2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency
= 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG
;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity
= 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency
= 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly
;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
93 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window
= 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
116 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
122 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
128 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling
) {
149 case SCHED_TUNABLESCALING_NONE
:
152 case SCHED_TUNABLESCALING_LINEAR
:
155 case SCHED_TUNABLESCALING_LOG
:
157 factor
= 1 + ilog2(cpus
);
164 static void update_sysctl(void)
166 unsigned int factor
= get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity
);
171 SET_SYSCTL(sched_latency
);
172 SET_SYSCTL(sched_wakeup_granularity
);
176 void sched_init_granularity(void)
181 #define WMULT_CONST (~0U)
182 #define WMULT_SHIFT 32
184 static void __update_inv_weight(struct load_weight
*lw
)
188 if (likely(lw
->inv_weight
))
191 w
= scale_load_down(lw
->weight
);
193 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
195 else if (unlikely(!w
))
196 lw
->inv_weight
= WMULT_CONST
;
198 lw
->inv_weight
= WMULT_CONST
/ w
;
202 * delta_exec * weight / lw.weight
204 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
206 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
207 * we're guaranteed shift stays positive because inv_weight is guaranteed to
208 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
210 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
211 * weight/lw.weight <= 1, and therefore our shift will also be positive.
213 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
215 u64 fact
= scale_load_down(weight
);
216 int shift
= WMULT_SHIFT
;
218 __update_inv_weight(lw
);
220 if (unlikely(fact
>> 32)) {
227 /* hint to use a 32x32->64 mul */
228 fact
= (u64
)(u32
)fact
* lw
->inv_weight
;
235 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
239 const struct sched_class fair_sched_class
;
241 /**************************************************************
242 * CFS operations on generic schedulable entities:
245 #ifdef CONFIG_FAIR_GROUP_SCHED
247 /* cpu runqueue to which this cfs_rq is attached */
248 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
253 /* An entity is a task if it doesn't "own" a runqueue */
254 #define entity_is_task(se) (!se->my_q)
256 static inline struct task_struct
*task_of(struct sched_entity
*se
)
258 #ifdef CONFIG_SCHED_DEBUG
259 WARN_ON_ONCE(!entity_is_task(se
));
261 return container_of(se
, struct task_struct
, se
);
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
268 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
285 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
288 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
290 if (!cfs_rq
->on_list
) {
292 * Ensure we either appear before our parent (if already
293 * enqueued) or force our parent to appear after us when it is
294 * enqueued. The fact that we always enqueue bottom-up
295 * reduces this to two cases.
297 if (cfs_rq
->tg
->parent
&&
298 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
299 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
300 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
302 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
303 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
307 /* We should have no load, but we need to update last_decay. */
308 update_cfs_rq_blocked_load(cfs_rq
, 0);
312 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
314 if (cfs_rq
->on_list
) {
315 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
320 /* Iterate thr' all leaf cfs_rq's on a runqueue */
321 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
322 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
324 /* Do the two (enqueued) entities belong to the same group ? */
325 static inline struct cfs_rq
*
326 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
328 if (se
->cfs_rq
== pse
->cfs_rq
)
334 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
340 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
342 int se_depth
, pse_depth
;
345 * preemption test can be made between sibling entities who are in the
346 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
347 * both tasks until we find their ancestors who are siblings of common
351 /* First walk up until both entities are at same depth */
352 se_depth
= (*se
)->depth
;
353 pse_depth
= (*pse
)->depth
;
355 while (se_depth
> pse_depth
) {
357 *se
= parent_entity(*se
);
360 while (pse_depth
> se_depth
) {
362 *pse
= parent_entity(*pse
);
365 while (!is_same_group(*se
, *pse
)) {
366 *se
= parent_entity(*se
);
367 *pse
= parent_entity(*pse
);
371 #else /* !CONFIG_FAIR_GROUP_SCHED */
373 static inline struct task_struct
*task_of(struct sched_entity
*se
)
375 return container_of(se
, struct task_struct
, se
);
378 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
380 return container_of(cfs_rq
, struct rq
, cfs
);
383 #define entity_is_task(se) 1
385 #define for_each_sched_entity(se) \
386 for (; se; se = NULL)
388 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
390 return &task_rq(p
)->cfs
;
393 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
395 struct task_struct
*p
= task_of(se
);
396 struct rq
*rq
= task_rq(p
);
401 /* runqueue "owned" by this group */
402 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
407 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
411 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
415 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
416 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
418 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
424 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
428 #endif /* CONFIG_FAIR_GROUP_SCHED */
430 static __always_inline
431 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
433 /**************************************************************
434 * Scheduling class tree data structure manipulation methods:
437 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
439 s64 delta
= (s64
)(vruntime
- max_vruntime
);
441 max_vruntime
= vruntime
;
446 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
448 s64 delta
= (s64
)(vruntime
- min_vruntime
);
450 min_vruntime
= vruntime
;
455 static inline int entity_before(struct sched_entity
*a
,
456 struct sched_entity
*b
)
458 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
461 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
463 u64 vruntime
= cfs_rq
->min_vruntime
;
466 vruntime
= cfs_rq
->curr
->vruntime
;
468 if (cfs_rq
->rb_leftmost
) {
469 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
474 vruntime
= se
->vruntime
;
476 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
479 /* ensure we never gain time by being placed backwards. */
480 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
483 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
488 * Enqueue an entity into the rb-tree:
490 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
492 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
493 struct rb_node
*parent
= NULL
;
494 struct sched_entity
*entry
;
498 * Find the right place in the rbtree:
502 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
504 * We dont care about collisions. Nodes with
505 * the same key stay together.
507 if (entity_before(se
, entry
)) {
508 link
= &parent
->rb_left
;
510 link
= &parent
->rb_right
;
516 * Maintain a cache of leftmost tree entries (it is frequently
520 cfs_rq
->rb_leftmost
= &se
->run_node
;
522 rb_link_node(&se
->run_node
, parent
, link
);
523 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
526 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
528 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
529 struct rb_node
*next_node
;
531 next_node
= rb_next(&se
->run_node
);
532 cfs_rq
->rb_leftmost
= next_node
;
535 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
538 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
540 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
545 return rb_entry(left
, struct sched_entity
, run_node
);
548 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
550 struct rb_node
*next
= rb_next(&se
->run_node
);
555 return rb_entry(next
, struct sched_entity
, run_node
);
558 #ifdef CONFIG_SCHED_DEBUG
559 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
561 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
566 return rb_entry(last
, struct sched_entity
, run_node
);
569 /**************************************************************
570 * Scheduling class statistics methods:
573 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
574 void __user
*buffer
, size_t *lenp
,
577 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
578 int factor
= get_update_sysctl_factor();
583 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
584 sysctl_sched_min_granularity
);
586 #define WRT_SYSCTL(name) \
587 (normalized_sysctl_##name = sysctl_##name / (factor))
588 WRT_SYSCTL(sched_min_granularity
);
589 WRT_SYSCTL(sched_latency
);
590 WRT_SYSCTL(sched_wakeup_granularity
);
600 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
602 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
603 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
609 * The idea is to set a period in which each task runs once.
611 * When there are too many tasks (sched_nr_latency) we have to stretch
612 * this period because otherwise the slices get too small.
614 * p = (nr <= nl) ? l : l*nr/nl
616 static u64
__sched_period(unsigned long nr_running
)
618 u64 period
= sysctl_sched_latency
;
619 unsigned long nr_latency
= sched_nr_latency
;
621 if (unlikely(nr_running
> nr_latency
)) {
622 period
= sysctl_sched_min_granularity
;
623 period
*= nr_running
;
630 * We calculate the wall-time slice from the period by taking a part
631 * proportional to the weight.
635 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
637 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
639 for_each_sched_entity(se
) {
640 struct load_weight
*load
;
641 struct load_weight lw
;
643 cfs_rq
= cfs_rq_of(se
);
644 load
= &cfs_rq
->load
;
646 if (unlikely(!se
->on_rq
)) {
649 update_load_add(&lw
, se
->load
.weight
);
652 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
658 * We calculate the vruntime slice of a to-be-inserted task.
662 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
664 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
668 static unsigned long task_h_load(struct task_struct
*p
);
670 static inline void __update_task_entity_contrib(struct sched_entity
*se
);
672 /* Give new task start runnable values to heavy its load in infant time */
673 void init_task_runnable_average(struct task_struct
*p
)
677 p
->se
.avg
.decay_count
= 0;
678 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
679 p
->se
.avg
.runnable_avg_sum
= slice
;
680 p
->se
.avg
.runnable_avg_period
= slice
;
681 __update_task_entity_contrib(&p
->se
);
684 void init_task_runnable_average(struct task_struct
*p
)
690 * Update the current task's runtime statistics.
692 static void update_curr(struct cfs_rq
*cfs_rq
)
694 struct sched_entity
*curr
= cfs_rq
->curr
;
695 u64 now
= rq_clock_task(rq_of(cfs_rq
));
701 delta_exec
= now
- curr
->exec_start
;
702 if (unlikely((s64
)delta_exec
<= 0))
705 curr
->exec_start
= now
;
707 schedstat_set(curr
->statistics
.exec_max
,
708 max(delta_exec
, curr
->statistics
.exec_max
));
710 curr
->sum_exec_runtime
+= delta_exec
;
711 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
713 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
714 update_min_vruntime(cfs_rq
);
716 if (entity_is_task(curr
)) {
717 struct task_struct
*curtask
= task_of(curr
);
719 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
720 cpuacct_charge(curtask
, delta_exec
);
721 account_group_exec_runtime(curtask
, delta_exec
);
724 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
728 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
730 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
734 * Task is being enqueued - update stats:
736 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
739 * Are we enqueueing a waiting task? (for current tasks
740 * a dequeue/enqueue event is a NOP)
742 if (se
!= cfs_rq
->curr
)
743 update_stats_wait_start(cfs_rq
, se
);
747 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
749 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
750 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
751 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
752 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
753 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
754 #ifdef CONFIG_SCHEDSTATS
755 if (entity_is_task(se
)) {
756 trace_sched_stat_wait(task_of(se
),
757 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
760 schedstat_set(se
->statistics
.wait_start
, 0);
764 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
767 * Mark the end of the wait period if dequeueing a
770 if (se
!= cfs_rq
->curr
)
771 update_stats_wait_end(cfs_rq
, se
);
775 * We are picking a new current task - update its stats:
778 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
781 * We are starting a new run period:
783 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
786 /**************************************************
787 * Scheduling class queueing methods:
790 #ifdef CONFIG_NUMA_BALANCING
792 * Approximate time to scan a full NUMA task in ms. The task scan period is
793 * calculated based on the tasks virtual memory size and
794 * numa_balancing_scan_size.
796 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
797 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
799 /* Portion of address space to scan in MB */
800 unsigned int sysctl_numa_balancing_scan_size
= 256;
802 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
803 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
805 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
807 unsigned long rss
= 0;
808 unsigned long nr_scan_pages
;
811 * Calculations based on RSS as non-present and empty pages are skipped
812 * by the PTE scanner and NUMA hinting faults should be trapped based
815 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
816 rss
= get_mm_rss(p
->mm
);
820 rss
= round_up(rss
, nr_scan_pages
);
821 return rss
/ nr_scan_pages
;
824 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
825 #define MAX_SCAN_WINDOW 2560
827 static unsigned int task_scan_min(struct task_struct
*p
)
829 unsigned int scan
, floor
;
830 unsigned int windows
= 1;
832 if (sysctl_numa_balancing_scan_size
< MAX_SCAN_WINDOW
)
833 windows
= MAX_SCAN_WINDOW
/ sysctl_numa_balancing_scan_size
;
834 floor
= 1000 / windows
;
836 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
837 return max_t(unsigned int, floor
, scan
);
840 static unsigned int task_scan_max(struct task_struct
*p
)
842 unsigned int smin
= task_scan_min(p
);
845 /* Watch for min being lower than max due to floor calculations */
846 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
847 return max(smin
, smax
);
850 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
852 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
853 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
856 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
858 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
859 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
865 spinlock_t lock
; /* nr_tasks, tasks */
868 struct list_head task_list
;
871 nodemask_t active_nodes
;
872 unsigned long total_faults
;
874 * Faults_cpu is used to decide whether memory should move
875 * towards the CPU. As a consequence, these stats are weighted
876 * more by CPU use than by memory faults.
878 unsigned long *faults_cpu
;
879 unsigned long faults
[0];
882 /* Shared or private faults. */
883 #define NR_NUMA_HINT_FAULT_TYPES 2
885 /* Memory and CPU locality */
886 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
888 /* Averaged statistics, and temporary buffers. */
889 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
891 pid_t
task_numa_group_id(struct task_struct
*p
)
893 return p
->numa_group
? p
->numa_group
->gid
: 0;
896 static inline int task_faults_idx(int nid
, int priv
)
898 return NR_NUMA_HINT_FAULT_TYPES
* nid
+ priv
;
901 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
903 if (!p
->numa_faults_memory
)
906 return p
->numa_faults_memory
[task_faults_idx(nid
, 0)] +
907 p
->numa_faults_memory
[task_faults_idx(nid
, 1)];
910 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
915 return p
->numa_group
->faults
[task_faults_idx(nid
, 0)] +
916 p
->numa_group
->faults
[task_faults_idx(nid
, 1)];
919 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
921 return group
->faults_cpu
[task_faults_idx(nid
, 0)] +
922 group
->faults_cpu
[task_faults_idx(nid
, 1)];
926 * These return the fraction of accesses done by a particular task, or
927 * task group, on a particular numa node. The group weight is given a
928 * larger multiplier, in order to group tasks together that are almost
929 * evenly spread out between numa nodes.
931 static inline unsigned long task_weight(struct task_struct
*p
, int nid
)
933 unsigned long total_faults
;
935 if (!p
->numa_faults_memory
)
938 total_faults
= p
->total_numa_faults
;
943 return 1000 * task_faults(p
, nid
) / total_faults
;
946 static inline unsigned long group_weight(struct task_struct
*p
, int nid
)
948 if (!p
->numa_group
|| !p
->numa_group
->total_faults
)
951 return 1000 * group_faults(p
, nid
) / p
->numa_group
->total_faults
;
954 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
955 int src_nid
, int dst_cpu
)
957 struct numa_group
*ng
= p
->numa_group
;
958 int dst_nid
= cpu_to_node(dst_cpu
);
959 int last_cpupid
, this_cpupid
;
961 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
964 * Multi-stage node selection is used in conjunction with a periodic
965 * migration fault to build a temporal task<->page relation. By using
966 * a two-stage filter we remove short/unlikely relations.
968 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
969 * a task's usage of a particular page (n_p) per total usage of this
970 * page (n_t) (in a given time-span) to a probability.
972 * Our periodic faults will sample this probability and getting the
973 * same result twice in a row, given these samples are fully
974 * independent, is then given by P(n)^2, provided our sample period
975 * is sufficiently short compared to the usage pattern.
977 * This quadric squishes small probabilities, making it less likely we
978 * act on an unlikely task<->page relation.
980 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
981 if (!cpupid_pid_unset(last_cpupid
) &&
982 cpupid_to_nid(last_cpupid
) != dst_nid
)
985 /* Always allow migrate on private faults */
986 if (cpupid_match_pid(p
, last_cpupid
))
989 /* A shared fault, but p->numa_group has not been set up yet. */
994 * Do not migrate if the destination is not a node that
995 * is actively used by this numa group.
997 if (!node_isset(dst_nid
, ng
->active_nodes
))
1001 * Source is a node that is not actively used by this
1002 * numa group, while the destination is. Migrate.
1004 if (!node_isset(src_nid
, ng
->active_nodes
))
1008 * Both source and destination are nodes in active
1009 * use by this numa group. Maximize memory bandwidth
1010 * by migrating from more heavily used groups, to less
1011 * heavily used ones, spreading the load around.
1012 * Use a 1/4 hysteresis to avoid spurious page movement.
1014 return group_faults(p
, dst_nid
) < (group_faults(p
, src_nid
) * 3 / 4);
1017 static unsigned long weighted_cpuload(const int cpu
);
1018 static unsigned long source_load(int cpu
, int type
);
1019 static unsigned long target_load(int cpu
, int type
);
1020 static unsigned long capacity_of(int cpu
);
1021 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
1023 /* Cached statistics for all CPUs within a node */
1025 unsigned long nr_running
;
1028 /* Total compute capacity of CPUs on a node */
1029 unsigned long compute_capacity
;
1031 /* Approximate capacity in terms of runnable tasks on a node */
1032 unsigned long task_capacity
;
1033 int has_free_capacity
;
1037 * XXX borrowed from update_sg_lb_stats
1039 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1043 memset(ns
, 0, sizeof(*ns
));
1044 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1045 struct rq
*rq
= cpu_rq(cpu
);
1047 ns
->nr_running
+= rq
->nr_running
;
1048 ns
->load
+= weighted_cpuload(cpu
);
1049 ns
->compute_capacity
+= capacity_of(cpu
);
1055 * If we raced with hotplug and there are no CPUs left in our mask
1056 * the @ns structure is NULL'ed and task_numa_compare() will
1057 * not find this node attractive.
1059 * We'll either bail at !has_free_capacity, or we'll detect a huge
1060 * imbalance and bail there.
1066 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
);
1067 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1070 struct task_numa_env
{
1071 struct task_struct
*p
;
1073 int src_cpu
, src_nid
;
1074 int dst_cpu
, dst_nid
;
1076 struct numa_stats src_stats
, dst_stats
;
1080 struct task_struct
*best_task
;
1085 static void task_numa_assign(struct task_numa_env
*env
,
1086 struct task_struct
*p
, long imp
)
1089 put_task_struct(env
->best_task
);
1094 env
->best_imp
= imp
;
1095 env
->best_cpu
= env
->dst_cpu
;
1098 static bool load_too_imbalanced(long src_load
, long dst_load
,
1099 struct task_numa_env
*env
)
1102 long orig_src_load
, orig_dst_load
;
1103 long src_capacity
, dst_capacity
;
1106 * The load is corrected for the CPU capacity available on each node.
1109 * ------------ vs ---------
1110 * src_capacity dst_capacity
1112 src_capacity
= env
->src_stats
.compute_capacity
;
1113 dst_capacity
= env
->dst_stats
.compute_capacity
;
1115 /* We care about the slope of the imbalance, not the direction. */
1116 if (dst_load
< src_load
)
1117 swap(dst_load
, src_load
);
1119 /* Is the difference below the threshold? */
1120 imb
= dst_load
* src_capacity
* 100 -
1121 src_load
* dst_capacity
* env
->imbalance_pct
;
1126 * The imbalance is above the allowed threshold.
1127 * Compare it with the old imbalance.
1129 orig_src_load
= env
->src_stats
.load
;
1130 orig_dst_load
= env
->dst_stats
.load
;
1132 if (orig_dst_load
< orig_src_load
)
1133 swap(orig_dst_load
, orig_src_load
);
1135 old_imb
= orig_dst_load
* src_capacity
* 100 -
1136 orig_src_load
* dst_capacity
* env
->imbalance_pct
;
1138 /* Would this change make things worse? */
1139 return (imb
> old_imb
);
1143 * This checks if the overall compute and NUMA accesses of the system would
1144 * be improved if the source tasks was migrated to the target dst_cpu taking
1145 * into account that it might be best if task running on the dst_cpu should
1146 * be exchanged with the source task
1148 static void task_numa_compare(struct task_numa_env
*env
,
1149 long taskimp
, long groupimp
)
1151 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1152 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1153 struct task_struct
*cur
;
1154 long src_load
, dst_load
;
1156 long imp
= (groupimp
> 0) ? groupimp
: taskimp
;
1159 cur
= ACCESS_ONCE(dst_rq
->curr
);
1160 if (cur
->pid
== 0) /* idle */
1164 * "imp" is the fault differential for the source task between the
1165 * source and destination node. Calculate the total differential for
1166 * the source task and potential destination task. The more negative
1167 * the value is, the more rmeote accesses that would be expected to
1168 * be incurred if the tasks were swapped.
1171 /* Skip this swap candidate if cannot move to the source cpu */
1172 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1176 * If dst and source tasks are in the same NUMA group, or not
1177 * in any group then look only at task weights.
1179 if (cur
->numa_group
== env
->p
->numa_group
) {
1180 imp
= taskimp
+ task_weight(cur
, env
->src_nid
) -
1181 task_weight(cur
, env
->dst_nid
);
1183 * Add some hysteresis to prevent swapping the
1184 * tasks within a group over tiny differences.
1186 if (cur
->numa_group
)
1190 * Compare the group weights. If a task is all by
1191 * itself (not part of a group), use the task weight
1194 if (env
->p
->numa_group
)
1199 if (cur
->numa_group
)
1200 imp
+= group_weight(cur
, env
->src_nid
) -
1201 group_weight(cur
, env
->dst_nid
);
1203 imp
+= task_weight(cur
, env
->src_nid
) -
1204 task_weight(cur
, env
->dst_nid
);
1208 if (imp
< env
->best_imp
)
1212 /* Is there capacity at our destination? */
1213 if (env
->src_stats
.has_free_capacity
&&
1214 !env
->dst_stats
.has_free_capacity
)
1220 /* Balance doesn't matter much if we're running a task per cpu */
1221 if (src_rq
->nr_running
== 1 && dst_rq
->nr_running
== 1)
1225 * In the overloaded case, try and keep the load balanced.
1228 load
= task_h_load(env
->p
);
1229 dst_load
= env
->dst_stats
.load
+ load
;
1230 src_load
= env
->src_stats
.load
- load
;
1233 load
= task_h_load(cur
);
1238 if (load_too_imbalanced(src_load
, dst_load
, env
))
1242 task_numa_assign(env
, cur
, imp
);
1247 static void task_numa_find_cpu(struct task_numa_env
*env
,
1248 long taskimp
, long groupimp
)
1252 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1253 /* Skip this CPU if the source task cannot migrate */
1254 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1258 task_numa_compare(env
, taskimp
, groupimp
);
1262 static int task_numa_migrate(struct task_struct
*p
)
1264 struct task_numa_env env
= {
1267 .src_cpu
= task_cpu(p
),
1268 .src_nid
= task_node(p
),
1270 .imbalance_pct
= 112,
1276 struct sched_domain
*sd
;
1277 unsigned long taskweight
, groupweight
;
1279 long taskimp
, groupimp
;
1282 * Pick the lowest SD_NUMA domain, as that would have the smallest
1283 * imbalance and would be the first to start moving tasks about.
1285 * And we want to avoid any moving of tasks about, as that would create
1286 * random movement of tasks -- counter the numa conditions we're trying
1290 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1292 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1296 * Cpusets can break the scheduler domain tree into smaller
1297 * balance domains, some of which do not cross NUMA boundaries.
1298 * Tasks that are "trapped" in such domains cannot be migrated
1299 * elsewhere, so there is no point in (re)trying.
1301 if (unlikely(!sd
)) {
1302 p
->numa_preferred_nid
= task_node(p
);
1306 taskweight
= task_weight(p
, env
.src_nid
);
1307 groupweight
= group_weight(p
, env
.src_nid
);
1308 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1309 env
.dst_nid
= p
->numa_preferred_nid
;
1310 taskimp
= task_weight(p
, env
.dst_nid
) - taskweight
;
1311 groupimp
= group_weight(p
, env
.dst_nid
) - groupweight
;
1312 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1314 /* Try to find a spot on the preferred nid. */
1315 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1317 /* No space available on the preferred nid. Look elsewhere. */
1318 if (env
.best_cpu
== -1) {
1319 for_each_online_node(nid
) {
1320 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1323 /* Only consider nodes where both task and groups benefit */
1324 taskimp
= task_weight(p
, nid
) - taskweight
;
1325 groupimp
= group_weight(p
, nid
) - groupweight
;
1326 if (taskimp
< 0 && groupimp
< 0)
1330 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1331 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1335 /* No better CPU than the current one was found. */
1336 if (env
.best_cpu
== -1)
1340 * If the task is part of a workload that spans multiple NUMA nodes,
1341 * and is migrating into one of the workload's active nodes, remember
1342 * this node as the task's preferred numa node, so the workload can
1344 * A task that migrated to a second choice node will be better off
1345 * trying for a better one later. Do not set the preferred node here.
1347 if (p
->numa_group
&& node_isset(env
.dst_nid
, p
->numa_group
->active_nodes
))
1348 sched_setnuma(p
, env
.dst_nid
);
1351 * Reset the scan period if the task is being rescheduled on an
1352 * alternative node to recheck if the tasks is now properly placed.
1354 p
->numa_scan_period
= task_scan_min(p
);
1356 if (env
.best_task
== NULL
) {
1357 ret
= migrate_task_to(p
, env
.best_cpu
);
1359 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1363 ret
= migrate_swap(p
, env
.best_task
);
1365 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1366 put_task_struct(env
.best_task
);
1370 /* Attempt to migrate a task to a CPU on the preferred node. */
1371 static void numa_migrate_preferred(struct task_struct
*p
)
1373 unsigned long interval
= HZ
;
1375 /* This task has no NUMA fault statistics yet */
1376 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults_memory
))
1379 /* Periodically retry migrating the task to the preferred node */
1380 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1381 p
->numa_migrate_retry
= jiffies
+ interval
;
1383 /* Success if task is already running on preferred CPU */
1384 if (task_node(p
) == p
->numa_preferred_nid
)
1387 /* Otherwise, try migrate to a CPU on the preferred node */
1388 task_numa_migrate(p
);
1392 * Find the nodes on which the workload is actively running. We do this by
1393 * tracking the nodes from which NUMA hinting faults are triggered. This can
1394 * be different from the set of nodes where the workload's memory is currently
1397 * The bitmask is used to make smarter decisions on when to do NUMA page
1398 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1399 * are added when they cause over 6/16 of the maximum number of faults, but
1400 * only removed when they drop below 3/16.
1402 static void update_numa_active_node_mask(struct numa_group
*numa_group
)
1404 unsigned long faults
, max_faults
= 0;
1407 for_each_online_node(nid
) {
1408 faults
= group_faults_cpu(numa_group
, nid
);
1409 if (faults
> max_faults
)
1410 max_faults
= faults
;
1413 for_each_online_node(nid
) {
1414 faults
= group_faults_cpu(numa_group
, nid
);
1415 if (!node_isset(nid
, numa_group
->active_nodes
)) {
1416 if (faults
> max_faults
* 6 / 16)
1417 node_set(nid
, numa_group
->active_nodes
);
1418 } else if (faults
< max_faults
* 3 / 16)
1419 node_clear(nid
, numa_group
->active_nodes
);
1424 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1425 * increments. The more local the fault statistics are, the higher the scan
1426 * period will be for the next scan window. If local/remote ratio is below
1427 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1428 * scan period will decrease
1430 #define NUMA_PERIOD_SLOTS 10
1431 #define NUMA_PERIOD_THRESHOLD 3
1434 * Increase the scan period (slow down scanning) if the majority of
1435 * our memory is already on our local node, or if the majority of
1436 * the page accesses are shared with other processes.
1437 * Otherwise, decrease the scan period.
1439 static void update_task_scan_period(struct task_struct
*p
,
1440 unsigned long shared
, unsigned long private)
1442 unsigned int period_slot
;
1446 unsigned long remote
= p
->numa_faults_locality
[0];
1447 unsigned long local
= p
->numa_faults_locality
[1];
1450 * If there were no record hinting faults then either the task is
1451 * completely idle or all activity is areas that are not of interest
1452 * to automatic numa balancing. Scan slower
1454 if (local
+ shared
== 0) {
1455 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1456 p
->numa_scan_period
<< 1);
1458 p
->mm
->numa_next_scan
= jiffies
+
1459 msecs_to_jiffies(p
->numa_scan_period
);
1465 * Prepare to scale scan period relative to the current period.
1466 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1467 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1468 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1470 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1471 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1472 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1473 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1476 diff
= slot
* period_slot
;
1478 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1481 * Scale scan rate increases based on sharing. There is an
1482 * inverse relationship between the degree of sharing and
1483 * the adjustment made to the scanning period. Broadly
1484 * speaking the intent is that there is little point
1485 * scanning faster if shared accesses dominate as it may
1486 * simply bounce migrations uselessly
1488 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
));
1489 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1492 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1493 task_scan_min(p
), task_scan_max(p
));
1494 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1498 * Get the fraction of time the task has been running since the last
1499 * NUMA placement cycle. The scheduler keeps similar statistics, but
1500 * decays those on a 32ms period, which is orders of magnitude off
1501 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1502 * stats only if the task is so new there are no NUMA statistics yet.
1504 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1506 u64 runtime
, delta
, now
;
1507 /* Use the start of this time slice to avoid calculations. */
1508 now
= p
->se
.exec_start
;
1509 runtime
= p
->se
.sum_exec_runtime
;
1511 if (p
->last_task_numa_placement
) {
1512 delta
= runtime
- p
->last_sum_exec_runtime
;
1513 *period
= now
- p
->last_task_numa_placement
;
1515 delta
= p
->se
.avg
.runnable_avg_sum
;
1516 *period
= p
->se
.avg
.runnable_avg_period
;
1519 p
->last_sum_exec_runtime
= runtime
;
1520 p
->last_task_numa_placement
= now
;
1525 static void task_numa_placement(struct task_struct
*p
)
1527 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1528 unsigned long max_faults
= 0, max_group_faults
= 0;
1529 unsigned long fault_types
[2] = { 0, 0 };
1530 unsigned long total_faults
;
1531 u64 runtime
, period
;
1532 spinlock_t
*group_lock
= NULL
;
1534 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1535 if (p
->numa_scan_seq
== seq
)
1537 p
->numa_scan_seq
= seq
;
1538 p
->numa_scan_period_max
= task_scan_max(p
);
1540 total_faults
= p
->numa_faults_locality
[0] +
1541 p
->numa_faults_locality
[1];
1542 runtime
= numa_get_avg_runtime(p
, &period
);
1544 /* If the task is part of a group prevent parallel updates to group stats */
1545 if (p
->numa_group
) {
1546 group_lock
= &p
->numa_group
->lock
;
1547 spin_lock_irq(group_lock
);
1550 /* Find the node with the highest number of faults */
1551 for_each_online_node(nid
) {
1552 unsigned long faults
= 0, group_faults
= 0;
1555 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
1556 long diff
, f_diff
, f_weight
;
1558 i
= task_faults_idx(nid
, priv
);
1560 /* Decay existing window, copy faults since last scan */
1561 diff
= p
->numa_faults_buffer_memory
[i
] - p
->numa_faults_memory
[i
] / 2;
1562 fault_types
[priv
] += p
->numa_faults_buffer_memory
[i
];
1563 p
->numa_faults_buffer_memory
[i
] = 0;
1566 * Normalize the faults_from, so all tasks in a group
1567 * count according to CPU use, instead of by the raw
1568 * number of faults. Tasks with little runtime have
1569 * little over-all impact on throughput, and thus their
1570 * faults are less important.
1572 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
1573 f_weight
= (f_weight
* p
->numa_faults_buffer_cpu
[i
]) /
1575 f_diff
= f_weight
- p
->numa_faults_cpu
[i
] / 2;
1576 p
->numa_faults_buffer_cpu
[i
] = 0;
1578 p
->numa_faults_memory
[i
] += diff
;
1579 p
->numa_faults_cpu
[i
] += f_diff
;
1580 faults
+= p
->numa_faults_memory
[i
];
1581 p
->total_numa_faults
+= diff
;
1582 if (p
->numa_group
) {
1583 /* safe because we can only change our own group */
1584 p
->numa_group
->faults
[i
] += diff
;
1585 p
->numa_group
->faults_cpu
[i
] += f_diff
;
1586 p
->numa_group
->total_faults
+= diff
;
1587 group_faults
+= p
->numa_group
->faults
[i
];
1591 if (faults
> max_faults
) {
1592 max_faults
= faults
;
1596 if (group_faults
> max_group_faults
) {
1597 max_group_faults
= group_faults
;
1598 max_group_nid
= nid
;
1602 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1604 if (p
->numa_group
) {
1605 update_numa_active_node_mask(p
->numa_group
);
1606 spin_unlock_irq(group_lock
);
1607 max_nid
= max_group_nid
;
1611 /* Set the new preferred node */
1612 if (max_nid
!= p
->numa_preferred_nid
)
1613 sched_setnuma(p
, max_nid
);
1615 if (task_node(p
) != p
->numa_preferred_nid
)
1616 numa_migrate_preferred(p
);
1620 static inline int get_numa_group(struct numa_group
*grp
)
1622 return atomic_inc_not_zero(&grp
->refcount
);
1625 static inline void put_numa_group(struct numa_group
*grp
)
1627 if (atomic_dec_and_test(&grp
->refcount
))
1628 kfree_rcu(grp
, rcu
);
1631 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1634 struct numa_group
*grp
, *my_grp
;
1635 struct task_struct
*tsk
;
1637 int cpu
= cpupid_to_cpu(cpupid
);
1640 if (unlikely(!p
->numa_group
)) {
1641 unsigned int size
= sizeof(struct numa_group
) +
1642 4*nr_node_ids
*sizeof(unsigned long);
1644 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1648 atomic_set(&grp
->refcount
, 1);
1649 spin_lock_init(&grp
->lock
);
1650 INIT_LIST_HEAD(&grp
->task_list
);
1652 /* Second half of the array tracks nids where faults happen */
1653 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
1656 node_set(task_node(current
), grp
->active_nodes
);
1658 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1659 grp
->faults
[i
] = p
->numa_faults_memory
[i
];
1661 grp
->total_faults
= p
->total_numa_faults
;
1663 list_add(&p
->numa_entry
, &grp
->task_list
);
1665 rcu_assign_pointer(p
->numa_group
, grp
);
1669 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1671 if (!cpupid_match_pid(tsk
, cpupid
))
1674 grp
= rcu_dereference(tsk
->numa_group
);
1678 my_grp
= p
->numa_group
;
1683 * Only join the other group if its bigger; if we're the bigger group,
1684 * the other task will join us.
1686 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1690 * Tie-break on the grp address.
1692 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1695 /* Always join threads in the same process. */
1696 if (tsk
->mm
== current
->mm
)
1699 /* Simple filter to avoid false positives due to PID collisions */
1700 if (flags
& TNF_SHARED
)
1703 /* Update priv based on whether false sharing was detected */
1706 if (join
&& !get_numa_group(grp
))
1714 BUG_ON(irqs_disabled());
1715 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
1717 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
1718 my_grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1719 grp
->faults
[i
] += p
->numa_faults_memory
[i
];
1721 my_grp
->total_faults
-= p
->total_numa_faults
;
1722 grp
->total_faults
+= p
->total_numa_faults
;
1724 list_move(&p
->numa_entry
, &grp
->task_list
);
1728 spin_unlock(&my_grp
->lock
);
1729 spin_unlock_irq(&grp
->lock
);
1731 rcu_assign_pointer(p
->numa_group
, grp
);
1733 put_numa_group(my_grp
);
1741 void task_numa_free(struct task_struct
*p
)
1743 struct numa_group
*grp
= p
->numa_group
;
1744 void *numa_faults
= p
->numa_faults_memory
;
1745 unsigned long flags
;
1749 spin_lock_irqsave(&grp
->lock
, flags
);
1750 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1751 grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1752 grp
->total_faults
-= p
->total_numa_faults
;
1754 list_del(&p
->numa_entry
);
1756 spin_unlock_irqrestore(&grp
->lock
, flags
);
1757 rcu_assign_pointer(p
->numa_group
, NULL
);
1758 put_numa_group(grp
);
1761 p
->numa_faults_memory
= NULL
;
1762 p
->numa_faults_buffer_memory
= NULL
;
1763 p
->numa_faults_cpu
= NULL
;
1764 p
->numa_faults_buffer_cpu
= NULL
;
1769 * Got a PROT_NONE fault for a page on @node.
1771 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
1773 struct task_struct
*p
= current
;
1774 bool migrated
= flags
& TNF_MIGRATED
;
1775 int cpu_node
= task_node(current
);
1776 int local
= !!(flags
& TNF_FAULT_LOCAL
);
1779 if (!numabalancing_enabled
)
1782 /* for example, ksmd faulting in a user's mm */
1786 /* Do not worry about placement if exiting */
1787 if (p
->state
== TASK_DEAD
)
1790 /* Allocate buffer to track faults on a per-node basis */
1791 if (unlikely(!p
->numa_faults_memory
)) {
1792 int size
= sizeof(*p
->numa_faults_memory
) *
1793 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
1795 p
->numa_faults_memory
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
1796 if (!p
->numa_faults_memory
)
1799 BUG_ON(p
->numa_faults_buffer_memory
);
1801 * The averaged statistics, shared & private, memory & cpu,
1802 * occupy the first half of the array. The second half of the
1803 * array is for current counters, which are averaged into the
1804 * first set by task_numa_placement.
1806 p
->numa_faults_cpu
= p
->numa_faults_memory
+ (2 * nr_node_ids
);
1807 p
->numa_faults_buffer_memory
= p
->numa_faults_memory
+ (4 * nr_node_ids
);
1808 p
->numa_faults_buffer_cpu
= p
->numa_faults_memory
+ (6 * nr_node_ids
);
1809 p
->total_numa_faults
= 0;
1810 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1814 * First accesses are treated as private, otherwise consider accesses
1815 * to be private if the accessing pid has not changed
1817 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
1820 priv
= cpupid_match_pid(p
, last_cpupid
);
1821 if (!priv
&& !(flags
& TNF_NO_GROUP
))
1822 task_numa_group(p
, last_cpupid
, flags
, &priv
);
1826 * If a workload spans multiple NUMA nodes, a shared fault that
1827 * occurs wholly within the set of nodes that the workload is
1828 * actively using should be counted as local. This allows the
1829 * scan rate to slow down when a workload has settled down.
1831 if (!priv
&& !local
&& p
->numa_group
&&
1832 node_isset(cpu_node
, p
->numa_group
->active_nodes
) &&
1833 node_isset(mem_node
, p
->numa_group
->active_nodes
))
1836 task_numa_placement(p
);
1839 * Retry task to preferred node migration periodically, in case it
1840 * case it previously failed, or the scheduler moved us.
1842 if (time_after(jiffies
, p
->numa_migrate_retry
))
1843 numa_migrate_preferred(p
);
1846 p
->numa_pages_migrated
+= pages
;
1848 p
->numa_faults_buffer_memory
[task_faults_idx(mem_node
, priv
)] += pages
;
1849 p
->numa_faults_buffer_cpu
[task_faults_idx(cpu_node
, priv
)] += pages
;
1850 p
->numa_faults_locality
[local
] += pages
;
1853 static void reset_ptenuma_scan(struct task_struct
*p
)
1855 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
1856 p
->mm
->numa_scan_offset
= 0;
1860 * The expensive part of numa migration is done from task_work context.
1861 * Triggered from task_tick_numa().
1863 void task_numa_work(struct callback_head
*work
)
1865 unsigned long migrate
, next_scan
, now
= jiffies
;
1866 struct task_struct
*p
= current
;
1867 struct mm_struct
*mm
= p
->mm
;
1868 struct vm_area_struct
*vma
;
1869 unsigned long start
, end
;
1870 unsigned long nr_pte_updates
= 0;
1873 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
1875 work
->next
= work
; /* protect against double add */
1877 * Who cares about NUMA placement when they're dying.
1879 * NOTE: make sure not to dereference p->mm before this check,
1880 * exit_task_work() happens _after_ exit_mm() so we could be called
1881 * without p->mm even though we still had it when we enqueued this
1884 if (p
->flags
& PF_EXITING
)
1887 if (!mm
->numa_next_scan
) {
1888 mm
->numa_next_scan
= now
+
1889 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1893 * Enforce maximal scan/migration frequency..
1895 migrate
= mm
->numa_next_scan
;
1896 if (time_before(now
, migrate
))
1899 if (p
->numa_scan_period
== 0) {
1900 p
->numa_scan_period_max
= task_scan_max(p
);
1901 p
->numa_scan_period
= task_scan_min(p
);
1904 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1905 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
1909 * Delay this task enough that another task of this mm will likely win
1910 * the next time around.
1912 p
->node_stamp
+= 2 * TICK_NSEC
;
1914 start
= mm
->numa_scan_offset
;
1915 pages
= sysctl_numa_balancing_scan_size
;
1916 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
1920 down_read(&mm
->mmap_sem
);
1921 vma
= find_vma(mm
, start
);
1923 reset_ptenuma_scan(p
);
1927 for (; vma
; vma
= vma
->vm_next
) {
1928 if (!vma_migratable(vma
) || !vma_policy_mof(p
, vma
))
1932 * Shared library pages mapped by multiple processes are not
1933 * migrated as it is expected they are cache replicated. Avoid
1934 * hinting faults in read-only file-backed mappings or the vdso
1935 * as migrating the pages will be of marginal benefit.
1938 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
1942 * Skip inaccessible VMAs to avoid any confusion between
1943 * PROT_NONE and NUMA hinting ptes
1945 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
1949 start
= max(start
, vma
->vm_start
);
1950 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
1951 end
= min(end
, vma
->vm_end
);
1952 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
1955 * Scan sysctl_numa_balancing_scan_size but ensure that
1956 * at least one PTE is updated so that unused virtual
1957 * address space is quickly skipped.
1960 pages
-= (end
- start
) >> PAGE_SHIFT
;
1967 } while (end
!= vma
->vm_end
);
1972 * It is possible to reach the end of the VMA list but the last few
1973 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1974 * would find the !migratable VMA on the next scan but not reset the
1975 * scanner to the start so check it now.
1978 mm
->numa_scan_offset
= start
;
1980 reset_ptenuma_scan(p
);
1981 up_read(&mm
->mmap_sem
);
1985 * Drive the periodic memory faults..
1987 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1989 struct callback_head
*work
= &curr
->numa_work
;
1993 * We don't care about NUMA placement if we don't have memory.
1995 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
1999 * Using runtime rather than walltime has the dual advantage that
2000 * we (mostly) drive the selection from busy threads and that the
2001 * task needs to have done some actual work before we bother with
2004 now
= curr
->se
.sum_exec_runtime
;
2005 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2007 if (now
- curr
->node_stamp
> period
) {
2008 if (!curr
->node_stamp
)
2009 curr
->numa_scan_period
= task_scan_min(curr
);
2010 curr
->node_stamp
+= period
;
2012 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2013 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2014 task_work_add(curr
, work
, true);
2019 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2023 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2027 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2030 #endif /* CONFIG_NUMA_BALANCING */
2033 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2035 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2036 if (!parent_entity(se
))
2037 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2039 if (entity_is_task(se
)) {
2040 struct rq
*rq
= rq_of(cfs_rq
);
2042 account_numa_enqueue(rq
, task_of(se
));
2043 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2046 cfs_rq
->nr_running
++;
2050 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2052 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2053 if (!parent_entity(se
))
2054 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2055 if (entity_is_task(se
)) {
2056 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2057 list_del_init(&se
->group_node
);
2059 cfs_rq
->nr_running
--;
2062 #ifdef CONFIG_FAIR_GROUP_SCHED
2064 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
2069 * Use this CPU's actual weight instead of the last load_contribution
2070 * to gain a more accurate current total weight. See
2071 * update_cfs_rq_load_contribution().
2073 tg_weight
= atomic_long_read(&tg
->load_avg
);
2074 tg_weight
-= cfs_rq
->tg_load_contrib
;
2075 tg_weight
+= cfs_rq
->load
.weight
;
2080 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2082 long tg_weight
, load
, shares
;
2084 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
2085 load
= cfs_rq
->load
.weight
;
2087 shares
= (tg
->shares
* load
);
2089 shares
/= tg_weight
;
2091 if (shares
< MIN_SHARES
)
2092 shares
= MIN_SHARES
;
2093 if (shares
> tg
->shares
)
2094 shares
= tg
->shares
;
2098 # else /* CONFIG_SMP */
2099 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2103 # endif /* CONFIG_SMP */
2104 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2105 unsigned long weight
)
2108 /* commit outstanding execution time */
2109 if (cfs_rq
->curr
== se
)
2110 update_curr(cfs_rq
);
2111 account_entity_dequeue(cfs_rq
, se
);
2114 update_load_set(&se
->load
, weight
);
2117 account_entity_enqueue(cfs_rq
, se
);
2120 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2122 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2124 struct task_group
*tg
;
2125 struct sched_entity
*se
;
2129 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2130 if (!se
|| throttled_hierarchy(cfs_rq
))
2133 if (likely(se
->load
.weight
== tg
->shares
))
2136 shares
= calc_cfs_shares(cfs_rq
, tg
);
2138 reweight_entity(cfs_rq_of(se
), se
, shares
);
2140 #else /* CONFIG_FAIR_GROUP_SCHED */
2141 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2144 #endif /* CONFIG_FAIR_GROUP_SCHED */
2148 * We choose a half-life close to 1 scheduling period.
2149 * Note: The tables below are dependent on this value.
2151 #define LOAD_AVG_PERIOD 32
2152 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2153 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2155 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2156 static const u32 runnable_avg_yN_inv
[] = {
2157 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2158 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2159 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2160 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2161 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2162 0x85aac367, 0x82cd8698,
2166 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2167 * over-estimates when re-combining.
2169 static const u32 runnable_avg_yN_sum
[] = {
2170 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2171 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2172 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2177 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2179 static __always_inline u64
decay_load(u64 val
, u64 n
)
2181 unsigned int local_n
;
2185 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2188 /* after bounds checking we can collapse to 32-bit */
2192 * As y^PERIOD = 1/2, we can combine
2193 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2194 * With a look-up table which covers k^n (n<PERIOD)
2196 * To achieve constant time decay_load.
2198 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2199 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2200 local_n
%= LOAD_AVG_PERIOD
;
2203 val
*= runnable_avg_yN_inv
[local_n
];
2204 /* We don't use SRR here since we always want to round down. */
2209 * For updates fully spanning n periods, the contribution to runnable
2210 * average will be: \Sum 1024*y^n
2212 * We can compute this reasonably efficiently by combining:
2213 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2215 static u32
__compute_runnable_contrib(u64 n
)
2219 if (likely(n
<= LOAD_AVG_PERIOD
))
2220 return runnable_avg_yN_sum
[n
];
2221 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2222 return LOAD_AVG_MAX
;
2224 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2226 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2227 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2229 n
-= LOAD_AVG_PERIOD
;
2230 } while (n
> LOAD_AVG_PERIOD
);
2232 contrib
= decay_load(contrib
, n
);
2233 return contrib
+ runnable_avg_yN_sum
[n
];
2237 * We can represent the historical contribution to runnable average as the
2238 * coefficients of a geometric series. To do this we sub-divide our runnable
2239 * history into segments of approximately 1ms (1024us); label the segment that
2240 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2242 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2244 * (now) (~1ms ago) (~2ms ago)
2246 * Let u_i denote the fraction of p_i that the entity was runnable.
2248 * We then designate the fractions u_i as our co-efficients, yielding the
2249 * following representation of historical load:
2250 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2252 * We choose y based on the with of a reasonably scheduling period, fixing:
2255 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2256 * approximately half as much as the contribution to load within the last ms
2259 * When a period "rolls over" and we have new u_0`, multiplying the previous
2260 * sum again by y is sufficient to update:
2261 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2262 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2264 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2265 struct sched_avg
*sa
,
2269 u32 runnable_contrib
;
2270 int delta_w
, decayed
= 0;
2272 delta
= now
- sa
->last_runnable_update
;
2274 * This should only happen when time goes backwards, which it
2275 * unfortunately does during sched clock init when we swap over to TSC.
2277 if ((s64
)delta
< 0) {
2278 sa
->last_runnable_update
= now
;
2283 * Use 1024ns as the unit of measurement since it's a reasonable
2284 * approximation of 1us and fast to compute.
2289 sa
->last_runnable_update
= now
;
2291 /* delta_w is the amount already accumulated against our next period */
2292 delta_w
= sa
->runnable_avg_period
% 1024;
2293 if (delta
+ delta_w
>= 1024) {
2294 /* period roll-over */
2298 * Now that we know we're crossing a period boundary, figure
2299 * out how much from delta we need to complete the current
2300 * period and accrue it.
2302 delta_w
= 1024 - delta_w
;
2304 sa
->runnable_avg_sum
+= delta_w
;
2305 sa
->runnable_avg_period
+= delta_w
;
2309 /* Figure out how many additional periods this update spans */
2310 periods
= delta
/ 1024;
2313 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2315 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2318 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2319 runnable_contrib
= __compute_runnable_contrib(periods
);
2321 sa
->runnable_avg_sum
+= runnable_contrib
;
2322 sa
->runnable_avg_period
+= runnable_contrib
;
2325 /* Remainder of delta accrued against u_0` */
2327 sa
->runnable_avg_sum
+= delta
;
2328 sa
->runnable_avg_period
+= delta
;
2333 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2334 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2336 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2337 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2339 decays
-= se
->avg
.decay_count
;
2343 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2344 se
->avg
.decay_count
= 0;
2349 #ifdef CONFIG_FAIR_GROUP_SCHED
2350 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2353 struct task_group
*tg
= cfs_rq
->tg
;
2356 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2357 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2359 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2360 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2361 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2366 * Aggregate cfs_rq runnable averages into an equivalent task_group
2367 * representation for computing load contributions.
2369 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2370 struct cfs_rq
*cfs_rq
)
2372 struct task_group
*tg
= cfs_rq
->tg
;
2375 /* The fraction of a cpu used by this cfs_rq */
2376 contrib
= div_u64((u64
)sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2377 sa
->runnable_avg_period
+ 1);
2378 contrib
-= cfs_rq
->tg_runnable_contrib
;
2380 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2381 atomic_add(contrib
, &tg
->runnable_avg
);
2382 cfs_rq
->tg_runnable_contrib
+= contrib
;
2386 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2388 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2389 struct task_group
*tg
= cfs_rq
->tg
;
2394 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2395 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2396 atomic_long_read(&tg
->load_avg
) + 1);
2399 * For group entities we need to compute a correction term in the case
2400 * that they are consuming <1 cpu so that we would contribute the same
2401 * load as a task of equal weight.
2403 * Explicitly co-ordinating this measurement would be expensive, but
2404 * fortunately the sum of each cpus contribution forms a usable
2405 * lower-bound on the true value.
2407 * Consider the aggregate of 2 contributions. Either they are disjoint
2408 * (and the sum represents true value) or they are disjoint and we are
2409 * understating by the aggregate of their overlap.
2411 * Extending this to N cpus, for a given overlap, the maximum amount we
2412 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2413 * cpus that overlap for this interval and w_i is the interval width.
2415 * On a small machine; the first term is well-bounded which bounds the
2416 * total error since w_i is a subset of the period. Whereas on a
2417 * larger machine, while this first term can be larger, if w_i is the
2418 * of consequential size guaranteed to see n_i*w_i quickly converge to
2419 * our upper bound of 1-cpu.
2421 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2422 if (runnable_avg
< NICE_0_LOAD
) {
2423 se
->avg
.load_avg_contrib
*= runnable_avg
;
2424 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2428 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2430 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2431 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2433 #else /* CONFIG_FAIR_GROUP_SCHED */
2434 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2435 int force_update
) {}
2436 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2437 struct cfs_rq
*cfs_rq
) {}
2438 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2439 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2440 #endif /* CONFIG_FAIR_GROUP_SCHED */
2442 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2446 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2447 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2448 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2449 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2452 /* Compute the current contribution to load_avg by se, return any delta */
2453 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2455 long old_contrib
= se
->avg
.load_avg_contrib
;
2457 if (entity_is_task(se
)) {
2458 __update_task_entity_contrib(se
);
2460 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2461 __update_group_entity_contrib(se
);
2464 return se
->avg
.load_avg_contrib
- old_contrib
;
2467 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2470 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2471 cfs_rq
->blocked_load_avg
-= load_contrib
;
2473 cfs_rq
->blocked_load_avg
= 0;
2476 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2478 /* Update a sched_entity's runnable average */
2479 static inline void update_entity_load_avg(struct sched_entity
*se
,
2482 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2487 * For a group entity we need to use their owned cfs_rq_clock_task() in
2488 * case they are the parent of a throttled hierarchy.
2490 if (entity_is_task(se
))
2491 now
= cfs_rq_clock_task(cfs_rq
);
2493 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2495 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2498 contrib_delta
= __update_entity_load_avg_contrib(se
);
2504 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2506 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2510 * Decay the load contributed by all blocked children and account this so that
2511 * their contribution may appropriately discounted when they wake up.
2513 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2515 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2518 decays
= now
- cfs_rq
->last_decay
;
2519 if (!decays
&& !force_update
)
2522 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2523 unsigned long removed_load
;
2524 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2525 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2529 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2531 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2532 cfs_rq
->last_decay
= now
;
2535 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2538 /* Add the load generated by se into cfs_rq's child load-average */
2539 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2540 struct sched_entity
*se
,
2544 * We track migrations using entity decay_count <= 0, on a wake-up
2545 * migration we use a negative decay count to track the remote decays
2546 * accumulated while sleeping.
2548 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2549 * are seen by enqueue_entity_load_avg() as a migration with an already
2550 * constructed load_avg_contrib.
2552 if (unlikely(se
->avg
.decay_count
<= 0)) {
2553 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2554 if (se
->avg
.decay_count
) {
2556 * In a wake-up migration we have to approximate the
2557 * time sleeping. This is because we can't synchronize
2558 * clock_task between the two cpus, and it is not
2559 * guaranteed to be read-safe. Instead, we can
2560 * approximate this using our carried decays, which are
2561 * explicitly atomically readable.
2563 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2565 update_entity_load_avg(se
, 0);
2566 /* Indicate that we're now synchronized and on-rq */
2567 se
->avg
.decay_count
= 0;
2571 __synchronize_entity_decay(se
);
2574 /* migrated tasks did not contribute to our blocked load */
2576 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2577 update_entity_load_avg(se
, 0);
2580 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2581 /* we force update consideration on load-balancer moves */
2582 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2586 * Remove se's load from this cfs_rq child load-average, if the entity is
2587 * transitioning to a blocked state we track its projected decay using
2590 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2591 struct sched_entity
*se
,
2594 update_entity_load_avg(se
, 1);
2595 /* we force update consideration on load-balancer moves */
2596 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2598 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2600 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2601 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2602 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2606 * Update the rq's load with the elapsed running time before entering
2607 * idle. if the last scheduled task is not a CFS task, idle_enter will
2608 * be the only way to update the runnable statistic.
2610 void idle_enter_fair(struct rq
*this_rq
)
2612 update_rq_runnable_avg(this_rq
, 1);
2616 * Update the rq's load with the elapsed idle time before a task is
2617 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2618 * be the only way to update the runnable statistic.
2620 void idle_exit_fair(struct rq
*this_rq
)
2622 update_rq_runnable_avg(this_rq
, 0);
2625 static int idle_balance(struct rq
*this_rq
);
2627 #else /* CONFIG_SMP */
2629 static inline void update_entity_load_avg(struct sched_entity
*se
,
2630 int update_cfs_rq
) {}
2631 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2632 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2633 struct sched_entity
*se
,
2635 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2636 struct sched_entity
*se
,
2638 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2639 int force_update
) {}
2641 static inline int idle_balance(struct rq
*rq
)
2646 #endif /* CONFIG_SMP */
2648 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2650 #ifdef CONFIG_SCHEDSTATS
2651 struct task_struct
*tsk
= NULL
;
2653 if (entity_is_task(se
))
2656 if (se
->statistics
.sleep_start
) {
2657 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2662 if (unlikely(delta
> se
->statistics
.sleep_max
))
2663 se
->statistics
.sleep_max
= delta
;
2665 se
->statistics
.sleep_start
= 0;
2666 se
->statistics
.sum_sleep_runtime
+= delta
;
2669 account_scheduler_latency(tsk
, delta
>> 10, 1);
2670 trace_sched_stat_sleep(tsk
, delta
);
2673 if (se
->statistics
.block_start
) {
2674 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2679 if (unlikely(delta
> se
->statistics
.block_max
))
2680 se
->statistics
.block_max
= delta
;
2682 se
->statistics
.block_start
= 0;
2683 se
->statistics
.sum_sleep_runtime
+= delta
;
2686 if (tsk
->in_iowait
) {
2687 se
->statistics
.iowait_sum
+= delta
;
2688 se
->statistics
.iowait_count
++;
2689 trace_sched_stat_iowait(tsk
, delta
);
2692 trace_sched_stat_blocked(tsk
, delta
);
2695 * Blocking time is in units of nanosecs, so shift by
2696 * 20 to get a milliseconds-range estimation of the
2697 * amount of time that the task spent sleeping:
2699 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2700 profile_hits(SLEEP_PROFILING
,
2701 (void *)get_wchan(tsk
),
2704 account_scheduler_latency(tsk
, delta
>> 10, 0);
2710 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2712 #ifdef CONFIG_SCHED_DEBUG
2713 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2718 if (d
> 3*sysctl_sched_latency
)
2719 schedstat_inc(cfs_rq
, nr_spread_over
);
2724 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2726 u64 vruntime
= cfs_rq
->min_vruntime
;
2729 * The 'current' period is already promised to the current tasks,
2730 * however the extra weight of the new task will slow them down a
2731 * little, place the new task so that it fits in the slot that
2732 * stays open at the end.
2734 if (initial
&& sched_feat(START_DEBIT
))
2735 vruntime
+= sched_vslice(cfs_rq
, se
);
2737 /* sleeps up to a single latency don't count. */
2739 unsigned long thresh
= sysctl_sched_latency
;
2742 * Halve their sleep time's effect, to allow
2743 * for a gentler effect of sleepers:
2745 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2751 /* ensure we never gain time by being placed backwards. */
2752 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
2755 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
2758 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2761 * Update the normalized vruntime before updating min_vruntime
2762 * through calling update_curr().
2764 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
2765 se
->vruntime
+= cfs_rq
->min_vruntime
;
2768 * Update run-time statistics of the 'current'.
2770 update_curr(cfs_rq
);
2771 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
2772 account_entity_enqueue(cfs_rq
, se
);
2773 update_cfs_shares(cfs_rq
);
2775 if (flags
& ENQUEUE_WAKEUP
) {
2776 place_entity(cfs_rq
, se
, 0);
2777 enqueue_sleeper(cfs_rq
, se
);
2780 update_stats_enqueue(cfs_rq
, se
);
2781 check_spread(cfs_rq
, se
);
2782 if (se
!= cfs_rq
->curr
)
2783 __enqueue_entity(cfs_rq
, se
);
2786 if (cfs_rq
->nr_running
== 1) {
2787 list_add_leaf_cfs_rq(cfs_rq
);
2788 check_enqueue_throttle(cfs_rq
);
2792 static void __clear_buddies_last(struct sched_entity
*se
)
2794 for_each_sched_entity(se
) {
2795 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2796 if (cfs_rq
->last
!= se
)
2799 cfs_rq
->last
= NULL
;
2803 static void __clear_buddies_next(struct sched_entity
*se
)
2805 for_each_sched_entity(se
) {
2806 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2807 if (cfs_rq
->next
!= se
)
2810 cfs_rq
->next
= NULL
;
2814 static void __clear_buddies_skip(struct sched_entity
*se
)
2816 for_each_sched_entity(se
) {
2817 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2818 if (cfs_rq
->skip
!= se
)
2821 cfs_rq
->skip
= NULL
;
2825 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2827 if (cfs_rq
->last
== se
)
2828 __clear_buddies_last(se
);
2830 if (cfs_rq
->next
== se
)
2831 __clear_buddies_next(se
);
2833 if (cfs_rq
->skip
== se
)
2834 __clear_buddies_skip(se
);
2837 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2840 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2843 * Update run-time statistics of the 'current'.
2845 update_curr(cfs_rq
);
2846 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
2848 update_stats_dequeue(cfs_rq
, se
);
2849 if (flags
& DEQUEUE_SLEEP
) {
2850 #ifdef CONFIG_SCHEDSTATS
2851 if (entity_is_task(se
)) {
2852 struct task_struct
*tsk
= task_of(se
);
2854 if (tsk
->state
& TASK_INTERRUPTIBLE
)
2855 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
2856 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
2857 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
2862 clear_buddies(cfs_rq
, se
);
2864 if (se
!= cfs_rq
->curr
)
2865 __dequeue_entity(cfs_rq
, se
);
2867 account_entity_dequeue(cfs_rq
, se
);
2870 * Normalize the entity after updating the min_vruntime because the
2871 * update can refer to the ->curr item and we need to reflect this
2872 * movement in our normalized position.
2874 if (!(flags
& DEQUEUE_SLEEP
))
2875 se
->vruntime
-= cfs_rq
->min_vruntime
;
2877 /* return excess runtime on last dequeue */
2878 return_cfs_rq_runtime(cfs_rq
);
2880 update_min_vruntime(cfs_rq
);
2881 update_cfs_shares(cfs_rq
);
2885 * Preempt the current task with a newly woken task if needed:
2888 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2890 unsigned long ideal_runtime
, delta_exec
;
2891 struct sched_entity
*se
;
2894 ideal_runtime
= sched_slice(cfs_rq
, curr
);
2895 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
2896 if (delta_exec
> ideal_runtime
) {
2897 resched_task(rq_of(cfs_rq
)->curr
);
2899 * The current task ran long enough, ensure it doesn't get
2900 * re-elected due to buddy favours.
2902 clear_buddies(cfs_rq
, curr
);
2907 * Ensure that a task that missed wakeup preemption by a
2908 * narrow margin doesn't have to wait for a full slice.
2909 * This also mitigates buddy induced latencies under load.
2911 if (delta_exec
< sysctl_sched_min_granularity
)
2914 se
= __pick_first_entity(cfs_rq
);
2915 delta
= curr
->vruntime
- se
->vruntime
;
2920 if (delta
> ideal_runtime
)
2921 resched_task(rq_of(cfs_rq
)->curr
);
2925 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2927 /* 'current' is not kept within the tree. */
2930 * Any task has to be enqueued before it get to execute on
2931 * a CPU. So account for the time it spent waiting on the
2934 update_stats_wait_end(cfs_rq
, se
);
2935 __dequeue_entity(cfs_rq
, se
);
2938 update_stats_curr_start(cfs_rq
, se
);
2940 #ifdef CONFIG_SCHEDSTATS
2942 * Track our maximum slice length, if the CPU's load is at
2943 * least twice that of our own weight (i.e. dont track it
2944 * when there are only lesser-weight tasks around):
2946 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
2947 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
2948 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
2951 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
2955 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
2958 * Pick the next process, keeping these things in mind, in this order:
2959 * 1) keep things fair between processes/task groups
2960 * 2) pick the "next" process, since someone really wants that to run
2961 * 3) pick the "last" process, for cache locality
2962 * 4) do not run the "skip" process, if something else is available
2964 static struct sched_entity
*
2965 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2967 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
2968 struct sched_entity
*se
;
2971 * If curr is set we have to see if its left of the leftmost entity
2972 * still in the tree, provided there was anything in the tree at all.
2974 if (!left
|| (curr
&& entity_before(curr
, left
)))
2977 se
= left
; /* ideally we run the leftmost entity */
2980 * Avoid running the skip buddy, if running something else can
2981 * be done without getting too unfair.
2983 if (cfs_rq
->skip
== se
) {
2984 struct sched_entity
*second
;
2987 second
= __pick_first_entity(cfs_rq
);
2989 second
= __pick_next_entity(se
);
2990 if (!second
|| (curr
&& entity_before(curr
, second
)))
2994 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
2999 * Prefer last buddy, try to return the CPU to a preempted task.
3001 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3005 * Someone really wants this to run. If it's not unfair, run it.
3007 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3010 clear_buddies(cfs_rq
, se
);
3015 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3017 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3020 * If still on the runqueue then deactivate_task()
3021 * was not called and update_curr() has to be done:
3024 update_curr(cfs_rq
);
3026 /* throttle cfs_rqs exceeding runtime */
3027 check_cfs_rq_runtime(cfs_rq
);
3029 check_spread(cfs_rq
, prev
);
3031 update_stats_wait_start(cfs_rq
, prev
);
3032 /* Put 'current' back into the tree. */
3033 __enqueue_entity(cfs_rq
, prev
);
3034 /* in !on_rq case, update occurred at dequeue */
3035 update_entity_load_avg(prev
, 1);
3037 cfs_rq
->curr
= NULL
;
3041 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3044 * Update run-time statistics of the 'current'.
3046 update_curr(cfs_rq
);
3049 * Ensure that runnable average is periodically updated.
3051 update_entity_load_avg(curr
, 1);
3052 update_cfs_rq_blocked_load(cfs_rq
, 1);
3053 update_cfs_shares(cfs_rq
);
3055 #ifdef CONFIG_SCHED_HRTICK
3057 * queued ticks are scheduled to match the slice, so don't bother
3058 * validating it and just reschedule.
3061 resched_task(rq_of(cfs_rq
)->curr
);
3065 * don't let the period tick interfere with the hrtick preemption
3067 if (!sched_feat(DOUBLE_TICK
) &&
3068 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3072 if (cfs_rq
->nr_running
> 1)
3073 check_preempt_tick(cfs_rq
, curr
);
3077 /**************************************************
3078 * CFS bandwidth control machinery
3081 #ifdef CONFIG_CFS_BANDWIDTH
3083 #ifdef HAVE_JUMP_LABEL
3084 static struct static_key __cfs_bandwidth_used
;
3086 static inline bool cfs_bandwidth_used(void)
3088 return static_key_false(&__cfs_bandwidth_used
);
3091 void cfs_bandwidth_usage_inc(void)
3093 static_key_slow_inc(&__cfs_bandwidth_used
);
3096 void cfs_bandwidth_usage_dec(void)
3098 static_key_slow_dec(&__cfs_bandwidth_used
);
3100 #else /* HAVE_JUMP_LABEL */
3101 static bool cfs_bandwidth_used(void)
3106 void cfs_bandwidth_usage_inc(void) {}
3107 void cfs_bandwidth_usage_dec(void) {}
3108 #endif /* HAVE_JUMP_LABEL */
3111 * default period for cfs group bandwidth.
3112 * default: 0.1s, units: nanoseconds
3114 static inline u64
default_cfs_period(void)
3116 return 100000000ULL;
3119 static inline u64
sched_cfs_bandwidth_slice(void)
3121 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3125 * Replenish runtime according to assigned quota and update expiration time.
3126 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3127 * additional synchronization around rq->lock.
3129 * requires cfs_b->lock
3131 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3135 if (cfs_b
->quota
== RUNTIME_INF
)
3138 now
= sched_clock_cpu(smp_processor_id());
3139 cfs_b
->runtime
= cfs_b
->quota
;
3140 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3143 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3145 return &tg
->cfs_bandwidth
;
3148 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3149 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3151 if (unlikely(cfs_rq
->throttle_count
))
3152 return cfs_rq
->throttled_clock_task
;
3154 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3157 /* returns 0 on failure to allocate runtime */
3158 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3160 struct task_group
*tg
= cfs_rq
->tg
;
3161 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3162 u64 amount
= 0, min_amount
, expires
;
3164 /* note: this is a positive sum as runtime_remaining <= 0 */
3165 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3167 raw_spin_lock(&cfs_b
->lock
);
3168 if (cfs_b
->quota
== RUNTIME_INF
)
3169 amount
= min_amount
;
3172 * If the bandwidth pool has become inactive, then at least one
3173 * period must have elapsed since the last consumption.
3174 * Refresh the global state and ensure bandwidth timer becomes
3177 if (!cfs_b
->timer_active
) {
3178 __refill_cfs_bandwidth_runtime(cfs_b
);
3179 __start_cfs_bandwidth(cfs_b
, false);
3182 if (cfs_b
->runtime
> 0) {
3183 amount
= min(cfs_b
->runtime
, min_amount
);
3184 cfs_b
->runtime
-= amount
;
3188 expires
= cfs_b
->runtime_expires
;
3189 raw_spin_unlock(&cfs_b
->lock
);
3191 cfs_rq
->runtime_remaining
+= amount
;
3193 * we may have advanced our local expiration to account for allowed
3194 * spread between our sched_clock and the one on which runtime was
3197 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3198 cfs_rq
->runtime_expires
= expires
;
3200 return cfs_rq
->runtime_remaining
> 0;
3204 * Note: This depends on the synchronization provided by sched_clock and the
3205 * fact that rq->clock snapshots this value.
3207 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3209 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3211 /* if the deadline is ahead of our clock, nothing to do */
3212 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3215 if (cfs_rq
->runtime_remaining
< 0)
3219 * If the local deadline has passed we have to consider the
3220 * possibility that our sched_clock is 'fast' and the global deadline
3221 * has not truly expired.
3223 * Fortunately we can check determine whether this the case by checking
3224 * whether the global deadline has advanced. It is valid to compare
3225 * cfs_b->runtime_expires without any locks since we only care about
3226 * exact equality, so a partial write will still work.
3229 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
3230 /* extend local deadline, drift is bounded above by 2 ticks */
3231 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3233 /* global deadline is ahead, expiration has passed */
3234 cfs_rq
->runtime_remaining
= 0;
3238 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3240 /* dock delta_exec before expiring quota (as it could span periods) */
3241 cfs_rq
->runtime_remaining
-= delta_exec
;
3242 expire_cfs_rq_runtime(cfs_rq
);
3244 if (likely(cfs_rq
->runtime_remaining
> 0))
3248 * if we're unable to extend our runtime we resched so that the active
3249 * hierarchy can be throttled
3251 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3252 resched_task(rq_of(cfs_rq
)->curr
);
3255 static __always_inline
3256 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3258 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3261 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3264 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3266 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3269 /* check whether cfs_rq, or any parent, is throttled */
3270 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3272 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3276 * Ensure that neither of the group entities corresponding to src_cpu or
3277 * dest_cpu are members of a throttled hierarchy when performing group
3278 * load-balance operations.
3280 static inline int throttled_lb_pair(struct task_group
*tg
,
3281 int src_cpu
, int dest_cpu
)
3283 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3285 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3286 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3288 return throttled_hierarchy(src_cfs_rq
) ||
3289 throttled_hierarchy(dest_cfs_rq
);
3292 /* updated child weight may affect parent so we have to do this bottom up */
3293 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3295 struct rq
*rq
= data
;
3296 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3298 cfs_rq
->throttle_count
--;
3300 if (!cfs_rq
->throttle_count
) {
3301 /* adjust cfs_rq_clock_task() */
3302 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3303 cfs_rq
->throttled_clock_task
;
3310 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3312 struct rq
*rq
= data
;
3313 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3315 /* group is entering throttled state, stop time */
3316 if (!cfs_rq
->throttle_count
)
3317 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3318 cfs_rq
->throttle_count
++;
3323 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3325 struct rq
*rq
= rq_of(cfs_rq
);
3326 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3327 struct sched_entity
*se
;
3328 long task_delta
, dequeue
= 1;
3330 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3332 /* freeze hierarchy runnable averages while throttled */
3334 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3337 task_delta
= cfs_rq
->h_nr_running
;
3338 for_each_sched_entity(se
) {
3339 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3340 /* throttled entity or throttle-on-deactivate */
3345 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3346 qcfs_rq
->h_nr_running
-= task_delta
;
3348 if (qcfs_rq
->load
.weight
)
3353 sub_nr_running(rq
, task_delta
);
3355 cfs_rq
->throttled
= 1;
3356 cfs_rq
->throttled_clock
= rq_clock(rq
);
3357 raw_spin_lock(&cfs_b
->lock
);
3359 * Add to the _head_ of the list, so that an already-started
3360 * distribute_cfs_runtime will not see us
3362 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3363 if (!cfs_b
->timer_active
)
3364 __start_cfs_bandwidth(cfs_b
, false);
3365 raw_spin_unlock(&cfs_b
->lock
);
3368 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3370 struct rq
*rq
= rq_of(cfs_rq
);
3371 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3372 struct sched_entity
*se
;
3376 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3378 cfs_rq
->throttled
= 0;
3380 update_rq_clock(rq
);
3382 raw_spin_lock(&cfs_b
->lock
);
3383 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3384 list_del_rcu(&cfs_rq
->throttled_list
);
3385 raw_spin_unlock(&cfs_b
->lock
);
3387 /* update hierarchical throttle state */
3388 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3390 if (!cfs_rq
->load
.weight
)
3393 task_delta
= cfs_rq
->h_nr_running
;
3394 for_each_sched_entity(se
) {
3398 cfs_rq
= cfs_rq_of(se
);
3400 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3401 cfs_rq
->h_nr_running
+= task_delta
;
3403 if (cfs_rq_throttled(cfs_rq
))
3408 add_nr_running(rq
, task_delta
);
3410 /* determine whether we need to wake up potentially idle cpu */
3411 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3412 resched_task(rq
->curr
);
3415 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3416 u64 remaining
, u64 expires
)
3418 struct cfs_rq
*cfs_rq
;
3420 u64 starting_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
);
3451 return starting_runtime
- remaining
;
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
;
3497 runtime_expires
= cfs_b
->runtime_expires
;
3500 * This check is repeated as we are holding onto the new bandwidth while
3501 * we unthrottle. This can potentially race with an unthrottled group
3502 * trying to acquire new bandwidth from the global pool. This can result
3503 * in us over-using our runtime if it is all used during this loop, but
3504 * only by limited amounts in that extreme case.
3506 while (throttled
&& cfs_b
->runtime
> 0) {
3507 runtime
= cfs_b
->runtime
;
3508 raw_spin_unlock(&cfs_b
->lock
);
3509 /* we can't nest cfs_b->lock while distributing bandwidth */
3510 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3512 raw_spin_lock(&cfs_b
->lock
);
3514 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3516 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3520 * While we are ensured activity in the period following an
3521 * unthrottle, this also covers the case in which the new bandwidth is
3522 * insufficient to cover the existing bandwidth deficit. (Forcing the
3523 * timer to remain active while there are any throttled entities.)
3530 cfs_b
->timer_active
= 0;
3534 /* a cfs_rq won't donate quota below this amount */
3535 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3536 /* minimum remaining period time to redistribute slack quota */
3537 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3538 /* how long we wait to gather additional slack before distributing */
3539 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3542 * Are we near the end of the current quota period?
3544 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3545 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3546 * migrate_hrtimers, base is never cleared, so we are fine.
3548 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3550 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3553 /* if the call-back is running a quota refresh is already occurring */
3554 if (hrtimer_callback_running(refresh_timer
))
3557 /* is a quota refresh about to occur? */
3558 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3559 if (remaining
< min_expire
)
3565 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3567 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3569 /* if there's a quota refresh soon don't bother with slack */
3570 if (runtime_refresh_within(cfs_b
, min_left
))
3573 start_bandwidth_timer(&cfs_b
->slack_timer
,
3574 ns_to_ktime(cfs_bandwidth_slack_period
));
3577 /* we know any runtime found here is valid as update_curr() precedes return */
3578 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3580 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3581 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3583 if (slack_runtime
<= 0)
3586 raw_spin_lock(&cfs_b
->lock
);
3587 if (cfs_b
->quota
!= RUNTIME_INF
&&
3588 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3589 cfs_b
->runtime
+= slack_runtime
;
3591 /* we are under rq->lock, defer unthrottling using a timer */
3592 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3593 !list_empty(&cfs_b
->throttled_cfs_rq
))
3594 start_cfs_slack_bandwidth(cfs_b
);
3596 raw_spin_unlock(&cfs_b
->lock
);
3598 /* even if it's not valid for return we don't want to try again */
3599 cfs_rq
->runtime_remaining
-= slack_runtime
;
3602 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3604 if (!cfs_bandwidth_used())
3607 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3610 __return_cfs_rq_runtime(cfs_rq
);
3614 * This is done with a timer (instead of inline with bandwidth return) since
3615 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3617 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3619 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3622 /* confirm we're still not at a refresh boundary */
3623 raw_spin_lock(&cfs_b
->lock
);
3624 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
3625 raw_spin_unlock(&cfs_b
->lock
);
3629 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
3630 runtime
= cfs_b
->runtime
;
3632 expires
= cfs_b
->runtime_expires
;
3633 raw_spin_unlock(&cfs_b
->lock
);
3638 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3640 raw_spin_lock(&cfs_b
->lock
);
3641 if (expires
== cfs_b
->runtime_expires
)
3642 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3643 raw_spin_unlock(&cfs_b
->lock
);
3647 * When a group wakes up we want to make sure that its quota is not already
3648 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3649 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3651 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3653 if (!cfs_bandwidth_used())
3656 /* an active group must be handled by the update_curr()->put() path */
3657 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3660 /* ensure the group is not already throttled */
3661 if (cfs_rq_throttled(cfs_rq
))
3664 /* update runtime allocation */
3665 account_cfs_rq_runtime(cfs_rq
, 0);
3666 if (cfs_rq
->runtime_remaining
<= 0)
3667 throttle_cfs_rq(cfs_rq
);
3670 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3671 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3673 if (!cfs_bandwidth_used())
3676 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3680 * it's possible for a throttled entity to be forced into a running
3681 * state (e.g. set_curr_task), in this case we're finished.
3683 if (cfs_rq_throttled(cfs_rq
))
3686 throttle_cfs_rq(cfs_rq
);
3690 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3692 struct cfs_bandwidth
*cfs_b
=
3693 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3694 do_sched_cfs_slack_timer(cfs_b
);
3696 return HRTIMER_NORESTART
;
3699 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3701 struct cfs_bandwidth
*cfs_b
=
3702 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3707 raw_spin_lock(&cfs_b
->lock
);
3709 now
= hrtimer_cb_get_time(timer
);
3710 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3715 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3717 raw_spin_unlock(&cfs_b
->lock
);
3719 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3722 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3724 raw_spin_lock_init(&cfs_b
->lock
);
3726 cfs_b
->quota
= RUNTIME_INF
;
3727 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3729 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3730 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3731 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3732 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3733 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3736 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3738 cfs_rq
->runtime_enabled
= 0;
3739 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3742 /* requires cfs_b->lock, may release to reprogram timer */
3743 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
, bool force
)
3746 * The timer may be active because we're trying to set a new bandwidth
3747 * period or because we're racing with the tear-down path
3748 * (timer_active==0 becomes visible before the hrtimer call-back
3749 * terminates). In either case we ensure that it's re-programmed
3751 while (unlikely(hrtimer_active(&cfs_b
->period_timer
)) &&
3752 hrtimer_try_to_cancel(&cfs_b
->period_timer
) < 0) {
3753 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3754 raw_spin_unlock(&cfs_b
->lock
);
3756 raw_spin_lock(&cfs_b
->lock
);
3757 /* if someone else restarted the timer then we're done */
3758 if (!force
&& cfs_b
->timer_active
)
3762 cfs_b
->timer_active
= 1;
3763 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
3766 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3768 hrtimer_cancel(&cfs_b
->period_timer
);
3769 hrtimer_cancel(&cfs_b
->slack_timer
);
3772 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
3774 struct cfs_rq
*cfs_rq
;
3776 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3777 if (!cfs_rq
->runtime_enabled
)
3781 * clock_task is not advancing so we just need to make sure
3782 * there's some valid quota amount
3784 cfs_rq
->runtime_remaining
= 1;
3785 if (cfs_rq_throttled(cfs_rq
))
3786 unthrottle_cfs_rq(cfs_rq
);
3790 #else /* CONFIG_CFS_BANDWIDTH */
3791 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3793 return rq_clock_task(rq_of(cfs_rq
));
3796 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
3797 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
3798 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
3799 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3801 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3806 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3811 static inline int throttled_lb_pair(struct task_group
*tg
,
3812 int src_cpu
, int dest_cpu
)
3817 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3819 #ifdef CONFIG_FAIR_GROUP_SCHED
3820 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3823 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3827 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3828 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
3830 #endif /* CONFIG_CFS_BANDWIDTH */
3832 /**************************************************
3833 * CFS operations on tasks:
3836 #ifdef CONFIG_SCHED_HRTICK
3837 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3839 struct sched_entity
*se
= &p
->se
;
3840 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3842 WARN_ON(task_rq(p
) != rq
);
3844 if (cfs_rq
->nr_running
> 1) {
3845 u64 slice
= sched_slice(cfs_rq
, se
);
3846 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
3847 s64 delta
= slice
- ran
;
3856 * Don't schedule slices shorter than 10000ns, that just
3857 * doesn't make sense. Rely on vruntime for fairness.
3860 delta
= max_t(s64
, 10000LL, delta
);
3862 hrtick_start(rq
, delta
);
3867 * called from enqueue/dequeue and updates the hrtick when the
3868 * current task is from our class and nr_running is low enough
3871 static void hrtick_update(struct rq
*rq
)
3873 struct task_struct
*curr
= rq
->curr
;
3875 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
3878 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
3879 hrtick_start_fair(rq
, curr
);
3881 #else /* !CONFIG_SCHED_HRTICK */
3883 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3887 static inline void hrtick_update(struct rq
*rq
)
3893 * The enqueue_task method is called before nr_running is
3894 * increased. Here we update the fair scheduling stats and
3895 * then put the task into the rbtree:
3898 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3900 struct cfs_rq
*cfs_rq
;
3901 struct sched_entity
*se
= &p
->se
;
3903 for_each_sched_entity(se
) {
3906 cfs_rq
= cfs_rq_of(se
);
3907 enqueue_entity(cfs_rq
, se
, flags
);
3910 * end evaluation on encountering a throttled cfs_rq
3912 * note: in the case of encountering a throttled cfs_rq we will
3913 * post the final h_nr_running increment below.
3915 if (cfs_rq_throttled(cfs_rq
))
3917 cfs_rq
->h_nr_running
++;
3919 flags
= ENQUEUE_WAKEUP
;
3922 for_each_sched_entity(se
) {
3923 cfs_rq
= cfs_rq_of(se
);
3924 cfs_rq
->h_nr_running
++;
3926 if (cfs_rq_throttled(cfs_rq
))
3929 update_cfs_shares(cfs_rq
);
3930 update_entity_load_avg(se
, 1);
3934 update_rq_runnable_avg(rq
, rq
->nr_running
);
3935 add_nr_running(rq
, 1);
3940 static void set_next_buddy(struct sched_entity
*se
);
3943 * The dequeue_task method is called before nr_running is
3944 * decreased. We remove the task from the rbtree and
3945 * update the fair scheduling stats:
3947 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3949 struct cfs_rq
*cfs_rq
;
3950 struct sched_entity
*se
= &p
->se
;
3951 int task_sleep
= flags
& DEQUEUE_SLEEP
;
3953 for_each_sched_entity(se
) {
3954 cfs_rq
= cfs_rq_of(se
);
3955 dequeue_entity(cfs_rq
, se
, flags
);
3958 * end evaluation on encountering a throttled cfs_rq
3960 * note: in the case of encountering a throttled cfs_rq we will
3961 * post the final h_nr_running decrement below.
3963 if (cfs_rq_throttled(cfs_rq
))
3965 cfs_rq
->h_nr_running
--;
3967 /* Don't dequeue parent if it has other entities besides us */
3968 if (cfs_rq
->load
.weight
) {
3970 * Bias pick_next to pick a task from this cfs_rq, as
3971 * p is sleeping when it is within its sched_slice.
3973 if (task_sleep
&& parent_entity(se
))
3974 set_next_buddy(parent_entity(se
));
3976 /* avoid re-evaluating load for this entity */
3977 se
= parent_entity(se
);
3980 flags
|= DEQUEUE_SLEEP
;
3983 for_each_sched_entity(se
) {
3984 cfs_rq
= cfs_rq_of(se
);
3985 cfs_rq
->h_nr_running
--;
3987 if (cfs_rq_throttled(cfs_rq
))
3990 update_cfs_shares(cfs_rq
);
3991 update_entity_load_avg(se
, 1);
3995 sub_nr_running(rq
, 1);
3996 update_rq_runnable_avg(rq
, 1);
4002 /* Used instead of source_load when we know the type == 0 */
4003 static unsigned long weighted_cpuload(const int cpu
)
4005 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
4009 * Return a low guess at the load of a migration-source cpu weighted
4010 * according to the scheduling class and "nice" value.
4012 * We want to under-estimate the load of migration sources, to
4013 * balance conservatively.
4015 static unsigned long source_load(int cpu
, int type
)
4017 struct rq
*rq
= cpu_rq(cpu
);
4018 unsigned long total
= weighted_cpuload(cpu
);
4020 if (type
== 0 || !sched_feat(LB_BIAS
))
4023 return min(rq
->cpu_load
[type
-1], total
);
4027 * Return a high guess at the load of a migration-target cpu weighted
4028 * according to the scheduling class and "nice" value.
4030 static unsigned long target_load(int cpu
, int type
)
4032 struct rq
*rq
= cpu_rq(cpu
);
4033 unsigned long total
= weighted_cpuload(cpu
);
4035 if (type
== 0 || !sched_feat(LB_BIAS
))
4038 return max(rq
->cpu_load
[type
-1], total
);
4041 static unsigned long capacity_of(int cpu
)
4043 return cpu_rq(cpu
)->cpu_capacity
;
4046 static unsigned long cpu_avg_load_per_task(int cpu
)
4048 struct rq
*rq
= cpu_rq(cpu
);
4049 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
4050 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
4053 return load_avg
/ nr_running
;
4058 static void record_wakee(struct task_struct
*p
)
4061 * Rough decay (wiping) for cost saving, don't worry
4062 * about the boundary, really active task won't care
4065 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
4066 current
->wakee_flips
>>= 1;
4067 current
->wakee_flip_decay_ts
= jiffies
;
4070 if (current
->last_wakee
!= p
) {
4071 current
->last_wakee
= p
;
4072 current
->wakee_flips
++;
4076 static void task_waking_fair(struct task_struct
*p
)
4078 struct sched_entity
*se
= &p
->se
;
4079 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4082 #ifndef CONFIG_64BIT
4083 u64 min_vruntime_copy
;
4086 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
4088 min_vruntime
= cfs_rq
->min_vruntime
;
4089 } while (min_vruntime
!= min_vruntime_copy
);
4091 min_vruntime
= cfs_rq
->min_vruntime
;
4094 se
->vruntime
-= min_vruntime
;
4098 #ifdef CONFIG_FAIR_GROUP_SCHED
4100 * effective_load() calculates the load change as seen from the root_task_group
4102 * Adding load to a group doesn't make a group heavier, but can cause movement
4103 * of group shares between cpus. Assuming the shares were perfectly aligned one
4104 * can calculate the shift in shares.
4106 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4107 * on this @cpu and results in a total addition (subtraction) of @wg to the
4108 * total group weight.
4110 * Given a runqueue weight distribution (rw_i) we can compute a shares
4111 * distribution (s_i) using:
4113 * s_i = rw_i / \Sum rw_j (1)
4115 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4116 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4117 * shares distribution (s_i):
4119 * rw_i = { 2, 4, 1, 0 }
4120 * s_i = { 2/7, 4/7, 1/7, 0 }
4122 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4123 * task used to run on and the CPU the waker is running on), we need to
4124 * compute the effect of waking a task on either CPU and, in case of a sync
4125 * wakeup, compute the effect of the current task going to sleep.
4127 * So for a change of @wl to the local @cpu with an overall group weight change
4128 * of @wl we can compute the new shares distribution (s'_i) using:
4130 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4132 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4133 * differences in waking a task to CPU 0. The additional task changes the
4134 * weight and shares distributions like:
4136 * rw'_i = { 3, 4, 1, 0 }
4137 * s'_i = { 3/8, 4/8, 1/8, 0 }
4139 * We can then compute the difference in effective weight by using:
4141 * dw_i = S * (s'_i - s_i) (3)
4143 * Where 'S' is the group weight as seen by its parent.
4145 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4146 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4147 * 4/7) times the weight of the group.
4149 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4151 struct sched_entity
*se
= tg
->se
[cpu
];
4153 if (!tg
->parent
) /* the trivial, non-cgroup case */
4156 for_each_sched_entity(se
) {
4162 * W = @wg + \Sum rw_j
4164 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4169 w
= se
->my_q
->load
.weight
+ wl
;
4172 * wl = S * s'_i; see (2)
4175 wl
= (w
* tg
->shares
) / W
;
4180 * Per the above, wl is the new se->load.weight value; since
4181 * those are clipped to [MIN_SHARES, ...) do so now. See
4182 * calc_cfs_shares().
4184 if (wl
< MIN_SHARES
)
4188 * wl = dw_i = S * (s'_i - s_i); see (3)
4190 wl
-= se
->load
.weight
;
4193 * Recursively apply this logic to all parent groups to compute
4194 * the final effective load change on the root group. Since
4195 * only the @tg group gets extra weight, all parent groups can
4196 * only redistribute existing shares. @wl is the shift in shares
4197 * resulting from this level per the above.
4206 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4213 static int wake_wide(struct task_struct
*p
)
4215 int factor
= this_cpu_read(sd_llc_size
);
4218 * Yeah, it's the switching-frequency, could means many wakee or
4219 * rapidly switch, use factor here will just help to automatically
4220 * adjust the loose-degree, so bigger node will lead to more pull.
4222 if (p
->wakee_flips
> factor
) {
4224 * wakee is somewhat hot, it needs certain amount of cpu
4225 * resource, so if waker is far more hot, prefer to leave
4228 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
4235 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4237 s64 this_load
, load
;
4238 int idx
, this_cpu
, prev_cpu
;
4239 unsigned long tl_per_task
;
4240 struct task_group
*tg
;
4241 unsigned long weight
;
4245 * If we wake multiple tasks be careful to not bounce
4246 * ourselves around too much.
4252 this_cpu
= smp_processor_id();
4253 prev_cpu
= task_cpu(p
);
4254 load
= source_load(prev_cpu
, idx
);
4255 this_load
= target_load(this_cpu
, idx
);
4258 * If sync wakeup then subtract the (maximum possible)
4259 * effect of the currently running task from the load
4260 * of the current CPU:
4263 tg
= task_group(current
);
4264 weight
= current
->se
.load
.weight
;
4266 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4267 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4271 weight
= p
->se
.load
.weight
;
4274 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4275 * due to the sync cause above having dropped this_load to 0, we'll
4276 * always have an imbalance, but there's really nothing you can do
4277 * about that, so that's good too.
4279 * Otherwise check if either cpus are near enough in load to allow this
4280 * task to be woken on this_cpu.
4282 if (this_load
> 0) {
4283 s64 this_eff_load
, prev_eff_load
;
4285 this_eff_load
= 100;
4286 this_eff_load
*= capacity_of(prev_cpu
);
4287 this_eff_load
*= this_load
+
4288 effective_load(tg
, this_cpu
, weight
, weight
);
4290 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4291 prev_eff_load
*= capacity_of(this_cpu
);
4292 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4294 balanced
= this_eff_load
<= prev_eff_load
;
4299 * If the currently running task will sleep within
4300 * a reasonable amount of time then attract this newly
4303 if (sync
&& balanced
)
4306 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4307 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
4310 (this_load
<= load
&&
4311 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
4313 * This domain has SD_WAKE_AFFINE and
4314 * p is cache cold in this domain, and
4315 * there is no bad imbalance.
4317 schedstat_inc(sd
, ttwu_move_affine
);
4318 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4326 * find_idlest_group finds and returns the least busy CPU group within the
4329 static struct sched_group
*
4330 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4331 int this_cpu
, int sd_flag
)
4333 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4334 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4335 int load_idx
= sd
->forkexec_idx
;
4336 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4338 if (sd_flag
& SD_BALANCE_WAKE
)
4339 load_idx
= sd
->wake_idx
;
4342 unsigned long load
, avg_load
;
4346 /* Skip over this group if it has no CPUs allowed */
4347 if (!cpumask_intersects(sched_group_cpus(group
),
4348 tsk_cpus_allowed(p
)))
4351 local_group
= cpumask_test_cpu(this_cpu
,
4352 sched_group_cpus(group
));
4354 /* Tally up the load of all CPUs in the group */
4357 for_each_cpu(i
, sched_group_cpus(group
)) {
4358 /* Bias balancing toward cpus of our domain */
4360 load
= source_load(i
, load_idx
);
4362 load
= target_load(i
, load_idx
);
4367 /* Adjust by relative CPU capacity of the group */
4368 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) / group
->sgc
->capacity
;
4371 this_load
= avg_load
;
4372 } else if (avg_load
< min_load
) {
4373 min_load
= avg_load
;
4376 } while (group
= group
->next
, group
!= sd
->groups
);
4378 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4384 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4387 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4389 unsigned long load
, min_load
= ULONG_MAX
;
4393 /* Traverse only the allowed CPUs */
4394 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4395 load
= weighted_cpuload(i
);
4397 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4407 * Try and locate an idle CPU in the sched_domain.
4409 static int select_idle_sibling(struct task_struct
*p
, int target
)
4411 struct sched_domain
*sd
;
4412 struct sched_group
*sg
;
4413 int i
= task_cpu(p
);
4415 if (idle_cpu(target
))
4419 * If the prevous cpu is cache affine and idle, don't be stupid.
4421 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4425 * Otherwise, iterate the domains and find an elegible idle cpu.
4427 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4428 for_each_lower_domain(sd
) {
4431 if (!cpumask_intersects(sched_group_cpus(sg
),
4432 tsk_cpus_allowed(p
)))
4435 for_each_cpu(i
, sched_group_cpus(sg
)) {
4436 if (i
== target
|| !idle_cpu(i
))
4440 target
= cpumask_first_and(sched_group_cpus(sg
),
4441 tsk_cpus_allowed(p
));
4445 } while (sg
!= sd
->groups
);
4452 * select_task_rq_fair: Select target runqueue for the waking task in domains
4453 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4454 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4456 * Balances load by selecting the idlest cpu in the idlest group, or under
4457 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4459 * Returns the target cpu number.
4461 * preempt must be disabled.
4464 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4466 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4467 int cpu
= smp_processor_id();
4469 int want_affine
= 0;
4470 int sync
= wake_flags
& WF_SYNC
;
4472 if (p
->nr_cpus_allowed
== 1)
4475 if (sd_flag
& SD_BALANCE_WAKE
) {
4476 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
4482 for_each_domain(cpu
, tmp
) {
4483 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4487 * If both cpu and prev_cpu are part of this domain,
4488 * cpu is a valid SD_WAKE_AFFINE target.
4490 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4491 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4496 if (tmp
->flags
& sd_flag
)
4500 if (affine_sd
&& cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4503 if (sd_flag
& SD_BALANCE_WAKE
) {
4504 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4509 struct sched_group
*group
;
4512 if (!(sd
->flags
& sd_flag
)) {
4517 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
4523 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4524 if (new_cpu
== -1 || new_cpu
== cpu
) {
4525 /* Now try balancing at a lower domain level of cpu */
4530 /* Now try balancing at a lower domain level of new_cpu */
4532 weight
= sd
->span_weight
;
4534 for_each_domain(cpu
, tmp
) {
4535 if (weight
<= tmp
->span_weight
)
4537 if (tmp
->flags
& sd_flag
)
4540 /* while loop will break here if sd == NULL */
4549 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4550 * cfs_rq_of(p) references at time of call are still valid and identify the
4551 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4552 * other assumptions, including the state of rq->lock, should be made.
4555 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4557 struct sched_entity
*se
= &p
->se
;
4558 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4561 * Load tracking: accumulate removed load so that it can be processed
4562 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4563 * to blocked load iff they have a positive decay-count. It can never
4564 * be negative here since on-rq tasks have decay-count == 0.
4566 if (se
->avg
.decay_count
) {
4567 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4568 atomic_long_add(se
->avg
.load_avg_contrib
,
4569 &cfs_rq
->removed_load
);
4572 /* We have migrated, no longer consider this task hot */
4575 #endif /* CONFIG_SMP */
4577 static unsigned long
4578 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4580 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4583 * Since its curr running now, convert the gran from real-time
4584 * to virtual-time in his units.
4586 * By using 'se' instead of 'curr' we penalize light tasks, so
4587 * they get preempted easier. That is, if 'se' < 'curr' then
4588 * the resulting gran will be larger, therefore penalizing the
4589 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4590 * be smaller, again penalizing the lighter task.
4592 * This is especially important for buddies when the leftmost
4593 * task is higher priority than the buddy.
4595 return calc_delta_fair(gran
, se
);
4599 * Should 'se' preempt 'curr'.
4613 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4615 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4620 gran
= wakeup_gran(curr
, se
);
4627 static void set_last_buddy(struct sched_entity
*se
)
4629 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4632 for_each_sched_entity(se
)
4633 cfs_rq_of(se
)->last
= se
;
4636 static void set_next_buddy(struct sched_entity
*se
)
4638 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4641 for_each_sched_entity(se
)
4642 cfs_rq_of(se
)->next
= se
;
4645 static void set_skip_buddy(struct sched_entity
*se
)
4647 for_each_sched_entity(se
)
4648 cfs_rq_of(se
)->skip
= se
;
4652 * Preempt the current task with a newly woken task if needed:
4654 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4656 struct task_struct
*curr
= rq
->curr
;
4657 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4658 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4659 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4660 int next_buddy_marked
= 0;
4662 if (unlikely(se
== pse
))
4666 * This is possible from callers such as move_task(), in which we
4667 * unconditionally check_prempt_curr() after an enqueue (which may have
4668 * lead to a throttle). This both saves work and prevents false
4669 * next-buddy nomination below.
4671 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4674 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4675 set_next_buddy(pse
);
4676 next_buddy_marked
= 1;
4680 * We can come here with TIF_NEED_RESCHED already set from new task
4683 * Note: this also catches the edge-case of curr being in a throttled
4684 * group (e.g. via set_curr_task), since update_curr() (in the
4685 * enqueue of curr) will have resulted in resched being set. This
4686 * prevents us from potentially nominating it as a false LAST_BUDDY
4689 if (test_tsk_need_resched(curr
))
4692 /* Idle tasks are by definition preempted by non-idle tasks. */
4693 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4694 likely(p
->policy
!= SCHED_IDLE
))
4698 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4699 * is driven by the tick):
4701 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4704 find_matching_se(&se
, &pse
);
4705 update_curr(cfs_rq_of(se
));
4707 if (wakeup_preempt_entity(se
, pse
) == 1) {
4709 * Bias pick_next to pick the sched entity that is
4710 * triggering this preemption.
4712 if (!next_buddy_marked
)
4713 set_next_buddy(pse
);
4722 * Only set the backward buddy when the current task is still
4723 * on the rq. This can happen when a wakeup gets interleaved
4724 * with schedule on the ->pre_schedule() or idle_balance()
4725 * point, either of which can * drop the rq lock.
4727 * Also, during early boot the idle thread is in the fair class,
4728 * for obvious reasons its a bad idea to schedule back to it.
4730 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4733 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
4737 static struct task_struct
*
4738 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4740 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
4741 struct sched_entity
*se
;
4742 struct task_struct
*p
;
4746 #ifdef CONFIG_FAIR_GROUP_SCHED
4747 if (!cfs_rq
->nr_running
)
4750 if (prev
->sched_class
!= &fair_sched_class
)
4754 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4755 * likely that a next task is from the same cgroup as the current.
4757 * Therefore attempt to avoid putting and setting the entire cgroup
4758 * hierarchy, only change the part that actually changes.
4762 struct sched_entity
*curr
= cfs_rq
->curr
;
4765 * Since we got here without doing put_prev_entity() we also
4766 * have to consider cfs_rq->curr. If it is still a runnable
4767 * entity, update_curr() will update its vruntime, otherwise
4768 * forget we've ever seen it.
4770 if (curr
&& curr
->on_rq
)
4771 update_curr(cfs_rq
);
4776 * This call to check_cfs_rq_runtime() will do the throttle and
4777 * dequeue its entity in the parent(s). Therefore the 'simple'
4778 * nr_running test will indeed be correct.
4780 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
4783 se
= pick_next_entity(cfs_rq
, curr
);
4784 cfs_rq
= group_cfs_rq(se
);
4790 * Since we haven't yet done put_prev_entity and if the selected task
4791 * is a different task than we started out with, try and touch the
4792 * least amount of cfs_rqs.
4795 struct sched_entity
*pse
= &prev
->se
;
4797 while (!(cfs_rq
= is_same_group(se
, pse
))) {
4798 int se_depth
= se
->depth
;
4799 int pse_depth
= pse
->depth
;
4801 if (se_depth
<= pse_depth
) {
4802 put_prev_entity(cfs_rq_of(pse
), pse
);
4803 pse
= parent_entity(pse
);
4805 if (se_depth
>= pse_depth
) {
4806 set_next_entity(cfs_rq_of(se
), se
);
4807 se
= parent_entity(se
);
4811 put_prev_entity(cfs_rq
, pse
);
4812 set_next_entity(cfs_rq
, se
);
4815 if (hrtick_enabled(rq
))
4816 hrtick_start_fair(rq
, p
);
4823 if (!cfs_rq
->nr_running
)
4826 put_prev_task(rq
, prev
);
4829 se
= pick_next_entity(cfs_rq
, NULL
);
4830 set_next_entity(cfs_rq
, se
);
4831 cfs_rq
= group_cfs_rq(se
);
4836 if (hrtick_enabled(rq
))
4837 hrtick_start_fair(rq
, p
);
4842 new_tasks
= idle_balance(rq
);
4844 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4845 * possible for any higher priority task to appear. In that case we
4846 * must re-start the pick_next_entity() loop.
4858 * Account for a descheduled task:
4860 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4862 struct sched_entity
*se
= &prev
->se
;
4863 struct cfs_rq
*cfs_rq
;
4865 for_each_sched_entity(se
) {
4866 cfs_rq
= cfs_rq_of(se
);
4867 put_prev_entity(cfs_rq
, se
);
4872 * sched_yield() is very simple
4874 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4876 static void yield_task_fair(struct rq
*rq
)
4878 struct task_struct
*curr
= rq
->curr
;
4879 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4880 struct sched_entity
*se
= &curr
->se
;
4883 * Are we the only task in the tree?
4885 if (unlikely(rq
->nr_running
== 1))
4888 clear_buddies(cfs_rq
, se
);
4890 if (curr
->policy
!= SCHED_BATCH
) {
4891 update_rq_clock(rq
);
4893 * Update run-time statistics of the 'current'.
4895 update_curr(cfs_rq
);
4897 * Tell update_rq_clock() that we've just updated,
4898 * so we don't do microscopic update in schedule()
4899 * and double the fastpath cost.
4901 rq
->skip_clock_update
= 1;
4907 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
4909 struct sched_entity
*se
= &p
->se
;
4911 /* throttled hierarchies are not runnable */
4912 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
4915 /* Tell the scheduler that we'd really like pse to run next. */
4918 yield_task_fair(rq
);
4924 /**************************************************
4925 * Fair scheduling class load-balancing methods.
4929 * The purpose of load-balancing is to achieve the same basic fairness the
4930 * per-cpu scheduler provides, namely provide a proportional amount of compute
4931 * time to each task. This is expressed in the following equation:
4933 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4935 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4936 * W_i,0 is defined as:
4938 * W_i,0 = \Sum_j w_i,j (2)
4940 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4941 * is derived from the nice value as per prio_to_weight[].
4943 * The weight average is an exponential decay average of the instantaneous
4946 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4948 * C_i is the compute capacity of cpu i, typically it is the
4949 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4950 * can also include other factors [XXX].
4952 * To achieve this balance we define a measure of imbalance which follows
4953 * directly from (1):
4955 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
4957 * We them move tasks around to minimize the imbalance. In the continuous
4958 * function space it is obvious this converges, in the discrete case we get
4959 * a few fun cases generally called infeasible weight scenarios.
4962 * - infeasible weights;
4963 * - local vs global optima in the discrete case. ]
4968 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4969 * for all i,j solution, we create a tree of cpus that follows the hardware
4970 * topology where each level pairs two lower groups (or better). This results
4971 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4972 * tree to only the first of the previous level and we decrease the frequency
4973 * of load-balance at each level inv. proportional to the number of cpus in
4979 * \Sum { --- * --- * 2^i } = O(n) (5)
4981 * `- size of each group
4982 * | | `- number of cpus doing load-balance
4984 * `- sum over all levels
4986 * Coupled with a limit on how many tasks we can migrate every balance pass,
4987 * this makes (5) the runtime complexity of the balancer.
4989 * An important property here is that each CPU is still (indirectly) connected
4990 * to every other cpu in at most O(log n) steps:
4992 * The adjacency matrix of the resulting graph is given by:
4995 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4998 * And you'll find that:
5000 * A^(log_2 n)_i,j != 0 for all i,j (7)
5002 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5003 * The task movement gives a factor of O(m), giving a convergence complexity
5006 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5011 * In order to avoid CPUs going idle while there's still work to do, new idle
5012 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5013 * tree itself instead of relying on other CPUs to bring it work.
5015 * This adds some complexity to both (5) and (8) but it reduces the total idle
5023 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5026 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5031 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5033 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5035 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5038 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5039 * rewrite all of this once again.]
5042 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
5044 enum fbq_type
{ regular
, remote
, all
};
5046 #define LBF_ALL_PINNED 0x01
5047 #define LBF_NEED_BREAK 0x02
5048 #define LBF_DST_PINNED 0x04
5049 #define LBF_SOME_PINNED 0x08
5052 struct sched_domain
*sd
;
5060 struct cpumask
*dst_grpmask
;
5062 enum cpu_idle_type idle
;
5064 /* The set of CPUs under consideration for load-balancing */
5065 struct cpumask
*cpus
;
5070 unsigned int loop_break
;
5071 unsigned int loop_max
;
5073 enum fbq_type fbq_type
;
5077 * move_task - move a task from one runqueue to another runqueue.
5078 * Both runqueues must be locked.
5080 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
5082 deactivate_task(env
->src_rq
, p
, 0);
5083 set_task_cpu(p
, env
->dst_cpu
);
5084 activate_task(env
->dst_rq
, p
, 0);
5085 check_preempt_curr(env
->dst_rq
, p
, 0);
5089 * Is this task likely cache-hot:
5091 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
5095 if (p
->sched_class
!= &fair_sched_class
)
5098 if (unlikely(p
->policy
== SCHED_IDLE
))
5102 * Buddy candidates are cache hot:
5104 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
5105 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
5106 &p
->se
== cfs_rq_of(&p
->se
)->last
))
5109 if (sysctl_sched_migration_cost
== -1)
5111 if (sysctl_sched_migration_cost
== 0)
5114 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
5116 return delta
< (s64
)sysctl_sched_migration_cost
;
5119 #ifdef CONFIG_NUMA_BALANCING
5120 /* Returns true if the destination node has incurred more faults */
5121 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
5123 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5124 int src_nid
, dst_nid
;
5126 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults_memory
||
5127 !(env
->sd
->flags
& SD_NUMA
)) {
5131 src_nid
= cpu_to_node(env
->src_cpu
);
5132 dst_nid
= cpu_to_node(env
->dst_cpu
);
5134 if (src_nid
== dst_nid
)
5138 /* Task is already in the group's interleave set. */
5139 if (node_isset(src_nid
, numa_group
->active_nodes
))
5142 /* Task is moving into the group's interleave set. */
5143 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5146 return group_faults(p
, dst_nid
) > group_faults(p
, src_nid
);
5149 /* Encourage migration to the preferred node. */
5150 if (dst_nid
== p
->numa_preferred_nid
)
5153 return task_faults(p
, dst_nid
) > task_faults(p
, src_nid
);
5157 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
5159 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5160 int src_nid
, dst_nid
;
5162 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
5165 if (!p
->numa_faults_memory
|| !(env
->sd
->flags
& SD_NUMA
))
5168 src_nid
= cpu_to_node(env
->src_cpu
);
5169 dst_nid
= cpu_to_node(env
->dst_cpu
);
5171 if (src_nid
== dst_nid
)
5175 /* Task is moving within/into the group's interleave set. */
5176 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5179 /* Task is moving out of the group's interleave set. */
5180 if (node_isset(src_nid
, numa_group
->active_nodes
))
5183 return group_faults(p
, dst_nid
) < group_faults(p
, src_nid
);
5186 /* Migrating away from the preferred node is always bad. */
5187 if (src_nid
== p
->numa_preferred_nid
)
5190 return task_faults(p
, dst_nid
) < task_faults(p
, src_nid
);
5194 static inline bool migrate_improves_locality(struct task_struct
*p
,
5200 static inline bool migrate_degrades_locality(struct task_struct
*p
,
5208 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5211 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
5213 int tsk_cache_hot
= 0;
5215 * We do not migrate tasks that are:
5216 * 1) throttled_lb_pair, or
5217 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5218 * 3) running (obviously), or
5219 * 4) are cache-hot on their current CPU.
5221 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
5224 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
5227 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
5229 env
->flags
|= LBF_SOME_PINNED
;
5232 * Remember if this task can be migrated to any other cpu in
5233 * our sched_group. We may want to revisit it if we couldn't
5234 * meet load balance goals by pulling other tasks on src_cpu.
5236 * Also avoid computing new_dst_cpu if we have already computed
5237 * one in current iteration.
5239 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
5242 /* Prevent to re-select dst_cpu via env's cpus */
5243 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
5244 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
5245 env
->flags
|= LBF_DST_PINNED
;
5246 env
->new_dst_cpu
= cpu
;
5254 /* Record that we found atleast one task that could run on dst_cpu */
5255 env
->flags
&= ~LBF_ALL_PINNED
;
5257 if (task_running(env
->src_rq
, p
)) {
5258 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
5263 * Aggressive migration if:
5264 * 1) destination numa is preferred
5265 * 2) task is cache cold, or
5266 * 3) too many balance attempts have failed.
5268 tsk_cache_hot
= task_hot(p
, env
);
5270 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
5272 if (migrate_improves_locality(p
, env
)) {
5273 #ifdef CONFIG_SCHEDSTATS
5274 if (tsk_cache_hot
) {
5275 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5276 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5282 if (!tsk_cache_hot
||
5283 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
5285 if (tsk_cache_hot
) {
5286 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5287 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5293 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
5298 * move_one_task tries to move exactly one task from busiest to this_rq, as
5299 * part of active balancing operations within "domain".
5300 * Returns 1 if successful and 0 otherwise.
5302 * Called with both runqueues locked.
5304 static int move_one_task(struct lb_env
*env
)
5306 struct task_struct
*p
, *n
;
5308 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
5309 if (!can_migrate_task(p
, env
))
5314 * Right now, this is only the second place move_task()
5315 * is called, so we can safely collect move_task()
5316 * stats here rather than inside move_task().
5318 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
5324 static const unsigned int sched_nr_migrate_break
= 32;
5327 * move_tasks tries to move up to imbalance weighted load from busiest to
5328 * this_rq, as part of a balancing operation within domain "sd".
5329 * Returns 1 if successful and 0 otherwise.
5331 * Called with both runqueues locked.
5333 static int move_tasks(struct lb_env
*env
)
5335 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
5336 struct task_struct
*p
;
5340 if (env
->imbalance
<= 0)
5343 while (!list_empty(tasks
)) {
5344 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5347 /* We've more or less seen every task there is, call it quits */
5348 if (env
->loop
> env
->loop_max
)
5351 /* take a breather every nr_migrate tasks */
5352 if (env
->loop
> env
->loop_break
) {
5353 env
->loop_break
+= sched_nr_migrate_break
;
5354 env
->flags
|= LBF_NEED_BREAK
;
5358 if (!can_migrate_task(p
, env
))
5361 load
= task_h_load(p
);
5363 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5366 if ((load
/ 2) > env
->imbalance
)
5371 env
->imbalance
-= load
;
5373 #ifdef CONFIG_PREEMPT
5375 * NEWIDLE balancing is a source of latency, so preemptible
5376 * kernels will stop after the first task is pulled to minimize
5377 * the critical section.
5379 if (env
->idle
== CPU_NEWLY_IDLE
)
5384 * We only want to steal up to the prescribed amount of
5387 if (env
->imbalance
<= 0)
5392 list_move_tail(&p
->se
.group_node
, tasks
);
5396 * Right now, this is one of only two places move_task() is called,
5397 * so we can safely collect move_task() stats here rather than
5398 * inside move_task().
5400 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
5405 #ifdef CONFIG_FAIR_GROUP_SCHED
5407 * update tg->load_weight by folding this cpu's load_avg
5409 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5411 struct sched_entity
*se
= tg
->se
[cpu
];
5412 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5414 /* throttled entities do not contribute to load */
5415 if (throttled_hierarchy(cfs_rq
))
5418 update_cfs_rq_blocked_load(cfs_rq
, 1);
5421 update_entity_load_avg(se
, 1);
5423 * We pivot on our runnable average having decayed to zero for
5424 * list removal. This generally implies that all our children
5425 * have also been removed (modulo rounding error or bandwidth
5426 * control); however, such cases are rare and we can fix these
5429 * TODO: fix up out-of-order children on enqueue.
5431 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5432 list_del_leaf_cfs_rq(cfs_rq
);
5434 struct rq
*rq
= rq_of(cfs_rq
);
5435 update_rq_runnable_avg(rq
, rq
->nr_running
);
5439 static void update_blocked_averages(int cpu
)
5441 struct rq
*rq
= cpu_rq(cpu
);
5442 struct cfs_rq
*cfs_rq
;
5443 unsigned long flags
;
5445 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5446 update_rq_clock(rq
);
5448 * Iterates the task_group tree in a bottom up fashion, see
5449 * list_add_leaf_cfs_rq() for details.
5451 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5453 * Note: We may want to consider periodically releasing
5454 * rq->lock about these updates so that creating many task
5455 * groups does not result in continually extending hold time.
5457 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5460 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5464 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5465 * This needs to be done in a top-down fashion because the load of a child
5466 * group is a fraction of its parents load.
5468 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5470 struct rq
*rq
= rq_of(cfs_rq
);
5471 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5472 unsigned long now
= jiffies
;
5475 if (cfs_rq
->last_h_load_update
== now
)
5478 cfs_rq
->h_load_next
= NULL
;
5479 for_each_sched_entity(se
) {
5480 cfs_rq
= cfs_rq_of(se
);
5481 cfs_rq
->h_load_next
= se
;
5482 if (cfs_rq
->last_h_load_update
== now
)
5487 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5488 cfs_rq
->last_h_load_update
= now
;
5491 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5492 load
= cfs_rq
->h_load
;
5493 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5494 cfs_rq
->runnable_load_avg
+ 1);
5495 cfs_rq
= group_cfs_rq(se
);
5496 cfs_rq
->h_load
= load
;
5497 cfs_rq
->last_h_load_update
= now
;
5501 static unsigned long task_h_load(struct task_struct
*p
)
5503 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5505 update_cfs_rq_h_load(cfs_rq
);
5506 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5507 cfs_rq
->runnable_load_avg
+ 1);
5510 static inline void update_blocked_averages(int cpu
)
5514 static unsigned long task_h_load(struct task_struct
*p
)
5516 return p
->se
.avg
.load_avg_contrib
;
5520 /********** Helpers for find_busiest_group ************************/
5522 * sg_lb_stats - stats of a sched_group required for load_balancing
5524 struct sg_lb_stats
{
5525 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5526 unsigned long group_load
; /* Total load over the CPUs of the group */
5527 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5528 unsigned long load_per_task
;
5529 unsigned long group_capacity
;
5530 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5531 unsigned int group_capacity_factor
;
5532 unsigned int idle_cpus
;
5533 unsigned int group_weight
;
5534 int group_imb
; /* Is there an imbalance in the group ? */
5535 int group_has_free_capacity
;
5536 #ifdef CONFIG_NUMA_BALANCING
5537 unsigned int nr_numa_running
;
5538 unsigned int nr_preferred_running
;
5543 * sd_lb_stats - Structure to store the statistics of a sched_domain
5544 * during load balancing.
5546 struct sd_lb_stats
{
5547 struct sched_group
*busiest
; /* Busiest group in this sd */
5548 struct sched_group
*local
; /* Local group in this sd */
5549 unsigned long total_load
; /* Total load of all groups in sd */
5550 unsigned long total_capacity
; /* Total capacity of all groups in sd */
5551 unsigned long avg_load
; /* Average load across all groups in sd */
5553 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5554 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5557 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5560 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5561 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5562 * We must however clear busiest_stat::avg_load because
5563 * update_sd_pick_busiest() reads this before assignment.
5565 *sds
= (struct sd_lb_stats
){
5569 .total_capacity
= 0UL,
5577 * get_sd_load_idx - Obtain the load index for a given sched domain.
5578 * @sd: The sched_domain whose load_idx is to be obtained.
5579 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5581 * Return: The load index.
5583 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5584 enum cpu_idle_type idle
)
5590 load_idx
= sd
->busy_idx
;
5593 case CPU_NEWLY_IDLE
:
5594 load_idx
= sd
->newidle_idx
;
5597 load_idx
= sd
->idle_idx
;
5604 static unsigned long default_scale_capacity(struct sched_domain
*sd
, int cpu
)
5606 return SCHED_CAPACITY_SCALE
;
5609 unsigned long __weak
arch_scale_freq_capacity(struct sched_domain
*sd
, int cpu
)
5611 return default_scale_capacity(sd
, cpu
);
5614 static unsigned long default_scale_smt_capacity(struct sched_domain
*sd
, int cpu
)
5616 unsigned long weight
= sd
->span_weight
;
5617 unsigned long smt_gain
= sd
->smt_gain
;
5624 unsigned long __weak
arch_scale_smt_capacity(struct sched_domain
*sd
, int cpu
)
5626 return default_scale_smt_capacity(sd
, cpu
);
5629 static unsigned long scale_rt_capacity(int cpu
)
5631 struct rq
*rq
= cpu_rq(cpu
);
5632 u64 total
, available
, age_stamp
, avg
;
5636 * Since we're reading these variables without serialization make sure
5637 * we read them once before doing sanity checks on them.
5639 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5640 avg
= ACCESS_ONCE(rq
->rt_avg
);
5642 delta
= rq_clock(rq
) - age_stamp
;
5643 if (unlikely(delta
< 0))
5646 total
= sched_avg_period() + delta
;
5648 if (unlikely(total
< avg
)) {
5649 /* Ensures that capacity won't end up being negative */
5652 available
= total
- avg
;
5655 if (unlikely((s64
)total
< SCHED_CAPACITY_SCALE
))
5656 total
= SCHED_CAPACITY_SCALE
;
5658 total
>>= SCHED_CAPACITY_SHIFT
;
5660 return div_u64(available
, total
);
5663 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5665 unsigned long weight
= sd
->span_weight
;
5666 unsigned long capacity
= SCHED_CAPACITY_SCALE
;
5667 struct sched_group
*sdg
= sd
->groups
;
5669 if ((sd
->flags
& SD_SHARE_CPUCAPACITY
) && weight
> 1) {
5670 if (sched_feat(ARCH_CAPACITY
))
5671 capacity
*= arch_scale_smt_capacity(sd
, cpu
);
5673 capacity
*= default_scale_smt_capacity(sd
, cpu
);
5675 capacity
>>= SCHED_CAPACITY_SHIFT
;
5678 sdg
->sgc
->capacity_orig
= capacity
;
5680 if (sched_feat(ARCH_CAPACITY
))
5681 capacity
*= arch_scale_freq_capacity(sd
, cpu
);
5683 capacity
*= default_scale_capacity(sd
, cpu
);
5685 capacity
>>= SCHED_CAPACITY_SHIFT
;
5687 capacity
*= scale_rt_capacity(cpu
);
5688 capacity
>>= SCHED_CAPACITY_SHIFT
;
5693 cpu_rq(cpu
)->cpu_capacity
= capacity
;
5694 sdg
->sgc
->capacity
= capacity
;
5697 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
5699 struct sched_domain
*child
= sd
->child
;
5700 struct sched_group
*group
, *sdg
= sd
->groups
;
5701 unsigned long capacity
, capacity_orig
;
5702 unsigned long interval
;
5704 interval
= msecs_to_jiffies(sd
->balance_interval
);
5705 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5706 sdg
->sgc
->next_update
= jiffies
+ interval
;
5709 update_cpu_capacity(sd
, cpu
);
5713 capacity_orig
= capacity
= 0;
5715 if (child
->flags
& SD_OVERLAP
) {
5717 * SD_OVERLAP domains cannot assume that child groups
5718 * span the current group.
5721 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
5722 struct sched_group_capacity
*sgc
;
5723 struct rq
*rq
= cpu_rq(cpu
);
5726 * build_sched_domains() -> init_sched_groups_capacity()
5727 * gets here before we've attached the domains to the
5730 * Use capacity_of(), which is set irrespective of domains
5731 * in update_cpu_capacity().
5733 * This avoids capacity/capacity_orig from being 0 and
5734 * causing divide-by-zero issues on boot.
5736 * Runtime updates will correct capacity_orig.
5738 if (unlikely(!rq
->sd
)) {
5739 capacity_orig
+= capacity_of(cpu
);
5740 capacity
+= capacity_of(cpu
);
5744 sgc
= rq
->sd
->groups
->sgc
;
5745 capacity_orig
+= sgc
->capacity_orig
;
5746 capacity
+= sgc
->capacity
;
5750 * !SD_OVERLAP domains can assume that child groups
5751 * span the current group.
5754 group
= child
->groups
;
5756 capacity_orig
+= group
->sgc
->capacity_orig
;
5757 capacity
+= group
->sgc
->capacity
;
5758 group
= group
->next
;
5759 } while (group
!= child
->groups
);
5762 sdg
->sgc
->capacity_orig
= capacity_orig
;
5763 sdg
->sgc
->capacity
= capacity
;
5767 * Try and fix up capacity for tiny siblings, this is needed when
5768 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5769 * which on its own isn't powerful enough.
5771 * See update_sd_pick_busiest() and check_asym_packing().
5774 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
5777 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
5779 if (!(sd
->flags
& SD_SHARE_CPUCAPACITY
))
5783 * If ~90% of the cpu_capacity is still there, we're good.
5785 if (group
->sgc
->capacity
* 32 > group
->sgc
->capacity_orig
* 29)
5792 * Group imbalance indicates (and tries to solve) the problem where balancing
5793 * groups is inadequate due to tsk_cpus_allowed() constraints.
5795 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5796 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5799 * { 0 1 2 3 } { 4 5 6 7 }
5802 * If we were to balance group-wise we'd place two tasks in the first group and
5803 * two tasks in the second group. Clearly this is undesired as it will overload
5804 * cpu 3 and leave one of the cpus in the second group unused.
5806 * The current solution to this issue is detecting the skew in the first group
5807 * by noticing the lower domain failed to reach balance and had difficulty
5808 * moving tasks due to affinity constraints.
5810 * When this is so detected; this group becomes a candidate for busiest; see
5811 * update_sd_pick_busiest(). And calculate_imbalance() and
5812 * find_busiest_group() avoid some of the usual balance conditions to allow it
5813 * to create an effective group imbalance.
5815 * This is a somewhat tricky proposition since the next run might not find the
5816 * group imbalance and decide the groups need to be balanced again. A most
5817 * subtle and fragile situation.
5820 static inline int sg_imbalanced(struct sched_group
*group
)
5822 return group
->sgc
->imbalance
;
5826 * Compute the group capacity factor.
5828 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
5829 * first dividing out the smt factor and computing the actual number of cores
5830 * and limit unit capacity with that.
5832 static inline int sg_capacity_factor(struct lb_env
*env
, struct sched_group
*group
)
5834 unsigned int capacity_factor
, smt
, cpus
;
5835 unsigned int capacity
, capacity_orig
;
5837 capacity
= group
->sgc
->capacity
;
5838 capacity_orig
= group
->sgc
->capacity_orig
;
5839 cpus
= group
->group_weight
;
5841 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
5842 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, capacity_orig
);
5843 capacity_factor
= cpus
/ smt
; /* cores */
5845 capacity_factor
= min_t(unsigned,
5846 capacity_factor
, DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
));
5847 if (!capacity_factor
)
5848 capacity_factor
= fix_small_capacity(env
->sd
, group
);
5850 return capacity_factor
;
5854 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5855 * @env: The load balancing environment.
5856 * @group: sched_group whose statistics are to be updated.
5857 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5858 * @local_group: Does group contain this_cpu.
5859 * @sgs: variable to hold the statistics for this group.
5861 static inline void update_sg_lb_stats(struct lb_env
*env
,
5862 struct sched_group
*group
, int load_idx
,
5863 int local_group
, struct sg_lb_stats
*sgs
,
5869 memset(sgs
, 0, sizeof(*sgs
));
5871 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5872 struct rq
*rq
= cpu_rq(i
);
5874 /* Bias balancing toward cpus of our domain */
5876 load
= target_load(i
, load_idx
);
5878 load
= source_load(i
, load_idx
);
5880 sgs
->group_load
+= load
;
5881 sgs
->sum_nr_running
+= rq
->nr_running
;
5883 if (rq
->nr_running
> 1)
5886 #ifdef CONFIG_NUMA_BALANCING
5887 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
5888 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
5890 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
5895 /* Adjust by relative CPU capacity of the group */
5896 sgs
->group_capacity
= group
->sgc
->capacity
;
5897 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
5899 if (sgs
->sum_nr_running
)
5900 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
5902 sgs
->group_weight
= group
->group_weight
;
5904 sgs
->group_imb
= sg_imbalanced(group
);
5905 sgs
->group_capacity_factor
= sg_capacity_factor(env
, group
);
5907 if (sgs
->group_capacity_factor
> sgs
->sum_nr_running
)
5908 sgs
->group_has_free_capacity
= 1;
5912 * update_sd_pick_busiest - return 1 on busiest group
5913 * @env: The load balancing environment.
5914 * @sds: sched_domain statistics
5915 * @sg: sched_group candidate to be checked for being the busiest
5916 * @sgs: sched_group statistics
5918 * Determine if @sg is a busier group than the previously selected
5921 * Return: %true if @sg is a busier group than the previously selected
5922 * busiest group. %false otherwise.
5924 static bool update_sd_pick_busiest(struct lb_env
*env
,
5925 struct sd_lb_stats
*sds
,
5926 struct sched_group
*sg
,
5927 struct sg_lb_stats
*sgs
)
5929 if (sgs
->avg_load
<= sds
->busiest_stat
.avg_load
)
5932 if (sgs
->sum_nr_running
> sgs
->group_capacity_factor
)
5939 * ASYM_PACKING needs to move all the work to the lowest
5940 * numbered CPUs in the group, therefore mark all groups
5941 * higher than ourself as busy.
5943 if ((env
->sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
5944 env
->dst_cpu
< group_first_cpu(sg
)) {
5948 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
5955 #ifdef CONFIG_NUMA_BALANCING
5956 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
5958 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
5960 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
5965 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
5967 if (rq
->nr_running
> rq
->nr_numa_running
)
5969 if (rq
->nr_running
> rq
->nr_preferred_running
)
5974 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
5979 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
5983 #endif /* CONFIG_NUMA_BALANCING */
5986 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5987 * @env: The load balancing environment.
5988 * @sds: variable to hold the statistics for this sched_domain.
5990 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5992 struct sched_domain
*child
= env
->sd
->child
;
5993 struct sched_group
*sg
= env
->sd
->groups
;
5994 struct sg_lb_stats tmp_sgs
;
5995 int load_idx
, prefer_sibling
= 0;
5996 bool overload
= false;
5998 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
6001 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
6004 struct sg_lb_stats
*sgs
= &tmp_sgs
;
6007 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
6010 sgs
= &sds
->local_stat
;
6012 if (env
->idle
!= CPU_NEWLY_IDLE
||
6013 time_after_eq(jiffies
, sg
->sgc
->next_update
))
6014 update_group_capacity(env
->sd
, env
->dst_cpu
);
6017 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
);
6053 if (!env
->sd
->parent
) {
6054 /* update overload indicator if we are at root domain */
6055 if (env
->dst_rq
->rd
->overload
!= overload
)
6056 env
->dst_rq
->rd
->overload
= overload
;
6062 * check_asym_packing - Check to see if the group is packed into the
6065 * This is primarily intended to used at the sibling level. Some
6066 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6067 * case of POWER7, it can move to lower SMT modes only when higher
6068 * threads are idle. When in lower SMT modes, the threads will
6069 * perform better since they share less core resources. Hence when we
6070 * have idle threads, we want them to be the higher ones.
6072 * This packing function is run on idle threads. It checks to see if
6073 * the busiest CPU in this domain (core in the P7 case) has a higher
6074 * CPU number than the packing function is being run on. Here we are
6075 * assuming lower CPU number will be equivalent to lower a SMT thread
6078 * Return: 1 when packing is required and a task should be moved to
6079 * this CPU. The amount of the imbalance is returned in *imbalance.
6081 * @env: The load balancing environment.
6082 * @sds: Statistics of the sched_domain which is to be packed
6084 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6088 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6094 busiest_cpu
= group_first_cpu(sds
->busiest
);
6095 if (env
->dst_cpu
> busiest_cpu
)
6098 env
->imbalance
= DIV_ROUND_CLOSEST(
6099 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
6100 SCHED_CAPACITY_SCALE
);
6106 * fix_small_imbalance - Calculate the minor imbalance that exists
6107 * amongst the groups of a sched_domain, during
6109 * @env: The load balancing environment.
6110 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6113 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6115 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
6116 unsigned int imbn
= 2;
6117 unsigned long scaled_busy_load_per_task
;
6118 struct sg_lb_stats
*local
, *busiest
;
6120 local
= &sds
->local_stat
;
6121 busiest
= &sds
->busiest_stat
;
6123 if (!local
->sum_nr_running
)
6124 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
6125 else if (busiest
->load_per_task
> local
->load_per_task
)
6128 scaled_busy_load_per_task
=
6129 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6130 busiest
->group_capacity
;
6132 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
6133 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
6134 env
->imbalance
= busiest
->load_per_task
;
6139 * OK, we don't have enough imbalance to justify moving tasks,
6140 * however we may be able to increase total CPU capacity used by
6144 capa_now
+= busiest
->group_capacity
*
6145 min(busiest
->load_per_task
, busiest
->avg_load
);
6146 capa_now
+= local
->group_capacity
*
6147 min(local
->load_per_task
, local
->avg_load
);
6148 capa_now
/= SCHED_CAPACITY_SCALE
;
6150 /* Amount of load we'd subtract */
6151 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
6152 capa_move
+= busiest
->group_capacity
*
6153 min(busiest
->load_per_task
,
6154 busiest
->avg_load
- scaled_busy_load_per_task
);
6157 /* Amount of load we'd add */
6158 if (busiest
->avg_load
* busiest
->group_capacity
<
6159 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
6160 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
6161 local
->group_capacity
;
6163 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6164 local
->group_capacity
;
6166 capa_move
+= local
->group_capacity
*
6167 min(local
->load_per_task
, local
->avg_load
+ tmp
);
6168 capa_move
/= SCHED_CAPACITY_SCALE
;
6170 /* Move if we gain throughput */
6171 if (capa_move
> capa_now
)
6172 env
->imbalance
= busiest
->load_per_task
;
6176 * calculate_imbalance - Calculate the amount of imbalance present within the
6177 * groups of a given sched_domain during load balance.
6178 * @env: load balance environment
6179 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6181 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6183 unsigned long max_pull
, load_above_capacity
= ~0UL;
6184 struct sg_lb_stats
*local
, *busiest
;
6186 local
= &sds
->local_stat
;
6187 busiest
= &sds
->busiest_stat
;
6189 if (busiest
->group_imb
) {
6191 * In the group_imb case we cannot rely on group-wide averages
6192 * to ensure cpu-load equilibrium, look at wider averages. XXX
6194 busiest
->load_per_task
=
6195 min(busiest
->load_per_task
, sds
->avg_load
);
6199 * In the presence of smp nice balancing, certain scenarios can have
6200 * max load less than avg load(as we skip the groups at or below
6201 * its cpu_capacity, while calculating max_load..)
6203 if (busiest
->avg_load
<= sds
->avg_load
||
6204 local
->avg_load
>= sds
->avg_load
) {
6206 return fix_small_imbalance(env
, sds
);
6209 if (!busiest
->group_imb
) {
6211 * Don't want to pull so many tasks that a group would go idle.
6212 * Except of course for the group_imb case, since then we might
6213 * have to drop below capacity to reach cpu-load equilibrium.
6215 load_above_capacity
=
6216 (busiest
->sum_nr_running
- busiest
->group_capacity_factor
);
6218 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_CAPACITY_SCALE
);
6219 load_above_capacity
/= busiest
->group_capacity
;
6223 * We're trying to get all the cpus to the average_load, so we don't
6224 * want to push ourselves above the average load, nor do we wish to
6225 * reduce the max loaded cpu below the average load. At the same time,
6226 * we also don't want to reduce the group load below the group capacity
6227 * (so that we can implement power-savings policies etc). Thus we look
6228 * for the minimum possible imbalance.
6230 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
6232 /* How much load to actually move to equalise the imbalance */
6233 env
->imbalance
= min(
6234 max_pull
* busiest
->group_capacity
,
6235 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
6236 ) / SCHED_CAPACITY_SCALE
;
6239 * if *imbalance is less than the average load per runnable task
6240 * there is no guarantee that any tasks will be moved so we'll have
6241 * a think about bumping its value to force at least one task to be
6244 if (env
->imbalance
< busiest
->load_per_task
)
6245 return fix_small_imbalance(env
, sds
);
6248 /******* find_busiest_group() helpers end here *********************/
6251 * find_busiest_group - Returns the busiest group within the sched_domain
6252 * if there is an imbalance. If there isn't an imbalance, and
6253 * the user has opted for power-savings, it returns a group whose
6254 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6255 * such a group exists.
6257 * Also calculates the amount of weighted load which should be moved
6258 * to restore balance.
6260 * @env: The load balancing environment.
6262 * Return: - The busiest group if imbalance exists.
6263 * - If no imbalance and user has opted for power-savings balance,
6264 * return the least loaded group whose CPUs can be
6265 * put to idle by rebalancing its tasks onto our group.
6267 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
6269 struct sg_lb_stats
*local
, *busiest
;
6270 struct sd_lb_stats sds
;
6272 init_sd_lb_stats(&sds
);
6275 * Compute the various statistics relavent for load balancing at
6278 update_sd_lb_stats(env
, &sds
);
6279 local
= &sds
.local_stat
;
6280 busiest
= &sds
.busiest_stat
;
6282 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
6283 check_asym_packing(env
, &sds
))
6286 /* There is no busy sibling group to pull tasks from */
6287 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
6290 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
6291 / sds
.total_capacity
;
6294 * If the busiest group is imbalanced the below checks don't
6295 * work because they assume all things are equal, which typically
6296 * isn't true due to cpus_allowed constraints and the like.
6298 if (busiest
->group_imb
)
6301 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6302 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_free_capacity
&&
6303 !busiest
->group_has_free_capacity
)
6307 * If the local group is more busy than the selected busiest group
6308 * don't try and pull any tasks.
6310 if (local
->avg_load
>= busiest
->avg_load
)
6314 * Don't pull any tasks if this group is already above the domain
6317 if (local
->avg_load
>= sds
.avg_load
)
6320 if (env
->idle
== CPU_IDLE
) {
6322 * This cpu is idle. If the busiest group load doesn't
6323 * have more tasks than the number of available cpu's and
6324 * there is no imbalance between this and busiest group
6325 * wrt to idle cpu's, it is balanced.
6327 if ((local
->idle_cpus
< busiest
->idle_cpus
) &&
6328 busiest
->sum_nr_running
<= busiest
->group_weight
)
6332 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6333 * imbalance_pct to be conservative.
6335 if (100 * busiest
->avg_load
<=
6336 env
->sd
->imbalance_pct
* local
->avg_load
)
6341 /* Looks like there is an imbalance. Compute it */
6342 calculate_imbalance(env
, &sds
);
6351 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6353 static struct rq
*find_busiest_queue(struct lb_env
*env
,
6354 struct sched_group
*group
)
6356 struct rq
*busiest
= NULL
, *rq
;
6357 unsigned long busiest_load
= 0, busiest_capacity
= 1;
6360 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6361 unsigned long capacity
, capacity_factor
, wl
;
6365 rt
= fbq_classify_rq(rq
);
6368 * We classify groups/runqueues into three groups:
6369 * - regular: there are !numa tasks
6370 * - remote: there are numa tasks that run on the 'wrong' node
6371 * - all: there is no distinction
6373 * In order to avoid migrating ideally placed numa tasks,
6374 * ignore those when there's better options.
6376 * If we ignore the actual busiest queue to migrate another
6377 * task, the next balance pass can still reduce the busiest
6378 * queue by moving tasks around inside the node.
6380 * If we cannot move enough load due to this classification
6381 * the next pass will adjust the group classification and
6382 * allow migration of more tasks.
6384 * Both cases only affect the total convergence complexity.
6386 if (rt
> env
->fbq_type
)
6389 capacity
= capacity_of(i
);
6390 capacity_factor
= DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
);
6391 if (!capacity_factor
)
6392 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6394 wl
= weighted_cpuload(i
);
6397 * When comparing with imbalance, use weighted_cpuload()
6398 * which is not scaled with the cpu capacity.
6400 if (capacity_factor
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6404 * For the load comparisons with the other cpu's, consider
6405 * the weighted_cpuload() scaled with the cpu capacity, so
6406 * that the load can be moved away from the cpu that is
6407 * potentially running at a lower capacity.
6409 * Thus we're looking for max(wl_i / capacity_i), crosswise
6410 * multiplication to rid ourselves of the division works out
6411 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6412 * our previous maximum.
6414 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
6416 busiest_capacity
= capacity
;
6425 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6426 * so long as it is large enough.
6428 #define MAX_PINNED_INTERVAL 512
6430 /* Working cpumask for load_balance and load_balance_newidle. */
6431 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6433 static int need_active_balance(struct lb_env
*env
)
6435 struct sched_domain
*sd
= env
->sd
;
6437 if (env
->idle
== CPU_NEWLY_IDLE
) {
6440 * ASYM_PACKING needs to force migrate tasks from busy but
6441 * higher numbered CPUs in order to pack all tasks in the
6442 * lowest numbered CPUs.
6444 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6448 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6451 static int active_load_balance_cpu_stop(void *data
);
6453 static int should_we_balance(struct lb_env
*env
)
6455 struct sched_group
*sg
= env
->sd
->groups
;
6456 struct cpumask
*sg_cpus
, *sg_mask
;
6457 int cpu
, balance_cpu
= -1;
6460 * In the newly idle case, we will allow all the cpu's
6461 * to do the newly idle load balance.
6463 if (env
->idle
== CPU_NEWLY_IDLE
)
6466 sg_cpus
= sched_group_cpus(sg
);
6467 sg_mask
= sched_group_mask(sg
);
6468 /* Try to find first idle cpu */
6469 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6470 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6477 if (balance_cpu
== -1)
6478 balance_cpu
= group_balance_cpu(sg
);
6481 * First idle cpu or the first cpu(busiest) in this sched group
6482 * is eligible for doing load balancing at this and above domains.
6484 return balance_cpu
== env
->dst_cpu
;
6488 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6489 * tasks if there is an imbalance.
6491 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6492 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6493 int *continue_balancing
)
6495 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6496 struct sched_domain
*sd_parent
= sd
->parent
;
6497 struct sched_group
*group
;
6499 unsigned long flags
;
6500 struct cpumask
*cpus
= __get_cpu_var(load_balance_mask
);
6502 struct lb_env env
= {
6504 .dst_cpu
= this_cpu
,
6506 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6508 .loop_break
= sched_nr_migrate_break
,
6514 * For NEWLY_IDLE load_balancing, we don't need to consider
6515 * other cpus in our group
6517 if (idle
== CPU_NEWLY_IDLE
)
6518 env
.dst_grpmask
= NULL
;
6520 cpumask_copy(cpus
, cpu_active_mask
);
6522 schedstat_inc(sd
, lb_count
[idle
]);
6525 if (!should_we_balance(&env
)) {
6526 *continue_balancing
= 0;
6530 group
= find_busiest_group(&env
);
6532 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6536 busiest
= find_busiest_queue(&env
, group
);
6538 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6542 BUG_ON(busiest
== env
.dst_rq
);
6544 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6547 if (busiest
->nr_running
> 1) {
6549 * Attempt to move tasks. If find_busiest_group has found
6550 * an imbalance but busiest->nr_running <= 1, the group is
6551 * still unbalanced. ld_moved simply stays zero, so it is
6552 * correctly treated as an imbalance.
6554 env
.flags
|= LBF_ALL_PINNED
;
6555 env
.src_cpu
= busiest
->cpu
;
6556 env
.src_rq
= busiest
;
6557 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6560 local_irq_save(flags
);
6561 double_rq_lock(env
.dst_rq
, busiest
);
6564 * cur_ld_moved - load moved in current iteration
6565 * ld_moved - cumulative load moved across iterations
6567 cur_ld_moved
= move_tasks(&env
);
6568 ld_moved
+= cur_ld_moved
;
6569 double_rq_unlock(env
.dst_rq
, busiest
);
6570 local_irq_restore(flags
);
6573 * some other cpu did the load balance for us.
6575 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
6576 resched_cpu(env
.dst_cpu
);
6578 if (env
.flags
& LBF_NEED_BREAK
) {
6579 env
.flags
&= ~LBF_NEED_BREAK
;
6584 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6585 * us and move them to an alternate dst_cpu in our sched_group
6586 * where they can run. The upper limit on how many times we
6587 * iterate on same src_cpu is dependent on number of cpus in our
6590 * This changes load balance semantics a bit on who can move
6591 * load to a given_cpu. In addition to the given_cpu itself
6592 * (or a ilb_cpu acting on its behalf where given_cpu is
6593 * nohz-idle), we now have balance_cpu in a position to move
6594 * load to given_cpu. In rare situations, this may cause
6595 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6596 * _independently_ and at _same_ time to move some load to
6597 * given_cpu) causing exceess load to be moved to given_cpu.
6598 * This however should not happen so much in practice and
6599 * moreover subsequent load balance cycles should correct the
6600 * excess load moved.
6602 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6604 /* Prevent to re-select dst_cpu via env's cpus */
6605 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6607 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6608 env
.dst_cpu
= env
.new_dst_cpu
;
6609 env
.flags
&= ~LBF_DST_PINNED
;
6611 env
.loop_break
= sched_nr_migrate_break
;
6614 * Go back to "more_balance" rather than "redo" since we
6615 * need to continue with same src_cpu.
6621 * We failed to reach balance because of affinity.
6624 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
6626 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0) {
6627 *group_imbalance
= 1;
6628 } else if (*group_imbalance
)
6629 *group_imbalance
= 0;
6632 /* All tasks on this runqueue were pinned by CPU affinity */
6633 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6634 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6635 if (!cpumask_empty(cpus
)) {
6637 env
.loop_break
= sched_nr_migrate_break
;
6645 schedstat_inc(sd
, lb_failed
[idle
]);
6647 * Increment the failure counter only on periodic balance.
6648 * We do not want newidle balance, which can be very
6649 * frequent, pollute the failure counter causing
6650 * excessive cache_hot migrations and active balances.
6652 if (idle
!= CPU_NEWLY_IDLE
)
6653 sd
->nr_balance_failed
++;
6655 if (need_active_balance(&env
)) {
6656 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6658 /* don't kick the active_load_balance_cpu_stop,
6659 * if the curr task on busiest cpu can't be
6662 if (!cpumask_test_cpu(this_cpu
,
6663 tsk_cpus_allowed(busiest
->curr
))) {
6664 raw_spin_unlock_irqrestore(&busiest
->lock
,
6666 env
.flags
|= LBF_ALL_PINNED
;
6667 goto out_one_pinned
;
6671 * ->active_balance synchronizes accesses to
6672 * ->active_balance_work. Once set, it's cleared
6673 * only after active load balance is finished.
6675 if (!busiest
->active_balance
) {
6676 busiest
->active_balance
= 1;
6677 busiest
->push_cpu
= this_cpu
;
6680 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
6682 if (active_balance
) {
6683 stop_one_cpu_nowait(cpu_of(busiest
),
6684 active_load_balance_cpu_stop
, busiest
,
6685 &busiest
->active_balance_work
);
6689 * We've kicked active balancing, reset the failure
6692 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
6695 sd
->nr_balance_failed
= 0;
6697 if (likely(!active_balance
)) {
6698 /* We were unbalanced, so reset the balancing interval */
6699 sd
->balance_interval
= sd
->min_interval
;
6702 * If we've begun active balancing, start to back off. This
6703 * case may not be covered by the all_pinned logic if there
6704 * is only 1 task on the busy runqueue (because we don't call
6707 if (sd
->balance_interval
< sd
->max_interval
)
6708 sd
->balance_interval
*= 2;
6714 schedstat_inc(sd
, lb_balanced
[idle
]);
6716 sd
->nr_balance_failed
= 0;
6719 /* tune up the balancing interval */
6720 if (((env
.flags
& LBF_ALL_PINNED
) &&
6721 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
6722 (sd
->balance_interval
< sd
->max_interval
))
6723 sd
->balance_interval
*= 2;
6730 static inline unsigned long
6731 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
6733 unsigned long interval
= sd
->balance_interval
;
6736 interval
*= sd
->busy_factor
;
6738 /* scale ms to jiffies */
6739 interval
= msecs_to_jiffies(interval
);
6740 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6746 update_next_balance(struct sched_domain
*sd
, int cpu_busy
, unsigned long *next_balance
)
6748 unsigned long interval
, next
;
6750 interval
= get_sd_balance_interval(sd
, cpu_busy
);
6751 next
= sd
->last_balance
+ interval
;
6753 if (time_after(*next_balance
, next
))
6754 *next_balance
= next
;
6758 * idle_balance is called by schedule() if this_cpu is about to become
6759 * idle. Attempts to pull tasks from other CPUs.
6761 static int idle_balance(struct rq
*this_rq
)
6763 unsigned long next_balance
= jiffies
+ HZ
;
6764 int this_cpu
= this_rq
->cpu
;
6765 struct sched_domain
*sd
;
6766 int pulled_task
= 0;
6769 idle_enter_fair(this_rq
);
6772 * We must set idle_stamp _before_ calling idle_balance(), such that we
6773 * measure the duration of idle_balance() as idle time.
6775 this_rq
->idle_stamp
= rq_clock(this_rq
);
6777 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
6778 !this_rq
->rd
->overload
) {
6780 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
6782 update_next_balance(sd
, 0, &next_balance
);
6789 * Drop the rq->lock, but keep IRQ/preempt disabled.
6791 raw_spin_unlock(&this_rq
->lock
);
6793 update_blocked_averages(this_cpu
);
6795 for_each_domain(this_cpu
, sd
) {
6796 int continue_balancing
= 1;
6797 u64 t0
, domain_cost
;
6799 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6802 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
6803 update_next_balance(sd
, 0, &next_balance
);
6807 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
6808 t0
= sched_clock_cpu(this_cpu
);
6810 pulled_task
= load_balance(this_cpu
, this_rq
,
6812 &continue_balancing
);
6814 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
6815 if (domain_cost
> sd
->max_newidle_lb_cost
)
6816 sd
->max_newidle_lb_cost
= domain_cost
;
6818 curr_cost
+= domain_cost
;
6821 update_next_balance(sd
, 0, &next_balance
);
6824 * Stop searching for tasks to pull if there are
6825 * now runnable tasks on this rq.
6827 if (pulled_task
|| this_rq
->nr_running
> 0)
6832 raw_spin_lock(&this_rq
->lock
);
6834 if (curr_cost
> this_rq
->max_idle_balance_cost
)
6835 this_rq
->max_idle_balance_cost
= curr_cost
;
6838 * While browsing the domains, we released the rq lock, a task could
6839 * have been enqueued in the meantime. Since we're not going idle,
6840 * pretend we pulled a task.
6842 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
6846 /* Move the next balance forward */
6847 if (time_after(this_rq
->next_balance
, next_balance
))
6848 this_rq
->next_balance
= next_balance
;
6850 /* Is there a task of a high priority class? */
6851 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
6855 idle_exit_fair(this_rq
);
6856 this_rq
->idle_stamp
= 0;
6863 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6864 * running tasks off the busiest CPU onto idle CPUs. It requires at
6865 * least 1 task to be running on each physical CPU where possible, and
6866 * avoids physical / logical imbalances.
6868 static int active_load_balance_cpu_stop(void *data
)
6870 struct rq
*busiest_rq
= data
;
6871 int busiest_cpu
= cpu_of(busiest_rq
);
6872 int target_cpu
= busiest_rq
->push_cpu
;
6873 struct rq
*target_rq
= cpu_rq(target_cpu
);
6874 struct sched_domain
*sd
;
6876 raw_spin_lock_irq(&busiest_rq
->lock
);
6878 /* make sure the requested cpu hasn't gone down in the meantime */
6879 if (unlikely(busiest_cpu
!= smp_processor_id() ||
6880 !busiest_rq
->active_balance
))
6883 /* Is there any task to move? */
6884 if (busiest_rq
->nr_running
<= 1)
6888 * This condition is "impossible", if it occurs
6889 * we need to fix it. Originally reported by
6890 * Bjorn Helgaas on a 128-cpu setup.
6892 BUG_ON(busiest_rq
== target_rq
);
6894 /* move a task from busiest_rq to target_rq */
6895 double_lock_balance(busiest_rq
, target_rq
);
6897 /* Search for an sd spanning us and the target CPU. */
6899 for_each_domain(target_cpu
, sd
) {
6900 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
6901 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
6906 struct lb_env env
= {
6908 .dst_cpu
= target_cpu
,
6909 .dst_rq
= target_rq
,
6910 .src_cpu
= busiest_rq
->cpu
,
6911 .src_rq
= busiest_rq
,
6915 schedstat_inc(sd
, alb_count
);
6917 if (move_one_task(&env
))
6918 schedstat_inc(sd
, alb_pushed
);
6920 schedstat_inc(sd
, alb_failed
);
6923 double_unlock_balance(busiest_rq
, target_rq
);
6925 busiest_rq
->active_balance
= 0;
6926 raw_spin_unlock_irq(&busiest_rq
->lock
);
6930 static inline int on_null_domain(struct rq
*rq
)
6932 return unlikely(!rcu_dereference_sched(rq
->sd
));
6935 #ifdef CONFIG_NO_HZ_COMMON
6937 * idle load balancing details
6938 * - When one of the busy CPUs notice that there may be an idle rebalancing
6939 * needed, they will kick the idle load balancer, which then does idle
6940 * load balancing for all the idle CPUs.
6943 cpumask_var_t idle_cpus_mask
;
6945 unsigned long next_balance
; /* in jiffy units */
6946 } nohz ____cacheline_aligned
;
6948 static inline int find_new_ilb(void)
6950 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
6952 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
6959 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6960 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6961 * CPU (if there is one).
6963 static void nohz_balancer_kick(void)
6967 nohz
.next_balance
++;
6969 ilb_cpu
= find_new_ilb();
6971 if (ilb_cpu
>= nr_cpu_ids
)
6974 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
6977 * Use smp_send_reschedule() instead of resched_cpu().
6978 * This way we generate a sched IPI on the target cpu which
6979 * is idle. And the softirq performing nohz idle load balance
6980 * will be run before returning from the IPI.
6982 smp_send_reschedule(ilb_cpu
);
6986 static inline void nohz_balance_exit_idle(int cpu
)
6988 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
6990 * Completely isolated CPUs don't ever set, so we must test.
6992 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
6993 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
6994 atomic_dec(&nohz
.nr_cpus
);
6996 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7000 static inline void set_cpu_sd_state_busy(void)
7002 struct sched_domain
*sd
;
7003 int cpu
= smp_processor_id();
7006 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7008 if (!sd
|| !sd
->nohz_idle
)
7012 atomic_inc(&sd
->groups
->sgc
->nr_busy_cpus
);
7017 void set_cpu_sd_state_idle(void)
7019 struct sched_domain
*sd
;
7020 int cpu
= smp_processor_id();
7023 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7025 if (!sd
|| sd
->nohz_idle
)
7029 atomic_dec(&sd
->groups
->sgc
->nr_busy_cpus
);
7035 * This routine will record that the cpu is going idle with tick stopped.
7036 * This info will be used in performing idle load balancing in the future.
7038 void nohz_balance_enter_idle(int cpu
)
7041 * If this cpu is going down, then nothing needs to be done.
7043 if (!cpu_active(cpu
))
7046 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
7050 * If we're a completely isolated CPU, we don't play.
7052 if (on_null_domain(cpu_rq(cpu
)))
7055 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
7056 atomic_inc(&nohz
.nr_cpus
);
7057 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7060 static int sched_ilb_notifier(struct notifier_block
*nfb
,
7061 unsigned long action
, void *hcpu
)
7063 switch (action
& ~CPU_TASKS_FROZEN
) {
7065 nohz_balance_exit_idle(smp_processor_id());
7073 static DEFINE_SPINLOCK(balancing
);
7076 * Scale the max load_balance interval with the number of CPUs in the system.
7077 * This trades load-balance latency on larger machines for less cross talk.
7079 void update_max_interval(void)
7081 max_load_balance_interval
= HZ
*num_online_cpus()/10;
7085 * It checks each scheduling domain to see if it is due to be balanced,
7086 * and initiates a balancing operation if so.
7088 * Balancing parameters are set up in init_sched_domains.
7090 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
7092 int continue_balancing
= 1;
7094 unsigned long interval
;
7095 struct sched_domain
*sd
;
7096 /* Earliest time when we have to do rebalance again */
7097 unsigned long next_balance
= jiffies
+ 60*HZ
;
7098 int update_next_balance
= 0;
7099 int need_serialize
, need_decay
= 0;
7102 update_blocked_averages(cpu
);
7105 for_each_domain(cpu
, sd
) {
7107 * Decay the newidle max times here because this is a regular
7108 * visit to all the domains. Decay ~1% per second.
7110 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
7111 sd
->max_newidle_lb_cost
=
7112 (sd
->max_newidle_lb_cost
* 253) / 256;
7113 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
7116 max_cost
+= sd
->max_newidle_lb_cost
;
7118 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7122 * Stop the load balance at this level. There is another
7123 * CPU in our sched group which is doing load balancing more
7126 if (!continue_balancing
) {
7132 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7134 need_serialize
= sd
->flags
& SD_SERIALIZE
;
7135 if (need_serialize
) {
7136 if (!spin_trylock(&balancing
))
7140 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
7141 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
7143 * The LBF_DST_PINNED logic could have changed
7144 * env->dst_cpu, so we can't know our idle
7145 * state even if we migrated tasks. Update it.
7147 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
7149 sd
->last_balance
= jiffies
;
7150 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7153 spin_unlock(&balancing
);
7155 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
7156 next_balance
= sd
->last_balance
+ interval
;
7157 update_next_balance
= 1;
7162 * Ensure the rq-wide value also decays but keep it at a
7163 * reasonable floor to avoid funnies with rq->avg_idle.
7165 rq
->max_idle_balance_cost
=
7166 max((u64
)sysctl_sched_migration_cost
, max_cost
);
7171 * next_balance will be updated only when there is a need.
7172 * When the cpu is attached to null domain for ex, it will not be
7175 if (likely(update_next_balance
))
7176 rq
->next_balance
= next_balance
;
7179 #ifdef CONFIG_NO_HZ_COMMON
7181 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7182 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7184 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
7186 int this_cpu
= this_rq
->cpu
;
7190 if (idle
!= CPU_IDLE
||
7191 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
7194 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
7195 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
7199 * If this cpu gets work to do, stop the load balancing
7200 * work being done for other cpus. Next load
7201 * balancing owner will pick it up.
7206 rq
= cpu_rq(balance_cpu
);
7209 * If time for next balance is due,
7212 if (time_after_eq(jiffies
, rq
->next_balance
)) {
7213 raw_spin_lock_irq(&rq
->lock
);
7214 update_rq_clock(rq
);
7215 update_idle_cpu_load(rq
);
7216 raw_spin_unlock_irq(&rq
->lock
);
7217 rebalance_domains(rq
, CPU_IDLE
);
7220 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
7221 this_rq
->next_balance
= rq
->next_balance
;
7223 nohz
.next_balance
= this_rq
->next_balance
;
7225 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
7229 * Current heuristic for kicking the idle load balancer in the presence
7230 * of an idle cpu is the system.
7231 * - This rq has more than one task.
7232 * - At any scheduler domain level, this cpu's scheduler group has multiple
7233 * busy cpu's exceeding the group's capacity.
7234 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7235 * domain span are idle.
7237 static inline int nohz_kick_needed(struct rq
*rq
)
7239 unsigned long now
= jiffies
;
7240 struct sched_domain
*sd
;
7241 struct sched_group_capacity
*sgc
;
7242 int nr_busy
, cpu
= rq
->cpu
;
7244 if (unlikely(rq
->idle_balance
))
7248 * We may be recently in ticked or tickless idle mode. At the first
7249 * busy tick after returning from idle, we will update the busy stats.
7251 set_cpu_sd_state_busy();
7252 nohz_balance_exit_idle(cpu
);
7255 * None are in tickless mode and hence no need for NOHZ idle load
7258 if (likely(!atomic_read(&nohz
.nr_cpus
)))
7261 if (time_before(now
, nohz
.next_balance
))
7264 if (rq
->nr_running
>= 2)
7268 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7271 sgc
= sd
->groups
->sgc
;
7272 nr_busy
= atomic_read(&sgc
->nr_busy_cpus
);
7275 goto need_kick_unlock
;
7278 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
7280 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
7281 sched_domain_span(sd
)) < cpu
))
7282 goto need_kick_unlock
;
7293 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
7297 * run_rebalance_domains is triggered when needed from the scheduler tick.
7298 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7300 static void run_rebalance_domains(struct softirq_action
*h
)
7302 struct rq
*this_rq
= this_rq();
7303 enum cpu_idle_type idle
= this_rq
->idle_balance
?
7304 CPU_IDLE
: CPU_NOT_IDLE
;
7306 rebalance_domains(this_rq
, idle
);
7309 * If this cpu has a pending nohz_balance_kick, then do the
7310 * balancing on behalf of the other idle cpus whose ticks are
7313 nohz_idle_balance(this_rq
, idle
);
7317 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7319 void trigger_load_balance(struct rq
*rq
)
7321 /* Don't need to rebalance while attached to NULL domain */
7322 if (unlikely(on_null_domain(rq
)))
7325 if (time_after_eq(jiffies
, rq
->next_balance
))
7326 raise_softirq(SCHED_SOFTIRQ
);
7327 #ifdef CONFIG_NO_HZ_COMMON
7328 if (nohz_kick_needed(rq
))
7329 nohz_balancer_kick();
7333 static void rq_online_fair(struct rq
*rq
)
7338 static void rq_offline_fair(struct rq
*rq
)
7342 /* Ensure any throttled groups are reachable by pick_next_task */
7343 unthrottle_offline_cfs_rqs(rq
);
7346 #endif /* CONFIG_SMP */
7349 * scheduler tick hitting a task of our scheduling class:
7351 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
7353 struct cfs_rq
*cfs_rq
;
7354 struct sched_entity
*se
= &curr
->se
;
7356 for_each_sched_entity(se
) {
7357 cfs_rq
= cfs_rq_of(se
);
7358 entity_tick(cfs_rq
, se
, queued
);
7361 if (numabalancing_enabled
)
7362 task_tick_numa(rq
, curr
);
7364 update_rq_runnable_avg(rq
, 1);
7368 * called on fork with the child task as argument from the parent's context
7369 * - child not yet on the tasklist
7370 * - preemption disabled
7372 static void task_fork_fair(struct task_struct
*p
)
7374 struct cfs_rq
*cfs_rq
;
7375 struct sched_entity
*se
= &p
->se
, *curr
;
7376 int this_cpu
= smp_processor_id();
7377 struct rq
*rq
= this_rq();
7378 unsigned long flags
;
7380 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7382 update_rq_clock(rq
);
7384 cfs_rq
= task_cfs_rq(current
);
7385 curr
= cfs_rq
->curr
;
7388 * Not only the cpu but also the task_group of the parent might have
7389 * been changed after parent->se.parent,cfs_rq were copied to
7390 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7391 * of child point to valid ones.
7394 __set_task_cpu(p
, this_cpu
);
7397 update_curr(cfs_rq
);
7400 se
->vruntime
= curr
->vruntime
;
7401 place_entity(cfs_rq
, se
, 1);
7403 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
7405 * Upon rescheduling, sched_class::put_prev_task() will place
7406 * 'current' within the tree based on its new key value.
7408 swap(curr
->vruntime
, se
->vruntime
);
7409 resched_task(rq
->curr
);
7412 se
->vruntime
-= cfs_rq
->min_vruntime
;
7414 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7418 * Priority of the task has changed. Check to see if we preempt
7422 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
7428 * Reschedule if we are currently running on this runqueue and
7429 * our priority decreased, or if we are not currently running on
7430 * this runqueue and our priority is higher than the current's
7432 if (rq
->curr
== p
) {
7433 if (p
->prio
> oldprio
)
7434 resched_task(rq
->curr
);
7436 check_preempt_curr(rq
, p
, 0);
7439 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
7441 struct sched_entity
*se
= &p
->se
;
7442 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7445 * Ensure the task's vruntime is normalized, so that when it's
7446 * switched back to the fair class the enqueue_entity(.flags=0) will
7447 * do the right thing.
7449 * If it's on_rq, then the dequeue_entity(.flags=0) will already
7450 * have normalized the vruntime, if it's !on_rq, then only when
7451 * the task is sleeping will it still have non-normalized vruntime.
7453 if (!p
->on_rq
&& p
->state
!= TASK_RUNNING
) {
7455 * Fix up our vruntime so that the current sleep doesn't
7456 * cause 'unlimited' sleep bonus.
7458 place_entity(cfs_rq
, se
, 0);
7459 se
->vruntime
-= cfs_rq
->min_vruntime
;
7464 * Remove our load from contribution when we leave sched_fair
7465 * and ensure we don't carry in an old decay_count if we
7468 if (se
->avg
.decay_count
) {
7469 __synchronize_entity_decay(se
);
7470 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7476 * We switched to the sched_fair class.
7478 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7480 struct sched_entity
*se
= &p
->se
;
7481 #ifdef CONFIG_FAIR_GROUP_SCHED
7483 * Since the real-depth could have been changed (only FAIR
7484 * class maintain depth value), reset depth properly.
7486 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7492 * We were most likely switched from sched_rt, so
7493 * kick off the schedule if running, otherwise just see
7494 * if we can still preempt the current task.
7497 resched_task(rq
->curr
);
7499 check_preempt_curr(rq
, p
, 0);
7502 /* Account for a task changing its policy or group.
7504 * This routine is mostly called to set cfs_rq->curr field when a task
7505 * migrates between groups/classes.
7507 static void set_curr_task_fair(struct rq
*rq
)
7509 struct sched_entity
*se
= &rq
->curr
->se
;
7511 for_each_sched_entity(se
) {
7512 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7514 set_next_entity(cfs_rq
, se
);
7515 /* ensure bandwidth has been allocated on our new cfs_rq */
7516 account_cfs_rq_runtime(cfs_rq
, 0);
7520 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7522 cfs_rq
->tasks_timeline
= RB_ROOT
;
7523 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7524 #ifndef CONFIG_64BIT
7525 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7528 atomic64_set(&cfs_rq
->decay_counter
, 1);
7529 atomic_long_set(&cfs_rq
->removed_load
, 0);
7533 #ifdef CONFIG_FAIR_GROUP_SCHED
7534 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
7536 struct sched_entity
*se
= &p
->se
;
7537 struct cfs_rq
*cfs_rq
;
7540 * If the task was not on the rq at the time of this cgroup movement
7541 * it must have been asleep, sleeping tasks keep their ->vruntime
7542 * absolute on their old rq until wakeup (needed for the fair sleeper
7543 * bonus in place_entity()).
7545 * If it was on the rq, we've just 'preempted' it, which does convert
7546 * ->vruntime to a relative base.
7548 * Make sure both cases convert their relative position when migrating
7549 * to another cgroup's rq. This does somewhat interfere with the
7550 * fair sleeper stuff for the first placement, but who cares.
7553 * When !on_rq, vruntime of the task has usually NOT been normalized.
7554 * But there are some cases where it has already been normalized:
7556 * - Moving a forked child which is waiting for being woken up by
7557 * wake_up_new_task().
7558 * - Moving a task which has been woken up by try_to_wake_up() and
7559 * waiting for actually being woken up by sched_ttwu_pending().
7561 * To prevent boost or penalty in the new cfs_rq caused by delta
7562 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7564 if (!on_rq
&& (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7568 se
->vruntime
-= cfs_rq_of(se
)->min_vruntime
;
7569 set_task_rq(p
, task_cpu(p
));
7570 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7572 cfs_rq
= cfs_rq_of(se
);
7573 se
->vruntime
+= cfs_rq
->min_vruntime
;
7576 * migrate_task_rq_fair() will have removed our previous
7577 * contribution, but we must synchronize for ongoing future
7580 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7581 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
7586 void free_fair_sched_group(struct task_group
*tg
)
7590 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7592 for_each_possible_cpu(i
) {
7594 kfree(tg
->cfs_rq
[i
]);
7603 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7605 struct cfs_rq
*cfs_rq
;
7606 struct sched_entity
*se
;
7609 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7612 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7616 tg
->shares
= NICE_0_LOAD
;
7618 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7620 for_each_possible_cpu(i
) {
7621 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7622 GFP_KERNEL
, cpu_to_node(i
));
7626 se
= kzalloc_node(sizeof(struct sched_entity
),
7627 GFP_KERNEL
, cpu_to_node(i
));
7631 init_cfs_rq(cfs_rq
);
7632 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
7643 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7645 struct rq
*rq
= cpu_rq(cpu
);
7646 unsigned long flags
;
7649 * Only empty task groups can be destroyed; so we can speculatively
7650 * check on_list without danger of it being re-added.
7652 if (!tg
->cfs_rq
[cpu
]->on_list
)
7655 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7656 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
7657 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7660 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7661 struct sched_entity
*se
, int cpu
,
7662 struct sched_entity
*parent
)
7664 struct rq
*rq
= cpu_rq(cpu
);
7668 init_cfs_rq_runtime(cfs_rq
);
7670 tg
->cfs_rq
[cpu
] = cfs_rq
;
7673 /* se could be NULL for root_task_group */
7678 se
->cfs_rq
= &rq
->cfs
;
7681 se
->cfs_rq
= parent
->my_q
;
7682 se
->depth
= parent
->depth
+ 1;
7686 /* guarantee group entities always have weight */
7687 update_load_set(&se
->load
, NICE_0_LOAD
);
7688 se
->parent
= parent
;
7691 static DEFINE_MUTEX(shares_mutex
);
7693 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7696 unsigned long flags
;
7699 * We can't change the weight of the root cgroup.
7704 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
7706 mutex_lock(&shares_mutex
);
7707 if (tg
->shares
== shares
)
7710 tg
->shares
= shares
;
7711 for_each_possible_cpu(i
) {
7712 struct rq
*rq
= cpu_rq(i
);
7713 struct sched_entity
*se
;
7716 /* Propagate contribution to hierarchy */
7717 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7719 /* Possible calls to update_curr() need rq clock */
7720 update_rq_clock(rq
);
7721 for_each_sched_entity(se
)
7722 update_cfs_shares(group_cfs_rq(se
));
7723 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7727 mutex_unlock(&shares_mutex
);
7730 #else /* CONFIG_FAIR_GROUP_SCHED */
7732 void free_fair_sched_group(struct task_group
*tg
) { }
7734 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7739 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
7741 #endif /* CONFIG_FAIR_GROUP_SCHED */
7744 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
7746 struct sched_entity
*se
= &task
->se
;
7747 unsigned int rr_interval
= 0;
7750 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7753 if (rq
->cfs
.load
.weight
)
7754 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
7760 * All the scheduling class methods:
7762 const struct sched_class fair_sched_class
= {
7763 .next
= &idle_sched_class
,
7764 .enqueue_task
= enqueue_task_fair
,
7765 .dequeue_task
= dequeue_task_fair
,
7766 .yield_task
= yield_task_fair
,
7767 .yield_to_task
= yield_to_task_fair
,
7769 .check_preempt_curr
= check_preempt_wakeup
,
7771 .pick_next_task
= pick_next_task_fair
,
7772 .put_prev_task
= put_prev_task_fair
,
7775 .select_task_rq
= select_task_rq_fair
,
7776 .migrate_task_rq
= migrate_task_rq_fair
,
7778 .rq_online
= rq_online_fair
,
7779 .rq_offline
= rq_offline_fair
,
7781 .task_waking
= task_waking_fair
,
7784 .set_curr_task
= set_curr_task_fair
,
7785 .task_tick
= task_tick_fair
,
7786 .task_fork
= task_fork_fair
,
7788 .prio_changed
= prio_changed_fair
,
7789 .switched_from
= switched_from_fair
,
7790 .switched_to
= switched_to_fair
,
7792 .get_rr_interval
= get_rr_interval_fair
,
7794 #ifdef CONFIG_FAIR_GROUP_SCHED
7795 .task_move_group
= task_move_group_fair
,
7799 #ifdef CONFIG_SCHED_DEBUG
7800 void print_cfs_stats(struct seq_file
*m
, int cpu
)
7802 struct cfs_rq
*cfs_rq
;
7805 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
7806 print_cfs_rq(m
, cpu
, cfs_rq
);
7811 __init
void init_sched_fair_class(void)
7814 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7816 #ifdef CONFIG_NO_HZ_COMMON
7817 nohz
.next_balance
= jiffies
;
7818 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7819 cpu_notifier(sched_ilb_notifier
, 0);