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 power_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
+= power_of(cpu
);
1055 * If we raced with hotplug and there are no CPUs left in our mask
1056 * the @ns structure is NULL'ed and task_numa_compare() will
1057 * not find this node attractive.
1059 * We'll either bail at !has_free_capacity, or we'll detect a huge
1060 * imbalance and bail there.
1065 ns
->load
= (ns
->load
* SCHED_POWER_SCALE
) / ns
->compute_capacity
;
1067 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_POWER_SCALE
);
1068 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1071 struct task_numa_env
{
1072 struct task_struct
*p
;
1074 int src_cpu
, src_nid
;
1075 int dst_cpu
, dst_nid
;
1077 struct numa_stats src_stats
, dst_stats
;
1081 struct task_struct
*best_task
;
1086 static void task_numa_assign(struct task_numa_env
*env
,
1087 struct task_struct
*p
, long imp
)
1090 put_task_struct(env
->best_task
);
1095 env
->best_imp
= imp
;
1096 env
->best_cpu
= env
->dst_cpu
;
1099 static bool load_too_imbalanced(long orig_src_load
, long orig_dst_load
,
1100 long src_load
, long dst_load
,
1101 struct task_numa_env
*env
)
1105 /* We care about the slope of the imbalance, not the direction. */
1106 if (dst_load
< src_load
)
1107 swap(dst_load
, src_load
);
1109 /* Is the difference below the threshold? */
1110 imb
= dst_load
* 100 - src_load
* env
->imbalance_pct
;
1115 * The imbalance is above the allowed threshold.
1116 * Compare it with the old imbalance.
1118 if (orig_dst_load
< orig_src_load
)
1119 swap(orig_dst_load
, orig_src_load
);
1121 old_imb
= orig_dst_load
* 100 - orig_src_load
* env
->imbalance_pct
;
1123 /* Would this change make things worse? */
1124 return (old_imb
> imb
);
1128 * This checks if the overall compute and NUMA accesses of the system would
1129 * be improved if the source tasks was migrated to the target dst_cpu taking
1130 * into account that it might be best if task running on the dst_cpu should
1131 * be exchanged with the source task
1133 static void task_numa_compare(struct task_numa_env
*env
,
1134 long taskimp
, long groupimp
)
1136 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1137 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1138 struct task_struct
*cur
;
1139 long orig_src_load
, src_load
;
1140 long orig_dst_load
, dst_load
;
1142 long imp
= (groupimp
> 0) ? groupimp
: taskimp
;
1145 cur
= ACCESS_ONCE(dst_rq
->curr
);
1146 if (cur
->pid
== 0) /* idle */
1150 * "imp" is the fault differential for the source task between the
1151 * source and destination node. Calculate the total differential for
1152 * the source task and potential destination task. The more negative
1153 * the value is, the more rmeote accesses that would be expected to
1154 * be incurred if the tasks were swapped.
1157 /* Skip this swap candidate if cannot move to the source cpu */
1158 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1162 * If dst and source tasks are in the same NUMA group, or not
1163 * in any group then look only at task weights.
1165 if (cur
->numa_group
== env
->p
->numa_group
) {
1166 imp
= taskimp
+ task_weight(cur
, env
->src_nid
) -
1167 task_weight(cur
, env
->dst_nid
);
1169 * Add some hysteresis to prevent swapping the
1170 * tasks within a group over tiny differences.
1172 if (cur
->numa_group
)
1176 * Compare the group weights. If a task is all by
1177 * itself (not part of a group), use the task weight
1180 if (env
->p
->numa_group
)
1185 if (cur
->numa_group
)
1186 imp
+= group_weight(cur
, env
->src_nid
) -
1187 group_weight(cur
, env
->dst_nid
);
1189 imp
+= task_weight(cur
, env
->src_nid
) -
1190 task_weight(cur
, env
->dst_nid
);
1194 if (imp
< env
->best_imp
)
1198 /* Is there capacity at our destination? */
1199 if (env
->src_stats
.has_free_capacity
&&
1200 !env
->dst_stats
.has_free_capacity
)
1206 /* Balance doesn't matter much if we're running a task per cpu */
1207 if (src_rq
->nr_running
== 1 && dst_rq
->nr_running
== 1)
1211 * In the overloaded case, try and keep the load balanced.
1214 orig_dst_load
= env
->dst_stats
.load
;
1215 orig_src_load
= env
->src_stats
.load
;
1217 /* XXX missing power terms */
1218 load
= task_h_load(env
->p
);
1219 dst_load
= orig_dst_load
+ load
;
1220 src_load
= orig_src_load
- load
;
1223 load
= task_h_load(cur
);
1228 if (load_too_imbalanced(orig_src_load
, orig_dst_load
,
1229 src_load
, dst_load
, env
))
1233 task_numa_assign(env
, cur
, imp
);
1238 static void task_numa_find_cpu(struct task_numa_env
*env
,
1239 long taskimp
, long groupimp
)
1243 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1244 /* Skip this CPU if the source task cannot migrate */
1245 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1249 task_numa_compare(env
, taskimp
, groupimp
);
1253 static int task_numa_migrate(struct task_struct
*p
)
1255 struct task_numa_env env
= {
1258 .src_cpu
= task_cpu(p
),
1259 .src_nid
= task_node(p
),
1261 .imbalance_pct
= 112,
1267 struct sched_domain
*sd
;
1268 unsigned long taskweight
, groupweight
;
1270 long taskimp
, groupimp
;
1273 * Pick the lowest SD_NUMA domain, as that would have the smallest
1274 * imbalance and would be the first to start moving tasks about.
1276 * And we want to avoid any moving of tasks about, as that would create
1277 * random movement of tasks -- counter the numa conditions we're trying
1281 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1283 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1287 * Cpusets can break the scheduler domain tree into smaller
1288 * balance domains, some of which do not cross NUMA boundaries.
1289 * Tasks that are "trapped" in such domains cannot be migrated
1290 * elsewhere, so there is no point in (re)trying.
1292 if (unlikely(!sd
)) {
1293 p
->numa_preferred_nid
= task_node(p
);
1297 taskweight
= task_weight(p
, env
.src_nid
);
1298 groupweight
= group_weight(p
, env
.src_nid
);
1299 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1300 env
.dst_nid
= p
->numa_preferred_nid
;
1301 taskimp
= task_weight(p
, env
.dst_nid
) - taskweight
;
1302 groupimp
= group_weight(p
, env
.dst_nid
) - groupweight
;
1303 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1305 /* If the preferred nid has free capacity, try to use it. */
1306 if (env
.dst_stats
.has_free_capacity
)
1307 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1309 /* No space available on the preferred nid. Look elsewhere. */
1310 if (env
.best_cpu
== -1) {
1311 for_each_online_node(nid
) {
1312 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1315 /* Only consider nodes where both task and groups benefit */
1316 taskimp
= task_weight(p
, nid
) - taskweight
;
1317 groupimp
= group_weight(p
, nid
) - groupweight
;
1318 if (taskimp
< 0 && groupimp
< 0)
1322 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1323 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1327 /* No better CPU than the current one was found. */
1328 if (env
.best_cpu
== -1)
1332 * If the task is part of a workload that spans multiple NUMA nodes,
1333 * and is migrating into one of the workload's active nodes, remember
1334 * this node as the task's preferred numa node, so the workload can
1336 * A task that migrated to a second choice node will be better off
1337 * trying for a better one later. Do not set the preferred node here.
1339 if (p
->numa_group
&& node_isset(env
.dst_nid
, p
->numa_group
->active_nodes
))
1340 sched_setnuma(p
, env
.dst_nid
);
1343 * Reset the scan period if the task is being rescheduled on an
1344 * alternative node to recheck if the tasks is now properly placed.
1346 p
->numa_scan_period
= task_scan_min(p
);
1348 if (env
.best_task
== NULL
) {
1349 ret
= migrate_task_to(p
, env
.best_cpu
);
1351 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1355 ret
= migrate_swap(p
, env
.best_task
);
1357 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1358 put_task_struct(env
.best_task
);
1362 /* Attempt to migrate a task to a CPU on the preferred node. */
1363 static void numa_migrate_preferred(struct task_struct
*p
)
1365 unsigned long interval
= HZ
;
1367 /* This task has no NUMA fault statistics yet */
1368 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults_memory
))
1371 /* Periodically retry migrating the task to the preferred node */
1372 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1373 p
->numa_migrate_retry
= jiffies
+ interval
;
1375 /* Success if task is already running on preferred CPU */
1376 if (task_node(p
) == p
->numa_preferred_nid
)
1379 /* Otherwise, try migrate to a CPU on the preferred node */
1380 task_numa_migrate(p
);
1384 * Find the nodes on which the workload is actively running. We do this by
1385 * tracking the nodes from which NUMA hinting faults are triggered. This can
1386 * be different from the set of nodes where the workload's memory is currently
1389 * The bitmask is used to make smarter decisions on when to do NUMA page
1390 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1391 * are added when they cause over 6/16 of the maximum number of faults, but
1392 * only removed when they drop below 3/16.
1394 static void update_numa_active_node_mask(struct numa_group
*numa_group
)
1396 unsigned long faults
, max_faults
= 0;
1399 for_each_online_node(nid
) {
1400 faults
= group_faults_cpu(numa_group
, nid
);
1401 if (faults
> max_faults
)
1402 max_faults
= faults
;
1405 for_each_online_node(nid
) {
1406 faults
= group_faults_cpu(numa_group
, nid
);
1407 if (!node_isset(nid
, numa_group
->active_nodes
)) {
1408 if (faults
> max_faults
* 6 / 16)
1409 node_set(nid
, numa_group
->active_nodes
);
1410 } else if (faults
< max_faults
* 3 / 16)
1411 node_clear(nid
, numa_group
->active_nodes
);
1416 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1417 * increments. The more local the fault statistics are, the higher the scan
1418 * period will be for the next scan window. If local/remote ratio is below
1419 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1420 * scan period will decrease
1422 #define NUMA_PERIOD_SLOTS 10
1423 #define NUMA_PERIOD_THRESHOLD 3
1426 * Increase the scan period (slow down scanning) if the majority of
1427 * our memory is already on our local node, or if the majority of
1428 * the page accesses are shared with other processes.
1429 * Otherwise, decrease the scan period.
1431 static void update_task_scan_period(struct task_struct
*p
,
1432 unsigned long shared
, unsigned long private)
1434 unsigned int period_slot
;
1438 unsigned long remote
= p
->numa_faults_locality
[0];
1439 unsigned long local
= p
->numa_faults_locality
[1];
1442 * If there were no record hinting faults then either the task is
1443 * completely idle or all activity is areas that are not of interest
1444 * to automatic numa balancing. Scan slower
1446 if (local
+ shared
== 0) {
1447 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1448 p
->numa_scan_period
<< 1);
1450 p
->mm
->numa_next_scan
= jiffies
+
1451 msecs_to_jiffies(p
->numa_scan_period
);
1457 * Prepare to scale scan period relative to the current period.
1458 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1459 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1460 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1462 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1463 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1464 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1465 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1468 diff
= slot
* period_slot
;
1470 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1473 * Scale scan rate increases based on sharing. There is an
1474 * inverse relationship between the degree of sharing and
1475 * the adjustment made to the scanning period. Broadly
1476 * speaking the intent is that there is little point
1477 * scanning faster if shared accesses dominate as it may
1478 * simply bounce migrations uselessly
1480 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
));
1481 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1484 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1485 task_scan_min(p
), task_scan_max(p
));
1486 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1490 * Get the fraction of time the task has been running since the last
1491 * NUMA placement cycle. The scheduler keeps similar statistics, but
1492 * decays those on a 32ms period, which is orders of magnitude off
1493 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1494 * stats only if the task is so new there are no NUMA statistics yet.
1496 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1498 u64 runtime
, delta
, now
;
1499 /* Use the start of this time slice to avoid calculations. */
1500 now
= p
->se
.exec_start
;
1501 runtime
= p
->se
.sum_exec_runtime
;
1503 if (p
->last_task_numa_placement
) {
1504 delta
= runtime
- p
->last_sum_exec_runtime
;
1505 *period
= now
- p
->last_task_numa_placement
;
1507 delta
= p
->se
.avg
.runnable_avg_sum
;
1508 *period
= p
->se
.avg
.runnable_avg_period
;
1511 p
->last_sum_exec_runtime
= runtime
;
1512 p
->last_task_numa_placement
= now
;
1517 static void task_numa_placement(struct task_struct
*p
)
1519 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1520 unsigned long max_faults
= 0, max_group_faults
= 0;
1521 unsigned long fault_types
[2] = { 0, 0 };
1522 unsigned long total_faults
;
1523 u64 runtime
, period
;
1524 spinlock_t
*group_lock
= NULL
;
1526 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1527 if (p
->numa_scan_seq
== seq
)
1529 p
->numa_scan_seq
= seq
;
1530 p
->numa_scan_period_max
= task_scan_max(p
);
1532 total_faults
= p
->numa_faults_locality
[0] +
1533 p
->numa_faults_locality
[1];
1534 runtime
= numa_get_avg_runtime(p
, &period
);
1536 /* If the task is part of a group prevent parallel updates to group stats */
1537 if (p
->numa_group
) {
1538 group_lock
= &p
->numa_group
->lock
;
1539 spin_lock_irq(group_lock
);
1542 /* Find the node with the highest number of faults */
1543 for_each_online_node(nid
) {
1544 unsigned long faults
= 0, group_faults
= 0;
1547 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
1548 long diff
, f_diff
, f_weight
;
1550 i
= task_faults_idx(nid
, priv
);
1552 /* Decay existing window, copy faults since last scan */
1553 diff
= p
->numa_faults_buffer_memory
[i
] - p
->numa_faults_memory
[i
] / 2;
1554 fault_types
[priv
] += p
->numa_faults_buffer_memory
[i
];
1555 p
->numa_faults_buffer_memory
[i
] = 0;
1558 * Normalize the faults_from, so all tasks in a group
1559 * count according to CPU use, instead of by the raw
1560 * number of faults. Tasks with little runtime have
1561 * little over-all impact on throughput, and thus their
1562 * faults are less important.
1564 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
1565 f_weight
= (f_weight
* p
->numa_faults_buffer_cpu
[i
]) /
1567 f_diff
= f_weight
- p
->numa_faults_cpu
[i
] / 2;
1568 p
->numa_faults_buffer_cpu
[i
] = 0;
1570 p
->numa_faults_memory
[i
] += diff
;
1571 p
->numa_faults_cpu
[i
] += f_diff
;
1572 faults
+= p
->numa_faults_memory
[i
];
1573 p
->total_numa_faults
+= diff
;
1574 if (p
->numa_group
) {
1575 /* safe because we can only change our own group */
1576 p
->numa_group
->faults
[i
] += diff
;
1577 p
->numa_group
->faults_cpu
[i
] += f_diff
;
1578 p
->numa_group
->total_faults
+= diff
;
1579 group_faults
+= p
->numa_group
->faults
[i
];
1583 if (faults
> max_faults
) {
1584 max_faults
= faults
;
1588 if (group_faults
> max_group_faults
) {
1589 max_group_faults
= group_faults
;
1590 max_group_nid
= nid
;
1594 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1596 if (p
->numa_group
) {
1597 update_numa_active_node_mask(p
->numa_group
);
1599 * If the preferred task and group nids are different,
1600 * iterate over the nodes again to find the best place.
1602 if (max_nid
!= max_group_nid
) {
1603 unsigned long weight
, max_weight
= 0;
1605 for_each_online_node(nid
) {
1606 weight
= task_weight(p
, nid
) + group_weight(p
, nid
);
1607 if (weight
> max_weight
) {
1608 max_weight
= weight
;
1614 spin_unlock_irq(group_lock
);
1617 /* Preferred node as the node with the most faults */
1618 if (max_faults
&& max_nid
!= p
->numa_preferred_nid
) {
1619 /* Update the preferred nid and migrate task if possible */
1620 sched_setnuma(p
, max_nid
);
1621 numa_migrate_preferred(p
);
1625 static inline int get_numa_group(struct numa_group
*grp
)
1627 return atomic_inc_not_zero(&grp
->refcount
);
1630 static inline void put_numa_group(struct numa_group
*grp
)
1632 if (atomic_dec_and_test(&grp
->refcount
))
1633 kfree_rcu(grp
, rcu
);
1636 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1639 struct numa_group
*grp
, *my_grp
;
1640 struct task_struct
*tsk
;
1642 int cpu
= cpupid_to_cpu(cpupid
);
1645 if (unlikely(!p
->numa_group
)) {
1646 unsigned int size
= sizeof(struct numa_group
) +
1647 4*nr_node_ids
*sizeof(unsigned long);
1649 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1653 atomic_set(&grp
->refcount
, 1);
1654 spin_lock_init(&grp
->lock
);
1655 INIT_LIST_HEAD(&grp
->task_list
);
1657 /* Second half of the array tracks nids where faults happen */
1658 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
1661 node_set(task_node(current
), grp
->active_nodes
);
1663 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1664 grp
->faults
[i
] = p
->numa_faults_memory
[i
];
1666 grp
->total_faults
= p
->total_numa_faults
;
1668 list_add(&p
->numa_entry
, &grp
->task_list
);
1670 rcu_assign_pointer(p
->numa_group
, grp
);
1674 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1676 if (!cpupid_match_pid(tsk
, cpupid
))
1679 grp
= rcu_dereference(tsk
->numa_group
);
1683 my_grp
= p
->numa_group
;
1688 * Only join the other group if its bigger; if we're the bigger group,
1689 * the other task will join us.
1691 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1695 * Tie-break on the grp address.
1697 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1700 /* Always join threads in the same process. */
1701 if (tsk
->mm
== current
->mm
)
1704 /* Simple filter to avoid false positives due to PID collisions */
1705 if (flags
& TNF_SHARED
)
1708 /* Update priv based on whether false sharing was detected */
1711 if (join
&& !get_numa_group(grp
))
1719 BUG_ON(irqs_disabled());
1720 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
1722 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
1723 my_grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1724 grp
->faults
[i
] += p
->numa_faults_memory
[i
];
1726 my_grp
->total_faults
-= p
->total_numa_faults
;
1727 grp
->total_faults
+= p
->total_numa_faults
;
1729 list_move(&p
->numa_entry
, &grp
->task_list
);
1733 spin_unlock(&my_grp
->lock
);
1734 spin_unlock_irq(&grp
->lock
);
1736 rcu_assign_pointer(p
->numa_group
, grp
);
1738 put_numa_group(my_grp
);
1746 void task_numa_free(struct task_struct
*p
)
1748 struct numa_group
*grp
= p
->numa_group
;
1750 void *numa_faults
= p
->numa_faults_memory
;
1753 spin_lock_irq(&grp
->lock
);
1754 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1755 grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1756 grp
->total_faults
-= p
->total_numa_faults
;
1758 list_del(&p
->numa_entry
);
1760 spin_unlock_irq(&grp
->lock
);
1761 rcu_assign_pointer(p
->numa_group
, NULL
);
1762 put_numa_group(grp
);
1765 p
->numa_faults_memory
= NULL
;
1766 p
->numa_faults_buffer_memory
= NULL
;
1767 p
->numa_faults_cpu
= NULL
;
1768 p
->numa_faults_buffer_cpu
= NULL
;
1773 * Got a PROT_NONE fault for a page on @node.
1775 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
1777 struct task_struct
*p
= current
;
1778 bool migrated
= flags
& TNF_MIGRATED
;
1779 int cpu_node
= task_node(current
);
1780 int local
= !!(flags
& TNF_FAULT_LOCAL
);
1783 if (!numabalancing_enabled
)
1786 /* for example, ksmd faulting in a user's mm */
1790 /* Do not worry about placement if exiting */
1791 if (p
->state
== TASK_DEAD
)
1794 /* Allocate buffer to track faults on a per-node basis */
1795 if (unlikely(!p
->numa_faults_memory
)) {
1796 int size
= sizeof(*p
->numa_faults_memory
) *
1797 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
1799 p
->numa_faults_memory
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
1800 if (!p
->numa_faults_memory
)
1803 BUG_ON(p
->numa_faults_buffer_memory
);
1805 * The averaged statistics, shared & private, memory & cpu,
1806 * occupy the first half of the array. The second half of the
1807 * array is for current counters, which are averaged into the
1808 * first set by task_numa_placement.
1810 p
->numa_faults_cpu
= p
->numa_faults_memory
+ (2 * nr_node_ids
);
1811 p
->numa_faults_buffer_memory
= p
->numa_faults_memory
+ (4 * nr_node_ids
);
1812 p
->numa_faults_buffer_cpu
= p
->numa_faults_memory
+ (6 * nr_node_ids
);
1813 p
->total_numa_faults
= 0;
1814 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1818 * First accesses are treated as private, otherwise consider accesses
1819 * to be private if the accessing pid has not changed
1821 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
1824 priv
= cpupid_match_pid(p
, last_cpupid
);
1825 if (!priv
&& !(flags
& TNF_NO_GROUP
))
1826 task_numa_group(p
, last_cpupid
, flags
, &priv
);
1830 * If a workload spans multiple NUMA nodes, a shared fault that
1831 * occurs wholly within the set of nodes that the workload is
1832 * actively using should be counted as local. This allows the
1833 * scan rate to slow down when a workload has settled down.
1835 if (!priv
&& !local
&& p
->numa_group
&&
1836 node_isset(cpu_node
, p
->numa_group
->active_nodes
) &&
1837 node_isset(mem_node
, p
->numa_group
->active_nodes
))
1840 task_numa_placement(p
);
1843 * Retry task to preferred node migration periodically, in case it
1844 * case it previously failed, or the scheduler moved us.
1846 if (time_after(jiffies
, p
->numa_migrate_retry
))
1847 numa_migrate_preferred(p
);
1850 p
->numa_pages_migrated
+= pages
;
1852 p
->numa_faults_buffer_memory
[task_faults_idx(mem_node
, priv
)] += pages
;
1853 p
->numa_faults_buffer_cpu
[task_faults_idx(cpu_node
, priv
)] += pages
;
1854 p
->numa_faults_locality
[local
] += pages
;
1857 static void reset_ptenuma_scan(struct task_struct
*p
)
1859 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
1860 p
->mm
->numa_scan_offset
= 0;
1864 * The expensive part of numa migration is done from task_work context.
1865 * Triggered from task_tick_numa().
1867 void task_numa_work(struct callback_head
*work
)
1869 unsigned long migrate
, next_scan
, now
= jiffies
;
1870 struct task_struct
*p
= current
;
1871 struct mm_struct
*mm
= p
->mm
;
1872 struct vm_area_struct
*vma
;
1873 unsigned long start
, end
;
1874 unsigned long nr_pte_updates
= 0;
1877 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
1879 work
->next
= work
; /* protect against double add */
1881 * Who cares about NUMA placement when they're dying.
1883 * NOTE: make sure not to dereference p->mm before this check,
1884 * exit_task_work() happens _after_ exit_mm() so we could be called
1885 * without p->mm even though we still had it when we enqueued this
1888 if (p
->flags
& PF_EXITING
)
1891 if (!mm
->numa_next_scan
) {
1892 mm
->numa_next_scan
= now
+
1893 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1897 * Enforce maximal scan/migration frequency..
1899 migrate
= mm
->numa_next_scan
;
1900 if (time_before(now
, migrate
))
1903 if (p
->numa_scan_period
== 0) {
1904 p
->numa_scan_period_max
= task_scan_max(p
);
1905 p
->numa_scan_period
= task_scan_min(p
);
1908 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1909 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
1913 * Delay this task enough that another task of this mm will likely win
1914 * the next time around.
1916 p
->node_stamp
+= 2 * TICK_NSEC
;
1918 start
= mm
->numa_scan_offset
;
1919 pages
= sysctl_numa_balancing_scan_size
;
1920 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
1924 down_read(&mm
->mmap_sem
);
1925 vma
= find_vma(mm
, start
);
1927 reset_ptenuma_scan(p
);
1931 for (; vma
; vma
= vma
->vm_next
) {
1932 if (!vma_migratable(vma
) || !vma_policy_mof(p
, vma
))
1936 * Shared library pages mapped by multiple processes are not
1937 * migrated as it is expected they are cache replicated. Avoid
1938 * hinting faults in read-only file-backed mappings or the vdso
1939 * as migrating the pages will be of marginal benefit.
1942 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
1946 * Skip inaccessible VMAs to avoid any confusion between
1947 * PROT_NONE and NUMA hinting ptes
1949 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
1953 start
= max(start
, vma
->vm_start
);
1954 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
1955 end
= min(end
, vma
->vm_end
);
1956 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
1959 * Scan sysctl_numa_balancing_scan_size but ensure that
1960 * at least one PTE is updated so that unused virtual
1961 * address space is quickly skipped.
1964 pages
-= (end
- start
) >> PAGE_SHIFT
;
1971 } while (end
!= vma
->vm_end
);
1976 * It is possible to reach the end of the VMA list but the last few
1977 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1978 * would find the !migratable VMA on the next scan but not reset the
1979 * scanner to the start so check it now.
1982 mm
->numa_scan_offset
= start
;
1984 reset_ptenuma_scan(p
);
1985 up_read(&mm
->mmap_sem
);
1989 * Drive the periodic memory faults..
1991 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1993 struct callback_head
*work
= &curr
->numa_work
;
1997 * We don't care about NUMA placement if we don't have memory.
1999 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2003 * Using runtime rather than walltime has the dual advantage that
2004 * we (mostly) drive the selection from busy threads and that the
2005 * task needs to have done some actual work before we bother with
2008 now
= curr
->se
.sum_exec_runtime
;
2009 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2011 if (now
- curr
->node_stamp
> period
) {
2012 if (!curr
->node_stamp
)
2013 curr
->numa_scan_period
= task_scan_min(curr
);
2014 curr
->node_stamp
+= period
;
2016 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2017 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2018 task_work_add(curr
, work
, true);
2023 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2027 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2031 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2034 #endif /* CONFIG_NUMA_BALANCING */
2037 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2039 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2040 if (!parent_entity(se
))
2041 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2043 if (entity_is_task(se
)) {
2044 struct rq
*rq
= rq_of(cfs_rq
);
2046 account_numa_enqueue(rq
, task_of(se
));
2047 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2050 cfs_rq
->nr_running
++;
2054 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2056 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2057 if (!parent_entity(se
))
2058 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2059 if (entity_is_task(se
)) {
2060 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2061 list_del_init(&se
->group_node
);
2063 cfs_rq
->nr_running
--;
2066 #ifdef CONFIG_FAIR_GROUP_SCHED
2068 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
2073 * Use this CPU's actual weight instead of the last load_contribution
2074 * to gain a more accurate current total weight. See
2075 * update_cfs_rq_load_contribution().
2077 tg_weight
= atomic_long_read(&tg
->load_avg
);
2078 tg_weight
-= cfs_rq
->tg_load_contrib
;
2079 tg_weight
+= cfs_rq
->load
.weight
;
2084 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2086 long tg_weight
, load
, shares
;
2088 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
2089 load
= cfs_rq
->load
.weight
;
2091 shares
= (tg
->shares
* load
);
2093 shares
/= tg_weight
;
2095 if (shares
< MIN_SHARES
)
2096 shares
= MIN_SHARES
;
2097 if (shares
> tg
->shares
)
2098 shares
= tg
->shares
;
2102 # else /* CONFIG_SMP */
2103 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2107 # endif /* CONFIG_SMP */
2108 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2109 unsigned long weight
)
2112 /* commit outstanding execution time */
2113 if (cfs_rq
->curr
== se
)
2114 update_curr(cfs_rq
);
2115 account_entity_dequeue(cfs_rq
, se
);
2118 update_load_set(&se
->load
, weight
);
2121 account_entity_enqueue(cfs_rq
, se
);
2124 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2126 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2128 struct task_group
*tg
;
2129 struct sched_entity
*se
;
2133 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2134 if (!se
|| throttled_hierarchy(cfs_rq
))
2137 if (likely(se
->load
.weight
== tg
->shares
))
2140 shares
= calc_cfs_shares(cfs_rq
, tg
);
2142 reweight_entity(cfs_rq_of(se
), se
, shares
);
2144 #else /* CONFIG_FAIR_GROUP_SCHED */
2145 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2148 #endif /* CONFIG_FAIR_GROUP_SCHED */
2152 * We choose a half-life close to 1 scheduling period.
2153 * Note: The tables below are dependent on this value.
2155 #define LOAD_AVG_PERIOD 32
2156 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2157 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2159 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2160 static const u32 runnable_avg_yN_inv
[] = {
2161 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2162 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2163 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2164 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2165 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2166 0x85aac367, 0x82cd8698,
2170 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2171 * over-estimates when re-combining.
2173 static const u32 runnable_avg_yN_sum
[] = {
2174 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2175 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2176 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2181 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2183 static __always_inline u64
decay_load(u64 val
, u64 n
)
2185 unsigned int local_n
;
2189 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2192 /* after bounds checking we can collapse to 32-bit */
2196 * As y^PERIOD = 1/2, we can combine
2197 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2198 * With a look-up table which covers k^n (n<PERIOD)
2200 * To achieve constant time decay_load.
2202 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2203 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2204 local_n
%= LOAD_AVG_PERIOD
;
2207 val
*= runnable_avg_yN_inv
[local_n
];
2208 /* We don't use SRR here since we always want to round down. */
2213 * For updates fully spanning n periods, the contribution to runnable
2214 * average will be: \Sum 1024*y^n
2216 * We can compute this reasonably efficiently by combining:
2217 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2219 static u32
__compute_runnable_contrib(u64 n
)
2223 if (likely(n
<= LOAD_AVG_PERIOD
))
2224 return runnable_avg_yN_sum
[n
];
2225 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2226 return LOAD_AVG_MAX
;
2228 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2230 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2231 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2233 n
-= LOAD_AVG_PERIOD
;
2234 } while (n
> LOAD_AVG_PERIOD
);
2236 contrib
= decay_load(contrib
, n
);
2237 return contrib
+ runnable_avg_yN_sum
[n
];
2241 * We can represent the historical contribution to runnable average as the
2242 * coefficients of a geometric series. To do this we sub-divide our runnable
2243 * history into segments of approximately 1ms (1024us); label the segment that
2244 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2246 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2248 * (now) (~1ms ago) (~2ms ago)
2250 * Let u_i denote the fraction of p_i that the entity was runnable.
2252 * We then designate the fractions u_i as our co-efficients, yielding the
2253 * following representation of historical load:
2254 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2256 * We choose y based on the with of a reasonably scheduling period, fixing:
2259 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2260 * approximately half as much as the contribution to load within the last ms
2263 * When a period "rolls over" and we have new u_0`, multiplying the previous
2264 * sum again by y is sufficient to update:
2265 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2266 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2268 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2269 struct sched_avg
*sa
,
2273 u32 runnable_contrib
;
2274 int delta_w
, decayed
= 0;
2276 delta
= now
- sa
->last_runnable_update
;
2278 * This should only happen when time goes backwards, which it
2279 * unfortunately does during sched clock init when we swap over to TSC.
2281 if ((s64
)delta
< 0) {
2282 sa
->last_runnable_update
= now
;
2287 * Use 1024ns as the unit of measurement since it's a reasonable
2288 * approximation of 1us and fast to compute.
2293 sa
->last_runnable_update
= now
;
2295 /* delta_w is the amount already accumulated against our next period */
2296 delta_w
= sa
->runnable_avg_period
% 1024;
2297 if (delta
+ delta_w
>= 1024) {
2298 /* period roll-over */
2302 * Now that we know we're crossing a period boundary, figure
2303 * out how much from delta we need to complete the current
2304 * period and accrue it.
2306 delta_w
= 1024 - delta_w
;
2308 sa
->runnable_avg_sum
+= delta_w
;
2309 sa
->runnable_avg_period
+= delta_w
;
2313 /* Figure out how many additional periods this update spans */
2314 periods
= delta
/ 1024;
2317 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2319 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2322 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2323 runnable_contrib
= __compute_runnable_contrib(periods
);
2325 sa
->runnable_avg_sum
+= runnable_contrib
;
2326 sa
->runnable_avg_period
+= runnable_contrib
;
2329 /* Remainder of delta accrued against u_0` */
2331 sa
->runnable_avg_sum
+= delta
;
2332 sa
->runnable_avg_period
+= delta
;
2337 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2338 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2340 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2341 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2343 decays
-= se
->avg
.decay_count
;
2347 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2348 se
->avg
.decay_count
= 0;
2353 #ifdef CONFIG_FAIR_GROUP_SCHED
2354 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2357 struct task_group
*tg
= cfs_rq
->tg
;
2360 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2361 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2363 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2364 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2365 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2370 * Aggregate cfs_rq runnable averages into an equivalent task_group
2371 * representation for computing load contributions.
2373 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2374 struct cfs_rq
*cfs_rq
)
2376 struct task_group
*tg
= cfs_rq
->tg
;
2379 /* The fraction of a cpu used by this cfs_rq */
2380 contrib
= div_u64((u64
)sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2381 sa
->runnable_avg_period
+ 1);
2382 contrib
-= cfs_rq
->tg_runnable_contrib
;
2384 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2385 atomic_add(contrib
, &tg
->runnable_avg
);
2386 cfs_rq
->tg_runnable_contrib
+= contrib
;
2390 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2392 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2393 struct task_group
*tg
= cfs_rq
->tg
;
2398 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2399 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2400 atomic_long_read(&tg
->load_avg
) + 1);
2403 * For group entities we need to compute a correction term in the case
2404 * that they are consuming <1 cpu so that we would contribute the same
2405 * load as a task of equal weight.
2407 * Explicitly co-ordinating this measurement would be expensive, but
2408 * fortunately the sum of each cpus contribution forms a usable
2409 * lower-bound on the true value.
2411 * Consider the aggregate of 2 contributions. Either they are disjoint
2412 * (and the sum represents true value) or they are disjoint and we are
2413 * understating by the aggregate of their overlap.
2415 * Extending this to N cpus, for a given overlap, the maximum amount we
2416 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2417 * cpus that overlap for this interval and w_i is the interval width.
2419 * On a small machine; the first term is well-bounded which bounds the
2420 * total error since w_i is a subset of the period. Whereas on a
2421 * larger machine, while this first term can be larger, if w_i is the
2422 * of consequential size guaranteed to see n_i*w_i quickly converge to
2423 * our upper bound of 1-cpu.
2425 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2426 if (runnable_avg
< NICE_0_LOAD
) {
2427 se
->avg
.load_avg_contrib
*= runnable_avg
;
2428 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2432 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2434 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2435 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2437 #else /* CONFIG_FAIR_GROUP_SCHED */
2438 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2439 int force_update
) {}
2440 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2441 struct cfs_rq
*cfs_rq
) {}
2442 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2443 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2444 #endif /* CONFIG_FAIR_GROUP_SCHED */
2446 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2450 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2451 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2452 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2453 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2456 /* Compute the current contribution to load_avg by se, return any delta */
2457 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2459 long old_contrib
= se
->avg
.load_avg_contrib
;
2461 if (entity_is_task(se
)) {
2462 __update_task_entity_contrib(se
);
2464 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2465 __update_group_entity_contrib(se
);
2468 return se
->avg
.load_avg_contrib
- old_contrib
;
2471 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2474 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2475 cfs_rq
->blocked_load_avg
-= load_contrib
;
2477 cfs_rq
->blocked_load_avg
= 0;
2480 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2482 /* Update a sched_entity's runnable average */
2483 static inline void update_entity_load_avg(struct sched_entity
*se
,
2486 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2491 * For a group entity we need to use their owned cfs_rq_clock_task() in
2492 * case they are the parent of a throttled hierarchy.
2494 if (entity_is_task(se
))
2495 now
= cfs_rq_clock_task(cfs_rq
);
2497 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2499 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2502 contrib_delta
= __update_entity_load_avg_contrib(se
);
2508 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2510 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2514 * Decay the load contributed by all blocked children and account this so that
2515 * their contribution may appropriately discounted when they wake up.
2517 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2519 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2522 decays
= now
- cfs_rq
->last_decay
;
2523 if (!decays
&& !force_update
)
2526 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2527 unsigned long removed_load
;
2528 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2529 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2533 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2535 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2536 cfs_rq
->last_decay
= now
;
2539 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2542 /* Add the load generated by se into cfs_rq's child load-average */
2543 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2544 struct sched_entity
*se
,
2548 * We track migrations using entity decay_count <= 0, on a wake-up
2549 * migration we use a negative decay count to track the remote decays
2550 * accumulated while sleeping.
2552 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2553 * are seen by enqueue_entity_load_avg() as a migration with an already
2554 * constructed load_avg_contrib.
2556 if (unlikely(se
->avg
.decay_count
<= 0)) {
2557 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2558 if (se
->avg
.decay_count
) {
2560 * In a wake-up migration we have to approximate the
2561 * time sleeping. This is because we can't synchronize
2562 * clock_task between the two cpus, and it is not
2563 * guaranteed to be read-safe. Instead, we can
2564 * approximate this using our carried decays, which are
2565 * explicitly atomically readable.
2567 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2569 update_entity_load_avg(se
, 0);
2570 /* Indicate that we're now synchronized and on-rq */
2571 se
->avg
.decay_count
= 0;
2575 __synchronize_entity_decay(se
);
2578 /* migrated tasks did not contribute to our blocked load */
2580 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2581 update_entity_load_avg(se
, 0);
2584 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2585 /* we force update consideration on load-balancer moves */
2586 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2590 * Remove se's load from this cfs_rq child load-average, if the entity is
2591 * transitioning to a blocked state we track its projected decay using
2594 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2595 struct sched_entity
*se
,
2598 update_entity_load_avg(se
, 1);
2599 /* we force update consideration on load-balancer moves */
2600 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2602 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2604 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2605 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2606 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2610 * Update the rq's load with the elapsed running time before entering
2611 * idle. if the last scheduled task is not a CFS task, idle_enter will
2612 * be the only way to update the runnable statistic.
2614 void idle_enter_fair(struct rq
*this_rq
)
2616 update_rq_runnable_avg(this_rq
, 1);
2620 * Update the rq's load with the elapsed idle time before a task is
2621 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2622 * be the only way to update the runnable statistic.
2624 void idle_exit_fair(struct rq
*this_rq
)
2626 update_rq_runnable_avg(this_rq
, 0);
2629 static int idle_balance(struct rq
*this_rq
);
2631 #else /* CONFIG_SMP */
2633 static inline void update_entity_load_avg(struct sched_entity
*se
,
2634 int update_cfs_rq
) {}
2635 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2636 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2637 struct sched_entity
*se
,
2639 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2640 struct sched_entity
*se
,
2642 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2643 int force_update
) {}
2645 static inline int idle_balance(struct rq
*rq
)
2650 #endif /* CONFIG_SMP */
2652 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2654 #ifdef CONFIG_SCHEDSTATS
2655 struct task_struct
*tsk
= NULL
;
2657 if (entity_is_task(se
))
2660 if (se
->statistics
.sleep_start
) {
2661 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2666 if (unlikely(delta
> se
->statistics
.sleep_max
))
2667 se
->statistics
.sleep_max
= delta
;
2669 se
->statistics
.sleep_start
= 0;
2670 se
->statistics
.sum_sleep_runtime
+= delta
;
2673 account_scheduler_latency(tsk
, delta
>> 10, 1);
2674 trace_sched_stat_sleep(tsk
, delta
);
2677 if (se
->statistics
.block_start
) {
2678 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2683 if (unlikely(delta
> se
->statistics
.block_max
))
2684 se
->statistics
.block_max
= delta
;
2686 se
->statistics
.block_start
= 0;
2687 se
->statistics
.sum_sleep_runtime
+= delta
;
2690 if (tsk
->in_iowait
) {
2691 se
->statistics
.iowait_sum
+= delta
;
2692 se
->statistics
.iowait_count
++;
2693 trace_sched_stat_iowait(tsk
, delta
);
2696 trace_sched_stat_blocked(tsk
, delta
);
2699 * Blocking time is in units of nanosecs, so shift by
2700 * 20 to get a milliseconds-range estimation of the
2701 * amount of time that the task spent sleeping:
2703 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2704 profile_hits(SLEEP_PROFILING
,
2705 (void *)get_wchan(tsk
),
2708 account_scheduler_latency(tsk
, delta
>> 10, 0);
2714 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2716 #ifdef CONFIG_SCHED_DEBUG
2717 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2722 if (d
> 3*sysctl_sched_latency
)
2723 schedstat_inc(cfs_rq
, nr_spread_over
);
2728 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2730 u64 vruntime
= cfs_rq
->min_vruntime
;
2733 * The 'current' period is already promised to the current tasks,
2734 * however the extra weight of the new task will slow them down a
2735 * little, place the new task so that it fits in the slot that
2736 * stays open at the end.
2738 if (initial
&& sched_feat(START_DEBIT
))
2739 vruntime
+= sched_vslice(cfs_rq
, se
);
2741 /* sleeps up to a single latency don't count. */
2743 unsigned long thresh
= sysctl_sched_latency
;
2746 * Halve their sleep time's effect, to allow
2747 * for a gentler effect of sleepers:
2749 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2755 /* ensure we never gain time by being placed backwards. */
2756 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
2759 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
2762 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2765 * Update the normalized vruntime before updating min_vruntime
2766 * through calling update_curr().
2768 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
2769 se
->vruntime
+= cfs_rq
->min_vruntime
;
2772 * Update run-time statistics of the 'current'.
2774 update_curr(cfs_rq
);
2775 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
2776 account_entity_enqueue(cfs_rq
, se
);
2777 update_cfs_shares(cfs_rq
);
2779 if (flags
& ENQUEUE_WAKEUP
) {
2780 place_entity(cfs_rq
, se
, 0);
2781 enqueue_sleeper(cfs_rq
, se
);
2784 update_stats_enqueue(cfs_rq
, se
);
2785 check_spread(cfs_rq
, se
);
2786 if (se
!= cfs_rq
->curr
)
2787 __enqueue_entity(cfs_rq
, se
);
2790 if (cfs_rq
->nr_running
== 1) {
2791 list_add_leaf_cfs_rq(cfs_rq
);
2792 check_enqueue_throttle(cfs_rq
);
2796 static void __clear_buddies_last(struct sched_entity
*se
)
2798 for_each_sched_entity(se
) {
2799 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2800 if (cfs_rq
->last
!= se
)
2803 cfs_rq
->last
= NULL
;
2807 static void __clear_buddies_next(struct sched_entity
*se
)
2809 for_each_sched_entity(se
) {
2810 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2811 if (cfs_rq
->next
!= se
)
2814 cfs_rq
->next
= NULL
;
2818 static void __clear_buddies_skip(struct sched_entity
*se
)
2820 for_each_sched_entity(se
) {
2821 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2822 if (cfs_rq
->skip
!= se
)
2825 cfs_rq
->skip
= NULL
;
2829 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2831 if (cfs_rq
->last
== se
)
2832 __clear_buddies_last(se
);
2834 if (cfs_rq
->next
== se
)
2835 __clear_buddies_next(se
);
2837 if (cfs_rq
->skip
== se
)
2838 __clear_buddies_skip(se
);
2841 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2844 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2847 * Update run-time statistics of the 'current'.
2849 update_curr(cfs_rq
);
2850 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
2852 update_stats_dequeue(cfs_rq
, se
);
2853 if (flags
& DEQUEUE_SLEEP
) {
2854 #ifdef CONFIG_SCHEDSTATS
2855 if (entity_is_task(se
)) {
2856 struct task_struct
*tsk
= task_of(se
);
2858 if (tsk
->state
& TASK_INTERRUPTIBLE
)
2859 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
2860 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
2861 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
2866 clear_buddies(cfs_rq
, se
);
2868 if (se
!= cfs_rq
->curr
)
2869 __dequeue_entity(cfs_rq
, se
);
2871 account_entity_dequeue(cfs_rq
, se
);
2874 * Normalize the entity after updating the min_vruntime because the
2875 * update can refer to the ->curr item and we need to reflect this
2876 * movement in our normalized position.
2878 if (!(flags
& DEQUEUE_SLEEP
))
2879 se
->vruntime
-= cfs_rq
->min_vruntime
;
2881 /* return excess runtime on last dequeue */
2882 return_cfs_rq_runtime(cfs_rq
);
2884 update_min_vruntime(cfs_rq
);
2885 update_cfs_shares(cfs_rq
);
2889 * Preempt the current task with a newly woken task if needed:
2892 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2894 unsigned long ideal_runtime
, delta_exec
;
2895 struct sched_entity
*se
;
2898 ideal_runtime
= sched_slice(cfs_rq
, curr
);
2899 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
2900 if (delta_exec
> ideal_runtime
) {
2901 resched_task(rq_of(cfs_rq
)->curr
);
2903 * The current task ran long enough, ensure it doesn't get
2904 * re-elected due to buddy favours.
2906 clear_buddies(cfs_rq
, curr
);
2911 * Ensure that a task that missed wakeup preemption by a
2912 * narrow margin doesn't have to wait for a full slice.
2913 * This also mitigates buddy induced latencies under load.
2915 if (delta_exec
< sysctl_sched_min_granularity
)
2918 se
= __pick_first_entity(cfs_rq
);
2919 delta
= curr
->vruntime
- se
->vruntime
;
2924 if (delta
> ideal_runtime
)
2925 resched_task(rq_of(cfs_rq
)->curr
);
2929 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2931 /* 'current' is not kept within the tree. */
2934 * Any task has to be enqueued before it get to execute on
2935 * a CPU. So account for the time it spent waiting on the
2938 update_stats_wait_end(cfs_rq
, se
);
2939 __dequeue_entity(cfs_rq
, se
);
2942 update_stats_curr_start(cfs_rq
, se
);
2944 #ifdef CONFIG_SCHEDSTATS
2946 * Track our maximum slice length, if the CPU's load is at
2947 * least twice that of our own weight (i.e. dont track it
2948 * when there are only lesser-weight tasks around):
2950 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
2951 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
2952 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
2955 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
2959 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
2962 * Pick the next process, keeping these things in mind, in this order:
2963 * 1) keep things fair between processes/task groups
2964 * 2) pick the "next" process, since someone really wants that to run
2965 * 3) pick the "last" process, for cache locality
2966 * 4) do not run the "skip" process, if something else is available
2968 static struct sched_entity
*
2969 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2971 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
2972 struct sched_entity
*se
;
2975 * If curr is set we have to see if its left of the leftmost entity
2976 * still in the tree, provided there was anything in the tree at all.
2978 if (!left
|| (curr
&& entity_before(curr
, left
)))
2981 se
= left
; /* ideally we run the leftmost entity */
2984 * Avoid running the skip buddy, if running something else can
2985 * be done without getting too unfair.
2987 if (cfs_rq
->skip
== se
) {
2988 struct sched_entity
*second
;
2991 second
= __pick_first_entity(cfs_rq
);
2993 second
= __pick_next_entity(se
);
2994 if (!second
|| (curr
&& entity_before(curr
, second
)))
2998 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3003 * Prefer last buddy, try to return the CPU to a preempted task.
3005 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3009 * Someone really wants this to run. If it's not unfair, run it.
3011 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3014 clear_buddies(cfs_rq
, se
);
3019 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3021 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3024 * If still on the runqueue then deactivate_task()
3025 * was not called and update_curr() has to be done:
3028 update_curr(cfs_rq
);
3030 /* throttle cfs_rqs exceeding runtime */
3031 check_cfs_rq_runtime(cfs_rq
);
3033 check_spread(cfs_rq
, prev
);
3035 update_stats_wait_start(cfs_rq
, prev
);
3036 /* Put 'current' back into the tree. */
3037 __enqueue_entity(cfs_rq
, prev
);
3038 /* in !on_rq case, update occurred at dequeue */
3039 update_entity_load_avg(prev
, 1);
3041 cfs_rq
->curr
= NULL
;
3045 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3048 * Update run-time statistics of the 'current'.
3050 update_curr(cfs_rq
);
3053 * Ensure that runnable average is periodically updated.
3055 update_entity_load_avg(curr
, 1);
3056 update_cfs_rq_blocked_load(cfs_rq
, 1);
3057 update_cfs_shares(cfs_rq
);
3059 #ifdef CONFIG_SCHED_HRTICK
3061 * queued ticks are scheduled to match the slice, so don't bother
3062 * validating it and just reschedule.
3065 resched_task(rq_of(cfs_rq
)->curr
);
3069 * don't let the period tick interfere with the hrtick preemption
3071 if (!sched_feat(DOUBLE_TICK
) &&
3072 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3076 if (cfs_rq
->nr_running
> 1)
3077 check_preempt_tick(cfs_rq
, curr
);
3081 /**************************************************
3082 * CFS bandwidth control machinery
3085 #ifdef CONFIG_CFS_BANDWIDTH
3087 #ifdef HAVE_JUMP_LABEL
3088 static struct static_key __cfs_bandwidth_used
;
3090 static inline bool cfs_bandwidth_used(void)
3092 return static_key_false(&__cfs_bandwidth_used
);
3095 void cfs_bandwidth_usage_inc(void)
3097 static_key_slow_inc(&__cfs_bandwidth_used
);
3100 void cfs_bandwidth_usage_dec(void)
3102 static_key_slow_dec(&__cfs_bandwidth_used
);
3104 #else /* HAVE_JUMP_LABEL */
3105 static bool cfs_bandwidth_used(void)
3110 void cfs_bandwidth_usage_inc(void) {}
3111 void cfs_bandwidth_usage_dec(void) {}
3112 #endif /* HAVE_JUMP_LABEL */
3115 * default period for cfs group bandwidth.
3116 * default: 0.1s, units: nanoseconds
3118 static inline u64
default_cfs_period(void)
3120 return 100000000ULL;
3123 static inline u64
sched_cfs_bandwidth_slice(void)
3125 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3129 * Replenish runtime according to assigned quota and update expiration time.
3130 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3131 * additional synchronization around rq->lock.
3133 * requires cfs_b->lock
3135 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3139 if (cfs_b
->quota
== RUNTIME_INF
)
3142 now
= sched_clock_cpu(smp_processor_id());
3143 cfs_b
->runtime
= cfs_b
->quota
;
3144 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3147 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3149 return &tg
->cfs_bandwidth
;
3152 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3153 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3155 if (unlikely(cfs_rq
->throttle_count
))
3156 return cfs_rq
->throttled_clock_task
;
3158 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3161 /* returns 0 on failure to allocate runtime */
3162 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3164 struct task_group
*tg
= cfs_rq
->tg
;
3165 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3166 u64 amount
= 0, min_amount
, expires
;
3168 /* note: this is a positive sum as runtime_remaining <= 0 */
3169 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3171 raw_spin_lock(&cfs_b
->lock
);
3172 if (cfs_b
->quota
== RUNTIME_INF
)
3173 amount
= min_amount
;
3176 * If the bandwidth pool has become inactive, then at least one
3177 * period must have elapsed since the last consumption.
3178 * Refresh the global state and ensure bandwidth timer becomes
3181 if (!cfs_b
->timer_active
) {
3182 __refill_cfs_bandwidth_runtime(cfs_b
);
3183 __start_cfs_bandwidth(cfs_b
);
3186 if (cfs_b
->runtime
> 0) {
3187 amount
= min(cfs_b
->runtime
, min_amount
);
3188 cfs_b
->runtime
-= amount
;
3192 expires
= cfs_b
->runtime_expires
;
3193 raw_spin_unlock(&cfs_b
->lock
);
3195 cfs_rq
->runtime_remaining
+= amount
;
3197 * we may have advanced our local expiration to account for allowed
3198 * spread between our sched_clock and the one on which runtime was
3201 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3202 cfs_rq
->runtime_expires
= expires
;
3204 return cfs_rq
->runtime_remaining
> 0;
3208 * Note: This depends on the synchronization provided by sched_clock and the
3209 * fact that rq->clock snapshots this value.
3211 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3213 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3215 /* if the deadline is ahead of our clock, nothing to do */
3216 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3219 if (cfs_rq
->runtime_remaining
< 0)
3223 * If the local deadline has passed we have to consider the
3224 * possibility that our sched_clock is 'fast' and the global deadline
3225 * has not truly expired.
3227 * Fortunately we can check determine whether this the case by checking
3228 * whether the global deadline has advanced. It is valid to compare
3229 * cfs_b->runtime_expires without any locks since we only care about
3230 * exact equality, so a partial write will still work.
3233 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
3234 /* extend local deadline, drift is bounded above by 2 ticks */
3235 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3237 /* global deadline is ahead, expiration has passed */
3238 cfs_rq
->runtime_remaining
= 0;
3242 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3244 /* dock delta_exec before expiring quota (as it could span periods) */
3245 cfs_rq
->runtime_remaining
-= delta_exec
;
3246 expire_cfs_rq_runtime(cfs_rq
);
3248 if (likely(cfs_rq
->runtime_remaining
> 0))
3252 * if we're unable to extend our runtime we resched so that the active
3253 * hierarchy can be throttled
3255 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3256 resched_task(rq_of(cfs_rq
)->curr
);
3259 static __always_inline
3260 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3262 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3265 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3268 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3270 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3273 /* check whether cfs_rq, or any parent, is throttled */
3274 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3276 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3280 * Ensure that neither of the group entities corresponding to src_cpu or
3281 * dest_cpu are members of a throttled hierarchy when performing group
3282 * load-balance operations.
3284 static inline int throttled_lb_pair(struct task_group
*tg
,
3285 int src_cpu
, int dest_cpu
)
3287 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3289 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3290 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3292 return throttled_hierarchy(src_cfs_rq
) ||
3293 throttled_hierarchy(dest_cfs_rq
);
3296 /* updated child weight may affect parent so we have to do this bottom up */
3297 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3299 struct rq
*rq
= data
;
3300 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3302 cfs_rq
->throttle_count
--;
3304 if (!cfs_rq
->throttle_count
) {
3305 /* adjust cfs_rq_clock_task() */
3306 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3307 cfs_rq
->throttled_clock_task
;
3314 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3316 struct rq
*rq
= data
;
3317 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3319 /* group is entering throttled state, stop time */
3320 if (!cfs_rq
->throttle_count
)
3321 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3322 cfs_rq
->throttle_count
++;
3327 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3329 struct rq
*rq
= rq_of(cfs_rq
);
3330 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3331 struct sched_entity
*se
;
3332 long task_delta
, dequeue
= 1;
3334 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3336 /* freeze hierarchy runnable averages while throttled */
3338 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3341 task_delta
= cfs_rq
->h_nr_running
;
3342 for_each_sched_entity(se
) {
3343 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3344 /* throttled entity or throttle-on-deactivate */
3349 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3350 qcfs_rq
->h_nr_running
-= task_delta
;
3352 if (qcfs_rq
->load
.weight
)
3357 sub_nr_running(rq
, task_delta
);
3359 cfs_rq
->throttled
= 1;
3360 cfs_rq
->throttled_clock
= rq_clock(rq
);
3361 raw_spin_lock(&cfs_b
->lock
);
3362 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3363 if (!cfs_b
->timer_active
)
3364 __start_cfs_bandwidth(cfs_b
);
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
;
3419 u64 runtime
= remaining
;
3422 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3424 struct rq
*rq
= rq_of(cfs_rq
);
3426 raw_spin_lock(&rq
->lock
);
3427 if (!cfs_rq_throttled(cfs_rq
))
3430 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3431 if (runtime
> remaining
)
3432 runtime
= remaining
;
3433 remaining
-= runtime
;
3435 cfs_rq
->runtime_remaining
+= runtime
;
3436 cfs_rq
->runtime_expires
= expires
;
3438 /* we check whether we're throttled above */
3439 if (cfs_rq
->runtime_remaining
> 0)
3440 unthrottle_cfs_rq(cfs_rq
);
3443 raw_spin_unlock(&rq
->lock
);
3454 * Responsible for refilling a task_group's bandwidth and unthrottling its
3455 * cfs_rqs as appropriate. If there has been no activity within the last
3456 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3457 * used to track this state.
3459 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3461 u64 runtime
, runtime_expires
;
3464 /* no need to continue the timer with no bandwidth constraint */
3465 if (cfs_b
->quota
== RUNTIME_INF
)
3466 goto out_deactivate
;
3468 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3469 cfs_b
->nr_periods
+= overrun
;
3472 * idle depends on !throttled (for the case of a large deficit), and if
3473 * we're going inactive then everything else can be deferred
3475 if (cfs_b
->idle
&& !throttled
)
3476 goto out_deactivate
;
3479 * if we have relooped after returning idle once, we need to update our
3480 * status as actually running, so that other cpus doing
3481 * __start_cfs_bandwidth will stop trying to cancel us.
3483 cfs_b
->timer_active
= 1;
3485 __refill_cfs_bandwidth_runtime(cfs_b
);
3488 /* mark as potentially idle for the upcoming period */
3493 /* account preceding periods in which throttling occurred */
3494 cfs_b
->nr_throttled
+= overrun
;
3497 * There are throttled entities so we must first use the new bandwidth
3498 * to unthrottle them before making it generally available. This
3499 * ensures that all existing debts will be paid before a new cfs_rq is
3502 runtime
= cfs_b
->runtime
;
3503 runtime_expires
= cfs_b
->runtime_expires
;
3507 * This check is repeated as we are holding onto the new bandwidth
3508 * while we unthrottle. This can potentially race with an unthrottled
3509 * group trying to acquire new bandwidth from the global pool.
3511 while (throttled
&& runtime
> 0) {
3512 raw_spin_unlock(&cfs_b
->lock
);
3513 /* we can't nest cfs_b->lock while distributing bandwidth */
3514 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3516 raw_spin_lock(&cfs_b
->lock
);
3518 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3521 /* return (any) remaining runtime */
3522 cfs_b
->runtime
= runtime
;
3524 * While we are ensured activity in the period following an
3525 * unthrottle, this also covers the case in which the new bandwidth is
3526 * insufficient to cover the existing bandwidth deficit. (Forcing the
3527 * timer to remain active while there are any throttled entities.)
3534 cfs_b
->timer_active
= 0;
3538 /* a cfs_rq won't donate quota below this amount */
3539 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3540 /* minimum remaining period time to redistribute slack quota */
3541 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3542 /* how long we wait to gather additional slack before distributing */
3543 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3546 * Are we near the end of the current quota period?
3548 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3549 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3550 * migrate_hrtimers, base is never cleared, so we are fine.
3552 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3554 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3557 /* if the call-back is running a quota refresh is already occurring */
3558 if (hrtimer_callback_running(refresh_timer
))
3561 /* is a quota refresh about to occur? */
3562 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3563 if (remaining
< min_expire
)
3569 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3571 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3573 /* if there's a quota refresh soon don't bother with slack */
3574 if (runtime_refresh_within(cfs_b
, min_left
))
3577 start_bandwidth_timer(&cfs_b
->slack_timer
,
3578 ns_to_ktime(cfs_bandwidth_slack_period
));
3581 /* we know any runtime found here is valid as update_curr() precedes return */
3582 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3584 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3585 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3587 if (slack_runtime
<= 0)
3590 raw_spin_lock(&cfs_b
->lock
);
3591 if (cfs_b
->quota
!= RUNTIME_INF
&&
3592 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3593 cfs_b
->runtime
+= slack_runtime
;
3595 /* we are under rq->lock, defer unthrottling using a timer */
3596 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3597 !list_empty(&cfs_b
->throttled_cfs_rq
))
3598 start_cfs_slack_bandwidth(cfs_b
);
3600 raw_spin_unlock(&cfs_b
->lock
);
3602 /* even if it's not valid for return we don't want to try again */
3603 cfs_rq
->runtime_remaining
-= slack_runtime
;
3606 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3608 if (!cfs_bandwidth_used())
3611 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3614 __return_cfs_rq_runtime(cfs_rq
);
3618 * This is done with a timer (instead of inline with bandwidth return) since
3619 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3621 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3623 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3626 /* confirm we're still not at a refresh boundary */
3627 raw_spin_lock(&cfs_b
->lock
);
3628 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
3629 raw_spin_unlock(&cfs_b
->lock
);
3633 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
) {
3634 runtime
= cfs_b
->runtime
;
3637 expires
= cfs_b
->runtime_expires
;
3638 raw_spin_unlock(&cfs_b
->lock
);
3643 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3645 raw_spin_lock(&cfs_b
->lock
);
3646 if (expires
== cfs_b
->runtime_expires
)
3647 cfs_b
->runtime
= runtime
;
3648 raw_spin_unlock(&cfs_b
->lock
);
3652 * When a group wakes up we want to make sure that its quota is not already
3653 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3654 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3656 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3658 if (!cfs_bandwidth_used())
3661 /* an active group must be handled by the update_curr()->put() path */
3662 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3665 /* ensure the group is not already throttled */
3666 if (cfs_rq_throttled(cfs_rq
))
3669 /* update runtime allocation */
3670 account_cfs_rq_runtime(cfs_rq
, 0);
3671 if (cfs_rq
->runtime_remaining
<= 0)
3672 throttle_cfs_rq(cfs_rq
);
3675 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3676 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3678 if (!cfs_bandwidth_used())
3681 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3685 * it's possible for a throttled entity to be forced into a running
3686 * state (e.g. set_curr_task), in this case we're finished.
3688 if (cfs_rq_throttled(cfs_rq
))
3691 throttle_cfs_rq(cfs_rq
);
3695 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3697 struct cfs_bandwidth
*cfs_b
=
3698 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3699 do_sched_cfs_slack_timer(cfs_b
);
3701 return HRTIMER_NORESTART
;
3704 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3706 struct cfs_bandwidth
*cfs_b
=
3707 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3712 raw_spin_lock(&cfs_b
->lock
);
3714 now
= hrtimer_cb_get_time(timer
);
3715 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3720 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3722 raw_spin_unlock(&cfs_b
->lock
);
3724 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3727 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3729 raw_spin_lock_init(&cfs_b
->lock
);
3731 cfs_b
->quota
= RUNTIME_INF
;
3732 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3734 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3735 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3736 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3737 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3738 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3741 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3743 cfs_rq
->runtime_enabled
= 0;
3744 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3747 /* requires cfs_b->lock, may release to reprogram timer */
3748 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3751 * The timer may be active because we're trying to set a new bandwidth
3752 * period or because we're racing with the tear-down path
3753 * (timer_active==0 becomes visible before the hrtimer call-back
3754 * terminates). In either case we ensure that it's re-programmed
3756 while (unlikely(hrtimer_active(&cfs_b
->period_timer
)) &&
3757 hrtimer_try_to_cancel(&cfs_b
->period_timer
) < 0) {
3758 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3759 raw_spin_unlock(&cfs_b
->lock
);
3761 raw_spin_lock(&cfs_b
->lock
);
3762 /* if someone else restarted the timer then we're done */
3763 if (cfs_b
->timer_active
)
3767 cfs_b
->timer_active
= 1;
3768 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
3771 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3773 hrtimer_cancel(&cfs_b
->period_timer
);
3774 hrtimer_cancel(&cfs_b
->slack_timer
);
3777 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
3779 struct cfs_rq
*cfs_rq
;
3781 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3782 if (!cfs_rq
->runtime_enabled
)
3786 * clock_task is not advancing so we just need to make sure
3787 * there's some valid quota amount
3789 cfs_rq
->runtime_remaining
= 1;
3790 if (cfs_rq_throttled(cfs_rq
))
3791 unthrottle_cfs_rq(cfs_rq
);
3795 #else /* CONFIG_CFS_BANDWIDTH */
3796 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3798 return rq_clock_task(rq_of(cfs_rq
));
3801 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
3802 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
3803 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
3804 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3806 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3811 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3816 static inline int throttled_lb_pair(struct task_group
*tg
,
3817 int src_cpu
, int dest_cpu
)
3822 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3824 #ifdef CONFIG_FAIR_GROUP_SCHED
3825 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3828 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3832 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3833 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
3835 #endif /* CONFIG_CFS_BANDWIDTH */
3837 /**************************************************
3838 * CFS operations on tasks:
3841 #ifdef CONFIG_SCHED_HRTICK
3842 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3844 struct sched_entity
*se
= &p
->se
;
3845 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3847 WARN_ON(task_rq(p
) != rq
);
3849 if (cfs_rq
->nr_running
> 1) {
3850 u64 slice
= sched_slice(cfs_rq
, se
);
3851 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
3852 s64 delta
= slice
- ran
;
3861 * Don't schedule slices shorter than 10000ns, that just
3862 * doesn't make sense. Rely on vruntime for fairness.
3865 delta
= max_t(s64
, 10000LL, delta
);
3867 hrtick_start(rq
, delta
);
3872 * called from enqueue/dequeue and updates the hrtick when the
3873 * current task is from our class and nr_running is low enough
3876 static void hrtick_update(struct rq
*rq
)
3878 struct task_struct
*curr
= rq
->curr
;
3880 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
3883 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
3884 hrtick_start_fair(rq
, curr
);
3886 #else /* !CONFIG_SCHED_HRTICK */
3888 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3892 static inline void hrtick_update(struct rq
*rq
)
3898 * The enqueue_task method is called before nr_running is
3899 * increased. Here we update the fair scheduling stats and
3900 * then put the task into the rbtree:
3903 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3905 struct cfs_rq
*cfs_rq
;
3906 struct sched_entity
*se
= &p
->se
;
3908 for_each_sched_entity(se
) {
3911 cfs_rq
= cfs_rq_of(se
);
3912 enqueue_entity(cfs_rq
, se
, flags
);
3915 * end evaluation on encountering a throttled cfs_rq
3917 * note: in the case of encountering a throttled cfs_rq we will
3918 * post the final h_nr_running increment below.
3920 if (cfs_rq_throttled(cfs_rq
))
3922 cfs_rq
->h_nr_running
++;
3924 flags
= ENQUEUE_WAKEUP
;
3927 for_each_sched_entity(se
) {
3928 cfs_rq
= cfs_rq_of(se
);
3929 cfs_rq
->h_nr_running
++;
3931 if (cfs_rq_throttled(cfs_rq
))
3934 update_cfs_shares(cfs_rq
);
3935 update_entity_load_avg(se
, 1);
3939 update_rq_runnable_avg(rq
, rq
->nr_running
);
3940 add_nr_running(rq
, 1);
3945 static void set_next_buddy(struct sched_entity
*se
);
3948 * The dequeue_task method is called before nr_running is
3949 * decreased. We remove the task from the rbtree and
3950 * update the fair scheduling stats:
3952 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3954 struct cfs_rq
*cfs_rq
;
3955 struct sched_entity
*se
= &p
->se
;
3956 int task_sleep
= flags
& DEQUEUE_SLEEP
;
3958 for_each_sched_entity(se
) {
3959 cfs_rq
= cfs_rq_of(se
);
3960 dequeue_entity(cfs_rq
, se
, flags
);
3963 * end evaluation on encountering a throttled cfs_rq
3965 * note: in the case of encountering a throttled cfs_rq we will
3966 * post the final h_nr_running decrement below.
3968 if (cfs_rq_throttled(cfs_rq
))
3970 cfs_rq
->h_nr_running
--;
3972 /* Don't dequeue parent if it has other entities besides us */
3973 if (cfs_rq
->load
.weight
) {
3975 * Bias pick_next to pick a task from this cfs_rq, as
3976 * p is sleeping when it is within its sched_slice.
3978 if (task_sleep
&& parent_entity(se
))
3979 set_next_buddy(parent_entity(se
));
3981 /* avoid re-evaluating load for this entity */
3982 se
= parent_entity(se
);
3985 flags
|= DEQUEUE_SLEEP
;
3988 for_each_sched_entity(se
) {
3989 cfs_rq
= cfs_rq_of(se
);
3990 cfs_rq
->h_nr_running
--;
3992 if (cfs_rq_throttled(cfs_rq
))
3995 update_cfs_shares(cfs_rq
);
3996 update_entity_load_avg(se
, 1);
4000 sub_nr_running(rq
, 1);
4001 update_rq_runnable_avg(rq
, 1);
4007 /* Used instead of source_load when we know the type == 0 */
4008 static unsigned long weighted_cpuload(const int cpu
)
4010 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
4014 * Return a low guess at the load of a migration-source cpu weighted
4015 * according to the scheduling class and "nice" value.
4017 * We want to under-estimate the load of migration sources, to
4018 * balance conservatively.
4020 static unsigned long source_load(int cpu
, int type
)
4022 struct rq
*rq
= cpu_rq(cpu
);
4023 unsigned long total
= weighted_cpuload(cpu
);
4025 if (type
== 0 || !sched_feat(LB_BIAS
))
4028 return min(rq
->cpu_load
[type
-1], total
);
4032 * Return a high guess at the load of a migration-target cpu weighted
4033 * according to the scheduling class and "nice" value.
4035 static unsigned long target_load(int cpu
, int type
)
4037 struct rq
*rq
= cpu_rq(cpu
);
4038 unsigned long total
= weighted_cpuload(cpu
);
4040 if (type
== 0 || !sched_feat(LB_BIAS
))
4043 return max(rq
->cpu_load
[type
-1], total
);
4046 static unsigned long power_of(int cpu
)
4048 return cpu_rq(cpu
)->cpu_power
;
4051 static unsigned long cpu_avg_load_per_task(int cpu
)
4053 struct rq
*rq
= cpu_rq(cpu
);
4054 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
4055 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
4058 return load_avg
/ nr_running
;
4063 static void record_wakee(struct task_struct
*p
)
4066 * Rough decay (wiping) for cost saving, don't worry
4067 * about the boundary, really active task won't care
4070 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
4071 current
->wakee_flips
>>= 1;
4072 current
->wakee_flip_decay_ts
= jiffies
;
4075 if (current
->last_wakee
!= p
) {
4076 current
->last_wakee
= p
;
4077 current
->wakee_flips
++;
4081 static void task_waking_fair(struct task_struct
*p
)
4083 struct sched_entity
*se
= &p
->se
;
4084 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4087 #ifndef CONFIG_64BIT
4088 u64 min_vruntime_copy
;
4091 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
4093 min_vruntime
= cfs_rq
->min_vruntime
;
4094 } while (min_vruntime
!= min_vruntime_copy
);
4096 min_vruntime
= cfs_rq
->min_vruntime
;
4099 se
->vruntime
-= min_vruntime
;
4103 #ifdef CONFIG_FAIR_GROUP_SCHED
4105 * effective_load() calculates the load change as seen from the root_task_group
4107 * Adding load to a group doesn't make a group heavier, but can cause movement
4108 * of group shares between cpus. Assuming the shares were perfectly aligned one
4109 * can calculate the shift in shares.
4111 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4112 * on this @cpu and results in a total addition (subtraction) of @wg to the
4113 * total group weight.
4115 * Given a runqueue weight distribution (rw_i) we can compute a shares
4116 * distribution (s_i) using:
4118 * s_i = rw_i / \Sum rw_j (1)
4120 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4121 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4122 * shares distribution (s_i):
4124 * rw_i = { 2, 4, 1, 0 }
4125 * s_i = { 2/7, 4/7, 1/7, 0 }
4127 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4128 * task used to run on and the CPU the waker is running on), we need to
4129 * compute the effect of waking a task on either CPU and, in case of a sync
4130 * wakeup, compute the effect of the current task going to sleep.
4132 * So for a change of @wl to the local @cpu with an overall group weight change
4133 * of @wl we can compute the new shares distribution (s'_i) using:
4135 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4137 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4138 * differences in waking a task to CPU 0. The additional task changes the
4139 * weight and shares distributions like:
4141 * rw'_i = { 3, 4, 1, 0 }
4142 * s'_i = { 3/8, 4/8, 1/8, 0 }
4144 * We can then compute the difference in effective weight by using:
4146 * dw_i = S * (s'_i - s_i) (3)
4148 * Where 'S' is the group weight as seen by its parent.
4150 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4151 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4152 * 4/7) times the weight of the group.
4154 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4156 struct sched_entity
*se
= tg
->se
[cpu
];
4158 if (!tg
->parent
) /* the trivial, non-cgroup case */
4161 for_each_sched_entity(se
) {
4167 * W = @wg + \Sum rw_j
4169 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4174 w
= se
->my_q
->load
.weight
+ wl
;
4177 * wl = S * s'_i; see (2)
4180 wl
= (w
* tg
->shares
) / W
;
4185 * Per the above, wl is the new se->load.weight value; since
4186 * those are clipped to [MIN_SHARES, ...) do so now. See
4187 * calc_cfs_shares().
4189 if (wl
< MIN_SHARES
)
4193 * wl = dw_i = S * (s'_i - s_i); see (3)
4195 wl
-= se
->load
.weight
;
4198 * Recursively apply this logic to all parent groups to compute
4199 * the final effective load change on the root group. Since
4200 * only the @tg group gets extra weight, all parent groups can
4201 * only redistribute existing shares. @wl is the shift in shares
4202 * resulting from this level per the above.
4211 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4218 static int wake_wide(struct task_struct
*p
)
4220 int factor
= this_cpu_read(sd_llc_size
);
4223 * Yeah, it's the switching-frequency, could means many wakee or
4224 * rapidly switch, use factor here will just help to automatically
4225 * adjust the loose-degree, so bigger node will lead to more pull.
4227 if (p
->wakee_flips
> factor
) {
4229 * wakee is somewhat hot, it needs certain amount of cpu
4230 * resource, so if waker is far more hot, prefer to leave
4233 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
4240 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4242 s64 this_load
, load
;
4243 int idx
, this_cpu
, prev_cpu
;
4244 unsigned long tl_per_task
;
4245 struct task_group
*tg
;
4246 unsigned long weight
;
4250 * If we wake multiple tasks be careful to not bounce
4251 * ourselves around too much.
4257 this_cpu
= smp_processor_id();
4258 prev_cpu
= task_cpu(p
);
4259 load
= source_load(prev_cpu
, idx
);
4260 this_load
= target_load(this_cpu
, idx
);
4263 * If sync wakeup then subtract the (maximum possible)
4264 * effect of the currently running task from the load
4265 * of the current CPU:
4268 tg
= task_group(current
);
4269 weight
= current
->se
.load
.weight
;
4271 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4272 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4276 weight
= p
->se
.load
.weight
;
4279 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4280 * due to the sync cause above having dropped this_load to 0, we'll
4281 * always have an imbalance, but there's really nothing you can do
4282 * about that, so that's good too.
4284 * Otherwise check if either cpus are near enough in load to allow this
4285 * task to be woken on this_cpu.
4287 if (this_load
> 0) {
4288 s64 this_eff_load
, prev_eff_load
;
4290 this_eff_load
= 100;
4291 this_eff_load
*= power_of(prev_cpu
);
4292 this_eff_load
*= this_load
+
4293 effective_load(tg
, this_cpu
, weight
, weight
);
4295 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4296 prev_eff_load
*= power_of(this_cpu
);
4297 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4299 balanced
= this_eff_load
<= prev_eff_load
;
4304 * If the currently running task will sleep within
4305 * a reasonable amount of time then attract this newly
4308 if (sync
&& balanced
)
4311 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4312 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
4315 (this_load
<= load
&&
4316 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
4318 * This domain has SD_WAKE_AFFINE and
4319 * p is cache cold in this domain, and
4320 * there is no bad imbalance.
4322 schedstat_inc(sd
, ttwu_move_affine
);
4323 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4331 * find_idlest_group finds and returns the least busy CPU group within the
4334 static struct sched_group
*
4335 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4336 int this_cpu
, int sd_flag
)
4338 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4339 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4340 int load_idx
= sd
->forkexec_idx
;
4341 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4343 if (sd_flag
& SD_BALANCE_WAKE
)
4344 load_idx
= sd
->wake_idx
;
4347 unsigned long load
, avg_load
;
4351 /* Skip over this group if it has no CPUs allowed */
4352 if (!cpumask_intersects(sched_group_cpus(group
),
4353 tsk_cpus_allowed(p
)))
4356 local_group
= cpumask_test_cpu(this_cpu
,
4357 sched_group_cpus(group
));
4359 /* Tally up the load of all CPUs in the group */
4362 for_each_cpu(i
, sched_group_cpus(group
)) {
4363 /* Bias balancing toward cpus of our domain */
4365 load
= source_load(i
, load_idx
);
4367 load
= target_load(i
, load_idx
);
4372 /* Adjust by relative CPU power of the group */
4373 avg_load
= (avg_load
* SCHED_POWER_SCALE
) / group
->sgp
->power
;
4376 this_load
= avg_load
;
4377 } else if (avg_load
< min_load
) {
4378 min_load
= avg_load
;
4381 } while (group
= group
->next
, group
!= sd
->groups
);
4383 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4389 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4392 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4394 unsigned long load
, min_load
= ULONG_MAX
;
4398 /* Traverse only the allowed CPUs */
4399 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4400 load
= weighted_cpuload(i
);
4402 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4412 * Try and locate an idle CPU in the sched_domain.
4414 static int select_idle_sibling(struct task_struct
*p
, int target
)
4416 struct sched_domain
*sd
;
4417 struct sched_group
*sg
;
4418 int i
= task_cpu(p
);
4420 if (idle_cpu(target
))
4424 * If the prevous cpu is cache affine and idle, don't be stupid.
4426 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4430 * Otherwise, iterate the domains and find an elegible idle cpu.
4432 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4433 for_each_lower_domain(sd
) {
4436 if (!cpumask_intersects(sched_group_cpus(sg
),
4437 tsk_cpus_allowed(p
)))
4440 for_each_cpu(i
, sched_group_cpus(sg
)) {
4441 if (i
== target
|| !idle_cpu(i
))
4445 target
= cpumask_first_and(sched_group_cpus(sg
),
4446 tsk_cpus_allowed(p
));
4450 } while (sg
!= sd
->groups
);
4457 * select_task_rq_fair: Select target runqueue for the waking task in domains
4458 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4459 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4461 * Balances load by selecting the idlest cpu in the idlest group, or under
4462 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4464 * Returns the target cpu number.
4466 * preempt must be disabled.
4469 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4471 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4472 int cpu
= smp_processor_id();
4474 int want_affine
= 0;
4475 int sync
= wake_flags
& WF_SYNC
;
4477 if (p
->nr_cpus_allowed
== 1)
4480 if (sd_flag
& SD_BALANCE_WAKE
) {
4481 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
4487 for_each_domain(cpu
, tmp
) {
4488 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4492 * If both cpu and prev_cpu are part of this domain,
4493 * cpu is a valid SD_WAKE_AFFINE target.
4495 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4496 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4501 if (tmp
->flags
& sd_flag
)
4505 if (affine_sd
&& cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4508 if (sd_flag
& SD_BALANCE_WAKE
) {
4509 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4514 struct sched_group
*group
;
4517 if (!(sd
->flags
& sd_flag
)) {
4522 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
4528 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4529 if (new_cpu
== -1 || new_cpu
== cpu
) {
4530 /* Now try balancing at a lower domain level of cpu */
4535 /* Now try balancing at a lower domain level of new_cpu */
4537 weight
= sd
->span_weight
;
4539 for_each_domain(cpu
, tmp
) {
4540 if (weight
<= tmp
->span_weight
)
4542 if (tmp
->flags
& sd_flag
)
4545 /* while loop will break here if sd == NULL */
4554 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4555 * cfs_rq_of(p) references at time of call are still valid and identify the
4556 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4557 * other assumptions, including the state of rq->lock, should be made.
4560 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4562 struct sched_entity
*se
= &p
->se
;
4563 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4566 * Load tracking: accumulate removed load so that it can be processed
4567 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4568 * to blocked load iff they have a positive decay-count. It can never
4569 * be negative here since on-rq tasks have decay-count == 0.
4571 if (se
->avg
.decay_count
) {
4572 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4573 atomic_long_add(se
->avg
.load_avg_contrib
,
4574 &cfs_rq
->removed_load
);
4577 /* We have migrated, no longer consider this task hot */
4580 #endif /* CONFIG_SMP */
4582 static unsigned long
4583 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4585 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4588 * Since its curr running now, convert the gran from real-time
4589 * to virtual-time in his units.
4591 * By using 'se' instead of 'curr' we penalize light tasks, so
4592 * they get preempted easier. That is, if 'se' < 'curr' then
4593 * the resulting gran will be larger, therefore penalizing the
4594 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4595 * be smaller, again penalizing the lighter task.
4597 * This is especially important for buddies when the leftmost
4598 * task is higher priority than the buddy.
4600 return calc_delta_fair(gran
, se
);
4604 * Should 'se' preempt 'curr'.
4618 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4620 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4625 gran
= wakeup_gran(curr
, se
);
4632 static void set_last_buddy(struct sched_entity
*se
)
4634 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4637 for_each_sched_entity(se
)
4638 cfs_rq_of(se
)->last
= se
;
4641 static void set_next_buddy(struct sched_entity
*se
)
4643 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4646 for_each_sched_entity(se
)
4647 cfs_rq_of(se
)->next
= se
;
4650 static void set_skip_buddy(struct sched_entity
*se
)
4652 for_each_sched_entity(se
)
4653 cfs_rq_of(se
)->skip
= se
;
4657 * Preempt the current task with a newly woken task if needed:
4659 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4661 struct task_struct
*curr
= rq
->curr
;
4662 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4663 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4664 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4665 int next_buddy_marked
= 0;
4667 if (unlikely(se
== pse
))
4671 * This is possible from callers such as move_task(), in which we
4672 * unconditionally check_prempt_curr() after an enqueue (which may have
4673 * lead to a throttle). This both saves work and prevents false
4674 * next-buddy nomination below.
4676 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4679 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4680 set_next_buddy(pse
);
4681 next_buddy_marked
= 1;
4685 * We can come here with TIF_NEED_RESCHED already set from new task
4688 * Note: this also catches the edge-case of curr being in a throttled
4689 * group (e.g. via set_curr_task), since update_curr() (in the
4690 * enqueue of curr) will have resulted in resched being set. This
4691 * prevents us from potentially nominating it as a false LAST_BUDDY
4694 if (test_tsk_need_resched(curr
))
4697 /* Idle tasks are by definition preempted by non-idle tasks. */
4698 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4699 likely(p
->policy
!= SCHED_IDLE
))
4703 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4704 * is driven by the tick):
4706 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4709 find_matching_se(&se
, &pse
);
4710 update_curr(cfs_rq_of(se
));
4712 if (wakeup_preempt_entity(se
, pse
) == 1) {
4714 * Bias pick_next to pick the sched entity that is
4715 * triggering this preemption.
4717 if (!next_buddy_marked
)
4718 set_next_buddy(pse
);
4727 * Only set the backward buddy when the current task is still
4728 * on the rq. This can happen when a wakeup gets interleaved
4729 * with schedule on the ->pre_schedule() or idle_balance()
4730 * point, either of which can * drop the rq lock.
4732 * Also, during early boot the idle thread is in the fair class,
4733 * for obvious reasons its a bad idea to schedule back to it.
4735 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4738 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
4742 static struct task_struct
*
4743 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4745 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
4746 struct sched_entity
*se
;
4747 struct task_struct
*p
;
4751 #ifdef CONFIG_FAIR_GROUP_SCHED
4752 if (!cfs_rq
->nr_running
)
4755 if (prev
->sched_class
!= &fair_sched_class
)
4759 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4760 * likely that a next task is from the same cgroup as the current.
4762 * Therefore attempt to avoid putting and setting the entire cgroup
4763 * hierarchy, only change the part that actually changes.
4767 struct sched_entity
*curr
= cfs_rq
->curr
;
4770 * Since we got here without doing put_prev_entity() we also
4771 * have to consider cfs_rq->curr. If it is still a runnable
4772 * entity, update_curr() will update its vruntime, otherwise
4773 * forget we've ever seen it.
4775 if (curr
&& curr
->on_rq
)
4776 update_curr(cfs_rq
);
4781 * This call to check_cfs_rq_runtime() will do the throttle and
4782 * dequeue its entity in the parent(s). Therefore the 'simple'
4783 * nr_running test will indeed be correct.
4785 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
4788 se
= pick_next_entity(cfs_rq
, curr
);
4789 cfs_rq
= group_cfs_rq(se
);
4795 * Since we haven't yet done put_prev_entity and if the selected task
4796 * is a different task than we started out with, try and touch the
4797 * least amount of cfs_rqs.
4800 struct sched_entity
*pse
= &prev
->se
;
4802 while (!(cfs_rq
= is_same_group(se
, pse
))) {
4803 int se_depth
= se
->depth
;
4804 int pse_depth
= pse
->depth
;
4806 if (se_depth
<= pse_depth
) {
4807 put_prev_entity(cfs_rq_of(pse
), pse
);
4808 pse
= parent_entity(pse
);
4810 if (se_depth
>= pse_depth
) {
4811 set_next_entity(cfs_rq_of(se
), se
);
4812 se
= parent_entity(se
);
4816 put_prev_entity(cfs_rq
, pse
);
4817 set_next_entity(cfs_rq
, se
);
4820 if (hrtick_enabled(rq
))
4821 hrtick_start_fair(rq
, p
);
4828 if (!cfs_rq
->nr_running
)
4831 put_prev_task(rq
, prev
);
4834 se
= pick_next_entity(cfs_rq
, NULL
);
4835 set_next_entity(cfs_rq
, se
);
4836 cfs_rq
= group_cfs_rq(se
);
4841 if (hrtick_enabled(rq
))
4842 hrtick_start_fair(rq
, p
);
4847 new_tasks
= idle_balance(rq
);
4849 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4850 * possible for any higher priority task to appear. In that case we
4851 * must re-start the pick_next_entity() loop.
4863 * Account for a descheduled task:
4865 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4867 struct sched_entity
*se
= &prev
->se
;
4868 struct cfs_rq
*cfs_rq
;
4870 for_each_sched_entity(se
) {
4871 cfs_rq
= cfs_rq_of(se
);
4872 put_prev_entity(cfs_rq
, se
);
4877 * sched_yield() is very simple
4879 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4881 static void yield_task_fair(struct rq
*rq
)
4883 struct task_struct
*curr
= rq
->curr
;
4884 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4885 struct sched_entity
*se
= &curr
->se
;
4888 * Are we the only task in the tree?
4890 if (unlikely(rq
->nr_running
== 1))
4893 clear_buddies(cfs_rq
, se
);
4895 if (curr
->policy
!= SCHED_BATCH
) {
4896 update_rq_clock(rq
);
4898 * Update run-time statistics of the 'current'.
4900 update_curr(cfs_rq
);
4902 * Tell update_rq_clock() that we've just updated,
4903 * so we don't do microscopic update in schedule()
4904 * and double the fastpath cost.
4906 rq
->skip_clock_update
= 1;
4912 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
4914 struct sched_entity
*se
= &p
->se
;
4916 /* throttled hierarchies are not runnable */
4917 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
4920 /* Tell the scheduler that we'd really like pse to run next. */
4923 yield_task_fair(rq
);
4929 /**************************************************
4930 * Fair scheduling class load-balancing methods.
4934 * The purpose of load-balancing is to achieve the same basic fairness the
4935 * per-cpu scheduler provides, namely provide a proportional amount of compute
4936 * time to each task. This is expressed in the following equation:
4938 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4940 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4941 * W_i,0 is defined as:
4943 * W_i,0 = \Sum_j w_i,j (2)
4945 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4946 * is derived from the nice value as per prio_to_weight[].
4948 * The weight average is an exponential decay average of the instantaneous
4951 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4953 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4954 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4955 * can also include other factors [XXX].
4957 * To achieve this balance we define a measure of imbalance which follows
4958 * directly from (1):
4960 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4962 * We them move tasks around to minimize the imbalance. In the continuous
4963 * function space it is obvious this converges, in the discrete case we get
4964 * a few fun cases generally called infeasible weight scenarios.
4967 * - infeasible weights;
4968 * - local vs global optima in the discrete case. ]
4973 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4974 * for all i,j solution, we create a tree of cpus that follows the hardware
4975 * topology where each level pairs two lower groups (or better). This results
4976 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4977 * tree to only the first of the previous level and we decrease the frequency
4978 * of load-balance at each level inv. proportional to the number of cpus in
4984 * \Sum { --- * --- * 2^i } = O(n) (5)
4986 * `- size of each group
4987 * | | `- number of cpus doing load-balance
4989 * `- sum over all levels
4991 * Coupled with a limit on how many tasks we can migrate every balance pass,
4992 * this makes (5) the runtime complexity of the balancer.
4994 * An important property here is that each CPU is still (indirectly) connected
4995 * to every other cpu in at most O(log n) steps:
4997 * The adjacency matrix of the resulting graph is given by:
5000 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5003 * And you'll find that:
5005 * A^(log_2 n)_i,j != 0 for all i,j (7)
5007 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5008 * The task movement gives a factor of O(m), giving a convergence complexity
5011 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5016 * In order to avoid CPUs going idle while there's still work to do, new idle
5017 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5018 * tree itself instead of relying on other CPUs to bring it work.
5020 * This adds some complexity to both (5) and (8) but it reduces the total idle
5028 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5031 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5036 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5038 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5040 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5043 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5044 * rewrite all of this once again.]
5047 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
5049 enum fbq_type
{ regular
, remote
, all
};
5051 #define LBF_ALL_PINNED 0x01
5052 #define LBF_NEED_BREAK 0x02
5053 #define LBF_DST_PINNED 0x04
5054 #define LBF_SOME_PINNED 0x08
5057 struct sched_domain
*sd
;
5065 struct cpumask
*dst_grpmask
;
5067 enum cpu_idle_type idle
;
5069 /* The set of CPUs under consideration for load-balancing */
5070 struct cpumask
*cpus
;
5075 unsigned int loop_break
;
5076 unsigned int loop_max
;
5078 enum fbq_type fbq_type
;
5082 * move_task - move a task from one runqueue to another runqueue.
5083 * Both runqueues must be locked.
5085 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
5087 deactivate_task(env
->src_rq
, p
, 0);
5088 set_task_cpu(p
, env
->dst_cpu
);
5089 activate_task(env
->dst_rq
, p
, 0);
5090 check_preempt_curr(env
->dst_rq
, p
, 0);
5094 * Is this task likely cache-hot:
5097 task_hot(struct task_struct
*p
, u64 now
)
5101 if (p
->sched_class
!= &fair_sched_class
)
5104 if (unlikely(p
->policy
== SCHED_IDLE
))
5108 * Buddy candidates are cache hot:
5110 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
5111 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
5112 &p
->se
== cfs_rq_of(&p
->se
)->last
))
5115 if (sysctl_sched_migration_cost
== -1)
5117 if (sysctl_sched_migration_cost
== 0)
5120 delta
= now
- p
->se
.exec_start
;
5122 return delta
< (s64
)sysctl_sched_migration_cost
;
5125 #ifdef CONFIG_NUMA_BALANCING
5126 /* Returns true if the destination node has incurred more faults */
5127 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
5129 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5130 int src_nid
, dst_nid
;
5132 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults_memory
||
5133 !(env
->sd
->flags
& SD_NUMA
)) {
5137 src_nid
= cpu_to_node(env
->src_cpu
);
5138 dst_nid
= cpu_to_node(env
->dst_cpu
);
5140 if (src_nid
== dst_nid
)
5144 /* Task is already in the group's interleave set. */
5145 if (node_isset(src_nid
, numa_group
->active_nodes
))
5148 /* Task is moving into the group's interleave set. */
5149 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5152 return group_faults(p
, dst_nid
) > group_faults(p
, src_nid
);
5155 /* Encourage migration to the preferred node. */
5156 if (dst_nid
== p
->numa_preferred_nid
)
5159 return task_faults(p
, dst_nid
) > task_faults(p
, src_nid
);
5163 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
5165 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5166 int src_nid
, dst_nid
;
5168 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
5171 if (!p
->numa_faults_memory
|| !(env
->sd
->flags
& SD_NUMA
))
5174 src_nid
= cpu_to_node(env
->src_cpu
);
5175 dst_nid
= cpu_to_node(env
->dst_cpu
);
5177 if (src_nid
== dst_nid
)
5181 /* Task is moving within/into the group's interleave set. */
5182 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5185 /* Task is moving out of the group's interleave set. */
5186 if (node_isset(src_nid
, numa_group
->active_nodes
))
5189 return group_faults(p
, dst_nid
) < group_faults(p
, src_nid
);
5192 /* Migrating away from the preferred node is always bad. */
5193 if (src_nid
== p
->numa_preferred_nid
)
5196 return task_faults(p
, dst_nid
) < task_faults(p
, src_nid
);
5200 static inline bool migrate_improves_locality(struct task_struct
*p
,
5206 static inline bool migrate_degrades_locality(struct task_struct
*p
,
5214 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5217 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
5219 int tsk_cache_hot
= 0;
5221 * We do not migrate tasks that are:
5222 * 1) throttled_lb_pair, or
5223 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5224 * 3) running (obviously), or
5225 * 4) are cache-hot on their current CPU.
5227 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
5230 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
5233 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
5235 env
->flags
|= LBF_SOME_PINNED
;
5238 * Remember if this task can be migrated to any other cpu in
5239 * our sched_group. We may want to revisit it if we couldn't
5240 * meet load balance goals by pulling other tasks on src_cpu.
5242 * Also avoid computing new_dst_cpu if we have already computed
5243 * one in current iteration.
5245 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
5248 /* Prevent to re-select dst_cpu via env's cpus */
5249 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
5250 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
5251 env
->flags
|= LBF_DST_PINNED
;
5252 env
->new_dst_cpu
= cpu
;
5260 /* Record that we found atleast one task that could run on dst_cpu */
5261 env
->flags
&= ~LBF_ALL_PINNED
;
5263 if (task_running(env
->src_rq
, p
)) {
5264 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
5269 * Aggressive migration if:
5270 * 1) destination numa is preferred
5271 * 2) task is cache cold, or
5272 * 3) too many balance attempts have failed.
5274 tsk_cache_hot
= task_hot(p
, rq_clock_task(env
->src_rq
));
5276 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
5278 if (migrate_improves_locality(p
, env
)) {
5279 #ifdef CONFIG_SCHEDSTATS
5280 if (tsk_cache_hot
) {
5281 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5282 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5288 if (!tsk_cache_hot
||
5289 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
5291 if (tsk_cache_hot
) {
5292 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5293 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5299 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
5304 * move_one_task tries to move exactly one task from busiest to this_rq, as
5305 * part of active balancing operations within "domain".
5306 * Returns 1 if successful and 0 otherwise.
5308 * Called with both runqueues locked.
5310 static int move_one_task(struct lb_env
*env
)
5312 struct task_struct
*p
, *n
;
5314 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
5315 if (!can_migrate_task(p
, env
))
5320 * Right now, this is only the second place move_task()
5321 * is called, so we can safely collect move_task()
5322 * stats here rather than inside move_task().
5324 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
5330 static const unsigned int sched_nr_migrate_break
= 32;
5333 * move_tasks tries to move up to imbalance weighted load from busiest to
5334 * this_rq, as part of a balancing operation within domain "sd".
5335 * Returns 1 if successful and 0 otherwise.
5337 * Called with both runqueues locked.
5339 static int move_tasks(struct lb_env
*env
)
5341 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
5342 struct task_struct
*p
;
5346 if (env
->imbalance
<= 0)
5349 while (!list_empty(tasks
)) {
5350 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5353 /* We've more or less seen every task there is, call it quits */
5354 if (env
->loop
> env
->loop_max
)
5357 /* take a breather every nr_migrate tasks */
5358 if (env
->loop
> env
->loop_break
) {
5359 env
->loop_break
+= sched_nr_migrate_break
;
5360 env
->flags
|= LBF_NEED_BREAK
;
5364 if (!can_migrate_task(p
, env
))
5367 load
= task_h_load(p
);
5369 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5372 if ((load
/ 2) > env
->imbalance
)
5377 env
->imbalance
-= load
;
5379 #ifdef CONFIG_PREEMPT
5381 * NEWIDLE balancing is a source of latency, so preemptible
5382 * kernels will stop after the first task is pulled to minimize
5383 * the critical section.
5385 if (env
->idle
== CPU_NEWLY_IDLE
)
5390 * We only want to steal up to the prescribed amount of
5393 if (env
->imbalance
<= 0)
5398 list_move_tail(&p
->se
.group_node
, tasks
);
5402 * Right now, this is one of only two places move_task() is called,
5403 * so we can safely collect move_task() stats here rather than
5404 * inside move_task().
5406 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
5411 #ifdef CONFIG_FAIR_GROUP_SCHED
5413 * update tg->load_weight by folding this cpu's load_avg
5415 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5417 struct sched_entity
*se
= tg
->se
[cpu
];
5418 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5420 /* throttled entities do not contribute to load */
5421 if (throttled_hierarchy(cfs_rq
))
5424 update_cfs_rq_blocked_load(cfs_rq
, 1);
5427 update_entity_load_avg(se
, 1);
5429 * We pivot on our runnable average having decayed to zero for
5430 * list removal. This generally implies that all our children
5431 * have also been removed (modulo rounding error or bandwidth
5432 * control); however, such cases are rare and we can fix these
5435 * TODO: fix up out-of-order children on enqueue.
5437 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5438 list_del_leaf_cfs_rq(cfs_rq
);
5440 struct rq
*rq
= rq_of(cfs_rq
);
5441 update_rq_runnable_avg(rq
, rq
->nr_running
);
5445 static void update_blocked_averages(int cpu
)
5447 struct rq
*rq
= cpu_rq(cpu
);
5448 struct cfs_rq
*cfs_rq
;
5449 unsigned long flags
;
5451 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5452 update_rq_clock(rq
);
5454 * Iterates the task_group tree in a bottom up fashion, see
5455 * list_add_leaf_cfs_rq() for details.
5457 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5459 * Note: We may want to consider periodically releasing
5460 * rq->lock about these updates so that creating many task
5461 * groups does not result in continually extending hold time.
5463 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5466 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5470 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5471 * This needs to be done in a top-down fashion because the load of a child
5472 * group is a fraction of its parents load.
5474 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5476 struct rq
*rq
= rq_of(cfs_rq
);
5477 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5478 unsigned long now
= jiffies
;
5481 if (cfs_rq
->last_h_load_update
== now
)
5484 cfs_rq
->h_load_next
= NULL
;
5485 for_each_sched_entity(se
) {
5486 cfs_rq
= cfs_rq_of(se
);
5487 cfs_rq
->h_load_next
= se
;
5488 if (cfs_rq
->last_h_load_update
== now
)
5493 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5494 cfs_rq
->last_h_load_update
= now
;
5497 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5498 load
= cfs_rq
->h_load
;
5499 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5500 cfs_rq
->runnable_load_avg
+ 1);
5501 cfs_rq
= group_cfs_rq(se
);
5502 cfs_rq
->h_load
= load
;
5503 cfs_rq
->last_h_load_update
= now
;
5507 static unsigned long task_h_load(struct task_struct
*p
)
5509 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5511 update_cfs_rq_h_load(cfs_rq
);
5512 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5513 cfs_rq
->runnable_load_avg
+ 1);
5516 static inline void update_blocked_averages(int cpu
)
5520 static unsigned long task_h_load(struct task_struct
*p
)
5522 return p
->se
.avg
.load_avg_contrib
;
5526 /********** Helpers for find_busiest_group ************************/
5528 * sg_lb_stats - stats of a sched_group required for load_balancing
5530 struct sg_lb_stats
{
5531 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5532 unsigned long group_load
; /* Total load over the CPUs of the group */
5533 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5534 unsigned long load_per_task
;
5535 unsigned long group_power
;
5536 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5537 unsigned int group_capacity_factor
;
5538 unsigned int idle_cpus
;
5539 unsigned int group_weight
;
5540 int group_imb
; /* Is there an imbalance in the group ? */
5541 int group_has_free_capacity
;
5542 #ifdef CONFIG_NUMA_BALANCING
5543 unsigned int nr_numa_running
;
5544 unsigned int nr_preferred_running
;
5549 * sd_lb_stats - Structure to store the statistics of a sched_domain
5550 * during load balancing.
5552 struct sd_lb_stats
{
5553 struct sched_group
*busiest
; /* Busiest group in this sd */
5554 struct sched_group
*local
; /* Local group in this sd */
5555 unsigned long total_load
; /* Total load of all groups in sd */
5556 unsigned long total_pwr
; /* Total power of all groups in sd */
5557 unsigned long avg_load
; /* Average load across all groups in sd */
5559 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5560 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5563 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5566 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5567 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5568 * We must however clear busiest_stat::avg_load because
5569 * update_sd_pick_busiest() reads this before assignment.
5571 *sds
= (struct sd_lb_stats
){
5583 * get_sd_load_idx - Obtain the load index for a given sched domain.
5584 * @sd: The sched_domain whose load_idx is to be obtained.
5585 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5587 * Return: The load index.
5589 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5590 enum cpu_idle_type idle
)
5596 load_idx
= sd
->busy_idx
;
5599 case CPU_NEWLY_IDLE
:
5600 load_idx
= sd
->newidle_idx
;
5603 load_idx
= sd
->idle_idx
;
5610 static unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
5612 return SCHED_POWER_SCALE
;
5615 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
5617 return default_scale_freq_power(sd
, cpu
);
5620 static unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
5622 unsigned long weight
= sd
->span_weight
;
5623 unsigned long smt_gain
= sd
->smt_gain
;
5630 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
5632 return default_scale_smt_power(sd
, cpu
);
5635 static unsigned long scale_rt_power(int cpu
)
5637 struct rq
*rq
= cpu_rq(cpu
);
5638 u64 total
, available
, age_stamp
, avg
;
5642 * Since we're reading these variables without serialization make sure
5643 * we read them once before doing sanity checks on them.
5645 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5646 avg
= ACCESS_ONCE(rq
->rt_avg
);
5648 delta
= rq_clock(rq
) - age_stamp
;
5649 if (unlikely(delta
< 0))
5652 total
= sched_avg_period() + delta
;
5654 if (unlikely(total
< avg
)) {
5655 /* Ensures that power won't end up being negative */
5658 available
= total
- avg
;
5661 if (unlikely((s64
)total
< SCHED_POWER_SCALE
))
5662 total
= SCHED_POWER_SCALE
;
5664 total
>>= SCHED_POWER_SHIFT
;
5666 return div_u64(available
, total
);
5669 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
5671 unsigned long weight
= sd
->span_weight
;
5672 unsigned long power
= SCHED_POWER_SCALE
;
5673 struct sched_group
*sdg
= sd
->groups
;
5675 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
5676 if (sched_feat(ARCH_POWER
))
5677 power
*= arch_scale_smt_power(sd
, cpu
);
5679 power
*= default_scale_smt_power(sd
, cpu
);
5681 power
>>= SCHED_POWER_SHIFT
;
5684 sdg
->sgp
->power_orig
= power
;
5686 if (sched_feat(ARCH_POWER
))
5687 power
*= arch_scale_freq_power(sd
, cpu
);
5689 power
*= default_scale_freq_power(sd
, cpu
);
5691 power
>>= SCHED_POWER_SHIFT
;
5693 power
*= scale_rt_power(cpu
);
5694 power
>>= SCHED_POWER_SHIFT
;
5699 cpu_rq(cpu
)->cpu_power
= power
;
5700 sdg
->sgp
->power
= power
;
5703 void update_group_power(struct sched_domain
*sd
, int cpu
)
5705 struct sched_domain
*child
= sd
->child
;
5706 struct sched_group
*group
, *sdg
= sd
->groups
;
5707 unsigned long power
, power_orig
;
5708 unsigned long interval
;
5710 interval
= msecs_to_jiffies(sd
->balance_interval
);
5711 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5712 sdg
->sgp
->next_update
= jiffies
+ interval
;
5715 update_cpu_power(sd
, cpu
);
5719 power_orig
= power
= 0;
5721 if (child
->flags
& SD_OVERLAP
) {
5723 * SD_OVERLAP domains cannot assume that child groups
5724 * span the current group.
5727 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
5728 struct sched_group_power
*sgp
;
5729 struct rq
*rq
= cpu_rq(cpu
);
5732 * build_sched_domains() -> init_sched_groups_power()
5733 * gets here before we've attached the domains to the
5736 * Use power_of(), which is set irrespective of domains
5737 * in update_cpu_power().
5739 * This avoids power/power_orig from being 0 and
5740 * causing divide-by-zero issues on boot.
5742 * Runtime updates will correct power_orig.
5744 if (unlikely(!rq
->sd
)) {
5745 power_orig
+= power_of(cpu
);
5746 power
+= power_of(cpu
);
5750 sgp
= rq
->sd
->groups
->sgp
;
5751 power_orig
+= sgp
->power_orig
;
5752 power
+= sgp
->power
;
5756 * !SD_OVERLAP domains can assume that child groups
5757 * span the current group.
5760 group
= child
->groups
;
5762 power_orig
+= group
->sgp
->power_orig
;
5763 power
+= group
->sgp
->power
;
5764 group
= group
->next
;
5765 } while (group
!= child
->groups
);
5768 sdg
->sgp
->power_orig
= power_orig
;
5769 sdg
->sgp
->power
= power
;
5773 * Try and fix up capacity for tiny siblings, this is needed when
5774 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5775 * which on its own isn't powerful enough.
5777 * See update_sd_pick_busiest() and check_asym_packing().
5780 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
5783 * Only siblings can have significantly less than SCHED_POWER_SCALE
5785 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
5789 * If ~90% of the cpu_power is still there, we're good.
5791 if (group
->sgp
->power
* 32 > group
->sgp
->power_orig
* 29)
5798 * Group imbalance indicates (and tries to solve) the problem where balancing
5799 * groups is inadequate due to tsk_cpus_allowed() constraints.
5801 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5802 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5805 * { 0 1 2 3 } { 4 5 6 7 }
5808 * If we were to balance group-wise we'd place two tasks in the first group and
5809 * two tasks in the second group. Clearly this is undesired as it will overload
5810 * cpu 3 and leave one of the cpus in the second group unused.
5812 * The current solution to this issue is detecting the skew in the first group
5813 * by noticing the lower domain failed to reach balance and had difficulty
5814 * moving tasks due to affinity constraints.
5816 * When this is so detected; this group becomes a candidate for busiest; see
5817 * update_sd_pick_busiest(). And calculate_imbalance() and
5818 * find_busiest_group() avoid some of the usual balance conditions to allow it
5819 * to create an effective group imbalance.
5821 * This is a somewhat tricky proposition since the next run might not find the
5822 * group imbalance and decide the groups need to be balanced again. A most
5823 * subtle and fragile situation.
5826 static inline int sg_imbalanced(struct sched_group
*group
)
5828 return group
->sgp
->imbalance
;
5832 * Compute the group capacity factor.
5834 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5835 * first dividing out the smt factor and computing the actual number of cores
5836 * and limit power unit capacity with that.
5838 static inline int sg_capacity_factor(struct lb_env
*env
, struct sched_group
*group
)
5840 unsigned int capacity_factor
, smt
, cpus
;
5841 unsigned int power
, power_orig
;
5843 power
= group
->sgp
->power
;
5844 power_orig
= group
->sgp
->power_orig
;
5845 cpus
= group
->group_weight
;
5847 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5848 smt
= DIV_ROUND_UP(SCHED_POWER_SCALE
* cpus
, power_orig
);
5849 capacity_factor
= cpus
/ smt
; /* cores */
5851 capacity_factor
= min_t(unsigned, capacity_factor
, DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
));
5852 if (!capacity_factor
)
5853 capacity_factor
= fix_small_capacity(env
->sd
, group
);
5855 return capacity_factor
;
5859 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5860 * @env: The load balancing environment.
5861 * @group: sched_group whose statistics are to be updated.
5862 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5863 * @local_group: Does group contain this_cpu.
5864 * @sgs: variable to hold the statistics for this group.
5866 static inline void update_sg_lb_stats(struct lb_env
*env
,
5867 struct sched_group
*group
, int load_idx
,
5868 int local_group
, struct sg_lb_stats
*sgs
)
5873 memset(sgs
, 0, sizeof(*sgs
));
5875 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5876 struct rq
*rq
= cpu_rq(i
);
5878 /* Bias balancing toward cpus of our domain */
5880 load
= target_load(i
, load_idx
);
5882 load
= source_load(i
, load_idx
);
5884 sgs
->group_load
+= load
;
5885 sgs
->sum_nr_running
+= rq
->nr_running
;
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 power of the group */
5896 sgs
->group_power
= group
->sgp
->power
;
5897 sgs
->avg_load
= (sgs
->group_load
*SCHED_POWER_SCALE
) / sgs
->group_power
;
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;
5997 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
6000 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
6003 struct sg_lb_stats
*sgs
= &tmp_sgs
;
6006 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
6009 sgs
= &sds
->local_stat
;
6011 if (env
->idle
!= CPU_NEWLY_IDLE
||
6012 time_after_eq(jiffies
, sg
->sgp
->next_update
))
6013 update_group_power(env
->sd
, env
->dst_cpu
);
6016 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
);
6022 * In case the child domain prefers tasks go to siblings
6023 * first, lower the sg capacity factor to one so that we'll try
6024 * and move all the excess tasks away. We lower the capacity
6025 * of a group only if the local group has the capacity to fit
6026 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6027 * extra check prevents the case where you always pull from the
6028 * heaviest group when it is already under-utilized (possible
6029 * with a large weight task outweighs the tasks on the system).
6031 if (prefer_sibling
&& sds
->local
&&
6032 sds
->local_stat
.group_has_free_capacity
)
6033 sgs
->group_capacity_factor
= min(sgs
->group_capacity_factor
, 1U);
6035 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
6037 sds
->busiest_stat
= *sgs
;
6041 /* Now, start updating sd_lb_stats */
6042 sds
->total_load
+= sgs
->group_load
;
6043 sds
->total_pwr
+= sgs
->group_power
;
6046 } while (sg
!= env
->sd
->groups
);
6048 if (env
->sd
->flags
& SD_NUMA
)
6049 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
6053 * check_asym_packing - Check to see if the group is packed into the
6056 * This is primarily intended to used at the sibling level. Some
6057 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6058 * case of POWER7, it can move to lower SMT modes only when higher
6059 * threads are idle. When in lower SMT modes, the threads will
6060 * perform better since they share less core resources. Hence when we
6061 * have idle threads, we want them to be the higher ones.
6063 * This packing function is run on idle threads. It checks to see if
6064 * the busiest CPU in this domain (core in the P7 case) has a higher
6065 * CPU number than the packing function is being run on. Here we are
6066 * assuming lower CPU number will be equivalent to lower a SMT thread
6069 * Return: 1 when packing is required and a task should be moved to
6070 * this CPU. The amount of the imbalance is returned in *imbalance.
6072 * @env: The load balancing environment.
6073 * @sds: Statistics of the sched_domain which is to be packed
6075 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6079 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6085 busiest_cpu
= group_first_cpu(sds
->busiest
);
6086 if (env
->dst_cpu
> busiest_cpu
)
6089 env
->imbalance
= DIV_ROUND_CLOSEST(
6090 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_power
,
6097 * fix_small_imbalance - Calculate the minor imbalance that exists
6098 * amongst the groups of a sched_domain, during
6100 * @env: The load balancing environment.
6101 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6104 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6106 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
6107 unsigned int imbn
= 2;
6108 unsigned long scaled_busy_load_per_task
;
6109 struct sg_lb_stats
*local
, *busiest
;
6111 local
= &sds
->local_stat
;
6112 busiest
= &sds
->busiest_stat
;
6114 if (!local
->sum_nr_running
)
6115 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
6116 else if (busiest
->load_per_task
> local
->load_per_task
)
6119 scaled_busy_load_per_task
=
6120 (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
6121 busiest
->group_power
;
6123 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
6124 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
6125 env
->imbalance
= busiest
->load_per_task
;
6130 * OK, we don't have enough imbalance to justify moving tasks,
6131 * however we may be able to increase total CPU power used by
6135 pwr_now
+= busiest
->group_power
*
6136 min(busiest
->load_per_task
, busiest
->avg_load
);
6137 pwr_now
+= local
->group_power
*
6138 min(local
->load_per_task
, local
->avg_load
);
6139 pwr_now
/= SCHED_POWER_SCALE
;
6141 /* Amount of load we'd subtract */
6142 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
6143 pwr_move
+= busiest
->group_power
*
6144 min(busiest
->load_per_task
,
6145 busiest
->avg_load
- scaled_busy_load_per_task
);
6148 /* Amount of load we'd add */
6149 if (busiest
->avg_load
* busiest
->group_power
<
6150 busiest
->load_per_task
* SCHED_POWER_SCALE
) {
6151 tmp
= (busiest
->avg_load
* busiest
->group_power
) /
6154 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
6157 pwr_move
+= local
->group_power
*
6158 min(local
->load_per_task
, local
->avg_load
+ tmp
);
6159 pwr_move
/= SCHED_POWER_SCALE
;
6161 /* Move if we gain throughput */
6162 if (pwr_move
> pwr_now
)
6163 env
->imbalance
= busiest
->load_per_task
;
6167 * calculate_imbalance - Calculate the amount of imbalance present within the
6168 * groups of a given sched_domain during load balance.
6169 * @env: load balance environment
6170 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6172 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6174 unsigned long max_pull
, load_above_capacity
= ~0UL;
6175 struct sg_lb_stats
*local
, *busiest
;
6177 local
= &sds
->local_stat
;
6178 busiest
= &sds
->busiest_stat
;
6180 if (busiest
->group_imb
) {
6182 * In the group_imb case we cannot rely on group-wide averages
6183 * to ensure cpu-load equilibrium, look at wider averages. XXX
6185 busiest
->load_per_task
=
6186 min(busiest
->load_per_task
, sds
->avg_load
);
6190 * In the presence of smp nice balancing, certain scenarios can have
6191 * max load less than avg load(as we skip the groups at or below
6192 * its cpu_power, while calculating max_load..)
6194 if (busiest
->avg_load
<= sds
->avg_load
||
6195 local
->avg_load
>= sds
->avg_load
) {
6197 return fix_small_imbalance(env
, sds
);
6200 if (!busiest
->group_imb
) {
6202 * Don't want to pull so many tasks that a group would go idle.
6203 * Except of course for the group_imb case, since then we might
6204 * have to drop below capacity to reach cpu-load equilibrium.
6206 load_above_capacity
=
6207 (busiest
->sum_nr_running
- busiest
->group_capacity_factor
);
6209 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_POWER_SCALE
);
6210 load_above_capacity
/= busiest
->group_power
;
6214 * We're trying to get all the cpus to the average_load, so we don't
6215 * want to push ourselves above the average load, nor do we wish to
6216 * reduce the max loaded cpu below the average load. At the same time,
6217 * we also don't want to reduce the group load below the group capacity
6218 * (so that we can implement power-savings policies etc). Thus we look
6219 * for the minimum possible imbalance.
6221 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
6223 /* How much load to actually move to equalise the imbalance */
6224 env
->imbalance
= min(
6225 max_pull
* busiest
->group_power
,
6226 (sds
->avg_load
- local
->avg_load
) * local
->group_power
6227 ) / SCHED_POWER_SCALE
;
6230 * if *imbalance is less than the average load per runnable task
6231 * there is no guarantee that any tasks will be moved so we'll have
6232 * a think about bumping its value to force at least one task to be
6235 if (env
->imbalance
< busiest
->load_per_task
)
6236 return fix_small_imbalance(env
, sds
);
6239 /******* find_busiest_group() helpers end here *********************/
6242 * find_busiest_group - Returns the busiest group within the sched_domain
6243 * if there is an imbalance. If there isn't an imbalance, and
6244 * the user has opted for power-savings, it returns a group whose
6245 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6246 * such a group exists.
6248 * Also calculates the amount of weighted load which should be moved
6249 * to restore balance.
6251 * @env: The load balancing environment.
6253 * Return: - The busiest group if imbalance exists.
6254 * - If no imbalance and user has opted for power-savings balance,
6255 * return the least loaded group whose CPUs can be
6256 * put to idle by rebalancing its tasks onto our group.
6258 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
6260 struct sg_lb_stats
*local
, *busiest
;
6261 struct sd_lb_stats sds
;
6263 init_sd_lb_stats(&sds
);
6266 * Compute the various statistics relavent for load balancing at
6269 update_sd_lb_stats(env
, &sds
);
6270 local
= &sds
.local_stat
;
6271 busiest
= &sds
.busiest_stat
;
6273 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
6274 check_asym_packing(env
, &sds
))
6277 /* There is no busy sibling group to pull tasks from */
6278 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
6281 sds
.avg_load
= (SCHED_POWER_SCALE
* sds
.total_load
) / sds
.total_pwr
;
6284 * If the busiest group is imbalanced the below checks don't
6285 * work because they assume all things are equal, which typically
6286 * isn't true due to cpus_allowed constraints and the like.
6288 if (busiest
->group_imb
)
6291 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6292 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_free_capacity
&&
6293 !busiest
->group_has_free_capacity
)
6297 * If the local group is more busy than the selected busiest group
6298 * don't try and pull any tasks.
6300 if (local
->avg_load
>= busiest
->avg_load
)
6304 * Don't pull any tasks if this group is already above the domain
6307 if (local
->avg_load
>= sds
.avg_load
)
6310 if (env
->idle
== CPU_IDLE
) {
6312 * This cpu is idle. If the busiest group load doesn't
6313 * have more tasks than the number of available cpu's and
6314 * there is no imbalance between this and busiest group
6315 * wrt to idle cpu's, it is balanced.
6317 if ((local
->idle_cpus
< busiest
->idle_cpus
) &&
6318 busiest
->sum_nr_running
<= busiest
->group_weight
)
6322 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6323 * imbalance_pct to be conservative.
6325 if (100 * busiest
->avg_load
<=
6326 env
->sd
->imbalance_pct
* local
->avg_load
)
6331 /* Looks like there is an imbalance. Compute it */
6332 calculate_imbalance(env
, &sds
);
6341 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6343 static struct rq
*find_busiest_queue(struct lb_env
*env
,
6344 struct sched_group
*group
)
6346 struct rq
*busiest
= NULL
, *rq
;
6347 unsigned long busiest_load
= 0, busiest_power
= 1;
6350 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6351 unsigned long power
, capacity_factor
, wl
;
6355 rt
= fbq_classify_rq(rq
);
6358 * We classify groups/runqueues into three groups:
6359 * - regular: there are !numa tasks
6360 * - remote: there are numa tasks that run on the 'wrong' node
6361 * - all: there is no distinction
6363 * In order to avoid migrating ideally placed numa tasks,
6364 * ignore those when there's better options.
6366 * If we ignore the actual busiest queue to migrate another
6367 * task, the next balance pass can still reduce the busiest
6368 * queue by moving tasks around inside the node.
6370 * If we cannot move enough load due to this classification
6371 * the next pass will adjust the group classification and
6372 * allow migration of more tasks.
6374 * Both cases only affect the total convergence complexity.
6376 if (rt
> env
->fbq_type
)
6379 power
= power_of(i
);
6380 capacity_factor
= DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
);
6381 if (!capacity_factor
)
6382 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6384 wl
= weighted_cpuload(i
);
6387 * When comparing with imbalance, use weighted_cpuload()
6388 * which is not scaled with the cpu power.
6390 if (capacity_factor
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6394 * For the load comparisons with the other cpu's, consider
6395 * the weighted_cpuload() scaled with the cpu power, so that
6396 * the load can be moved away from the cpu that is potentially
6397 * running at a lower capacity.
6399 * Thus we're looking for max(wl_i / power_i), crosswise
6400 * multiplication to rid ourselves of the division works out
6401 * to: wl_i * power_j > wl_j * power_i; where j is our
6404 if (wl
* busiest_power
> busiest_load
* power
) {
6406 busiest_power
= power
;
6415 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6416 * so long as it is large enough.
6418 #define MAX_PINNED_INTERVAL 512
6420 /* Working cpumask for load_balance and load_balance_newidle. */
6421 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6423 static int need_active_balance(struct lb_env
*env
)
6425 struct sched_domain
*sd
= env
->sd
;
6427 if (env
->idle
== CPU_NEWLY_IDLE
) {
6430 * ASYM_PACKING needs to force migrate tasks from busy but
6431 * higher numbered CPUs in order to pack all tasks in the
6432 * lowest numbered CPUs.
6434 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6438 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6441 static int active_load_balance_cpu_stop(void *data
);
6443 static int should_we_balance(struct lb_env
*env
)
6445 struct sched_group
*sg
= env
->sd
->groups
;
6446 struct cpumask
*sg_cpus
, *sg_mask
;
6447 int cpu
, balance_cpu
= -1;
6450 * In the newly idle case, we will allow all the cpu's
6451 * to do the newly idle load balance.
6453 if (env
->idle
== CPU_NEWLY_IDLE
)
6456 sg_cpus
= sched_group_cpus(sg
);
6457 sg_mask
= sched_group_mask(sg
);
6458 /* Try to find first idle cpu */
6459 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6460 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6467 if (balance_cpu
== -1)
6468 balance_cpu
= group_balance_cpu(sg
);
6471 * First idle cpu or the first cpu(busiest) in this sched group
6472 * is eligible for doing load balancing at this and above domains.
6474 return balance_cpu
== env
->dst_cpu
;
6478 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6479 * tasks if there is an imbalance.
6481 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6482 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6483 int *continue_balancing
)
6485 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6486 struct sched_domain
*sd_parent
= sd
->parent
;
6487 struct sched_group
*group
;
6489 unsigned long flags
;
6490 struct cpumask
*cpus
= __get_cpu_var(load_balance_mask
);
6492 struct lb_env env
= {
6494 .dst_cpu
= this_cpu
,
6496 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6498 .loop_break
= sched_nr_migrate_break
,
6504 * For NEWLY_IDLE load_balancing, we don't need to consider
6505 * other cpus in our group
6507 if (idle
== CPU_NEWLY_IDLE
)
6508 env
.dst_grpmask
= NULL
;
6510 cpumask_copy(cpus
, cpu_active_mask
);
6512 schedstat_inc(sd
, lb_count
[idle
]);
6515 if (!should_we_balance(&env
)) {
6516 *continue_balancing
= 0;
6520 group
= find_busiest_group(&env
);
6522 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6526 busiest
= find_busiest_queue(&env
, group
);
6528 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6532 BUG_ON(busiest
== env
.dst_rq
);
6534 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6537 if (busiest
->nr_running
> 1) {
6539 * Attempt to move tasks. If find_busiest_group has found
6540 * an imbalance but busiest->nr_running <= 1, the group is
6541 * still unbalanced. ld_moved simply stays zero, so it is
6542 * correctly treated as an imbalance.
6544 env
.flags
|= LBF_ALL_PINNED
;
6545 env
.src_cpu
= busiest
->cpu
;
6546 env
.src_rq
= busiest
;
6547 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6550 local_irq_save(flags
);
6551 double_rq_lock(env
.dst_rq
, busiest
);
6554 * cur_ld_moved - load moved in current iteration
6555 * ld_moved - cumulative load moved across iterations
6557 cur_ld_moved
= move_tasks(&env
);
6558 ld_moved
+= cur_ld_moved
;
6559 double_rq_unlock(env
.dst_rq
, busiest
);
6560 local_irq_restore(flags
);
6563 * some other cpu did the load balance for us.
6565 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
6566 resched_cpu(env
.dst_cpu
);
6568 if (env
.flags
& LBF_NEED_BREAK
) {
6569 env
.flags
&= ~LBF_NEED_BREAK
;
6574 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6575 * us and move them to an alternate dst_cpu in our sched_group
6576 * where they can run. The upper limit on how many times we
6577 * iterate on same src_cpu is dependent on number of cpus in our
6580 * This changes load balance semantics a bit on who can move
6581 * load to a given_cpu. In addition to the given_cpu itself
6582 * (or a ilb_cpu acting on its behalf where given_cpu is
6583 * nohz-idle), we now have balance_cpu in a position to move
6584 * load to given_cpu. In rare situations, this may cause
6585 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6586 * _independently_ and at _same_ time to move some load to
6587 * given_cpu) causing exceess load to be moved to given_cpu.
6588 * This however should not happen so much in practice and
6589 * moreover subsequent load balance cycles should correct the
6590 * excess load moved.
6592 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6594 /* Prevent to re-select dst_cpu via env's cpus */
6595 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6597 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6598 env
.dst_cpu
= env
.new_dst_cpu
;
6599 env
.flags
&= ~LBF_DST_PINNED
;
6601 env
.loop_break
= sched_nr_migrate_break
;
6604 * Go back to "more_balance" rather than "redo" since we
6605 * need to continue with same src_cpu.
6611 * We failed to reach balance because of affinity.
6614 int *group_imbalance
= &sd_parent
->groups
->sgp
->imbalance
;
6616 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0) {
6617 *group_imbalance
= 1;
6618 } else if (*group_imbalance
)
6619 *group_imbalance
= 0;
6622 /* All tasks on this runqueue were pinned by CPU affinity */
6623 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6624 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6625 if (!cpumask_empty(cpus
)) {
6627 env
.loop_break
= sched_nr_migrate_break
;
6635 schedstat_inc(sd
, lb_failed
[idle
]);
6637 * Increment the failure counter only on periodic balance.
6638 * We do not want newidle balance, which can be very
6639 * frequent, pollute the failure counter causing
6640 * excessive cache_hot migrations and active balances.
6642 if (idle
!= CPU_NEWLY_IDLE
)
6643 sd
->nr_balance_failed
++;
6645 if (need_active_balance(&env
)) {
6646 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6648 /* don't kick the active_load_balance_cpu_stop,
6649 * if the curr task on busiest cpu can't be
6652 if (!cpumask_test_cpu(this_cpu
,
6653 tsk_cpus_allowed(busiest
->curr
))) {
6654 raw_spin_unlock_irqrestore(&busiest
->lock
,
6656 env
.flags
|= LBF_ALL_PINNED
;
6657 goto out_one_pinned
;
6661 * ->active_balance synchronizes accesses to
6662 * ->active_balance_work. Once set, it's cleared
6663 * only after active load balance is finished.
6665 if (!busiest
->active_balance
) {
6666 busiest
->active_balance
= 1;
6667 busiest
->push_cpu
= this_cpu
;
6670 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
6672 if (active_balance
) {
6673 stop_one_cpu_nowait(cpu_of(busiest
),
6674 active_load_balance_cpu_stop
, busiest
,
6675 &busiest
->active_balance_work
);
6679 * We've kicked active balancing, reset the failure
6682 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
6685 sd
->nr_balance_failed
= 0;
6687 if (likely(!active_balance
)) {
6688 /* We were unbalanced, so reset the balancing interval */
6689 sd
->balance_interval
= sd
->min_interval
;
6692 * If we've begun active balancing, start to back off. This
6693 * case may not be covered by the all_pinned logic if there
6694 * is only 1 task on the busy runqueue (because we don't call
6697 if (sd
->balance_interval
< sd
->max_interval
)
6698 sd
->balance_interval
*= 2;
6704 schedstat_inc(sd
, lb_balanced
[idle
]);
6706 sd
->nr_balance_failed
= 0;
6709 /* tune up the balancing interval */
6710 if (((env
.flags
& LBF_ALL_PINNED
) &&
6711 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
6712 (sd
->balance_interval
< sd
->max_interval
))
6713 sd
->balance_interval
*= 2;
6720 static inline unsigned long
6721 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
6723 unsigned long interval
= sd
->balance_interval
;
6726 interval
*= sd
->busy_factor
;
6728 /* scale ms to jiffies */
6729 interval
= msecs_to_jiffies(interval
);
6730 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6736 update_next_balance(struct sched_domain
*sd
, int cpu_busy
, unsigned long *next_balance
)
6738 unsigned long interval
, next
;
6740 interval
= get_sd_balance_interval(sd
, cpu_busy
);
6741 next
= sd
->last_balance
+ interval
;
6743 if (time_after(*next_balance
, next
))
6744 *next_balance
= next
;
6748 * idle_balance is called by schedule() if this_cpu is about to become
6749 * idle. Attempts to pull tasks from other CPUs.
6751 static int idle_balance(struct rq
*this_rq
)
6753 unsigned long next_balance
= jiffies
+ HZ
;
6754 int this_cpu
= this_rq
->cpu
;
6755 struct sched_domain
*sd
;
6756 int pulled_task
= 0;
6759 idle_enter_fair(this_rq
);
6762 * We must set idle_stamp _before_ calling idle_balance(), such that we
6763 * measure the duration of idle_balance() as idle time.
6765 this_rq
->idle_stamp
= rq_clock(this_rq
);
6767 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
) {
6769 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
6771 update_next_balance(sd
, 0, &next_balance
);
6778 * Drop the rq->lock, but keep IRQ/preempt disabled.
6780 raw_spin_unlock(&this_rq
->lock
);
6782 update_blocked_averages(this_cpu
);
6784 for_each_domain(this_cpu
, sd
) {
6785 int continue_balancing
= 1;
6786 u64 t0
, domain_cost
;
6788 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6791 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
6792 update_next_balance(sd
, 0, &next_balance
);
6796 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
6797 t0
= sched_clock_cpu(this_cpu
);
6799 pulled_task
= load_balance(this_cpu
, this_rq
,
6801 &continue_balancing
);
6803 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
6804 if (domain_cost
> sd
->max_newidle_lb_cost
)
6805 sd
->max_newidle_lb_cost
= domain_cost
;
6807 curr_cost
+= domain_cost
;
6810 update_next_balance(sd
, 0, &next_balance
);
6813 * Stop searching for tasks to pull if there are
6814 * now runnable tasks on this rq.
6816 if (pulled_task
|| this_rq
->nr_running
> 0)
6821 raw_spin_lock(&this_rq
->lock
);
6823 if (curr_cost
> this_rq
->max_idle_balance_cost
)
6824 this_rq
->max_idle_balance_cost
= curr_cost
;
6827 * While browsing the domains, we released the rq lock, a task could
6828 * have been enqueued in the meantime. Since we're not going idle,
6829 * pretend we pulled a task.
6831 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
6835 /* Move the next balance forward */
6836 if (time_after(this_rq
->next_balance
, next_balance
))
6837 this_rq
->next_balance
= next_balance
;
6839 /* Is there a task of a high priority class? */
6840 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
6844 idle_exit_fair(this_rq
);
6845 this_rq
->idle_stamp
= 0;
6852 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6853 * running tasks off the busiest CPU onto idle CPUs. It requires at
6854 * least 1 task to be running on each physical CPU where possible, and
6855 * avoids physical / logical imbalances.
6857 static int active_load_balance_cpu_stop(void *data
)
6859 struct rq
*busiest_rq
= data
;
6860 int busiest_cpu
= cpu_of(busiest_rq
);
6861 int target_cpu
= busiest_rq
->push_cpu
;
6862 struct rq
*target_rq
= cpu_rq(target_cpu
);
6863 struct sched_domain
*sd
;
6865 raw_spin_lock_irq(&busiest_rq
->lock
);
6867 /* make sure the requested cpu hasn't gone down in the meantime */
6868 if (unlikely(busiest_cpu
!= smp_processor_id() ||
6869 !busiest_rq
->active_balance
))
6872 /* Is there any task to move? */
6873 if (busiest_rq
->nr_running
<= 1)
6877 * This condition is "impossible", if it occurs
6878 * we need to fix it. Originally reported by
6879 * Bjorn Helgaas on a 128-cpu setup.
6881 BUG_ON(busiest_rq
== target_rq
);
6883 /* move a task from busiest_rq to target_rq */
6884 double_lock_balance(busiest_rq
, target_rq
);
6886 /* Search for an sd spanning us and the target CPU. */
6888 for_each_domain(target_cpu
, sd
) {
6889 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
6890 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
6895 struct lb_env env
= {
6897 .dst_cpu
= target_cpu
,
6898 .dst_rq
= target_rq
,
6899 .src_cpu
= busiest_rq
->cpu
,
6900 .src_rq
= busiest_rq
,
6904 schedstat_inc(sd
, alb_count
);
6906 if (move_one_task(&env
))
6907 schedstat_inc(sd
, alb_pushed
);
6909 schedstat_inc(sd
, alb_failed
);
6912 double_unlock_balance(busiest_rq
, target_rq
);
6914 busiest_rq
->active_balance
= 0;
6915 raw_spin_unlock_irq(&busiest_rq
->lock
);
6919 static inline int on_null_domain(struct rq
*rq
)
6921 return unlikely(!rcu_dereference_sched(rq
->sd
));
6924 #ifdef CONFIG_NO_HZ_COMMON
6926 * idle load balancing details
6927 * - When one of the busy CPUs notice that there may be an idle rebalancing
6928 * needed, they will kick the idle load balancer, which then does idle
6929 * load balancing for all the idle CPUs.
6932 cpumask_var_t idle_cpus_mask
;
6934 unsigned long next_balance
; /* in jiffy units */
6935 } nohz ____cacheline_aligned
;
6937 static inline int find_new_ilb(void)
6939 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
6941 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
6948 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6949 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6950 * CPU (if there is one).
6952 static void nohz_balancer_kick(void)
6956 nohz
.next_balance
++;
6958 ilb_cpu
= find_new_ilb();
6960 if (ilb_cpu
>= nr_cpu_ids
)
6963 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
6966 * Use smp_send_reschedule() instead of resched_cpu().
6967 * This way we generate a sched IPI on the target cpu which
6968 * is idle. And the softirq performing nohz idle load balance
6969 * will be run before returning from the IPI.
6971 smp_send_reschedule(ilb_cpu
);
6975 static inline void nohz_balance_exit_idle(int cpu
)
6977 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
6979 * Completely isolated CPUs don't ever set, so we must test.
6981 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
6982 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
6983 atomic_dec(&nohz
.nr_cpus
);
6985 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
6989 static inline void set_cpu_sd_state_busy(void)
6991 struct sched_domain
*sd
;
6992 int cpu
= smp_processor_id();
6995 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
6997 if (!sd
|| !sd
->nohz_idle
)
7001 atomic_inc(&sd
->groups
->sgp
->nr_busy_cpus
);
7006 void set_cpu_sd_state_idle(void)
7008 struct sched_domain
*sd
;
7009 int cpu
= smp_processor_id();
7012 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7014 if (!sd
|| sd
->nohz_idle
)
7018 atomic_dec(&sd
->groups
->sgp
->nr_busy_cpus
);
7024 * This routine will record that the cpu is going idle with tick stopped.
7025 * This info will be used in performing idle load balancing in the future.
7027 void nohz_balance_enter_idle(int cpu
)
7030 * If this cpu is going down, then nothing needs to be done.
7032 if (!cpu_active(cpu
))
7035 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
7039 * If we're a completely isolated CPU, we don't play.
7041 if (on_null_domain(cpu_rq(cpu
)))
7044 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
7045 atomic_inc(&nohz
.nr_cpus
);
7046 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7049 static int sched_ilb_notifier(struct notifier_block
*nfb
,
7050 unsigned long action
, void *hcpu
)
7052 switch (action
& ~CPU_TASKS_FROZEN
) {
7054 nohz_balance_exit_idle(smp_processor_id());
7062 static DEFINE_SPINLOCK(balancing
);
7065 * Scale the max load_balance interval with the number of CPUs in the system.
7066 * This trades load-balance latency on larger machines for less cross talk.
7068 void update_max_interval(void)
7070 max_load_balance_interval
= HZ
*num_online_cpus()/10;
7074 * It checks each scheduling domain to see if it is due to be balanced,
7075 * and initiates a balancing operation if so.
7077 * Balancing parameters are set up in init_sched_domains.
7079 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
7081 int continue_balancing
= 1;
7083 unsigned long interval
;
7084 struct sched_domain
*sd
;
7085 /* Earliest time when we have to do rebalance again */
7086 unsigned long next_balance
= jiffies
+ 60*HZ
;
7087 int update_next_balance
= 0;
7088 int need_serialize
, need_decay
= 0;
7091 update_blocked_averages(cpu
);
7094 for_each_domain(cpu
, sd
) {
7096 * Decay the newidle max times here because this is a regular
7097 * visit to all the domains. Decay ~1% per second.
7099 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
7100 sd
->max_newidle_lb_cost
=
7101 (sd
->max_newidle_lb_cost
* 253) / 256;
7102 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
7105 max_cost
+= sd
->max_newidle_lb_cost
;
7107 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7111 * Stop the load balance at this level. There is another
7112 * CPU in our sched group which is doing load balancing more
7115 if (!continue_balancing
) {
7121 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7123 need_serialize
= sd
->flags
& SD_SERIALIZE
;
7124 if (need_serialize
) {
7125 if (!spin_trylock(&balancing
))
7129 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
7130 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
7132 * The LBF_DST_PINNED logic could have changed
7133 * env->dst_cpu, so we can't know our idle
7134 * state even if we migrated tasks. Update it.
7136 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
7138 sd
->last_balance
= jiffies
;
7139 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7142 spin_unlock(&balancing
);
7144 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
7145 next_balance
= sd
->last_balance
+ interval
;
7146 update_next_balance
= 1;
7151 * Ensure the rq-wide value also decays but keep it at a
7152 * reasonable floor to avoid funnies with rq->avg_idle.
7154 rq
->max_idle_balance_cost
=
7155 max((u64
)sysctl_sched_migration_cost
, max_cost
);
7160 * next_balance will be updated only when there is a need.
7161 * When the cpu is attached to null domain for ex, it will not be
7164 if (likely(update_next_balance
))
7165 rq
->next_balance
= next_balance
;
7168 #ifdef CONFIG_NO_HZ_COMMON
7170 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7171 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7173 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
7175 int this_cpu
= this_rq
->cpu
;
7179 if (idle
!= CPU_IDLE
||
7180 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
7183 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
7184 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
7188 * If this cpu gets work to do, stop the load balancing
7189 * work being done for other cpus. Next load
7190 * balancing owner will pick it up.
7195 rq
= cpu_rq(balance_cpu
);
7198 * If time for next balance is due,
7201 if (time_after_eq(jiffies
, rq
->next_balance
)) {
7202 raw_spin_lock_irq(&rq
->lock
);
7203 update_rq_clock(rq
);
7204 update_idle_cpu_load(rq
);
7205 raw_spin_unlock_irq(&rq
->lock
);
7206 rebalance_domains(rq
, CPU_IDLE
);
7209 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
7210 this_rq
->next_balance
= rq
->next_balance
;
7212 nohz
.next_balance
= this_rq
->next_balance
;
7214 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
7218 * Current heuristic for kicking the idle load balancer in the presence
7219 * of an idle cpu is the system.
7220 * - This rq has more than one task.
7221 * - At any scheduler domain level, this cpu's scheduler group has multiple
7222 * busy cpu's exceeding the group's power.
7223 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7224 * domain span are idle.
7226 static inline int nohz_kick_needed(struct rq
*rq
)
7228 unsigned long now
= jiffies
;
7229 struct sched_domain
*sd
;
7230 struct sched_group_power
*sgp
;
7231 int nr_busy
, cpu
= rq
->cpu
;
7233 if (unlikely(rq
->idle_balance
))
7237 * We may be recently in ticked or tickless idle mode. At the first
7238 * busy tick after returning from idle, we will update the busy stats.
7240 set_cpu_sd_state_busy();
7241 nohz_balance_exit_idle(cpu
);
7244 * None are in tickless mode and hence no need for NOHZ idle load
7247 if (likely(!atomic_read(&nohz
.nr_cpus
)))
7250 if (time_before(now
, nohz
.next_balance
))
7253 if (rq
->nr_running
>= 2)
7257 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7260 sgp
= sd
->groups
->sgp
;
7261 nr_busy
= atomic_read(&sgp
->nr_busy_cpus
);
7264 goto need_kick_unlock
;
7267 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
7269 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
7270 sched_domain_span(sd
)) < cpu
))
7271 goto need_kick_unlock
;
7282 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
7286 * run_rebalance_domains is triggered when needed from the scheduler tick.
7287 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7289 static void run_rebalance_domains(struct softirq_action
*h
)
7291 struct rq
*this_rq
= this_rq();
7292 enum cpu_idle_type idle
= this_rq
->idle_balance
?
7293 CPU_IDLE
: CPU_NOT_IDLE
;
7295 rebalance_domains(this_rq
, idle
);
7298 * If this cpu has a pending nohz_balance_kick, then do the
7299 * balancing on behalf of the other idle cpus whose ticks are
7302 nohz_idle_balance(this_rq
, idle
);
7306 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7308 void trigger_load_balance(struct rq
*rq
)
7310 /* Don't need to rebalance while attached to NULL domain */
7311 if (unlikely(on_null_domain(rq
)))
7314 if (time_after_eq(jiffies
, rq
->next_balance
))
7315 raise_softirq(SCHED_SOFTIRQ
);
7316 #ifdef CONFIG_NO_HZ_COMMON
7317 if (nohz_kick_needed(rq
))
7318 nohz_balancer_kick();
7322 static void rq_online_fair(struct rq
*rq
)
7327 static void rq_offline_fair(struct rq
*rq
)
7331 /* Ensure any throttled groups are reachable by pick_next_task */
7332 unthrottle_offline_cfs_rqs(rq
);
7335 #endif /* CONFIG_SMP */
7338 * scheduler tick hitting a task of our scheduling class:
7340 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
7342 struct cfs_rq
*cfs_rq
;
7343 struct sched_entity
*se
= &curr
->se
;
7345 for_each_sched_entity(se
) {
7346 cfs_rq
= cfs_rq_of(se
);
7347 entity_tick(cfs_rq
, se
, queued
);
7350 if (numabalancing_enabled
)
7351 task_tick_numa(rq
, curr
);
7353 update_rq_runnable_avg(rq
, 1);
7357 * called on fork with the child task as argument from the parent's context
7358 * - child not yet on the tasklist
7359 * - preemption disabled
7361 static void task_fork_fair(struct task_struct
*p
)
7363 struct cfs_rq
*cfs_rq
;
7364 struct sched_entity
*se
= &p
->se
, *curr
;
7365 int this_cpu
= smp_processor_id();
7366 struct rq
*rq
= this_rq();
7367 unsigned long flags
;
7369 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7371 update_rq_clock(rq
);
7373 cfs_rq
= task_cfs_rq(current
);
7374 curr
= cfs_rq
->curr
;
7377 * Not only the cpu but also the task_group of the parent might have
7378 * been changed after parent->se.parent,cfs_rq were copied to
7379 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7380 * of child point to valid ones.
7383 __set_task_cpu(p
, this_cpu
);
7386 update_curr(cfs_rq
);
7389 se
->vruntime
= curr
->vruntime
;
7390 place_entity(cfs_rq
, se
, 1);
7392 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
7394 * Upon rescheduling, sched_class::put_prev_task() will place
7395 * 'current' within the tree based on its new key value.
7397 swap(curr
->vruntime
, se
->vruntime
);
7398 resched_task(rq
->curr
);
7401 se
->vruntime
-= cfs_rq
->min_vruntime
;
7403 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7407 * Priority of the task has changed. Check to see if we preempt
7411 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
7417 * Reschedule if we are currently running on this runqueue and
7418 * our priority decreased, or if we are not currently running on
7419 * this runqueue and our priority is higher than the current's
7421 if (rq
->curr
== p
) {
7422 if (p
->prio
> oldprio
)
7423 resched_task(rq
->curr
);
7425 check_preempt_curr(rq
, p
, 0);
7428 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
7430 struct sched_entity
*se
= &p
->se
;
7431 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7434 * Ensure the task's vruntime is normalized, so that when it's
7435 * switched back to the fair class the enqueue_entity(.flags=0) will
7436 * do the right thing.
7438 * If it's on_rq, then the dequeue_entity(.flags=0) will already
7439 * have normalized the vruntime, if it's !on_rq, then only when
7440 * the task is sleeping will it still have non-normalized vruntime.
7442 if (!p
->on_rq
&& p
->state
!= TASK_RUNNING
) {
7444 * Fix up our vruntime so that the current sleep doesn't
7445 * cause 'unlimited' sleep bonus.
7447 place_entity(cfs_rq
, se
, 0);
7448 se
->vruntime
-= cfs_rq
->min_vruntime
;
7453 * Remove our load from contribution when we leave sched_fair
7454 * and ensure we don't carry in an old decay_count if we
7457 if (se
->avg
.decay_count
) {
7458 __synchronize_entity_decay(se
);
7459 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7465 * We switched to the sched_fair class.
7467 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7469 struct sched_entity
*se
= &p
->se
;
7470 #ifdef CONFIG_FAIR_GROUP_SCHED
7472 * Since the real-depth could have been changed (only FAIR
7473 * class maintain depth value), reset depth properly.
7475 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7481 * We were most likely switched from sched_rt, so
7482 * kick off the schedule if running, otherwise just see
7483 * if we can still preempt the current task.
7486 resched_task(rq
->curr
);
7488 check_preempt_curr(rq
, p
, 0);
7491 /* Account for a task changing its policy or group.
7493 * This routine is mostly called to set cfs_rq->curr field when a task
7494 * migrates between groups/classes.
7496 static void set_curr_task_fair(struct rq
*rq
)
7498 struct sched_entity
*se
= &rq
->curr
->se
;
7500 for_each_sched_entity(se
) {
7501 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7503 set_next_entity(cfs_rq
, se
);
7504 /* ensure bandwidth has been allocated on our new cfs_rq */
7505 account_cfs_rq_runtime(cfs_rq
, 0);
7509 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7511 cfs_rq
->tasks_timeline
= RB_ROOT
;
7512 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7513 #ifndef CONFIG_64BIT
7514 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7517 atomic64_set(&cfs_rq
->decay_counter
, 1);
7518 atomic_long_set(&cfs_rq
->removed_load
, 0);
7522 #ifdef CONFIG_FAIR_GROUP_SCHED
7523 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
7525 struct sched_entity
*se
= &p
->se
;
7526 struct cfs_rq
*cfs_rq
;
7529 * If the task was not on the rq at the time of this cgroup movement
7530 * it must have been asleep, sleeping tasks keep their ->vruntime
7531 * absolute on their old rq until wakeup (needed for the fair sleeper
7532 * bonus in place_entity()).
7534 * If it was on the rq, we've just 'preempted' it, which does convert
7535 * ->vruntime to a relative base.
7537 * Make sure both cases convert their relative position when migrating
7538 * to another cgroup's rq. This does somewhat interfere with the
7539 * fair sleeper stuff for the first placement, but who cares.
7542 * When !on_rq, vruntime of the task has usually NOT been normalized.
7543 * But there are some cases where it has already been normalized:
7545 * - Moving a forked child which is waiting for being woken up by
7546 * wake_up_new_task().
7547 * - Moving a task which has been woken up by try_to_wake_up() and
7548 * waiting for actually being woken up by sched_ttwu_pending().
7550 * To prevent boost or penalty in the new cfs_rq caused by delta
7551 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7553 if (!on_rq
&& (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7557 se
->vruntime
-= cfs_rq_of(se
)->min_vruntime
;
7558 set_task_rq(p
, task_cpu(p
));
7559 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7561 cfs_rq
= cfs_rq_of(se
);
7562 se
->vruntime
+= cfs_rq
->min_vruntime
;
7565 * migrate_task_rq_fair() will have removed our previous
7566 * contribution, but we must synchronize for ongoing future
7569 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7570 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
7575 void free_fair_sched_group(struct task_group
*tg
)
7579 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7581 for_each_possible_cpu(i
) {
7583 kfree(tg
->cfs_rq
[i
]);
7592 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7594 struct cfs_rq
*cfs_rq
;
7595 struct sched_entity
*se
;
7598 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7601 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7605 tg
->shares
= NICE_0_LOAD
;
7607 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7609 for_each_possible_cpu(i
) {
7610 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7611 GFP_KERNEL
, cpu_to_node(i
));
7615 se
= kzalloc_node(sizeof(struct sched_entity
),
7616 GFP_KERNEL
, cpu_to_node(i
));
7620 init_cfs_rq(cfs_rq
);
7621 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
7632 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7634 struct rq
*rq
= cpu_rq(cpu
);
7635 unsigned long flags
;
7638 * Only empty task groups can be destroyed; so we can speculatively
7639 * check on_list without danger of it being re-added.
7641 if (!tg
->cfs_rq
[cpu
]->on_list
)
7644 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7645 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
7646 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7649 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7650 struct sched_entity
*se
, int cpu
,
7651 struct sched_entity
*parent
)
7653 struct rq
*rq
= cpu_rq(cpu
);
7657 init_cfs_rq_runtime(cfs_rq
);
7659 tg
->cfs_rq
[cpu
] = cfs_rq
;
7662 /* se could be NULL for root_task_group */
7667 se
->cfs_rq
= &rq
->cfs
;
7670 se
->cfs_rq
= parent
->my_q
;
7671 se
->depth
= parent
->depth
+ 1;
7675 /* guarantee group entities always have weight */
7676 update_load_set(&se
->load
, NICE_0_LOAD
);
7677 se
->parent
= parent
;
7680 static DEFINE_MUTEX(shares_mutex
);
7682 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7685 unsigned long flags
;
7688 * We can't change the weight of the root cgroup.
7693 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
7695 mutex_lock(&shares_mutex
);
7696 if (tg
->shares
== shares
)
7699 tg
->shares
= shares
;
7700 for_each_possible_cpu(i
) {
7701 struct rq
*rq
= cpu_rq(i
);
7702 struct sched_entity
*se
;
7705 /* Propagate contribution to hierarchy */
7706 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7708 /* Possible calls to update_curr() need rq clock */
7709 update_rq_clock(rq
);
7710 for_each_sched_entity(se
)
7711 update_cfs_shares(group_cfs_rq(se
));
7712 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7716 mutex_unlock(&shares_mutex
);
7719 #else /* CONFIG_FAIR_GROUP_SCHED */
7721 void free_fair_sched_group(struct task_group
*tg
) { }
7723 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7728 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
7730 #endif /* CONFIG_FAIR_GROUP_SCHED */
7733 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
7735 struct sched_entity
*se
= &task
->se
;
7736 unsigned int rr_interval
= 0;
7739 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7742 if (rq
->cfs
.load
.weight
)
7743 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
7749 * All the scheduling class methods:
7751 const struct sched_class fair_sched_class
= {
7752 .next
= &idle_sched_class
,
7753 .enqueue_task
= enqueue_task_fair
,
7754 .dequeue_task
= dequeue_task_fair
,
7755 .yield_task
= yield_task_fair
,
7756 .yield_to_task
= yield_to_task_fair
,
7758 .check_preempt_curr
= check_preempt_wakeup
,
7760 .pick_next_task
= pick_next_task_fair
,
7761 .put_prev_task
= put_prev_task_fair
,
7764 .select_task_rq
= select_task_rq_fair
,
7765 .migrate_task_rq
= migrate_task_rq_fair
,
7767 .rq_online
= rq_online_fair
,
7768 .rq_offline
= rq_offline_fair
,
7770 .task_waking
= task_waking_fair
,
7773 .set_curr_task
= set_curr_task_fair
,
7774 .task_tick
= task_tick_fair
,
7775 .task_fork
= task_fork_fair
,
7777 .prio_changed
= prio_changed_fair
,
7778 .switched_from
= switched_from_fair
,
7779 .switched_to
= switched_to_fair
,
7781 .get_rr_interval
= get_rr_interval_fair
,
7783 #ifdef CONFIG_FAIR_GROUP_SCHED
7784 .task_move_group
= task_move_group_fair
,
7788 #ifdef CONFIG_SCHED_DEBUG
7789 void print_cfs_stats(struct seq_file
*m
, int cpu
)
7791 struct cfs_rq
*cfs_rq
;
7794 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
7795 print_cfs_rq(m
, cpu
, cfs_rq
);
7800 __init
void init_sched_fair_class(void)
7803 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7805 #ifdef CONFIG_NO_HZ_COMMON
7806 nohz
.next_balance
= jiffies
;
7807 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7808 cpu_notifier(sched_ilb_notifier
, 0);