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 power
;
1031 /* Approximate capacity in terms of runnable tasks on a node */
1032 unsigned long 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
->power
+= 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_capacity, or we'll detect a huge imbalance
1065 ns
->load
= (ns
->load
* SCHED_POWER_SCALE
) / ns
->power
;
1066 ns
->capacity
= DIV_ROUND_CLOSEST(ns
->power
, SCHED_POWER_SCALE
);
1067 ns
->has_capacity
= (ns
->nr_running
< ns
->capacity
);
1070 struct task_numa_env
{
1071 struct task_struct
*p
;
1073 int src_cpu
, src_nid
;
1074 int dst_cpu
, dst_nid
;
1076 struct numa_stats src_stats
, dst_stats
;
1080 struct task_struct
*best_task
;
1085 static void task_numa_assign(struct task_numa_env
*env
,
1086 struct task_struct
*p
, long imp
)
1089 put_task_struct(env
->best_task
);
1094 env
->best_imp
= imp
;
1095 env
->best_cpu
= env
->dst_cpu
;
1099 * This checks if the overall compute and NUMA accesses of the system would
1100 * be improved if the source tasks was migrated to the target dst_cpu taking
1101 * into account that it might be best if task running on the dst_cpu should
1102 * be exchanged with the source task
1104 static void task_numa_compare(struct task_numa_env
*env
,
1105 long taskimp
, long groupimp
)
1107 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1108 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1109 struct task_struct
*cur
;
1110 long dst_load
, src_load
;
1112 long imp
= (groupimp
> 0) ? groupimp
: taskimp
;
1115 cur
= ACCESS_ONCE(dst_rq
->curr
);
1116 if (cur
->pid
== 0) /* idle */
1120 * "imp" is the fault differential for the source task between the
1121 * source and destination node. Calculate the total differential for
1122 * the source task and potential destination task. The more negative
1123 * the value is, the more rmeote accesses that would be expected to
1124 * be incurred if the tasks were swapped.
1127 /* Skip this swap candidate if cannot move to the source cpu */
1128 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1132 * If dst and source tasks are in the same NUMA group, or not
1133 * in any group then look only at task weights.
1135 if (cur
->numa_group
== env
->p
->numa_group
) {
1136 imp
= taskimp
+ task_weight(cur
, env
->src_nid
) -
1137 task_weight(cur
, env
->dst_nid
);
1139 * Add some hysteresis to prevent swapping the
1140 * tasks within a group over tiny differences.
1142 if (cur
->numa_group
)
1146 * Compare the group weights. If a task is all by
1147 * itself (not part of a group), use the task weight
1150 if (env
->p
->numa_group
)
1155 if (cur
->numa_group
)
1156 imp
+= group_weight(cur
, env
->src_nid
) -
1157 group_weight(cur
, env
->dst_nid
);
1159 imp
+= task_weight(cur
, env
->src_nid
) -
1160 task_weight(cur
, env
->dst_nid
);
1164 if (imp
< env
->best_imp
)
1168 /* Is there capacity at our destination? */
1169 if (env
->src_stats
.has_capacity
&&
1170 !env
->dst_stats
.has_capacity
)
1176 /* Balance doesn't matter much if we're running a task per cpu */
1177 if (src_rq
->nr_running
== 1 && dst_rq
->nr_running
== 1)
1181 * In the overloaded case, try and keep the load balanced.
1184 dst_load
= env
->dst_stats
.load
;
1185 src_load
= env
->src_stats
.load
;
1187 /* XXX missing power terms */
1188 load
= task_h_load(env
->p
);
1193 load
= task_h_load(cur
);
1198 /* make src_load the smaller */
1199 if (dst_load
< src_load
)
1200 swap(dst_load
, src_load
);
1202 if (src_load
* env
->imbalance_pct
< dst_load
* 100)
1206 task_numa_assign(env
, cur
, imp
);
1211 static void task_numa_find_cpu(struct task_numa_env
*env
,
1212 long taskimp
, long groupimp
)
1216 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1217 /* Skip this CPU if the source task cannot migrate */
1218 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1222 task_numa_compare(env
, taskimp
, groupimp
);
1226 static int task_numa_migrate(struct task_struct
*p
)
1228 struct task_numa_env env
= {
1231 .src_cpu
= task_cpu(p
),
1232 .src_nid
= task_node(p
),
1234 .imbalance_pct
= 112,
1240 struct sched_domain
*sd
;
1241 unsigned long taskweight
, groupweight
;
1243 long taskimp
, groupimp
;
1246 * Pick the lowest SD_NUMA domain, as that would have the smallest
1247 * imbalance and would be the first to start moving tasks about.
1249 * And we want to avoid any moving of tasks about, as that would create
1250 * random movement of tasks -- counter the numa conditions we're trying
1254 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1256 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1260 * Cpusets can break the scheduler domain tree into smaller
1261 * balance domains, some of which do not cross NUMA boundaries.
1262 * Tasks that are "trapped" in such domains cannot be migrated
1263 * elsewhere, so there is no point in (re)trying.
1265 if (unlikely(!sd
)) {
1266 p
->numa_preferred_nid
= task_node(p
);
1270 taskweight
= task_weight(p
, env
.src_nid
);
1271 groupweight
= group_weight(p
, env
.src_nid
);
1272 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1273 env
.dst_nid
= p
->numa_preferred_nid
;
1274 taskimp
= task_weight(p
, env
.dst_nid
) - taskweight
;
1275 groupimp
= group_weight(p
, env
.dst_nid
) - groupweight
;
1276 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1278 /* If the preferred nid has capacity, try to use it. */
1279 if (env
.dst_stats
.has_capacity
)
1280 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1282 /* No space available on the preferred nid. Look elsewhere. */
1283 if (env
.best_cpu
== -1) {
1284 for_each_online_node(nid
) {
1285 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1288 /* Only consider nodes where both task and groups benefit */
1289 taskimp
= task_weight(p
, nid
) - taskweight
;
1290 groupimp
= group_weight(p
, nid
) - groupweight
;
1291 if (taskimp
< 0 && groupimp
< 0)
1295 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1296 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1300 /* No better CPU than the current one was found. */
1301 if (env
.best_cpu
== -1)
1304 sched_setnuma(p
, env
.dst_nid
);
1307 * Reset the scan period if the task is being rescheduled on an
1308 * alternative node to recheck if the tasks is now properly placed.
1310 p
->numa_scan_period
= task_scan_min(p
);
1312 if (env
.best_task
== NULL
) {
1313 ret
= migrate_task_to(p
, env
.best_cpu
);
1315 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1319 ret
= migrate_swap(p
, env
.best_task
);
1321 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1322 put_task_struct(env
.best_task
);
1326 /* Attempt to migrate a task to a CPU on the preferred node. */
1327 static void numa_migrate_preferred(struct task_struct
*p
)
1329 /* This task has no NUMA fault statistics yet */
1330 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults_memory
))
1333 /* Periodically retry migrating the task to the preferred node */
1334 p
->numa_migrate_retry
= jiffies
+ HZ
;
1336 /* Success if task is already running on preferred CPU */
1337 if (task_node(p
) == p
->numa_preferred_nid
)
1340 /* Otherwise, try migrate to a CPU on the preferred node */
1341 task_numa_migrate(p
);
1345 * Find the nodes on which the workload is actively running. We do this by
1346 * tracking the nodes from which NUMA hinting faults are triggered. This can
1347 * be different from the set of nodes where the workload's memory is currently
1350 * The bitmask is used to make smarter decisions on when to do NUMA page
1351 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1352 * are added when they cause over 6/16 of the maximum number of faults, but
1353 * only removed when they drop below 3/16.
1355 static void update_numa_active_node_mask(struct numa_group
*numa_group
)
1357 unsigned long faults
, max_faults
= 0;
1360 for_each_online_node(nid
) {
1361 faults
= group_faults_cpu(numa_group
, nid
);
1362 if (faults
> max_faults
)
1363 max_faults
= faults
;
1366 for_each_online_node(nid
) {
1367 faults
= group_faults_cpu(numa_group
, nid
);
1368 if (!node_isset(nid
, numa_group
->active_nodes
)) {
1369 if (faults
> max_faults
* 6 / 16)
1370 node_set(nid
, numa_group
->active_nodes
);
1371 } else if (faults
< max_faults
* 3 / 16)
1372 node_clear(nid
, numa_group
->active_nodes
);
1377 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1378 * increments. The more local the fault statistics are, the higher the scan
1379 * period will be for the next scan window. If local/remote ratio is below
1380 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1381 * scan period will decrease
1383 #define NUMA_PERIOD_SLOTS 10
1384 #define NUMA_PERIOD_THRESHOLD 3
1387 * Increase the scan period (slow down scanning) if the majority of
1388 * our memory is already on our local node, or if the majority of
1389 * the page accesses are shared with other processes.
1390 * Otherwise, decrease the scan period.
1392 static void update_task_scan_period(struct task_struct
*p
,
1393 unsigned long shared
, unsigned long private)
1395 unsigned int period_slot
;
1399 unsigned long remote
= p
->numa_faults_locality
[0];
1400 unsigned long local
= p
->numa_faults_locality
[1];
1403 * If there were no record hinting faults then either the task is
1404 * completely idle or all activity is areas that are not of interest
1405 * to automatic numa balancing. Scan slower
1407 if (local
+ shared
== 0) {
1408 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1409 p
->numa_scan_period
<< 1);
1411 p
->mm
->numa_next_scan
= jiffies
+
1412 msecs_to_jiffies(p
->numa_scan_period
);
1418 * Prepare to scale scan period relative to the current period.
1419 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1420 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1421 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1423 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1424 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1425 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1426 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1429 diff
= slot
* period_slot
;
1431 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1434 * Scale scan rate increases based on sharing. There is an
1435 * inverse relationship between the degree of sharing and
1436 * the adjustment made to the scanning period. Broadly
1437 * speaking the intent is that there is little point
1438 * scanning faster if shared accesses dominate as it may
1439 * simply bounce migrations uselessly
1441 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
));
1442 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1445 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1446 task_scan_min(p
), task_scan_max(p
));
1447 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1451 * Get the fraction of time the task has been running since the last
1452 * NUMA placement cycle. The scheduler keeps similar statistics, but
1453 * decays those on a 32ms period, which is orders of magnitude off
1454 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1455 * stats only if the task is so new there are no NUMA statistics yet.
1457 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1459 u64 runtime
, delta
, now
;
1460 /* Use the start of this time slice to avoid calculations. */
1461 now
= p
->se
.exec_start
;
1462 runtime
= p
->se
.sum_exec_runtime
;
1464 if (p
->last_task_numa_placement
) {
1465 delta
= runtime
- p
->last_sum_exec_runtime
;
1466 *period
= now
- p
->last_task_numa_placement
;
1468 delta
= p
->se
.avg
.runnable_avg_sum
;
1469 *period
= p
->se
.avg
.runnable_avg_period
;
1472 p
->last_sum_exec_runtime
= runtime
;
1473 p
->last_task_numa_placement
= now
;
1478 static void task_numa_placement(struct task_struct
*p
)
1480 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1481 unsigned long max_faults
= 0, max_group_faults
= 0;
1482 unsigned long fault_types
[2] = { 0, 0 };
1483 unsigned long total_faults
;
1484 u64 runtime
, period
;
1485 spinlock_t
*group_lock
= NULL
;
1487 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1488 if (p
->numa_scan_seq
== seq
)
1490 p
->numa_scan_seq
= seq
;
1491 p
->numa_scan_period_max
= task_scan_max(p
);
1493 total_faults
= p
->numa_faults_locality
[0] +
1494 p
->numa_faults_locality
[1];
1495 runtime
= numa_get_avg_runtime(p
, &period
);
1497 /* If the task is part of a group prevent parallel updates to group stats */
1498 if (p
->numa_group
) {
1499 group_lock
= &p
->numa_group
->lock
;
1500 spin_lock(group_lock
);
1503 /* Find the node with the highest number of faults */
1504 for_each_online_node(nid
) {
1505 unsigned long faults
= 0, group_faults
= 0;
1508 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
1509 long diff
, f_diff
, f_weight
;
1511 i
= task_faults_idx(nid
, priv
);
1513 /* Decay existing window, copy faults since last scan */
1514 diff
= p
->numa_faults_buffer_memory
[i
] - p
->numa_faults_memory
[i
] / 2;
1515 fault_types
[priv
] += p
->numa_faults_buffer_memory
[i
];
1516 p
->numa_faults_buffer_memory
[i
] = 0;
1519 * Normalize the faults_from, so all tasks in a group
1520 * count according to CPU use, instead of by the raw
1521 * number of faults. Tasks with little runtime have
1522 * little over-all impact on throughput, and thus their
1523 * faults are less important.
1525 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
1526 f_weight
= (f_weight
* p
->numa_faults_buffer_cpu
[i
]) /
1528 f_diff
= f_weight
- p
->numa_faults_cpu
[i
] / 2;
1529 p
->numa_faults_buffer_cpu
[i
] = 0;
1531 p
->numa_faults_memory
[i
] += diff
;
1532 p
->numa_faults_cpu
[i
] += f_diff
;
1533 faults
+= p
->numa_faults_memory
[i
];
1534 p
->total_numa_faults
+= diff
;
1535 if (p
->numa_group
) {
1536 /* safe because we can only change our own group */
1537 p
->numa_group
->faults
[i
] += diff
;
1538 p
->numa_group
->faults_cpu
[i
] += f_diff
;
1539 p
->numa_group
->total_faults
+= diff
;
1540 group_faults
+= p
->numa_group
->faults
[i
];
1544 if (faults
> max_faults
) {
1545 max_faults
= faults
;
1549 if (group_faults
> max_group_faults
) {
1550 max_group_faults
= group_faults
;
1551 max_group_nid
= nid
;
1555 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1557 if (p
->numa_group
) {
1558 update_numa_active_node_mask(p
->numa_group
);
1560 * If the preferred task and group nids are different,
1561 * iterate over the nodes again to find the best place.
1563 if (max_nid
!= max_group_nid
) {
1564 unsigned long weight
, max_weight
= 0;
1566 for_each_online_node(nid
) {
1567 weight
= task_weight(p
, nid
) + group_weight(p
, nid
);
1568 if (weight
> max_weight
) {
1569 max_weight
= weight
;
1575 spin_unlock(group_lock
);
1578 /* Preferred node as the node with the most faults */
1579 if (max_faults
&& max_nid
!= p
->numa_preferred_nid
) {
1580 /* Update the preferred nid and migrate task if possible */
1581 sched_setnuma(p
, max_nid
);
1582 numa_migrate_preferred(p
);
1586 static inline int get_numa_group(struct numa_group
*grp
)
1588 return atomic_inc_not_zero(&grp
->refcount
);
1591 static inline void put_numa_group(struct numa_group
*grp
)
1593 if (atomic_dec_and_test(&grp
->refcount
))
1594 kfree_rcu(grp
, rcu
);
1597 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1600 struct numa_group
*grp
, *my_grp
;
1601 struct task_struct
*tsk
;
1603 int cpu
= cpupid_to_cpu(cpupid
);
1606 if (unlikely(!p
->numa_group
)) {
1607 unsigned int size
= sizeof(struct numa_group
) +
1608 4*nr_node_ids
*sizeof(unsigned long);
1610 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1614 atomic_set(&grp
->refcount
, 1);
1615 spin_lock_init(&grp
->lock
);
1616 INIT_LIST_HEAD(&grp
->task_list
);
1618 /* Second half of the array tracks nids where faults happen */
1619 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
1622 node_set(task_node(current
), grp
->active_nodes
);
1624 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1625 grp
->faults
[i
] = p
->numa_faults_memory
[i
];
1627 grp
->total_faults
= p
->total_numa_faults
;
1629 list_add(&p
->numa_entry
, &grp
->task_list
);
1631 rcu_assign_pointer(p
->numa_group
, grp
);
1635 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1637 if (!cpupid_match_pid(tsk
, cpupid
))
1640 grp
= rcu_dereference(tsk
->numa_group
);
1644 my_grp
= p
->numa_group
;
1649 * Only join the other group if its bigger; if we're the bigger group,
1650 * the other task will join us.
1652 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1656 * Tie-break on the grp address.
1658 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1661 /* Always join threads in the same process. */
1662 if (tsk
->mm
== current
->mm
)
1665 /* Simple filter to avoid false positives due to PID collisions */
1666 if (flags
& TNF_SHARED
)
1669 /* Update priv based on whether false sharing was detected */
1672 if (join
&& !get_numa_group(grp
))
1680 double_lock(&my_grp
->lock
, &grp
->lock
);
1682 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
1683 my_grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1684 grp
->faults
[i
] += p
->numa_faults_memory
[i
];
1686 my_grp
->total_faults
-= p
->total_numa_faults
;
1687 grp
->total_faults
+= p
->total_numa_faults
;
1689 list_move(&p
->numa_entry
, &grp
->task_list
);
1693 spin_unlock(&my_grp
->lock
);
1694 spin_unlock(&grp
->lock
);
1696 rcu_assign_pointer(p
->numa_group
, grp
);
1698 put_numa_group(my_grp
);
1706 void task_numa_free(struct task_struct
*p
)
1708 struct numa_group
*grp
= p
->numa_group
;
1710 void *numa_faults
= p
->numa_faults_memory
;
1713 spin_lock(&grp
->lock
);
1714 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1715 grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1716 grp
->total_faults
-= p
->total_numa_faults
;
1718 list_del(&p
->numa_entry
);
1720 spin_unlock(&grp
->lock
);
1721 rcu_assign_pointer(p
->numa_group
, NULL
);
1722 put_numa_group(grp
);
1725 p
->numa_faults_memory
= NULL
;
1726 p
->numa_faults_buffer_memory
= NULL
;
1727 p
->numa_faults_cpu
= NULL
;
1728 p
->numa_faults_buffer_cpu
= NULL
;
1733 * Got a PROT_NONE fault for a page on @node.
1735 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
1737 struct task_struct
*p
= current
;
1738 bool migrated
= flags
& TNF_MIGRATED
;
1739 int cpu_node
= task_node(current
);
1742 if (!numabalancing_enabled
)
1745 /* for example, ksmd faulting in a user's mm */
1749 /* Do not worry about placement if exiting */
1750 if (p
->state
== TASK_DEAD
)
1753 /* Allocate buffer to track faults on a per-node basis */
1754 if (unlikely(!p
->numa_faults_memory
)) {
1755 int size
= sizeof(*p
->numa_faults_memory
) *
1756 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
1758 p
->numa_faults_memory
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
1759 if (!p
->numa_faults_memory
)
1762 BUG_ON(p
->numa_faults_buffer_memory
);
1764 * The averaged statistics, shared & private, memory & cpu,
1765 * occupy the first half of the array. The second half of the
1766 * array is for current counters, which are averaged into the
1767 * first set by task_numa_placement.
1769 p
->numa_faults_cpu
= p
->numa_faults_memory
+ (2 * nr_node_ids
);
1770 p
->numa_faults_buffer_memory
= p
->numa_faults_memory
+ (4 * nr_node_ids
);
1771 p
->numa_faults_buffer_cpu
= p
->numa_faults_memory
+ (6 * nr_node_ids
);
1772 p
->total_numa_faults
= 0;
1773 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1777 * First accesses are treated as private, otherwise consider accesses
1778 * to be private if the accessing pid has not changed
1780 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
1783 priv
= cpupid_match_pid(p
, last_cpupid
);
1784 if (!priv
&& !(flags
& TNF_NO_GROUP
))
1785 task_numa_group(p
, last_cpupid
, flags
, &priv
);
1788 task_numa_placement(p
);
1791 * Retry task to preferred node migration periodically, in case it
1792 * case it previously failed, or the scheduler moved us.
1794 if (time_after(jiffies
, p
->numa_migrate_retry
))
1795 numa_migrate_preferred(p
);
1798 p
->numa_pages_migrated
+= pages
;
1800 p
->numa_faults_buffer_memory
[task_faults_idx(mem_node
, priv
)] += pages
;
1801 p
->numa_faults_buffer_cpu
[task_faults_idx(cpu_node
, priv
)] += pages
;
1802 p
->numa_faults_locality
[!!(flags
& TNF_FAULT_LOCAL
)] += pages
;
1805 static void reset_ptenuma_scan(struct task_struct
*p
)
1807 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
1808 p
->mm
->numa_scan_offset
= 0;
1812 * The expensive part of numa migration is done from task_work context.
1813 * Triggered from task_tick_numa().
1815 void task_numa_work(struct callback_head
*work
)
1817 unsigned long migrate
, next_scan
, now
= jiffies
;
1818 struct task_struct
*p
= current
;
1819 struct mm_struct
*mm
= p
->mm
;
1820 struct vm_area_struct
*vma
;
1821 unsigned long start
, end
;
1822 unsigned long nr_pte_updates
= 0;
1825 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
1827 work
->next
= work
; /* protect against double add */
1829 * Who cares about NUMA placement when they're dying.
1831 * NOTE: make sure not to dereference p->mm before this check,
1832 * exit_task_work() happens _after_ exit_mm() so we could be called
1833 * without p->mm even though we still had it when we enqueued this
1836 if (p
->flags
& PF_EXITING
)
1839 if (!mm
->numa_next_scan
) {
1840 mm
->numa_next_scan
= now
+
1841 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1845 * Enforce maximal scan/migration frequency..
1847 migrate
= mm
->numa_next_scan
;
1848 if (time_before(now
, migrate
))
1851 if (p
->numa_scan_period
== 0) {
1852 p
->numa_scan_period_max
= task_scan_max(p
);
1853 p
->numa_scan_period
= task_scan_min(p
);
1856 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1857 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
1861 * Delay this task enough that another task of this mm will likely win
1862 * the next time around.
1864 p
->node_stamp
+= 2 * TICK_NSEC
;
1866 start
= mm
->numa_scan_offset
;
1867 pages
= sysctl_numa_balancing_scan_size
;
1868 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
1872 down_read(&mm
->mmap_sem
);
1873 vma
= find_vma(mm
, start
);
1875 reset_ptenuma_scan(p
);
1879 for (; vma
; vma
= vma
->vm_next
) {
1880 if (!vma_migratable(vma
) || !vma_policy_mof(p
, vma
))
1884 * Shared library pages mapped by multiple processes are not
1885 * migrated as it is expected they are cache replicated. Avoid
1886 * hinting faults in read-only file-backed mappings or the vdso
1887 * as migrating the pages will be of marginal benefit.
1890 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
1894 * Skip inaccessible VMAs to avoid any confusion between
1895 * PROT_NONE and NUMA hinting ptes
1897 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
1901 start
= max(start
, vma
->vm_start
);
1902 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
1903 end
= min(end
, vma
->vm_end
);
1904 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
1907 * Scan sysctl_numa_balancing_scan_size but ensure that
1908 * at least one PTE is updated so that unused virtual
1909 * address space is quickly skipped.
1912 pages
-= (end
- start
) >> PAGE_SHIFT
;
1917 } while (end
!= vma
->vm_end
);
1922 * It is possible to reach the end of the VMA list but the last few
1923 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1924 * would find the !migratable VMA on the next scan but not reset the
1925 * scanner to the start so check it now.
1928 mm
->numa_scan_offset
= start
;
1930 reset_ptenuma_scan(p
);
1931 up_read(&mm
->mmap_sem
);
1935 * Drive the periodic memory faults..
1937 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1939 struct callback_head
*work
= &curr
->numa_work
;
1943 * We don't care about NUMA placement if we don't have memory.
1945 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
1949 * Using runtime rather than walltime has the dual advantage that
1950 * we (mostly) drive the selection from busy threads and that the
1951 * task needs to have done some actual work before we bother with
1954 now
= curr
->se
.sum_exec_runtime
;
1955 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
1957 if (now
- curr
->node_stamp
> period
) {
1958 if (!curr
->node_stamp
)
1959 curr
->numa_scan_period
= task_scan_min(curr
);
1960 curr
->node_stamp
+= period
;
1962 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
1963 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
1964 task_work_add(curr
, work
, true);
1969 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1973 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
1977 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
1980 #endif /* CONFIG_NUMA_BALANCING */
1983 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1985 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
1986 if (!parent_entity(se
))
1987 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1989 if (entity_is_task(se
)) {
1990 struct rq
*rq
= rq_of(cfs_rq
);
1992 account_numa_enqueue(rq
, task_of(se
));
1993 list_add(&se
->group_node
, &rq
->cfs_tasks
);
1996 cfs_rq
->nr_running
++;
2000 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2002 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2003 if (!parent_entity(se
))
2004 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2005 if (entity_is_task(se
)) {
2006 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2007 list_del_init(&se
->group_node
);
2009 cfs_rq
->nr_running
--;
2012 #ifdef CONFIG_FAIR_GROUP_SCHED
2014 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
2019 * Use this CPU's actual weight instead of the last load_contribution
2020 * to gain a more accurate current total weight. See
2021 * update_cfs_rq_load_contribution().
2023 tg_weight
= atomic_long_read(&tg
->load_avg
);
2024 tg_weight
-= cfs_rq
->tg_load_contrib
;
2025 tg_weight
+= cfs_rq
->load
.weight
;
2030 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2032 long tg_weight
, load
, shares
;
2034 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
2035 load
= cfs_rq
->load
.weight
;
2037 shares
= (tg
->shares
* load
);
2039 shares
/= tg_weight
;
2041 if (shares
< MIN_SHARES
)
2042 shares
= MIN_SHARES
;
2043 if (shares
> tg
->shares
)
2044 shares
= tg
->shares
;
2048 # else /* CONFIG_SMP */
2049 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2053 # endif /* CONFIG_SMP */
2054 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2055 unsigned long weight
)
2058 /* commit outstanding execution time */
2059 if (cfs_rq
->curr
== se
)
2060 update_curr(cfs_rq
);
2061 account_entity_dequeue(cfs_rq
, se
);
2064 update_load_set(&se
->load
, weight
);
2067 account_entity_enqueue(cfs_rq
, se
);
2070 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2072 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2074 struct task_group
*tg
;
2075 struct sched_entity
*se
;
2079 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2080 if (!se
|| throttled_hierarchy(cfs_rq
))
2083 if (likely(se
->load
.weight
== tg
->shares
))
2086 shares
= calc_cfs_shares(cfs_rq
, tg
);
2088 reweight_entity(cfs_rq_of(se
), se
, shares
);
2090 #else /* CONFIG_FAIR_GROUP_SCHED */
2091 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2094 #endif /* CONFIG_FAIR_GROUP_SCHED */
2098 * We choose a half-life close to 1 scheduling period.
2099 * Note: The tables below are dependent on this value.
2101 #define LOAD_AVG_PERIOD 32
2102 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2103 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2105 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2106 static const u32 runnable_avg_yN_inv
[] = {
2107 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2108 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2109 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2110 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2111 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2112 0x85aac367, 0x82cd8698,
2116 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2117 * over-estimates when re-combining.
2119 static const u32 runnable_avg_yN_sum
[] = {
2120 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2121 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2122 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2127 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2129 static __always_inline u64
decay_load(u64 val
, u64 n
)
2131 unsigned int local_n
;
2135 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2138 /* after bounds checking we can collapse to 32-bit */
2142 * As y^PERIOD = 1/2, we can combine
2143 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2144 * With a look-up table which covers k^n (n<PERIOD)
2146 * To achieve constant time decay_load.
2148 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2149 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2150 local_n
%= LOAD_AVG_PERIOD
;
2153 val
*= runnable_avg_yN_inv
[local_n
];
2154 /* We don't use SRR here since we always want to round down. */
2159 * For updates fully spanning n periods, the contribution to runnable
2160 * average will be: \Sum 1024*y^n
2162 * We can compute this reasonably efficiently by combining:
2163 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2165 static u32
__compute_runnable_contrib(u64 n
)
2169 if (likely(n
<= LOAD_AVG_PERIOD
))
2170 return runnable_avg_yN_sum
[n
];
2171 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2172 return LOAD_AVG_MAX
;
2174 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2176 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2177 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2179 n
-= LOAD_AVG_PERIOD
;
2180 } while (n
> LOAD_AVG_PERIOD
);
2182 contrib
= decay_load(contrib
, n
);
2183 return contrib
+ runnable_avg_yN_sum
[n
];
2187 * We can represent the historical contribution to runnable average as the
2188 * coefficients of a geometric series. To do this we sub-divide our runnable
2189 * history into segments of approximately 1ms (1024us); label the segment that
2190 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2192 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2194 * (now) (~1ms ago) (~2ms ago)
2196 * Let u_i denote the fraction of p_i that the entity was runnable.
2198 * We then designate the fractions u_i as our co-efficients, yielding the
2199 * following representation of historical load:
2200 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2202 * We choose y based on the with of a reasonably scheduling period, fixing:
2205 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2206 * approximately half as much as the contribution to load within the last ms
2209 * When a period "rolls over" and we have new u_0`, multiplying the previous
2210 * sum again by y is sufficient to update:
2211 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2212 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2214 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2215 struct sched_avg
*sa
,
2219 u32 runnable_contrib
;
2220 int delta_w
, decayed
= 0;
2222 delta
= now
- sa
->last_runnable_update
;
2224 * This should only happen when time goes backwards, which it
2225 * unfortunately does during sched clock init when we swap over to TSC.
2227 if ((s64
)delta
< 0) {
2228 sa
->last_runnable_update
= now
;
2233 * Use 1024ns as the unit of measurement since it's a reasonable
2234 * approximation of 1us and fast to compute.
2239 sa
->last_runnable_update
= now
;
2241 /* delta_w is the amount already accumulated against our next period */
2242 delta_w
= sa
->runnable_avg_period
% 1024;
2243 if (delta
+ delta_w
>= 1024) {
2244 /* period roll-over */
2248 * Now that we know we're crossing a period boundary, figure
2249 * out how much from delta we need to complete the current
2250 * period and accrue it.
2252 delta_w
= 1024 - delta_w
;
2254 sa
->runnable_avg_sum
+= delta_w
;
2255 sa
->runnable_avg_period
+= delta_w
;
2259 /* Figure out how many additional periods this update spans */
2260 periods
= delta
/ 1024;
2263 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2265 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2268 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2269 runnable_contrib
= __compute_runnable_contrib(periods
);
2271 sa
->runnable_avg_sum
+= runnable_contrib
;
2272 sa
->runnable_avg_period
+= runnable_contrib
;
2275 /* Remainder of delta accrued against u_0` */
2277 sa
->runnable_avg_sum
+= delta
;
2278 sa
->runnable_avg_period
+= delta
;
2283 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2284 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2286 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2287 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2289 decays
-= se
->avg
.decay_count
;
2293 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2294 se
->avg
.decay_count
= 0;
2299 #ifdef CONFIG_FAIR_GROUP_SCHED
2300 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2303 struct task_group
*tg
= cfs_rq
->tg
;
2306 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2307 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2309 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2310 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2311 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2316 * Aggregate cfs_rq runnable averages into an equivalent task_group
2317 * representation for computing load contributions.
2319 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2320 struct cfs_rq
*cfs_rq
)
2322 struct task_group
*tg
= cfs_rq
->tg
;
2325 /* The fraction of a cpu used by this cfs_rq */
2326 contrib
= div_u64((u64
)sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2327 sa
->runnable_avg_period
+ 1);
2328 contrib
-= cfs_rq
->tg_runnable_contrib
;
2330 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2331 atomic_add(contrib
, &tg
->runnable_avg
);
2332 cfs_rq
->tg_runnable_contrib
+= contrib
;
2336 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2338 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2339 struct task_group
*tg
= cfs_rq
->tg
;
2344 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2345 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2346 atomic_long_read(&tg
->load_avg
) + 1);
2349 * For group entities we need to compute a correction term in the case
2350 * that they are consuming <1 cpu so that we would contribute the same
2351 * load as a task of equal weight.
2353 * Explicitly co-ordinating this measurement would be expensive, but
2354 * fortunately the sum of each cpus contribution forms a usable
2355 * lower-bound on the true value.
2357 * Consider the aggregate of 2 contributions. Either they are disjoint
2358 * (and the sum represents true value) or they are disjoint and we are
2359 * understating by the aggregate of their overlap.
2361 * Extending this to N cpus, for a given overlap, the maximum amount we
2362 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2363 * cpus that overlap for this interval and w_i is the interval width.
2365 * On a small machine; the first term is well-bounded which bounds the
2366 * total error since w_i is a subset of the period. Whereas on a
2367 * larger machine, while this first term can be larger, if w_i is the
2368 * of consequential size guaranteed to see n_i*w_i quickly converge to
2369 * our upper bound of 1-cpu.
2371 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2372 if (runnable_avg
< NICE_0_LOAD
) {
2373 se
->avg
.load_avg_contrib
*= runnable_avg
;
2374 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2377 #else /* CONFIG_FAIR_GROUP_SCHED */
2378 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2379 int force_update
) {}
2380 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2381 struct cfs_rq
*cfs_rq
) {}
2382 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2383 #endif /* CONFIG_FAIR_GROUP_SCHED */
2385 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2389 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2390 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2391 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2392 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2395 /* Compute the current contribution to load_avg by se, return any delta */
2396 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2398 long old_contrib
= se
->avg
.load_avg_contrib
;
2400 if (entity_is_task(se
)) {
2401 __update_task_entity_contrib(se
);
2403 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2404 __update_group_entity_contrib(se
);
2407 return se
->avg
.load_avg_contrib
- old_contrib
;
2410 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2413 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2414 cfs_rq
->blocked_load_avg
-= load_contrib
;
2416 cfs_rq
->blocked_load_avg
= 0;
2419 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2421 /* Update a sched_entity's runnable average */
2422 static inline void update_entity_load_avg(struct sched_entity
*se
,
2425 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2430 * For a group entity we need to use their owned cfs_rq_clock_task() in
2431 * case they are the parent of a throttled hierarchy.
2433 if (entity_is_task(se
))
2434 now
= cfs_rq_clock_task(cfs_rq
);
2436 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2438 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2441 contrib_delta
= __update_entity_load_avg_contrib(se
);
2447 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2449 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2453 * Decay the load contributed by all blocked children and account this so that
2454 * their contribution may appropriately discounted when they wake up.
2456 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2458 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2461 decays
= now
- cfs_rq
->last_decay
;
2462 if (!decays
&& !force_update
)
2465 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2466 unsigned long removed_load
;
2467 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2468 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2472 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2474 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2475 cfs_rq
->last_decay
= now
;
2478 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2481 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2483 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2484 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2487 /* Add the load generated by se into cfs_rq's child load-average */
2488 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2489 struct sched_entity
*se
,
2493 * We track migrations using entity decay_count <= 0, on a wake-up
2494 * migration we use a negative decay count to track the remote decays
2495 * accumulated while sleeping.
2497 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2498 * are seen by enqueue_entity_load_avg() as a migration with an already
2499 * constructed load_avg_contrib.
2501 if (unlikely(se
->avg
.decay_count
<= 0)) {
2502 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2503 if (se
->avg
.decay_count
) {
2505 * In a wake-up migration we have to approximate the
2506 * time sleeping. This is because we can't synchronize
2507 * clock_task between the two cpus, and it is not
2508 * guaranteed to be read-safe. Instead, we can
2509 * approximate this using our carried decays, which are
2510 * explicitly atomically readable.
2512 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2514 update_entity_load_avg(se
, 0);
2515 /* Indicate that we're now synchronized and on-rq */
2516 se
->avg
.decay_count
= 0;
2520 __synchronize_entity_decay(se
);
2523 /* migrated tasks did not contribute to our blocked load */
2525 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2526 update_entity_load_avg(se
, 0);
2529 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2530 /* we force update consideration on load-balancer moves */
2531 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2535 * Remove se's load from this cfs_rq child load-average, if the entity is
2536 * transitioning to a blocked state we track its projected decay using
2539 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2540 struct sched_entity
*se
,
2543 update_entity_load_avg(se
, 1);
2544 /* we force update consideration on load-balancer moves */
2545 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2547 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2549 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2550 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2551 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2555 * Update the rq's load with the elapsed running time before entering
2556 * idle. if the last scheduled task is not a CFS task, idle_enter will
2557 * be the only way to update the runnable statistic.
2559 void idle_enter_fair(struct rq
*this_rq
)
2561 update_rq_runnable_avg(this_rq
, 1);
2565 * Update the rq's load with the elapsed idle time before a task is
2566 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2567 * be the only way to update the runnable statistic.
2569 void idle_exit_fair(struct rq
*this_rq
)
2571 update_rq_runnable_avg(this_rq
, 0);
2574 static int idle_balance(struct rq
*this_rq
);
2576 #else /* CONFIG_SMP */
2578 static inline void update_entity_load_avg(struct sched_entity
*se
,
2579 int update_cfs_rq
) {}
2580 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2581 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2582 struct sched_entity
*se
,
2584 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2585 struct sched_entity
*se
,
2587 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2588 int force_update
) {}
2590 static inline int idle_balance(struct rq
*rq
)
2595 #endif /* CONFIG_SMP */
2597 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2599 #ifdef CONFIG_SCHEDSTATS
2600 struct task_struct
*tsk
= NULL
;
2602 if (entity_is_task(se
))
2605 if (se
->statistics
.sleep_start
) {
2606 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2611 if (unlikely(delta
> se
->statistics
.sleep_max
))
2612 se
->statistics
.sleep_max
= delta
;
2614 se
->statistics
.sleep_start
= 0;
2615 se
->statistics
.sum_sleep_runtime
+= delta
;
2618 account_scheduler_latency(tsk
, delta
>> 10, 1);
2619 trace_sched_stat_sleep(tsk
, delta
);
2622 if (se
->statistics
.block_start
) {
2623 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2628 if (unlikely(delta
> se
->statistics
.block_max
))
2629 se
->statistics
.block_max
= delta
;
2631 se
->statistics
.block_start
= 0;
2632 se
->statistics
.sum_sleep_runtime
+= delta
;
2635 if (tsk
->in_iowait
) {
2636 se
->statistics
.iowait_sum
+= delta
;
2637 se
->statistics
.iowait_count
++;
2638 trace_sched_stat_iowait(tsk
, delta
);
2641 trace_sched_stat_blocked(tsk
, delta
);
2644 * Blocking time is in units of nanosecs, so shift by
2645 * 20 to get a milliseconds-range estimation of the
2646 * amount of time that the task spent sleeping:
2648 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2649 profile_hits(SLEEP_PROFILING
,
2650 (void *)get_wchan(tsk
),
2653 account_scheduler_latency(tsk
, delta
>> 10, 0);
2659 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2661 #ifdef CONFIG_SCHED_DEBUG
2662 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2667 if (d
> 3*sysctl_sched_latency
)
2668 schedstat_inc(cfs_rq
, nr_spread_over
);
2673 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2675 u64 vruntime
= cfs_rq
->min_vruntime
;
2678 * The 'current' period is already promised to the current tasks,
2679 * however the extra weight of the new task will slow them down a
2680 * little, place the new task so that it fits in the slot that
2681 * stays open at the end.
2683 if (initial
&& sched_feat(START_DEBIT
))
2684 vruntime
+= sched_vslice(cfs_rq
, se
);
2686 /* sleeps up to a single latency don't count. */
2688 unsigned long thresh
= sysctl_sched_latency
;
2691 * Halve their sleep time's effect, to allow
2692 * for a gentler effect of sleepers:
2694 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2700 /* ensure we never gain time by being placed backwards. */
2701 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
2704 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
2707 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2710 * Update the normalized vruntime before updating min_vruntime
2711 * through calling update_curr().
2713 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
2714 se
->vruntime
+= cfs_rq
->min_vruntime
;
2717 * Update run-time statistics of the 'current'.
2719 update_curr(cfs_rq
);
2720 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
2721 account_entity_enqueue(cfs_rq
, se
);
2722 update_cfs_shares(cfs_rq
);
2724 if (flags
& ENQUEUE_WAKEUP
) {
2725 place_entity(cfs_rq
, se
, 0);
2726 enqueue_sleeper(cfs_rq
, se
);
2729 update_stats_enqueue(cfs_rq
, se
);
2730 check_spread(cfs_rq
, se
);
2731 if (se
!= cfs_rq
->curr
)
2732 __enqueue_entity(cfs_rq
, se
);
2735 if (cfs_rq
->nr_running
== 1) {
2736 list_add_leaf_cfs_rq(cfs_rq
);
2737 check_enqueue_throttle(cfs_rq
);
2741 static void __clear_buddies_last(struct sched_entity
*se
)
2743 for_each_sched_entity(se
) {
2744 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2745 if (cfs_rq
->last
!= se
)
2748 cfs_rq
->last
= NULL
;
2752 static void __clear_buddies_next(struct sched_entity
*se
)
2754 for_each_sched_entity(se
) {
2755 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2756 if (cfs_rq
->next
!= se
)
2759 cfs_rq
->next
= NULL
;
2763 static void __clear_buddies_skip(struct sched_entity
*se
)
2765 for_each_sched_entity(se
) {
2766 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2767 if (cfs_rq
->skip
!= se
)
2770 cfs_rq
->skip
= NULL
;
2774 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2776 if (cfs_rq
->last
== se
)
2777 __clear_buddies_last(se
);
2779 if (cfs_rq
->next
== se
)
2780 __clear_buddies_next(se
);
2782 if (cfs_rq
->skip
== se
)
2783 __clear_buddies_skip(se
);
2786 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2789 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2792 * Update run-time statistics of the 'current'.
2794 update_curr(cfs_rq
);
2795 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
2797 update_stats_dequeue(cfs_rq
, se
);
2798 if (flags
& DEQUEUE_SLEEP
) {
2799 #ifdef CONFIG_SCHEDSTATS
2800 if (entity_is_task(se
)) {
2801 struct task_struct
*tsk
= task_of(se
);
2803 if (tsk
->state
& TASK_INTERRUPTIBLE
)
2804 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
2805 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
2806 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
2811 clear_buddies(cfs_rq
, se
);
2813 if (se
!= cfs_rq
->curr
)
2814 __dequeue_entity(cfs_rq
, se
);
2816 account_entity_dequeue(cfs_rq
, se
);
2819 * Normalize the entity after updating the min_vruntime because the
2820 * update can refer to the ->curr item and we need to reflect this
2821 * movement in our normalized position.
2823 if (!(flags
& DEQUEUE_SLEEP
))
2824 se
->vruntime
-= cfs_rq
->min_vruntime
;
2826 /* return excess runtime on last dequeue */
2827 return_cfs_rq_runtime(cfs_rq
);
2829 update_min_vruntime(cfs_rq
);
2830 update_cfs_shares(cfs_rq
);
2834 * Preempt the current task with a newly woken task if needed:
2837 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2839 unsigned long ideal_runtime
, delta_exec
;
2840 struct sched_entity
*se
;
2843 ideal_runtime
= sched_slice(cfs_rq
, curr
);
2844 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
2845 if (delta_exec
> ideal_runtime
) {
2846 resched_task(rq_of(cfs_rq
)->curr
);
2848 * The current task ran long enough, ensure it doesn't get
2849 * re-elected due to buddy favours.
2851 clear_buddies(cfs_rq
, curr
);
2856 * Ensure that a task that missed wakeup preemption by a
2857 * narrow margin doesn't have to wait for a full slice.
2858 * This also mitigates buddy induced latencies under load.
2860 if (delta_exec
< sysctl_sched_min_granularity
)
2863 se
= __pick_first_entity(cfs_rq
);
2864 delta
= curr
->vruntime
- se
->vruntime
;
2869 if (delta
> ideal_runtime
)
2870 resched_task(rq_of(cfs_rq
)->curr
);
2874 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2876 /* 'current' is not kept within the tree. */
2879 * Any task has to be enqueued before it get to execute on
2880 * a CPU. So account for the time it spent waiting on the
2883 update_stats_wait_end(cfs_rq
, se
);
2884 __dequeue_entity(cfs_rq
, se
);
2887 update_stats_curr_start(cfs_rq
, se
);
2889 #ifdef CONFIG_SCHEDSTATS
2891 * Track our maximum slice length, if the CPU's load is at
2892 * least twice that of our own weight (i.e. dont track it
2893 * when there are only lesser-weight tasks around):
2895 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
2896 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
2897 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
2900 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
2904 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
2907 * Pick the next process, keeping these things in mind, in this order:
2908 * 1) keep things fair between processes/task groups
2909 * 2) pick the "next" process, since someone really wants that to run
2910 * 3) pick the "last" process, for cache locality
2911 * 4) do not run the "skip" process, if something else is available
2913 static struct sched_entity
*
2914 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2916 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
2917 struct sched_entity
*se
;
2920 * If curr is set we have to see if its left of the leftmost entity
2921 * still in the tree, provided there was anything in the tree at all.
2923 if (!left
|| (curr
&& entity_before(curr
, left
)))
2926 se
= left
; /* ideally we run the leftmost entity */
2929 * Avoid running the skip buddy, if running something else can
2930 * be done without getting too unfair.
2932 if (cfs_rq
->skip
== se
) {
2933 struct sched_entity
*second
;
2936 second
= __pick_first_entity(cfs_rq
);
2938 second
= __pick_next_entity(se
);
2939 if (!second
|| (curr
&& entity_before(curr
, second
)))
2943 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
2948 * Prefer last buddy, try to return the CPU to a preempted task.
2950 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
2954 * Someone really wants this to run. If it's not unfair, run it.
2956 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
2959 clear_buddies(cfs_rq
, se
);
2964 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2966 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
2969 * If still on the runqueue then deactivate_task()
2970 * was not called and update_curr() has to be done:
2973 update_curr(cfs_rq
);
2975 /* throttle cfs_rqs exceeding runtime */
2976 check_cfs_rq_runtime(cfs_rq
);
2978 check_spread(cfs_rq
, prev
);
2980 update_stats_wait_start(cfs_rq
, prev
);
2981 /* Put 'current' back into the tree. */
2982 __enqueue_entity(cfs_rq
, prev
);
2983 /* in !on_rq case, update occurred at dequeue */
2984 update_entity_load_avg(prev
, 1);
2986 cfs_rq
->curr
= NULL
;
2990 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
2993 * Update run-time statistics of the 'current'.
2995 update_curr(cfs_rq
);
2998 * Ensure that runnable average is periodically updated.
3000 update_entity_load_avg(curr
, 1);
3001 update_cfs_rq_blocked_load(cfs_rq
, 1);
3002 update_cfs_shares(cfs_rq
);
3004 #ifdef CONFIG_SCHED_HRTICK
3006 * queued ticks are scheduled to match the slice, so don't bother
3007 * validating it and just reschedule.
3010 resched_task(rq_of(cfs_rq
)->curr
);
3014 * don't let the period tick interfere with the hrtick preemption
3016 if (!sched_feat(DOUBLE_TICK
) &&
3017 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3021 if (cfs_rq
->nr_running
> 1)
3022 check_preempt_tick(cfs_rq
, curr
);
3026 /**************************************************
3027 * CFS bandwidth control machinery
3030 #ifdef CONFIG_CFS_BANDWIDTH
3032 #ifdef HAVE_JUMP_LABEL
3033 static struct static_key __cfs_bandwidth_used
;
3035 static inline bool cfs_bandwidth_used(void)
3037 return static_key_false(&__cfs_bandwidth_used
);
3040 void cfs_bandwidth_usage_inc(void)
3042 static_key_slow_inc(&__cfs_bandwidth_used
);
3045 void cfs_bandwidth_usage_dec(void)
3047 static_key_slow_dec(&__cfs_bandwidth_used
);
3049 #else /* HAVE_JUMP_LABEL */
3050 static bool cfs_bandwidth_used(void)
3055 void cfs_bandwidth_usage_inc(void) {}
3056 void cfs_bandwidth_usage_dec(void) {}
3057 #endif /* HAVE_JUMP_LABEL */
3060 * default period for cfs group bandwidth.
3061 * default: 0.1s, units: nanoseconds
3063 static inline u64
default_cfs_period(void)
3065 return 100000000ULL;
3068 static inline u64
sched_cfs_bandwidth_slice(void)
3070 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3074 * Replenish runtime according to assigned quota and update expiration time.
3075 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3076 * additional synchronization around rq->lock.
3078 * requires cfs_b->lock
3080 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3084 if (cfs_b
->quota
== RUNTIME_INF
)
3087 now
= sched_clock_cpu(smp_processor_id());
3088 cfs_b
->runtime
= cfs_b
->quota
;
3089 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3092 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3094 return &tg
->cfs_bandwidth
;
3097 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3098 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3100 if (unlikely(cfs_rq
->throttle_count
))
3101 return cfs_rq
->throttled_clock_task
;
3103 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3106 /* returns 0 on failure to allocate runtime */
3107 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3109 struct task_group
*tg
= cfs_rq
->tg
;
3110 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3111 u64 amount
= 0, min_amount
, expires
;
3113 /* note: this is a positive sum as runtime_remaining <= 0 */
3114 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3116 raw_spin_lock(&cfs_b
->lock
);
3117 if (cfs_b
->quota
== RUNTIME_INF
)
3118 amount
= min_amount
;
3121 * If the bandwidth pool has become inactive, then at least one
3122 * period must have elapsed since the last consumption.
3123 * Refresh the global state and ensure bandwidth timer becomes
3126 if (!cfs_b
->timer_active
) {
3127 __refill_cfs_bandwidth_runtime(cfs_b
);
3128 __start_cfs_bandwidth(cfs_b
);
3131 if (cfs_b
->runtime
> 0) {
3132 amount
= min(cfs_b
->runtime
, min_amount
);
3133 cfs_b
->runtime
-= amount
;
3137 expires
= cfs_b
->runtime_expires
;
3138 raw_spin_unlock(&cfs_b
->lock
);
3140 cfs_rq
->runtime_remaining
+= amount
;
3142 * we may have advanced our local expiration to account for allowed
3143 * spread between our sched_clock and the one on which runtime was
3146 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3147 cfs_rq
->runtime_expires
= expires
;
3149 return cfs_rq
->runtime_remaining
> 0;
3153 * Note: This depends on the synchronization provided by sched_clock and the
3154 * fact that rq->clock snapshots this value.
3156 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3158 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3160 /* if the deadline is ahead of our clock, nothing to do */
3161 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3164 if (cfs_rq
->runtime_remaining
< 0)
3168 * If the local deadline has passed we have to consider the
3169 * possibility that our sched_clock is 'fast' and the global deadline
3170 * has not truly expired.
3172 * Fortunately we can check determine whether this the case by checking
3173 * whether the global deadline has advanced.
3176 if ((s64
)(cfs_rq
->runtime_expires
- cfs_b
->runtime_expires
) >= 0) {
3177 /* extend local deadline, drift is bounded above by 2 ticks */
3178 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3180 /* global deadline is ahead, expiration has passed */
3181 cfs_rq
->runtime_remaining
= 0;
3185 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3187 /* dock delta_exec before expiring quota (as it could span periods) */
3188 cfs_rq
->runtime_remaining
-= delta_exec
;
3189 expire_cfs_rq_runtime(cfs_rq
);
3191 if (likely(cfs_rq
->runtime_remaining
> 0))
3195 * if we're unable to extend our runtime we resched so that the active
3196 * hierarchy can be throttled
3198 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3199 resched_task(rq_of(cfs_rq
)->curr
);
3202 static __always_inline
3203 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3205 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3208 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3211 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3213 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3216 /* check whether cfs_rq, or any parent, is throttled */
3217 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3219 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3223 * Ensure that neither of the group entities corresponding to src_cpu or
3224 * dest_cpu are members of a throttled hierarchy when performing group
3225 * load-balance operations.
3227 static inline int throttled_lb_pair(struct task_group
*tg
,
3228 int src_cpu
, int dest_cpu
)
3230 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3232 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3233 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3235 return throttled_hierarchy(src_cfs_rq
) ||
3236 throttled_hierarchy(dest_cfs_rq
);
3239 /* updated child weight may affect parent so we have to do this bottom up */
3240 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3242 struct rq
*rq
= data
;
3243 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3245 cfs_rq
->throttle_count
--;
3247 if (!cfs_rq
->throttle_count
) {
3248 /* adjust cfs_rq_clock_task() */
3249 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3250 cfs_rq
->throttled_clock_task
;
3257 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3259 struct rq
*rq
= data
;
3260 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3262 /* group is entering throttled state, stop time */
3263 if (!cfs_rq
->throttle_count
)
3264 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3265 cfs_rq
->throttle_count
++;
3270 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3272 struct rq
*rq
= rq_of(cfs_rq
);
3273 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3274 struct sched_entity
*se
;
3275 long task_delta
, dequeue
= 1;
3277 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3279 /* freeze hierarchy runnable averages while throttled */
3281 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3284 task_delta
= cfs_rq
->h_nr_running
;
3285 for_each_sched_entity(se
) {
3286 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3287 /* throttled entity or throttle-on-deactivate */
3292 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3293 qcfs_rq
->h_nr_running
-= task_delta
;
3295 if (qcfs_rq
->load
.weight
)
3300 rq
->nr_running
-= task_delta
;
3302 cfs_rq
->throttled
= 1;
3303 cfs_rq
->throttled_clock
= rq_clock(rq
);
3304 raw_spin_lock(&cfs_b
->lock
);
3305 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3306 if (!cfs_b
->timer_active
)
3307 __start_cfs_bandwidth(cfs_b
);
3308 raw_spin_unlock(&cfs_b
->lock
);
3311 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3313 struct rq
*rq
= rq_of(cfs_rq
);
3314 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3315 struct sched_entity
*se
;
3319 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3321 cfs_rq
->throttled
= 0;
3323 update_rq_clock(rq
);
3325 raw_spin_lock(&cfs_b
->lock
);
3326 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3327 list_del_rcu(&cfs_rq
->throttled_list
);
3328 raw_spin_unlock(&cfs_b
->lock
);
3330 /* update hierarchical throttle state */
3331 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3333 if (!cfs_rq
->load
.weight
)
3336 task_delta
= cfs_rq
->h_nr_running
;
3337 for_each_sched_entity(se
) {
3341 cfs_rq
= cfs_rq_of(se
);
3343 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3344 cfs_rq
->h_nr_running
+= task_delta
;
3346 if (cfs_rq_throttled(cfs_rq
))
3351 rq
->nr_running
+= task_delta
;
3353 /* determine whether we need to wake up potentially idle cpu */
3354 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3355 resched_task(rq
->curr
);
3358 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3359 u64 remaining
, u64 expires
)
3361 struct cfs_rq
*cfs_rq
;
3362 u64 runtime
= remaining
;
3365 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3367 struct rq
*rq
= rq_of(cfs_rq
);
3369 raw_spin_lock(&rq
->lock
);
3370 if (!cfs_rq_throttled(cfs_rq
))
3373 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3374 if (runtime
> remaining
)
3375 runtime
= remaining
;
3376 remaining
-= runtime
;
3378 cfs_rq
->runtime_remaining
+= runtime
;
3379 cfs_rq
->runtime_expires
= expires
;
3381 /* we check whether we're throttled above */
3382 if (cfs_rq
->runtime_remaining
> 0)
3383 unthrottle_cfs_rq(cfs_rq
);
3386 raw_spin_unlock(&rq
->lock
);
3397 * Responsible for refilling a task_group's bandwidth and unthrottling its
3398 * cfs_rqs as appropriate. If there has been no activity within the last
3399 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3400 * used to track this state.
3402 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3404 u64 runtime
, runtime_expires
;
3405 int idle
= 1, throttled
;
3407 raw_spin_lock(&cfs_b
->lock
);
3408 /* no need to continue the timer with no bandwidth constraint */
3409 if (cfs_b
->quota
== RUNTIME_INF
)
3412 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3413 /* idle depends on !throttled (for the case of a large deficit) */
3414 idle
= cfs_b
->idle
&& !throttled
;
3415 cfs_b
->nr_periods
+= overrun
;
3417 /* if we're going inactive then everything else can be deferred */
3422 * if we have relooped after returning idle once, we need to update our
3423 * status as actually running, so that other cpus doing
3424 * __start_cfs_bandwidth will stop trying to cancel us.
3426 cfs_b
->timer_active
= 1;
3428 __refill_cfs_bandwidth_runtime(cfs_b
);
3431 /* mark as potentially idle for the upcoming period */
3436 /* account preceding periods in which throttling occurred */
3437 cfs_b
->nr_throttled
+= overrun
;
3440 * There are throttled entities so we must first use the new bandwidth
3441 * to unthrottle them before making it generally available. This
3442 * ensures that all existing debts will be paid before a new cfs_rq is
3445 runtime
= cfs_b
->runtime
;
3446 runtime_expires
= cfs_b
->runtime_expires
;
3450 * This check is repeated as we are holding onto the new bandwidth
3451 * while we unthrottle. This can potentially race with an unthrottled
3452 * group trying to acquire new bandwidth from the global pool.
3454 while (throttled
&& runtime
> 0) {
3455 raw_spin_unlock(&cfs_b
->lock
);
3456 /* we can't nest cfs_b->lock while distributing bandwidth */
3457 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3459 raw_spin_lock(&cfs_b
->lock
);
3461 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3464 /* return (any) remaining runtime */
3465 cfs_b
->runtime
= runtime
;
3467 * While we are ensured activity in the period following an
3468 * unthrottle, this also covers the case in which the new bandwidth is
3469 * insufficient to cover the existing bandwidth deficit. (Forcing the
3470 * timer to remain active while there are any throttled entities.)
3475 cfs_b
->timer_active
= 0;
3476 raw_spin_unlock(&cfs_b
->lock
);
3481 /* a cfs_rq won't donate quota below this amount */
3482 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3483 /* minimum remaining period time to redistribute slack quota */
3484 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3485 /* how long we wait to gather additional slack before distributing */
3486 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3489 * Are we near the end of the current quota period?
3491 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3492 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3493 * migrate_hrtimers, base is never cleared, so we are fine.
3495 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3497 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3500 /* if the call-back is running a quota refresh is already occurring */
3501 if (hrtimer_callback_running(refresh_timer
))
3504 /* is a quota refresh about to occur? */
3505 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3506 if (remaining
< min_expire
)
3512 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3514 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3516 /* if there's a quota refresh soon don't bother with slack */
3517 if (runtime_refresh_within(cfs_b
, min_left
))
3520 start_bandwidth_timer(&cfs_b
->slack_timer
,
3521 ns_to_ktime(cfs_bandwidth_slack_period
));
3524 /* we know any runtime found here is valid as update_curr() precedes return */
3525 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3527 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3528 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3530 if (slack_runtime
<= 0)
3533 raw_spin_lock(&cfs_b
->lock
);
3534 if (cfs_b
->quota
!= RUNTIME_INF
&&
3535 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3536 cfs_b
->runtime
+= slack_runtime
;
3538 /* we are under rq->lock, defer unthrottling using a timer */
3539 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3540 !list_empty(&cfs_b
->throttled_cfs_rq
))
3541 start_cfs_slack_bandwidth(cfs_b
);
3543 raw_spin_unlock(&cfs_b
->lock
);
3545 /* even if it's not valid for return we don't want to try again */
3546 cfs_rq
->runtime_remaining
-= slack_runtime
;
3549 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3551 if (!cfs_bandwidth_used())
3554 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3557 __return_cfs_rq_runtime(cfs_rq
);
3561 * This is done with a timer (instead of inline with bandwidth return) since
3562 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3564 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3566 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3569 /* confirm we're still not at a refresh boundary */
3570 raw_spin_lock(&cfs_b
->lock
);
3571 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
3572 raw_spin_unlock(&cfs_b
->lock
);
3576 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
) {
3577 runtime
= cfs_b
->runtime
;
3580 expires
= cfs_b
->runtime_expires
;
3581 raw_spin_unlock(&cfs_b
->lock
);
3586 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3588 raw_spin_lock(&cfs_b
->lock
);
3589 if (expires
== cfs_b
->runtime_expires
)
3590 cfs_b
->runtime
= runtime
;
3591 raw_spin_unlock(&cfs_b
->lock
);
3595 * When a group wakes up we want to make sure that its quota is not already
3596 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3597 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3599 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3601 if (!cfs_bandwidth_used())
3604 /* an active group must be handled by the update_curr()->put() path */
3605 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3608 /* ensure the group is not already throttled */
3609 if (cfs_rq_throttled(cfs_rq
))
3612 /* update runtime allocation */
3613 account_cfs_rq_runtime(cfs_rq
, 0);
3614 if (cfs_rq
->runtime_remaining
<= 0)
3615 throttle_cfs_rq(cfs_rq
);
3618 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3619 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3621 if (!cfs_bandwidth_used())
3624 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3628 * it's possible for a throttled entity to be forced into a running
3629 * state (e.g. set_curr_task), in this case we're finished.
3631 if (cfs_rq_throttled(cfs_rq
))
3634 throttle_cfs_rq(cfs_rq
);
3638 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3640 struct cfs_bandwidth
*cfs_b
=
3641 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3642 do_sched_cfs_slack_timer(cfs_b
);
3644 return HRTIMER_NORESTART
;
3647 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3649 struct cfs_bandwidth
*cfs_b
=
3650 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3656 now
= hrtimer_cb_get_time(timer
);
3657 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3662 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3665 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3668 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3670 raw_spin_lock_init(&cfs_b
->lock
);
3672 cfs_b
->quota
= RUNTIME_INF
;
3673 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3675 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3676 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3677 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3678 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3679 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3682 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3684 cfs_rq
->runtime_enabled
= 0;
3685 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3688 /* requires cfs_b->lock, may release to reprogram timer */
3689 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3692 * The timer may be active because we're trying to set a new bandwidth
3693 * period or because we're racing with the tear-down path
3694 * (timer_active==0 becomes visible before the hrtimer call-back
3695 * terminates). In either case we ensure that it's re-programmed
3697 while (unlikely(hrtimer_active(&cfs_b
->period_timer
)) &&
3698 hrtimer_try_to_cancel(&cfs_b
->period_timer
) < 0) {
3699 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3700 raw_spin_unlock(&cfs_b
->lock
);
3702 raw_spin_lock(&cfs_b
->lock
);
3703 /* if someone else restarted the timer then we're done */
3704 if (cfs_b
->timer_active
)
3708 cfs_b
->timer_active
= 1;
3709 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
3712 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3714 hrtimer_cancel(&cfs_b
->period_timer
);
3715 hrtimer_cancel(&cfs_b
->slack_timer
);
3718 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
3720 struct cfs_rq
*cfs_rq
;
3722 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3723 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3725 if (!cfs_rq
->runtime_enabled
)
3729 * clock_task is not advancing so we just need to make sure
3730 * there's some valid quota amount
3732 cfs_rq
->runtime_remaining
= cfs_b
->quota
;
3733 if (cfs_rq_throttled(cfs_rq
))
3734 unthrottle_cfs_rq(cfs_rq
);
3738 #else /* CONFIG_CFS_BANDWIDTH */
3739 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3741 return rq_clock_task(rq_of(cfs_rq
));
3744 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
3745 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
3746 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
3747 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3749 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3754 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3759 static inline int throttled_lb_pair(struct task_group
*tg
,
3760 int src_cpu
, int dest_cpu
)
3765 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3767 #ifdef CONFIG_FAIR_GROUP_SCHED
3768 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3771 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3775 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3776 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
3778 #endif /* CONFIG_CFS_BANDWIDTH */
3780 /**************************************************
3781 * CFS operations on tasks:
3784 #ifdef CONFIG_SCHED_HRTICK
3785 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3787 struct sched_entity
*se
= &p
->se
;
3788 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3790 WARN_ON(task_rq(p
) != rq
);
3792 if (cfs_rq
->nr_running
> 1) {
3793 u64 slice
= sched_slice(cfs_rq
, se
);
3794 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
3795 s64 delta
= slice
- ran
;
3804 * Don't schedule slices shorter than 10000ns, that just
3805 * doesn't make sense. Rely on vruntime for fairness.
3808 delta
= max_t(s64
, 10000LL, delta
);
3810 hrtick_start(rq
, delta
);
3815 * called from enqueue/dequeue and updates the hrtick when the
3816 * current task is from our class and nr_running is low enough
3819 static void hrtick_update(struct rq
*rq
)
3821 struct task_struct
*curr
= rq
->curr
;
3823 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
3826 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
3827 hrtick_start_fair(rq
, curr
);
3829 #else /* !CONFIG_SCHED_HRTICK */
3831 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3835 static inline void hrtick_update(struct rq
*rq
)
3841 * The enqueue_task method is called before nr_running is
3842 * increased. Here we update the fair scheduling stats and
3843 * then put the task into the rbtree:
3846 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3848 struct cfs_rq
*cfs_rq
;
3849 struct sched_entity
*se
= &p
->se
;
3851 for_each_sched_entity(se
) {
3854 cfs_rq
= cfs_rq_of(se
);
3855 enqueue_entity(cfs_rq
, se
, flags
);
3858 * end evaluation on encountering a throttled cfs_rq
3860 * note: in the case of encountering a throttled cfs_rq we will
3861 * post the final h_nr_running increment below.
3863 if (cfs_rq_throttled(cfs_rq
))
3865 cfs_rq
->h_nr_running
++;
3867 flags
= ENQUEUE_WAKEUP
;
3870 for_each_sched_entity(se
) {
3871 cfs_rq
= cfs_rq_of(se
);
3872 cfs_rq
->h_nr_running
++;
3874 if (cfs_rq_throttled(cfs_rq
))
3877 update_cfs_shares(cfs_rq
);
3878 update_entity_load_avg(se
, 1);
3882 update_rq_runnable_avg(rq
, rq
->nr_running
);
3888 static void set_next_buddy(struct sched_entity
*se
);
3891 * The dequeue_task method is called before nr_running is
3892 * decreased. We remove the task from the rbtree and
3893 * update the fair scheduling stats:
3895 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3897 struct cfs_rq
*cfs_rq
;
3898 struct sched_entity
*se
= &p
->se
;
3899 int task_sleep
= flags
& DEQUEUE_SLEEP
;
3901 for_each_sched_entity(se
) {
3902 cfs_rq
= cfs_rq_of(se
);
3903 dequeue_entity(cfs_rq
, se
, flags
);
3906 * end evaluation on encountering a throttled cfs_rq
3908 * note: in the case of encountering a throttled cfs_rq we will
3909 * post the final h_nr_running decrement below.
3911 if (cfs_rq_throttled(cfs_rq
))
3913 cfs_rq
->h_nr_running
--;
3915 /* Don't dequeue parent if it has other entities besides us */
3916 if (cfs_rq
->load
.weight
) {
3918 * Bias pick_next to pick a task from this cfs_rq, as
3919 * p is sleeping when it is within its sched_slice.
3921 if (task_sleep
&& parent_entity(se
))
3922 set_next_buddy(parent_entity(se
));
3924 /* avoid re-evaluating load for this entity */
3925 se
= parent_entity(se
);
3928 flags
|= DEQUEUE_SLEEP
;
3931 for_each_sched_entity(se
) {
3932 cfs_rq
= cfs_rq_of(se
);
3933 cfs_rq
->h_nr_running
--;
3935 if (cfs_rq_throttled(cfs_rq
))
3938 update_cfs_shares(cfs_rq
);
3939 update_entity_load_avg(se
, 1);
3944 update_rq_runnable_avg(rq
, 1);
3950 /* Used instead of source_load when we know the type == 0 */
3951 static unsigned long weighted_cpuload(const int cpu
)
3953 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
3957 * Return a low guess at the load of a migration-source cpu weighted
3958 * according to the scheduling class and "nice" value.
3960 * We want to under-estimate the load of migration sources, to
3961 * balance conservatively.
3963 static unsigned long source_load(int cpu
, int type
)
3965 struct rq
*rq
= cpu_rq(cpu
);
3966 unsigned long total
= weighted_cpuload(cpu
);
3968 if (type
== 0 || !sched_feat(LB_BIAS
))
3971 return min(rq
->cpu_load
[type
-1], total
);
3975 * Return a high guess at the load of a migration-target cpu weighted
3976 * according to the scheduling class and "nice" value.
3978 static unsigned long target_load(int cpu
, int type
)
3980 struct rq
*rq
= cpu_rq(cpu
);
3981 unsigned long total
= weighted_cpuload(cpu
);
3983 if (type
== 0 || !sched_feat(LB_BIAS
))
3986 return max(rq
->cpu_load
[type
-1], total
);
3989 static unsigned long power_of(int cpu
)
3991 return cpu_rq(cpu
)->cpu_power
;
3994 static unsigned long cpu_avg_load_per_task(int cpu
)
3996 struct rq
*rq
= cpu_rq(cpu
);
3997 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
3998 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
4001 return load_avg
/ nr_running
;
4006 static void record_wakee(struct task_struct
*p
)
4009 * Rough decay (wiping) for cost saving, don't worry
4010 * about the boundary, really active task won't care
4013 if (jiffies
> current
->wakee_flip_decay_ts
+ HZ
) {
4014 current
->wakee_flips
= 0;
4015 current
->wakee_flip_decay_ts
= jiffies
;
4018 if (current
->last_wakee
!= p
) {
4019 current
->last_wakee
= p
;
4020 current
->wakee_flips
++;
4024 static void task_waking_fair(struct task_struct
*p
)
4026 struct sched_entity
*se
= &p
->se
;
4027 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4030 #ifndef CONFIG_64BIT
4031 u64 min_vruntime_copy
;
4034 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
4036 min_vruntime
= cfs_rq
->min_vruntime
;
4037 } while (min_vruntime
!= min_vruntime_copy
);
4039 min_vruntime
= cfs_rq
->min_vruntime
;
4042 se
->vruntime
-= min_vruntime
;
4046 #ifdef CONFIG_FAIR_GROUP_SCHED
4048 * effective_load() calculates the load change as seen from the root_task_group
4050 * Adding load to a group doesn't make a group heavier, but can cause movement
4051 * of group shares between cpus. Assuming the shares were perfectly aligned one
4052 * can calculate the shift in shares.
4054 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4055 * on this @cpu and results in a total addition (subtraction) of @wg to the
4056 * total group weight.
4058 * Given a runqueue weight distribution (rw_i) we can compute a shares
4059 * distribution (s_i) using:
4061 * s_i = rw_i / \Sum rw_j (1)
4063 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4064 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4065 * shares distribution (s_i):
4067 * rw_i = { 2, 4, 1, 0 }
4068 * s_i = { 2/7, 4/7, 1/7, 0 }
4070 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4071 * task used to run on and the CPU the waker is running on), we need to
4072 * compute the effect of waking a task on either CPU and, in case of a sync
4073 * wakeup, compute the effect of the current task going to sleep.
4075 * So for a change of @wl to the local @cpu with an overall group weight change
4076 * of @wl we can compute the new shares distribution (s'_i) using:
4078 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4080 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4081 * differences in waking a task to CPU 0. The additional task changes the
4082 * weight and shares distributions like:
4084 * rw'_i = { 3, 4, 1, 0 }
4085 * s'_i = { 3/8, 4/8, 1/8, 0 }
4087 * We can then compute the difference in effective weight by using:
4089 * dw_i = S * (s'_i - s_i) (3)
4091 * Where 'S' is the group weight as seen by its parent.
4093 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4094 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4095 * 4/7) times the weight of the group.
4097 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4099 struct sched_entity
*se
= tg
->se
[cpu
];
4101 if (!tg
->parent
) /* the trivial, non-cgroup case */
4104 for_each_sched_entity(se
) {
4110 * W = @wg + \Sum rw_j
4112 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4117 w
= se
->my_q
->load
.weight
+ wl
;
4120 * wl = S * s'_i; see (2)
4123 wl
= (w
* tg
->shares
) / W
;
4128 * Per the above, wl is the new se->load.weight value; since
4129 * those are clipped to [MIN_SHARES, ...) do so now. See
4130 * calc_cfs_shares().
4132 if (wl
< MIN_SHARES
)
4136 * wl = dw_i = S * (s'_i - s_i); see (3)
4138 wl
-= se
->load
.weight
;
4141 * Recursively apply this logic to all parent groups to compute
4142 * the final effective load change on the root group. Since
4143 * only the @tg group gets extra weight, all parent groups can
4144 * only redistribute existing shares. @wl is the shift in shares
4145 * resulting from this level per the above.
4154 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4161 static int wake_wide(struct task_struct
*p
)
4163 int factor
= this_cpu_read(sd_llc_size
);
4166 * Yeah, it's the switching-frequency, could means many wakee or
4167 * rapidly switch, use factor here will just help to automatically
4168 * adjust the loose-degree, so bigger node will lead to more pull.
4170 if (p
->wakee_flips
> factor
) {
4172 * wakee is somewhat hot, it needs certain amount of cpu
4173 * resource, so if waker is far more hot, prefer to leave
4176 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
4183 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4185 s64 this_load
, load
;
4186 int idx
, this_cpu
, prev_cpu
;
4187 unsigned long tl_per_task
;
4188 struct task_group
*tg
;
4189 unsigned long weight
;
4193 * If we wake multiple tasks be careful to not bounce
4194 * ourselves around too much.
4200 this_cpu
= smp_processor_id();
4201 prev_cpu
= task_cpu(p
);
4202 load
= source_load(prev_cpu
, idx
);
4203 this_load
= target_load(this_cpu
, idx
);
4206 * If sync wakeup then subtract the (maximum possible)
4207 * effect of the currently running task from the load
4208 * of the current CPU:
4211 tg
= task_group(current
);
4212 weight
= current
->se
.load
.weight
;
4214 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4215 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4219 weight
= p
->se
.load
.weight
;
4222 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4223 * due to the sync cause above having dropped this_load to 0, we'll
4224 * always have an imbalance, but there's really nothing you can do
4225 * about that, so that's good too.
4227 * Otherwise check if either cpus are near enough in load to allow this
4228 * task to be woken on this_cpu.
4230 if (this_load
> 0) {
4231 s64 this_eff_load
, prev_eff_load
;
4233 this_eff_load
= 100;
4234 this_eff_load
*= power_of(prev_cpu
);
4235 this_eff_load
*= this_load
+
4236 effective_load(tg
, this_cpu
, weight
, weight
);
4238 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4239 prev_eff_load
*= power_of(this_cpu
);
4240 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4242 balanced
= this_eff_load
<= prev_eff_load
;
4247 * If the currently running task will sleep within
4248 * a reasonable amount of time then attract this newly
4251 if (sync
&& balanced
)
4254 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4255 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
4258 (this_load
<= load
&&
4259 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
4261 * This domain has SD_WAKE_AFFINE and
4262 * p is cache cold in this domain, and
4263 * there is no bad imbalance.
4265 schedstat_inc(sd
, ttwu_move_affine
);
4266 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4274 * find_idlest_group finds and returns the least busy CPU group within the
4277 static struct sched_group
*
4278 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4279 int this_cpu
, int sd_flag
)
4281 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4282 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4283 int load_idx
= sd
->forkexec_idx
;
4284 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4286 if (sd_flag
& SD_BALANCE_WAKE
)
4287 load_idx
= sd
->wake_idx
;
4290 unsigned long load
, avg_load
;
4294 /* Skip over this group if it has no CPUs allowed */
4295 if (!cpumask_intersects(sched_group_cpus(group
),
4296 tsk_cpus_allowed(p
)))
4299 local_group
= cpumask_test_cpu(this_cpu
,
4300 sched_group_cpus(group
));
4302 /* Tally up the load of all CPUs in the group */
4305 for_each_cpu(i
, sched_group_cpus(group
)) {
4306 /* Bias balancing toward cpus of our domain */
4308 load
= source_load(i
, load_idx
);
4310 load
= target_load(i
, load_idx
);
4315 /* Adjust by relative CPU power of the group */
4316 avg_load
= (avg_load
* SCHED_POWER_SCALE
) / group
->sgp
->power
;
4319 this_load
= avg_load
;
4320 } else if (avg_load
< min_load
) {
4321 min_load
= avg_load
;
4324 } while (group
= group
->next
, group
!= sd
->groups
);
4326 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4332 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4335 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4337 unsigned long load
, min_load
= ULONG_MAX
;
4341 /* Traverse only the allowed CPUs */
4342 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4343 load
= weighted_cpuload(i
);
4345 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4355 * Try and locate an idle CPU in the sched_domain.
4357 static int select_idle_sibling(struct task_struct
*p
, int target
)
4359 struct sched_domain
*sd
;
4360 struct sched_group
*sg
;
4361 int i
= task_cpu(p
);
4363 if (idle_cpu(target
))
4367 * If the prevous cpu is cache affine and idle, don't be stupid.
4369 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4373 * Otherwise, iterate the domains and find an elegible idle cpu.
4375 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4376 for_each_lower_domain(sd
) {
4379 if (!cpumask_intersects(sched_group_cpus(sg
),
4380 tsk_cpus_allowed(p
)))
4383 for_each_cpu(i
, sched_group_cpus(sg
)) {
4384 if (i
== target
|| !idle_cpu(i
))
4388 target
= cpumask_first_and(sched_group_cpus(sg
),
4389 tsk_cpus_allowed(p
));
4393 } while (sg
!= sd
->groups
);
4400 * select_task_rq_fair: Select target runqueue for the waking task in domains
4401 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4402 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4404 * Balances load by selecting the idlest cpu in the idlest group, or under
4405 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4407 * Returns the target cpu number.
4409 * preempt must be disabled.
4412 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4414 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4415 int cpu
= smp_processor_id();
4417 int want_affine
= 0;
4418 int sync
= wake_flags
& WF_SYNC
;
4420 if (p
->nr_cpus_allowed
== 1)
4423 if (sd_flag
& SD_BALANCE_WAKE
) {
4424 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
4430 for_each_domain(cpu
, tmp
) {
4431 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4435 * If both cpu and prev_cpu are part of this domain,
4436 * cpu is a valid SD_WAKE_AFFINE target.
4438 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4439 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4444 if (tmp
->flags
& sd_flag
)
4449 if (cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4452 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4457 struct sched_group
*group
;
4460 if (!(sd
->flags
& sd_flag
)) {
4465 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
4471 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4472 if (new_cpu
== -1 || new_cpu
== cpu
) {
4473 /* Now try balancing at a lower domain level of cpu */
4478 /* Now try balancing at a lower domain level of new_cpu */
4480 weight
= sd
->span_weight
;
4482 for_each_domain(cpu
, tmp
) {
4483 if (weight
<= tmp
->span_weight
)
4485 if (tmp
->flags
& sd_flag
)
4488 /* while loop will break here if sd == NULL */
4497 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4498 * cfs_rq_of(p) references at time of call are still valid and identify the
4499 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4500 * other assumptions, including the state of rq->lock, should be made.
4503 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4505 struct sched_entity
*se
= &p
->se
;
4506 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4509 * Load tracking: accumulate removed load so that it can be processed
4510 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4511 * to blocked load iff they have a positive decay-count. It can never
4512 * be negative here since on-rq tasks have decay-count == 0.
4514 if (se
->avg
.decay_count
) {
4515 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4516 atomic_long_add(se
->avg
.load_avg_contrib
,
4517 &cfs_rq
->removed_load
);
4520 #endif /* CONFIG_SMP */
4522 static unsigned long
4523 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4525 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4528 * Since its curr running now, convert the gran from real-time
4529 * to virtual-time in his units.
4531 * By using 'se' instead of 'curr' we penalize light tasks, so
4532 * they get preempted easier. That is, if 'se' < 'curr' then
4533 * the resulting gran will be larger, therefore penalizing the
4534 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4535 * be smaller, again penalizing the lighter task.
4537 * This is especially important for buddies when the leftmost
4538 * task is higher priority than the buddy.
4540 return calc_delta_fair(gran
, se
);
4544 * Should 'se' preempt 'curr'.
4558 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4560 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4565 gran
= wakeup_gran(curr
, se
);
4572 static void set_last_buddy(struct sched_entity
*se
)
4574 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4577 for_each_sched_entity(se
)
4578 cfs_rq_of(se
)->last
= se
;
4581 static void set_next_buddy(struct sched_entity
*se
)
4583 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4586 for_each_sched_entity(se
)
4587 cfs_rq_of(se
)->next
= se
;
4590 static void set_skip_buddy(struct sched_entity
*se
)
4592 for_each_sched_entity(se
)
4593 cfs_rq_of(se
)->skip
= se
;
4597 * Preempt the current task with a newly woken task if needed:
4599 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4601 struct task_struct
*curr
= rq
->curr
;
4602 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4603 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4604 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4605 int next_buddy_marked
= 0;
4607 if (unlikely(se
== pse
))
4611 * This is possible from callers such as move_task(), in which we
4612 * unconditionally check_prempt_curr() after an enqueue (which may have
4613 * lead to a throttle). This both saves work and prevents false
4614 * next-buddy nomination below.
4616 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4619 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4620 set_next_buddy(pse
);
4621 next_buddy_marked
= 1;
4625 * We can come here with TIF_NEED_RESCHED already set from new task
4628 * Note: this also catches the edge-case of curr being in a throttled
4629 * group (e.g. via set_curr_task), since update_curr() (in the
4630 * enqueue of curr) will have resulted in resched being set. This
4631 * prevents us from potentially nominating it as a false LAST_BUDDY
4634 if (test_tsk_need_resched(curr
))
4637 /* Idle tasks are by definition preempted by non-idle tasks. */
4638 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4639 likely(p
->policy
!= SCHED_IDLE
))
4643 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4644 * is driven by the tick):
4646 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4649 find_matching_se(&se
, &pse
);
4650 update_curr(cfs_rq_of(se
));
4652 if (wakeup_preempt_entity(se
, pse
) == 1) {
4654 * Bias pick_next to pick the sched entity that is
4655 * triggering this preemption.
4657 if (!next_buddy_marked
)
4658 set_next_buddy(pse
);
4667 * Only set the backward buddy when the current task is still
4668 * on the rq. This can happen when a wakeup gets interleaved
4669 * with schedule on the ->pre_schedule() or idle_balance()
4670 * point, either of which can * drop the rq lock.
4672 * Also, during early boot the idle thread is in the fair class,
4673 * for obvious reasons its a bad idea to schedule back to it.
4675 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4678 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
4682 static struct task_struct
*
4683 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4685 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
4686 struct sched_entity
*se
;
4687 struct task_struct
*p
;
4690 #ifdef CONFIG_FAIR_GROUP_SCHED
4691 if (!cfs_rq
->nr_running
)
4694 if (prev
->sched_class
!= &fair_sched_class
)
4698 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4699 * likely that a next task is from the same cgroup as the current.
4701 * Therefore attempt to avoid putting and setting the entire cgroup
4702 * hierarchy, only change the part that actually changes.
4706 struct sched_entity
*curr
= cfs_rq
->curr
;
4709 * Since we got here without doing put_prev_entity() we also
4710 * have to consider cfs_rq->curr. If it is still a runnable
4711 * entity, update_curr() will update its vruntime, otherwise
4712 * forget we've ever seen it.
4714 if (curr
&& curr
->on_rq
)
4715 update_curr(cfs_rq
);
4720 * This call to check_cfs_rq_runtime() will do the throttle and
4721 * dequeue its entity in the parent(s). Therefore the 'simple'
4722 * nr_running test will indeed be correct.
4724 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
4727 se
= pick_next_entity(cfs_rq
, curr
);
4728 cfs_rq
= group_cfs_rq(se
);
4734 * Since we haven't yet done put_prev_entity and if the selected task
4735 * is a different task than we started out with, try and touch the
4736 * least amount of cfs_rqs.
4739 struct sched_entity
*pse
= &prev
->se
;
4741 while (!(cfs_rq
= is_same_group(se
, pse
))) {
4742 int se_depth
= se
->depth
;
4743 int pse_depth
= pse
->depth
;
4745 if (se_depth
<= pse_depth
) {
4746 put_prev_entity(cfs_rq_of(pse
), pse
);
4747 pse
= parent_entity(pse
);
4749 if (se_depth
>= pse_depth
) {
4750 set_next_entity(cfs_rq_of(se
), se
);
4751 se
= parent_entity(se
);
4755 put_prev_entity(cfs_rq
, pse
);
4756 set_next_entity(cfs_rq
, se
);
4759 if (hrtick_enabled(rq
))
4760 hrtick_start_fair(rq
, p
);
4767 if (!cfs_rq
->nr_running
)
4770 put_prev_task(rq
, prev
);
4773 se
= pick_next_entity(cfs_rq
, NULL
);
4774 set_next_entity(cfs_rq
, se
);
4775 cfs_rq
= group_cfs_rq(se
);
4780 if (hrtick_enabled(rq
))
4781 hrtick_start_fair(rq
, p
);
4786 if (idle_balance(rq
)) /* drops rq->lock */
4793 * Account for a descheduled task:
4795 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4797 struct sched_entity
*se
= &prev
->se
;
4798 struct cfs_rq
*cfs_rq
;
4800 for_each_sched_entity(se
) {
4801 cfs_rq
= cfs_rq_of(se
);
4802 put_prev_entity(cfs_rq
, se
);
4807 * sched_yield() is very simple
4809 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4811 static void yield_task_fair(struct rq
*rq
)
4813 struct task_struct
*curr
= rq
->curr
;
4814 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4815 struct sched_entity
*se
= &curr
->se
;
4818 * Are we the only task in the tree?
4820 if (unlikely(rq
->nr_running
== 1))
4823 clear_buddies(cfs_rq
, se
);
4825 if (curr
->policy
!= SCHED_BATCH
) {
4826 update_rq_clock(rq
);
4828 * Update run-time statistics of the 'current'.
4830 update_curr(cfs_rq
);
4832 * Tell update_rq_clock() that we've just updated,
4833 * so we don't do microscopic update in schedule()
4834 * and double the fastpath cost.
4836 rq
->skip_clock_update
= 1;
4842 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
4844 struct sched_entity
*se
= &p
->se
;
4846 /* throttled hierarchies are not runnable */
4847 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
4850 /* Tell the scheduler that we'd really like pse to run next. */
4853 yield_task_fair(rq
);
4859 /**************************************************
4860 * Fair scheduling class load-balancing methods.
4864 * The purpose of load-balancing is to achieve the same basic fairness the
4865 * per-cpu scheduler provides, namely provide a proportional amount of compute
4866 * time to each task. This is expressed in the following equation:
4868 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4870 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4871 * W_i,0 is defined as:
4873 * W_i,0 = \Sum_j w_i,j (2)
4875 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4876 * is derived from the nice value as per prio_to_weight[].
4878 * The weight average is an exponential decay average of the instantaneous
4881 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4883 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4884 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4885 * can also include other factors [XXX].
4887 * To achieve this balance we define a measure of imbalance which follows
4888 * directly from (1):
4890 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4892 * We them move tasks around to minimize the imbalance. In the continuous
4893 * function space it is obvious this converges, in the discrete case we get
4894 * a few fun cases generally called infeasible weight scenarios.
4897 * - infeasible weights;
4898 * - local vs global optima in the discrete case. ]
4903 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4904 * for all i,j solution, we create a tree of cpus that follows the hardware
4905 * topology where each level pairs two lower groups (or better). This results
4906 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4907 * tree to only the first of the previous level and we decrease the frequency
4908 * of load-balance at each level inv. proportional to the number of cpus in
4914 * \Sum { --- * --- * 2^i } = O(n) (5)
4916 * `- size of each group
4917 * | | `- number of cpus doing load-balance
4919 * `- sum over all levels
4921 * Coupled with a limit on how many tasks we can migrate every balance pass,
4922 * this makes (5) the runtime complexity of the balancer.
4924 * An important property here is that each CPU is still (indirectly) connected
4925 * to every other cpu in at most O(log n) steps:
4927 * The adjacency matrix of the resulting graph is given by:
4930 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4933 * And you'll find that:
4935 * A^(log_2 n)_i,j != 0 for all i,j (7)
4937 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4938 * The task movement gives a factor of O(m), giving a convergence complexity
4941 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4946 * In order to avoid CPUs going idle while there's still work to do, new idle
4947 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4948 * tree itself instead of relying on other CPUs to bring it work.
4950 * This adds some complexity to both (5) and (8) but it reduces the total idle
4958 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4961 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4966 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4968 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4970 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4973 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4974 * rewrite all of this once again.]
4977 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
4979 enum fbq_type
{ regular
, remote
, all
};
4981 #define LBF_ALL_PINNED 0x01
4982 #define LBF_NEED_BREAK 0x02
4983 #define LBF_DST_PINNED 0x04
4984 #define LBF_SOME_PINNED 0x08
4987 struct sched_domain
*sd
;
4995 struct cpumask
*dst_grpmask
;
4997 enum cpu_idle_type idle
;
4999 /* The set of CPUs under consideration for load-balancing */
5000 struct cpumask
*cpus
;
5005 unsigned int loop_break
;
5006 unsigned int loop_max
;
5008 enum fbq_type fbq_type
;
5012 * move_task - move a task from one runqueue to another runqueue.
5013 * Both runqueues must be locked.
5015 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
5017 deactivate_task(env
->src_rq
, p
, 0);
5018 set_task_cpu(p
, env
->dst_cpu
);
5019 activate_task(env
->dst_rq
, p
, 0);
5020 check_preempt_curr(env
->dst_rq
, p
, 0);
5024 * Is this task likely cache-hot:
5027 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
5031 if (p
->sched_class
!= &fair_sched_class
)
5034 if (unlikely(p
->policy
== SCHED_IDLE
))
5038 * Buddy candidates are cache hot:
5040 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
5041 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
5042 &p
->se
== cfs_rq_of(&p
->se
)->last
))
5045 if (sysctl_sched_migration_cost
== -1)
5047 if (sysctl_sched_migration_cost
== 0)
5050 delta
= now
- p
->se
.exec_start
;
5052 return delta
< (s64
)sysctl_sched_migration_cost
;
5055 #ifdef CONFIG_NUMA_BALANCING
5056 /* Returns true if the destination node has incurred more faults */
5057 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
5059 int src_nid
, dst_nid
;
5061 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults_memory
||
5062 !(env
->sd
->flags
& SD_NUMA
)) {
5066 src_nid
= cpu_to_node(env
->src_cpu
);
5067 dst_nid
= cpu_to_node(env
->dst_cpu
);
5069 if (src_nid
== dst_nid
)
5072 /* Always encourage migration to the preferred node. */
5073 if (dst_nid
== p
->numa_preferred_nid
)
5076 /* If both task and group weight improve, this move is a winner. */
5077 if (task_weight(p
, dst_nid
) > task_weight(p
, src_nid
) &&
5078 group_weight(p
, dst_nid
) > group_weight(p
, src_nid
))
5085 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
5087 int src_nid
, dst_nid
;
5089 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
5092 if (!p
->numa_faults_memory
|| !(env
->sd
->flags
& SD_NUMA
))
5095 src_nid
= cpu_to_node(env
->src_cpu
);
5096 dst_nid
= cpu_to_node(env
->dst_cpu
);
5098 if (src_nid
== dst_nid
)
5101 /* Migrating away from the preferred node is always bad. */
5102 if (src_nid
== p
->numa_preferred_nid
)
5105 /* If either task or group weight get worse, don't do it. */
5106 if (task_weight(p
, dst_nid
) < task_weight(p
, src_nid
) ||
5107 group_weight(p
, dst_nid
) < group_weight(p
, src_nid
))
5114 static inline bool migrate_improves_locality(struct task_struct
*p
,
5120 static inline bool migrate_degrades_locality(struct task_struct
*p
,
5128 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5131 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
5133 int tsk_cache_hot
= 0;
5135 * We do not migrate tasks that are:
5136 * 1) throttled_lb_pair, or
5137 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5138 * 3) running (obviously), or
5139 * 4) are cache-hot on their current CPU.
5141 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
5144 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
5147 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
5149 env
->flags
|= LBF_SOME_PINNED
;
5152 * Remember if this task can be migrated to any other cpu in
5153 * our sched_group. We may want to revisit it if we couldn't
5154 * meet load balance goals by pulling other tasks on src_cpu.
5156 * Also avoid computing new_dst_cpu if we have already computed
5157 * one in current iteration.
5159 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
5162 /* Prevent to re-select dst_cpu via env's cpus */
5163 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
5164 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
5165 env
->flags
|= LBF_DST_PINNED
;
5166 env
->new_dst_cpu
= cpu
;
5174 /* Record that we found atleast one task that could run on dst_cpu */
5175 env
->flags
&= ~LBF_ALL_PINNED
;
5177 if (task_running(env
->src_rq
, p
)) {
5178 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
5183 * Aggressive migration if:
5184 * 1) destination numa is preferred
5185 * 2) task is cache cold, or
5186 * 3) too many balance attempts have failed.
5188 tsk_cache_hot
= task_hot(p
, rq_clock_task(env
->src_rq
), env
->sd
);
5190 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
5192 if (migrate_improves_locality(p
, env
)) {
5193 #ifdef CONFIG_SCHEDSTATS
5194 if (tsk_cache_hot
) {
5195 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5196 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5202 if (!tsk_cache_hot
||
5203 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
5205 if (tsk_cache_hot
) {
5206 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5207 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5213 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
5218 * move_one_task tries to move exactly one task from busiest to this_rq, as
5219 * part of active balancing operations within "domain".
5220 * Returns 1 if successful and 0 otherwise.
5222 * Called with both runqueues locked.
5224 static int move_one_task(struct lb_env
*env
)
5226 struct task_struct
*p
, *n
;
5228 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
5229 if (!can_migrate_task(p
, env
))
5234 * Right now, this is only the second place move_task()
5235 * is called, so we can safely collect move_task()
5236 * stats here rather than inside move_task().
5238 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
5244 static const unsigned int sched_nr_migrate_break
= 32;
5247 * move_tasks tries to move up to imbalance weighted load from busiest to
5248 * this_rq, as part of a balancing operation within domain "sd".
5249 * Returns 1 if successful and 0 otherwise.
5251 * Called with both runqueues locked.
5253 static int move_tasks(struct lb_env
*env
)
5255 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
5256 struct task_struct
*p
;
5260 if (env
->imbalance
<= 0)
5263 while (!list_empty(tasks
)) {
5264 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5267 /* We've more or less seen every task there is, call it quits */
5268 if (env
->loop
> env
->loop_max
)
5271 /* take a breather every nr_migrate tasks */
5272 if (env
->loop
> env
->loop_break
) {
5273 env
->loop_break
+= sched_nr_migrate_break
;
5274 env
->flags
|= LBF_NEED_BREAK
;
5278 if (!can_migrate_task(p
, env
))
5281 load
= task_h_load(p
);
5283 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5286 if ((load
/ 2) > env
->imbalance
)
5291 env
->imbalance
-= load
;
5293 #ifdef CONFIG_PREEMPT
5295 * NEWIDLE balancing is a source of latency, so preemptible
5296 * kernels will stop after the first task is pulled to minimize
5297 * the critical section.
5299 if (env
->idle
== CPU_NEWLY_IDLE
)
5304 * We only want to steal up to the prescribed amount of
5307 if (env
->imbalance
<= 0)
5312 list_move_tail(&p
->se
.group_node
, tasks
);
5316 * Right now, this is one of only two places move_task() is called,
5317 * so we can safely collect move_task() stats here rather than
5318 * inside move_task().
5320 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
5325 #ifdef CONFIG_FAIR_GROUP_SCHED
5327 * update tg->load_weight by folding this cpu's load_avg
5329 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5331 struct sched_entity
*se
= tg
->se
[cpu
];
5332 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5334 /* throttled entities do not contribute to load */
5335 if (throttled_hierarchy(cfs_rq
))
5338 update_cfs_rq_blocked_load(cfs_rq
, 1);
5341 update_entity_load_avg(se
, 1);
5343 * We pivot on our runnable average having decayed to zero for
5344 * list removal. This generally implies that all our children
5345 * have also been removed (modulo rounding error or bandwidth
5346 * control); however, such cases are rare and we can fix these
5349 * TODO: fix up out-of-order children on enqueue.
5351 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5352 list_del_leaf_cfs_rq(cfs_rq
);
5354 struct rq
*rq
= rq_of(cfs_rq
);
5355 update_rq_runnable_avg(rq
, rq
->nr_running
);
5359 static void update_blocked_averages(int cpu
)
5361 struct rq
*rq
= cpu_rq(cpu
);
5362 struct cfs_rq
*cfs_rq
;
5363 unsigned long flags
;
5365 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5366 update_rq_clock(rq
);
5368 * Iterates the task_group tree in a bottom up fashion, see
5369 * list_add_leaf_cfs_rq() for details.
5371 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5373 * Note: We may want to consider periodically releasing
5374 * rq->lock about these updates so that creating many task
5375 * groups does not result in continually extending hold time.
5377 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5380 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5384 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5385 * This needs to be done in a top-down fashion because the load of a child
5386 * group is a fraction of its parents load.
5388 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5390 struct rq
*rq
= rq_of(cfs_rq
);
5391 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5392 unsigned long now
= jiffies
;
5395 if (cfs_rq
->last_h_load_update
== now
)
5398 cfs_rq
->h_load_next
= NULL
;
5399 for_each_sched_entity(se
) {
5400 cfs_rq
= cfs_rq_of(se
);
5401 cfs_rq
->h_load_next
= se
;
5402 if (cfs_rq
->last_h_load_update
== now
)
5407 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5408 cfs_rq
->last_h_load_update
= now
;
5411 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5412 load
= cfs_rq
->h_load
;
5413 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5414 cfs_rq
->runnable_load_avg
+ 1);
5415 cfs_rq
= group_cfs_rq(se
);
5416 cfs_rq
->h_load
= load
;
5417 cfs_rq
->last_h_load_update
= now
;
5421 static unsigned long task_h_load(struct task_struct
*p
)
5423 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5425 update_cfs_rq_h_load(cfs_rq
);
5426 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5427 cfs_rq
->runnable_load_avg
+ 1);
5430 static inline void update_blocked_averages(int cpu
)
5434 static unsigned long task_h_load(struct task_struct
*p
)
5436 return p
->se
.avg
.load_avg_contrib
;
5440 /********** Helpers for find_busiest_group ************************/
5442 * sg_lb_stats - stats of a sched_group required for load_balancing
5444 struct sg_lb_stats
{
5445 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5446 unsigned long group_load
; /* Total load over the CPUs of the group */
5447 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5448 unsigned long load_per_task
;
5449 unsigned long group_power
;
5450 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5451 unsigned int group_capacity
;
5452 unsigned int idle_cpus
;
5453 unsigned int group_weight
;
5454 int group_imb
; /* Is there an imbalance in the group ? */
5455 int group_has_capacity
; /* Is there extra capacity in the group? */
5456 #ifdef CONFIG_NUMA_BALANCING
5457 unsigned int nr_numa_running
;
5458 unsigned int nr_preferred_running
;
5463 * sd_lb_stats - Structure to store the statistics of a sched_domain
5464 * during load balancing.
5466 struct sd_lb_stats
{
5467 struct sched_group
*busiest
; /* Busiest group in this sd */
5468 struct sched_group
*local
; /* Local group in this sd */
5469 unsigned long total_load
; /* Total load of all groups in sd */
5470 unsigned long total_pwr
; /* Total power of all groups in sd */
5471 unsigned long avg_load
; /* Average load across all groups in sd */
5473 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5474 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5477 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5480 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5481 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5482 * We must however clear busiest_stat::avg_load because
5483 * update_sd_pick_busiest() reads this before assignment.
5485 *sds
= (struct sd_lb_stats
){
5497 * get_sd_load_idx - Obtain the load index for a given sched domain.
5498 * @sd: The sched_domain whose load_idx is to be obtained.
5499 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5501 * Return: The load index.
5503 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5504 enum cpu_idle_type idle
)
5510 load_idx
= sd
->busy_idx
;
5513 case CPU_NEWLY_IDLE
:
5514 load_idx
= sd
->newidle_idx
;
5517 load_idx
= sd
->idle_idx
;
5524 static unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
5526 return SCHED_POWER_SCALE
;
5529 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
5531 return default_scale_freq_power(sd
, cpu
);
5534 static unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
5536 unsigned long weight
= sd
->span_weight
;
5537 unsigned long smt_gain
= sd
->smt_gain
;
5544 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
5546 return default_scale_smt_power(sd
, cpu
);
5549 static unsigned long scale_rt_power(int cpu
)
5551 struct rq
*rq
= cpu_rq(cpu
);
5552 u64 total
, available
, age_stamp
, avg
;
5555 * Since we're reading these variables without serialization make sure
5556 * we read them once before doing sanity checks on them.
5558 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5559 avg
= ACCESS_ONCE(rq
->rt_avg
);
5561 total
= sched_avg_period() + (rq_clock(rq
) - age_stamp
);
5563 if (unlikely(total
< avg
)) {
5564 /* Ensures that power won't end up being negative */
5567 available
= total
- avg
;
5570 if (unlikely((s64
)total
< SCHED_POWER_SCALE
))
5571 total
= SCHED_POWER_SCALE
;
5573 total
>>= SCHED_POWER_SHIFT
;
5575 return div_u64(available
, total
);
5578 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
5580 unsigned long weight
= sd
->span_weight
;
5581 unsigned long power
= SCHED_POWER_SCALE
;
5582 struct sched_group
*sdg
= sd
->groups
;
5584 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
5585 if (sched_feat(ARCH_POWER
))
5586 power
*= arch_scale_smt_power(sd
, cpu
);
5588 power
*= default_scale_smt_power(sd
, cpu
);
5590 power
>>= SCHED_POWER_SHIFT
;
5593 sdg
->sgp
->power_orig
= power
;
5595 if (sched_feat(ARCH_POWER
))
5596 power
*= arch_scale_freq_power(sd
, cpu
);
5598 power
*= default_scale_freq_power(sd
, cpu
);
5600 power
>>= SCHED_POWER_SHIFT
;
5602 power
*= scale_rt_power(cpu
);
5603 power
>>= SCHED_POWER_SHIFT
;
5608 cpu_rq(cpu
)->cpu_power
= power
;
5609 sdg
->sgp
->power
= power
;
5612 void update_group_power(struct sched_domain
*sd
, int cpu
)
5614 struct sched_domain
*child
= sd
->child
;
5615 struct sched_group
*group
, *sdg
= sd
->groups
;
5616 unsigned long power
, power_orig
;
5617 unsigned long interval
;
5619 interval
= msecs_to_jiffies(sd
->balance_interval
);
5620 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5621 sdg
->sgp
->next_update
= jiffies
+ interval
;
5624 update_cpu_power(sd
, cpu
);
5628 power_orig
= power
= 0;
5630 if (child
->flags
& SD_OVERLAP
) {
5632 * SD_OVERLAP domains cannot assume that child groups
5633 * span the current group.
5636 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
5637 struct sched_group_power
*sgp
;
5638 struct rq
*rq
= cpu_rq(cpu
);
5641 * build_sched_domains() -> init_sched_groups_power()
5642 * gets here before we've attached the domains to the
5645 * Use power_of(), which is set irrespective of domains
5646 * in update_cpu_power().
5648 * This avoids power/power_orig from being 0 and
5649 * causing divide-by-zero issues on boot.
5651 * Runtime updates will correct power_orig.
5653 if (unlikely(!rq
->sd
)) {
5654 power_orig
+= power_of(cpu
);
5655 power
+= power_of(cpu
);
5659 sgp
= rq
->sd
->groups
->sgp
;
5660 power_orig
+= sgp
->power_orig
;
5661 power
+= sgp
->power
;
5665 * !SD_OVERLAP domains can assume that child groups
5666 * span the current group.
5669 group
= child
->groups
;
5671 power_orig
+= group
->sgp
->power_orig
;
5672 power
+= group
->sgp
->power
;
5673 group
= group
->next
;
5674 } while (group
!= child
->groups
);
5677 sdg
->sgp
->power_orig
= power_orig
;
5678 sdg
->sgp
->power
= power
;
5682 * Try and fix up capacity for tiny siblings, this is needed when
5683 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5684 * which on its own isn't powerful enough.
5686 * See update_sd_pick_busiest() and check_asym_packing().
5689 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
5692 * Only siblings can have significantly less than SCHED_POWER_SCALE
5694 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
5698 * If ~90% of the cpu_power is still there, we're good.
5700 if (group
->sgp
->power
* 32 > group
->sgp
->power_orig
* 29)
5707 * Group imbalance indicates (and tries to solve) the problem where balancing
5708 * groups is inadequate due to tsk_cpus_allowed() constraints.
5710 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5711 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5714 * { 0 1 2 3 } { 4 5 6 7 }
5717 * If we were to balance group-wise we'd place two tasks in the first group and
5718 * two tasks in the second group. Clearly this is undesired as it will overload
5719 * cpu 3 and leave one of the cpus in the second group unused.
5721 * The current solution to this issue is detecting the skew in the first group
5722 * by noticing the lower domain failed to reach balance and had difficulty
5723 * moving tasks due to affinity constraints.
5725 * When this is so detected; this group becomes a candidate for busiest; see
5726 * update_sd_pick_busiest(). And calculate_imbalance() and
5727 * find_busiest_group() avoid some of the usual balance conditions to allow it
5728 * to create an effective group imbalance.
5730 * This is a somewhat tricky proposition since the next run might not find the
5731 * group imbalance and decide the groups need to be balanced again. A most
5732 * subtle and fragile situation.
5735 static inline int sg_imbalanced(struct sched_group
*group
)
5737 return group
->sgp
->imbalance
;
5741 * Compute the group capacity.
5743 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5744 * first dividing out the smt factor and computing the actual number of cores
5745 * and limit power unit capacity with that.
5747 static inline int sg_capacity(struct lb_env
*env
, struct sched_group
*group
)
5749 unsigned int capacity
, smt
, cpus
;
5750 unsigned int power
, power_orig
;
5752 power
= group
->sgp
->power
;
5753 power_orig
= group
->sgp
->power_orig
;
5754 cpus
= group
->group_weight
;
5756 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5757 smt
= DIV_ROUND_UP(SCHED_POWER_SCALE
* cpus
, power_orig
);
5758 capacity
= cpus
/ smt
; /* cores */
5760 capacity
= min_t(unsigned, capacity
, DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
));
5762 capacity
= fix_small_capacity(env
->sd
, group
);
5768 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5769 * @env: The load balancing environment.
5770 * @group: sched_group whose statistics are to be updated.
5771 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5772 * @local_group: Does group contain this_cpu.
5773 * @sgs: variable to hold the statistics for this group.
5775 static inline void update_sg_lb_stats(struct lb_env
*env
,
5776 struct sched_group
*group
, int load_idx
,
5777 int local_group
, struct sg_lb_stats
*sgs
)
5782 memset(sgs
, 0, sizeof(*sgs
));
5784 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5785 struct rq
*rq
= cpu_rq(i
);
5787 /* Bias balancing toward cpus of our domain */
5789 load
= target_load(i
, load_idx
);
5791 load
= source_load(i
, load_idx
);
5793 sgs
->group_load
+= load
;
5794 sgs
->sum_nr_running
+= rq
->nr_running
;
5795 #ifdef CONFIG_NUMA_BALANCING
5796 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
5797 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
5799 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
5804 /* Adjust by relative CPU power of the group */
5805 sgs
->group_power
= group
->sgp
->power
;
5806 sgs
->avg_load
= (sgs
->group_load
*SCHED_POWER_SCALE
) / sgs
->group_power
;
5808 if (sgs
->sum_nr_running
)
5809 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
5811 sgs
->group_weight
= group
->group_weight
;
5813 sgs
->group_imb
= sg_imbalanced(group
);
5814 sgs
->group_capacity
= sg_capacity(env
, group
);
5816 if (sgs
->group_capacity
> sgs
->sum_nr_running
)
5817 sgs
->group_has_capacity
= 1;
5821 * update_sd_pick_busiest - return 1 on busiest group
5822 * @env: The load balancing environment.
5823 * @sds: sched_domain statistics
5824 * @sg: sched_group candidate to be checked for being the busiest
5825 * @sgs: sched_group statistics
5827 * Determine if @sg is a busier group than the previously selected
5830 * Return: %true if @sg is a busier group than the previously selected
5831 * busiest group. %false otherwise.
5833 static bool update_sd_pick_busiest(struct lb_env
*env
,
5834 struct sd_lb_stats
*sds
,
5835 struct sched_group
*sg
,
5836 struct sg_lb_stats
*sgs
)
5838 if (sgs
->avg_load
<= sds
->busiest_stat
.avg_load
)
5841 if (sgs
->sum_nr_running
> sgs
->group_capacity
)
5848 * ASYM_PACKING needs to move all the work to the lowest
5849 * numbered CPUs in the group, therefore mark all groups
5850 * higher than ourself as busy.
5852 if ((env
->sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
5853 env
->dst_cpu
< group_first_cpu(sg
)) {
5857 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
5864 #ifdef CONFIG_NUMA_BALANCING
5865 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
5867 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
5869 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
5874 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
5876 if (rq
->nr_running
> rq
->nr_numa_running
)
5878 if (rq
->nr_running
> rq
->nr_preferred_running
)
5883 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
5888 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
5892 #endif /* CONFIG_NUMA_BALANCING */
5895 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5896 * @env: The load balancing environment.
5897 * @sds: variable to hold the statistics for this sched_domain.
5899 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5901 struct sched_domain
*child
= env
->sd
->child
;
5902 struct sched_group
*sg
= env
->sd
->groups
;
5903 struct sg_lb_stats tmp_sgs
;
5904 int load_idx
, prefer_sibling
= 0;
5906 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
5909 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
5912 struct sg_lb_stats
*sgs
= &tmp_sgs
;
5915 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
5918 sgs
= &sds
->local_stat
;
5920 if (env
->idle
!= CPU_NEWLY_IDLE
||
5921 time_after_eq(jiffies
, sg
->sgp
->next_update
))
5922 update_group_power(env
->sd
, env
->dst_cpu
);
5925 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
);
5931 * In case the child domain prefers tasks go to siblings
5932 * first, lower the sg capacity to one so that we'll try
5933 * and move all the excess tasks away. We lower the capacity
5934 * of a group only if the local group has the capacity to fit
5935 * these excess tasks, i.e. nr_running < group_capacity. The
5936 * extra check prevents the case where you always pull from the
5937 * heaviest group when it is already under-utilized (possible
5938 * with a large weight task outweighs the tasks on the system).
5940 if (prefer_sibling
&& sds
->local
&&
5941 sds
->local_stat
.group_has_capacity
)
5942 sgs
->group_capacity
= min(sgs
->group_capacity
, 1U);
5944 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
5946 sds
->busiest_stat
= *sgs
;
5950 /* Now, start updating sd_lb_stats */
5951 sds
->total_load
+= sgs
->group_load
;
5952 sds
->total_pwr
+= sgs
->group_power
;
5955 } while (sg
!= env
->sd
->groups
);
5957 if (env
->sd
->flags
& SD_NUMA
)
5958 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
5962 * check_asym_packing - Check to see if the group is packed into the
5965 * This is primarily intended to used at the sibling level. Some
5966 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5967 * case of POWER7, it can move to lower SMT modes only when higher
5968 * threads are idle. When in lower SMT modes, the threads will
5969 * perform better since they share less core resources. Hence when we
5970 * have idle threads, we want them to be the higher ones.
5972 * This packing function is run on idle threads. It checks to see if
5973 * the busiest CPU in this domain (core in the P7 case) has a higher
5974 * CPU number than the packing function is being run on. Here we are
5975 * assuming lower CPU number will be equivalent to lower a SMT thread
5978 * Return: 1 when packing is required and a task should be moved to
5979 * this CPU. The amount of the imbalance is returned in *imbalance.
5981 * @env: The load balancing environment.
5982 * @sds: Statistics of the sched_domain which is to be packed
5984 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5988 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
5994 busiest_cpu
= group_first_cpu(sds
->busiest
);
5995 if (env
->dst_cpu
> busiest_cpu
)
5998 env
->imbalance
= DIV_ROUND_CLOSEST(
5999 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_power
,
6006 * fix_small_imbalance - Calculate the minor imbalance that exists
6007 * amongst the groups of a sched_domain, during
6009 * @env: The load balancing environment.
6010 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6013 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6015 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
6016 unsigned int imbn
= 2;
6017 unsigned long scaled_busy_load_per_task
;
6018 struct sg_lb_stats
*local
, *busiest
;
6020 local
= &sds
->local_stat
;
6021 busiest
= &sds
->busiest_stat
;
6023 if (!local
->sum_nr_running
)
6024 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
6025 else if (busiest
->load_per_task
> local
->load_per_task
)
6028 scaled_busy_load_per_task
=
6029 (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
6030 busiest
->group_power
;
6032 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
6033 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
6034 env
->imbalance
= busiest
->load_per_task
;
6039 * OK, we don't have enough imbalance to justify moving tasks,
6040 * however we may be able to increase total CPU power used by
6044 pwr_now
+= busiest
->group_power
*
6045 min(busiest
->load_per_task
, busiest
->avg_load
);
6046 pwr_now
+= local
->group_power
*
6047 min(local
->load_per_task
, local
->avg_load
);
6048 pwr_now
/= SCHED_POWER_SCALE
;
6050 /* Amount of load we'd subtract */
6051 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
6052 busiest
->group_power
;
6053 if (busiest
->avg_load
> tmp
) {
6054 pwr_move
+= busiest
->group_power
*
6055 min(busiest
->load_per_task
,
6056 busiest
->avg_load
- tmp
);
6059 /* Amount of load we'd add */
6060 if (busiest
->avg_load
* busiest
->group_power
<
6061 busiest
->load_per_task
* SCHED_POWER_SCALE
) {
6062 tmp
= (busiest
->avg_load
* busiest
->group_power
) /
6065 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
6068 pwr_move
+= local
->group_power
*
6069 min(local
->load_per_task
, local
->avg_load
+ tmp
);
6070 pwr_move
/= SCHED_POWER_SCALE
;
6072 /* Move if we gain throughput */
6073 if (pwr_move
> pwr_now
)
6074 env
->imbalance
= busiest
->load_per_task
;
6078 * calculate_imbalance - Calculate the amount of imbalance present within the
6079 * groups of a given sched_domain during load balance.
6080 * @env: load balance environment
6081 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6083 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6085 unsigned long max_pull
, load_above_capacity
= ~0UL;
6086 struct sg_lb_stats
*local
, *busiest
;
6088 local
= &sds
->local_stat
;
6089 busiest
= &sds
->busiest_stat
;
6091 if (busiest
->group_imb
) {
6093 * In the group_imb case we cannot rely on group-wide averages
6094 * to ensure cpu-load equilibrium, look at wider averages. XXX
6096 busiest
->load_per_task
=
6097 min(busiest
->load_per_task
, sds
->avg_load
);
6101 * In the presence of smp nice balancing, certain scenarios can have
6102 * max load less than avg load(as we skip the groups at or below
6103 * its cpu_power, while calculating max_load..)
6105 if (busiest
->avg_load
<= sds
->avg_load
||
6106 local
->avg_load
>= sds
->avg_load
) {
6108 return fix_small_imbalance(env
, sds
);
6111 if (!busiest
->group_imb
) {
6113 * Don't want to pull so many tasks that a group would go idle.
6114 * Except of course for the group_imb case, since then we might
6115 * have to drop below capacity to reach cpu-load equilibrium.
6117 load_above_capacity
=
6118 (busiest
->sum_nr_running
- busiest
->group_capacity
);
6120 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_POWER_SCALE
);
6121 load_above_capacity
/= busiest
->group_power
;
6125 * We're trying to get all the cpus to the average_load, so we don't
6126 * want to push ourselves above the average load, nor do we wish to
6127 * reduce the max loaded cpu below the average load. At the same time,
6128 * we also don't want to reduce the group load below the group capacity
6129 * (so that we can implement power-savings policies etc). Thus we look
6130 * for the minimum possible imbalance.
6132 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
6134 /* How much load to actually move to equalise the imbalance */
6135 env
->imbalance
= min(
6136 max_pull
* busiest
->group_power
,
6137 (sds
->avg_load
- local
->avg_load
) * local
->group_power
6138 ) / SCHED_POWER_SCALE
;
6141 * if *imbalance is less than the average load per runnable task
6142 * there is no guarantee that any tasks will be moved so we'll have
6143 * a think about bumping its value to force at least one task to be
6146 if (env
->imbalance
< busiest
->load_per_task
)
6147 return fix_small_imbalance(env
, sds
);
6150 /******* find_busiest_group() helpers end here *********************/
6153 * find_busiest_group - Returns the busiest group within the sched_domain
6154 * if there is an imbalance. If there isn't an imbalance, and
6155 * the user has opted for power-savings, it returns a group whose
6156 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6157 * such a group exists.
6159 * Also calculates the amount of weighted load which should be moved
6160 * to restore balance.
6162 * @env: The load balancing environment.
6164 * Return: - The busiest group if imbalance exists.
6165 * - If no imbalance and user has opted for power-savings balance,
6166 * return the least loaded group whose CPUs can be
6167 * put to idle by rebalancing its tasks onto our group.
6169 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
6171 struct sg_lb_stats
*local
, *busiest
;
6172 struct sd_lb_stats sds
;
6174 init_sd_lb_stats(&sds
);
6177 * Compute the various statistics relavent for load balancing at
6180 update_sd_lb_stats(env
, &sds
);
6181 local
= &sds
.local_stat
;
6182 busiest
= &sds
.busiest_stat
;
6184 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
6185 check_asym_packing(env
, &sds
))
6188 /* There is no busy sibling group to pull tasks from */
6189 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
6192 sds
.avg_load
= (SCHED_POWER_SCALE
* sds
.total_load
) / sds
.total_pwr
;
6195 * If the busiest group is imbalanced the below checks don't
6196 * work because they assume all things are equal, which typically
6197 * isn't true due to cpus_allowed constraints and the like.
6199 if (busiest
->group_imb
)
6202 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6203 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_capacity
&&
6204 !busiest
->group_has_capacity
)
6208 * If the local group is more busy than the selected busiest group
6209 * don't try and pull any tasks.
6211 if (local
->avg_load
>= busiest
->avg_load
)
6215 * Don't pull any tasks if this group is already above the domain
6218 if (local
->avg_load
>= sds
.avg_load
)
6221 if (env
->idle
== CPU_IDLE
) {
6223 * This cpu is idle. If the busiest group load doesn't
6224 * have more tasks than the number of available cpu's and
6225 * there is no imbalance between this and busiest group
6226 * wrt to idle cpu's, it is balanced.
6228 if ((local
->idle_cpus
< busiest
->idle_cpus
) &&
6229 busiest
->sum_nr_running
<= busiest
->group_weight
)
6233 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6234 * imbalance_pct to be conservative.
6236 if (100 * busiest
->avg_load
<=
6237 env
->sd
->imbalance_pct
* local
->avg_load
)
6242 /* Looks like there is an imbalance. Compute it */
6243 calculate_imbalance(env
, &sds
);
6252 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6254 static struct rq
*find_busiest_queue(struct lb_env
*env
,
6255 struct sched_group
*group
)
6257 struct rq
*busiest
= NULL
, *rq
;
6258 unsigned long busiest_load
= 0, busiest_power
= 1;
6261 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6262 unsigned long power
, capacity
, wl
;
6266 rt
= fbq_classify_rq(rq
);
6269 * We classify groups/runqueues into three groups:
6270 * - regular: there are !numa tasks
6271 * - remote: there are numa tasks that run on the 'wrong' node
6272 * - all: there is no distinction
6274 * In order to avoid migrating ideally placed numa tasks,
6275 * ignore those when there's better options.
6277 * If we ignore the actual busiest queue to migrate another
6278 * task, the next balance pass can still reduce the busiest
6279 * queue by moving tasks around inside the node.
6281 * If we cannot move enough load due to this classification
6282 * the next pass will adjust the group classification and
6283 * allow migration of more tasks.
6285 * Both cases only affect the total convergence complexity.
6287 if (rt
> env
->fbq_type
)
6290 power
= power_of(i
);
6291 capacity
= DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
);
6293 capacity
= fix_small_capacity(env
->sd
, group
);
6295 wl
= weighted_cpuload(i
);
6298 * When comparing with imbalance, use weighted_cpuload()
6299 * which is not scaled with the cpu power.
6301 if (capacity
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6305 * For the load comparisons with the other cpu's, consider
6306 * the weighted_cpuload() scaled with the cpu power, so that
6307 * the load can be moved away from the cpu that is potentially
6308 * running at a lower capacity.
6310 * Thus we're looking for max(wl_i / power_i), crosswise
6311 * multiplication to rid ourselves of the division works out
6312 * to: wl_i * power_j > wl_j * power_i; where j is our
6315 if (wl
* busiest_power
> busiest_load
* power
) {
6317 busiest_power
= power
;
6326 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6327 * so long as it is large enough.
6329 #define MAX_PINNED_INTERVAL 512
6331 /* Working cpumask for load_balance and load_balance_newidle. */
6332 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6334 static int need_active_balance(struct lb_env
*env
)
6336 struct sched_domain
*sd
= env
->sd
;
6338 if (env
->idle
== CPU_NEWLY_IDLE
) {
6341 * ASYM_PACKING needs to force migrate tasks from busy but
6342 * higher numbered CPUs in order to pack all tasks in the
6343 * lowest numbered CPUs.
6345 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6349 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6352 static int active_load_balance_cpu_stop(void *data
);
6354 static int should_we_balance(struct lb_env
*env
)
6356 struct sched_group
*sg
= env
->sd
->groups
;
6357 struct cpumask
*sg_cpus
, *sg_mask
;
6358 int cpu
, balance_cpu
= -1;
6361 * In the newly idle case, we will allow all the cpu's
6362 * to do the newly idle load balance.
6364 if (env
->idle
== CPU_NEWLY_IDLE
)
6367 sg_cpus
= sched_group_cpus(sg
);
6368 sg_mask
= sched_group_mask(sg
);
6369 /* Try to find first idle cpu */
6370 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6371 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6378 if (balance_cpu
== -1)
6379 balance_cpu
= group_balance_cpu(sg
);
6382 * First idle cpu or the first cpu(busiest) in this sched group
6383 * is eligible for doing load balancing at this and above domains.
6385 return balance_cpu
== env
->dst_cpu
;
6389 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6390 * tasks if there is an imbalance.
6392 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6393 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6394 int *continue_balancing
)
6396 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6397 struct sched_domain
*sd_parent
= sd
->parent
;
6398 struct sched_group
*group
;
6400 unsigned long flags
;
6401 struct cpumask
*cpus
= __get_cpu_var(load_balance_mask
);
6403 struct lb_env env
= {
6405 .dst_cpu
= this_cpu
,
6407 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6409 .loop_break
= sched_nr_migrate_break
,
6415 * For NEWLY_IDLE load_balancing, we don't need to consider
6416 * other cpus in our group
6418 if (idle
== CPU_NEWLY_IDLE
)
6419 env
.dst_grpmask
= NULL
;
6421 cpumask_copy(cpus
, cpu_active_mask
);
6423 schedstat_inc(sd
, lb_count
[idle
]);
6426 if (!should_we_balance(&env
)) {
6427 *continue_balancing
= 0;
6431 group
= find_busiest_group(&env
);
6433 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6437 busiest
= find_busiest_queue(&env
, group
);
6439 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6443 BUG_ON(busiest
== env
.dst_rq
);
6445 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6448 if (busiest
->nr_running
> 1) {
6450 * Attempt to move tasks. If find_busiest_group has found
6451 * an imbalance but busiest->nr_running <= 1, the group is
6452 * still unbalanced. ld_moved simply stays zero, so it is
6453 * correctly treated as an imbalance.
6455 env
.flags
|= LBF_ALL_PINNED
;
6456 env
.src_cpu
= busiest
->cpu
;
6457 env
.src_rq
= busiest
;
6458 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6461 local_irq_save(flags
);
6462 double_rq_lock(env
.dst_rq
, busiest
);
6465 * cur_ld_moved - load moved in current iteration
6466 * ld_moved - cumulative load moved across iterations
6468 cur_ld_moved
= move_tasks(&env
);
6469 ld_moved
+= cur_ld_moved
;
6470 double_rq_unlock(env
.dst_rq
, busiest
);
6471 local_irq_restore(flags
);
6474 * some other cpu did the load balance for us.
6476 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
6477 resched_cpu(env
.dst_cpu
);
6479 if (env
.flags
& LBF_NEED_BREAK
) {
6480 env
.flags
&= ~LBF_NEED_BREAK
;
6485 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6486 * us and move them to an alternate dst_cpu in our sched_group
6487 * where they can run. The upper limit on how many times we
6488 * iterate on same src_cpu is dependent on number of cpus in our
6491 * This changes load balance semantics a bit on who can move
6492 * load to a given_cpu. In addition to the given_cpu itself
6493 * (or a ilb_cpu acting on its behalf where given_cpu is
6494 * nohz-idle), we now have balance_cpu in a position to move
6495 * load to given_cpu. In rare situations, this may cause
6496 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6497 * _independently_ and at _same_ time to move some load to
6498 * given_cpu) causing exceess load to be moved to given_cpu.
6499 * This however should not happen so much in practice and
6500 * moreover subsequent load balance cycles should correct the
6501 * excess load moved.
6503 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6505 /* Prevent to re-select dst_cpu via env's cpus */
6506 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6508 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6509 env
.dst_cpu
= env
.new_dst_cpu
;
6510 env
.flags
&= ~LBF_DST_PINNED
;
6512 env
.loop_break
= sched_nr_migrate_break
;
6515 * Go back to "more_balance" rather than "redo" since we
6516 * need to continue with same src_cpu.
6522 * We failed to reach balance because of affinity.
6525 int *group_imbalance
= &sd_parent
->groups
->sgp
->imbalance
;
6527 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0) {
6528 *group_imbalance
= 1;
6529 } else if (*group_imbalance
)
6530 *group_imbalance
= 0;
6533 /* All tasks on this runqueue were pinned by CPU affinity */
6534 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6535 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6536 if (!cpumask_empty(cpus
)) {
6538 env
.loop_break
= sched_nr_migrate_break
;
6546 schedstat_inc(sd
, lb_failed
[idle
]);
6548 * Increment the failure counter only on periodic balance.
6549 * We do not want newidle balance, which can be very
6550 * frequent, pollute the failure counter causing
6551 * excessive cache_hot migrations and active balances.
6553 if (idle
!= CPU_NEWLY_IDLE
)
6554 sd
->nr_balance_failed
++;
6556 if (need_active_balance(&env
)) {
6557 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6559 /* don't kick the active_load_balance_cpu_stop,
6560 * if the curr task on busiest cpu can't be
6563 if (!cpumask_test_cpu(this_cpu
,
6564 tsk_cpus_allowed(busiest
->curr
))) {
6565 raw_spin_unlock_irqrestore(&busiest
->lock
,
6567 env
.flags
|= LBF_ALL_PINNED
;
6568 goto out_one_pinned
;
6572 * ->active_balance synchronizes accesses to
6573 * ->active_balance_work. Once set, it's cleared
6574 * only after active load balance is finished.
6576 if (!busiest
->active_balance
) {
6577 busiest
->active_balance
= 1;
6578 busiest
->push_cpu
= this_cpu
;
6581 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
6583 if (active_balance
) {
6584 stop_one_cpu_nowait(cpu_of(busiest
),
6585 active_load_balance_cpu_stop
, busiest
,
6586 &busiest
->active_balance_work
);
6590 * We've kicked active balancing, reset the failure
6593 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
6596 sd
->nr_balance_failed
= 0;
6598 if (likely(!active_balance
)) {
6599 /* We were unbalanced, so reset the balancing interval */
6600 sd
->balance_interval
= sd
->min_interval
;
6603 * If we've begun active balancing, start to back off. This
6604 * case may not be covered by the all_pinned logic if there
6605 * is only 1 task on the busy runqueue (because we don't call
6608 if (sd
->balance_interval
< sd
->max_interval
)
6609 sd
->balance_interval
*= 2;
6615 schedstat_inc(sd
, lb_balanced
[idle
]);
6617 sd
->nr_balance_failed
= 0;
6620 /* tune up the balancing interval */
6621 if (((env
.flags
& LBF_ALL_PINNED
) &&
6622 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
6623 (sd
->balance_interval
< sd
->max_interval
))
6624 sd
->balance_interval
*= 2;
6632 * idle_balance is called by schedule() if this_cpu is about to become
6633 * idle. Attempts to pull tasks from other CPUs.
6635 static int idle_balance(struct rq
*this_rq
)
6637 struct sched_domain
*sd
;
6638 int pulled_task
= 0;
6639 unsigned long next_balance
= jiffies
+ HZ
;
6641 int this_cpu
= this_rq
->cpu
;
6643 idle_enter_fair(this_rq
);
6645 * We must set idle_stamp _before_ calling idle_balance(), such that we
6646 * measure the duration of idle_balance() as idle time.
6648 this_rq
->idle_stamp
= rq_clock(this_rq
);
6650 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
6654 * Drop the rq->lock, but keep IRQ/preempt disabled.
6656 raw_spin_unlock(&this_rq
->lock
);
6658 update_blocked_averages(this_cpu
);
6660 for_each_domain(this_cpu
, sd
) {
6661 unsigned long interval
;
6662 int continue_balancing
= 1;
6663 u64 t0
, domain_cost
;
6665 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6668 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
)
6671 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
6672 t0
= sched_clock_cpu(this_cpu
);
6674 /* If we've pulled tasks over stop searching: */
6675 pulled_task
= load_balance(this_cpu
, this_rq
,
6677 &continue_balancing
);
6679 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
6680 if (domain_cost
> sd
->max_newidle_lb_cost
)
6681 sd
->max_newidle_lb_cost
= domain_cost
;
6683 curr_cost
+= domain_cost
;
6686 interval
= msecs_to_jiffies(sd
->balance_interval
);
6687 if (time_after(next_balance
, sd
->last_balance
+ interval
))
6688 next_balance
= sd
->last_balance
+ interval
;
6694 raw_spin_lock(&this_rq
->lock
);
6697 * While browsing the domains, we released the rq lock.
6698 * A task could have be enqueued in the meantime
6700 if (this_rq
->nr_running
&& !pulled_task
) {
6705 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
6707 * We are going idle. next_balance may be set based on
6708 * a busy processor. So reset next_balance.
6710 this_rq
->next_balance
= next_balance
;
6713 if (curr_cost
> this_rq
->max_idle_balance_cost
)
6714 this_rq
->max_idle_balance_cost
= curr_cost
;
6718 this_rq
->idle_stamp
= 0;
6724 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6725 * running tasks off the busiest CPU onto idle CPUs. It requires at
6726 * least 1 task to be running on each physical CPU where possible, and
6727 * avoids physical / logical imbalances.
6729 static int active_load_balance_cpu_stop(void *data
)
6731 struct rq
*busiest_rq
= data
;
6732 int busiest_cpu
= cpu_of(busiest_rq
);
6733 int target_cpu
= busiest_rq
->push_cpu
;
6734 struct rq
*target_rq
= cpu_rq(target_cpu
);
6735 struct sched_domain
*sd
;
6737 raw_spin_lock_irq(&busiest_rq
->lock
);
6739 /* make sure the requested cpu hasn't gone down in the meantime */
6740 if (unlikely(busiest_cpu
!= smp_processor_id() ||
6741 !busiest_rq
->active_balance
))
6744 /* Is there any task to move? */
6745 if (busiest_rq
->nr_running
<= 1)
6749 * This condition is "impossible", if it occurs
6750 * we need to fix it. Originally reported by
6751 * Bjorn Helgaas on a 128-cpu setup.
6753 BUG_ON(busiest_rq
== target_rq
);
6755 /* move a task from busiest_rq to target_rq */
6756 double_lock_balance(busiest_rq
, target_rq
);
6758 /* Search for an sd spanning us and the target CPU. */
6760 for_each_domain(target_cpu
, sd
) {
6761 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
6762 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
6767 struct lb_env env
= {
6769 .dst_cpu
= target_cpu
,
6770 .dst_rq
= target_rq
,
6771 .src_cpu
= busiest_rq
->cpu
,
6772 .src_rq
= busiest_rq
,
6776 schedstat_inc(sd
, alb_count
);
6778 if (move_one_task(&env
))
6779 schedstat_inc(sd
, alb_pushed
);
6781 schedstat_inc(sd
, alb_failed
);
6784 double_unlock_balance(busiest_rq
, target_rq
);
6786 busiest_rq
->active_balance
= 0;
6787 raw_spin_unlock_irq(&busiest_rq
->lock
);
6791 static inline int on_null_domain(struct rq
*rq
)
6793 return unlikely(!rcu_dereference_sched(rq
->sd
));
6796 #ifdef CONFIG_NO_HZ_COMMON
6798 * idle load balancing details
6799 * - When one of the busy CPUs notice that there may be an idle rebalancing
6800 * needed, they will kick the idle load balancer, which then does idle
6801 * load balancing for all the idle CPUs.
6804 cpumask_var_t idle_cpus_mask
;
6806 unsigned long next_balance
; /* in jiffy units */
6807 } nohz ____cacheline_aligned
;
6809 static inline int find_new_ilb(void)
6811 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
6813 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
6820 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6821 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6822 * CPU (if there is one).
6824 static void nohz_balancer_kick(void)
6828 nohz
.next_balance
++;
6830 ilb_cpu
= find_new_ilb();
6832 if (ilb_cpu
>= nr_cpu_ids
)
6835 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
6838 * Use smp_send_reschedule() instead of resched_cpu().
6839 * This way we generate a sched IPI on the target cpu which
6840 * is idle. And the softirq performing nohz idle load balance
6841 * will be run before returning from the IPI.
6843 smp_send_reschedule(ilb_cpu
);
6847 static inline void nohz_balance_exit_idle(int cpu
)
6849 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
6851 * Completely isolated CPUs don't ever set, so we must test.
6853 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
6854 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
6855 atomic_dec(&nohz
.nr_cpus
);
6857 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
6861 static inline void set_cpu_sd_state_busy(void)
6863 struct sched_domain
*sd
;
6864 int cpu
= smp_processor_id();
6867 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
6869 if (!sd
|| !sd
->nohz_idle
)
6873 atomic_inc(&sd
->groups
->sgp
->nr_busy_cpus
);
6878 void set_cpu_sd_state_idle(void)
6880 struct sched_domain
*sd
;
6881 int cpu
= smp_processor_id();
6884 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
6886 if (!sd
|| sd
->nohz_idle
)
6890 atomic_dec(&sd
->groups
->sgp
->nr_busy_cpus
);
6896 * This routine will record that the cpu is going idle with tick stopped.
6897 * This info will be used in performing idle load balancing in the future.
6899 void nohz_balance_enter_idle(int cpu
)
6902 * If this cpu is going down, then nothing needs to be done.
6904 if (!cpu_active(cpu
))
6907 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
6911 * If we're a completely isolated CPU, we don't play.
6913 if (on_null_domain(cpu_rq(cpu
)))
6916 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
6917 atomic_inc(&nohz
.nr_cpus
);
6918 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
6921 static int sched_ilb_notifier(struct notifier_block
*nfb
,
6922 unsigned long action
, void *hcpu
)
6924 switch (action
& ~CPU_TASKS_FROZEN
) {
6926 nohz_balance_exit_idle(smp_processor_id());
6934 static DEFINE_SPINLOCK(balancing
);
6937 * Scale the max load_balance interval with the number of CPUs in the system.
6938 * This trades load-balance latency on larger machines for less cross talk.
6940 void update_max_interval(void)
6942 max_load_balance_interval
= HZ
*num_online_cpus()/10;
6946 * It checks each scheduling domain to see if it is due to be balanced,
6947 * and initiates a balancing operation if so.
6949 * Balancing parameters are set up in init_sched_domains.
6951 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
6953 int continue_balancing
= 1;
6955 unsigned long interval
;
6956 struct sched_domain
*sd
;
6957 /* Earliest time when we have to do rebalance again */
6958 unsigned long next_balance
= jiffies
+ 60*HZ
;
6959 int update_next_balance
= 0;
6960 int need_serialize
, need_decay
= 0;
6963 update_blocked_averages(cpu
);
6966 for_each_domain(cpu
, sd
) {
6968 * Decay the newidle max times here because this is a regular
6969 * visit to all the domains. Decay ~1% per second.
6971 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
6972 sd
->max_newidle_lb_cost
=
6973 (sd
->max_newidle_lb_cost
* 253) / 256;
6974 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
6977 max_cost
+= sd
->max_newidle_lb_cost
;
6979 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6983 * Stop the load balance at this level. There is another
6984 * CPU in our sched group which is doing load balancing more
6987 if (!continue_balancing
) {
6993 interval
= sd
->balance_interval
;
6994 if (idle
!= CPU_IDLE
)
6995 interval
*= sd
->busy_factor
;
6997 /* scale ms to jiffies */
6998 interval
= msecs_to_jiffies(interval
);
6999 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
7001 need_serialize
= sd
->flags
& SD_SERIALIZE
;
7003 if (need_serialize
) {
7004 if (!spin_trylock(&balancing
))
7008 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
7009 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
7011 * The LBF_DST_PINNED logic could have changed
7012 * env->dst_cpu, so we can't know our idle
7013 * state even if we migrated tasks. Update it.
7015 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
7017 sd
->last_balance
= jiffies
;
7020 spin_unlock(&balancing
);
7022 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
7023 next_balance
= sd
->last_balance
+ interval
;
7024 update_next_balance
= 1;
7029 * Ensure the rq-wide value also decays but keep it at a
7030 * reasonable floor to avoid funnies with rq->avg_idle.
7032 rq
->max_idle_balance_cost
=
7033 max((u64
)sysctl_sched_migration_cost
, max_cost
);
7038 * next_balance will be updated only when there is a need.
7039 * When the cpu is attached to null domain for ex, it will not be
7042 if (likely(update_next_balance
))
7043 rq
->next_balance
= next_balance
;
7046 #ifdef CONFIG_NO_HZ_COMMON
7048 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7049 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7051 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
7053 int this_cpu
= this_rq
->cpu
;
7057 if (idle
!= CPU_IDLE
||
7058 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
7061 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
7062 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
7066 * If this cpu gets work to do, stop the load balancing
7067 * work being done for other cpus. Next load
7068 * balancing owner will pick it up.
7073 rq
= cpu_rq(balance_cpu
);
7075 raw_spin_lock_irq(&rq
->lock
);
7076 update_rq_clock(rq
);
7077 update_idle_cpu_load(rq
);
7078 raw_spin_unlock_irq(&rq
->lock
);
7080 rebalance_domains(rq
, CPU_IDLE
);
7082 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
7083 this_rq
->next_balance
= rq
->next_balance
;
7085 nohz
.next_balance
= this_rq
->next_balance
;
7087 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
7091 * Current heuristic for kicking the idle load balancer in the presence
7092 * of an idle cpu is the system.
7093 * - This rq has more than one task.
7094 * - At any scheduler domain level, this cpu's scheduler group has multiple
7095 * busy cpu's exceeding the group's power.
7096 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7097 * domain span are idle.
7099 static inline int nohz_kick_needed(struct rq
*rq
)
7101 unsigned long now
= jiffies
;
7102 struct sched_domain
*sd
;
7103 struct sched_group_power
*sgp
;
7104 int nr_busy
, cpu
= rq
->cpu
;
7106 if (unlikely(rq
->idle_balance
))
7110 * We may be recently in ticked or tickless idle mode. At the first
7111 * busy tick after returning from idle, we will update the busy stats.
7113 set_cpu_sd_state_busy();
7114 nohz_balance_exit_idle(cpu
);
7117 * None are in tickless mode and hence no need for NOHZ idle load
7120 if (likely(!atomic_read(&nohz
.nr_cpus
)))
7123 if (time_before(now
, nohz
.next_balance
))
7126 if (rq
->nr_running
>= 2)
7130 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7133 sgp
= sd
->groups
->sgp
;
7134 nr_busy
= atomic_read(&sgp
->nr_busy_cpus
);
7137 goto need_kick_unlock
;
7140 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
7142 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
7143 sched_domain_span(sd
)) < cpu
))
7144 goto need_kick_unlock
;
7155 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
7159 * run_rebalance_domains is triggered when needed from the scheduler tick.
7160 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7162 static void run_rebalance_domains(struct softirq_action
*h
)
7164 struct rq
*this_rq
= this_rq();
7165 enum cpu_idle_type idle
= this_rq
->idle_balance
?
7166 CPU_IDLE
: CPU_NOT_IDLE
;
7168 rebalance_domains(this_rq
, idle
);
7171 * If this cpu has a pending nohz_balance_kick, then do the
7172 * balancing on behalf of the other idle cpus whose ticks are
7175 nohz_idle_balance(this_rq
, idle
);
7179 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7181 void trigger_load_balance(struct rq
*rq
)
7183 /* Don't need to rebalance while attached to NULL domain */
7184 if (unlikely(on_null_domain(rq
)))
7187 if (time_after_eq(jiffies
, rq
->next_balance
))
7188 raise_softirq(SCHED_SOFTIRQ
);
7189 #ifdef CONFIG_NO_HZ_COMMON
7190 if (nohz_kick_needed(rq
))
7191 nohz_balancer_kick();
7195 static void rq_online_fair(struct rq
*rq
)
7200 static void rq_offline_fair(struct rq
*rq
)
7204 /* Ensure any throttled groups are reachable by pick_next_task */
7205 unthrottle_offline_cfs_rqs(rq
);
7208 #endif /* CONFIG_SMP */
7211 * scheduler tick hitting a task of our scheduling class:
7213 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
7215 struct cfs_rq
*cfs_rq
;
7216 struct sched_entity
*se
= &curr
->se
;
7218 for_each_sched_entity(se
) {
7219 cfs_rq
= cfs_rq_of(se
);
7220 entity_tick(cfs_rq
, se
, queued
);
7223 if (numabalancing_enabled
)
7224 task_tick_numa(rq
, curr
);
7226 update_rq_runnable_avg(rq
, 1);
7230 * called on fork with the child task as argument from the parent's context
7231 * - child not yet on the tasklist
7232 * - preemption disabled
7234 static void task_fork_fair(struct task_struct
*p
)
7236 struct cfs_rq
*cfs_rq
;
7237 struct sched_entity
*se
= &p
->se
, *curr
;
7238 int this_cpu
= smp_processor_id();
7239 struct rq
*rq
= this_rq();
7240 unsigned long flags
;
7242 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7244 update_rq_clock(rq
);
7246 cfs_rq
= task_cfs_rq(current
);
7247 curr
= cfs_rq
->curr
;
7250 * Not only the cpu but also the task_group of the parent might have
7251 * been changed after parent->se.parent,cfs_rq were copied to
7252 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7253 * of child point to valid ones.
7256 __set_task_cpu(p
, this_cpu
);
7259 update_curr(cfs_rq
);
7262 se
->vruntime
= curr
->vruntime
;
7263 place_entity(cfs_rq
, se
, 1);
7265 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
7267 * Upon rescheduling, sched_class::put_prev_task() will place
7268 * 'current' within the tree based on its new key value.
7270 swap(curr
->vruntime
, se
->vruntime
);
7271 resched_task(rq
->curr
);
7274 se
->vruntime
-= cfs_rq
->min_vruntime
;
7276 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7280 * Priority of the task has changed. Check to see if we preempt
7284 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
7290 * Reschedule if we are currently running on this runqueue and
7291 * our priority decreased, or if we are not currently running on
7292 * this runqueue and our priority is higher than the current's
7294 if (rq
->curr
== p
) {
7295 if (p
->prio
> oldprio
)
7296 resched_task(rq
->curr
);
7298 check_preempt_curr(rq
, p
, 0);
7301 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
7303 struct sched_entity
*se
= &p
->se
;
7304 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7307 * Ensure the task's vruntime is normalized, so that when its
7308 * switched back to the fair class the enqueue_entity(.flags=0) will
7309 * do the right thing.
7311 * If it was on_rq, then the dequeue_entity(.flags=0) will already
7312 * have normalized the vruntime, if it was !on_rq, then only when
7313 * the task is sleeping will it still have non-normalized vruntime.
7315 if (!se
->on_rq
&& p
->state
!= TASK_RUNNING
) {
7317 * Fix up our vruntime so that the current sleep doesn't
7318 * cause 'unlimited' sleep bonus.
7320 place_entity(cfs_rq
, se
, 0);
7321 se
->vruntime
-= cfs_rq
->min_vruntime
;
7326 * Remove our load from contribution when we leave sched_fair
7327 * and ensure we don't carry in an old decay_count if we
7330 if (se
->avg
.decay_count
) {
7331 __synchronize_entity_decay(se
);
7332 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7338 * We switched to the sched_fair class.
7340 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7342 struct sched_entity
*se
= &p
->se
;
7343 #ifdef CONFIG_FAIR_GROUP_SCHED
7345 * Since the real-depth could have been changed (only FAIR
7346 * class maintain depth value), reset depth properly.
7348 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7354 * We were most likely switched from sched_rt, so
7355 * kick off the schedule if running, otherwise just see
7356 * if we can still preempt the current task.
7359 resched_task(rq
->curr
);
7361 check_preempt_curr(rq
, p
, 0);
7364 /* Account for a task changing its policy or group.
7366 * This routine is mostly called to set cfs_rq->curr field when a task
7367 * migrates between groups/classes.
7369 static void set_curr_task_fair(struct rq
*rq
)
7371 struct sched_entity
*se
= &rq
->curr
->se
;
7373 for_each_sched_entity(se
) {
7374 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7376 set_next_entity(cfs_rq
, se
);
7377 /* ensure bandwidth has been allocated on our new cfs_rq */
7378 account_cfs_rq_runtime(cfs_rq
, 0);
7382 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7384 cfs_rq
->tasks_timeline
= RB_ROOT
;
7385 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7386 #ifndef CONFIG_64BIT
7387 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7390 atomic64_set(&cfs_rq
->decay_counter
, 1);
7391 atomic_long_set(&cfs_rq
->removed_load
, 0);
7395 #ifdef CONFIG_FAIR_GROUP_SCHED
7396 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
7398 struct sched_entity
*se
= &p
->se
;
7399 struct cfs_rq
*cfs_rq
;
7402 * If the task was not on the rq at the time of this cgroup movement
7403 * it must have been asleep, sleeping tasks keep their ->vruntime
7404 * absolute on their old rq until wakeup (needed for the fair sleeper
7405 * bonus in place_entity()).
7407 * If it was on the rq, we've just 'preempted' it, which does convert
7408 * ->vruntime to a relative base.
7410 * Make sure both cases convert their relative position when migrating
7411 * to another cgroup's rq. This does somewhat interfere with the
7412 * fair sleeper stuff for the first placement, but who cares.
7415 * When !on_rq, vruntime of the task has usually NOT been normalized.
7416 * But there are some cases where it has already been normalized:
7418 * - Moving a forked child which is waiting for being woken up by
7419 * wake_up_new_task().
7420 * - Moving a task which has been woken up by try_to_wake_up() and
7421 * waiting for actually being woken up by sched_ttwu_pending().
7423 * To prevent boost or penalty in the new cfs_rq caused by delta
7424 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7426 if (!on_rq
&& (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7430 se
->vruntime
-= cfs_rq_of(se
)->min_vruntime
;
7431 set_task_rq(p
, task_cpu(p
));
7432 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7434 cfs_rq
= cfs_rq_of(se
);
7435 se
->vruntime
+= cfs_rq
->min_vruntime
;
7438 * migrate_task_rq_fair() will have removed our previous
7439 * contribution, but we must synchronize for ongoing future
7442 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7443 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
7448 void free_fair_sched_group(struct task_group
*tg
)
7452 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7454 for_each_possible_cpu(i
) {
7456 kfree(tg
->cfs_rq
[i
]);
7465 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7467 struct cfs_rq
*cfs_rq
;
7468 struct sched_entity
*se
;
7471 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7474 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7478 tg
->shares
= NICE_0_LOAD
;
7480 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7482 for_each_possible_cpu(i
) {
7483 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7484 GFP_KERNEL
, cpu_to_node(i
));
7488 se
= kzalloc_node(sizeof(struct sched_entity
),
7489 GFP_KERNEL
, cpu_to_node(i
));
7493 init_cfs_rq(cfs_rq
);
7494 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
7505 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7507 struct rq
*rq
= cpu_rq(cpu
);
7508 unsigned long flags
;
7511 * Only empty task groups can be destroyed; so we can speculatively
7512 * check on_list without danger of it being re-added.
7514 if (!tg
->cfs_rq
[cpu
]->on_list
)
7517 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7518 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
7519 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7522 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7523 struct sched_entity
*se
, int cpu
,
7524 struct sched_entity
*parent
)
7526 struct rq
*rq
= cpu_rq(cpu
);
7530 init_cfs_rq_runtime(cfs_rq
);
7532 tg
->cfs_rq
[cpu
] = cfs_rq
;
7535 /* se could be NULL for root_task_group */
7540 se
->cfs_rq
= &rq
->cfs
;
7543 se
->cfs_rq
= parent
->my_q
;
7544 se
->depth
= parent
->depth
+ 1;
7548 /* guarantee group entities always have weight */
7549 update_load_set(&se
->load
, NICE_0_LOAD
);
7550 se
->parent
= parent
;
7553 static DEFINE_MUTEX(shares_mutex
);
7555 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7558 unsigned long flags
;
7561 * We can't change the weight of the root cgroup.
7566 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
7568 mutex_lock(&shares_mutex
);
7569 if (tg
->shares
== shares
)
7572 tg
->shares
= shares
;
7573 for_each_possible_cpu(i
) {
7574 struct rq
*rq
= cpu_rq(i
);
7575 struct sched_entity
*se
;
7578 /* Propagate contribution to hierarchy */
7579 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7581 /* Possible calls to update_curr() need rq clock */
7582 update_rq_clock(rq
);
7583 for_each_sched_entity(se
)
7584 update_cfs_shares(group_cfs_rq(se
));
7585 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7589 mutex_unlock(&shares_mutex
);
7592 #else /* CONFIG_FAIR_GROUP_SCHED */
7594 void free_fair_sched_group(struct task_group
*tg
) { }
7596 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7601 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
7603 #endif /* CONFIG_FAIR_GROUP_SCHED */
7606 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
7608 struct sched_entity
*se
= &task
->se
;
7609 unsigned int rr_interval
= 0;
7612 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7615 if (rq
->cfs
.load
.weight
)
7616 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
7622 * All the scheduling class methods:
7624 const struct sched_class fair_sched_class
= {
7625 .next
= &idle_sched_class
,
7626 .enqueue_task
= enqueue_task_fair
,
7627 .dequeue_task
= dequeue_task_fair
,
7628 .yield_task
= yield_task_fair
,
7629 .yield_to_task
= yield_to_task_fair
,
7631 .check_preempt_curr
= check_preempt_wakeup
,
7633 .pick_next_task
= pick_next_task_fair
,
7634 .put_prev_task
= put_prev_task_fair
,
7637 .select_task_rq
= select_task_rq_fair
,
7638 .migrate_task_rq
= migrate_task_rq_fair
,
7640 .rq_online
= rq_online_fair
,
7641 .rq_offline
= rq_offline_fair
,
7643 .task_waking
= task_waking_fair
,
7646 .set_curr_task
= set_curr_task_fair
,
7647 .task_tick
= task_tick_fair
,
7648 .task_fork
= task_fork_fair
,
7650 .prio_changed
= prio_changed_fair
,
7651 .switched_from
= switched_from_fair
,
7652 .switched_to
= switched_to_fair
,
7654 .get_rr_interval
= get_rr_interval_fair
,
7656 #ifdef CONFIG_FAIR_GROUP_SCHED
7657 .task_move_group
= task_move_group_fair
,
7661 #ifdef CONFIG_SCHED_DEBUG
7662 void print_cfs_stats(struct seq_file
*m
, int cpu
)
7664 struct cfs_rq
*cfs_rq
;
7667 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
7668 print_cfs_rq(m
, cpu
, cfs_rq
);
7673 __init
void init_sched_fair_class(void)
7676 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7678 #ifdef CONFIG_NO_HZ_COMMON
7679 nohz
.next_balance
= jiffies
;
7680 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7681 cpu_notifier(sched_ilb_notifier
, 0);