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 ? */
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
)
339 /* return depth at which a sched entity is present in the hierarchy */
340 static inline int depth_se(struct sched_entity
*se
)
344 for_each_sched_entity(se
)
351 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
353 int se_depth
, pse_depth
;
356 * preemption test can be made between sibling entities who are in the
357 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
358 * both tasks until we find their ancestors who are siblings of common
362 /* First walk up until both entities are at same depth */
363 se_depth
= depth_se(*se
);
364 pse_depth
= depth_se(*pse
);
366 while (se_depth
> pse_depth
) {
368 *se
= parent_entity(*se
);
371 while (pse_depth
> se_depth
) {
373 *pse
= parent_entity(*pse
);
376 while (!is_same_group(*se
, *pse
)) {
377 *se
= parent_entity(*se
);
378 *pse
= parent_entity(*pse
);
382 #else /* !CONFIG_FAIR_GROUP_SCHED */
384 static inline struct task_struct
*task_of(struct sched_entity
*se
)
386 return container_of(se
, struct task_struct
, se
);
389 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
391 return container_of(cfs_rq
, struct rq
, cfs
);
394 #define entity_is_task(se) 1
396 #define for_each_sched_entity(se) \
397 for (; se; se = NULL)
399 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
401 return &task_rq(p
)->cfs
;
404 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
406 struct task_struct
*p
= task_of(se
);
407 struct rq
*rq
= task_rq(p
);
412 /* runqueue "owned" by this group */
413 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
418 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
422 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
426 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
427 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
430 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
435 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
441 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
445 #endif /* CONFIG_FAIR_GROUP_SCHED */
447 static __always_inline
448 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
450 /**************************************************************
451 * Scheduling class tree data structure manipulation methods:
454 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
456 s64 delta
= (s64
)(vruntime
- max_vruntime
);
458 max_vruntime
= vruntime
;
463 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
465 s64 delta
= (s64
)(vruntime
- min_vruntime
);
467 min_vruntime
= vruntime
;
472 static inline int entity_before(struct sched_entity
*a
,
473 struct sched_entity
*b
)
475 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
478 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
480 u64 vruntime
= cfs_rq
->min_vruntime
;
483 vruntime
= cfs_rq
->curr
->vruntime
;
485 if (cfs_rq
->rb_leftmost
) {
486 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
491 vruntime
= se
->vruntime
;
493 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
496 /* ensure we never gain time by being placed backwards. */
497 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
500 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
505 * Enqueue an entity into the rb-tree:
507 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
509 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
510 struct rb_node
*parent
= NULL
;
511 struct sched_entity
*entry
;
515 * Find the right place in the rbtree:
519 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
521 * We dont care about collisions. Nodes with
522 * the same key stay together.
524 if (entity_before(se
, entry
)) {
525 link
= &parent
->rb_left
;
527 link
= &parent
->rb_right
;
533 * Maintain a cache of leftmost tree entries (it is frequently
537 cfs_rq
->rb_leftmost
= &se
->run_node
;
539 rb_link_node(&se
->run_node
, parent
, link
);
540 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
543 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
545 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
546 struct rb_node
*next_node
;
548 next_node
= rb_next(&se
->run_node
);
549 cfs_rq
->rb_leftmost
= next_node
;
552 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
555 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
557 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
562 return rb_entry(left
, struct sched_entity
, run_node
);
565 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
567 struct rb_node
*next
= rb_next(&se
->run_node
);
572 return rb_entry(next
, struct sched_entity
, run_node
);
575 #ifdef CONFIG_SCHED_DEBUG
576 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
578 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
583 return rb_entry(last
, struct sched_entity
, run_node
);
586 /**************************************************************
587 * Scheduling class statistics methods:
590 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
591 void __user
*buffer
, size_t *lenp
,
594 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
595 int factor
= get_update_sysctl_factor();
600 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
601 sysctl_sched_min_granularity
);
603 #define WRT_SYSCTL(name) \
604 (normalized_sysctl_##name = sysctl_##name / (factor))
605 WRT_SYSCTL(sched_min_granularity
);
606 WRT_SYSCTL(sched_latency
);
607 WRT_SYSCTL(sched_wakeup_granularity
);
617 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
619 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
620 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
626 * The idea is to set a period in which each task runs once.
628 * When there are too many tasks (sched_nr_latency) we have to stretch
629 * this period because otherwise the slices get too small.
631 * p = (nr <= nl) ? l : l*nr/nl
633 static u64
__sched_period(unsigned long nr_running
)
635 u64 period
= sysctl_sched_latency
;
636 unsigned long nr_latency
= sched_nr_latency
;
638 if (unlikely(nr_running
> nr_latency
)) {
639 period
= sysctl_sched_min_granularity
;
640 period
*= nr_running
;
647 * We calculate the wall-time slice from the period by taking a part
648 * proportional to the weight.
652 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
654 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
656 for_each_sched_entity(se
) {
657 struct load_weight
*load
;
658 struct load_weight lw
;
660 cfs_rq
= cfs_rq_of(se
);
661 load
= &cfs_rq
->load
;
663 if (unlikely(!se
->on_rq
)) {
666 update_load_add(&lw
, se
->load
.weight
);
669 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
675 * We calculate the vruntime slice of a to-be-inserted task.
679 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
681 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
685 static unsigned long task_h_load(struct task_struct
*p
);
687 static inline void __update_task_entity_contrib(struct sched_entity
*se
);
689 /* Give new task start runnable values to heavy its load in infant time */
690 void init_task_runnable_average(struct task_struct
*p
)
694 p
->se
.avg
.decay_count
= 0;
695 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
696 p
->se
.avg
.runnable_avg_sum
= slice
;
697 p
->se
.avg
.runnable_avg_period
= slice
;
698 __update_task_entity_contrib(&p
->se
);
701 void init_task_runnable_average(struct task_struct
*p
)
707 * Update the current task's runtime statistics.
709 static void update_curr(struct cfs_rq
*cfs_rq
)
711 struct sched_entity
*curr
= cfs_rq
->curr
;
712 u64 now
= rq_clock_task(rq_of(cfs_rq
));
718 delta_exec
= now
- curr
->exec_start
;
719 if (unlikely((s64
)delta_exec
<= 0))
722 curr
->exec_start
= now
;
724 schedstat_set(curr
->statistics
.exec_max
,
725 max(delta_exec
, curr
->statistics
.exec_max
));
727 curr
->sum_exec_runtime
+= delta_exec
;
728 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
730 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
731 update_min_vruntime(cfs_rq
);
733 if (entity_is_task(curr
)) {
734 struct task_struct
*curtask
= task_of(curr
);
736 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
737 cpuacct_charge(curtask
, delta_exec
);
738 account_group_exec_runtime(curtask
, delta_exec
);
741 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
745 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
747 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
751 * Task is being enqueued - update stats:
753 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
756 * Are we enqueueing a waiting task? (for current tasks
757 * a dequeue/enqueue event is a NOP)
759 if (se
!= cfs_rq
->curr
)
760 update_stats_wait_start(cfs_rq
, se
);
764 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
766 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
767 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
768 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
769 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
770 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
771 #ifdef CONFIG_SCHEDSTATS
772 if (entity_is_task(se
)) {
773 trace_sched_stat_wait(task_of(se
),
774 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
777 schedstat_set(se
->statistics
.wait_start
, 0);
781 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
784 * Mark the end of the wait period if dequeueing a
787 if (se
!= cfs_rq
->curr
)
788 update_stats_wait_end(cfs_rq
, se
);
792 * We are picking a new current task - update its stats:
795 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
798 * We are starting a new run period:
800 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
803 /**************************************************
804 * Scheduling class queueing methods:
807 #ifdef CONFIG_NUMA_BALANCING
809 * Approximate time to scan a full NUMA task in ms. The task scan period is
810 * calculated based on the tasks virtual memory size and
811 * numa_balancing_scan_size.
813 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
814 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
816 /* Portion of address space to scan in MB */
817 unsigned int sysctl_numa_balancing_scan_size
= 256;
819 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
820 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
822 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
824 unsigned long rss
= 0;
825 unsigned long nr_scan_pages
;
828 * Calculations based on RSS as non-present and empty pages are skipped
829 * by the PTE scanner and NUMA hinting faults should be trapped based
832 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
833 rss
= get_mm_rss(p
->mm
);
837 rss
= round_up(rss
, nr_scan_pages
);
838 return rss
/ nr_scan_pages
;
841 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
842 #define MAX_SCAN_WINDOW 2560
844 static unsigned int task_scan_min(struct task_struct
*p
)
846 unsigned int scan
, floor
;
847 unsigned int windows
= 1;
849 if (sysctl_numa_balancing_scan_size
< MAX_SCAN_WINDOW
)
850 windows
= MAX_SCAN_WINDOW
/ sysctl_numa_balancing_scan_size
;
851 floor
= 1000 / windows
;
853 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
854 return max_t(unsigned int, floor
, scan
);
857 static unsigned int task_scan_max(struct task_struct
*p
)
859 unsigned int smin
= task_scan_min(p
);
862 /* Watch for min being lower than max due to floor calculations */
863 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
864 return max(smin
, smax
);
867 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
869 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
870 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
873 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
875 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
876 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
882 spinlock_t lock
; /* nr_tasks, tasks */
885 struct list_head task_list
;
888 nodemask_t active_nodes
;
889 unsigned long total_faults
;
890 unsigned long *faults_cpu
;
891 unsigned long faults
[0];
894 pid_t
task_numa_group_id(struct task_struct
*p
)
896 return p
->numa_group
? p
->numa_group
->gid
: 0;
899 static inline int task_faults_idx(int nid
, int priv
)
901 return 2 * nid
+ priv
;
904 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
906 if (!p
->numa_faults_memory
)
909 return p
->numa_faults_memory
[task_faults_idx(nid
, 0)] +
910 p
->numa_faults_memory
[task_faults_idx(nid
, 1)];
913 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
918 return p
->numa_group
->faults
[task_faults_idx(nid
, 0)] +
919 p
->numa_group
->faults
[task_faults_idx(nid
, 1)];
922 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
924 return group
->faults_cpu
[task_faults_idx(nid
, 0)] +
925 group
->faults_cpu
[task_faults_idx(nid
, 1)];
929 * These return the fraction of accesses done by a particular task, or
930 * task group, on a particular numa node. The group weight is given a
931 * larger multiplier, in order to group tasks together that are almost
932 * evenly spread out between numa nodes.
934 static inline unsigned long task_weight(struct task_struct
*p
, int nid
)
936 unsigned long total_faults
;
938 if (!p
->numa_faults_memory
)
941 total_faults
= p
->total_numa_faults
;
946 return 1000 * task_faults(p
, nid
) / total_faults
;
949 static inline unsigned long group_weight(struct task_struct
*p
, int nid
)
951 if (!p
->numa_group
|| !p
->numa_group
->total_faults
)
954 return 1000 * group_faults(p
, nid
) / p
->numa_group
->total_faults
;
957 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
958 int src_nid
, int dst_cpu
)
960 struct numa_group
*ng
= p
->numa_group
;
961 int dst_nid
= cpu_to_node(dst_cpu
);
962 int last_cpupid
, this_cpupid
;
964 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
967 * Multi-stage node selection is used in conjunction with a periodic
968 * migration fault to build a temporal task<->page relation. By using
969 * a two-stage filter we remove short/unlikely relations.
971 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
972 * a task's usage of a particular page (n_p) per total usage of this
973 * page (n_t) (in a given time-span) to a probability.
975 * Our periodic faults will sample this probability and getting the
976 * same result twice in a row, given these samples are fully
977 * independent, is then given by P(n)^2, provided our sample period
978 * is sufficiently short compared to the usage pattern.
980 * This quadric squishes small probabilities, making it less likely we
981 * act on an unlikely task<->page relation.
983 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
984 if (!cpupid_pid_unset(last_cpupid
) &&
985 cpupid_to_nid(last_cpupid
) != dst_nid
)
988 /* Always allow migrate on private faults */
989 if (cpupid_match_pid(p
, last_cpupid
))
992 /* A shared fault, but p->numa_group has not been set up yet. */
997 * Do not migrate if the destination is not a node that
998 * is actively used by this numa group.
1000 if (!node_isset(dst_nid
, ng
->active_nodes
))
1004 * Source is a node that is not actively used by this
1005 * numa group, while the destination is. Migrate.
1007 if (!node_isset(src_nid
, ng
->active_nodes
))
1011 * Both source and destination are nodes in active
1012 * use by this numa group. Maximize memory bandwidth
1013 * by migrating from more heavily used groups, to less
1014 * heavily used ones, spreading the load around.
1015 * Use a 1/4 hysteresis to avoid spurious page movement.
1017 return group_faults(p
, dst_nid
) < (group_faults(p
, src_nid
) * 3 / 4);
1020 static unsigned long weighted_cpuload(const int cpu
);
1021 static unsigned long source_load(int cpu
, int type
);
1022 static unsigned long target_load(int cpu
, int type
);
1023 static unsigned long power_of(int cpu
);
1024 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
1026 /* Cached statistics for all CPUs within a node */
1028 unsigned long nr_running
;
1031 /* Total compute capacity of CPUs on a node */
1032 unsigned long power
;
1034 /* Approximate capacity in terms of runnable tasks on a node */
1035 unsigned long capacity
;
1040 * XXX borrowed from update_sg_lb_stats
1042 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1046 memset(ns
, 0, sizeof(*ns
));
1047 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1048 struct rq
*rq
= cpu_rq(cpu
);
1050 ns
->nr_running
+= rq
->nr_running
;
1051 ns
->load
+= weighted_cpuload(cpu
);
1052 ns
->power
+= power_of(cpu
);
1058 * If we raced with hotplug and there are no CPUs left in our mask
1059 * the @ns structure is NULL'ed and task_numa_compare() will
1060 * not find this node attractive.
1062 * We'll either bail at !has_capacity, or we'll detect a huge imbalance
1068 ns
->load
= (ns
->load
* SCHED_POWER_SCALE
) / ns
->power
;
1069 ns
->capacity
= DIV_ROUND_CLOSEST(ns
->power
, SCHED_POWER_SCALE
);
1070 ns
->has_capacity
= (ns
->nr_running
< ns
->capacity
);
1073 struct task_numa_env
{
1074 struct task_struct
*p
;
1076 int src_cpu
, src_nid
;
1077 int dst_cpu
, dst_nid
;
1079 struct numa_stats src_stats
, dst_stats
;
1083 struct task_struct
*best_task
;
1088 static void task_numa_assign(struct task_numa_env
*env
,
1089 struct task_struct
*p
, long imp
)
1092 put_task_struct(env
->best_task
);
1097 env
->best_imp
= imp
;
1098 env
->best_cpu
= env
->dst_cpu
;
1102 * This checks if the overall compute and NUMA accesses of the system would
1103 * be improved if the source tasks was migrated to the target dst_cpu taking
1104 * into account that it might be best if task running on the dst_cpu should
1105 * be exchanged with the source task
1107 static void task_numa_compare(struct task_numa_env
*env
,
1108 long taskimp
, long groupimp
)
1110 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1111 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1112 struct task_struct
*cur
;
1113 long dst_load
, src_load
;
1115 long imp
= (groupimp
> 0) ? groupimp
: taskimp
;
1118 cur
= ACCESS_ONCE(dst_rq
->curr
);
1119 if (cur
->pid
== 0) /* idle */
1123 * "imp" is the fault differential for the source task between the
1124 * source and destination node. Calculate the total differential for
1125 * the source task and potential destination task. The more negative
1126 * the value is, the more rmeote accesses that would be expected to
1127 * be incurred if the tasks were swapped.
1130 /* Skip this swap candidate if cannot move to the source cpu */
1131 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1135 * If dst and source tasks are in the same NUMA group, or not
1136 * in any group then look only at task weights.
1138 if (cur
->numa_group
== env
->p
->numa_group
) {
1139 imp
= taskimp
+ task_weight(cur
, env
->src_nid
) -
1140 task_weight(cur
, env
->dst_nid
);
1142 * Add some hysteresis to prevent swapping the
1143 * tasks within a group over tiny differences.
1145 if (cur
->numa_group
)
1149 * Compare the group weights. If a task is all by
1150 * itself (not part of a group), use the task weight
1153 if (env
->p
->numa_group
)
1158 if (cur
->numa_group
)
1159 imp
+= group_weight(cur
, env
->src_nid
) -
1160 group_weight(cur
, env
->dst_nid
);
1162 imp
+= task_weight(cur
, env
->src_nid
) -
1163 task_weight(cur
, env
->dst_nid
);
1167 if (imp
< env
->best_imp
)
1171 /* Is there capacity at our destination? */
1172 if (env
->src_stats
.has_capacity
&&
1173 !env
->dst_stats
.has_capacity
)
1179 /* Balance doesn't matter much if we're running a task per cpu */
1180 if (src_rq
->nr_running
== 1 && dst_rq
->nr_running
== 1)
1184 * In the overloaded case, try and keep the load balanced.
1187 dst_load
= env
->dst_stats
.load
;
1188 src_load
= env
->src_stats
.load
;
1190 /* XXX missing power terms */
1191 load
= task_h_load(env
->p
);
1196 load
= task_h_load(cur
);
1201 /* make src_load the smaller */
1202 if (dst_load
< src_load
)
1203 swap(dst_load
, src_load
);
1205 if (src_load
* env
->imbalance_pct
< dst_load
* 100)
1209 task_numa_assign(env
, cur
, imp
);
1214 static void task_numa_find_cpu(struct task_numa_env
*env
,
1215 long taskimp
, long groupimp
)
1219 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1220 /* Skip this CPU if the source task cannot migrate */
1221 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1225 task_numa_compare(env
, taskimp
, groupimp
);
1229 static int task_numa_migrate(struct task_struct
*p
)
1231 struct task_numa_env env
= {
1234 .src_cpu
= task_cpu(p
),
1235 .src_nid
= task_node(p
),
1237 .imbalance_pct
= 112,
1243 struct sched_domain
*sd
;
1244 unsigned long taskweight
, groupweight
;
1246 long taskimp
, groupimp
;
1249 * Pick the lowest SD_NUMA domain, as that would have the smallest
1250 * imbalance and would be the first to start moving tasks about.
1252 * And we want to avoid any moving of tasks about, as that would create
1253 * random movement of tasks -- counter the numa conditions we're trying
1257 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1259 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1263 * Cpusets can break the scheduler domain tree into smaller
1264 * balance domains, some of which do not cross NUMA boundaries.
1265 * Tasks that are "trapped" in such domains cannot be migrated
1266 * elsewhere, so there is no point in (re)trying.
1268 if (unlikely(!sd
)) {
1269 p
->numa_preferred_nid
= task_node(p
);
1273 taskweight
= task_weight(p
, env
.src_nid
);
1274 groupweight
= group_weight(p
, env
.src_nid
);
1275 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1276 env
.dst_nid
= p
->numa_preferred_nid
;
1277 taskimp
= task_weight(p
, env
.dst_nid
) - taskweight
;
1278 groupimp
= group_weight(p
, env
.dst_nid
) - groupweight
;
1279 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1281 /* If the preferred nid has capacity, try to use it. */
1282 if (env
.dst_stats
.has_capacity
)
1283 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1285 /* No space available on the preferred nid. Look elsewhere. */
1286 if (env
.best_cpu
== -1) {
1287 for_each_online_node(nid
) {
1288 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1291 /* Only consider nodes where both task and groups benefit */
1292 taskimp
= task_weight(p
, nid
) - taskweight
;
1293 groupimp
= group_weight(p
, nid
) - groupweight
;
1294 if (taskimp
< 0 && groupimp
< 0)
1298 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1299 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1303 /* No better CPU than the current one was found. */
1304 if (env
.best_cpu
== -1)
1307 sched_setnuma(p
, env
.dst_nid
);
1310 * Reset the scan period if the task is being rescheduled on an
1311 * alternative node to recheck if the tasks is now properly placed.
1313 p
->numa_scan_period
= task_scan_min(p
);
1315 if (env
.best_task
== NULL
) {
1316 int ret
= migrate_task_to(p
, env
.best_cpu
);
1320 ret
= migrate_swap(p
, env
.best_task
);
1321 put_task_struct(env
.best_task
);
1325 /* Attempt to migrate a task to a CPU on the preferred node. */
1326 static void numa_migrate_preferred(struct task_struct
*p
)
1328 /* This task has no NUMA fault statistics yet */
1329 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults_memory
))
1332 /* Periodically retry migrating the task to the preferred node */
1333 p
->numa_migrate_retry
= jiffies
+ HZ
;
1335 /* Success if task is already running on preferred CPU */
1336 if (task_node(p
) == p
->numa_preferred_nid
)
1339 /* Otherwise, try migrate to a CPU on the preferred node */
1340 task_numa_migrate(p
);
1344 * Find the nodes on which the workload is actively running. We do this by
1345 * tracking the nodes from which NUMA hinting faults are triggered. This can
1346 * be different from the set of nodes where the workload's memory is currently
1349 * The bitmask is used to make smarter decisions on when to do NUMA page
1350 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1351 * are added when they cause over 6/16 of the maximum number of faults, but
1352 * only removed when they drop below 3/16.
1354 static void update_numa_active_node_mask(struct numa_group
*numa_group
)
1356 unsigned long faults
, max_faults
= 0;
1359 for_each_online_node(nid
) {
1360 faults
= group_faults_cpu(numa_group
, nid
);
1361 if (faults
> max_faults
)
1362 max_faults
= faults
;
1365 for_each_online_node(nid
) {
1366 faults
= group_faults_cpu(numa_group
, nid
);
1367 if (!node_isset(nid
, numa_group
->active_nodes
)) {
1368 if (faults
> max_faults
* 6 / 16)
1369 node_set(nid
, numa_group
->active_nodes
);
1370 } else if (faults
< max_faults
* 3 / 16)
1371 node_clear(nid
, numa_group
->active_nodes
);
1376 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1377 * increments. The more local the fault statistics are, the higher the scan
1378 * period will be for the next scan window. If local/remote ratio is below
1379 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1380 * scan period will decrease
1382 #define NUMA_PERIOD_SLOTS 10
1383 #define NUMA_PERIOD_THRESHOLD 3
1386 * Increase the scan period (slow down scanning) if the majority of
1387 * our memory is already on our local node, or if the majority of
1388 * the page accesses are shared with other processes.
1389 * Otherwise, decrease the scan period.
1391 static void update_task_scan_period(struct task_struct
*p
,
1392 unsigned long shared
, unsigned long private)
1394 unsigned int period_slot
;
1398 unsigned long remote
= p
->numa_faults_locality
[0];
1399 unsigned long local
= p
->numa_faults_locality
[1];
1402 * If there were no record hinting faults then either the task is
1403 * completely idle or all activity is areas that are not of interest
1404 * to automatic numa balancing. Scan slower
1406 if (local
+ shared
== 0) {
1407 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1408 p
->numa_scan_period
<< 1);
1410 p
->mm
->numa_next_scan
= jiffies
+
1411 msecs_to_jiffies(p
->numa_scan_period
);
1417 * Prepare to scale scan period relative to the current period.
1418 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1419 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1420 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1422 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1423 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1424 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1425 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1428 diff
= slot
* period_slot
;
1430 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1433 * Scale scan rate increases based on sharing. There is an
1434 * inverse relationship between the degree of sharing and
1435 * the adjustment made to the scanning period. Broadly
1436 * speaking the intent is that there is little point
1437 * scanning faster if shared accesses dominate as it may
1438 * simply bounce migrations uselessly
1440 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
));
1441 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1444 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1445 task_scan_min(p
), task_scan_max(p
));
1446 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1449 static void task_numa_placement(struct task_struct
*p
)
1451 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1452 unsigned long max_faults
= 0, max_group_faults
= 0;
1453 unsigned long fault_types
[2] = { 0, 0 };
1454 spinlock_t
*group_lock
= NULL
;
1456 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1457 if (p
->numa_scan_seq
== seq
)
1459 p
->numa_scan_seq
= seq
;
1460 p
->numa_scan_period_max
= task_scan_max(p
);
1462 /* If the task is part of a group prevent parallel updates to group stats */
1463 if (p
->numa_group
) {
1464 group_lock
= &p
->numa_group
->lock
;
1465 spin_lock(group_lock
);
1468 /* Find the node with the highest number of faults */
1469 for_each_online_node(nid
) {
1470 unsigned long faults
= 0, group_faults
= 0;
1473 for (priv
= 0; priv
< 2; priv
++) {
1476 i
= task_faults_idx(nid
, priv
);
1477 diff
= -p
->numa_faults_memory
[i
];
1478 f_diff
= -p
->numa_faults_cpu
[i
];
1480 /* Decay existing window, copy faults since last scan */
1481 p
->numa_faults_memory
[i
] >>= 1;
1482 p
->numa_faults_memory
[i
] += p
->numa_faults_buffer_memory
[i
];
1483 fault_types
[priv
] += p
->numa_faults_buffer_memory
[i
];
1484 p
->numa_faults_buffer_memory
[i
] = 0;
1486 p
->numa_faults_cpu
[i
] >>= 1;
1487 p
->numa_faults_cpu
[i
] += p
->numa_faults_buffer_cpu
[i
];
1488 p
->numa_faults_buffer_cpu
[i
] = 0;
1490 faults
+= p
->numa_faults_memory
[i
];
1491 diff
+= p
->numa_faults_memory
[i
];
1492 f_diff
+= p
->numa_faults_cpu
[i
];
1493 p
->total_numa_faults
+= diff
;
1494 if (p
->numa_group
) {
1495 /* safe because we can only change our own group */
1496 p
->numa_group
->faults
[i
] += diff
;
1497 p
->numa_group
->faults_cpu
[i
] += f_diff
;
1498 p
->numa_group
->total_faults
+= diff
;
1499 group_faults
+= p
->numa_group
->faults
[i
];
1503 if (faults
> max_faults
) {
1504 max_faults
= faults
;
1508 if (group_faults
> max_group_faults
) {
1509 max_group_faults
= group_faults
;
1510 max_group_nid
= nid
;
1514 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1516 if (p
->numa_group
) {
1517 update_numa_active_node_mask(p
->numa_group
);
1519 * If the preferred task and group nids are different,
1520 * iterate over the nodes again to find the best place.
1522 if (max_nid
!= max_group_nid
) {
1523 unsigned long weight
, max_weight
= 0;
1525 for_each_online_node(nid
) {
1526 weight
= task_weight(p
, nid
) + group_weight(p
, nid
);
1527 if (weight
> max_weight
) {
1528 max_weight
= weight
;
1534 spin_unlock(group_lock
);
1537 /* Preferred node as the node with the most faults */
1538 if (max_faults
&& max_nid
!= p
->numa_preferred_nid
) {
1539 /* Update the preferred nid and migrate task if possible */
1540 sched_setnuma(p
, max_nid
);
1541 numa_migrate_preferred(p
);
1545 static inline int get_numa_group(struct numa_group
*grp
)
1547 return atomic_inc_not_zero(&grp
->refcount
);
1550 static inline void put_numa_group(struct numa_group
*grp
)
1552 if (atomic_dec_and_test(&grp
->refcount
))
1553 kfree_rcu(grp
, rcu
);
1556 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1559 struct numa_group
*grp
, *my_grp
;
1560 struct task_struct
*tsk
;
1562 int cpu
= cpupid_to_cpu(cpupid
);
1565 if (unlikely(!p
->numa_group
)) {
1566 unsigned int size
= sizeof(struct numa_group
) +
1567 4*nr_node_ids
*sizeof(unsigned long);
1569 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1573 atomic_set(&grp
->refcount
, 1);
1574 spin_lock_init(&grp
->lock
);
1575 INIT_LIST_HEAD(&grp
->task_list
);
1577 /* Second half of the array tracks nids where faults happen */
1578 grp
->faults_cpu
= grp
->faults
+ 2 * nr_node_ids
;
1580 node_set(task_node(current
), grp
->active_nodes
);
1582 for (i
= 0; i
< 4*nr_node_ids
; i
++)
1583 grp
->faults
[i
] = p
->numa_faults_memory
[i
];
1585 grp
->total_faults
= p
->total_numa_faults
;
1587 list_add(&p
->numa_entry
, &grp
->task_list
);
1589 rcu_assign_pointer(p
->numa_group
, grp
);
1593 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1595 if (!cpupid_match_pid(tsk
, cpupid
))
1598 grp
= rcu_dereference(tsk
->numa_group
);
1602 my_grp
= p
->numa_group
;
1607 * Only join the other group if its bigger; if we're the bigger group,
1608 * the other task will join us.
1610 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1614 * Tie-break on the grp address.
1616 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1619 /* Always join threads in the same process. */
1620 if (tsk
->mm
== current
->mm
)
1623 /* Simple filter to avoid false positives due to PID collisions */
1624 if (flags
& TNF_SHARED
)
1627 /* Update priv based on whether false sharing was detected */
1630 if (join
&& !get_numa_group(grp
))
1638 double_lock(&my_grp
->lock
, &grp
->lock
);
1640 for (i
= 0; i
< 4*nr_node_ids
; i
++) {
1641 my_grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1642 grp
->faults
[i
] += p
->numa_faults_memory
[i
];
1644 my_grp
->total_faults
-= p
->total_numa_faults
;
1645 grp
->total_faults
+= p
->total_numa_faults
;
1647 list_move(&p
->numa_entry
, &grp
->task_list
);
1651 spin_unlock(&my_grp
->lock
);
1652 spin_unlock(&grp
->lock
);
1654 rcu_assign_pointer(p
->numa_group
, grp
);
1656 put_numa_group(my_grp
);
1664 void task_numa_free(struct task_struct
*p
)
1666 struct numa_group
*grp
= p
->numa_group
;
1668 void *numa_faults
= p
->numa_faults_memory
;
1671 spin_lock(&grp
->lock
);
1672 for (i
= 0; i
< 4*nr_node_ids
; i
++)
1673 grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1674 grp
->total_faults
-= p
->total_numa_faults
;
1676 list_del(&p
->numa_entry
);
1678 spin_unlock(&grp
->lock
);
1679 rcu_assign_pointer(p
->numa_group
, NULL
);
1680 put_numa_group(grp
);
1683 p
->numa_faults_memory
= NULL
;
1684 p
->numa_faults_buffer_memory
= NULL
;
1685 p
->numa_faults_cpu
= NULL
;
1686 p
->numa_faults_buffer_cpu
= NULL
;
1691 * Got a PROT_NONE fault for a page on @node.
1693 void task_numa_fault(int last_cpupid
, int node
, int pages
, int flags
)
1695 struct task_struct
*p
= current
;
1696 bool migrated
= flags
& TNF_MIGRATED
;
1697 int this_node
= task_node(current
);
1700 if (!numabalancing_enabled
)
1703 /* for example, ksmd faulting in a user's mm */
1707 /* Do not worry about placement if exiting */
1708 if (p
->state
== TASK_DEAD
)
1711 /* Allocate buffer to track faults on a per-node basis */
1712 if (unlikely(!p
->numa_faults_memory
)) {
1713 int size
= sizeof(*p
->numa_faults_memory
) * 4 * nr_node_ids
;
1715 /* numa_faults and numa_faults_buffer share the allocation */
1716 p
->numa_faults_memory
= kzalloc(size
* 2, GFP_KERNEL
|__GFP_NOWARN
);
1717 if (!p
->numa_faults_memory
)
1720 BUG_ON(p
->numa_faults_buffer_memory
);
1721 p
->numa_faults_cpu
= p
->numa_faults_memory
+ (2 * nr_node_ids
);
1722 p
->numa_faults_buffer_memory
= p
->numa_faults_memory
+ (4 * nr_node_ids
);
1723 p
->numa_faults_buffer_cpu
= p
->numa_faults_memory
+ (6 * nr_node_ids
);
1724 p
->total_numa_faults
= 0;
1725 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1729 * First accesses are treated as private, otherwise consider accesses
1730 * to be private if the accessing pid has not changed
1732 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
1735 priv
= cpupid_match_pid(p
, last_cpupid
);
1736 if (!priv
&& !(flags
& TNF_NO_GROUP
))
1737 task_numa_group(p
, last_cpupid
, flags
, &priv
);
1740 task_numa_placement(p
);
1743 * Retry task to preferred node migration periodically, in case it
1744 * case it previously failed, or the scheduler moved us.
1746 if (time_after(jiffies
, p
->numa_migrate_retry
))
1747 numa_migrate_preferred(p
);
1750 p
->numa_pages_migrated
+= pages
;
1752 p
->numa_faults_buffer_memory
[task_faults_idx(node
, priv
)] += pages
;
1753 p
->numa_faults_buffer_cpu
[task_faults_idx(this_node
, priv
)] += pages
;
1754 p
->numa_faults_locality
[!!(flags
& TNF_FAULT_LOCAL
)] += pages
;
1757 static void reset_ptenuma_scan(struct task_struct
*p
)
1759 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
1760 p
->mm
->numa_scan_offset
= 0;
1764 * The expensive part of numa migration is done from task_work context.
1765 * Triggered from task_tick_numa().
1767 void task_numa_work(struct callback_head
*work
)
1769 unsigned long migrate
, next_scan
, now
= jiffies
;
1770 struct task_struct
*p
= current
;
1771 struct mm_struct
*mm
= p
->mm
;
1772 struct vm_area_struct
*vma
;
1773 unsigned long start
, end
;
1774 unsigned long nr_pte_updates
= 0;
1777 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
1779 work
->next
= work
; /* protect against double add */
1781 * Who cares about NUMA placement when they're dying.
1783 * NOTE: make sure not to dereference p->mm before this check,
1784 * exit_task_work() happens _after_ exit_mm() so we could be called
1785 * without p->mm even though we still had it when we enqueued this
1788 if (p
->flags
& PF_EXITING
)
1791 if (!mm
->numa_next_scan
) {
1792 mm
->numa_next_scan
= now
+
1793 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1797 * Enforce maximal scan/migration frequency..
1799 migrate
= mm
->numa_next_scan
;
1800 if (time_before(now
, migrate
))
1803 if (p
->numa_scan_period
== 0) {
1804 p
->numa_scan_period_max
= task_scan_max(p
);
1805 p
->numa_scan_period
= task_scan_min(p
);
1808 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1809 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
1813 * Delay this task enough that another task of this mm will likely win
1814 * the next time around.
1816 p
->node_stamp
+= 2 * TICK_NSEC
;
1818 start
= mm
->numa_scan_offset
;
1819 pages
= sysctl_numa_balancing_scan_size
;
1820 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
1824 down_read(&mm
->mmap_sem
);
1825 vma
= find_vma(mm
, start
);
1827 reset_ptenuma_scan(p
);
1831 for (; vma
; vma
= vma
->vm_next
) {
1832 if (!vma_migratable(vma
) || !vma_policy_mof(p
, vma
))
1836 * Shared library pages mapped by multiple processes are not
1837 * migrated as it is expected they are cache replicated. Avoid
1838 * hinting faults in read-only file-backed mappings or the vdso
1839 * as migrating the pages will be of marginal benefit.
1842 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
1846 * Skip inaccessible VMAs to avoid any confusion between
1847 * PROT_NONE and NUMA hinting ptes
1849 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
1853 start
= max(start
, vma
->vm_start
);
1854 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
1855 end
= min(end
, vma
->vm_end
);
1856 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
1859 * Scan sysctl_numa_balancing_scan_size but ensure that
1860 * at least one PTE is updated so that unused virtual
1861 * address space is quickly skipped.
1864 pages
-= (end
- start
) >> PAGE_SHIFT
;
1869 } while (end
!= vma
->vm_end
);
1874 * It is possible to reach the end of the VMA list but the last few
1875 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1876 * would find the !migratable VMA on the next scan but not reset the
1877 * scanner to the start so check it now.
1880 mm
->numa_scan_offset
= start
;
1882 reset_ptenuma_scan(p
);
1883 up_read(&mm
->mmap_sem
);
1887 * Drive the periodic memory faults..
1889 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1891 struct callback_head
*work
= &curr
->numa_work
;
1895 * We don't care about NUMA placement if we don't have memory.
1897 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
1901 * Using runtime rather than walltime has the dual advantage that
1902 * we (mostly) drive the selection from busy threads and that the
1903 * task needs to have done some actual work before we bother with
1906 now
= curr
->se
.sum_exec_runtime
;
1907 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
1909 if (now
- curr
->node_stamp
> period
) {
1910 if (!curr
->node_stamp
)
1911 curr
->numa_scan_period
= task_scan_min(curr
);
1912 curr
->node_stamp
+= period
;
1914 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
1915 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
1916 task_work_add(curr
, work
, true);
1921 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1925 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
1929 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
1932 #endif /* CONFIG_NUMA_BALANCING */
1935 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1937 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
1938 if (!parent_entity(se
))
1939 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1941 if (entity_is_task(se
)) {
1942 struct rq
*rq
= rq_of(cfs_rq
);
1944 account_numa_enqueue(rq
, task_of(se
));
1945 list_add(&se
->group_node
, &rq
->cfs_tasks
);
1948 cfs_rq
->nr_running
++;
1952 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1954 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
1955 if (!parent_entity(se
))
1956 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1957 if (entity_is_task(se
)) {
1958 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
1959 list_del_init(&se
->group_node
);
1961 cfs_rq
->nr_running
--;
1964 #ifdef CONFIG_FAIR_GROUP_SCHED
1966 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
1971 * Use this CPU's actual weight instead of the last load_contribution
1972 * to gain a more accurate current total weight. See
1973 * update_cfs_rq_load_contribution().
1975 tg_weight
= atomic_long_read(&tg
->load_avg
);
1976 tg_weight
-= cfs_rq
->tg_load_contrib
;
1977 tg_weight
+= cfs_rq
->load
.weight
;
1982 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1984 long tg_weight
, load
, shares
;
1986 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
1987 load
= cfs_rq
->load
.weight
;
1989 shares
= (tg
->shares
* load
);
1991 shares
/= tg_weight
;
1993 if (shares
< MIN_SHARES
)
1994 shares
= MIN_SHARES
;
1995 if (shares
> tg
->shares
)
1996 shares
= tg
->shares
;
2000 # else /* CONFIG_SMP */
2001 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2005 # endif /* CONFIG_SMP */
2006 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2007 unsigned long weight
)
2010 /* commit outstanding execution time */
2011 if (cfs_rq
->curr
== se
)
2012 update_curr(cfs_rq
);
2013 account_entity_dequeue(cfs_rq
, se
);
2016 update_load_set(&se
->load
, weight
);
2019 account_entity_enqueue(cfs_rq
, se
);
2022 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2024 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2026 struct task_group
*tg
;
2027 struct sched_entity
*se
;
2031 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2032 if (!se
|| throttled_hierarchy(cfs_rq
))
2035 if (likely(se
->load
.weight
== tg
->shares
))
2038 shares
= calc_cfs_shares(cfs_rq
, tg
);
2040 reweight_entity(cfs_rq_of(se
), se
, shares
);
2042 #else /* CONFIG_FAIR_GROUP_SCHED */
2043 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2046 #endif /* CONFIG_FAIR_GROUP_SCHED */
2050 * We choose a half-life close to 1 scheduling period.
2051 * Note: The tables below are dependent on this value.
2053 #define LOAD_AVG_PERIOD 32
2054 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2055 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2057 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2058 static const u32 runnable_avg_yN_inv
[] = {
2059 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2060 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2061 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2062 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2063 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2064 0x85aac367, 0x82cd8698,
2068 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2069 * over-estimates when re-combining.
2071 static const u32 runnable_avg_yN_sum
[] = {
2072 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2073 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2074 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2079 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2081 static __always_inline u64
decay_load(u64 val
, u64 n
)
2083 unsigned int local_n
;
2087 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2090 /* after bounds checking we can collapse to 32-bit */
2094 * As y^PERIOD = 1/2, we can combine
2095 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2096 * With a look-up table which covers k^n (n<PERIOD)
2098 * To achieve constant time decay_load.
2100 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2101 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2102 local_n
%= LOAD_AVG_PERIOD
;
2105 val
*= runnable_avg_yN_inv
[local_n
];
2106 /* We don't use SRR here since we always want to round down. */
2111 * For updates fully spanning n periods, the contribution to runnable
2112 * average will be: \Sum 1024*y^n
2114 * We can compute this reasonably efficiently by combining:
2115 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2117 static u32
__compute_runnable_contrib(u64 n
)
2121 if (likely(n
<= LOAD_AVG_PERIOD
))
2122 return runnable_avg_yN_sum
[n
];
2123 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2124 return LOAD_AVG_MAX
;
2126 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2128 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2129 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2131 n
-= LOAD_AVG_PERIOD
;
2132 } while (n
> LOAD_AVG_PERIOD
);
2134 contrib
= decay_load(contrib
, n
);
2135 return contrib
+ runnable_avg_yN_sum
[n
];
2139 * We can represent the historical contribution to runnable average as the
2140 * coefficients of a geometric series. To do this we sub-divide our runnable
2141 * history into segments of approximately 1ms (1024us); label the segment that
2142 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2144 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2146 * (now) (~1ms ago) (~2ms ago)
2148 * Let u_i denote the fraction of p_i that the entity was runnable.
2150 * We then designate the fractions u_i as our co-efficients, yielding the
2151 * following representation of historical load:
2152 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2154 * We choose y based on the with of a reasonably scheduling period, fixing:
2157 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2158 * approximately half as much as the contribution to load within the last ms
2161 * When a period "rolls over" and we have new u_0`, multiplying the previous
2162 * sum again by y is sufficient to update:
2163 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2164 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2166 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2167 struct sched_avg
*sa
,
2171 u32 runnable_contrib
;
2172 int delta_w
, decayed
= 0;
2174 delta
= now
- sa
->last_runnable_update
;
2176 * This should only happen when time goes backwards, which it
2177 * unfortunately does during sched clock init when we swap over to TSC.
2179 if ((s64
)delta
< 0) {
2180 sa
->last_runnable_update
= now
;
2185 * Use 1024ns as the unit of measurement since it's a reasonable
2186 * approximation of 1us and fast to compute.
2191 sa
->last_runnable_update
= now
;
2193 /* delta_w is the amount already accumulated against our next period */
2194 delta_w
= sa
->runnable_avg_period
% 1024;
2195 if (delta
+ delta_w
>= 1024) {
2196 /* period roll-over */
2200 * Now that we know we're crossing a period boundary, figure
2201 * out how much from delta we need to complete the current
2202 * period and accrue it.
2204 delta_w
= 1024 - delta_w
;
2206 sa
->runnable_avg_sum
+= delta_w
;
2207 sa
->runnable_avg_period
+= delta_w
;
2211 /* Figure out how many additional periods this update spans */
2212 periods
= delta
/ 1024;
2215 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2217 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2220 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2221 runnable_contrib
= __compute_runnable_contrib(periods
);
2223 sa
->runnable_avg_sum
+= runnable_contrib
;
2224 sa
->runnable_avg_period
+= runnable_contrib
;
2227 /* Remainder of delta accrued against u_0` */
2229 sa
->runnable_avg_sum
+= delta
;
2230 sa
->runnable_avg_period
+= delta
;
2235 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2236 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2238 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2239 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2241 decays
-= se
->avg
.decay_count
;
2245 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2246 se
->avg
.decay_count
= 0;
2251 #ifdef CONFIG_FAIR_GROUP_SCHED
2252 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2255 struct task_group
*tg
= cfs_rq
->tg
;
2258 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2259 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2261 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2262 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2263 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2268 * Aggregate cfs_rq runnable averages into an equivalent task_group
2269 * representation for computing load contributions.
2271 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2272 struct cfs_rq
*cfs_rq
)
2274 struct task_group
*tg
= cfs_rq
->tg
;
2277 /* The fraction of a cpu used by this cfs_rq */
2278 contrib
= div_u64((u64
)sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2279 sa
->runnable_avg_period
+ 1);
2280 contrib
-= cfs_rq
->tg_runnable_contrib
;
2282 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2283 atomic_add(contrib
, &tg
->runnable_avg
);
2284 cfs_rq
->tg_runnable_contrib
+= contrib
;
2288 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2290 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2291 struct task_group
*tg
= cfs_rq
->tg
;
2296 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2297 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2298 atomic_long_read(&tg
->load_avg
) + 1);
2301 * For group entities we need to compute a correction term in the case
2302 * that they are consuming <1 cpu so that we would contribute the same
2303 * load as a task of equal weight.
2305 * Explicitly co-ordinating this measurement would be expensive, but
2306 * fortunately the sum of each cpus contribution forms a usable
2307 * lower-bound on the true value.
2309 * Consider the aggregate of 2 contributions. Either they are disjoint
2310 * (and the sum represents true value) or they are disjoint and we are
2311 * understating by the aggregate of their overlap.
2313 * Extending this to N cpus, for a given overlap, the maximum amount we
2314 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2315 * cpus that overlap for this interval and w_i is the interval width.
2317 * On a small machine; the first term is well-bounded which bounds the
2318 * total error since w_i is a subset of the period. Whereas on a
2319 * larger machine, while this first term can be larger, if w_i is the
2320 * of consequential size guaranteed to see n_i*w_i quickly converge to
2321 * our upper bound of 1-cpu.
2323 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2324 if (runnable_avg
< NICE_0_LOAD
) {
2325 se
->avg
.load_avg_contrib
*= runnable_avg
;
2326 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2330 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2331 int force_update
) {}
2332 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2333 struct cfs_rq
*cfs_rq
) {}
2334 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2337 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2341 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2342 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2343 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2344 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2347 /* Compute the current contribution to load_avg by se, return any delta */
2348 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2350 long old_contrib
= se
->avg
.load_avg_contrib
;
2352 if (entity_is_task(se
)) {
2353 __update_task_entity_contrib(se
);
2355 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2356 __update_group_entity_contrib(se
);
2359 return se
->avg
.load_avg_contrib
- old_contrib
;
2362 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2365 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2366 cfs_rq
->blocked_load_avg
-= load_contrib
;
2368 cfs_rq
->blocked_load_avg
= 0;
2371 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2373 /* Update a sched_entity's runnable average */
2374 static inline void update_entity_load_avg(struct sched_entity
*se
,
2377 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2382 * For a group entity we need to use their owned cfs_rq_clock_task() in
2383 * case they are the parent of a throttled hierarchy.
2385 if (entity_is_task(se
))
2386 now
= cfs_rq_clock_task(cfs_rq
);
2388 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2390 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2393 contrib_delta
= __update_entity_load_avg_contrib(se
);
2399 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2401 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2405 * Decay the load contributed by all blocked children and account this so that
2406 * their contribution may appropriately discounted when they wake up.
2408 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2410 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2413 decays
= now
- cfs_rq
->last_decay
;
2414 if (!decays
&& !force_update
)
2417 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2418 unsigned long removed_load
;
2419 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2420 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2424 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2426 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2427 cfs_rq
->last_decay
= now
;
2430 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2433 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2435 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2436 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2439 /* Add the load generated by se into cfs_rq's child load-average */
2440 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2441 struct sched_entity
*se
,
2445 * We track migrations using entity decay_count <= 0, on a wake-up
2446 * migration we use a negative decay count to track the remote decays
2447 * accumulated while sleeping.
2449 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2450 * are seen by enqueue_entity_load_avg() as a migration with an already
2451 * constructed load_avg_contrib.
2453 if (unlikely(se
->avg
.decay_count
<= 0)) {
2454 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2455 if (se
->avg
.decay_count
) {
2457 * In a wake-up migration we have to approximate the
2458 * time sleeping. This is because we can't synchronize
2459 * clock_task between the two cpus, and it is not
2460 * guaranteed to be read-safe. Instead, we can
2461 * approximate this using our carried decays, which are
2462 * explicitly atomically readable.
2464 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2466 update_entity_load_avg(se
, 0);
2467 /* Indicate that we're now synchronized and on-rq */
2468 se
->avg
.decay_count
= 0;
2472 __synchronize_entity_decay(se
);
2475 /* migrated tasks did not contribute to our blocked load */
2477 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2478 update_entity_load_avg(se
, 0);
2481 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2482 /* we force update consideration on load-balancer moves */
2483 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2487 * Remove se's load from this cfs_rq child load-average, if the entity is
2488 * transitioning to a blocked state we track its projected decay using
2491 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2492 struct sched_entity
*se
,
2495 update_entity_load_avg(se
, 1);
2496 /* we force update consideration on load-balancer moves */
2497 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2499 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2501 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2502 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2503 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2507 * Update the rq's load with the elapsed running time before entering
2508 * idle. if the last scheduled task is not a CFS task, idle_enter will
2509 * be the only way to update the runnable statistic.
2511 void idle_enter_fair(struct rq
*this_rq
)
2513 update_rq_runnable_avg(this_rq
, 1);
2517 * Update the rq's load with the elapsed idle time before a task is
2518 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2519 * be the only way to update the runnable statistic.
2521 void idle_exit_fair(struct rq
*this_rq
)
2523 update_rq_runnable_avg(this_rq
, 0);
2527 static inline void update_entity_load_avg(struct sched_entity
*se
,
2528 int update_cfs_rq
) {}
2529 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2530 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2531 struct sched_entity
*se
,
2533 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2534 struct sched_entity
*se
,
2536 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2537 int force_update
) {}
2540 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2542 #ifdef CONFIG_SCHEDSTATS
2543 struct task_struct
*tsk
= NULL
;
2545 if (entity_is_task(se
))
2548 if (se
->statistics
.sleep_start
) {
2549 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2554 if (unlikely(delta
> se
->statistics
.sleep_max
))
2555 se
->statistics
.sleep_max
= delta
;
2557 se
->statistics
.sleep_start
= 0;
2558 se
->statistics
.sum_sleep_runtime
+= delta
;
2561 account_scheduler_latency(tsk
, delta
>> 10, 1);
2562 trace_sched_stat_sleep(tsk
, delta
);
2565 if (se
->statistics
.block_start
) {
2566 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2571 if (unlikely(delta
> se
->statistics
.block_max
))
2572 se
->statistics
.block_max
= delta
;
2574 se
->statistics
.block_start
= 0;
2575 se
->statistics
.sum_sleep_runtime
+= delta
;
2578 if (tsk
->in_iowait
) {
2579 se
->statistics
.iowait_sum
+= delta
;
2580 se
->statistics
.iowait_count
++;
2581 trace_sched_stat_iowait(tsk
, delta
);
2584 trace_sched_stat_blocked(tsk
, delta
);
2587 * Blocking time is in units of nanosecs, so shift by
2588 * 20 to get a milliseconds-range estimation of the
2589 * amount of time that the task spent sleeping:
2591 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2592 profile_hits(SLEEP_PROFILING
,
2593 (void *)get_wchan(tsk
),
2596 account_scheduler_latency(tsk
, delta
>> 10, 0);
2602 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2604 #ifdef CONFIG_SCHED_DEBUG
2605 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2610 if (d
> 3*sysctl_sched_latency
)
2611 schedstat_inc(cfs_rq
, nr_spread_over
);
2616 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2618 u64 vruntime
= cfs_rq
->min_vruntime
;
2621 * The 'current' period is already promised to the current tasks,
2622 * however the extra weight of the new task will slow them down a
2623 * little, place the new task so that it fits in the slot that
2624 * stays open at the end.
2626 if (initial
&& sched_feat(START_DEBIT
))
2627 vruntime
+= sched_vslice(cfs_rq
, se
);
2629 /* sleeps up to a single latency don't count. */
2631 unsigned long thresh
= sysctl_sched_latency
;
2634 * Halve their sleep time's effect, to allow
2635 * for a gentler effect of sleepers:
2637 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2643 /* ensure we never gain time by being placed backwards. */
2644 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
2647 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
2650 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2653 * Update the normalized vruntime before updating min_vruntime
2654 * through calling update_curr().
2656 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
2657 se
->vruntime
+= cfs_rq
->min_vruntime
;
2660 * Update run-time statistics of the 'current'.
2662 update_curr(cfs_rq
);
2663 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
2664 account_entity_enqueue(cfs_rq
, se
);
2665 update_cfs_shares(cfs_rq
);
2667 if (flags
& ENQUEUE_WAKEUP
) {
2668 place_entity(cfs_rq
, se
, 0);
2669 enqueue_sleeper(cfs_rq
, se
);
2672 update_stats_enqueue(cfs_rq
, se
);
2673 check_spread(cfs_rq
, se
);
2674 if (se
!= cfs_rq
->curr
)
2675 __enqueue_entity(cfs_rq
, se
);
2678 if (cfs_rq
->nr_running
== 1) {
2679 list_add_leaf_cfs_rq(cfs_rq
);
2680 check_enqueue_throttle(cfs_rq
);
2684 static void __clear_buddies_last(struct sched_entity
*se
)
2686 for_each_sched_entity(se
) {
2687 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2688 if (cfs_rq
->last
== se
)
2689 cfs_rq
->last
= NULL
;
2695 static void __clear_buddies_next(struct sched_entity
*se
)
2697 for_each_sched_entity(se
) {
2698 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2699 if (cfs_rq
->next
== se
)
2700 cfs_rq
->next
= NULL
;
2706 static void __clear_buddies_skip(struct sched_entity
*se
)
2708 for_each_sched_entity(se
) {
2709 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2710 if (cfs_rq
->skip
== se
)
2711 cfs_rq
->skip
= NULL
;
2717 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2719 if (cfs_rq
->last
== se
)
2720 __clear_buddies_last(se
);
2722 if (cfs_rq
->next
== se
)
2723 __clear_buddies_next(se
);
2725 if (cfs_rq
->skip
== se
)
2726 __clear_buddies_skip(se
);
2729 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2732 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2735 * Update run-time statistics of the 'current'.
2737 update_curr(cfs_rq
);
2738 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
2740 update_stats_dequeue(cfs_rq
, se
);
2741 if (flags
& DEQUEUE_SLEEP
) {
2742 #ifdef CONFIG_SCHEDSTATS
2743 if (entity_is_task(se
)) {
2744 struct task_struct
*tsk
= task_of(se
);
2746 if (tsk
->state
& TASK_INTERRUPTIBLE
)
2747 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
2748 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
2749 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
2754 clear_buddies(cfs_rq
, se
);
2756 if (se
!= cfs_rq
->curr
)
2757 __dequeue_entity(cfs_rq
, se
);
2759 account_entity_dequeue(cfs_rq
, se
);
2762 * Normalize the entity after updating the min_vruntime because the
2763 * update can refer to the ->curr item and we need to reflect this
2764 * movement in our normalized position.
2766 if (!(flags
& DEQUEUE_SLEEP
))
2767 se
->vruntime
-= cfs_rq
->min_vruntime
;
2769 /* return excess runtime on last dequeue */
2770 return_cfs_rq_runtime(cfs_rq
);
2772 update_min_vruntime(cfs_rq
);
2773 update_cfs_shares(cfs_rq
);
2777 * Preempt the current task with a newly woken task if needed:
2780 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2782 unsigned long ideal_runtime
, delta_exec
;
2783 struct sched_entity
*se
;
2786 ideal_runtime
= sched_slice(cfs_rq
, curr
);
2787 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
2788 if (delta_exec
> ideal_runtime
) {
2789 resched_task(rq_of(cfs_rq
)->curr
);
2791 * The current task ran long enough, ensure it doesn't get
2792 * re-elected due to buddy favours.
2794 clear_buddies(cfs_rq
, curr
);
2799 * Ensure that a task that missed wakeup preemption by a
2800 * narrow margin doesn't have to wait for a full slice.
2801 * This also mitigates buddy induced latencies under load.
2803 if (delta_exec
< sysctl_sched_min_granularity
)
2806 se
= __pick_first_entity(cfs_rq
);
2807 delta
= curr
->vruntime
- se
->vruntime
;
2812 if (delta
> ideal_runtime
)
2813 resched_task(rq_of(cfs_rq
)->curr
);
2817 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2819 /* 'current' is not kept within the tree. */
2822 * Any task has to be enqueued before it get to execute on
2823 * a CPU. So account for the time it spent waiting on the
2826 update_stats_wait_end(cfs_rq
, se
);
2827 __dequeue_entity(cfs_rq
, se
);
2830 update_stats_curr_start(cfs_rq
, se
);
2832 #ifdef CONFIG_SCHEDSTATS
2834 * Track our maximum slice length, if the CPU's load is at
2835 * least twice that of our own weight (i.e. dont track it
2836 * when there are only lesser-weight tasks around):
2838 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
2839 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
2840 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
2843 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
2847 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
2850 * Pick the next process, keeping these things in mind, in this order:
2851 * 1) keep things fair between processes/task groups
2852 * 2) pick the "next" process, since someone really wants that to run
2853 * 3) pick the "last" process, for cache locality
2854 * 4) do not run the "skip" process, if something else is available
2856 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
2858 struct sched_entity
*se
= __pick_first_entity(cfs_rq
);
2859 struct sched_entity
*left
= se
;
2862 * Avoid running the skip buddy, if running something else can
2863 * be done without getting too unfair.
2865 if (cfs_rq
->skip
== se
) {
2866 struct sched_entity
*second
= __pick_next_entity(se
);
2867 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
2872 * Prefer last buddy, try to return the CPU to a preempted task.
2874 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
2878 * Someone really wants this to run. If it's not unfair, run it.
2880 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
2883 clear_buddies(cfs_rq
, se
);
2888 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2890 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
2893 * If still on the runqueue then deactivate_task()
2894 * was not called and update_curr() has to be done:
2897 update_curr(cfs_rq
);
2899 /* throttle cfs_rqs exceeding runtime */
2900 check_cfs_rq_runtime(cfs_rq
);
2902 check_spread(cfs_rq
, prev
);
2904 update_stats_wait_start(cfs_rq
, prev
);
2905 /* Put 'current' back into the tree. */
2906 __enqueue_entity(cfs_rq
, prev
);
2907 /* in !on_rq case, update occurred at dequeue */
2908 update_entity_load_avg(prev
, 1);
2910 cfs_rq
->curr
= NULL
;
2914 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
2917 * Update run-time statistics of the 'current'.
2919 update_curr(cfs_rq
);
2922 * Ensure that runnable average is periodically updated.
2924 update_entity_load_avg(curr
, 1);
2925 update_cfs_rq_blocked_load(cfs_rq
, 1);
2926 update_cfs_shares(cfs_rq
);
2928 #ifdef CONFIG_SCHED_HRTICK
2930 * queued ticks are scheduled to match the slice, so don't bother
2931 * validating it and just reschedule.
2934 resched_task(rq_of(cfs_rq
)->curr
);
2938 * don't let the period tick interfere with the hrtick preemption
2940 if (!sched_feat(DOUBLE_TICK
) &&
2941 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
2945 if (cfs_rq
->nr_running
> 1)
2946 check_preempt_tick(cfs_rq
, curr
);
2950 /**************************************************
2951 * CFS bandwidth control machinery
2954 #ifdef CONFIG_CFS_BANDWIDTH
2956 #ifdef HAVE_JUMP_LABEL
2957 static struct static_key __cfs_bandwidth_used
;
2959 static inline bool cfs_bandwidth_used(void)
2961 return static_key_false(&__cfs_bandwidth_used
);
2964 void cfs_bandwidth_usage_inc(void)
2966 static_key_slow_inc(&__cfs_bandwidth_used
);
2969 void cfs_bandwidth_usage_dec(void)
2971 static_key_slow_dec(&__cfs_bandwidth_used
);
2973 #else /* HAVE_JUMP_LABEL */
2974 static bool cfs_bandwidth_used(void)
2979 void cfs_bandwidth_usage_inc(void) {}
2980 void cfs_bandwidth_usage_dec(void) {}
2981 #endif /* HAVE_JUMP_LABEL */
2984 * default period for cfs group bandwidth.
2985 * default: 0.1s, units: nanoseconds
2987 static inline u64
default_cfs_period(void)
2989 return 100000000ULL;
2992 static inline u64
sched_cfs_bandwidth_slice(void)
2994 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
2998 * Replenish runtime according to assigned quota and update expiration time.
2999 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3000 * additional synchronization around rq->lock.
3002 * requires cfs_b->lock
3004 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3008 if (cfs_b
->quota
== RUNTIME_INF
)
3011 now
= sched_clock_cpu(smp_processor_id());
3012 cfs_b
->runtime
= cfs_b
->quota
;
3013 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3016 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3018 return &tg
->cfs_bandwidth
;
3021 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3022 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3024 if (unlikely(cfs_rq
->throttle_count
))
3025 return cfs_rq
->throttled_clock_task
;
3027 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3030 /* returns 0 on failure to allocate runtime */
3031 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3033 struct task_group
*tg
= cfs_rq
->tg
;
3034 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3035 u64 amount
= 0, min_amount
, expires
;
3037 /* note: this is a positive sum as runtime_remaining <= 0 */
3038 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3040 raw_spin_lock(&cfs_b
->lock
);
3041 if (cfs_b
->quota
== RUNTIME_INF
)
3042 amount
= min_amount
;
3045 * If the bandwidth pool has become inactive, then at least one
3046 * period must have elapsed since the last consumption.
3047 * Refresh the global state and ensure bandwidth timer becomes
3050 if (!cfs_b
->timer_active
) {
3051 __refill_cfs_bandwidth_runtime(cfs_b
);
3052 __start_cfs_bandwidth(cfs_b
);
3055 if (cfs_b
->runtime
> 0) {
3056 amount
= min(cfs_b
->runtime
, min_amount
);
3057 cfs_b
->runtime
-= amount
;
3061 expires
= cfs_b
->runtime_expires
;
3062 raw_spin_unlock(&cfs_b
->lock
);
3064 cfs_rq
->runtime_remaining
+= amount
;
3066 * we may have advanced our local expiration to account for allowed
3067 * spread between our sched_clock and the one on which runtime was
3070 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3071 cfs_rq
->runtime_expires
= expires
;
3073 return cfs_rq
->runtime_remaining
> 0;
3077 * Note: This depends on the synchronization provided by sched_clock and the
3078 * fact that rq->clock snapshots this value.
3080 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3082 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3084 /* if the deadline is ahead of our clock, nothing to do */
3085 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3088 if (cfs_rq
->runtime_remaining
< 0)
3092 * If the local deadline has passed we have to consider the
3093 * possibility that our sched_clock is 'fast' and the global deadline
3094 * has not truly expired.
3096 * Fortunately we can check determine whether this the case by checking
3097 * whether the global deadline has advanced.
3100 if ((s64
)(cfs_rq
->runtime_expires
- cfs_b
->runtime_expires
) >= 0) {
3101 /* extend local deadline, drift is bounded above by 2 ticks */
3102 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3104 /* global deadline is ahead, expiration has passed */
3105 cfs_rq
->runtime_remaining
= 0;
3109 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3111 /* dock delta_exec before expiring quota (as it could span periods) */
3112 cfs_rq
->runtime_remaining
-= delta_exec
;
3113 expire_cfs_rq_runtime(cfs_rq
);
3115 if (likely(cfs_rq
->runtime_remaining
> 0))
3119 * if we're unable to extend our runtime we resched so that the active
3120 * hierarchy can be throttled
3122 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3123 resched_task(rq_of(cfs_rq
)->curr
);
3126 static __always_inline
3127 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3129 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3132 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3135 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3137 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3140 /* check whether cfs_rq, or any parent, is throttled */
3141 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3143 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3147 * Ensure that neither of the group entities corresponding to src_cpu or
3148 * dest_cpu are members of a throttled hierarchy when performing group
3149 * load-balance operations.
3151 static inline int throttled_lb_pair(struct task_group
*tg
,
3152 int src_cpu
, int dest_cpu
)
3154 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3156 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3157 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3159 return throttled_hierarchy(src_cfs_rq
) ||
3160 throttled_hierarchy(dest_cfs_rq
);
3163 /* updated child weight may affect parent so we have to do this bottom up */
3164 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3166 struct rq
*rq
= data
;
3167 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3169 cfs_rq
->throttle_count
--;
3171 if (!cfs_rq
->throttle_count
) {
3172 /* adjust cfs_rq_clock_task() */
3173 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3174 cfs_rq
->throttled_clock_task
;
3181 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3183 struct rq
*rq
= data
;
3184 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3186 /* group is entering throttled state, stop time */
3187 if (!cfs_rq
->throttle_count
)
3188 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3189 cfs_rq
->throttle_count
++;
3194 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3196 struct rq
*rq
= rq_of(cfs_rq
);
3197 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3198 struct sched_entity
*se
;
3199 long task_delta
, dequeue
= 1;
3201 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3203 /* freeze hierarchy runnable averages while throttled */
3205 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3208 task_delta
= cfs_rq
->h_nr_running
;
3209 for_each_sched_entity(se
) {
3210 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3211 /* throttled entity or throttle-on-deactivate */
3216 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3217 qcfs_rq
->h_nr_running
-= task_delta
;
3219 if (qcfs_rq
->load
.weight
)
3224 rq
->nr_running
-= task_delta
;
3226 cfs_rq
->throttled
= 1;
3227 cfs_rq
->throttled_clock
= rq_clock(rq
);
3228 raw_spin_lock(&cfs_b
->lock
);
3229 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3230 if (!cfs_b
->timer_active
)
3231 __start_cfs_bandwidth(cfs_b
);
3232 raw_spin_unlock(&cfs_b
->lock
);
3235 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3237 struct rq
*rq
= rq_of(cfs_rq
);
3238 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3239 struct sched_entity
*se
;
3243 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3245 cfs_rq
->throttled
= 0;
3247 update_rq_clock(rq
);
3249 raw_spin_lock(&cfs_b
->lock
);
3250 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3251 list_del_rcu(&cfs_rq
->throttled_list
);
3252 raw_spin_unlock(&cfs_b
->lock
);
3254 /* update hierarchical throttle state */
3255 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3257 if (!cfs_rq
->load
.weight
)
3260 task_delta
= cfs_rq
->h_nr_running
;
3261 for_each_sched_entity(se
) {
3265 cfs_rq
= cfs_rq_of(se
);
3267 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3268 cfs_rq
->h_nr_running
+= task_delta
;
3270 if (cfs_rq_throttled(cfs_rq
))
3275 rq
->nr_running
+= task_delta
;
3277 /* determine whether we need to wake up potentially idle cpu */
3278 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3279 resched_task(rq
->curr
);
3282 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3283 u64 remaining
, u64 expires
)
3285 struct cfs_rq
*cfs_rq
;
3286 u64 runtime
= remaining
;
3289 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3291 struct rq
*rq
= rq_of(cfs_rq
);
3293 raw_spin_lock(&rq
->lock
);
3294 if (!cfs_rq_throttled(cfs_rq
))
3297 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3298 if (runtime
> remaining
)
3299 runtime
= remaining
;
3300 remaining
-= runtime
;
3302 cfs_rq
->runtime_remaining
+= runtime
;
3303 cfs_rq
->runtime_expires
= expires
;
3305 /* we check whether we're throttled above */
3306 if (cfs_rq
->runtime_remaining
> 0)
3307 unthrottle_cfs_rq(cfs_rq
);
3310 raw_spin_unlock(&rq
->lock
);
3321 * Responsible for refilling a task_group's bandwidth and unthrottling its
3322 * cfs_rqs as appropriate. If there has been no activity within the last
3323 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3324 * used to track this state.
3326 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3328 u64 runtime
, runtime_expires
;
3329 int idle
= 1, throttled
;
3331 raw_spin_lock(&cfs_b
->lock
);
3332 /* no need to continue the timer with no bandwidth constraint */
3333 if (cfs_b
->quota
== RUNTIME_INF
)
3336 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3337 /* idle depends on !throttled (for the case of a large deficit) */
3338 idle
= cfs_b
->idle
&& !throttled
;
3339 cfs_b
->nr_periods
+= overrun
;
3341 /* if we're going inactive then everything else can be deferred */
3346 * if we have relooped after returning idle once, we need to update our
3347 * status as actually running, so that other cpus doing
3348 * __start_cfs_bandwidth will stop trying to cancel us.
3350 cfs_b
->timer_active
= 1;
3352 __refill_cfs_bandwidth_runtime(cfs_b
);
3355 /* mark as potentially idle for the upcoming period */
3360 /* account preceding periods in which throttling occurred */
3361 cfs_b
->nr_throttled
+= overrun
;
3364 * There are throttled entities so we must first use the new bandwidth
3365 * to unthrottle them before making it generally available. This
3366 * ensures that all existing debts will be paid before a new cfs_rq is
3369 runtime
= cfs_b
->runtime
;
3370 runtime_expires
= cfs_b
->runtime_expires
;
3374 * This check is repeated as we are holding onto the new bandwidth
3375 * while we unthrottle. This can potentially race with an unthrottled
3376 * group trying to acquire new bandwidth from the global pool.
3378 while (throttled
&& runtime
> 0) {
3379 raw_spin_unlock(&cfs_b
->lock
);
3380 /* we can't nest cfs_b->lock while distributing bandwidth */
3381 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3383 raw_spin_lock(&cfs_b
->lock
);
3385 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3388 /* return (any) remaining runtime */
3389 cfs_b
->runtime
= runtime
;
3391 * While we are ensured activity in the period following an
3392 * unthrottle, this also covers the case in which the new bandwidth is
3393 * insufficient to cover the existing bandwidth deficit. (Forcing the
3394 * timer to remain active while there are any throttled entities.)
3399 cfs_b
->timer_active
= 0;
3400 raw_spin_unlock(&cfs_b
->lock
);
3405 /* a cfs_rq won't donate quota below this amount */
3406 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3407 /* minimum remaining period time to redistribute slack quota */
3408 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3409 /* how long we wait to gather additional slack before distributing */
3410 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3413 * Are we near the end of the current quota period?
3415 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3416 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3417 * migrate_hrtimers, base is never cleared, so we are fine.
3419 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3421 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3424 /* if the call-back is running a quota refresh is already occurring */
3425 if (hrtimer_callback_running(refresh_timer
))
3428 /* is a quota refresh about to occur? */
3429 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3430 if (remaining
< min_expire
)
3436 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3438 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3440 /* if there's a quota refresh soon don't bother with slack */
3441 if (runtime_refresh_within(cfs_b
, min_left
))
3444 start_bandwidth_timer(&cfs_b
->slack_timer
,
3445 ns_to_ktime(cfs_bandwidth_slack_period
));
3448 /* we know any runtime found here is valid as update_curr() precedes return */
3449 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3451 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3452 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3454 if (slack_runtime
<= 0)
3457 raw_spin_lock(&cfs_b
->lock
);
3458 if (cfs_b
->quota
!= RUNTIME_INF
&&
3459 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3460 cfs_b
->runtime
+= slack_runtime
;
3462 /* we are under rq->lock, defer unthrottling using a timer */
3463 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3464 !list_empty(&cfs_b
->throttled_cfs_rq
))
3465 start_cfs_slack_bandwidth(cfs_b
);
3467 raw_spin_unlock(&cfs_b
->lock
);
3469 /* even if it's not valid for return we don't want to try again */
3470 cfs_rq
->runtime_remaining
-= slack_runtime
;
3473 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3475 if (!cfs_bandwidth_used())
3478 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3481 __return_cfs_rq_runtime(cfs_rq
);
3485 * This is done with a timer (instead of inline with bandwidth return) since
3486 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3488 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3490 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3493 /* confirm we're still not at a refresh boundary */
3494 raw_spin_lock(&cfs_b
->lock
);
3495 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
3496 raw_spin_unlock(&cfs_b
->lock
);
3500 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
) {
3501 runtime
= cfs_b
->runtime
;
3504 expires
= cfs_b
->runtime_expires
;
3505 raw_spin_unlock(&cfs_b
->lock
);
3510 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3512 raw_spin_lock(&cfs_b
->lock
);
3513 if (expires
== cfs_b
->runtime_expires
)
3514 cfs_b
->runtime
= runtime
;
3515 raw_spin_unlock(&cfs_b
->lock
);
3519 * When a group wakes up we want to make sure that its quota is not already
3520 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3521 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3523 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3525 if (!cfs_bandwidth_used())
3528 /* an active group must be handled by the update_curr()->put() path */
3529 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3532 /* ensure the group is not already throttled */
3533 if (cfs_rq_throttled(cfs_rq
))
3536 /* update runtime allocation */
3537 account_cfs_rq_runtime(cfs_rq
, 0);
3538 if (cfs_rq
->runtime_remaining
<= 0)
3539 throttle_cfs_rq(cfs_rq
);
3542 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3543 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3545 if (!cfs_bandwidth_used())
3548 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3552 * it's possible for a throttled entity to be forced into a running
3553 * state (e.g. set_curr_task), in this case we're finished.
3555 if (cfs_rq_throttled(cfs_rq
))
3558 throttle_cfs_rq(cfs_rq
);
3561 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3563 struct cfs_bandwidth
*cfs_b
=
3564 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3565 do_sched_cfs_slack_timer(cfs_b
);
3567 return HRTIMER_NORESTART
;
3570 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3572 struct cfs_bandwidth
*cfs_b
=
3573 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3579 now
= hrtimer_cb_get_time(timer
);
3580 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3585 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3588 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3591 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3593 raw_spin_lock_init(&cfs_b
->lock
);
3595 cfs_b
->quota
= RUNTIME_INF
;
3596 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3598 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3599 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3600 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3601 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3602 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3605 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3607 cfs_rq
->runtime_enabled
= 0;
3608 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3611 /* requires cfs_b->lock, may release to reprogram timer */
3612 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3615 * The timer may be active because we're trying to set a new bandwidth
3616 * period or because we're racing with the tear-down path
3617 * (timer_active==0 becomes visible before the hrtimer call-back
3618 * terminates). In either case we ensure that it's re-programmed
3620 while (unlikely(hrtimer_active(&cfs_b
->period_timer
)) &&
3621 hrtimer_try_to_cancel(&cfs_b
->period_timer
) < 0) {
3622 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3623 raw_spin_unlock(&cfs_b
->lock
);
3625 raw_spin_lock(&cfs_b
->lock
);
3626 /* if someone else restarted the timer then we're done */
3627 if (cfs_b
->timer_active
)
3631 cfs_b
->timer_active
= 1;
3632 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
3635 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3637 hrtimer_cancel(&cfs_b
->period_timer
);
3638 hrtimer_cancel(&cfs_b
->slack_timer
);
3641 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
3643 struct cfs_rq
*cfs_rq
;
3645 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3646 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3648 if (!cfs_rq
->runtime_enabled
)
3652 * clock_task is not advancing so we just need to make sure
3653 * there's some valid quota amount
3655 cfs_rq
->runtime_remaining
= cfs_b
->quota
;
3656 if (cfs_rq_throttled(cfs_rq
))
3657 unthrottle_cfs_rq(cfs_rq
);
3661 #else /* CONFIG_CFS_BANDWIDTH */
3662 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3664 return rq_clock_task(rq_of(cfs_rq
));
3667 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
3668 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3669 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
3670 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3672 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3677 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3682 static inline int throttled_lb_pair(struct task_group
*tg
,
3683 int src_cpu
, int dest_cpu
)
3688 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3690 #ifdef CONFIG_FAIR_GROUP_SCHED
3691 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3694 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3698 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3699 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
3701 #endif /* CONFIG_CFS_BANDWIDTH */
3703 /**************************************************
3704 * CFS operations on tasks:
3707 #ifdef CONFIG_SCHED_HRTICK
3708 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3710 struct sched_entity
*se
= &p
->se
;
3711 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3713 WARN_ON(task_rq(p
) != rq
);
3715 if (cfs_rq
->nr_running
> 1) {
3716 u64 slice
= sched_slice(cfs_rq
, se
);
3717 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
3718 s64 delta
= slice
- ran
;
3727 * Don't schedule slices shorter than 10000ns, that just
3728 * doesn't make sense. Rely on vruntime for fairness.
3731 delta
= max_t(s64
, 10000LL, delta
);
3733 hrtick_start(rq
, delta
);
3738 * called from enqueue/dequeue and updates the hrtick when the
3739 * current task is from our class and nr_running is low enough
3742 static void hrtick_update(struct rq
*rq
)
3744 struct task_struct
*curr
= rq
->curr
;
3746 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
3749 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
3750 hrtick_start_fair(rq
, curr
);
3752 #else /* !CONFIG_SCHED_HRTICK */
3754 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3758 static inline void hrtick_update(struct rq
*rq
)
3764 * The enqueue_task method is called before nr_running is
3765 * increased. Here we update the fair scheduling stats and
3766 * then put the task into the rbtree:
3769 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3771 struct cfs_rq
*cfs_rq
;
3772 struct sched_entity
*se
= &p
->se
;
3774 for_each_sched_entity(se
) {
3777 cfs_rq
= cfs_rq_of(se
);
3778 enqueue_entity(cfs_rq
, se
, flags
);
3781 * end evaluation on encountering a throttled cfs_rq
3783 * note: in the case of encountering a throttled cfs_rq we will
3784 * post the final h_nr_running increment below.
3786 if (cfs_rq_throttled(cfs_rq
))
3788 cfs_rq
->h_nr_running
++;
3790 flags
= ENQUEUE_WAKEUP
;
3793 for_each_sched_entity(se
) {
3794 cfs_rq
= cfs_rq_of(se
);
3795 cfs_rq
->h_nr_running
++;
3797 if (cfs_rq_throttled(cfs_rq
))
3800 update_cfs_shares(cfs_rq
);
3801 update_entity_load_avg(se
, 1);
3805 update_rq_runnable_avg(rq
, rq
->nr_running
);
3811 static void set_next_buddy(struct sched_entity
*se
);
3814 * The dequeue_task method is called before nr_running is
3815 * decreased. We remove the task from the rbtree and
3816 * update the fair scheduling stats:
3818 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3820 struct cfs_rq
*cfs_rq
;
3821 struct sched_entity
*se
= &p
->se
;
3822 int task_sleep
= flags
& DEQUEUE_SLEEP
;
3824 for_each_sched_entity(se
) {
3825 cfs_rq
= cfs_rq_of(se
);
3826 dequeue_entity(cfs_rq
, se
, flags
);
3829 * end evaluation on encountering a throttled cfs_rq
3831 * note: in the case of encountering a throttled cfs_rq we will
3832 * post the final h_nr_running decrement below.
3834 if (cfs_rq_throttled(cfs_rq
))
3836 cfs_rq
->h_nr_running
--;
3838 /* Don't dequeue parent if it has other entities besides us */
3839 if (cfs_rq
->load
.weight
) {
3841 * Bias pick_next to pick a task from this cfs_rq, as
3842 * p is sleeping when it is within its sched_slice.
3844 if (task_sleep
&& parent_entity(se
))
3845 set_next_buddy(parent_entity(se
));
3847 /* avoid re-evaluating load for this entity */
3848 se
= parent_entity(se
);
3851 flags
|= DEQUEUE_SLEEP
;
3854 for_each_sched_entity(se
) {
3855 cfs_rq
= cfs_rq_of(se
);
3856 cfs_rq
->h_nr_running
--;
3858 if (cfs_rq_throttled(cfs_rq
))
3861 update_cfs_shares(cfs_rq
);
3862 update_entity_load_avg(se
, 1);
3867 update_rq_runnable_avg(rq
, 1);
3873 /* Used instead of source_load when we know the type == 0 */
3874 static unsigned long weighted_cpuload(const int cpu
)
3876 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
3880 * Return a low guess at the load of a migration-source cpu weighted
3881 * according to the scheduling class and "nice" value.
3883 * We want to under-estimate the load of migration sources, to
3884 * balance conservatively.
3886 static unsigned long source_load(int cpu
, int type
)
3888 struct rq
*rq
= cpu_rq(cpu
);
3889 unsigned long total
= weighted_cpuload(cpu
);
3891 if (type
== 0 || !sched_feat(LB_BIAS
))
3894 return min(rq
->cpu_load
[type
-1], total
);
3898 * Return a high guess at the load of a migration-target cpu weighted
3899 * according to the scheduling class and "nice" value.
3901 static unsigned long target_load(int cpu
, int type
)
3903 struct rq
*rq
= cpu_rq(cpu
);
3904 unsigned long total
= weighted_cpuload(cpu
);
3906 if (type
== 0 || !sched_feat(LB_BIAS
))
3909 return max(rq
->cpu_load
[type
-1], total
);
3912 static unsigned long power_of(int cpu
)
3914 return cpu_rq(cpu
)->cpu_power
;
3917 static unsigned long cpu_avg_load_per_task(int cpu
)
3919 struct rq
*rq
= cpu_rq(cpu
);
3920 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
3921 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
3924 return load_avg
/ nr_running
;
3929 static void record_wakee(struct task_struct
*p
)
3932 * Rough decay (wiping) for cost saving, don't worry
3933 * about the boundary, really active task won't care
3936 if (jiffies
> current
->wakee_flip_decay_ts
+ HZ
) {
3937 current
->wakee_flips
= 0;
3938 current
->wakee_flip_decay_ts
= jiffies
;
3941 if (current
->last_wakee
!= p
) {
3942 current
->last_wakee
= p
;
3943 current
->wakee_flips
++;
3947 static void task_waking_fair(struct task_struct
*p
)
3949 struct sched_entity
*se
= &p
->se
;
3950 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3953 #ifndef CONFIG_64BIT
3954 u64 min_vruntime_copy
;
3957 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
3959 min_vruntime
= cfs_rq
->min_vruntime
;
3960 } while (min_vruntime
!= min_vruntime_copy
);
3962 min_vruntime
= cfs_rq
->min_vruntime
;
3965 se
->vruntime
-= min_vruntime
;
3969 #ifdef CONFIG_FAIR_GROUP_SCHED
3971 * effective_load() calculates the load change as seen from the root_task_group
3973 * Adding load to a group doesn't make a group heavier, but can cause movement
3974 * of group shares between cpus. Assuming the shares were perfectly aligned one
3975 * can calculate the shift in shares.
3977 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3978 * on this @cpu and results in a total addition (subtraction) of @wg to the
3979 * total group weight.
3981 * Given a runqueue weight distribution (rw_i) we can compute a shares
3982 * distribution (s_i) using:
3984 * s_i = rw_i / \Sum rw_j (1)
3986 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3987 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3988 * shares distribution (s_i):
3990 * rw_i = { 2, 4, 1, 0 }
3991 * s_i = { 2/7, 4/7, 1/7, 0 }
3993 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3994 * task used to run on and the CPU the waker is running on), we need to
3995 * compute the effect of waking a task on either CPU and, in case of a sync
3996 * wakeup, compute the effect of the current task going to sleep.
3998 * So for a change of @wl to the local @cpu with an overall group weight change
3999 * of @wl we can compute the new shares distribution (s'_i) using:
4001 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4003 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4004 * differences in waking a task to CPU 0. The additional task changes the
4005 * weight and shares distributions like:
4007 * rw'_i = { 3, 4, 1, 0 }
4008 * s'_i = { 3/8, 4/8, 1/8, 0 }
4010 * We can then compute the difference in effective weight by using:
4012 * dw_i = S * (s'_i - s_i) (3)
4014 * Where 'S' is the group weight as seen by its parent.
4016 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4017 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4018 * 4/7) times the weight of the group.
4020 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4022 struct sched_entity
*se
= tg
->se
[cpu
];
4024 if (!tg
->parent
) /* the trivial, non-cgroup case */
4027 for_each_sched_entity(se
) {
4033 * W = @wg + \Sum rw_j
4035 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4040 w
= se
->my_q
->load
.weight
+ wl
;
4043 * wl = S * s'_i; see (2)
4046 wl
= (w
* tg
->shares
) / W
;
4051 * Per the above, wl is the new se->load.weight value; since
4052 * those are clipped to [MIN_SHARES, ...) do so now. See
4053 * calc_cfs_shares().
4055 if (wl
< MIN_SHARES
)
4059 * wl = dw_i = S * (s'_i - s_i); see (3)
4061 wl
-= se
->load
.weight
;
4064 * Recursively apply this logic to all parent groups to compute
4065 * the final effective load change on the root group. Since
4066 * only the @tg group gets extra weight, all parent groups can
4067 * only redistribute existing shares. @wl is the shift in shares
4068 * resulting from this level per the above.
4077 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4084 static int wake_wide(struct task_struct
*p
)
4086 int factor
= this_cpu_read(sd_llc_size
);
4089 * Yeah, it's the switching-frequency, could means many wakee or
4090 * rapidly switch, use factor here will just help to automatically
4091 * adjust the loose-degree, so bigger node will lead to more pull.
4093 if (p
->wakee_flips
> factor
) {
4095 * wakee is somewhat hot, it needs certain amount of cpu
4096 * resource, so if waker is far more hot, prefer to leave
4099 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
4106 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4108 s64 this_load
, load
;
4109 int idx
, this_cpu
, prev_cpu
;
4110 unsigned long tl_per_task
;
4111 struct task_group
*tg
;
4112 unsigned long weight
;
4116 * If we wake multiple tasks be careful to not bounce
4117 * ourselves around too much.
4123 this_cpu
= smp_processor_id();
4124 prev_cpu
= task_cpu(p
);
4125 load
= source_load(prev_cpu
, idx
);
4126 this_load
= target_load(this_cpu
, idx
);
4129 * If sync wakeup then subtract the (maximum possible)
4130 * effect of the currently running task from the load
4131 * of the current CPU:
4134 tg
= task_group(current
);
4135 weight
= current
->se
.load
.weight
;
4137 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4138 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4142 weight
= p
->se
.load
.weight
;
4145 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4146 * due to the sync cause above having dropped this_load to 0, we'll
4147 * always have an imbalance, but there's really nothing you can do
4148 * about that, so that's good too.
4150 * Otherwise check if either cpus are near enough in load to allow this
4151 * task to be woken on this_cpu.
4153 if (this_load
> 0) {
4154 s64 this_eff_load
, prev_eff_load
;
4156 this_eff_load
= 100;
4157 this_eff_load
*= power_of(prev_cpu
);
4158 this_eff_load
*= this_load
+
4159 effective_load(tg
, this_cpu
, weight
, weight
);
4161 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4162 prev_eff_load
*= power_of(this_cpu
);
4163 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4165 balanced
= this_eff_load
<= prev_eff_load
;
4170 * If the currently running task will sleep within
4171 * a reasonable amount of time then attract this newly
4174 if (sync
&& balanced
)
4177 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4178 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
4181 (this_load
<= load
&&
4182 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
4184 * This domain has SD_WAKE_AFFINE and
4185 * p is cache cold in this domain, and
4186 * there is no bad imbalance.
4188 schedstat_inc(sd
, ttwu_move_affine
);
4189 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4197 * find_idlest_group finds and returns the least busy CPU group within the
4200 static struct sched_group
*
4201 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4202 int this_cpu
, int sd_flag
)
4204 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4205 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4206 int load_idx
= sd
->forkexec_idx
;
4207 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4209 if (sd_flag
& SD_BALANCE_WAKE
)
4210 load_idx
= sd
->wake_idx
;
4213 unsigned long load
, avg_load
;
4217 /* Skip over this group if it has no CPUs allowed */
4218 if (!cpumask_intersects(sched_group_cpus(group
),
4219 tsk_cpus_allowed(p
)))
4222 local_group
= cpumask_test_cpu(this_cpu
,
4223 sched_group_cpus(group
));
4225 /* Tally up the load of all CPUs in the group */
4228 for_each_cpu(i
, sched_group_cpus(group
)) {
4229 /* Bias balancing toward cpus of our domain */
4231 load
= source_load(i
, load_idx
);
4233 load
= target_load(i
, load_idx
);
4238 /* Adjust by relative CPU power of the group */
4239 avg_load
= (avg_load
* SCHED_POWER_SCALE
) / group
->sgp
->power
;
4242 this_load
= avg_load
;
4243 } else if (avg_load
< min_load
) {
4244 min_load
= avg_load
;
4247 } while (group
= group
->next
, group
!= sd
->groups
);
4249 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4255 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4258 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4260 unsigned long load
, min_load
= ULONG_MAX
;
4264 /* Traverse only the allowed CPUs */
4265 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4266 load
= weighted_cpuload(i
);
4268 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4278 * Try and locate an idle CPU in the sched_domain.
4280 static int select_idle_sibling(struct task_struct
*p
, int target
)
4282 struct sched_domain
*sd
;
4283 struct sched_group
*sg
;
4284 int i
= task_cpu(p
);
4286 if (idle_cpu(target
))
4290 * If the prevous cpu is cache affine and idle, don't be stupid.
4292 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4296 * Otherwise, iterate the domains and find an elegible idle cpu.
4298 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4299 for_each_lower_domain(sd
) {
4302 if (!cpumask_intersects(sched_group_cpus(sg
),
4303 tsk_cpus_allowed(p
)))
4306 for_each_cpu(i
, sched_group_cpus(sg
)) {
4307 if (i
== target
|| !idle_cpu(i
))
4311 target
= cpumask_first_and(sched_group_cpus(sg
),
4312 tsk_cpus_allowed(p
));
4316 } while (sg
!= sd
->groups
);
4323 * sched_balance_self: balance the current task (running on cpu) in domains
4324 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4327 * Balance, ie. select the least loaded group.
4329 * Returns the target CPU number, or the same CPU if no balancing is needed.
4331 * preempt must be disabled.
4334 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4336 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4337 int cpu
= smp_processor_id();
4339 int want_affine
= 0;
4340 int sync
= wake_flags
& WF_SYNC
;
4342 if (p
->nr_cpus_allowed
== 1)
4345 if (sd_flag
& SD_BALANCE_WAKE
) {
4346 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
4352 for_each_domain(cpu
, tmp
) {
4353 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4357 * If both cpu and prev_cpu are part of this domain,
4358 * cpu is a valid SD_WAKE_AFFINE target.
4360 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4361 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4366 if (tmp
->flags
& sd_flag
)
4371 if (cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4374 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4379 struct sched_group
*group
;
4382 if (!(sd
->flags
& sd_flag
)) {
4387 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
4393 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4394 if (new_cpu
== -1 || new_cpu
== cpu
) {
4395 /* Now try balancing at a lower domain level of cpu */
4400 /* Now try balancing at a lower domain level of new_cpu */
4402 weight
= sd
->span_weight
;
4404 for_each_domain(cpu
, tmp
) {
4405 if (weight
<= tmp
->span_weight
)
4407 if (tmp
->flags
& sd_flag
)
4410 /* while loop will break here if sd == NULL */
4419 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4420 * cfs_rq_of(p) references at time of call are still valid and identify the
4421 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4422 * other assumptions, including the state of rq->lock, should be made.
4425 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4427 struct sched_entity
*se
= &p
->se
;
4428 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4431 * Load tracking: accumulate removed load so that it can be processed
4432 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4433 * to blocked load iff they have a positive decay-count. It can never
4434 * be negative here since on-rq tasks have decay-count == 0.
4436 if (se
->avg
.decay_count
) {
4437 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4438 atomic_long_add(se
->avg
.load_avg_contrib
,
4439 &cfs_rq
->removed_load
);
4442 #endif /* CONFIG_SMP */
4444 static unsigned long
4445 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4447 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4450 * Since its curr running now, convert the gran from real-time
4451 * to virtual-time in his units.
4453 * By using 'se' instead of 'curr' we penalize light tasks, so
4454 * they get preempted easier. That is, if 'se' < 'curr' then
4455 * the resulting gran will be larger, therefore penalizing the
4456 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4457 * be smaller, again penalizing the lighter task.
4459 * This is especially important for buddies when the leftmost
4460 * task is higher priority than the buddy.
4462 return calc_delta_fair(gran
, se
);
4466 * Should 'se' preempt 'curr'.
4480 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4482 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4487 gran
= wakeup_gran(curr
, se
);
4494 static void set_last_buddy(struct sched_entity
*se
)
4496 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4499 for_each_sched_entity(se
)
4500 cfs_rq_of(se
)->last
= se
;
4503 static void set_next_buddy(struct sched_entity
*se
)
4505 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4508 for_each_sched_entity(se
)
4509 cfs_rq_of(se
)->next
= se
;
4512 static void set_skip_buddy(struct sched_entity
*se
)
4514 for_each_sched_entity(se
)
4515 cfs_rq_of(se
)->skip
= se
;
4519 * Preempt the current task with a newly woken task if needed:
4521 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4523 struct task_struct
*curr
= rq
->curr
;
4524 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4525 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4526 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4527 int next_buddy_marked
= 0;
4529 if (unlikely(se
== pse
))
4533 * This is possible from callers such as move_task(), in which we
4534 * unconditionally check_prempt_curr() after an enqueue (which may have
4535 * lead to a throttle). This both saves work and prevents false
4536 * next-buddy nomination below.
4538 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4541 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4542 set_next_buddy(pse
);
4543 next_buddy_marked
= 1;
4547 * We can come here with TIF_NEED_RESCHED already set from new task
4550 * Note: this also catches the edge-case of curr being in a throttled
4551 * group (e.g. via set_curr_task), since update_curr() (in the
4552 * enqueue of curr) will have resulted in resched being set. This
4553 * prevents us from potentially nominating it as a false LAST_BUDDY
4556 if (test_tsk_need_resched(curr
))
4559 /* Idle tasks are by definition preempted by non-idle tasks. */
4560 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4561 likely(p
->policy
!= SCHED_IDLE
))
4565 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4566 * is driven by the tick):
4568 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4571 find_matching_se(&se
, &pse
);
4572 update_curr(cfs_rq_of(se
));
4574 if (wakeup_preempt_entity(se
, pse
) == 1) {
4576 * Bias pick_next to pick the sched entity that is
4577 * triggering this preemption.
4579 if (!next_buddy_marked
)
4580 set_next_buddy(pse
);
4589 * Only set the backward buddy when the current task is still
4590 * on the rq. This can happen when a wakeup gets interleaved
4591 * with schedule on the ->pre_schedule() or idle_balance()
4592 * point, either of which can * drop the rq lock.
4594 * Also, during early boot the idle thread is in the fair class,
4595 * for obvious reasons its a bad idea to schedule back to it.
4597 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4600 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
4604 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
4606 struct task_struct
*p
;
4607 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
4608 struct sched_entity
*se
;
4610 if (!cfs_rq
->nr_running
)
4614 se
= pick_next_entity(cfs_rq
);
4615 set_next_entity(cfs_rq
, se
);
4616 cfs_rq
= group_cfs_rq(se
);
4620 if (hrtick_enabled(rq
))
4621 hrtick_start_fair(rq
, p
);
4627 * Account for a descheduled task:
4629 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4631 struct sched_entity
*se
= &prev
->se
;
4632 struct cfs_rq
*cfs_rq
;
4634 for_each_sched_entity(se
) {
4635 cfs_rq
= cfs_rq_of(se
);
4636 put_prev_entity(cfs_rq
, se
);
4641 * sched_yield() is very simple
4643 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4645 static void yield_task_fair(struct rq
*rq
)
4647 struct task_struct
*curr
= rq
->curr
;
4648 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4649 struct sched_entity
*se
= &curr
->se
;
4652 * Are we the only task in the tree?
4654 if (unlikely(rq
->nr_running
== 1))
4657 clear_buddies(cfs_rq
, se
);
4659 if (curr
->policy
!= SCHED_BATCH
) {
4660 update_rq_clock(rq
);
4662 * Update run-time statistics of the 'current'.
4664 update_curr(cfs_rq
);
4666 * Tell update_rq_clock() that we've just updated,
4667 * so we don't do microscopic update in schedule()
4668 * and double the fastpath cost.
4670 rq
->skip_clock_update
= 1;
4676 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
4678 struct sched_entity
*se
= &p
->se
;
4680 /* throttled hierarchies are not runnable */
4681 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
4684 /* Tell the scheduler that we'd really like pse to run next. */
4687 yield_task_fair(rq
);
4693 /**************************************************
4694 * Fair scheduling class load-balancing methods.
4698 * The purpose of load-balancing is to achieve the same basic fairness the
4699 * per-cpu scheduler provides, namely provide a proportional amount of compute
4700 * time to each task. This is expressed in the following equation:
4702 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4704 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4705 * W_i,0 is defined as:
4707 * W_i,0 = \Sum_j w_i,j (2)
4709 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4710 * is derived from the nice value as per prio_to_weight[].
4712 * The weight average is an exponential decay average of the instantaneous
4715 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4717 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4718 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4719 * can also include other factors [XXX].
4721 * To achieve this balance we define a measure of imbalance which follows
4722 * directly from (1):
4724 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4726 * We them move tasks around to minimize the imbalance. In the continuous
4727 * function space it is obvious this converges, in the discrete case we get
4728 * a few fun cases generally called infeasible weight scenarios.
4731 * - infeasible weights;
4732 * - local vs global optima in the discrete case. ]
4737 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4738 * for all i,j solution, we create a tree of cpus that follows the hardware
4739 * topology where each level pairs two lower groups (or better). This results
4740 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4741 * tree to only the first of the previous level and we decrease the frequency
4742 * of load-balance at each level inv. proportional to the number of cpus in
4748 * \Sum { --- * --- * 2^i } = O(n) (5)
4750 * `- size of each group
4751 * | | `- number of cpus doing load-balance
4753 * `- sum over all levels
4755 * Coupled with a limit on how many tasks we can migrate every balance pass,
4756 * this makes (5) the runtime complexity of the balancer.
4758 * An important property here is that each CPU is still (indirectly) connected
4759 * to every other cpu in at most O(log n) steps:
4761 * The adjacency matrix of the resulting graph is given by:
4764 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4767 * And you'll find that:
4769 * A^(log_2 n)_i,j != 0 for all i,j (7)
4771 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4772 * The task movement gives a factor of O(m), giving a convergence complexity
4775 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4780 * In order to avoid CPUs going idle while there's still work to do, new idle
4781 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4782 * tree itself instead of relying on other CPUs to bring it work.
4784 * This adds some complexity to both (5) and (8) but it reduces the total idle
4792 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4795 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4800 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4802 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4804 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4807 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4808 * rewrite all of this once again.]
4811 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
4813 enum fbq_type
{ regular
, remote
, all
};
4815 #define LBF_ALL_PINNED 0x01
4816 #define LBF_NEED_BREAK 0x02
4817 #define LBF_DST_PINNED 0x04
4818 #define LBF_SOME_PINNED 0x08
4821 struct sched_domain
*sd
;
4829 struct cpumask
*dst_grpmask
;
4831 enum cpu_idle_type idle
;
4833 /* The set of CPUs under consideration for load-balancing */
4834 struct cpumask
*cpus
;
4839 unsigned int loop_break
;
4840 unsigned int loop_max
;
4842 enum fbq_type fbq_type
;
4846 * move_task - move a task from one runqueue to another runqueue.
4847 * Both runqueues must be locked.
4849 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
4851 deactivate_task(env
->src_rq
, p
, 0);
4852 set_task_cpu(p
, env
->dst_cpu
);
4853 activate_task(env
->dst_rq
, p
, 0);
4854 check_preempt_curr(env
->dst_rq
, p
, 0);
4858 * Is this task likely cache-hot:
4861 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
4865 if (p
->sched_class
!= &fair_sched_class
)
4868 if (unlikely(p
->policy
== SCHED_IDLE
))
4872 * Buddy candidates are cache hot:
4874 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
4875 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
4876 &p
->se
== cfs_rq_of(&p
->se
)->last
))
4879 if (sysctl_sched_migration_cost
== -1)
4881 if (sysctl_sched_migration_cost
== 0)
4884 delta
= now
- p
->se
.exec_start
;
4886 return delta
< (s64
)sysctl_sched_migration_cost
;
4889 #ifdef CONFIG_NUMA_BALANCING
4890 /* Returns true if the destination node has incurred more faults */
4891 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
4893 int src_nid
, dst_nid
;
4895 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults_memory
||
4896 !(env
->sd
->flags
& SD_NUMA
)) {
4900 src_nid
= cpu_to_node(env
->src_cpu
);
4901 dst_nid
= cpu_to_node(env
->dst_cpu
);
4903 if (src_nid
== dst_nid
)
4906 /* Always encourage migration to the preferred node. */
4907 if (dst_nid
== p
->numa_preferred_nid
)
4910 /* If both task and group weight improve, this move is a winner. */
4911 if (task_weight(p
, dst_nid
) > task_weight(p
, src_nid
) &&
4912 group_weight(p
, dst_nid
) > group_weight(p
, src_nid
))
4919 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
4921 int src_nid
, dst_nid
;
4923 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
4926 if (!p
->numa_faults_memory
|| !(env
->sd
->flags
& SD_NUMA
))
4929 src_nid
= cpu_to_node(env
->src_cpu
);
4930 dst_nid
= cpu_to_node(env
->dst_cpu
);
4932 if (src_nid
== dst_nid
)
4935 /* Migrating away from the preferred node is always bad. */
4936 if (src_nid
== p
->numa_preferred_nid
)
4939 /* If either task or group weight get worse, don't do it. */
4940 if (task_weight(p
, dst_nid
) < task_weight(p
, src_nid
) ||
4941 group_weight(p
, dst_nid
) < group_weight(p
, src_nid
))
4948 static inline bool migrate_improves_locality(struct task_struct
*p
,
4954 static inline bool migrate_degrades_locality(struct task_struct
*p
,
4962 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4965 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
4967 int tsk_cache_hot
= 0;
4969 * We do not migrate tasks that are:
4970 * 1) throttled_lb_pair, or
4971 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4972 * 3) running (obviously), or
4973 * 4) are cache-hot on their current CPU.
4975 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
4978 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
4981 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
4983 env
->flags
|= LBF_SOME_PINNED
;
4986 * Remember if this task can be migrated to any other cpu in
4987 * our sched_group. We may want to revisit it if we couldn't
4988 * meet load balance goals by pulling other tasks on src_cpu.
4990 * Also avoid computing new_dst_cpu if we have already computed
4991 * one in current iteration.
4993 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
4996 /* Prevent to re-select dst_cpu via env's cpus */
4997 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
4998 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
4999 env
->flags
|= LBF_DST_PINNED
;
5000 env
->new_dst_cpu
= cpu
;
5008 /* Record that we found atleast one task that could run on dst_cpu */
5009 env
->flags
&= ~LBF_ALL_PINNED
;
5011 if (task_running(env
->src_rq
, p
)) {
5012 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
5017 * Aggressive migration if:
5018 * 1) destination numa is preferred
5019 * 2) task is cache cold, or
5020 * 3) too many balance attempts have failed.
5022 tsk_cache_hot
= task_hot(p
, rq_clock_task(env
->src_rq
), env
->sd
);
5024 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
5026 if (migrate_improves_locality(p
, env
)) {
5027 #ifdef CONFIG_SCHEDSTATS
5028 if (tsk_cache_hot
) {
5029 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5030 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5036 if (!tsk_cache_hot
||
5037 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
5039 if (tsk_cache_hot
) {
5040 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5041 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5047 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
5052 * move_one_task tries to move exactly one task from busiest to this_rq, as
5053 * part of active balancing operations within "domain".
5054 * Returns 1 if successful and 0 otherwise.
5056 * Called with both runqueues locked.
5058 static int move_one_task(struct lb_env
*env
)
5060 struct task_struct
*p
, *n
;
5062 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
5063 if (!can_migrate_task(p
, env
))
5068 * Right now, this is only the second place move_task()
5069 * is called, so we can safely collect move_task()
5070 * stats here rather than inside move_task().
5072 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
5078 static const unsigned int sched_nr_migrate_break
= 32;
5081 * move_tasks tries to move up to imbalance weighted load from busiest to
5082 * this_rq, as part of a balancing operation within domain "sd".
5083 * Returns 1 if successful and 0 otherwise.
5085 * Called with both runqueues locked.
5087 static int move_tasks(struct lb_env
*env
)
5089 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
5090 struct task_struct
*p
;
5094 if (env
->imbalance
<= 0)
5097 while (!list_empty(tasks
)) {
5098 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5101 /* We've more or less seen every task there is, call it quits */
5102 if (env
->loop
> env
->loop_max
)
5105 /* take a breather every nr_migrate tasks */
5106 if (env
->loop
> env
->loop_break
) {
5107 env
->loop_break
+= sched_nr_migrate_break
;
5108 env
->flags
|= LBF_NEED_BREAK
;
5112 if (!can_migrate_task(p
, env
))
5115 load
= task_h_load(p
);
5117 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5120 if ((load
/ 2) > env
->imbalance
)
5125 env
->imbalance
-= load
;
5127 #ifdef CONFIG_PREEMPT
5129 * NEWIDLE balancing is a source of latency, so preemptible
5130 * kernels will stop after the first task is pulled to minimize
5131 * the critical section.
5133 if (env
->idle
== CPU_NEWLY_IDLE
)
5138 * We only want to steal up to the prescribed amount of
5141 if (env
->imbalance
<= 0)
5146 list_move_tail(&p
->se
.group_node
, tasks
);
5150 * Right now, this is one of only two places move_task() is called,
5151 * so we can safely collect move_task() stats here rather than
5152 * inside move_task().
5154 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
5159 #ifdef CONFIG_FAIR_GROUP_SCHED
5161 * update tg->load_weight by folding this cpu's load_avg
5163 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5165 struct sched_entity
*se
= tg
->se
[cpu
];
5166 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5168 /* throttled entities do not contribute to load */
5169 if (throttled_hierarchy(cfs_rq
))
5172 update_cfs_rq_blocked_load(cfs_rq
, 1);
5175 update_entity_load_avg(se
, 1);
5177 * We pivot on our runnable average having decayed to zero for
5178 * list removal. This generally implies that all our children
5179 * have also been removed (modulo rounding error or bandwidth
5180 * control); however, such cases are rare and we can fix these
5183 * TODO: fix up out-of-order children on enqueue.
5185 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5186 list_del_leaf_cfs_rq(cfs_rq
);
5188 struct rq
*rq
= rq_of(cfs_rq
);
5189 update_rq_runnable_avg(rq
, rq
->nr_running
);
5193 static void update_blocked_averages(int cpu
)
5195 struct rq
*rq
= cpu_rq(cpu
);
5196 struct cfs_rq
*cfs_rq
;
5197 unsigned long flags
;
5199 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5200 update_rq_clock(rq
);
5202 * Iterates the task_group tree in a bottom up fashion, see
5203 * list_add_leaf_cfs_rq() for details.
5205 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5207 * Note: We may want to consider periodically releasing
5208 * rq->lock about these updates so that creating many task
5209 * groups does not result in continually extending hold time.
5211 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5214 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5218 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5219 * This needs to be done in a top-down fashion because the load of a child
5220 * group is a fraction of its parents load.
5222 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5224 struct rq
*rq
= rq_of(cfs_rq
);
5225 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5226 unsigned long now
= jiffies
;
5229 if (cfs_rq
->last_h_load_update
== now
)
5232 cfs_rq
->h_load_next
= NULL
;
5233 for_each_sched_entity(se
) {
5234 cfs_rq
= cfs_rq_of(se
);
5235 cfs_rq
->h_load_next
= se
;
5236 if (cfs_rq
->last_h_load_update
== now
)
5241 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5242 cfs_rq
->last_h_load_update
= now
;
5245 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5246 load
= cfs_rq
->h_load
;
5247 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5248 cfs_rq
->runnable_load_avg
+ 1);
5249 cfs_rq
= group_cfs_rq(se
);
5250 cfs_rq
->h_load
= load
;
5251 cfs_rq
->last_h_load_update
= now
;
5255 static unsigned long task_h_load(struct task_struct
*p
)
5257 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5259 update_cfs_rq_h_load(cfs_rq
);
5260 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5261 cfs_rq
->runnable_load_avg
+ 1);
5264 static inline void update_blocked_averages(int cpu
)
5268 static unsigned long task_h_load(struct task_struct
*p
)
5270 return p
->se
.avg
.load_avg_contrib
;
5274 /********** Helpers for find_busiest_group ************************/
5276 * sg_lb_stats - stats of a sched_group required for load_balancing
5278 struct sg_lb_stats
{
5279 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5280 unsigned long group_load
; /* Total load over the CPUs of the group */
5281 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5282 unsigned long load_per_task
;
5283 unsigned long group_power
;
5284 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5285 unsigned int group_capacity
;
5286 unsigned int idle_cpus
;
5287 unsigned int group_weight
;
5288 int group_imb
; /* Is there an imbalance in the group ? */
5289 int group_has_capacity
; /* Is there extra capacity in the group? */
5290 #ifdef CONFIG_NUMA_BALANCING
5291 unsigned int nr_numa_running
;
5292 unsigned int nr_preferred_running
;
5297 * sd_lb_stats - Structure to store the statistics of a sched_domain
5298 * during load balancing.
5300 struct sd_lb_stats
{
5301 struct sched_group
*busiest
; /* Busiest group in this sd */
5302 struct sched_group
*local
; /* Local group in this sd */
5303 unsigned long total_load
; /* Total load of all groups in sd */
5304 unsigned long total_pwr
; /* Total power of all groups in sd */
5305 unsigned long avg_load
; /* Average load across all groups in sd */
5307 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5308 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5311 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5314 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5315 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5316 * We must however clear busiest_stat::avg_load because
5317 * update_sd_pick_busiest() reads this before assignment.
5319 *sds
= (struct sd_lb_stats
){
5331 * get_sd_load_idx - Obtain the load index for a given sched domain.
5332 * @sd: The sched_domain whose load_idx is to be obtained.
5333 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5335 * Return: The load index.
5337 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5338 enum cpu_idle_type idle
)
5344 load_idx
= sd
->busy_idx
;
5347 case CPU_NEWLY_IDLE
:
5348 load_idx
= sd
->newidle_idx
;
5351 load_idx
= sd
->idle_idx
;
5358 static unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
5360 return SCHED_POWER_SCALE
;
5363 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
5365 return default_scale_freq_power(sd
, cpu
);
5368 static unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
5370 unsigned long weight
= sd
->span_weight
;
5371 unsigned long smt_gain
= sd
->smt_gain
;
5378 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
5380 return default_scale_smt_power(sd
, cpu
);
5383 static unsigned long scale_rt_power(int cpu
)
5385 struct rq
*rq
= cpu_rq(cpu
);
5386 u64 total
, available
, age_stamp
, avg
;
5389 * Since we're reading these variables without serialization make sure
5390 * we read them once before doing sanity checks on them.
5392 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5393 avg
= ACCESS_ONCE(rq
->rt_avg
);
5395 total
= sched_avg_period() + (rq_clock(rq
) - age_stamp
);
5397 if (unlikely(total
< avg
)) {
5398 /* Ensures that power won't end up being negative */
5401 available
= total
- avg
;
5404 if (unlikely((s64
)total
< SCHED_POWER_SCALE
))
5405 total
= SCHED_POWER_SCALE
;
5407 total
>>= SCHED_POWER_SHIFT
;
5409 return div_u64(available
, total
);
5412 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
5414 unsigned long weight
= sd
->span_weight
;
5415 unsigned long power
= SCHED_POWER_SCALE
;
5416 struct sched_group
*sdg
= sd
->groups
;
5418 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
5419 if (sched_feat(ARCH_POWER
))
5420 power
*= arch_scale_smt_power(sd
, cpu
);
5422 power
*= default_scale_smt_power(sd
, cpu
);
5424 power
>>= SCHED_POWER_SHIFT
;
5427 sdg
->sgp
->power_orig
= power
;
5429 if (sched_feat(ARCH_POWER
))
5430 power
*= arch_scale_freq_power(sd
, cpu
);
5432 power
*= default_scale_freq_power(sd
, cpu
);
5434 power
>>= SCHED_POWER_SHIFT
;
5436 power
*= scale_rt_power(cpu
);
5437 power
>>= SCHED_POWER_SHIFT
;
5442 cpu_rq(cpu
)->cpu_power
= power
;
5443 sdg
->sgp
->power
= power
;
5446 void update_group_power(struct sched_domain
*sd
, int cpu
)
5448 struct sched_domain
*child
= sd
->child
;
5449 struct sched_group
*group
, *sdg
= sd
->groups
;
5450 unsigned long power
, power_orig
;
5451 unsigned long interval
;
5453 interval
= msecs_to_jiffies(sd
->balance_interval
);
5454 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5455 sdg
->sgp
->next_update
= jiffies
+ interval
;
5458 update_cpu_power(sd
, cpu
);
5462 power_orig
= power
= 0;
5464 if (child
->flags
& SD_OVERLAP
) {
5466 * SD_OVERLAP domains cannot assume that child groups
5467 * span the current group.
5470 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
5471 struct sched_group_power
*sgp
;
5472 struct rq
*rq
= cpu_rq(cpu
);
5475 * build_sched_domains() -> init_sched_groups_power()
5476 * gets here before we've attached the domains to the
5479 * Use power_of(), which is set irrespective of domains
5480 * in update_cpu_power().
5482 * This avoids power/power_orig from being 0 and
5483 * causing divide-by-zero issues on boot.
5485 * Runtime updates will correct power_orig.
5487 if (unlikely(!rq
->sd
)) {
5488 power_orig
+= power_of(cpu
);
5489 power
+= power_of(cpu
);
5493 sgp
= rq
->sd
->groups
->sgp
;
5494 power_orig
+= sgp
->power_orig
;
5495 power
+= sgp
->power
;
5499 * !SD_OVERLAP domains can assume that child groups
5500 * span the current group.
5503 group
= child
->groups
;
5505 power_orig
+= group
->sgp
->power_orig
;
5506 power
+= group
->sgp
->power
;
5507 group
= group
->next
;
5508 } while (group
!= child
->groups
);
5511 sdg
->sgp
->power_orig
= power_orig
;
5512 sdg
->sgp
->power
= power
;
5516 * Try and fix up capacity for tiny siblings, this is needed when
5517 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5518 * which on its own isn't powerful enough.
5520 * See update_sd_pick_busiest() and check_asym_packing().
5523 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
5526 * Only siblings can have significantly less than SCHED_POWER_SCALE
5528 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
5532 * If ~90% of the cpu_power is still there, we're good.
5534 if (group
->sgp
->power
* 32 > group
->sgp
->power_orig
* 29)
5541 * Group imbalance indicates (and tries to solve) the problem where balancing
5542 * groups is inadequate due to tsk_cpus_allowed() constraints.
5544 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5545 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5548 * { 0 1 2 3 } { 4 5 6 7 }
5551 * If we were to balance group-wise we'd place two tasks in the first group and
5552 * two tasks in the second group. Clearly this is undesired as it will overload
5553 * cpu 3 and leave one of the cpus in the second group unused.
5555 * The current solution to this issue is detecting the skew in the first group
5556 * by noticing the lower domain failed to reach balance and had difficulty
5557 * moving tasks due to affinity constraints.
5559 * When this is so detected; this group becomes a candidate for busiest; see
5560 * update_sd_pick_busiest(). And calculate_imbalance() and
5561 * find_busiest_group() avoid some of the usual balance conditions to allow it
5562 * to create an effective group imbalance.
5564 * This is a somewhat tricky proposition since the next run might not find the
5565 * group imbalance and decide the groups need to be balanced again. A most
5566 * subtle and fragile situation.
5569 static inline int sg_imbalanced(struct sched_group
*group
)
5571 return group
->sgp
->imbalance
;
5575 * Compute the group capacity.
5577 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5578 * first dividing out the smt factor and computing the actual number of cores
5579 * and limit power unit capacity with that.
5581 static inline int sg_capacity(struct lb_env
*env
, struct sched_group
*group
)
5583 unsigned int capacity
, smt
, cpus
;
5584 unsigned int power
, power_orig
;
5586 power
= group
->sgp
->power
;
5587 power_orig
= group
->sgp
->power_orig
;
5588 cpus
= group
->group_weight
;
5590 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5591 smt
= DIV_ROUND_UP(SCHED_POWER_SCALE
* cpus
, power_orig
);
5592 capacity
= cpus
/ smt
; /* cores */
5594 capacity
= min_t(unsigned, capacity
, DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
));
5596 capacity
= fix_small_capacity(env
->sd
, group
);
5602 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5603 * @env: The load balancing environment.
5604 * @group: sched_group whose statistics are to be updated.
5605 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5606 * @local_group: Does group contain this_cpu.
5607 * @sgs: variable to hold the statistics for this group.
5609 static inline void update_sg_lb_stats(struct lb_env
*env
,
5610 struct sched_group
*group
, int load_idx
,
5611 int local_group
, struct sg_lb_stats
*sgs
)
5616 memset(sgs
, 0, sizeof(*sgs
));
5618 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5619 struct rq
*rq
= cpu_rq(i
);
5621 /* Bias balancing toward cpus of our domain */
5623 load
= target_load(i
, load_idx
);
5625 load
= source_load(i
, load_idx
);
5627 sgs
->group_load
+= load
;
5628 sgs
->sum_nr_running
+= rq
->nr_running
;
5629 #ifdef CONFIG_NUMA_BALANCING
5630 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
5631 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
5633 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
5638 /* Adjust by relative CPU power of the group */
5639 sgs
->group_power
= group
->sgp
->power
;
5640 sgs
->avg_load
= (sgs
->group_load
*SCHED_POWER_SCALE
) / sgs
->group_power
;
5642 if (sgs
->sum_nr_running
)
5643 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
5645 sgs
->group_weight
= group
->group_weight
;
5647 sgs
->group_imb
= sg_imbalanced(group
);
5648 sgs
->group_capacity
= sg_capacity(env
, group
);
5650 if (sgs
->group_capacity
> sgs
->sum_nr_running
)
5651 sgs
->group_has_capacity
= 1;
5655 * update_sd_pick_busiest - return 1 on busiest group
5656 * @env: The load balancing environment.
5657 * @sds: sched_domain statistics
5658 * @sg: sched_group candidate to be checked for being the busiest
5659 * @sgs: sched_group statistics
5661 * Determine if @sg is a busier group than the previously selected
5664 * Return: %true if @sg is a busier group than the previously selected
5665 * busiest group. %false otherwise.
5667 static bool update_sd_pick_busiest(struct lb_env
*env
,
5668 struct sd_lb_stats
*sds
,
5669 struct sched_group
*sg
,
5670 struct sg_lb_stats
*sgs
)
5672 if (sgs
->avg_load
<= sds
->busiest_stat
.avg_load
)
5675 if (sgs
->sum_nr_running
> sgs
->group_capacity
)
5682 * ASYM_PACKING needs to move all the work to the lowest
5683 * numbered CPUs in the group, therefore mark all groups
5684 * higher than ourself as busy.
5686 if ((env
->sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
5687 env
->dst_cpu
< group_first_cpu(sg
)) {
5691 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
5698 #ifdef CONFIG_NUMA_BALANCING
5699 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
5701 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
5703 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
5708 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
5710 if (rq
->nr_running
> rq
->nr_numa_running
)
5712 if (rq
->nr_running
> rq
->nr_preferred_running
)
5717 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
5722 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
5726 #endif /* CONFIG_NUMA_BALANCING */
5729 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5730 * @env: The load balancing environment.
5731 * @sds: variable to hold the statistics for this sched_domain.
5733 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5735 struct sched_domain
*child
= env
->sd
->child
;
5736 struct sched_group
*sg
= env
->sd
->groups
;
5737 struct sg_lb_stats tmp_sgs
;
5738 int load_idx
, prefer_sibling
= 0;
5740 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
5743 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
5746 struct sg_lb_stats
*sgs
= &tmp_sgs
;
5749 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
5752 sgs
= &sds
->local_stat
;
5754 if (env
->idle
!= CPU_NEWLY_IDLE
||
5755 time_after_eq(jiffies
, sg
->sgp
->next_update
))
5756 update_group_power(env
->sd
, env
->dst_cpu
);
5759 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
);
5765 * In case the child domain prefers tasks go to siblings
5766 * first, lower the sg capacity to one so that we'll try
5767 * and move all the excess tasks away. We lower the capacity
5768 * of a group only if the local group has the capacity to fit
5769 * these excess tasks, i.e. nr_running < group_capacity. The
5770 * extra check prevents the case where you always pull from the
5771 * heaviest group when it is already under-utilized (possible
5772 * with a large weight task outweighs the tasks on the system).
5774 if (prefer_sibling
&& sds
->local
&&
5775 sds
->local_stat
.group_has_capacity
)
5776 sgs
->group_capacity
= min(sgs
->group_capacity
, 1U);
5778 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
5780 sds
->busiest_stat
= *sgs
;
5784 /* Now, start updating sd_lb_stats */
5785 sds
->total_load
+= sgs
->group_load
;
5786 sds
->total_pwr
+= sgs
->group_power
;
5789 } while (sg
!= env
->sd
->groups
);
5791 if (env
->sd
->flags
& SD_NUMA
)
5792 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
5796 * check_asym_packing - Check to see if the group is packed into the
5799 * This is primarily intended to used at the sibling level. Some
5800 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5801 * case of POWER7, it can move to lower SMT modes only when higher
5802 * threads are idle. When in lower SMT modes, the threads will
5803 * perform better since they share less core resources. Hence when we
5804 * have idle threads, we want them to be the higher ones.
5806 * This packing function is run on idle threads. It checks to see if
5807 * the busiest CPU in this domain (core in the P7 case) has a higher
5808 * CPU number than the packing function is being run on. Here we are
5809 * assuming lower CPU number will be equivalent to lower a SMT thread
5812 * Return: 1 when packing is required and a task should be moved to
5813 * this CPU. The amount of the imbalance is returned in *imbalance.
5815 * @env: The load balancing environment.
5816 * @sds: Statistics of the sched_domain which is to be packed
5818 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5822 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
5828 busiest_cpu
= group_first_cpu(sds
->busiest
);
5829 if (env
->dst_cpu
> busiest_cpu
)
5832 env
->imbalance
= DIV_ROUND_CLOSEST(
5833 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_power
,
5840 * fix_small_imbalance - Calculate the minor imbalance that exists
5841 * amongst the groups of a sched_domain, during
5843 * @env: The load balancing environment.
5844 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5847 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5849 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
5850 unsigned int imbn
= 2;
5851 unsigned long scaled_busy_load_per_task
;
5852 struct sg_lb_stats
*local
, *busiest
;
5854 local
= &sds
->local_stat
;
5855 busiest
= &sds
->busiest_stat
;
5857 if (!local
->sum_nr_running
)
5858 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
5859 else if (busiest
->load_per_task
> local
->load_per_task
)
5862 scaled_busy_load_per_task
=
5863 (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5864 busiest
->group_power
;
5866 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
5867 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
5868 env
->imbalance
= busiest
->load_per_task
;
5873 * OK, we don't have enough imbalance to justify moving tasks,
5874 * however we may be able to increase total CPU power used by
5878 pwr_now
+= busiest
->group_power
*
5879 min(busiest
->load_per_task
, busiest
->avg_load
);
5880 pwr_now
+= local
->group_power
*
5881 min(local
->load_per_task
, local
->avg_load
);
5882 pwr_now
/= SCHED_POWER_SCALE
;
5884 /* Amount of load we'd subtract */
5885 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5886 busiest
->group_power
;
5887 if (busiest
->avg_load
> tmp
) {
5888 pwr_move
+= busiest
->group_power
*
5889 min(busiest
->load_per_task
,
5890 busiest
->avg_load
- tmp
);
5893 /* Amount of load we'd add */
5894 if (busiest
->avg_load
* busiest
->group_power
<
5895 busiest
->load_per_task
* SCHED_POWER_SCALE
) {
5896 tmp
= (busiest
->avg_load
* busiest
->group_power
) /
5899 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5902 pwr_move
+= local
->group_power
*
5903 min(local
->load_per_task
, local
->avg_load
+ tmp
);
5904 pwr_move
/= SCHED_POWER_SCALE
;
5906 /* Move if we gain throughput */
5907 if (pwr_move
> pwr_now
)
5908 env
->imbalance
= busiest
->load_per_task
;
5912 * calculate_imbalance - Calculate the amount of imbalance present within the
5913 * groups of a given sched_domain during load balance.
5914 * @env: load balance environment
5915 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5917 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5919 unsigned long max_pull
, load_above_capacity
= ~0UL;
5920 struct sg_lb_stats
*local
, *busiest
;
5922 local
= &sds
->local_stat
;
5923 busiest
= &sds
->busiest_stat
;
5925 if (busiest
->group_imb
) {
5927 * In the group_imb case we cannot rely on group-wide averages
5928 * to ensure cpu-load equilibrium, look at wider averages. XXX
5930 busiest
->load_per_task
=
5931 min(busiest
->load_per_task
, sds
->avg_load
);
5935 * In the presence of smp nice balancing, certain scenarios can have
5936 * max load less than avg load(as we skip the groups at or below
5937 * its cpu_power, while calculating max_load..)
5939 if (busiest
->avg_load
<= sds
->avg_load
||
5940 local
->avg_load
>= sds
->avg_load
) {
5942 return fix_small_imbalance(env
, sds
);
5945 if (!busiest
->group_imb
) {
5947 * Don't want to pull so many tasks that a group would go idle.
5948 * Except of course for the group_imb case, since then we might
5949 * have to drop below capacity to reach cpu-load equilibrium.
5951 load_above_capacity
=
5952 (busiest
->sum_nr_running
- busiest
->group_capacity
);
5954 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_POWER_SCALE
);
5955 load_above_capacity
/= busiest
->group_power
;
5959 * We're trying to get all the cpus to the average_load, so we don't
5960 * want to push ourselves above the average load, nor do we wish to
5961 * reduce the max loaded cpu below the average load. At the same time,
5962 * we also don't want to reduce the group load below the group capacity
5963 * (so that we can implement power-savings policies etc). Thus we look
5964 * for the minimum possible imbalance.
5966 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
5968 /* How much load to actually move to equalise the imbalance */
5969 env
->imbalance
= min(
5970 max_pull
* busiest
->group_power
,
5971 (sds
->avg_load
- local
->avg_load
) * local
->group_power
5972 ) / SCHED_POWER_SCALE
;
5975 * if *imbalance is less than the average load per runnable task
5976 * there is no guarantee that any tasks will be moved so we'll have
5977 * a think about bumping its value to force at least one task to be
5980 if (env
->imbalance
< busiest
->load_per_task
)
5981 return fix_small_imbalance(env
, sds
);
5984 /******* find_busiest_group() helpers end here *********************/
5987 * find_busiest_group - Returns the busiest group within the sched_domain
5988 * if there is an imbalance. If there isn't an imbalance, and
5989 * the user has opted for power-savings, it returns a group whose
5990 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5991 * such a group exists.
5993 * Also calculates the amount of weighted load which should be moved
5994 * to restore balance.
5996 * @env: The load balancing environment.
5998 * Return: - The busiest group if imbalance exists.
5999 * - If no imbalance and user has opted for power-savings balance,
6000 * return the least loaded group whose CPUs can be
6001 * put to idle by rebalancing its tasks onto our group.
6003 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
6005 struct sg_lb_stats
*local
, *busiest
;
6006 struct sd_lb_stats sds
;
6008 init_sd_lb_stats(&sds
);
6011 * Compute the various statistics relavent for load balancing at
6014 update_sd_lb_stats(env
, &sds
);
6015 local
= &sds
.local_stat
;
6016 busiest
= &sds
.busiest_stat
;
6018 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
6019 check_asym_packing(env
, &sds
))
6022 /* There is no busy sibling group to pull tasks from */
6023 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
6026 sds
.avg_load
= (SCHED_POWER_SCALE
* sds
.total_load
) / sds
.total_pwr
;
6029 * If the busiest group is imbalanced the below checks don't
6030 * work because they assume all things are equal, which typically
6031 * isn't true due to cpus_allowed constraints and the like.
6033 if (busiest
->group_imb
)
6036 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6037 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_capacity
&&
6038 !busiest
->group_has_capacity
)
6042 * If the local group is more busy than the selected busiest group
6043 * don't try and pull any tasks.
6045 if (local
->avg_load
>= busiest
->avg_load
)
6049 * Don't pull any tasks if this group is already above the domain
6052 if (local
->avg_load
>= sds
.avg_load
)
6055 if (env
->idle
== CPU_IDLE
) {
6057 * This cpu is idle. If the busiest group load doesn't
6058 * have more tasks than the number of available cpu's and
6059 * there is no imbalance between this and busiest group
6060 * wrt to idle cpu's, it is balanced.
6062 if ((local
->idle_cpus
< busiest
->idle_cpus
) &&
6063 busiest
->sum_nr_running
<= busiest
->group_weight
)
6067 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6068 * imbalance_pct to be conservative.
6070 if (100 * busiest
->avg_load
<=
6071 env
->sd
->imbalance_pct
* local
->avg_load
)
6076 /* Looks like there is an imbalance. Compute it */
6077 calculate_imbalance(env
, &sds
);
6086 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6088 static struct rq
*find_busiest_queue(struct lb_env
*env
,
6089 struct sched_group
*group
)
6091 struct rq
*busiest
= NULL
, *rq
;
6092 unsigned long busiest_load
= 0, busiest_power
= 1;
6095 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6096 unsigned long power
, capacity
, wl
;
6100 rt
= fbq_classify_rq(rq
);
6103 * We classify groups/runqueues into three groups:
6104 * - regular: there are !numa tasks
6105 * - remote: there are numa tasks that run on the 'wrong' node
6106 * - all: there is no distinction
6108 * In order to avoid migrating ideally placed numa tasks,
6109 * ignore those when there's better options.
6111 * If we ignore the actual busiest queue to migrate another
6112 * task, the next balance pass can still reduce the busiest
6113 * queue by moving tasks around inside the node.
6115 * If we cannot move enough load due to this classification
6116 * the next pass will adjust the group classification and
6117 * allow migration of more tasks.
6119 * Both cases only affect the total convergence complexity.
6121 if (rt
> env
->fbq_type
)
6124 power
= power_of(i
);
6125 capacity
= DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
);
6127 capacity
= fix_small_capacity(env
->sd
, group
);
6129 wl
= weighted_cpuload(i
);
6132 * When comparing with imbalance, use weighted_cpuload()
6133 * which is not scaled with the cpu power.
6135 if (capacity
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6139 * For the load comparisons with the other cpu's, consider
6140 * the weighted_cpuload() scaled with the cpu power, so that
6141 * the load can be moved away from the cpu that is potentially
6142 * running at a lower capacity.
6144 * Thus we're looking for max(wl_i / power_i), crosswise
6145 * multiplication to rid ourselves of the division works out
6146 * to: wl_i * power_j > wl_j * power_i; where j is our
6149 if (wl
* busiest_power
> busiest_load
* power
) {
6151 busiest_power
= power
;
6160 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6161 * so long as it is large enough.
6163 #define MAX_PINNED_INTERVAL 512
6165 /* Working cpumask for load_balance and load_balance_newidle. */
6166 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6168 static int need_active_balance(struct lb_env
*env
)
6170 struct sched_domain
*sd
= env
->sd
;
6172 if (env
->idle
== CPU_NEWLY_IDLE
) {
6175 * ASYM_PACKING needs to force migrate tasks from busy but
6176 * higher numbered CPUs in order to pack all tasks in the
6177 * lowest numbered CPUs.
6179 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6183 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6186 static int active_load_balance_cpu_stop(void *data
);
6188 static int should_we_balance(struct lb_env
*env
)
6190 struct sched_group
*sg
= env
->sd
->groups
;
6191 struct cpumask
*sg_cpus
, *sg_mask
;
6192 int cpu
, balance_cpu
= -1;
6195 * In the newly idle case, we will allow all the cpu's
6196 * to do the newly idle load balance.
6198 if (env
->idle
== CPU_NEWLY_IDLE
)
6201 sg_cpus
= sched_group_cpus(sg
);
6202 sg_mask
= sched_group_mask(sg
);
6203 /* Try to find first idle cpu */
6204 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6205 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6212 if (balance_cpu
== -1)
6213 balance_cpu
= group_balance_cpu(sg
);
6216 * First idle cpu or the first cpu(busiest) in this sched group
6217 * is eligible for doing load balancing at this and above domains.
6219 return balance_cpu
== env
->dst_cpu
;
6223 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6224 * tasks if there is an imbalance.
6226 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6227 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6228 int *continue_balancing
)
6230 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6231 struct sched_domain
*sd_parent
= sd
->parent
;
6232 struct sched_group
*group
;
6234 unsigned long flags
;
6235 struct cpumask
*cpus
= __get_cpu_var(load_balance_mask
);
6237 struct lb_env env
= {
6239 .dst_cpu
= this_cpu
,
6241 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6243 .loop_break
= sched_nr_migrate_break
,
6249 * For NEWLY_IDLE load_balancing, we don't need to consider
6250 * other cpus in our group
6252 if (idle
== CPU_NEWLY_IDLE
)
6253 env
.dst_grpmask
= NULL
;
6255 cpumask_copy(cpus
, cpu_active_mask
);
6257 schedstat_inc(sd
, lb_count
[idle
]);
6260 if (!should_we_balance(&env
)) {
6261 *continue_balancing
= 0;
6265 group
= find_busiest_group(&env
);
6267 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6271 busiest
= find_busiest_queue(&env
, group
);
6273 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6277 BUG_ON(busiest
== env
.dst_rq
);
6279 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6282 if (busiest
->nr_running
> 1) {
6284 * Attempt to move tasks. If find_busiest_group has found
6285 * an imbalance but busiest->nr_running <= 1, the group is
6286 * still unbalanced. ld_moved simply stays zero, so it is
6287 * correctly treated as an imbalance.
6289 env
.flags
|= LBF_ALL_PINNED
;
6290 env
.src_cpu
= busiest
->cpu
;
6291 env
.src_rq
= busiest
;
6292 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6295 local_irq_save(flags
);
6296 double_rq_lock(env
.dst_rq
, busiest
);
6299 * cur_ld_moved - load moved in current iteration
6300 * ld_moved - cumulative load moved across iterations
6302 cur_ld_moved
= move_tasks(&env
);
6303 ld_moved
+= cur_ld_moved
;
6304 double_rq_unlock(env
.dst_rq
, busiest
);
6305 local_irq_restore(flags
);
6308 * some other cpu did the load balance for us.
6310 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
6311 resched_cpu(env
.dst_cpu
);
6313 if (env
.flags
& LBF_NEED_BREAK
) {
6314 env
.flags
&= ~LBF_NEED_BREAK
;
6319 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6320 * us and move them to an alternate dst_cpu in our sched_group
6321 * where they can run. The upper limit on how many times we
6322 * iterate on same src_cpu is dependent on number of cpus in our
6325 * This changes load balance semantics a bit on who can move
6326 * load to a given_cpu. In addition to the given_cpu itself
6327 * (or a ilb_cpu acting on its behalf where given_cpu is
6328 * nohz-idle), we now have balance_cpu in a position to move
6329 * load to given_cpu. In rare situations, this may cause
6330 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6331 * _independently_ and at _same_ time to move some load to
6332 * given_cpu) causing exceess load to be moved to given_cpu.
6333 * This however should not happen so much in practice and
6334 * moreover subsequent load balance cycles should correct the
6335 * excess load moved.
6337 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6339 /* Prevent to re-select dst_cpu via env's cpus */
6340 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6342 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6343 env
.dst_cpu
= env
.new_dst_cpu
;
6344 env
.flags
&= ~LBF_DST_PINNED
;
6346 env
.loop_break
= sched_nr_migrate_break
;
6349 * Go back to "more_balance" rather than "redo" since we
6350 * need to continue with same src_cpu.
6356 * We failed to reach balance because of affinity.
6359 int *group_imbalance
= &sd_parent
->groups
->sgp
->imbalance
;
6361 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0) {
6362 *group_imbalance
= 1;
6363 } else if (*group_imbalance
)
6364 *group_imbalance
= 0;
6367 /* All tasks on this runqueue were pinned by CPU affinity */
6368 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6369 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6370 if (!cpumask_empty(cpus
)) {
6372 env
.loop_break
= sched_nr_migrate_break
;
6380 schedstat_inc(sd
, lb_failed
[idle
]);
6382 * Increment the failure counter only on periodic balance.
6383 * We do not want newidle balance, which can be very
6384 * frequent, pollute the failure counter causing
6385 * excessive cache_hot migrations and active balances.
6387 if (idle
!= CPU_NEWLY_IDLE
)
6388 sd
->nr_balance_failed
++;
6390 if (need_active_balance(&env
)) {
6391 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6393 /* don't kick the active_load_balance_cpu_stop,
6394 * if the curr task on busiest cpu can't be
6397 if (!cpumask_test_cpu(this_cpu
,
6398 tsk_cpus_allowed(busiest
->curr
))) {
6399 raw_spin_unlock_irqrestore(&busiest
->lock
,
6401 env
.flags
|= LBF_ALL_PINNED
;
6402 goto out_one_pinned
;
6406 * ->active_balance synchronizes accesses to
6407 * ->active_balance_work. Once set, it's cleared
6408 * only after active load balance is finished.
6410 if (!busiest
->active_balance
) {
6411 busiest
->active_balance
= 1;
6412 busiest
->push_cpu
= this_cpu
;
6415 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
6417 if (active_balance
) {
6418 stop_one_cpu_nowait(cpu_of(busiest
),
6419 active_load_balance_cpu_stop
, busiest
,
6420 &busiest
->active_balance_work
);
6424 * We've kicked active balancing, reset the failure
6427 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
6430 sd
->nr_balance_failed
= 0;
6432 if (likely(!active_balance
)) {
6433 /* We were unbalanced, so reset the balancing interval */
6434 sd
->balance_interval
= sd
->min_interval
;
6437 * If we've begun active balancing, start to back off. This
6438 * case may not be covered by the all_pinned logic if there
6439 * is only 1 task on the busy runqueue (because we don't call
6442 if (sd
->balance_interval
< sd
->max_interval
)
6443 sd
->balance_interval
*= 2;
6449 schedstat_inc(sd
, lb_balanced
[idle
]);
6451 sd
->nr_balance_failed
= 0;
6454 /* tune up the balancing interval */
6455 if (((env
.flags
& LBF_ALL_PINNED
) &&
6456 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
6457 (sd
->balance_interval
< sd
->max_interval
))
6458 sd
->balance_interval
*= 2;
6466 * idle_balance is called by schedule() if this_cpu is about to become
6467 * idle. Attempts to pull tasks from other CPUs.
6469 void idle_balance(int this_cpu
, struct rq
*this_rq
)
6471 struct sched_domain
*sd
;
6472 int pulled_task
= 0;
6473 unsigned long next_balance
= jiffies
+ HZ
;
6476 this_rq
->idle_stamp
= rq_clock(this_rq
);
6478 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
6482 * Drop the rq->lock, but keep IRQ/preempt disabled.
6484 raw_spin_unlock(&this_rq
->lock
);
6486 update_blocked_averages(this_cpu
);
6488 for_each_domain(this_cpu
, sd
) {
6489 unsigned long interval
;
6490 int continue_balancing
= 1;
6491 u64 t0
, domain_cost
;
6493 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6496 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
)
6499 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
6500 t0
= sched_clock_cpu(this_cpu
);
6502 /* If we've pulled tasks over stop searching: */
6503 pulled_task
= load_balance(this_cpu
, this_rq
,
6505 &continue_balancing
);
6507 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
6508 if (domain_cost
> sd
->max_newidle_lb_cost
)
6509 sd
->max_newidle_lb_cost
= domain_cost
;
6511 curr_cost
+= domain_cost
;
6514 interval
= msecs_to_jiffies(sd
->balance_interval
);
6515 if (time_after(next_balance
, sd
->last_balance
+ interval
))
6516 next_balance
= sd
->last_balance
+ interval
;
6518 this_rq
->idle_stamp
= 0;
6524 raw_spin_lock(&this_rq
->lock
);
6526 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
6528 * We are going idle. next_balance may be set based on
6529 * a busy processor. So reset next_balance.
6531 this_rq
->next_balance
= next_balance
;
6534 if (curr_cost
> this_rq
->max_idle_balance_cost
)
6535 this_rq
->max_idle_balance_cost
= curr_cost
;
6539 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6540 * running tasks off the busiest CPU onto idle CPUs. It requires at
6541 * least 1 task to be running on each physical CPU where possible, and
6542 * avoids physical / logical imbalances.
6544 static int active_load_balance_cpu_stop(void *data
)
6546 struct rq
*busiest_rq
= data
;
6547 int busiest_cpu
= cpu_of(busiest_rq
);
6548 int target_cpu
= busiest_rq
->push_cpu
;
6549 struct rq
*target_rq
= cpu_rq(target_cpu
);
6550 struct sched_domain
*sd
;
6552 raw_spin_lock_irq(&busiest_rq
->lock
);
6554 /* make sure the requested cpu hasn't gone down in the meantime */
6555 if (unlikely(busiest_cpu
!= smp_processor_id() ||
6556 !busiest_rq
->active_balance
))
6559 /* Is there any task to move? */
6560 if (busiest_rq
->nr_running
<= 1)
6564 * This condition is "impossible", if it occurs
6565 * we need to fix it. Originally reported by
6566 * Bjorn Helgaas on a 128-cpu setup.
6568 BUG_ON(busiest_rq
== target_rq
);
6570 /* move a task from busiest_rq to target_rq */
6571 double_lock_balance(busiest_rq
, target_rq
);
6573 /* Search for an sd spanning us and the target CPU. */
6575 for_each_domain(target_cpu
, sd
) {
6576 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
6577 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
6582 struct lb_env env
= {
6584 .dst_cpu
= target_cpu
,
6585 .dst_rq
= target_rq
,
6586 .src_cpu
= busiest_rq
->cpu
,
6587 .src_rq
= busiest_rq
,
6591 schedstat_inc(sd
, alb_count
);
6593 if (move_one_task(&env
))
6594 schedstat_inc(sd
, alb_pushed
);
6596 schedstat_inc(sd
, alb_failed
);
6599 double_unlock_balance(busiest_rq
, target_rq
);
6601 busiest_rq
->active_balance
= 0;
6602 raw_spin_unlock_irq(&busiest_rq
->lock
);
6606 #ifdef CONFIG_NO_HZ_COMMON
6608 * idle load balancing details
6609 * - When one of the busy CPUs notice that there may be an idle rebalancing
6610 * needed, they will kick the idle load balancer, which then does idle
6611 * load balancing for all the idle CPUs.
6614 cpumask_var_t idle_cpus_mask
;
6616 unsigned long next_balance
; /* in jiffy units */
6617 } nohz ____cacheline_aligned
;
6619 static inline int find_new_ilb(void)
6621 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
6623 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
6630 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6631 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6632 * CPU (if there is one).
6634 static void nohz_balancer_kick(void)
6638 nohz
.next_balance
++;
6640 ilb_cpu
= find_new_ilb();
6642 if (ilb_cpu
>= nr_cpu_ids
)
6645 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
6648 * Use smp_send_reschedule() instead of resched_cpu().
6649 * This way we generate a sched IPI on the target cpu which
6650 * is idle. And the softirq performing nohz idle load balance
6651 * will be run before returning from the IPI.
6653 smp_send_reschedule(ilb_cpu
);
6657 static inline void nohz_balance_exit_idle(int cpu
)
6659 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
6660 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
6661 atomic_dec(&nohz
.nr_cpus
);
6662 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
6666 static inline void set_cpu_sd_state_busy(void)
6668 struct sched_domain
*sd
;
6669 int cpu
= smp_processor_id();
6672 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
6674 if (!sd
|| !sd
->nohz_idle
)
6678 atomic_inc(&sd
->groups
->sgp
->nr_busy_cpus
);
6683 void set_cpu_sd_state_idle(void)
6685 struct sched_domain
*sd
;
6686 int cpu
= smp_processor_id();
6689 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
6691 if (!sd
|| sd
->nohz_idle
)
6695 atomic_dec(&sd
->groups
->sgp
->nr_busy_cpus
);
6701 * This routine will record that the cpu is going idle with tick stopped.
6702 * This info will be used in performing idle load balancing in the future.
6704 void nohz_balance_enter_idle(int cpu
)
6707 * If this cpu is going down, then nothing needs to be done.
6709 if (!cpu_active(cpu
))
6712 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
6715 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
6716 atomic_inc(&nohz
.nr_cpus
);
6717 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
6720 static int sched_ilb_notifier(struct notifier_block
*nfb
,
6721 unsigned long action
, void *hcpu
)
6723 switch (action
& ~CPU_TASKS_FROZEN
) {
6725 nohz_balance_exit_idle(smp_processor_id());
6733 static DEFINE_SPINLOCK(balancing
);
6736 * Scale the max load_balance interval with the number of CPUs in the system.
6737 * This trades load-balance latency on larger machines for less cross talk.
6739 void update_max_interval(void)
6741 max_load_balance_interval
= HZ
*num_online_cpus()/10;
6745 * It checks each scheduling domain to see if it is due to be balanced,
6746 * and initiates a balancing operation if so.
6748 * Balancing parameters are set up in init_sched_domains.
6750 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
6752 int continue_balancing
= 1;
6754 unsigned long interval
;
6755 struct sched_domain
*sd
;
6756 /* Earliest time when we have to do rebalance again */
6757 unsigned long next_balance
= jiffies
+ 60*HZ
;
6758 int update_next_balance
= 0;
6759 int need_serialize
, need_decay
= 0;
6762 update_blocked_averages(cpu
);
6765 for_each_domain(cpu
, sd
) {
6767 * Decay the newidle max times here because this is a regular
6768 * visit to all the domains. Decay ~1% per second.
6770 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
6771 sd
->max_newidle_lb_cost
=
6772 (sd
->max_newidle_lb_cost
* 253) / 256;
6773 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
6776 max_cost
+= sd
->max_newidle_lb_cost
;
6778 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6782 * Stop the load balance at this level. There is another
6783 * CPU in our sched group which is doing load balancing more
6786 if (!continue_balancing
) {
6792 interval
= sd
->balance_interval
;
6793 if (idle
!= CPU_IDLE
)
6794 interval
*= sd
->busy_factor
;
6796 /* scale ms to jiffies */
6797 interval
= msecs_to_jiffies(interval
);
6798 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6800 need_serialize
= sd
->flags
& SD_SERIALIZE
;
6802 if (need_serialize
) {
6803 if (!spin_trylock(&balancing
))
6807 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
6808 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
6810 * The LBF_DST_PINNED logic could have changed
6811 * env->dst_cpu, so we can't know our idle
6812 * state even if we migrated tasks. Update it.
6814 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
6816 sd
->last_balance
= jiffies
;
6819 spin_unlock(&balancing
);
6821 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
6822 next_balance
= sd
->last_balance
+ interval
;
6823 update_next_balance
= 1;
6828 * Ensure the rq-wide value also decays but keep it at a
6829 * reasonable floor to avoid funnies with rq->avg_idle.
6831 rq
->max_idle_balance_cost
=
6832 max((u64
)sysctl_sched_migration_cost
, max_cost
);
6837 * next_balance will be updated only when there is a need.
6838 * When the cpu is attached to null domain for ex, it will not be
6841 if (likely(update_next_balance
))
6842 rq
->next_balance
= next_balance
;
6845 #ifdef CONFIG_NO_HZ_COMMON
6847 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6848 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6850 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
6852 int this_cpu
= this_rq
->cpu
;
6856 if (idle
!= CPU_IDLE
||
6857 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
6860 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
6861 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
6865 * If this cpu gets work to do, stop the load balancing
6866 * work being done for other cpus. Next load
6867 * balancing owner will pick it up.
6872 rq
= cpu_rq(balance_cpu
);
6874 raw_spin_lock_irq(&rq
->lock
);
6875 update_rq_clock(rq
);
6876 update_idle_cpu_load(rq
);
6877 raw_spin_unlock_irq(&rq
->lock
);
6879 rebalance_domains(rq
, CPU_IDLE
);
6881 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
6882 this_rq
->next_balance
= rq
->next_balance
;
6884 nohz
.next_balance
= this_rq
->next_balance
;
6886 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
6890 * Current heuristic for kicking the idle load balancer in the presence
6891 * of an idle cpu is the system.
6892 * - This rq has more than one task.
6893 * - At any scheduler domain level, this cpu's scheduler group has multiple
6894 * busy cpu's exceeding the group's power.
6895 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6896 * domain span are idle.
6898 static inline int nohz_kick_needed(struct rq
*rq
)
6900 unsigned long now
= jiffies
;
6901 struct sched_domain
*sd
;
6902 struct sched_group_power
*sgp
;
6903 int nr_busy
, cpu
= rq
->cpu
;
6905 if (unlikely(rq
->idle_balance
))
6909 * We may be recently in ticked or tickless idle mode. At the first
6910 * busy tick after returning from idle, we will update the busy stats.
6912 set_cpu_sd_state_busy();
6913 nohz_balance_exit_idle(cpu
);
6916 * None are in tickless mode and hence no need for NOHZ idle load
6919 if (likely(!atomic_read(&nohz
.nr_cpus
)))
6922 if (time_before(now
, nohz
.next_balance
))
6925 if (rq
->nr_running
>= 2)
6929 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
6932 sgp
= sd
->groups
->sgp
;
6933 nr_busy
= atomic_read(&sgp
->nr_busy_cpus
);
6936 goto need_kick_unlock
;
6939 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
6941 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
6942 sched_domain_span(sd
)) < cpu
))
6943 goto need_kick_unlock
;
6954 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
6958 * run_rebalance_domains is triggered when needed from the scheduler tick.
6959 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6961 static void run_rebalance_domains(struct softirq_action
*h
)
6963 struct rq
*this_rq
= this_rq();
6964 enum cpu_idle_type idle
= this_rq
->idle_balance
?
6965 CPU_IDLE
: CPU_NOT_IDLE
;
6967 rebalance_domains(this_rq
, idle
);
6970 * If this cpu has a pending nohz_balance_kick, then do the
6971 * balancing on behalf of the other idle cpus whose ticks are
6974 nohz_idle_balance(this_rq
, idle
);
6977 static inline int on_null_domain(struct rq
*rq
)
6979 return !rcu_dereference_sched(rq
->sd
);
6983 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6985 void trigger_load_balance(struct rq
*rq
)
6987 /* Don't need to rebalance while attached to NULL domain */
6988 if (unlikely(on_null_domain(rq
)))
6991 if (time_after_eq(jiffies
, rq
->next_balance
))
6992 raise_softirq(SCHED_SOFTIRQ
);
6993 #ifdef CONFIG_NO_HZ_COMMON
6994 if (nohz_kick_needed(rq
))
6995 nohz_balancer_kick();
6999 static void rq_online_fair(struct rq
*rq
)
7004 static void rq_offline_fair(struct rq
*rq
)
7008 /* Ensure any throttled groups are reachable by pick_next_task */
7009 unthrottle_offline_cfs_rqs(rq
);
7012 #endif /* CONFIG_SMP */
7015 * scheduler tick hitting a task of our scheduling class:
7017 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
7019 struct cfs_rq
*cfs_rq
;
7020 struct sched_entity
*se
= &curr
->se
;
7022 for_each_sched_entity(se
) {
7023 cfs_rq
= cfs_rq_of(se
);
7024 entity_tick(cfs_rq
, se
, queued
);
7027 if (numabalancing_enabled
)
7028 task_tick_numa(rq
, curr
);
7030 update_rq_runnable_avg(rq
, 1);
7034 * called on fork with the child task as argument from the parent's context
7035 * - child not yet on the tasklist
7036 * - preemption disabled
7038 static void task_fork_fair(struct task_struct
*p
)
7040 struct cfs_rq
*cfs_rq
;
7041 struct sched_entity
*se
= &p
->se
, *curr
;
7042 int this_cpu
= smp_processor_id();
7043 struct rq
*rq
= this_rq();
7044 unsigned long flags
;
7046 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7048 update_rq_clock(rq
);
7050 cfs_rq
= task_cfs_rq(current
);
7051 curr
= cfs_rq
->curr
;
7054 * Not only the cpu but also the task_group of the parent might have
7055 * been changed after parent->se.parent,cfs_rq were copied to
7056 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7057 * of child point to valid ones.
7060 __set_task_cpu(p
, this_cpu
);
7063 update_curr(cfs_rq
);
7066 se
->vruntime
= curr
->vruntime
;
7067 place_entity(cfs_rq
, se
, 1);
7069 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
7071 * Upon rescheduling, sched_class::put_prev_task() will place
7072 * 'current' within the tree based on its new key value.
7074 swap(curr
->vruntime
, se
->vruntime
);
7075 resched_task(rq
->curr
);
7078 se
->vruntime
-= cfs_rq
->min_vruntime
;
7080 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7084 * Priority of the task has changed. Check to see if we preempt
7088 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
7094 * Reschedule if we are currently running on this runqueue and
7095 * our priority decreased, or if we are not currently running on
7096 * this runqueue and our priority is higher than the current's
7098 if (rq
->curr
== p
) {
7099 if (p
->prio
> oldprio
)
7100 resched_task(rq
->curr
);
7102 check_preempt_curr(rq
, p
, 0);
7105 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
7107 struct sched_entity
*se
= &p
->se
;
7108 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7111 * Ensure the task's vruntime is normalized, so that when its
7112 * switched back to the fair class the enqueue_entity(.flags=0) will
7113 * do the right thing.
7115 * If it was on_rq, then the dequeue_entity(.flags=0) will already
7116 * have normalized the vruntime, if it was !on_rq, then only when
7117 * the task is sleeping will it still have non-normalized vruntime.
7119 if (!se
->on_rq
&& p
->state
!= TASK_RUNNING
) {
7121 * Fix up our vruntime so that the current sleep doesn't
7122 * cause 'unlimited' sleep bonus.
7124 place_entity(cfs_rq
, se
, 0);
7125 se
->vruntime
-= cfs_rq
->min_vruntime
;
7130 * Remove our load from contribution when we leave sched_fair
7131 * and ensure we don't carry in an old decay_count if we
7134 if (se
->avg
.decay_count
) {
7135 __synchronize_entity_decay(se
);
7136 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7142 * We switched to the sched_fair class.
7144 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7150 * We were most likely switched from sched_rt, so
7151 * kick off the schedule if running, otherwise just see
7152 * if we can still preempt the current task.
7155 resched_task(rq
->curr
);
7157 check_preempt_curr(rq
, p
, 0);
7160 /* Account for a task changing its policy or group.
7162 * This routine is mostly called to set cfs_rq->curr field when a task
7163 * migrates between groups/classes.
7165 static void set_curr_task_fair(struct rq
*rq
)
7167 struct sched_entity
*se
= &rq
->curr
->se
;
7169 for_each_sched_entity(se
) {
7170 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7172 set_next_entity(cfs_rq
, se
);
7173 /* ensure bandwidth has been allocated on our new cfs_rq */
7174 account_cfs_rq_runtime(cfs_rq
, 0);
7178 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7180 cfs_rq
->tasks_timeline
= RB_ROOT
;
7181 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7182 #ifndef CONFIG_64BIT
7183 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7186 atomic64_set(&cfs_rq
->decay_counter
, 1);
7187 atomic_long_set(&cfs_rq
->removed_load
, 0);
7191 #ifdef CONFIG_FAIR_GROUP_SCHED
7192 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
7194 struct cfs_rq
*cfs_rq
;
7196 * If the task was not on the rq at the time of this cgroup movement
7197 * it must have been asleep, sleeping tasks keep their ->vruntime
7198 * absolute on their old rq until wakeup (needed for the fair sleeper
7199 * bonus in place_entity()).
7201 * If it was on the rq, we've just 'preempted' it, which does convert
7202 * ->vruntime to a relative base.
7204 * Make sure both cases convert their relative position when migrating
7205 * to another cgroup's rq. This does somewhat interfere with the
7206 * fair sleeper stuff for the first placement, but who cares.
7209 * When !on_rq, vruntime of the task has usually NOT been normalized.
7210 * But there are some cases where it has already been normalized:
7212 * - Moving a forked child which is waiting for being woken up by
7213 * wake_up_new_task().
7214 * - Moving a task which has been woken up by try_to_wake_up() and
7215 * waiting for actually being woken up by sched_ttwu_pending().
7217 * To prevent boost or penalty in the new cfs_rq caused by delta
7218 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7220 if (!on_rq
&& (!p
->se
.sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7224 p
->se
.vruntime
-= cfs_rq_of(&p
->se
)->min_vruntime
;
7225 set_task_rq(p
, task_cpu(p
));
7227 cfs_rq
= cfs_rq_of(&p
->se
);
7228 p
->se
.vruntime
+= cfs_rq
->min_vruntime
;
7231 * migrate_task_rq_fair() will have removed our previous
7232 * contribution, but we must synchronize for ongoing future
7235 p
->se
.avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7236 cfs_rq
->blocked_load_avg
+= p
->se
.avg
.load_avg_contrib
;
7241 void free_fair_sched_group(struct task_group
*tg
)
7245 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7247 for_each_possible_cpu(i
) {
7249 kfree(tg
->cfs_rq
[i
]);
7258 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7260 struct cfs_rq
*cfs_rq
;
7261 struct sched_entity
*se
;
7264 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7267 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7271 tg
->shares
= NICE_0_LOAD
;
7273 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7275 for_each_possible_cpu(i
) {
7276 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7277 GFP_KERNEL
, cpu_to_node(i
));
7281 se
= kzalloc_node(sizeof(struct sched_entity
),
7282 GFP_KERNEL
, cpu_to_node(i
));
7286 init_cfs_rq(cfs_rq
);
7287 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
7298 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7300 struct rq
*rq
= cpu_rq(cpu
);
7301 unsigned long flags
;
7304 * Only empty task groups can be destroyed; so we can speculatively
7305 * check on_list without danger of it being re-added.
7307 if (!tg
->cfs_rq
[cpu
]->on_list
)
7310 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7311 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
7312 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7315 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7316 struct sched_entity
*se
, int cpu
,
7317 struct sched_entity
*parent
)
7319 struct rq
*rq
= cpu_rq(cpu
);
7323 init_cfs_rq_runtime(cfs_rq
);
7325 tg
->cfs_rq
[cpu
] = cfs_rq
;
7328 /* se could be NULL for root_task_group */
7333 se
->cfs_rq
= &rq
->cfs
;
7335 se
->cfs_rq
= parent
->my_q
;
7338 /* guarantee group entities always have weight */
7339 update_load_set(&se
->load
, NICE_0_LOAD
);
7340 se
->parent
= parent
;
7343 static DEFINE_MUTEX(shares_mutex
);
7345 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7348 unsigned long flags
;
7351 * We can't change the weight of the root cgroup.
7356 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
7358 mutex_lock(&shares_mutex
);
7359 if (tg
->shares
== shares
)
7362 tg
->shares
= shares
;
7363 for_each_possible_cpu(i
) {
7364 struct rq
*rq
= cpu_rq(i
);
7365 struct sched_entity
*se
;
7368 /* Propagate contribution to hierarchy */
7369 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7371 /* Possible calls to update_curr() need rq clock */
7372 update_rq_clock(rq
);
7373 for_each_sched_entity(se
)
7374 update_cfs_shares(group_cfs_rq(se
));
7375 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7379 mutex_unlock(&shares_mutex
);
7382 #else /* CONFIG_FAIR_GROUP_SCHED */
7384 void free_fair_sched_group(struct task_group
*tg
) { }
7386 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7391 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
7393 #endif /* CONFIG_FAIR_GROUP_SCHED */
7396 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
7398 struct sched_entity
*se
= &task
->se
;
7399 unsigned int rr_interval
= 0;
7402 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7405 if (rq
->cfs
.load
.weight
)
7406 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
7412 * All the scheduling class methods:
7414 const struct sched_class fair_sched_class
= {
7415 .next
= &idle_sched_class
,
7416 .enqueue_task
= enqueue_task_fair
,
7417 .dequeue_task
= dequeue_task_fair
,
7418 .yield_task
= yield_task_fair
,
7419 .yield_to_task
= yield_to_task_fair
,
7421 .check_preempt_curr
= check_preempt_wakeup
,
7423 .pick_next_task
= pick_next_task_fair
,
7424 .put_prev_task
= put_prev_task_fair
,
7427 .select_task_rq
= select_task_rq_fair
,
7428 .migrate_task_rq
= migrate_task_rq_fair
,
7430 .rq_online
= rq_online_fair
,
7431 .rq_offline
= rq_offline_fair
,
7433 .task_waking
= task_waking_fair
,
7436 .set_curr_task
= set_curr_task_fair
,
7437 .task_tick
= task_tick_fair
,
7438 .task_fork
= task_fork_fair
,
7440 .prio_changed
= prio_changed_fair
,
7441 .switched_from
= switched_from_fair
,
7442 .switched_to
= switched_to_fair
,
7444 .get_rr_interval
= get_rr_interval_fair
,
7446 #ifdef CONFIG_FAIR_GROUP_SCHED
7447 .task_move_group
= task_move_group_fair
,
7451 #ifdef CONFIG_SCHED_DEBUG
7452 void print_cfs_stats(struct seq_file
*m
, int cpu
)
7454 struct cfs_rq
*cfs_rq
;
7457 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
7458 print_cfs_rq(m
, cpu
, cfs_rq
);
7463 __init
void init_sched_fair_class(void)
7466 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7468 #ifdef CONFIG_NO_HZ_COMMON
7469 nohz
.next_balance
= jiffies
;
7470 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7471 cpu_notifier(sched_ilb_notifier
, 0);