
->flags
& SD_SHARE_CPUCAPACITY
))
5790 * If ~90% of the cpu_capacity is still there, we're good.
5792 if (group
->sgc
->capacity
* 32 > group
->sgc
->capacity_orig
* 29)
5799 * Group imbalance indicates (and tries to solve) the problem where balancing
5800 * groups is inadequate due to tsk_cpus_allowed() constraints.
5802 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5803 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5806 * { 0 1 2 3 } { 4 5 6 7 }
5809 * If we were to balance group-wise we'd place two tasks in the first group and
5810 * two tasks in the second group. Clearly this is undesired as it will overload
5811 * cpu 3 and leave one of the cpus in the second group unused.
5813 * The current solution to this issue is detecting the skew in the first group
5814 * by noticing the lower domain failed to reach balance and had difficulty
5815 * moving tasks due to affinity constraints.
5817 * When this is so detected; this group becomes a candidate for busiest; see
5818 * update_sd_pick_busiest(). And calculate_imbalance() and
5819 * find_busiest_group() avoid some of the usual balance conditions to allow it
5820 * to create an effective group imbalance.
5822 * This is a somewhat tricky proposition since the next run might not find the
5823 * group imbalance and decide the groups need to be balanced again. A most
5824 * subtle and fragile situation.
5827 static inline int sg_imbalanced(struct sched_group
*group
)
5829 return group
->sgc
->imbalance
;
5833 * Compute the group capacity factor.
5835 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
5836 * first dividing out the smt factor and computing the actual number of cores
5837 * and limit unit capacity with that.
5839 static inline int sg_capacity_factor(struct lb_env
*env
, struct sched_group
*group
)
5841 unsigned int capacity_factor
, smt
, cpus
;
5842 unsigned int capacity
, capacity_orig
;
5844 capacity
= group
->sgc
->capacity
;
5845 capacity_orig
= group
->sgc
->capacity_orig
;
5846 cpus
= group
->group_weight
;
5848 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
5849 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, capacity_orig
);
5850 capacity_factor
= cpus
/ smt
; /* cores */
5852 capacity_factor
= min_t(unsigned,
5853 capacity_factor
, DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
));
5854 if (!capacity_factor
)
5855 capacity_factor
= fix_small_capacity(env
->sd
, group
);
5857 return capacity_factor
;
5861 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5862 * @env: The load balancing environment.
5863 * @group: sched_group whose statistics are to be updated.
5864 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5865 * @local_group: Does group contain this_cpu.
5866 * @sgs: variable to hold the statistics for this group.
5868 static inline void update_sg_lb_stats(struct lb_env
*env
,
5869 struct sched_group
*group
, int load_idx
,
5870 int local_group
, struct sg_lb_stats
*sgs
)
5875 memset(sgs
, 0, sizeof(*sgs
));
5877 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5878 struct rq
*rq
= cpu_rq(i
);
5880 /* Bias balancing toward cpus of our domain */
5882 load
= target_load(i
, load_idx
);
5884 load
= source_load(i
, load_idx
);
5886 sgs
->group_load
+= load
;
5887 sgs
->sum_nr_running
+= rq
->nr_running
;
5888 #ifdef CONFIG_NUMA_BALANCING
5889 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
5890 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
5892 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
5897 /* Adjust by relative CPU capacity of the group */
5898 sgs
->group_capacity
= group
->sgc
->capacity
;
5899 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
5901 if (sgs
->sum_nr_running
)
5902 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
5904 sgs
->group_weight
= group
->group_weight
;
5906 sgs
->group_imb
= sg_imbalanced(group
);
5907 sgs
->group_capacity_factor
= sg_capacity_factor(env
, group
);
5909 if (sgs
->group_capacity_factor
> sgs
->sum_nr_running
)
5910 sgs
->group_has_free_capacity
= 1;
5914 * update_sd_pick_busiest - return 1 on busiest group
5915 * @env: The load balancing environment.
5916 * @sds: sched_domain statistics
5917 * @sg: sched_group candidate to be checked for being the busiest
5918 * @sgs: sched_group statistics
5920 * Determine if @sg is a busier group than the previously selected
5923 * Return: %true if @sg is a busier group than the previously selected
5924 * busiest group. %false otherwise.
5926 static bool update_sd_pick_busiest(struct lb_env
*env
,
5927 struct sd_lb_stats
*sds
,
5928 struct sched_group
*sg
,
5929 struct sg_lb_stats
*sgs
)
5931 if (sgs
->avg_load
<= sds
->busiest_stat
.avg_load
)
5934 if (sgs
->sum_nr_running
> sgs
->group_capacity_factor
)
5941 * ASYM_PACKING needs to move all the work to the lowest
5942 * numbered CPUs in the group, therefore mark all groups
5943 * higher than ourself as busy.
5945 if ((env
->sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
5946 env
->dst_cpu
< group_first_cpu(sg
)) {
5950 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
5957 #ifdef CONFIG_NUMA_BALANCING
5958 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
5960 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
5962 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
5967 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
5969 if (rq
->nr_running
> rq
->nr_numa_running
)
5971 if (rq
->nr_running
> rq
->nr_preferred_running
)
5976 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
5981 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
5985 #endif /* CONFIG_NUMA_BALANCING */
5988 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5989 * @env: The load balancing environment.
5990 * @sds: variable to hold the statistics for this sched_domain.
5992 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5994 struct sched_domain
*child
= env
->sd
->child
;
5995 struct sched_group
*sg
= env
->sd
->groups
;
5996 struct sg_lb_stats tmp_sgs
;
5997 int load_idx
, prefer_sibling
= 0;
5999 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
6002 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
6005 struct sg_lb_stats
*sgs
= &tmp_sgs
;
6008 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
6011 sgs
= &sds
->local_stat
;
6013 if (env
->idle
!= CPU_NEWLY_IDLE
||
6014 time_after_eq(jiffies
, sg
->sgc
->next_update
))
6015 update_group_capacity(env
->sd
, env
->dst_cpu
);
6018 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
);
6024 * In case the child domain prefers tasks go to siblings
6025 * first, lower the sg capacity factor to one so that we'll try
6026 * and move all the excess tasks away. We lower the capacity
6027 * of a group only if the local group has the capacity to fit
6028 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6029 * extra check prevents the case where you always pull from the
6030 * heaviest group when it is already under-utilized (possible
6031 * with a large weight task outweighs the tasks on the system).
6033 if (prefer_sibling
&& sds
->local
&&
6034 sds
->local_stat
.group_has_free_capacity
)
6035 sgs
->group_capacity_factor
= min(sgs
->group_capacity_factor
, 1U);
6037 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
6039 sds
->busiest_stat
= *sgs
;
6043 /* Now, start updating sd_lb_stats */
6044 sds
->total_load
+= sgs
->group_load
;
6045 sds
->total_capacity
+= sgs
->group_capacity
;
6048 } while (sg
!= env
->sd
->groups
);
6050 if (env
->sd
->flags
& SD_NUMA
)
6051 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
6055 * check_asym_packing - Check to see if the group is packed into the
6058 * This is primarily intended to used at the sibling level. Some
6059 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6060 * case of POWER7, it can move to lower SMT modes only when higher
6061 * threads are idle. When in lower SMT modes, the threads will
6062 * perform better since they share less core resources. Hence when we
6063 * have idle threads, we want them to be the higher ones.
6065 * This packing function is run on idle threads. It checks to see if
6066 * the busiest CPU in this domain (core in the P7 case) has a higher
6067 * CPU number than the packing function is being run on. Here we are
6068 * assuming lower CPU number will be equivalent to lower a SMT thread
6071 * Return: 1 when packing is required and a task should be moved to
6072 * this CPU. The amount of the imbalance is returned in *imbalance.
6074 * @env: The load balancing environment.
6075 * @sds: Statistics of the sched_domain which is to be packed
6077 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6081 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6087 busiest_cpu
= group_first_cpu(sds
->busiest
);
6088 if (env
->dst_cpu
> busiest_cpu
)
6091 env
->imbalance
= DIV_ROUND_CLOSEST(
6092 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
6093 SCHED_CAPACITY_SCALE
);
6099 * fix_small_imbalance - Calculate the minor imbalance that exists
6100 * amongst the groups of a sched_domain, during
6102 * @env: The load balancing environment.
6103 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6106 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6108 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
6109 unsigned int imbn
= 2;
6110 unsigned long scaled_busy_load_per_task
;
6111 struct sg_lb_stats
*local
, *busiest
;
6113 local
= &sds
->local_stat
;
6114 busiest
= &sds
->busiest_stat
;
6116 if (!local
->sum_nr_running
)
6117 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
6118 else if (busiest
->load_per_task
> local
->load_per_task
)
6121 scaled_busy_load_per_task
=
6122 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6123 busiest
->group_capacity
;
6125 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
6126 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
6127 env
->imbalance
= busiest
->load_per_task
;
6132 * OK, we don't have enough imbalance to justify moving tasks,
6133 * however we may be able to increase total CPU capacity used by
6137 capa_now
+= busiest
->group_capacity
*
6138 min(busiest
->load_per_task
, busiest
->avg_load
);
6139 capa_now
+= local
->group_capacity
*
6140 min(local
->load_per_task
, local
->avg_load
);
6141 capa_now
/= SCHED_CAPACITY_SCALE
;
6143 /* Amount of load we'd subtract */
6144 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
6145 capa_move
+= busiest
->group_capacity
*
6146 min(busiest
->load_per_task
,
6147 busiest
->avg_load
- scaled_busy_load_per_task
);
6150 /* Amount of load we'd add */
6151 if (busiest
->avg_load
* busiest
->group_capacity
<
6152 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
6153 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
6154 local
->group_capacity
;
6156 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6157 local
->group_capacity
;
6159 capa_move
+= local
->group_capacity
*
6160 min(local
->load_per_task
, local
->avg_load
+ tmp
);
6161 capa_move
/= SCHED_CAPACITY_SCALE
;
6163 /* Move if we gain throughput */
6164 if (capa_move
> capa_now
)
6165 env
->imbalance
= busiest
->load_per_task
;
6169 * calculate_imbalance - Calculate the amount of imbalance present within the
6170 * groups of a given sched_domain during load balance.
6171 * @env: load balance environment
6172 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6174 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6176 unsigned long max_pull
, load_above_capacity
= ~0UL;
6177 struct sg_lb_stats
*local
, *busiest
;
6179 local
= &sds
->local_stat
;
6180 busiest
= &sds
->busiest_stat
;
6182 if (busiest
->group_imb
) {
6184 * In the group_imb case we cannot rely on group-wide averages
6185 * to ensure cpu-load equilibrium, look at wider averages. XXX
6187 busiest
->load_per_task
=
6188 min(busiest
->load_per_task
, sds
->avg_load
);
6192 * In the presence of smp nice balancing, certain scenarios can have
6193 * max load less than avg load(as we skip the groups at or below
6194 * its cpu_capacity, while calculating max_load..)
6196 if (busiest
->avg_load
<= sds
->avg_load
||
6197 local
->avg_load
>= sds
->avg_load
) {
6199 return fix_small_imbalance(env
, sds
);
6202 if (!busiest
->group_imb
) {
6204 * Don't want to pull so many tasks that a group would go idle.
6205 * Except of course for the group_imb case, since then we might
6206 * have to drop below capacity to reach cpu-load equilibrium.
6208 load_above_capacity
=
6209 (busiest
->sum_nr_running
- busiest
->group_capacity_factor
);
6211 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_CAPACITY_SCALE
);
6212 load_above_capacity
/= busiest
->group_capacity
;
6216 * We're trying to get all the cpus to the average_load, so we don't
6217 * want to push ourselves above the average load, nor do we wish to
6218 * reduce the max loaded cpu below the average load. At the same time,
6219 * we also don't want to reduce the group load below the group capacity
6220 * (so that we can implement power-savings policies etc). Thus we look
6221 * for the minimum possible imbalance.
6223 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
6225 /* How much load to actually move to equalise the imbalance */
6226 env
->imbalance
= min(
6227 max_pull
* busiest
->group_capacity
,
6228 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
6229 ) / SCHED_CAPACITY_SCALE
;
6232 * if *imbalance is less than the average load per runnable task
6233 * there is no guarantee that any tasks will be moved so we'll have
6234 * a think about bumping its value to force at least one task to be
6237 if (env
->imbalance
< busiest
->load_per_task
)
6238 return fix_small_imbalance(env
, sds
);
6241 /******* find_busiest_group() helpers end here *********************/
6244 * find_busiest_group - Returns the busiest group within the sched_domain
6245 * if there is an imbalance. If there isn't an imbalance, and
6246 * the user has opted for power-savings, it returns a group whose
6247 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6248 * such a group exists.
6250 * Also calculates the amount of weighted load which should be moved
6251 * to restore balance.
6253 * @env: The load balancing environment.
6255 * Return: - The busiest group if imbalance exists.
6256 * - If no imbalance and user has opted for power-savings balance,
6257 * return the least loaded group whose CPUs can be
6258 * put to idle by rebalancing its tasks onto our group.
6260 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
6262 struct sg_lb_stats
*local
, *busiest
;
6263 struct sd_lb_stats sds
;
6265 init_sd_lb_stats(&sds
);
6268 * Compute the various statistics relavent for load balancing at
6271 update_sd_lb_stats(env
, &sds
);
6272 local
= &sds
.local_stat
;
6273 busiest
= &sds
.busiest_stat
;
6275 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
6276 check_asym_packing(env
, &sds
))
6279 /* There is no busy sibling group to pull tasks from */
6280 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
6283 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
6284 / sds
.total_capacity
;
6287 * If the busiest group is imbalanced the below checks don't
6288 * work because they assume all things are equal, which typically
6289 * isn't true due to cpus_allowed constraints and the like.
6291 if (busiest
->group_imb
)
6294 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6295 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_free_capacity
&&
6296 !busiest
->group_has_free_capacity
)
6300 * If the local group is more busy than the selected busiest group
6301 * don't try and pull any tasks.
6303 if (local
->avg_load
>= busiest
->avg_load
)
6307 * Don't pull any tasks if this group is already above the domain
6310 if (local
->avg_load
>= sds
.avg_load
)
6313 if (env
->idle
== CPU_IDLE
) {
6315 * This cpu is idle. If the busiest group load doesn't
6316 * have more tasks than the number of available cpu's and
6317 * there is no imbalance between this and busiest group
6318 * wrt to idle cpu's, it is balanced.
6320 if ((local
->idle_cpus
< busiest
->idle_cpus
) &&
6321 busiest
->sum_nr_running
<= busiest
->group_weight
)
6325 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6326 * imbalance_pct to be conservative.
6328 if (100 * busiest
->avg_load
<=
6329 env
->sd
->imbalance_pct
* local
->avg_load
)
6334 /* Looks like there is an imbalance. Compute it */
6335 calculate_imbalance(env
, &sds
);
6344 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6346 static struct rq
*find_busiest_queue(struct lb_env
*env
,
6347 struct sched_group
*group
)
6349 struct rq
*busiest
= NULL
, *rq
;
6350 unsigned long busiest_load
= 0, busiest_capacity
= 1;
6353 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6354 unsigned long capacity
, capacity_factor
, wl
;
6358 rt
= fbq_classify_rq(rq
);
6361 * We classify groups/runqueues into three groups:
6362 * - regular: there are !numa tasks
6363 * - remote: there are numa tasks that run on the 'wrong' node
6364 * - all: there is no distinction
6366 * In order to avoid migrating ideally placed numa tasks,
6367 * ignore those when there's better options.
6369 * If we ignore the actual busiest queue to migrate another
6370 * task, the next balance pass can still reduce the busiest
6371 * queue by moving tasks around inside the node.
6373 * If we cannot move enough load due to this classification
6374 * the next pass will adjust the group classification and
6375 * allow migration of more tasks.
6377 * Both cases only affect the total convergence complexity.
6379 if (rt
> env
->fbq_type
)
6382 capacity
= capacity_of(i
);
6383 capacity_factor
= DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
);
6384 if (!capacity_factor
)
6385 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6387 wl
= weighted_cpuload(i
);
6390 * When comparing with imbalance, use weighted_cpuload()
6391 * which is not scaled with the cpu capacity.
6393 if (capacity_factor
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6397 * For the load comparisons with the other cpu's, consider
6398 * the weighted_cpuload() scaled with the cpu capacity, so
6399 * that the load can be moved away from the cpu that is
6400 * potentially running at a lower capacity.
6402 * Thus we're looking for max(wl_i / capacity_i), crosswise
6403 * multiplication to rid ourselves of the division works out
6404 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6405 * our previous maximum.
6407 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
6409 busiest_capacity
= capacity
;
6418 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6419 * so long as it is large enough.
6421 #define MAX_PINNED_INTERVAL 512
6423 /* Working cpumask for load_balance and load_balance_newidle. */
6424 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6426 static int need_active_balance(struct lb_env
*env
)
6428 struct sched_domain
*sd
= env
->sd
;
6430 if (env
->idle
== CPU_NEWLY_IDLE
) {
6433 * ASYM_PACKING needs to force migrate tasks from busy but
6434 * higher numbered CPUs in order to pack all tasks in the
6435 * lowest numbered CPUs.
6437 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6441 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6444 static int active_load_balance_cpu_stop(void *data
);
6446 static int should_we_balance(struct lb_env
*env
)
6448 struct sched_group
*sg
= env
->sd
->groups
;
6449 struct cpumask
*sg_cpus
, *sg_mask
;
6450 int cpu
, balance_cpu
= -1;
6453 * In the newly idle case, we will allow all the cpu's
6454 * to do the newly idle load balance.
6456 if (env
->idle
== CPU_NEWLY_IDLE
)
6459 sg_cpus
= sched_group_cpus(sg
);
6460 sg_mask
= sched_group_mask(sg
);
6461 /* Try to find first idle cpu */
6462 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6463 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6470 if (balance_cpu
== -1)
6471 balance_cpu
= group_balance_cpu(sg
);
6474 * First idle cpu or the first cpu(busiest) in this sched group
6475 * is eligible for doing load balancing at this and above domains.
6477 return balance_cpu
== env
->dst_cpu
;
6481 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6482 * tasks if there is an imbalance.
6484 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6485 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6486 int *continue_balancing
)
6488 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6489 struct sched_domain
*sd_parent
= sd
->parent
;
6490 struct sched_group
*group
;
6492 unsigned long flags
;
6493 struct cpumask
*cpus
= __get_cpu_var(load_balance_mask
);
6495 struct lb_env env
= {
6497 .dst_cpu
= this_cpu
,
6499 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6501 .loop_break
= sched_nr_migrate_break
,
6507 * For NEWLY_IDLE load_balancing, we don't need to consider
6508 * other cpus in our group
6510 if (idle
== CPU_NEWLY_IDLE
)
6511 env
.dst_grpmask
= NULL
;
6513 cpumask_copy(cpus
, cpu_active_mask
);
6515 schedstat_inc(sd
, lb_count
[idle
]);
6518 if (!should_we_balance(&env
)) {
6519 *continue_balancing
= 0;
6523 group
= find_busiest_group(&env
);
6525 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6529 busiest
= find_busiest_queue(&env
, group
);
6531 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6535 BUG_ON(busiest
== env
.dst_rq
);
6537 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6540 if (busiest
->nr_running
> 1) {
6542 * Attempt to move tasks. If find_busiest_group has found
6543 * an imbalance but busiest->nr_running <= 1, the group is
6544 * still unbalanced. ld_moved simply stays zero, so it is
6545 * correctly treated as an imbalance.
6547 env
.flags
|= LBF_ALL_PINNED
;
6548 env
.src_cpu
= busiest
->cpu
;
6549 env
.src_rq
= busiest
;
6550 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6553 local_irq_save(flags
);
6554 double_rq_lock(env
.dst_rq
, busiest
);
6557 * cur_ld_moved - load moved in current iteration
6558 * ld_moved - cumulative load moved across iterations
6560 cur_ld_moved
= move_tasks(&env
);
6561 ld_moved
+= cur_ld_moved
;
6562 double_rq_unlock(env
.dst_rq
, busiest
);
6563 local_irq_restore(flags
);
6566 * some other cpu did the load balance for us.
6568 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
6569 resched_cpu(env
.dst_cpu
);
6571 if (env
.flags
& LBF_NEED_BREAK
) {
6572 env
.flags
&= ~LBF_NEED_BREAK
;
6577 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6578 * us and move them to an alternate dst_cpu in our sched_group
6579 * where they can run. The upper limit on how many times we
6580 * iterate on same src_cpu is dependent on number of cpus in our
6583 * This changes load balance semantics a bit on who can move
6584 * load to a given_cpu. In addition to the given_cpu itself
6585 * (or a ilb_cpu acting on its behalf where given_cpu is
6586 * nohz-idle), we now have balance_cpu in a position to move
6587 * load to given_cpu. In rare situations, this may cause
6588 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6589 * _independently_ and at _same_ time to move some load to
6590 * given_cpu) causing exceess load to be moved to given_cpu.
6591 * This however should not happen so much in practice and
6592 * moreover subsequent load balance cycles should correct the
6593 * excess load moved.
6595 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6597 /* Prevent to re-select dst_cpu via env's cpus */
6598 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6600 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6601 env
.dst_cpu
= env
.new_dst_cpu
;
6602 env
.flags
&= ~LBF_DST_PINNED
;
6604 env
.loop_break
= sched_nr_migrate_break
;
6607 * Go back to "more_balance" rather than "redo" since we
6608 * need to continue with same src_cpu.
6614 * We failed to reach balance because of affinity.
6617 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
6619 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0) {
6620 *group_imbalance
= 1;
6621 } else if (*group_imbalance
)
6622 *group_imbalance
= 0;
6625 /* All tasks on this runqueue were pinned by CPU affinity */
6626 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6627 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6628 if (!cpumask_empty(cpus
)) {
6630 env
.loop_break
= sched_nr_migrate_break
;
6638 schedstat_inc(sd
, lb_failed
[idle
]);
6640 * Increment the failure counter only on periodic balance.
6641 * We do not want newidle balance, which can be very
6642 * frequent, pollute the failure counter causing
6643 * excessive cache_hot migrations and active balances.
6645 if (idle
!= CPU_NEWLY_IDLE
)
6646 sd
->nr_balance_failed
++;
6648 if (need_active_balance(&env
)) {
6649 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6651 /* don't kick the active_load_balance_cpu_stop,
6652 * if the curr task on busiest cpu can't be
6655 if (!cpumask_test_cpu(this_cpu
,
6656 tsk_cpus_allowed(busiest
->curr
))) {
6657 raw_spin_unlock_irqrestore(&busiest
->lock
,
6659 env
.flags
|= LBF_ALL_PINNED
;
6660 goto out_one_pinned
;
6664 * ->active_balance synchronizes accesses to
6665 * ->active_balance_work. Once set, it's cleared
6666 * only after active load balance is finished.
6668 if (!busiest
->active_balance
) {
6669 busiest
->active_balance
= 1;
6670 busiest
->push_cpu
= this_cpu
;
6673 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
6675 if (active_balance
) {
6676 stop_one_cpu_nowait(cpu_of(busiest
),
6677 active_load_balance_cpu_stop
, busiest
,
6678 &busiest
->active_balance_work
);
6682 * We've kicked active balancing, reset the failure
6685 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
6688 sd
->nr_balance_failed
= 0;
6690 if (likely(!active_balance
)) {
6691 /* We were unbalanced, so reset the balancing interval */
6692 sd
->balance_interval
= sd
->min_interval
;
6695 * If we've begun active balancing, start to back off. This
6696 * case may not be covered by the all_pinned logic if there
6697 * is only 1 task on the busy runqueue (because we don't call
6700 if (sd
->balance_interval
< sd
->max_interval
)
6701 sd
->balance_interval
*= 2;
6707 schedstat_inc(sd
, lb_balanced
[idle
]);
6709 sd
->nr_balance_failed
= 0;
6712 /* tune up the balancing interval */
6713 if (((env
.flags
& LBF_ALL_PINNED
) &&
6714 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
6715 (sd
->balance_interval
< sd
->max_interval
))
6716 sd
->balance_interval
*= 2;
6723 static inline unsigned long
6724 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
6726 unsigned long interval
= sd
->balance_interval
;
6729 interval
*= sd
->busy_factor
;
6731 /* scale ms to jiffies */
6732 interval
= msecs_to_jiffies(interval
);
6733 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6739 update_next_balance(struct sched_domain
*sd
, int cpu_busy
, unsigned long *next_balance
)
6741 unsigned long interval
, next
;
6743 interval
= get_sd_balance_interval(sd
, cpu_busy
);
6744 next
= sd
->last_balance
+ interval
;
6746 if (time_after(*next_balance
, next
))
6747 *next_balance
= next
;
6751 * idle_balance is called by schedule() if this_cpu is about to become
6752 * idle. Attempts to pull tasks from other CPUs.
6754 static int idle_balance(struct rq
*this_rq
)
6756 unsigned long next_balance
= jiffies
+ HZ
;
6757 int this_cpu
= this_rq
->cpu
;
6758 struct sched_domain
*sd
;
6759 int pulled_task
= 0;
6762 idle_enter_fair(this_rq
);
6765 * We must set idle_stamp _before_ calling idle_balance(), such that we
6766 * measure the duration of idle_balance() as idle time.
6768 this_rq
->idle_stamp
= rq_clock(this_rq
);
6770 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
) {
6772 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
6774 update_next_balance(sd
, 0, &next_balance
);
6781 * Drop the rq->lock, but keep IRQ/preempt disabled.
6783 raw_spin_unlock(&this_rq
->lock
);
6785 update_blocked_averages(this_cpu
);
6787 for_each_domain(this_cpu
, sd
) {
6788 int continue_balancing
= 1;
6789 u64 t0
, domain_cost
;
6791 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6794 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
6795 update_next_balance(sd
, 0, &next_balance
);
6799 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
6800 t0
= sched_clock_cpu(this_cpu
);
6802 pulled_task
= load_balance(this_cpu
, this_rq
,
6804 &continue_balancing
);
6806 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
6807 if (domain_cost
> sd
->max_newidle_lb_cost
)
6808 sd
->max_newidle_lb_cost
= domain_cost
;
6810 curr_cost
+= domain_cost
;
6813 update_next_balance(sd
, 0, &next_balance
);
6816 * Stop searching for tasks to pull if there are
6817 * now runnable tasks on this rq.
6819 if (pulled_task
|| this_rq
->nr_running
> 0)
6824 raw_spin_lock(&this_rq
->lock
);
6826 if (curr_cost
> this_rq
->max_idle_balance_cost
)
6827 this_rq
->max_idle_balance_cost
= curr_cost
;
6830 * While browsing the domains, we released the rq lock, a task could
6831 * have been enqueued in the meantime. Since we're not going idle,
6832 * pretend we pulled a task.
6834 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
6838 /* Move the next balance forward */
6839 if (time_after(this_rq
->next_balance
, next_balance
))
6840 this_rq
->next_balance
= next_balance
;
6842 /* Is there a task of a high priority class? */
6843 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
6847 idle_exit_fair(this_rq
);
6848 this_rq
->idle_stamp
= 0;
6855 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6856 * running tasks off the busiest CPU onto idle CPUs. It requires at
6857 * least 1 task to be running on each physical CPU where possible, and
6858 * avoids physical / logical imbalances.
6860 static int active_load_balance_cpu_stop(void *data
)
6862 struct rq
*busiest_rq
= data
;
6863 int busiest_cpu
= cpu_of(busiest_rq
);
6864 int target_cpu
= busiest_rq
->push_cpu
;
6865 struct rq
*target_rq
= cpu_rq(target_cpu
);
6866 struct sched_domain
*sd
;
6868 raw_spin_lock_irq(&busiest_rq
->lock
);
6870 /* make sure the requested cpu hasn't gone down in the meantime */
6871 if (unlikely(busiest_cpu
!= smp_processor_id() ||
6872 !busiest_rq
->active_balance
))
6875 /* Is there any task to move? */
6876 if (busiest_rq
->nr_running
<= 1)
6880 * This condition is "impossible", if it occurs
6881 * we need to fix it. Originally reported by
6882 * Bjorn Helgaas on a 128-cpu setup.
6884 BUG_ON(busiest_rq
== target_rq
);
6886 /* move a task from busiest_rq to target_rq */
6887 double_lock_balance(busiest_rq
, target_rq
);
6889 /* Search for an sd spanning us and the target CPU. */
6891 for_each_domain(target_cpu
, sd
) {
6892 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
6893 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
6898 struct lb_env env
= {
6900 .dst_cpu
= target_cpu
,
6901 .dst_rq
= target_rq
,
6902 .src_cpu
= busiest_rq
->cpu
,
6903 .src_rq
= busiest_rq
,
6907 schedstat_inc(sd
, alb_count
);
6909 if (move_one_task(&env
))
6910 schedstat_inc(sd
, alb_pushed
);
6912 schedstat_inc(sd
, alb_failed
);
6915 double_unlock_balance(busiest_rq
, target_rq
);
6917 busiest_rq
->active_balance
= 0;
6918 raw_spin_unlock_irq(&busiest_rq
->lock
);
6922 static inline int on_null_domain(struct rq
*rq
)
6924 return unlikely(!rcu_dereference_sched(rq
->sd
));
6927 #ifdef CONFIG_NO_HZ_COMMON
6929 * idle load balancing details
6930 * - When one of the busy CPUs notice that there may be an idle rebalancing
6931 * needed, they will kick the idle load balancer, which then does idle
6932 * load balancing for all the idle CPUs.
6935 cpumask_var_t idle_cpus_mask
;
6937 unsigned long next_balance
; /* in jiffy units */
6938 } nohz ____cacheline_aligned
;
6940 static inline int find_new_ilb(void)
6942 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
6944 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
6951 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6952 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6953 * CPU (if there is one).
6955 static void nohz_balancer_kick(void)
6959 nohz
.next_balance
++;
6961 ilb_cpu
= find_new_ilb();
6963 if (ilb_cpu
>= nr_cpu_ids
)
6966 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
6969 * Use smp_send_reschedule() instead of resched_cpu().
6970 * This way we generate a sched IPI on the target cpu which
6971 * is idle. And the softirq performing nohz idle load balance
6972 * will be run before returning from the IPI.
6974 smp_send_reschedule(ilb_cpu
);
6978 static inline void nohz_balance_exit_idle(int cpu
)
6980 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
6982 * Completely isolated CPUs don't ever set, so we must test.
6984 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
6985 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
6986 atomic_dec(&nohz
.nr_cpus
);
6988 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
6992 static inline void set_cpu_sd_state_busy(void)
6994 struct sched_domain
*sd
;
6995 int cpu
= smp_processor_id();
6998 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7000 if (!sd
|| !sd
->nohz_idle
)
7004 atomic_inc(&sd
->groups
->sgc
->nr_busy_cpus
);
7009 void set_cpu_sd_state_idle(void)
7011 struct sched_domain
*sd
;
7012 int cpu
= smp_processor_id();
7015 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7017 if (!sd
|| sd
->nohz_idle
)
7021 atomic_dec(&sd
->groups
->sgc
->nr_busy_cpus
);
7027 * This routine will record that the cpu is going idle with tick stopped.
7028 * This info will be used in performing idle load balancing in the future.
7030 void nohz_balance_enter_idle(int cpu
)
7033 * If this cpu is going down, then nothing needs to be done.
7035 if (!cpu_active(cpu
))
7038 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
7042 * If we're a completely isolated CPU, we don't play.
7044 if (on_null_domain(cpu_rq(cpu
)))
7047 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
7048 atomic_inc(&nohz
.nr_cpus
);
7049 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7052 static int sched_ilb_notifier(struct notifier_block
*nfb
,
7053 unsigned long action
, void *hcpu
)
7055 switch (action
& ~CPU_TASKS_FROZEN
) {
7057 nohz_balance_exit_idle(smp_processor_id());
7065 static DEFINE_SPINLOCK(balancing
);
7068 * Scale the max load_balance interval with the number of CPUs in the system.
7069 * This trades load-balance latency on larger machines for less cross talk.
7071 void update_max_interval(void)
7073 max_load_balance_interval
= HZ
*num_online_cpus()/10;
7077 * It checks each scheduling domain to see if it is due to be balanced,
7078 * and initiates a balancing operation if so.
7080 * Balancing parameters are set up in init_sched_domains.
7082 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
7084 int continue_balancing
= 1;
7086 unsigned long interval
;
7087 struct sched_domain
*sd
;
7088 /* Earliest time when we have to do rebalance again */
7089 unsigned long next_balance
= jiffies
+ 60*HZ
;
7090 int update_next_balance
= 0;
7091 int need_serialize
, need_decay
= 0;
7094 update_blocked_averages(cpu
);
7097 for_each_domain(cpu
, sd
) {
7099 * Decay the newidle max times here because this is a regular
7100 * visit to all the domains. Decay ~1% per second.
7102 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
7103 sd
->max_newidle_lb_cost
=
7104 (sd
->max_newidle_lb_cost
* 253) / 256;
7105 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
7108 max_cost
+= sd
->max_newidle_lb_cost
;
7110 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7114 * Stop the load balance at this level. There is another
7115 * CPU in our sched group which is doing load balancing more
7118 if (!continue_balancing
) {
7124 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7126 need_serialize
= sd
->flags
& SD_SERIALIZE
;
7127 if (need_serialize
) {
7128 if (!spin_trylock(&balancing
))
7132 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
7133 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
7135 * The LBF_DST_PINNED logic could have changed
7136 * env->dst_cpu, so we can't know our idle
7137 * state even if we migrated tasks. Update it.
7139 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
7141 sd
->last_balance
= jiffies
;
7142 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7145 spin_unlock(&balancing
);
7147 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
7148 next_balance
= sd
->last_balance
+ interval
;
7149 update_next_balance
= 1;
7154 * Ensure the rq-wide value also decays but keep it at a
7155 * reasonable floor to avoid funnies with rq->avg_idle.
7157 rq
->max_idle_balance_cost
=
7158 max((u64
)sysctl_sched_migration_cost
, max_cost
);
7163 * next_balance will be updated only when there is a need.
7164 * When the cpu is attached to null domain for ex, it will not be
7167 if (likely(update_next_balance
))
7168 rq
->next_balance
= next_balance
;
7171 #ifdef CONFIG_NO_HZ_COMMON
7173 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7174 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7176 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
7178 int this_cpu
= this_rq
->cpu
;
7182 if (idle
!= CPU_IDLE
||
7183 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
7186 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
7187 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
7191 * If this cpu gets work to do, stop the load balancing
7192 * work being done for other cpus. Next load
7193 * balancing owner will pick it up.
7198 rq
= cpu_rq(balance_cpu
);
7201 * If time for next balance is due,
7204 if (time_after_eq(jiffies
, rq
->next_balance
)) {
7205 raw_spin_lock_irq(&rq
->lock
);
7206 update_rq_clock(rq
);
7207 update_idle_cpu_load(rq
);
7208 raw_spin_unlock_irq(&rq
->lock
);
7209 rebalance_domains(rq
, CPU_IDLE
);
7212 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
7213 this_rq
->next_balance
= rq
->next_balance
;
7215 nohz
.next_balance
= this_rq
->next_balance
;
7217 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
7221 * Current heuristic for kicking the idle load balancer in the presence
7222 * of an idle cpu is the system.
7223 * - This rq has more than one task.
7224 * - At any scheduler domain level, this cpu's scheduler group has multiple
7225 * busy cpu's exceeding the group's capacity.
7226 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7227 * domain span are idle.
7229 static inline int nohz_kick_needed(struct rq
*rq
)
7231 unsigned long now
= jiffies
;
7232 struct sched_domain
*sd
;
7233 struct sched_group_capacity
*sgc
;
7234 int nr_busy
, cpu
= rq
->cpu
;
7236 if (unlikely(rq
->idle_balance
))
7240 * We may be recently in ticked or tickless idle mode. At the first
7241 * busy tick after returning from idle, we will update the busy stats.
7243 set_cpu_sd_state_busy();
7244 nohz_balance_exit_idle(cpu
);
7247 * None are in tickless mode and hence no need for NOHZ idle load
7250 if (likely(!atomic_read(&nohz
.nr_cpus
)))
7253 if (time_before(now
, nohz
.next_balance
))
7256 if (rq
->nr_running
>= 2)
7260 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7263 sgc
= sd
->groups
->sgc
;
7264 nr_busy
= atomic_read(&sgc
->nr_busy_cpus
);
7267 goto need_kick_unlock
;
7270 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
7272 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
7273 sched_domain_span(sd
)) < cpu
))
7274 goto need_kick_unlock
;
7285 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
7289 * run_rebalance_domains is triggered when needed from the scheduler tick.
7290 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7292 static void run_rebalance_domains(struct softirq_action
*h
)
7294 struct rq
*this_rq
= this_rq();
7295 enum cpu_idle_type idle
= this_rq
->idle_balance
?
7296 CPU_IDLE
: CPU_NOT_IDLE
;
7298 rebalance_domains(this_rq
, idle
);
7301 * If this cpu has a pending nohz_balance_kick, then do the
7302 * balancing on behalf of the other idle cpus whose ticks are
7305 nohz_idle_balance(this_rq
, idle
);
7309 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7311 void trigger_load_balance(struct rq
*rq
)
7313 /* Don't need to rebalance while attached to NULL domain */
7314 if (unlikely(on_null_domain(rq
)))
7317 if (time_after_eq(jiffies
, rq
->next_balance
))
7318 raise_softirq(SCHED_SOFTIRQ
);
7319 #ifdef CONFIG_NO_HZ_COMMON
7320 if (nohz_kick_needed(rq
))
7321 nohz_balancer_kick();
7325 static void rq_online_fair(struct rq
*rq
)
7330 static void rq_offline_fair(struct rq
*rq
)
7334 /* Ensure any throttled groups are reachable by pick_next_task */
7335 unthrottle_offline_cfs_rqs(rq
);
7338 #endif /* CONFIG_SMP */
7341 * scheduler tick hitting a task of our scheduling class:
7343 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
7345 struct cfs_rq
*cfs_rq
;
7346 struct sched_entity
*se
= &curr
->se
;
7348 for_each_sched_entity(se
) {
7349 cfs_rq
= cfs_rq_of(se
);
7350 entity_tick(cfs_rq
, se
, queued
);
7353 if (numabalancing_enabled
)
7354 task_tick_numa(rq
, curr
);
7356 update_rq_runnable_avg(rq
, 1);
7360 * called on fork with the child task as argument from the parent's context
7361 * - child not yet on the tasklist
7362 * - preemption disabled
7364 static void task_fork_fair(struct task_struct
*p
)
7366 struct cfs_rq
*cfs_rq
;
7367 struct sched_entity
*se
= &p
->se
, *curr
;
7368 int this_cpu
= smp_processor_id();
7369 struct rq
*rq
= this_rq();
7370 unsigned long flags
;
7372 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7374 update_rq_clock(rq
);
7376 cfs_rq
= task_cfs_rq(current
);
7377 curr
= cfs_rq
->curr
;
7380 * Not only the cpu but also the task_group of the parent might have
7381 * been changed after parent->se.parent,cfs_rq were copied to
7382 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7383 * of child point to valid ones.
7386 __set_task_cpu(p
, this_cpu
);
7389 update_curr(cfs_rq
);
7392 se
->vruntime
= curr
->vruntime
;
7393 place_entity(cfs_rq
, se
, 1);
7395 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
7397 * Upon rescheduling, sched_class::put_prev_task() will place
7398 * 'current' within the tree based on its new key value.
7400 swap(curr
->vruntime
, se
->vruntime
);
7401 resched_task(rq
->curr
);
7404 se
->vruntime
-= cfs_rq
->min_vruntime
;
7406 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7410 * Priority of the task has changed. Check to see if we preempt
7414 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
7420 * Reschedule if we are currently running on this runqueue and
7421 * our priority decreased, or if we are not currently running on
7422 * this runqueue and our priority is higher than the current's
7424 if (rq
->curr
== p
) {
7425 if (p
->prio
> oldprio
)
7426 resched_task(rq
->curr
);
7428 check_preempt_curr(rq
, p
, 0);
7431 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
7433 struct sched_entity
*se
= &p
->se
;
7434 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7437 * Ensure the task's vruntime is normalized, so that when it's
7438 * switched back to the fair class the enqueue_entity(.flags=0) will
7439 * do the right thing.
7441 * If it's on_rq, then the dequeue_entity(.flags=0) will already
7442 * have normalized the vruntime, if it's !on_rq, then only when
7443 * the task is sleeping will it still have non-normalized vruntime.
7445 if (!p
->on_rq
&& p
->state
!= TASK_RUNNING
) {
7447 * Fix up our vruntime so that the current sleep doesn't
7448 * cause 'unlimited' sleep bonus.
7450 place_entity(cfs_rq
, se
, 0);
7451 se
->vruntime
-= cfs_rq
->min_vruntime
;
7456 * Remove our load from contribution when we leave sched_fair
7457 * and ensure we don't carry in an old decay_count if we
7460 if (se
->avg
.decay_count
) {
7461 __synchronize_entity_decay(se
);
7462 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7468 * We switched to the sched_fair class.
7470 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7472 struct sched_entity
*se
= &p
->se
;
7473 #ifdef CONFIG_FAIR_GROUP_SCHED
7475 * Since the real-depth could have been changed (only FAIR
7476 * class maintain depth value), reset depth properly.
7478 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7484 * We were most likely switched from sched_rt, so
7485 * kick off the schedule if running, otherwise just see
7486 * if we can still preempt the current task.
7489 resched_task(rq
->curr
);
7491 check_preempt_curr(rq
, p
, 0);
7494 /* Account for a task changing its policy or group.
7496 * This routine is mostly called to set cfs_rq->curr field when a task
7497 * migrates between groups/classes.
7499 static void set_curr_task_fair(struct rq
*rq
)
7501 struct sched_entity
*se
= &rq
->curr
->se
;
7503 for_each_sched_entity(se
) {
7504 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7506 set_next_entity(cfs_rq
, se
);
7507 /* ensure bandwidth has been allocated on our new cfs_rq */
7508 account_cfs_rq_runtime(cfs_rq
, 0);
7512 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7514 cfs_rq
->tasks_timeline
= RB_ROOT
;
7515 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7516 #ifndef CONFIG_64BIT
7517 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7520 atomic64_set(&cfs_rq
->decay_counter
, 1);
7521 atomic_long_set(&cfs_rq
->removed_load
, 0);
7525 #ifdef CONFIG_FAIR_GROUP_SCHED
7526 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
7528 struct sched_entity
*se
= &p
->se
;
7529 struct cfs_rq
*cfs_rq
;
7532 * If the task was not on the rq at the time of this cgroup movement
7533 * it must have been asleep, sleeping tasks keep their ->vruntime
7534 * absolute on their old rq until wakeup (needed for the fair sleeper
7535 * bonus in place_entity()).
7537 * If it was on the rq, we've just 'preempted' it, which does convert
7538 * ->vruntime to a relative base.
7540 * Make sure both cases convert their relative position when migrating
7541 * to another cgroup's rq. This does somewhat interfere with the
7542 * fair sleeper stuff for the first placement, but who cares.
7545 * When !on_rq, vruntime of the task has usually NOT been normalized.
7546 * But there are some cases where it has already been normalized:
7548 * - Moving a forked child which is waiting for being woken up by
7549 * wake_up_new_task().
7550 * - Moving a task which has been woken up by try_to_wake_up() and
7551 * waiting for actually being woken up by sched_ttwu_pending().
7553 * To prevent boost or penalty in the new cfs_rq caused by delta
7554 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7556 if (!on_rq
&& (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7560 se
->vruntime
-= cfs_rq_of(se
)->min_vruntime
;
7561 set_task_rq(p
, task_cpu(p
));
7562 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7564 cfs_rq
= cfs_rq_of(se
);
7565 se
->vruntime
+= cfs_rq
->min_vruntime
;
7568 * migrate_task_rq_fair() will have removed our previous
7569 * contribution, but we must synchronize for ongoing future
7572 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7573 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
7578 void free_fair_sched_group(struct task_group
*tg
)
7582 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7584 for_each_possible_cpu(i
) {
7586 kfree(tg
->cfs_rq
[i
]);
7595 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7597 struct cfs_rq
*cfs_rq
;
7598 struct sched_entity
*se
;
7601 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7604 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7608 tg
->shares
= NICE_0_LOAD
;
7610 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7612 for_each_possible_cpu(i
) {
7613 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7614 GFP_KERNEL
, cpu_to_node(i
));
7618 se
= kzalloc_node(sizeof(struct sched_entity
),
7619 GFP_KERNEL
, cpu_to_node(i
));
7623 init_cfs_rq(cfs_rq
);
7624 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
7635 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7637 struct rq
*rq
= cpu_rq(cpu
);
7638 unsigned long flags
;
7641 * Only empty task groups can be destroyed; so we can speculatively
7642 * check on_list without danger of it being re-added.
7644 if (!tg
->cfs_rq
[cpu
]->on_list
)
7647 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7648 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
7649 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7652 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7653 struct sched_entity
*se
, int cpu
,
7654 struct sched_entity
*parent
)
7656 struct rq
*rq
= cpu_rq(cpu
);
7660 init_cfs_rq_runtime(cfs_rq
);
7662 tg
->cfs_rq
[cpu
] = cfs_rq
;
7665 /* se could be NULL for root_task_group */
7670 se
->cfs_rq
= &rq
->cfs
;
7673 se
->cfs_rq
= parent
->my_q
;
7674 se
->depth
= parent
->depth
+ 1;
7678 /* guarantee group entities always have weight */
7679 update_load_set(&se
->load
, NICE_0_LOAD
);
7680 se
->parent
= parent
;
7683 static DEFINE_MUTEX(shares_mutex
);
7685 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7688 unsigned long flags
;
7691 * We can't change the weight of the root cgroup.
7696 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
7698 mutex_lock(&shares_mutex
);
7699 if (tg
->shares
== shares
)
7702 tg
->shares
= shares
;
7703 for_each_possible_cpu(i
) {
7704 struct rq
*rq
= cpu_rq(i
);
7705 struct sched_entity
*se
;
7708 /* Propagate contribution to hierarchy */
7709 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7711 /* Possible calls to update_curr() need rq clock */
7712 update_rq_clock(rq
);
7713 for_each_sched_entity(se
)
7714 update_cfs_shares(group_cfs_rq(se
));
7715 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7719 mutex_unlock(&shares_mutex
);
7722 #else /* CONFIG_FAIR_GROUP_SCHED */
7724 void free_fair_sched_group(struct task_group
*tg
) { }
7726 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7731 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
7733 #endif /* CONFIG_FAIR_GROUP_SCHED */
7736 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
7738 struct sched_entity
*se
= &task
->se
;
7739 unsigned int rr_interval
= 0;
7742 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7745 if (rq
->cfs
.load
.weight
)
7746 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
7752 * All the scheduling class methods:
7754 const struct sched_class fair_sched_class
= {
7755 .next
= &idle_sched_class
,
7756 .enqueue_task
= enqueue_task_fair
,
7757 .dequeue_task
= dequeue_task_fair
,
7758 .yield_task
= yield_task_fair
,
7759 .yield_to_task
= yield_to_task_fair
,
7761 .check_preempt_curr
= check_preempt_wakeup
,
7763 .pick_next_task
= pick_next_task_fair
,
7764 .put_prev_task
= put_prev_task_fair
,
7767 .select_task_rq
= select_task_rq_fair
,
7768 .migrate_task_rq
= migrate_task_rq_fair
,
7770 .rq_online
= rq_online_fair
,
7771 .rq_offline
= rq_offline_fair
,
7773 .task_waking
= task_waking_fair
,
7776 .set_curr_task
= set_curr_task_fair
,
7777 .task_tick
= task_tick_fair
,
7778 .task_fork
= task_fork_fair
,
7780 .prio_changed
= prio_changed_fair
,
7781 .switched_from
= switched_from_fair
,
7782 .switched_to
= switched_to_fair
,
7784 .get_rr_interval
= get_rr_interval_fair
,
7786 #ifdef CONFIG_FAIR_GROUP_SCHED
7787 .task_move_group
= task_move_group_fair
,
7791 #ifdef CONFIG_SCHED_DEBUG
7792 void print_cfs_stats(struct seq_file
*m
, int cpu
)
7794 struct cfs_rq
*cfs_rq
;
7797 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
7798 print_cfs_rq(m
, cpu
, cfs_rq
);
7803 __init
void init_sched_fair_class(void)
7806 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7808 #ifdef CONFIG_NO_HZ_COMMON
7809 nohz
.next_balance
= jiffies
;
7810 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7811 cpu_notifier(sched_ilb_notifier
, 0);