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/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency
= 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG
;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity
= 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency
= 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly
;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
94 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window
= 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
117 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
123 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
129 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static int get_update_sysctl_factor(void)
146 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
149 switch (sysctl_sched_tunable_scaling
) {
150 case SCHED_TUNABLESCALING_NONE
:
153 case SCHED_TUNABLESCALING_LINEAR
:
156 case SCHED_TUNABLESCALING_LOG
:
158 factor
= 1 + ilog2(cpus
);
165 static void update_sysctl(void)
167 unsigned int factor
= get_update_sysctl_factor();
169 #define SET_SYSCTL(name) \
170 (sysctl_##name = (factor) * normalized_sysctl_##name)
171 SET_SYSCTL(sched_min_granularity
);
172 SET_SYSCTL(sched_latency
);
173 SET_SYSCTL(sched_wakeup_granularity
);
177 void sched_init_granularity(void)
182 #define WMULT_CONST (~0U)
183 #define WMULT_SHIFT 32
185 static void __update_inv_weight(struct load_weight
*lw
)
189 if (likely(lw
->inv_weight
))
192 w
= scale_load_down(lw
->weight
);
194 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
196 else if (unlikely(!w
))
197 lw
->inv_weight
= WMULT_CONST
;
199 lw
->inv_weight
= WMULT_CONST
/ w
;
203 * delta_exec * weight / lw.weight
205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
207 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
208 * we're guaranteed shift stays positive because inv_weight is guaranteed to
209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212 * weight/lw.weight <= 1, and therefore our shift will also be positive.
214 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
216 u64 fact
= scale_load_down(weight
);
217 int shift
= WMULT_SHIFT
;
219 __update_inv_weight(lw
);
221 if (unlikely(fact
>> 32)) {
228 /* hint to use a 32x32->64 mul */
229 fact
= (u64
)(u32
)fact
* lw
->inv_weight
;
236 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
240 const struct sched_class fair_sched_class
;
242 /**************************************************************
243 * CFS operations on generic schedulable entities:
246 #ifdef CONFIG_FAIR_GROUP_SCHED
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se) (!se->my_q)
257 static inline struct task_struct
*task_of(struct sched_entity
*se
)
259 #ifdef CONFIG_SCHED_DEBUG
260 WARN_ON_ONCE(!entity_is_task(se
));
262 return container_of(se
, struct task_struct
, se
);
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 for (; se; se = se->parent)
269 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
286 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
289 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
291 if (!cfs_rq
->on_list
) {
293 * Ensure we either appear before our parent (if already
294 * enqueued) or force our parent to appear after us when it is
295 * enqueued. The fact that we always enqueue bottom-up
296 * reduces this to two cases.
298 if (cfs_rq
->tg
->parent
&&
299 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
300 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
301 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
303 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
304 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
308 /* We should have no load, but we need to update last_decay. */
309 update_cfs_rq_blocked_load(cfs_rq
, 0);
313 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
315 if (cfs_rq
->on_list
) {
316 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
321 /* Iterate thr' all leaf cfs_rq's on a runqueue */
322 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
323 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
325 /* Do the two (enqueued) entities belong to the same group ? */
326 static inline struct cfs_rq
*
327 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
329 if (se
->cfs_rq
== pse
->cfs_rq
)
335 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
341 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
343 int se_depth
, pse_depth
;
346 * preemption test can be made between sibling entities who are in the
347 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
348 * both tasks until we find their ancestors who are siblings of common
352 /* First walk up until both entities are at same depth */
353 se_depth
= (*se
)->depth
;
354 pse_depth
= (*pse
)->depth
;
356 while (se_depth
> pse_depth
) {
358 *se
= parent_entity(*se
);
361 while (pse_depth
> se_depth
) {
363 *pse
= parent_entity(*pse
);
366 while (!is_same_group(*se
, *pse
)) {
367 *se
= parent_entity(*se
);
368 *pse
= parent_entity(*pse
);
372 #else /* !CONFIG_FAIR_GROUP_SCHED */
374 static inline struct task_struct
*task_of(struct sched_entity
*se
)
376 return container_of(se
, struct task_struct
, se
);
379 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
381 return container_of(cfs_rq
, struct rq
, cfs
);
384 #define entity_is_task(se) 1
386 #define for_each_sched_entity(se) \
387 for (; se; se = NULL)
389 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
391 return &task_rq(p
)->cfs
;
394 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
396 struct task_struct
*p
= task_of(se
);
397 struct rq
*rq
= task_rq(p
);
402 /* runqueue "owned" by this group */
403 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
408 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
412 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
416 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
417 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
419 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
425 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
429 #endif /* CONFIG_FAIR_GROUP_SCHED */
431 static __always_inline
432 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
434 /**************************************************************
435 * Scheduling class tree data structure manipulation methods:
438 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
440 s64 delta
= (s64
)(vruntime
- max_vruntime
);
442 max_vruntime
= vruntime
;
447 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
449 s64 delta
= (s64
)(vruntime
- min_vruntime
);
451 min_vruntime
= vruntime
;
456 static inline int entity_before(struct sched_entity
*a
,
457 struct sched_entity
*b
)
459 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
462 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
464 u64 vruntime
= cfs_rq
->min_vruntime
;
467 vruntime
= cfs_rq
->curr
->vruntime
;
469 if (cfs_rq
->rb_leftmost
) {
470 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
475 vruntime
= se
->vruntime
;
477 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
480 /* ensure we never gain time by being placed backwards. */
481 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
484 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
489 * Enqueue an entity into the rb-tree:
491 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
493 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
494 struct rb_node
*parent
= NULL
;
495 struct sched_entity
*entry
;
499 * Find the right place in the rbtree:
503 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
505 * We dont care about collisions. Nodes with
506 * the same key stay together.
508 if (entity_before(se
, entry
)) {
509 link
= &parent
->rb_left
;
511 link
= &parent
->rb_right
;
517 * Maintain a cache of leftmost tree entries (it is frequently
521 cfs_rq
->rb_leftmost
= &se
->run_node
;
523 rb_link_node(&se
->run_node
, parent
, link
);
524 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
527 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
529 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
530 struct rb_node
*next_node
;
532 next_node
= rb_next(&se
->run_node
);
533 cfs_rq
->rb_leftmost
= next_node
;
536 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
539 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
541 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
546 return rb_entry(left
, struct sched_entity
, run_node
);
549 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
551 struct rb_node
*next
= rb_next(&se
->run_node
);
556 return rb_entry(next
, struct sched_entity
, run_node
);
559 #ifdef CONFIG_SCHED_DEBUG
560 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
562 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
567 return rb_entry(last
, struct sched_entity
, run_node
);
570 /**************************************************************
571 * Scheduling class statistics methods:
574 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
575 void __user
*buffer
, size_t *lenp
,
578 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
579 int factor
= get_update_sysctl_factor();
584 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
585 sysctl_sched_min_granularity
);
587 #define WRT_SYSCTL(name) \
588 (normalized_sysctl_##name = sysctl_##name / (factor))
589 WRT_SYSCTL(sched_min_granularity
);
590 WRT_SYSCTL(sched_latency
);
591 WRT_SYSCTL(sched_wakeup_granularity
);
601 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
603 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
604 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
610 * The idea is to set a period in which each task runs once.
612 * When there are too many tasks (sched_nr_latency) we have to stretch
613 * this period because otherwise the slices get too small.
615 * p = (nr <= nl) ? l : l*nr/nl
617 static u64
__sched_period(unsigned long nr_running
)
619 u64 period
= sysctl_sched_latency
;
620 unsigned long nr_latency
= sched_nr_latency
;
622 if (unlikely(nr_running
> nr_latency
)) {
623 period
= sysctl_sched_min_granularity
;
624 period
*= nr_running
;
631 * We calculate the wall-time slice from the period by taking a part
632 * proportional to the weight.
636 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
638 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
640 for_each_sched_entity(se
) {
641 struct load_weight
*load
;
642 struct load_weight lw
;
644 cfs_rq
= cfs_rq_of(se
);
645 load
= &cfs_rq
->load
;
647 if (unlikely(!se
->on_rq
)) {
650 update_load_add(&lw
, se
->load
.weight
);
653 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
659 * We calculate the vruntime slice of a to-be-inserted task.
663 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
665 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
669 static int select_idle_sibling(struct task_struct
*p
, int cpu
);
670 static unsigned long task_h_load(struct task_struct
*p
);
672 static inline void __update_task_entity_contrib(struct sched_entity
*se
);
674 /* Give new task start runnable values to heavy its load in infant time */
675 void init_task_runnable_average(struct task_struct
*p
)
679 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
680 p
->se
.avg
.runnable_avg_sum
= slice
;
681 p
->se
.avg
.runnable_avg_period
= slice
;
682 __update_task_entity_contrib(&p
->se
);
685 void init_task_runnable_average(struct task_struct
*p
)
691 * Update the current task's runtime statistics.
693 static void update_curr(struct cfs_rq
*cfs_rq
)
695 struct sched_entity
*curr
= cfs_rq
->curr
;
696 u64 now
= rq_clock_task(rq_of(cfs_rq
));
702 delta_exec
= now
- curr
->exec_start
;
703 if (unlikely((s64
)delta_exec
<= 0))
706 curr
->exec_start
= now
;
708 schedstat_set(curr
->statistics
.exec_max
,
709 max(delta_exec
, curr
->statistics
.exec_max
));
711 curr
->sum_exec_runtime
+= delta_exec
;
712 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
714 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
715 update_min_vruntime(cfs_rq
);
717 if (entity_is_task(curr
)) {
718 struct task_struct
*curtask
= task_of(curr
);
720 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
721 cpuacct_charge(curtask
, delta_exec
);
722 account_group_exec_runtime(curtask
, delta_exec
);
725 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
728 static void update_curr_fair(struct rq
*rq
)
730 update_curr(cfs_rq_of(&rq
->curr
->se
));
734 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
736 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
740 * Task is being enqueued - update stats:
742 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
745 * Are we enqueueing a waiting task? (for current tasks
746 * a dequeue/enqueue event is a NOP)
748 if (se
!= cfs_rq
->curr
)
749 update_stats_wait_start(cfs_rq
, se
);
753 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
755 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
756 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
757 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
758 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
759 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
760 #ifdef CONFIG_SCHEDSTATS
761 if (entity_is_task(se
)) {
762 trace_sched_stat_wait(task_of(se
),
763 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
766 schedstat_set(se
->statistics
.wait_start
, 0);
770 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
773 * Mark the end of the wait period if dequeueing a
776 if (se
!= cfs_rq
->curr
)
777 update_stats_wait_end(cfs_rq
, se
);
781 * We are picking a new current task - update its stats:
784 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
787 * We are starting a new run period:
789 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
792 /**************************************************
793 * Scheduling class queueing methods:
796 #ifdef CONFIG_NUMA_BALANCING
798 * Approximate time to scan a full NUMA task in ms. The task scan period is
799 * calculated based on the tasks virtual memory size and
800 * numa_balancing_scan_size.
802 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
803 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
805 /* Portion of address space to scan in MB */
806 unsigned int sysctl_numa_balancing_scan_size
= 256;
808 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
809 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
811 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
813 unsigned long rss
= 0;
814 unsigned long nr_scan_pages
;
817 * Calculations based on RSS as non-present and empty pages are skipped
818 * by the PTE scanner and NUMA hinting faults should be trapped based
821 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
822 rss
= get_mm_rss(p
->mm
);
826 rss
= round_up(rss
, nr_scan_pages
);
827 return rss
/ nr_scan_pages
;
830 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
831 #define MAX_SCAN_WINDOW 2560
833 static unsigned int task_scan_min(struct task_struct
*p
)
835 unsigned int scan_size
= ACCESS_ONCE(sysctl_numa_balancing_scan_size
);
836 unsigned int scan
, floor
;
837 unsigned int windows
= 1;
839 if (scan_size
< MAX_SCAN_WINDOW
)
840 windows
= MAX_SCAN_WINDOW
/ scan_size
;
841 floor
= 1000 / windows
;
843 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
844 return max_t(unsigned int, floor
, scan
);
847 static unsigned int task_scan_max(struct task_struct
*p
)
849 unsigned int smin
= task_scan_min(p
);
852 /* Watch for min being lower than max due to floor calculations */
853 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
854 return max(smin
, smax
);
857 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
859 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
860 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
863 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
865 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
866 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
872 spinlock_t lock
; /* nr_tasks, tasks */
877 nodemask_t active_nodes
;
878 unsigned long total_faults
;
880 * Faults_cpu is used to decide whether memory should move
881 * towards the CPU. As a consequence, these stats are weighted
882 * more by CPU use than by memory faults.
884 unsigned long *faults_cpu
;
885 unsigned long faults
[0];
888 /* Shared or private faults. */
889 #define NR_NUMA_HINT_FAULT_TYPES 2
891 /* Memory and CPU locality */
892 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
894 /* Averaged statistics, and temporary buffers. */
895 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
897 pid_t
task_numa_group_id(struct task_struct
*p
)
899 return p
->numa_group
? p
->numa_group
->gid
: 0;
903 * The averaged statistics, shared & private, memory & cpu,
904 * occupy the first half of the array. The second half of the
905 * array is for current counters, which are averaged into the
906 * first set by task_numa_placement.
908 static inline int task_faults_idx(enum numa_faults_stats s
, int nid
, int priv
)
910 return NR_NUMA_HINT_FAULT_TYPES
* (s
* nr_node_ids
+ nid
) + priv
;
913 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
918 return p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
919 p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
922 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
927 return p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
928 p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
931 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
933 return group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
934 group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 1)];
937 /* Handle placement on systems where not all nodes are directly connected. */
938 static unsigned long score_nearby_nodes(struct task_struct
*p
, int nid
,
939 int maxdist
, bool task
)
941 unsigned long score
= 0;
945 * All nodes are directly connected, and the same distance
946 * from each other. No need for fancy placement algorithms.
948 if (sched_numa_topology_type
== NUMA_DIRECT
)
952 * This code is called for each node, introducing N^2 complexity,
953 * which should be ok given the number of nodes rarely exceeds 8.
955 for_each_online_node(node
) {
956 unsigned long faults
;
957 int dist
= node_distance(nid
, node
);
960 * The furthest away nodes in the system are not interesting
961 * for placement; nid was already counted.
963 if (dist
== sched_max_numa_distance
|| node
== nid
)
967 * On systems with a backplane NUMA topology, compare groups
968 * of nodes, and move tasks towards the group with the most
969 * memory accesses. When comparing two nodes at distance
970 * "hoplimit", only nodes closer by than "hoplimit" are part
971 * of each group. Skip other nodes.
973 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
977 /* Add up the faults from nearby nodes. */
979 faults
= task_faults(p
, node
);
981 faults
= group_faults(p
, node
);
984 * On systems with a glueless mesh NUMA topology, there are
985 * no fixed "groups of nodes". Instead, nodes that are not
986 * directly connected bounce traffic through intermediate
987 * nodes; a numa_group can occupy any set of nodes.
988 * The further away a node is, the less the faults count.
989 * This seems to result in good task placement.
991 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
992 faults
*= (sched_max_numa_distance
- dist
);
993 faults
/= (sched_max_numa_distance
- LOCAL_DISTANCE
);
1003 * These return the fraction of accesses done by a particular task, or
1004 * task group, on a particular numa node. The group weight is given a
1005 * larger multiplier, in order to group tasks together that are almost
1006 * evenly spread out between numa nodes.
1008 static inline unsigned long task_weight(struct task_struct
*p
, int nid
,
1011 unsigned long faults
, total_faults
;
1013 if (!p
->numa_faults
)
1016 total_faults
= p
->total_numa_faults
;
1021 faults
= task_faults(p
, nid
);
1022 faults
+= score_nearby_nodes(p
, nid
, dist
, true);
1024 return 1000 * faults
/ total_faults
;
1027 static inline unsigned long group_weight(struct task_struct
*p
, int nid
,
1030 unsigned long faults
, total_faults
;
1035 total_faults
= p
->numa_group
->total_faults
;
1040 faults
= group_faults(p
, nid
);
1041 faults
+= score_nearby_nodes(p
, nid
, dist
, false);
1043 return 1000 * faults
/ total_faults
;
1046 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
1047 int src_nid
, int dst_cpu
)
1049 struct numa_group
*ng
= p
->numa_group
;
1050 int dst_nid
= cpu_to_node(dst_cpu
);
1051 int last_cpupid
, this_cpupid
;
1053 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
1056 * Multi-stage node selection is used in conjunction with a periodic
1057 * migration fault to build a temporal task<->page relation. By using
1058 * a two-stage filter we remove short/unlikely relations.
1060 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1061 * a task's usage of a particular page (n_p) per total usage of this
1062 * page (n_t) (in a given time-span) to a probability.
1064 * Our periodic faults will sample this probability and getting the
1065 * same result twice in a row, given these samples are fully
1066 * independent, is then given by P(n)^2, provided our sample period
1067 * is sufficiently short compared to the usage pattern.
1069 * This quadric squishes small probabilities, making it less likely we
1070 * act on an unlikely task<->page relation.
1072 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
1073 if (!cpupid_pid_unset(last_cpupid
) &&
1074 cpupid_to_nid(last_cpupid
) != dst_nid
)
1077 /* Always allow migrate on private faults */
1078 if (cpupid_match_pid(p
, last_cpupid
))
1081 /* A shared fault, but p->numa_group has not been set up yet. */
1086 * Do not migrate if the destination is not a node that
1087 * is actively used by this numa group.
1089 if (!node_isset(dst_nid
, ng
->active_nodes
))
1093 * Source is a node that is not actively used by this
1094 * numa group, while the destination is. Migrate.
1096 if (!node_isset(src_nid
, ng
->active_nodes
))
1100 * Both source and destination are nodes in active
1101 * use by this numa group. Maximize memory bandwidth
1102 * by migrating from more heavily used groups, to less
1103 * heavily used ones, spreading the load around.
1104 * Use a 1/4 hysteresis to avoid spurious page movement.
1106 return group_faults(p
, dst_nid
) < (group_faults(p
, src_nid
) * 3 / 4);
1109 static unsigned long weighted_cpuload(const int cpu
);
1110 static unsigned long source_load(int cpu
, int type
);
1111 static unsigned long target_load(int cpu
, int type
);
1112 static unsigned long capacity_of(int cpu
);
1113 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
1115 /* Cached statistics for all CPUs within a node */
1117 unsigned long nr_running
;
1120 /* Total compute capacity of CPUs on a node */
1121 unsigned long compute_capacity
;
1123 /* Approximate capacity in terms of runnable tasks on a node */
1124 unsigned long task_capacity
;
1125 int has_free_capacity
;
1129 * XXX borrowed from update_sg_lb_stats
1131 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1133 int smt
, cpu
, cpus
= 0;
1134 unsigned long capacity
;
1136 memset(ns
, 0, sizeof(*ns
));
1137 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1138 struct rq
*rq
= cpu_rq(cpu
);
1140 ns
->nr_running
+= rq
->nr_running
;
1141 ns
->load
+= weighted_cpuload(cpu
);
1142 ns
->compute_capacity
+= capacity_of(cpu
);
1148 * If we raced with hotplug and there are no CPUs left in our mask
1149 * the @ns structure is NULL'ed and task_numa_compare() will
1150 * not find this node attractive.
1152 * We'll either bail at !has_free_capacity, or we'll detect a huge
1153 * imbalance and bail there.
1158 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1159 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, ns
->compute_capacity
);
1160 capacity
= cpus
/ smt
; /* cores */
1162 ns
->task_capacity
= min_t(unsigned, capacity
,
1163 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
));
1164 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1167 struct task_numa_env
{
1168 struct task_struct
*p
;
1170 int src_cpu
, src_nid
;
1171 int dst_cpu
, dst_nid
;
1173 struct numa_stats src_stats
, dst_stats
;
1178 struct task_struct
*best_task
;
1183 static void task_numa_assign(struct task_numa_env
*env
,
1184 struct task_struct
*p
, long imp
)
1187 put_task_struct(env
->best_task
);
1192 env
->best_imp
= imp
;
1193 env
->best_cpu
= env
->dst_cpu
;
1196 static bool load_too_imbalanced(long src_load
, long dst_load
,
1197 struct task_numa_env
*env
)
1200 long orig_src_load
, orig_dst_load
;
1201 long src_capacity
, dst_capacity
;
1204 * The load is corrected for the CPU capacity available on each node.
1207 * ------------ vs ---------
1208 * src_capacity dst_capacity
1210 src_capacity
= env
->src_stats
.compute_capacity
;
1211 dst_capacity
= env
->dst_stats
.compute_capacity
;
1213 /* We care about the slope of the imbalance, not the direction. */
1214 if (dst_load
< src_load
)
1215 swap(dst_load
, src_load
);
1217 /* Is the difference below the threshold? */
1218 imb
= dst_load
* src_capacity
* 100 -
1219 src_load
* dst_capacity
* env
->imbalance_pct
;
1224 * The imbalance is above the allowed threshold.
1225 * Compare it with the old imbalance.
1227 orig_src_load
= env
->src_stats
.load
;
1228 orig_dst_load
= env
->dst_stats
.load
;
1230 if (orig_dst_load
< orig_src_load
)
1231 swap(orig_dst_load
, orig_src_load
);
1233 old_imb
= orig_dst_load
* src_capacity
* 100 -
1234 orig_src_load
* dst_capacity
* env
->imbalance_pct
;
1236 /* Would this change make things worse? */
1237 return (imb
> old_imb
);
1241 * This checks if the overall compute and NUMA accesses of the system would
1242 * be improved if the source tasks was migrated to the target dst_cpu taking
1243 * into account that it might be best if task running on the dst_cpu should
1244 * be exchanged with the source task
1246 static void task_numa_compare(struct task_numa_env
*env
,
1247 long taskimp
, long groupimp
)
1249 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1250 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1251 struct task_struct
*cur
;
1252 long src_load
, dst_load
;
1254 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1256 int dist
= env
->dist
;
1260 raw_spin_lock_irq(&dst_rq
->lock
);
1263 * No need to move the exiting task, and this ensures that ->curr
1264 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1265 * is safe under RCU read lock.
1266 * Note that rcu_read_lock() itself can't protect from the final
1267 * put_task_struct() after the last schedule().
1269 if ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
))
1271 raw_spin_unlock_irq(&dst_rq
->lock
);
1274 * Because we have preemption enabled we can get migrated around and
1275 * end try selecting ourselves (current == env->p) as a swap candidate.
1281 * "imp" is the fault differential for the source task between the
1282 * source and destination node. Calculate the total differential for
1283 * the source task and potential destination task. The more negative
1284 * the value is, the more rmeote accesses that would be expected to
1285 * be incurred if the tasks were swapped.
1288 /* Skip this swap candidate if cannot move to the source cpu */
1289 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1293 * If dst and source tasks are in the same NUMA group, or not
1294 * in any group then look only at task weights.
1296 if (cur
->numa_group
== env
->p
->numa_group
) {
1297 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1298 task_weight(cur
, env
->dst_nid
, dist
);
1300 * Add some hysteresis to prevent swapping the
1301 * tasks within a group over tiny differences.
1303 if (cur
->numa_group
)
1307 * Compare the group weights. If a task is all by
1308 * itself (not part of a group), use the task weight
1311 if (cur
->numa_group
)
1312 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1313 group_weight(cur
, env
->dst_nid
, dist
);
1315 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1316 task_weight(cur
, env
->dst_nid
, dist
);
1320 if (imp
<= env
->best_imp
&& moveimp
<= env
->best_imp
)
1324 /* Is there capacity at our destination? */
1325 if (env
->src_stats
.nr_running
<= env
->src_stats
.task_capacity
&&
1326 !env
->dst_stats
.has_free_capacity
)
1332 /* Balance doesn't matter much if we're running a task per cpu */
1333 if (imp
> env
->best_imp
&& src_rq
->nr_running
== 1 &&
1334 dst_rq
->nr_running
== 1)
1338 * In the overloaded case, try and keep the load balanced.
1341 load
= task_h_load(env
->p
);
1342 dst_load
= env
->dst_stats
.load
+ load
;
1343 src_load
= env
->src_stats
.load
- load
;
1345 if (moveimp
> imp
&& moveimp
> env
->best_imp
) {
1347 * If the improvement from just moving env->p direction is
1348 * better than swapping tasks around, check if a move is
1349 * possible. Store a slightly smaller score than moveimp,
1350 * so an actually idle CPU will win.
1352 if (!load_too_imbalanced(src_load
, dst_load
, env
)) {
1359 if (imp
<= env
->best_imp
)
1363 load
= task_h_load(cur
);
1368 if (load_too_imbalanced(src_load
, dst_load
, env
))
1372 * One idle CPU per node is evaluated for a task numa move.
1373 * Call select_idle_sibling to maybe find a better one.
1376 env
->dst_cpu
= select_idle_sibling(env
->p
, env
->dst_cpu
);
1379 task_numa_assign(env
, cur
, imp
);
1384 static void task_numa_find_cpu(struct task_numa_env
*env
,
1385 long taskimp
, long groupimp
)
1389 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1390 /* Skip this CPU if the source task cannot migrate */
1391 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1395 task_numa_compare(env
, taskimp
, groupimp
);
1399 static int task_numa_migrate(struct task_struct
*p
)
1401 struct task_numa_env env
= {
1404 .src_cpu
= task_cpu(p
),
1405 .src_nid
= task_node(p
),
1407 .imbalance_pct
= 112,
1413 struct sched_domain
*sd
;
1414 unsigned long taskweight
, groupweight
;
1416 long taskimp
, groupimp
;
1419 * Pick the lowest SD_NUMA domain, as that would have the smallest
1420 * imbalance and would be the first to start moving tasks about.
1422 * And we want to avoid any moving of tasks about, as that would create
1423 * random movement of tasks -- counter the numa conditions we're trying
1427 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1429 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1433 * Cpusets can break the scheduler domain tree into smaller
1434 * balance domains, some of which do not cross NUMA boundaries.
1435 * Tasks that are "trapped" in such domains cannot be migrated
1436 * elsewhere, so there is no point in (re)trying.
1438 if (unlikely(!sd
)) {
1439 p
->numa_preferred_nid
= task_node(p
);
1443 env
.dst_nid
= p
->numa_preferred_nid
;
1444 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1445 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1446 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1447 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1448 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
1449 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
1450 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1452 /* Try to find a spot on the preferred nid. */
1453 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1456 * Look at other nodes in these cases:
1457 * - there is no space available on the preferred_nid
1458 * - the task is part of a numa_group that is interleaved across
1459 * multiple NUMA nodes; in order to better consolidate the group,
1460 * we need to check other locations.
1462 if (env
.best_cpu
== -1 || (p
->numa_group
&&
1463 nodes_weight(p
->numa_group
->active_nodes
) > 1)) {
1464 for_each_online_node(nid
) {
1465 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1468 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1469 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1471 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1472 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1475 /* Only consider nodes where both task and groups benefit */
1476 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
1477 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
1478 if (taskimp
< 0 && groupimp
< 0)
1483 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1484 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1489 * If the task is part of a workload that spans multiple NUMA nodes,
1490 * and is migrating into one of the workload's active nodes, remember
1491 * this node as the task's preferred numa node, so the workload can
1493 * A task that migrated to a second choice node will be better off
1494 * trying for a better one later. Do not set the preferred node here.
1496 if (p
->numa_group
) {
1497 if (env
.best_cpu
== -1)
1502 if (node_isset(nid
, p
->numa_group
->active_nodes
))
1503 sched_setnuma(p
, env
.dst_nid
);
1506 /* No better CPU than the current one was found. */
1507 if (env
.best_cpu
== -1)
1511 * Reset the scan period if the task is being rescheduled on an
1512 * alternative node to recheck if the tasks is now properly placed.
1514 p
->numa_scan_period
= task_scan_min(p
);
1516 if (env
.best_task
== NULL
) {
1517 ret
= migrate_task_to(p
, env
.best_cpu
);
1519 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1523 ret
= migrate_swap(p
, env
.best_task
);
1525 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1526 put_task_struct(env
.best_task
);
1530 /* Attempt to migrate a task to a CPU on the preferred node. */
1531 static void numa_migrate_preferred(struct task_struct
*p
)
1533 unsigned long interval
= HZ
;
1535 /* This task has no NUMA fault statistics yet */
1536 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults
))
1539 /* Periodically retry migrating the task to the preferred node */
1540 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1541 p
->numa_migrate_retry
= jiffies
+ interval
;
1543 /* Success if task is already running on preferred CPU */
1544 if (task_node(p
) == p
->numa_preferred_nid
)
1547 /* Otherwise, try migrate to a CPU on the preferred node */
1548 task_numa_migrate(p
);
1552 * Find the nodes on which the workload is actively running. We do this by
1553 * tracking the nodes from which NUMA hinting faults are triggered. This can
1554 * be different from the set of nodes where the workload's memory is currently
1557 * The bitmask is used to make smarter decisions on when to do NUMA page
1558 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1559 * are added when they cause over 6/16 of the maximum number of faults, but
1560 * only removed when they drop below 3/16.
1562 static void update_numa_active_node_mask(struct numa_group
*numa_group
)
1564 unsigned long faults
, max_faults
= 0;
1567 for_each_online_node(nid
) {
1568 faults
= group_faults_cpu(numa_group
, nid
);
1569 if (faults
> max_faults
)
1570 max_faults
= faults
;
1573 for_each_online_node(nid
) {
1574 faults
= group_faults_cpu(numa_group
, nid
);
1575 if (!node_isset(nid
, numa_group
->active_nodes
)) {
1576 if (faults
> max_faults
* 6 / 16)
1577 node_set(nid
, numa_group
->active_nodes
);
1578 } else if (faults
< max_faults
* 3 / 16)
1579 node_clear(nid
, numa_group
->active_nodes
);
1584 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1585 * increments. The more local the fault statistics are, the higher the scan
1586 * period will be for the next scan window. If local/(local+remote) ratio is
1587 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1588 * the scan period will decrease. Aim for 70% local accesses.
1590 #define NUMA_PERIOD_SLOTS 10
1591 #define NUMA_PERIOD_THRESHOLD 7
1594 * Increase the scan period (slow down scanning) if the majority of
1595 * our memory is already on our local node, or if the majority of
1596 * the page accesses are shared with other processes.
1597 * Otherwise, decrease the scan period.
1599 static void update_task_scan_period(struct task_struct
*p
,
1600 unsigned long shared
, unsigned long private)
1602 unsigned int period_slot
;
1606 unsigned long remote
= p
->numa_faults_locality
[0];
1607 unsigned long local
= p
->numa_faults_locality
[1];
1610 * If there were no record hinting faults then either the task is
1611 * completely idle or all activity is areas that are not of interest
1612 * to automatic numa balancing. Related to that, if there were failed
1613 * migration then it implies we are migrating too quickly or the local
1614 * node is overloaded. In either case, scan slower
1616 if (local
+ shared
== 0 || p
->numa_faults_locality
[2]) {
1617 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1618 p
->numa_scan_period
<< 1);
1620 p
->mm
->numa_next_scan
= jiffies
+
1621 msecs_to_jiffies(p
->numa_scan_period
);
1627 * Prepare to scale scan period relative to the current period.
1628 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1629 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1630 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1632 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1633 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1634 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1635 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1638 diff
= slot
* period_slot
;
1640 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1643 * Scale scan rate increases based on sharing. There is an
1644 * inverse relationship between the degree of sharing and
1645 * the adjustment made to the scanning period. Broadly
1646 * speaking the intent is that there is little point
1647 * scanning faster if shared accesses dominate as it may
1648 * simply bounce migrations uselessly
1650 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
+ 1));
1651 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1654 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1655 task_scan_min(p
), task_scan_max(p
));
1656 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1660 * Get the fraction of time the task has been running since the last
1661 * NUMA placement cycle. The scheduler keeps similar statistics, but
1662 * decays those on a 32ms period, which is orders of magnitude off
1663 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1664 * stats only if the task is so new there are no NUMA statistics yet.
1666 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1668 u64 runtime
, delta
, now
;
1669 /* Use the start of this time slice to avoid calculations. */
1670 now
= p
->se
.exec_start
;
1671 runtime
= p
->se
.sum_exec_runtime
;
1673 if (p
->last_task_numa_placement
) {
1674 delta
= runtime
- p
->last_sum_exec_runtime
;
1675 *period
= now
- p
->last_task_numa_placement
;
1677 delta
= p
->se
.avg
.runnable_avg_sum
;
1678 *period
= p
->se
.avg
.runnable_avg_period
;
1681 p
->last_sum_exec_runtime
= runtime
;
1682 p
->last_task_numa_placement
= now
;
1688 * Determine the preferred nid for a task in a numa_group. This needs to
1689 * be done in a way that produces consistent results with group_weight,
1690 * otherwise workloads might not converge.
1692 static int preferred_group_nid(struct task_struct
*p
, int nid
)
1697 /* Direct connections between all NUMA nodes. */
1698 if (sched_numa_topology_type
== NUMA_DIRECT
)
1702 * On a system with glueless mesh NUMA topology, group_weight
1703 * scores nodes according to the number of NUMA hinting faults on
1704 * both the node itself, and on nearby nodes.
1706 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1707 unsigned long score
, max_score
= 0;
1708 int node
, max_node
= nid
;
1710 dist
= sched_max_numa_distance
;
1712 for_each_online_node(node
) {
1713 score
= group_weight(p
, node
, dist
);
1714 if (score
> max_score
) {
1723 * Finding the preferred nid in a system with NUMA backplane
1724 * interconnect topology is more involved. The goal is to locate
1725 * tasks from numa_groups near each other in the system, and
1726 * untangle workloads from different sides of the system. This requires
1727 * searching down the hierarchy of node groups, recursively searching
1728 * inside the highest scoring group of nodes. The nodemask tricks
1729 * keep the complexity of the search down.
1731 nodes
= node_online_map
;
1732 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
1733 unsigned long max_faults
= 0;
1734 nodemask_t max_group
= NODE_MASK_NONE
;
1737 /* Are there nodes at this distance from each other? */
1738 if (!find_numa_distance(dist
))
1741 for_each_node_mask(a
, nodes
) {
1742 unsigned long faults
= 0;
1743 nodemask_t this_group
;
1744 nodes_clear(this_group
);
1746 /* Sum group's NUMA faults; includes a==b case. */
1747 for_each_node_mask(b
, nodes
) {
1748 if (node_distance(a
, b
) < dist
) {
1749 faults
+= group_faults(p
, b
);
1750 node_set(b
, this_group
);
1751 node_clear(b
, nodes
);
1755 /* Remember the top group. */
1756 if (faults
> max_faults
) {
1757 max_faults
= faults
;
1758 max_group
= this_group
;
1760 * subtle: at the smallest distance there is
1761 * just one node left in each "group", the
1762 * winner is the preferred nid.
1767 /* Next round, evaluate the nodes within max_group. */
1773 static void task_numa_placement(struct task_struct
*p
)
1775 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1776 unsigned long max_faults
= 0, max_group_faults
= 0;
1777 unsigned long fault_types
[2] = { 0, 0 };
1778 unsigned long total_faults
;
1779 u64 runtime
, period
;
1780 spinlock_t
*group_lock
= NULL
;
1782 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1783 if (p
->numa_scan_seq
== seq
)
1785 p
->numa_scan_seq
= seq
;
1786 p
->numa_scan_period_max
= task_scan_max(p
);
1788 total_faults
= p
->numa_faults_locality
[0] +
1789 p
->numa_faults_locality
[1];
1790 runtime
= numa_get_avg_runtime(p
, &period
);
1792 /* If the task is part of a group prevent parallel updates to group stats */
1793 if (p
->numa_group
) {
1794 group_lock
= &p
->numa_group
->lock
;
1795 spin_lock_irq(group_lock
);
1798 /* Find the node with the highest number of faults */
1799 for_each_online_node(nid
) {
1800 /* Keep track of the offsets in numa_faults array */
1801 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
1802 unsigned long faults
= 0, group_faults
= 0;
1805 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
1806 long diff
, f_diff
, f_weight
;
1808 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
1809 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
1810 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
1811 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
1813 /* Decay existing window, copy faults since last scan */
1814 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
1815 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
1816 p
->numa_faults
[membuf_idx
] = 0;
1819 * Normalize the faults_from, so all tasks in a group
1820 * count according to CPU use, instead of by the raw
1821 * number of faults. Tasks with little runtime have
1822 * little over-all impact on throughput, and thus their
1823 * faults are less important.
1825 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
1826 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
1828 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
1829 p
->numa_faults
[cpubuf_idx
] = 0;
1831 p
->numa_faults
[mem_idx
] += diff
;
1832 p
->numa_faults
[cpu_idx
] += f_diff
;
1833 faults
+= p
->numa_faults
[mem_idx
];
1834 p
->total_numa_faults
+= diff
;
1835 if (p
->numa_group
) {
1837 * safe because we can only change our own group
1839 * mem_idx represents the offset for a given
1840 * nid and priv in a specific region because it
1841 * is at the beginning of the numa_faults array.
1843 p
->numa_group
->faults
[mem_idx
] += diff
;
1844 p
->numa_group
->faults_cpu
[mem_idx
] += f_diff
;
1845 p
->numa_group
->total_faults
+= diff
;
1846 group_faults
+= p
->numa_group
->faults
[mem_idx
];
1850 if (faults
> max_faults
) {
1851 max_faults
= faults
;
1855 if (group_faults
> max_group_faults
) {
1856 max_group_faults
= group_faults
;
1857 max_group_nid
= nid
;
1861 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1863 if (p
->numa_group
) {
1864 update_numa_active_node_mask(p
->numa_group
);
1865 spin_unlock_irq(group_lock
);
1866 max_nid
= preferred_group_nid(p
, max_group_nid
);
1870 /* Set the new preferred node */
1871 if (max_nid
!= p
->numa_preferred_nid
)
1872 sched_setnuma(p
, max_nid
);
1874 if (task_node(p
) != p
->numa_preferred_nid
)
1875 numa_migrate_preferred(p
);
1879 static inline int get_numa_group(struct numa_group
*grp
)
1881 return atomic_inc_not_zero(&grp
->refcount
);
1884 static inline void put_numa_group(struct numa_group
*grp
)
1886 if (atomic_dec_and_test(&grp
->refcount
))
1887 kfree_rcu(grp
, rcu
);
1890 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1893 struct numa_group
*grp
, *my_grp
;
1894 struct task_struct
*tsk
;
1896 int cpu
= cpupid_to_cpu(cpupid
);
1899 if (unlikely(!p
->numa_group
)) {
1900 unsigned int size
= sizeof(struct numa_group
) +
1901 4*nr_node_ids
*sizeof(unsigned long);
1903 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1907 atomic_set(&grp
->refcount
, 1);
1908 spin_lock_init(&grp
->lock
);
1910 /* Second half of the array tracks nids where faults happen */
1911 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
1914 node_set(task_node(current
), grp
->active_nodes
);
1916 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1917 grp
->faults
[i
] = p
->numa_faults
[i
];
1919 grp
->total_faults
= p
->total_numa_faults
;
1922 rcu_assign_pointer(p
->numa_group
, grp
);
1926 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1928 if (!cpupid_match_pid(tsk
, cpupid
))
1931 grp
= rcu_dereference(tsk
->numa_group
);
1935 my_grp
= p
->numa_group
;
1940 * Only join the other group if its bigger; if we're the bigger group,
1941 * the other task will join us.
1943 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1947 * Tie-break on the grp address.
1949 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1952 /* Always join threads in the same process. */
1953 if (tsk
->mm
== current
->mm
)
1956 /* Simple filter to avoid false positives due to PID collisions */
1957 if (flags
& TNF_SHARED
)
1960 /* Update priv based on whether false sharing was detected */
1963 if (join
&& !get_numa_group(grp
))
1971 BUG_ON(irqs_disabled());
1972 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
1974 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
1975 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
1976 grp
->faults
[i
] += p
->numa_faults
[i
];
1978 my_grp
->total_faults
-= p
->total_numa_faults
;
1979 grp
->total_faults
+= p
->total_numa_faults
;
1984 spin_unlock(&my_grp
->lock
);
1985 spin_unlock_irq(&grp
->lock
);
1987 rcu_assign_pointer(p
->numa_group
, grp
);
1989 put_numa_group(my_grp
);
1997 void task_numa_free(struct task_struct
*p
)
1999 struct numa_group
*grp
= p
->numa_group
;
2000 void *numa_faults
= p
->numa_faults
;
2001 unsigned long flags
;
2005 spin_lock_irqsave(&grp
->lock
, flags
);
2006 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2007 grp
->faults
[i
] -= p
->numa_faults
[i
];
2008 grp
->total_faults
-= p
->total_numa_faults
;
2011 spin_unlock_irqrestore(&grp
->lock
, flags
);
2012 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2013 put_numa_group(grp
);
2016 p
->numa_faults
= NULL
;
2021 * Got a PROT_NONE fault for a page on @node.
2023 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2025 struct task_struct
*p
= current
;
2026 bool migrated
= flags
& TNF_MIGRATED
;
2027 int cpu_node
= task_node(current
);
2028 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2031 if (!numabalancing_enabled
)
2034 /* for example, ksmd faulting in a user's mm */
2038 /* Allocate buffer to track faults on a per-node basis */
2039 if (unlikely(!p
->numa_faults
)) {
2040 int size
= sizeof(*p
->numa_faults
) *
2041 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2043 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2044 if (!p
->numa_faults
)
2047 p
->total_numa_faults
= 0;
2048 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2052 * First accesses are treated as private, otherwise consider accesses
2053 * to be private if the accessing pid has not changed
2055 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2058 priv
= cpupid_match_pid(p
, last_cpupid
);
2059 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2060 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2064 * If a workload spans multiple NUMA nodes, a shared fault that
2065 * occurs wholly within the set of nodes that the workload is
2066 * actively using should be counted as local. This allows the
2067 * scan rate to slow down when a workload has settled down.
2069 if (!priv
&& !local
&& p
->numa_group
&&
2070 node_isset(cpu_node
, p
->numa_group
->active_nodes
) &&
2071 node_isset(mem_node
, p
->numa_group
->active_nodes
))
2074 task_numa_placement(p
);
2077 * Retry task to preferred node migration periodically, in case it
2078 * case it previously failed, or the scheduler moved us.
2080 if (time_after(jiffies
, p
->numa_migrate_retry
))
2081 numa_migrate_preferred(p
);
2084 p
->numa_pages_migrated
+= pages
;
2085 if (flags
& TNF_MIGRATE_FAIL
)
2086 p
->numa_faults_locality
[2] += pages
;
2088 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2089 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2090 p
->numa_faults_locality
[local
] += pages
;
2093 static void reset_ptenuma_scan(struct task_struct
*p
)
2095 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
2096 p
->mm
->numa_scan_offset
= 0;
2100 * The expensive part of numa migration is done from task_work context.
2101 * Triggered from task_tick_numa().
2103 void task_numa_work(struct callback_head
*work
)
2105 unsigned long migrate
, next_scan
, now
= jiffies
;
2106 struct task_struct
*p
= current
;
2107 struct mm_struct
*mm
= p
->mm
;
2108 struct vm_area_struct
*vma
;
2109 unsigned long start
, end
;
2110 unsigned long nr_pte_updates
= 0;
2113 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
2115 work
->next
= work
; /* protect against double add */
2117 * Who cares about NUMA placement when they're dying.
2119 * NOTE: make sure not to dereference p->mm before this check,
2120 * exit_task_work() happens _after_ exit_mm() so we could be called
2121 * without p->mm even though we still had it when we enqueued this
2124 if (p
->flags
& PF_EXITING
)
2127 if (!mm
->numa_next_scan
) {
2128 mm
->numa_next_scan
= now
+
2129 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2133 * Enforce maximal scan/migration frequency..
2135 migrate
= mm
->numa_next_scan
;
2136 if (time_before(now
, migrate
))
2139 if (p
->numa_scan_period
== 0) {
2140 p
->numa_scan_period_max
= task_scan_max(p
);
2141 p
->numa_scan_period
= task_scan_min(p
);
2144 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2145 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2149 * Delay this task enough that another task of this mm will likely win
2150 * the next time around.
2152 p
->node_stamp
+= 2 * TICK_NSEC
;
2154 start
= mm
->numa_scan_offset
;
2155 pages
= sysctl_numa_balancing_scan_size
;
2156 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2160 down_read(&mm
->mmap_sem
);
2161 vma
= find_vma(mm
, start
);
2163 reset_ptenuma_scan(p
);
2167 for (; vma
; vma
= vma
->vm_next
) {
2168 if (!vma_migratable(vma
) || !vma_policy_mof(vma
) ||
2169 is_vm_hugetlb_page(vma
)) {
2174 * Shared library pages mapped by multiple processes are not
2175 * migrated as it is expected they are cache replicated. Avoid
2176 * hinting faults in read-only file-backed mappings or the vdso
2177 * as migrating the pages will be of marginal benefit.
2180 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2184 * Skip inaccessible VMAs to avoid any confusion between
2185 * PROT_NONE and NUMA hinting ptes
2187 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
2191 start
= max(start
, vma
->vm_start
);
2192 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2193 end
= min(end
, vma
->vm_end
);
2194 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
2197 * Scan sysctl_numa_balancing_scan_size but ensure that
2198 * at least one PTE is updated so that unused virtual
2199 * address space is quickly skipped.
2202 pages
-= (end
- start
) >> PAGE_SHIFT
;
2209 } while (end
!= vma
->vm_end
);
2214 * It is possible to reach the end of the VMA list but the last few
2215 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2216 * would find the !migratable VMA on the next scan but not reset the
2217 * scanner to the start so check it now.
2220 mm
->numa_scan_offset
= start
;
2222 reset_ptenuma_scan(p
);
2223 up_read(&mm
->mmap_sem
);
2227 * Drive the periodic memory faults..
2229 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2231 struct callback_head
*work
= &curr
->numa_work
;
2235 * We don't care about NUMA placement if we don't have memory.
2237 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2241 * Using runtime rather than walltime has the dual advantage that
2242 * we (mostly) drive the selection from busy threads and that the
2243 * task needs to have done some actual work before we bother with
2246 now
= curr
->se
.sum_exec_runtime
;
2247 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2249 if (now
- curr
->node_stamp
> period
) {
2250 if (!curr
->node_stamp
)
2251 curr
->numa_scan_period
= task_scan_min(curr
);
2252 curr
->node_stamp
+= period
;
2254 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2255 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2256 task_work_add(curr
, work
, true);
2261 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2265 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2269 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2272 #endif /* CONFIG_NUMA_BALANCING */
2275 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2277 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2278 if (!parent_entity(se
))
2279 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2281 if (entity_is_task(se
)) {
2282 struct rq
*rq
= rq_of(cfs_rq
);
2284 account_numa_enqueue(rq
, task_of(se
));
2285 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2288 cfs_rq
->nr_running
++;
2292 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2294 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2295 if (!parent_entity(se
))
2296 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2297 if (entity_is_task(se
)) {
2298 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2299 list_del_init(&se
->group_node
);
2301 cfs_rq
->nr_running
--;
2304 #ifdef CONFIG_FAIR_GROUP_SCHED
2306 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
2311 * Use this CPU's actual weight instead of the last load_contribution
2312 * to gain a more accurate current total weight. See
2313 * update_cfs_rq_load_contribution().
2315 tg_weight
= atomic_long_read(&tg
->load_avg
);
2316 tg_weight
-= cfs_rq
->tg_load_contrib
;
2317 tg_weight
+= cfs_rq
->load
.weight
;
2322 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2324 long tg_weight
, load
, shares
;
2326 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
2327 load
= cfs_rq
->load
.weight
;
2329 shares
= (tg
->shares
* load
);
2331 shares
/= tg_weight
;
2333 if (shares
< MIN_SHARES
)
2334 shares
= MIN_SHARES
;
2335 if (shares
> tg
->shares
)
2336 shares
= tg
->shares
;
2340 # else /* CONFIG_SMP */
2341 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2345 # endif /* CONFIG_SMP */
2346 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2347 unsigned long weight
)
2350 /* commit outstanding execution time */
2351 if (cfs_rq
->curr
== se
)
2352 update_curr(cfs_rq
);
2353 account_entity_dequeue(cfs_rq
, se
);
2356 update_load_set(&se
->load
, weight
);
2359 account_entity_enqueue(cfs_rq
, se
);
2362 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2364 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2366 struct task_group
*tg
;
2367 struct sched_entity
*se
;
2371 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2372 if (!se
|| throttled_hierarchy(cfs_rq
))
2375 if (likely(se
->load
.weight
== tg
->shares
))
2378 shares
= calc_cfs_shares(cfs_rq
, tg
);
2380 reweight_entity(cfs_rq_of(se
), se
, shares
);
2382 #else /* CONFIG_FAIR_GROUP_SCHED */
2383 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2386 #endif /* CONFIG_FAIR_GROUP_SCHED */
2390 * We choose a half-life close to 1 scheduling period.
2391 * Note: The tables below are dependent on this value.
2393 #define LOAD_AVG_PERIOD 32
2394 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2395 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2397 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2398 static const u32 runnable_avg_yN_inv
[] = {
2399 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2400 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2401 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2402 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2403 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2404 0x85aac367, 0x82cd8698,
2408 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2409 * over-estimates when re-combining.
2411 static const u32 runnable_avg_yN_sum
[] = {
2412 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2413 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2414 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2419 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2421 static __always_inline u64
decay_load(u64 val
, u64 n
)
2423 unsigned int local_n
;
2427 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2430 /* after bounds checking we can collapse to 32-bit */
2434 * As y^PERIOD = 1/2, we can combine
2435 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2436 * With a look-up table which covers y^n (n<PERIOD)
2438 * To achieve constant time decay_load.
2440 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2441 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2442 local_n
%= LOAD_AVG_PERIOD
;
2445 val
*= runnable_avg_yN_inv
[local_n
];
2446 /* We don't use SRR here since we always want to round down. */
2451 * For updates fully spanning n periods, the contribution to runnable
2452 * average will be: \Sum 1024*y^n
2454 * We can compute this reasonably efficiently by combining:
2455 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2457 static u32
__compute_runnable_contrib(u64 n
)
2461 if (likely(n
<= LOAD_AVG_PERIOD
))
2462 return runnable_avg_yN_sum
[n
];
2463 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2464 return LOAD_AVG_MAX
;
2466 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2468 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2469 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2471 n
-= LOAD_AVG_PERIOD
;
2472 } while (n
> LOAD_AVG_PERIOD
);
2474 contrib
= decay_load(contrib
, n
);
2475 return contrib
+ runnable_avg_yN_sum
[n
];
2479 * We can represent the historical contribution to runnable average as the
2480 * coefficients of a geometric series. To do this we sub-divide our runnable
2481 * history into segments of approximately 1ms (1024us); label the segment that
2482 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2484 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2486 * (now) (~1ms ago) (~2ms ago)
2488 * Let u_i denote the fraction of p_i that the entity was runnable.
2490 * We then designate the fractions u_i as our co-efficients, yielding the
2491 * following representation of historical load:
2492 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2494 * We choose y based on the with of a reasonably scheduling period, fixing:
2497 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2498 * approximately half as much as the contribution to load within the last ms
2501 * When a period "rolls over" and we have new u_0`, multiplying the previous
2502 * sum again by y is sufficient to update:
2503 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2504 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2506 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2507 struct sched_avg
*sa
,
2511 u32 runnable_contrib
;
2512 int delta_w
, decayed
= 0;
2514 delta
= now
- sa
->last_runnable_update
;
2516 * This should only happen when time goes backwards, which it
2517 * unfortunately does during sched clock init when we swap over to TSC.
2519 if ((s64
)delta
< 0) {
2520 sa
->last_runnable_update
= now
;
2525 * Use 1024ns as the unit of measurement since it's a reasonable
2526 * approximation of 1us and fast to compute.
2531 sa
->last_runnable_update
= now
;
2533 /* delta_w is the amount already accumulated against our next period */
2534 delta_w
= sa
->runnable_avg_period
% 1024;
2535 if (delta
+ delta_w
>= 1024) {
2536 /* period roll-over */
2540 * Now that we know we're crossing a period boundary, figure
2541 * out how much from delta we need to complete the current
2542 * period and accrue it.
2544 delta_w
= 1024 - delta_w
;
2546 sa
->runnable_avg_sum
+= delta_w
;
2547 sa
->runnable_avg_period
+= delta_w
;
2551 /* Figure out how many additional periods this update spans */
2552 periods
= delta
/ 1024;
2555 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2557 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2560 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2561 runnable_contrib
= __compute_runnable_contrib(periods
);
2563 sa
->runnable_avg_sum
+= runnable_contrib
;
2564 sa
->runnable_avg_period
+= runnable_contrib
;
2567 /* Remainder of delta accrued against u_0` */
2569 sa
->runnable_avg_sum
+= delta
;
2570 sa
->runnable_avg_period
+= delta
;
2575 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2576 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2578 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2579 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2581 decays
-= se
->avg
.decay_count
;
2582 se
->avg
.decay_count
= 0;
2586 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2591 #ifdef CONFIG_FAIR_GROUP_SCHED
2592 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2595 struct task_group
*tg
= cfs_rq
->tg
;
2598 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2599 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2604 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2605 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2606 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2611 * Aggregate cfs_rq runnable averages into an equivalent task_group
2612 * representation for computing load contributions.
2614 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2615 struct cfs_rq
*cfs_rq
)
2617 struct task_group
*tg
= cfs_rq
->tg
;
2620 /* The fraction of a cpu used by this cfs_rq */
2621 contrib
= div_u64((u64
)sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2622 sa
->runnable_avg_period
+ 1);
2623 contrib
-= cfs_rq
->tg_runnable_contrib
;
2625 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2626 atomic_add(contrib
, &tg
->runnable_avg
);
2627 cfs_rq
->tg_runnable_contrib
+= contrib
;
2631 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2633 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2634 struct task_group
*tg
= cfs_rq
->tg
;
2639 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2640 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2641 atomic_long_read(&tg
->load_avg
) + 1);
2644 * For group entities we need to compute a correction term in the case
2645 * that they are consuming <1 cpu so that we would contribute the same
2646 * load as a task of equal weight.
2648 * Explicitly co-ordinating this measurement would be expensive, but
2649 * fortunately the sum of each cpus contribution forms a usable
2650 * lower-bound on the true value.
2652 * Consider the aggregate of 2 contributions. Either they are disjoint
2653 * (and the sum represents true value) or they are disjoint and we are
2654 * understating by the aggregate of their overlap.
2656 * Extending this to N cpus, for a given overlap, the maximum amount we
2657 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2658 * cpus that overlap for this interval and w_i is the interval width.
2660 * On a small machine; the first term is well-bounded which bounds the
2661 * total error since w_i is a subset of the period. Whereas on a
2662 * larger machine, while this first term can be larger, if w_i is the
2663 * of consequential size guaranteed to see n_i*w_i quickly converge to
2664 * our upper bound of 1-cpu.
2666 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2667 if (runnable_avg
< NICE_0_LOAD
) {
2668 se
->avg
.load_avg_contrib
*= runnable_avg
;
2669 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2673 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2675 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2676 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2678 #else /* CONFIG_FAIR_GROUP_SCHED */
2679 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2680 int force_update
) {}
2681 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2682 struct cfs_rq
*cfs_rq
) {}
2683 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2684 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2685 #endif /* CONFIG_FAIR_GROUP_SCHED */
2687 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2691 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2692 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2693 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2694 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2697 /* Compute the current contribution to load_avg by se, return any delta */
2698 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2700 long old_contrib
= se
->avg
.load_avg_contrib
;
2702 if (entity_is_task(se
)) {
2703 __update_task_entity_contrib(se
);
2705 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2706 __update_group_entity_contrib(se
);
2709 return se
->avg
.load_avg_contrib
- old_contrib
;
2712 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2715 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2716 cfs_rq
->blocked_load_avg
-= load_contrib
;
2718 cfs_rq
->blocked_load_avg
= 0;
2721 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2723 /* Update a sched_entity's runnable average */
2724 static inline void update_entity_load_avg(struct sched_entity
*se
,
2727 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2732 * For a group entity we need to use their owned cfs_rq_clock_task() in
2733 * case they are the parent of a throttled hierarchy.
2735 if (entity_is_task(se
))
2736 now
= cfs_rq_clock_task(cfs_rq
);
2738 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2740 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2743 contrib_delta
= __update_entity_load_avg_contrib(se
);
2749 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2751 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2755 * Decay the load contributed by all blocked children and account this so that
2756 * their contribution may appropriately discounted when they wake up.
2758 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2760 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2763 decays
= now
- cfs_rq
->last_decay
;
2764 if (!decays
&& !force_update
)
2767 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2768 unsigned long removed_load
;
2769 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2770 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2774 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2776 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2777 cfs_rq
->last_decay
= now
;
2780 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2783 /* Add the load generated by se into cfs_rq's child load-average */
2784 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2785 struct sched_entity
*se
,
2789 * We track migrations using entity decay_count <= 0, on a wake-up
2790 * migration we use a negative decay count to track the remote decays
2791 * accumulated while sleeping.
2793 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2794 * are seen by enqueue_entity_load_avg() as a migration with an already
2795 * constructed load_avg_contrib.
2797 if (unlikely(se
->avg
.decay_count
<= 0)) {
2798 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2799 if (se
->avg
.decay_count
) {
2801 * In a wake-up migration we have to approximate the
2802 * time sleeping. This is because we can't synchronize
2803 * clock_task between the two cpus, and it is not
2804 * guaranteed to be read-safe. Instead, we can
2805 * approximate this using our carried decays, which are
2806 * explicitly atomically readable.
2808 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2810 update_entity_load_avg(se
, 0);
2811 /* Indicate that we're now synchronized and on-rq */
2812 se
->avg
.decay_count
= 0;
2816 __synchronize_entity_decay(se
);
2819 /* migrated tasks did not contribute to our blocked load */
2821 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2822 update_entity_load_avg(se
, 0);
2825 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2826 /* we force update consideration on load-balancer moves */
2827 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2831 * Remove se's load from this cfs_rq child load-average, if the entity is
2832 * transitioning to a blocked state we track its projected decay using
2835 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2836 struct sched_entity
*se
,
2839 update_entity_load_avg(se
, 1);
2840 /* we force update consideration on load-balancer moves */
2841 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2843 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2845 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2846 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2847 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2851 * Update the rq's load with the elapsed running time before entering
2852 * idle. if the last scheduled task is not a CFS task, idle_enter will
2853 * be the only way to update the runnable statistic.
2855 void idle_enter_fair(struct rq
*this_rq
)
2857 update_rq_runnable_avg(this_rq
, 1);
2861 * Update the rq's load with the elapsed idle time before a task is
2862 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2863 * be the only way to update the runnable statistic.
2865 void idle_exit_fair(struct rq
*this_rq
)
2867 update_rq_runnable_avg(this_rq
, 0);
2870 static int idle_balance(struct rq
*this_rq
);
2872 #else /* CONFIG_SMP */
2874 static inline void update_entity_load_avg(struct sched_entity
*se
,
2875 int update_cfs_rq
) {}
2876 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2877 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2878 struct sched_entity
*se
,
2880 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2881 struct sched_entity
*se
,
2883 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2884 int force_update
) {}
2886 static inline int idle_balance(struct rq
*rq
)
2891 #endif /* CONFIG_SMP */
2893 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2895 #ifdef CONFIG_SCHEDSTATS
2896 struct task_struct
*tsk
= NULL
;
2898 if (entity_is_task(se
))
2901 if (se
->statistics
.sleep_start
) {
2902 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2907 if (unlikely(delta
> se
->statistics
.sleep_max
))
2908 se
->statistics
.sleep_max
= delta
;
2910 se
->statistics
.sleep_start
= 0;
2911 se
->statistics
.sum_sleep_runtime
+= delta
;
2914 account_scheduler_latency(tsk
, delta
>> 10, 1);
2915 trace_sched_stat_sleep(tsk
, delta
);
2918 if (se
->statistics
.block_start
) {
2919 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2924 if (unlikely(delta
> se
->statistics
.block_max
))
2925 se
->statistics
.block_max
= delta
;
2927 se
->statistics
.block_start
= 0;
2928 se
->statistics
.sum_sleep_runtime
+= delta
;
2931 if (tsk
->in_iowait
) {
2932 se
->statistics
.iowait_sum
+= delta
;
2933 se
->statistics
.iowait_count
++;
2934 trace_sched_stat_iowait(tsk
, delta
);
2937 trace_sched_stat_blocked(tsk
, delta
);
2940 * Blocking time is in units of nanosecs, so shift by
2941 * 20 to get a milliseconds-range estimation of the
2942 * amount of time that the task spent sleeping:
2944 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2945 profile_hits(SLEEP_PROFILING
,
2946 (void *)get_wchan(tsk
),
2949 account_scheduler_latency(tsk
, delta
>> 10, 0);
2955 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2957 #ifdef CONFIG_SCHED_DEBUG
2958 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2963 if (d
> 3*sysctl_sched_latency
)
2964 schedstat_inc(cfs_rq
, nr_spread_over
);
2969 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2971 u64 vruntime
= cfs_rq
->min_vruntime
;
2974 * The 'current' period is already promised to the current tasks,
2975 * however the extra weight of the new task will slow them down a
2976 * little, place the new task so that it fits in the slot that
2977 * stays open at the end.
2979 if (initial
&& sched_feat(START_DEBIT
))
2980 vruntime
+= sched_vslice(cfs_rq
, se
);
2982 /* sleeps up to a single latency don't count. */
2984 unsigned long thresh
= sysctl_sched_latency
;
2987 * Halve their sleep time's effect, to allow
2988 * for a gentler effect of sleepers:
2990 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2996 /* ensure we never gain time by being placed backwards. */
2997 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
3000 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
3003 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3006 * Update the normalized vruntime before updating min_vruntime
3007 * through calling update_curr().
3009 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
3010 se
->vruntime
+= cfs_rq
->min_vruntime
;
3013 * Update run-time statistics of the 'current'.
3015 update_curr(cfs_rq
);
3016 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
3017 account_entity_enqueue(cfs_rq
, se
);
3018 update_cfs_shares(cfs_rq
);
3020 if (flags
& ENQUEUE_WAKEUP
) {
3021 place_entity(cfs_rq
, se
, 0);
3022 enqueue_sleeper(cfs_rq
, se
);
3025 update_stats_enqueue(cfs_rq
, se
);
3026 check_spread(cfs_rq
, se
);
3027 if (se
!= cfs_rq
->curr
)
3028 __enqueue_entity(cfs_rq
, se
);
3031 if (cfs_rq
->nr_running
== 1) {
3032 list_add_leaf_cfs_rq(cfs_rq
);
3033 check_enqueue_throttle(cfs_rq
);
3037 static void __clear_buddies_last(struct sched_entity
*se
)
3039 for_each_sched_entity(se
) {
3040 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3041 if (cfs_rq
->last
!= se
)
3044 cfs_rq
->last
= NULL
;
3048 static void __clear_buddies_next(struct sched_entity
*se
)
3050 for_each_sched_entity(se
) {
3051 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3052 if (cfs_rq
->next
!= se
)
3055 cfs_rq
->next
= NULL
;
3059 static void __clear_buddies_skip(struct sched_entity
*se
)
3061 for_each_sched_entity(se
) {
3062 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3063 if (cfs_rq
->skip
!= se
)
3066 cfs_rq
->skip
= NULL
;
3070 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3072 if (cfs_rq
->last
== se
)
3073 __clear_buddies_last(se
);
3075 if (cfs_rq
->next
== se
)
3076 __clear_buddies_next(se
);
3078 if (cfs_rq
->skip
== se
)
3079 __clear_buddies_skip(se
);
3082 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3085 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3088 * Update run-time statistics of the 'current'.
3090 update_curr(cfs_rq
);
3091 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
3093 update_stats_dequeue(cfs_rq
, se
);
3094 if (flags
& DEQUEUE_SLEEP
) {
3095 #ifdef CONFIG_SCHEDSTATS
3096 if (entity_is_task(se
)) {
3097 struct task_struct
*tsk
= task_of(se
);
3099 if (tsk
->state
& TASK_INTERRUPTIBLE
)
3100 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
3101 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
3102 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
3107 clear_buddies(cfs_rq
, se
);
3109 if (se
!= cfs_rq
->curr
)
3110 __dequeue_entity(cfs_rq
, se
);
3112 account_entity_dequeue(cfs_rq
, se
);
3115 * Normalize the entity after updating the min_vruntime because the
3116 * update can refer to the ->curr item and we need to reflect this
3117 * movement in our normalized position.
3119 if (!(flags
& DEQUEUE_SLEEP
))
3120 se
->vruntime
-= cfs_rq
->min_vruntime
;
3122 /* return excess runtime on last dequeue */
3123 return_cfs_rq_runtime(cfs_rq
);
3125 update_min_vruntime(cfs_rq
);
3126 update_cfs_shares(cfs_rq
);
3130 * Preempt the current task with a newly woken task if needed:
3133 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3135 unsigned long ideal_runtime
, delta_exec
;
3136 struct sched_entity
*se
;
3139 ideal_runtime
= sched_slice(cfs_rq
, curr
);
3140 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
3141 if (delta_exec
> ideal_runtime
) {
3142 resched_curr(rq_of(cfs_rq
));
3144 * The current task ran long enough, ensure it doesn't get
3145 * re-elected due to buddy favours.
3147 clear_buddies(cfs_rq
, curr
);
3152 * Ensure that a task that missed wakeup preemption by a
3153 * narrow margin doesn't have to wait for a full slice.
3154 * This also mitigates buddy induced latencies under load.
3156 if (delta_exec
< sysctl_sched_min_granularity
)
3159 se
= __pick_first_entity(cfs_rq
);
3160 delta
= curr
->vruntime
- se
->vruntime
;
3165 if (delta
> ideal_runtime
)
3166 resched_curr(rq_of(cfs_rq
));
3170 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3172 /* 'current' is not kept within the tree. */
3175 * Any task has to be enqueued before it get to execute on
3176 * a CPU. So account for the time it spent waiting on the
3179 update_stats_wait_end(cfs_rq
, se
);
3180 __dequeue_entity(cfs_rq
, se
);
3183 update_stats_curr_start(cfs_rq
, se
);
3185 #ifdef CONFIG_SCHEDSTATS
3187 * Track our maximum slice length, if the CPU's load is at
3188 * least twice that of our own weight (i.e. dont track it
3189 * when there are only lesser-weight tasks around):
3191 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
3192 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
3193 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
3196 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
3200 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
3203 * Pick the next process, keeping these things in mind, in this order:
3204 * 1) keep things fair between processes/task groups
3205 * 2) pick the "next" process, since someone really wants that to run
3206 * 3) pick the "last" process, for cache locality
3207 * 4) do not run the "skip" process, if something else is available
3209 static struct sched_entity
*
3210 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3212 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
3213 struct sched_entity
*se
;
3216 * If curr is set we have to see if its left of the leftmost entity
3217 * still in the tree, provided there was anything in the tree at all.
3219 if (!left
|| (curr
&& entity_before(curr
, left
)))
3222 se
= left
; /* ideally we run the leftmost entity */
3225 * Avoid running the skip buddy, if running something else can
3226 * be done without getting too unfair.
3228 if (cfs_rq
->skip
== se
) {
3229 struct sched_entity
*second
;
3232 second
= __pick_first_entity(cfs_rq
);
3234 second
= __pick_next_entity(se
);
3235 if (!second
|| (curr
&& entity_before(curr
, second
)))
3239 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3244 * Prefer last buddy, try to return the CPU to a preempted task.
3246 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3250 * Someone really wants this to run. If it's not unfair, run it.
3252 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3255 clear_buddies(cfs_rq
, se
);
3260 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3262 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3265 * If still on the runqueue then deactivate_task()
3266 * was not called and update_curr() has to be done:
3269 update_curr(cfs_rq
);
3271 /* throttle cfs_rqs exceeding runtime */
3272 check_cfs_rq_runtime(cfs_rq
);
3274 check_spread(cfs_rq
, prev
);
3276 update_stats_wait_start(cfs_rq
, prev
);
3277 /* Put 'current' back into the tree. */
3278 __enqueue_entity(cfs_rq
, prev
);
3279 /* in !on_rq case, update occurred at dequeue */
3280 update_entity_load_avg(prev
, 1);
3282 cfs_rq
->curr
= NULL
;
3286 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3289 * Update run-time statistics of the 'current'.
3291 update_curr(cfs_rq
);
3294 * Ensure that runnable average is periodically updated.
3296 update_entity_load_avg(curr
, 1);
3297 update_cfs_rq_blocked_load(cfs_rq
, 1);
3298 update_cfs_shares(cfs_rq
);
3300 #ifdef CONFIG_SCHED_HRTICK
3302 * queued ticks are scheduled to match the slice, so don't bother
3303 * validating it and just reschedule.
3306 resched_curr(rq_of(cfs_rq
));
3310 * don't let the period tick interfere with the hrtick preemption
3312 if (!sched_feat(DOUBLE_TICK
) &&
3313 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3317 if (cfs_rq
->nr_running
> 1)
3318 check_preempt_tick(cfs_rq
, curr
);
3322 /**************************************************
3323 * CFS bandwidth control machinery
3326 #ifdef CONFIG_CFS_BANDWIDTH
3328 #ifdef HAVE_JUMP_LABEL
3329 static struct static_key __cfs_bandwidth_used
;
3331 static inline bool cfs_bandwidth_used(void)
3333 return static_key_false(&__cfs_bandwidth_used
);
3336 void cfs_bandwidth_usage_inc(void)
3338 static_key_slow_inc(&__cfs_bandwidth_used
);
3341 void cfs_bandwidth_usage_dec(void)
3343 static_key_slow_dec(&__cfs_bandwidth_used
);
3345 #else /* HAVE_JUMP_LABEL */
3346 static bool cfs_bandwidth_used(void)
3351 void cfs_bandwidth_usage_inc(void) {}
3352 void cfs_bandwidth_usage_dec(void) {}
3353 #endif /* HAVE_JUMP_LABEL */
3356 * default period for cfs group bandwidth.
3357 * default: 0.1s, units: nanoseconds
3359 static inline u64
default_cfs_period(void)
3361 return 100000000ULL;
3364 static inline u64
sched_cfs_bandwidth_slice(void)
3366 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3370 * Replenish runtime according to assigned quota and update expiration time.
3371 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3372 * additional synchronization around rq->lock.
3374 * requires cfs_b->lock
3376 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3380 if (cfs_b
->quota
== RUNTIME_INF
)
3383 now
= sched_clock_cpu(smp_processor_id());
3384 cfs_b
->runtime
= cfs_b
->quota
;
3385 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3388 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3390 return &tg
->cfs_bandwidth
;
3393 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3394 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3396 if (unlikely(cfs_rq
->throttle_count
))
3397 return cfs_rq
->throttled_clock_task
;
3399 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3402 /* returns 0 on failure to allocate runtime */
3403 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3405 struct task_group
*tg
= cfs_rq
->tg
;
3406 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3407 u64 amount
= 0, min_amount
, expires
;
3409 /* note: this is a positive sum as runtime_remaining <= 0 */
3410 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3412 raw_spin_lock(&cfs_b
->lock
);
3413 if (cfs_b
->quota
== RUNTIME_INF
)
3414 amount
= min_amount
;
3417 * If the bandwidth pool has become inactive, then at least one
3418 * period must have elapsed since the last consumption.
3419 * Refresh the global state and ensure bandwidth timer becomes
3422 if (!cfs_b
->timer_active
) {
3423 __refill_cfs_bandwidth_runtime(cfs_b
);
3424 __start_cfs_bandwidth(cfs_b
, false);
3427 if (cfs_b
->runtime
> 0) {
3428 amount
= min(cfs_b
->runtime
, min_amount
);
3429 cfs_b
->runtime
-= amount
;
3433 expires
= cfs_b
->runtime_expires
;
3434 raw_spin_unlock(&cfs_b
->lock
);
3436 cfs_rq
->runtime_remaining
+= amount
;
3438 * we may have advanced our local expiration to account for allowed
3439 * spread between our sched_clock and the one on which runtime was
3442 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3443 cfs_rq
->runtime_expires
= expires
;
3445 return cfs_rq
->runtime_remaining
> 0;
3449 * Note: This depends on the synchronization provided by sched_clock and the
3450 * fact that rq->clock snapshots this value.
3452 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3454 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3456 /* if the deadline is ahead of our clock, nothing to do */
3457 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3460 if (cfs_rq
->runtime_remaining
< 0)
3464 * If the local deadline has passed we have to consider the
3465 * possibility that our sched_clock is 'fast' and the global deadline
3466 * has not truly expired.
3468 * Fortunately we can check determine whether this the case by checking
3469 * whether the global deadline has advanced. It is valid to compare
3470 * cfs_b->runtime_expires without any locks since we only care about
3471 * exact equality, so a partial write will still work.
3474 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
3475 /* extend local deadline, drift is bounded above by 2 ticks */
3476 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3478 /* global deadline is ahead, expiration has passed */
3479 cfs_rq
->runtime_remaining
= 0;
3483 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3485 /* dock delta_exec before expiring quota (as it could span periods) */
3486 cfs_rq
->runtime_remaining
-= delta_exec
;
3487 expire_cfs_rq_runtime(cfs_rq
);
3489 if (likely(cfs_rq
->runtime_remaining
> 0))
3493 * if we're unable to extend our runtime we resched so that the active
3494 * hierarchy can be throttled
3496 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3497 resched_curr(rq_of(cfs_rq
));
3500 static __always_inline
3501 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3503 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3506 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3509 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3511 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3514 /* check whether cfs_rq, or any parent, is throttled */
3515 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3517 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3521 * Ensure that neither of the group entities corresponding to src_cpu or
3522 * dest_cpu are members of a throttled hierarchy when performing group
3523 * load-balance operations.
3525 static inline int throttled_lb_pair(struct task_group
*tg
,
3526 int src_cpu
, int dest_cpu
)
3528 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3530 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3531 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3533 return throttled_hierarchy(src_cfs_rq
) ||
3534 throttled_hierarchy(dest_cfs_rq
);
3537 /* updated child weight may affect parent so we have to do this bottom up */
3538 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3540 struct rq
*rq
= data
;
3541 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3543 cfs_rq
->throttle_count
--;
3545 if (!cfs_rq
->throttle_count
) {
3546 /* adjust cfs_rq_clock_task() */
3547 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3548 cfs_rq
->throttled_clock_task
;
3555 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3557 struct rq
*rq
= data
;
3558 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3560 /* group is entering throttled state, stop time */
3561 if (!cfs_rq
->throttle_count
)
3562 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3563 cfs_rq
->throttle_count
++;
3568 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3570 struct rq
*rq
= rq_of(cfs_rq
);
3571 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3572 struct sched_entity
*se
;
3573 long task_delta
, dequeue
= 1;
3575 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3577 /* freeze hierarchy runnable averages while throttled */
3579 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3582 task_delta
= cfs_rq
->h_nr_running
;
3583 for_each_sched_entity(se
) {
3584 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3585 /* throttled entity or throttle-on-deactivate */
3590 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3591 qcfs_rq
->h_nr_running
-= task_delta
;
3593 if (qcfs_rq
->load
.weight
)
3598 sub_nr_running(rq
, task_delta
);
3600 cfs_rq
->throttled
= 1;
3601 cfs_rq
->throttled_clock
= rq_clock(rq
);
3602 raw_spin_lock(&cfs_b
->lock
);
3604 * Add to the _head_ of the list, so that an already-started
3605 * distribute_cfs_runtime will not see us
3607 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3608 if (!cfs_b
->timer_active
)
3609 __start_cfs_bandwidth(cfs_b
, false);
3610 raw_spin_unlock(&cfs_b
->lock
);
3613 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3615 struct rq
*rq
= rq_of(cfs_rq
);
3616 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3617 struct sched_entity
*se
;
3621 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3623 cfs_rq
->throttled
= 0;
3625 update_rq_clock(rq
);
3627 raw_spin_lock(&cfs_b
->lock
);
3628 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3629 list_del_rcu(&cfs_rq
->throttled_list
);
3630 raw_spin_unlock(&cfs_b
->lock
);
3632 /* update hierarchical throttle state */
3633 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3635 if (!cfs_rq
->load
.weight
)
3638 task_delta
= cfs_rq
->h_nr_running
;
3639 for_each_sched_entity(se
) {
3643 cfs_rq
= cfs_rq_of(se
);
3645 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3646 cfs_rq
->h_nr_running
+= task_delta
;
3648 if (cfs_rq_throttled(cfs_rq
))
3653 add_nr_running(rq
, task_delta
);
3655 /* determine whether we need to wake up potentially idle cpu */
3656 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3660 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3661 u64 remaining
, u64 expires
)
3663 struct cfs_rq
*cfs_rq
;
3665 u64 starting_runtime
= remaining
;
3668 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3670 struct rq
*rq
= rq_of(cfs_rq
);
3672 raw_spin_lock(&rq
->lock
);
3673 if (!cfs_rq_throttled(cfs_rq
))
3676 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3677 if (runtime
> remaining
)
3678 runtime
= remaining
;
3679 remaining
-= runtime
;
3681 cfs_rq
->runtime_remaining
+= runtime
;
3682 cfs_rq
->runtime_expires
= expires
;
3684 /* we check whether we're throttled above */
3685 if (cfs_rq
->runtime_remaining
> 0)
3686 unthrottle_cfs_rq(cfs_rq
);
3689 raw_spin_unlock(&rq
->lock
);
3696 return starting_runtime
- remaining
;
3700 * Responsible for refilling a task_group's bandwidth and unthrottling its
3701 * cfs_rqs as appropriate. If there has been no activity within the last
3702 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3703 * used to track this state.
3705 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3707 u64 runtime
, runtime_expires
;
3710 /* no need to continue the timer with no bandwidth constraint */
3711 if (cfs_b
->quota
== RUNTIME_INF
)
3712 goto out_deactivate
;
3714 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3715 cfs_b
->nr_periods
+= overrun
;
3718 * idle depends on !throttled (for the case of a large deficit), and if
3719 * we're going inactive then everything else can be deferred
3721 if (cfs_b
->idle
&& !throttled
)
3722 goto out_deactivate
;
3725 * if we have relooped after returning idle once, we need to update our
3726 * status as actually running, so that other cpus doing
3727 * __start_cfs_bandwidth will stop trying to cancel us.
3729 cfs_b
->timer_active
= 1;
3731 __refill_cfs_bandwidth_runtime(cfs_b
);
3734 /* mark as potentially idle for the upcoming period */
3739 /* account preceding periods in which throttling occurred */
3740 cfs_b
->nr_throttled
+= overrun
;
3742 runtime_expires
= cfs_b
->runtime_expires
;
3745 * This check is repeated as we are holding onto the new bandwidth while
3746 * we unthrottle. This can potentially race with an unthrottled group
3747 * trying to acquire new bandwidth from the global pool. This can result
3748 * in us over-using our runtime if it is all used during this loop, but
3749 * only by limited amounts in that extreme case.
3751 while (throttled
&& cfs_b
->runtime
> 0) {
3752 runtime
= cfs_b
->runtime
;
3753 raw_spin_unlock(&cfs_b
->lock
);
3754 /* we can't nest cfs_b->lock while distributing bandwidth */
3755 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3757 raw_spin_lock(&cfs_b
->lock
);
3759 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3761 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3765 * While we are ensured activity in the period following an
3766 * unthrottle, this also covers the case in which the new bandwidth is
3767 * insufficient to cover the existing bandwidth deficit. (Forcing the
3768 * timer to remain active while there are any throttled entities.)
3775 cfs_b
->timer_active
= 0;
3779 /* a cfs_rq won't donate quota below this amount */
3780 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3781 /* minimum remaining period time to redistribute slack quota */
3782 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3783 /* how long we wait to gather additional slack before distributing */
3784 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3787 * Are we near the end of the current quota period?
3789 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3790 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3791 * migrate_hrtimers, base is never cleared, so we are fine.
3793 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3795 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3798 /* if the call-back is running a quota refresh is already occurring */
3799 if (hrtimer_callback_running(refresh_timer
))
3802 /* is a quota refresh about to occur? */
3803 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3804 if (remaining
< min_expire
)
3810 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3812 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3814 /* if there's a quota refresh soon don't bother with slack */
3815 if (runtime_refresh_within(cfs_b
, min_left
))
3818 start_bandwidth_timer(&cfs_b
->slack_timer
,
3819 ns_to_ktime(cfs_bandwidth_slack_period
));
3822 /* we know any runtime found here is valid as update_curr() precedes return */
3823 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3825 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3826 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3828 if (slack_runtime
<= 0)
3831 raw_spin_lock(&cfs_b
->lock
);
3832 if (cfs_b
->quota
!= RUNTIME_INF
&&
3833 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3834 cfs_b
->runtime
+= slack_runtime
;
3836 /* we are under rq->lock, defer unthrottling using a timer */
3837 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3838 !list_empty(&cfs_b
->throttled_cfs_rq
))
3839 start_cfs_slack_bandwidth(cfs_b
);
3841 raw_spin_unlock(&cfs_b
->lock
);
3843 /* even if it's not valid for return we don't want to try again */
3844 cfs_rq
->runtime_remaining
-= slack_runtime
;
3847 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3849 if (!cfs_bandwidth_used())
3852 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3855 __return_cfs_rq_runtime(cfs_rq
);
3859 * This is done with a timer (instead of inline with bandwidth return) since
3860 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3862 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3864 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3867 /* confirm we're still not at a refresh boundary */
3868 raw_spin_lock(&cfs_b
->lock
);
3869 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
3870 raw_spin_unlock(&cfs_b
->lock
);
3874 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
3875 runtime
= cfs_b
->runtime
;
3877 expires
= cfs_b
->runtime_expires
;
3878 raw_spin_unlock(&cfs_b
->lock
);
3883 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3885 raw_spin_lock(&cfs_b
->lock
);
3886 if (expires
== cfs_b
->runtime_expires
)
3887 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3888 raw_spin_unlock(&cfs_b
->lock
);
3892 * When a group wakes up we want to make sure that its quota is not already
3893 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3894 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3896 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3898 if (!cfs_bandwidth_used())
3901 /* an active group must be handled by the update_curr()->put() path */
3902 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3905 /* ensure the group is not already throttled */
3906 if (cfs_rq_throttled(cfs_rq
))
3909 /* update runtime allocation */
3910 account_cfs_rq_runtime(cfs_rq
, 0);
3911 if (cfs_rq
->runtime_remaining
<= 0)
3912 throttle_cfs_rq(cfs_rq
);
3915 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3916 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3918 if (!cfs_bandwidth_used())
3921 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3925 * it's possible for a throttled entity to be forced into a running
3926 * state (e.g. set_curr_task), in this case we're finished.
3928 if (cfs_rq_throttled(cfs_rq
))
3931 throttle_cfs_rq(cfs_rq
);
3935 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3937 struct cfs_bandwidth
*cfs_b
=
3938 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3939 do_sched_cfs_slack_timer(cfs_b
);
3941 return HRTIMER_NORESTART
;
3944 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3946 struct cfs_bandwidth
*cfs_b
=
3947 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3952 raw_spin_lock(&cfs_b
->lock
);
3954 now
= hrtimer_cb_get_time(timer
);
3955 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3960 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3962 raw_spin_unlock(&cfs_b
->lock
);
3964 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3967 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3969 raw_spin_lock_init(&cfs_b
->lock
);
3971 cfs_b
->quota
= RUNTIME_INF
;
3972 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3974 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3975 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3976 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3977 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3978 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3981 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3983 cfs_rq
->runtime_enabled
= 0;
3984 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3987 /* requires cfs_b->lock, may release to reprogram timer */
3988 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
, bool force
)
3991 * The timer may be active because we're trying to set a new bandwidth
3992 * period or because we're racing with the tear-down path
3993 * (timer_active==0 becomes visible before the hrtimer call-back
3994 * terminates). In either case we ensure that it's re-programmed
3996 while (unlikely(hrtimer_active(&cfs_b
->period_timer
)) &&
3997 hrtimer_try_to_cancel(&cfs_b
->period_timer
) < 0) {
3998 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3999 raw_spin_unlock(&cfs_b
->lock
);
4001 raw_spin_lock(&cfs_b
->lock
);
4002 /* if someone else restarted the timer then we're done */
4003 if (!force
&& cfs_b
->timer_active
)
4007 cfs_b
->timer_active
= 1;
4008 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
4011 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4013 /* init_cfs_bandwidth() was not called */
4014 if (!cfs_b
->throttled_cfs_rq
.next
)
4017 hrtimer_cancel(&cfs_b
->period_timer
);
4018 hrtimer_cancel(&cfs_b
->slack_timer
);
4021 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
4023 struct cfs_rq
*cfs_rq
;
4025 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4026 struct cfs_bandwidth
*cfs_b
= &cfs_rq
->tg
->cfs_bandwidth
;
4028 raw_spin_lock(&cfs_b
->lock
);
4029 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
4030 raw_spin_unlock(&cfs_b
->lock
);
4034 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
4036 struct cfs_rq
*cfs_rq
;
4038 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4039 if (!cfs_rq
->runtime_enabled
)
4043 * clock_task is not advancing so we just need to make sure
4044 * there's some valid quota amount
4046 cfs_rq
->runtime_remaining
= 1;
4048 * Offline rq is schedulable till cpu is completely disabled
4049 * in take_cpu_down(), so we prevent new cfs throttling here.
4051 cfs_rq
->runtime_enabled
= 0;
4053 if (cfs_rq_throttled(cfs_rq
))
4054 unthrottle_cfs_rq(cfs_rq
);
4058 #else /* CONFIG_CFS_BANDWIDTH */
4059 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
4061 return rq_clock_task(rq_of(cfs_rq
));
4064 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
4065 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
4066 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
4067 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4069 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4074 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4079 static inline int throttled_lb_pair(struct task_group
*tg
,
4080 int src_cpu
, int dest_cpu
)
4085 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4087 #ifdef CONFIG_FAIR_GROUP_SCHED
4088 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4091 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4095 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4096 static inline void update_runtime_enabled(struct rq
*rq
) {}
4097 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
4099 #endif /* CONFIG_CFS_BANDWIDTH */
4101 /**************************************************
4102 * CFS operations on tasks:
4105 #ifdef CONFIG_SCHED_HRTICK
4106 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4108 struct sched_entity
*se
= &p
->se
;
4109 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4111 WARN_ON(task_rq(p
) != rq
);
4113 if (cfs_rq
->nr_running
> 1) {
4114 u64 slice
= sched_slice(cfs_rq
, se
);
4115 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
4116 s64 delta
= slice
- ran
;
4123 hrtick_start(rq
, delta
);
4128 * called from enqueue/dequeue and updates the hrtick when the
4129 * current task is from our class and nr_running is low enough
4132 static void hrtick_update(struct rq
*rq
)
4134 struct task_struct
*curr
= rq
->curr
;
4136 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
4139 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
4140 hrtick_start_fair(rq
, curr
);
4142 #else /* !CONFIG_SCHED_HRTICK */
4144 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4148 static inline void hrtick_update(struct rq
*rq
)
4154 * The enqueue_task method is called before nr_running is
4155 * increased. Here we update the fair scheduling stats and
4156 * then put the task into the rbtree:
4159 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4161 struct cfs_rq
*cfs_rq
;
4162 struct sched_entity
*se
= &p
->se
;
4164 for_each_sched_entity(se
) {
4167 cfs_rq
= cfs_rq_of(se
);
4168 enqueue_entity(cfs_rq
, se
, flags
);
4171 * end evaluation on encountering a throttled cfs_rq
4173 * note: in the case of encountering a throttled cfs_rq we will
4174 * post the final h_nr_running increment below.
4176 if (cfs_rq_throttled(cfs_rq
))
4178 cfs_rq
->h_nr_running
++;
4180 flags
= ENQUEUE_WAKEUP
;
4183 for_each_sched_entity(se
) {
4184 cfs_rq
= cfs_rq_of(se
);
4185 cfs_rq
->h_nr_running
++;
4187 if (cfs_rq_throttled(cfs_rq
))
4190 update_cfs_shares(cfs_rq
);
4191 update_entity_load_avg(se
, 1);
4195 update_rq_runnable_avg(rq
, rq
->nr_running
);
4196 add_nr_running(rq
, 1);
4201 static void set_next_buddy(struct sched_entity
*se
);
4204 * The dequeue_task method is called before nr_running is
4205 * decreased. We remove the task from the rbtree and
4206 * update the fair scheduling stats:
4208 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4210 struct cfs_rq
*cfs_rq
;
4211 struct sched_entity
*se
= &p
->se
;
4212 int task_sleep
= flags
& DEQUEUE_SLEEP
;
4214 for_each_sched_entity(se
) {
4215 cfs_rq
= cfs_rq_of(se
);
4216 dequeue_entity(cfs_rq
, se
, flags
);
4219 * end evaluation on encountering a throttled cfs_rq
4221 * note: in the case of encountering a throttled cfs_rq we will
4222 * post the final h_nr_running decrement below.
4224 if (cfs_rq_throttled(cfs_rq
))
4226 cfs_rq
->h_nr_running
--;
4228 /* Don't dequeue parent if it has other entities besides us */
4229 if (cfs_rq
->load
.weight
) {
4231 * Bias pick_next to pick a task from this cfs_rq, as
4232 * p is sleeping when it is within its sched_slice.
4234 if (task_sleep
&& parent_entity(se
))
4235 set_next_buddy(parent_entity(se
));
4237 /* avoid re-evaluating load for this entity */
4238 se
= parent_entity(se
);
4241 flags
|= DEQUEUE_SLEEP
;
4244 for_each_sched_entity(se
) {
4245 cfs_rq
= cfs_rq_of(se
);
4246 cfs_rq
->h_nr_running
--;
4248 if (cfs_rq_throttled(cfs_rq
))
4251 update_cfs_shares(cfs_rq
);
4252 update_entity_load_avg(se
, 1);
4256 sub_nr_running(rq
, 1);
4257 update_rq_runnable_avg(rq
, 1);
4263 /* Used instead of source_load when we know the type == 0 */
4264 static unsigned long weighted_cpuload(const int cpu
)
4266 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
4270 * Return a low guess at the load of a migration-source cpu weighted
4271 * according to the scheduling class and "nice" value.
4273 * We want to under-estimate the load of migration sources, to
4274 * balance conservatively.
4276 static unsigned long source_load(int cpu
, int type
)
4278 struct rq
*rq
= cpu_rq(cpu
);
4279 unsigned long total
= weighted_cpuload(cpu
);
4281 if (type
== 0 || !sched_feat(LB_BIAS
))
4284 return min(rq
->cpu_load
[type
-1], total
);
4288 * Return a high guess at the load of a migration-target cpu weighted
4289 * according to the scheduling class and "nice" value.
4291 static unsigned long target_load(int cpu
, int type
)
4293 struct rq
*rq
= cpu_rq(cpu
);
4294 unsigned long total
= weighted_cpuload(cpu
);
4296 if (type
== 0 || !sched_feat(LB_BIAS
))
4299 return max(rq
->cpu_load
[type
-1], total
);
4302 static unsigned long capacity_of(int cpu
)
4304 return cpu_rq(cpu
)->cpu_capacity
;
4307 static unsigned long cpu_avg_load_per_task(int cpu
)
4309 struct rq
*rq
= cpu_rq(cpu
);
4310 unsigned long nr_running
= ACCESS_ONCE(rq
->cfs
.h_nr_running
);
4311 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
4314 return load_avg
/ nr_running
;
4319 static void record_wakee(struct task_struct
*p
)
4322 * Rough decay (wiping) for cost saving, don't worry
4323 * about the boundary, really active task won't care
4326 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
4327 current
->wakee_flips
>>= 1;
4328 current
->wakee_flip_decay_ts
= jiffies
;
4331 if (current
->last_wakee
!= p
) {
4332 current
->last_wakee
= p
;
4333 current
->wakee_flips
++;
4337 static void task_waking_fair(struct task_struct
*p
)
4339 struct sched_entity
*se
= &p
->se
;
4340 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4343 #ifndef CONFIG_64BIT
4344 u64 min_vruntime_copy
;
4347 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
4349 min_vruntime
= cfs_rq
->min_vruntime
;
4350 } while (min_vruntime
!= min_vruntime_copy
);
4352 min_vruntime
= cfs_rq
->min_vruntime
;
4355 se
->vruntime
-= min_vruntime
;
4359 #ifdef CONFIG_FAIR_GROUP_SCHED
4361 * effective_load() calculates the load change as seen from the root_task_group
4363 * Adding load to a group doesn't make a group heavier, but can cause movement
4364 * of group shares between cpus. Assuming the shares were perfectly aligned one
4365 * can calculate the shift in shares.
4367 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4368 * on this @cpu and results in a total addition (subtraction) of @wg to the
4369 * total group weight.
4371 * Given a runqueue weight distribution (rw_i) we can compute a shares
4372 * distribution (s_i) using:
4374 * s_i = rw_i / \Sum rw_j (1)
4376 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4377 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4378 * shares distribution (s_i):
4380 * rw_i = { 2, 4, 1, 0 }
4381 * s_i = { 2/7, 4/7, 1/7, 0 }
4383 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4384 * task used to run on and the CPU the waker is running on), we need to
4385 * compute the effect of waking a task on either CPU and, in case of a sync
4386 * wakeup, compute the effect of the current task going to sleep.
4388 * So for a change of @wl to the local @cpu with an overall group weight change
4389 * of @wl we can compute the new shares distribution (s'_i) using:
4391 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4393 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4394 * differences in waking a task to CPU 0. The additional task changes the
4395 * weight and shares distributions like:
4397 * rw'_i = { 3, 4, 1, 0 }
4398 * s'_i = { 3/8, 4/8, 1/8, 0 }
4400 * We can then compute the difference in effective weight by using:
4402 * dw_i = S * (s'_i - s_i) (3)
4404 * Where 'S' is the group weight as seen by its parent.
4406 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4407 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4408 * 4/7) times the weight of the group.
4410 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4412 struct sched_entity
*se
= tg
->se
[cpu
];
4414 if (!tg
->parent
) /* the trivial, non-cgroup case */
4417 for_each_sched_entity(se
) {
4423 * W = @wg + \Sum rw_j
4425 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4430 w
= se
->my_q
->load
.weight
+ wl
;
4433 * wl = S * s'_i; see (2)
4436 wl
= (w
* (long)tg
->shares
) / W
;
4441 * Per the above, wl is the new se->load.weight value; since
4442 * those are clipped to [MIN_SHARES, ...) do so now. See
4443 * calc_cfs_shares().
4445 if (wl
< MIN_SHARES
)
4449 * wl = dw_i = S * (s'_i - s_i); see (3)
4451 wl
-= se
->load
.weight
;
4454 * Recursively apply this logic to all parent groups to compute
4455 * the final effective load change on the root group. Since
4456 * only the @tg group gets extra weight, all parent groups can
4457 * only redistribute existing shares. @wl is the shift in shares
4458 * resulting from this level per the above.
4467 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4474 static int wake_wide(struct task_struct
*p
)
4476 int factor
= this_cpu_read(sd_llc_size
);
4479 * Yeah, it's the switching-frequency, could means many wakee or
4480 * rapidly switch, use factor here will just help to automatically
4481 * adjust the loose-degree, so bigger node will lead to more pull.
4483 if (p
->wakee_flips
> factor
) {
4485 * wakee is somewhat hot, it needs certain amount of cpu
4486 * resource, so if waker is far more hot, prefer to leave
4489 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
4496 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4498 s64 this_load
, load
;
4499 s64 this_eff_load
, prev_eff_load
;
4500 int idx
, this_cpu
, prev_cpu
;
4501 struct task_group
*tg
;
4502 unsigned long weight
;
4506 * If we wake multiple tasks be careful to not bounce
4507 * ourselves around too much.
4513 this_cpu
= smp_processor_id();
4514 prev_cpu
= task_cpu(p
);
4515 load
= source_load(prev_cpu
, idx
);
4516 this_load
= target_load(this_cpu
, idx
);
4519 * If sync wakeup then subtract the (maximum possible)
4520 * effect of the currently running task from the load
4521 * of the current CPU:
4524 tg
= task_group(current
);
4525 weight
= current
->se
.load
.weight
;
4527 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4528 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4532 weight
= p
->se
.load
.weight
;
4535 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4536 * due to the sync cause above having dropped this_load to 0, we'll
4537 * always have an imbalance, but there's really nothing you can do
4538 * about that, so that's good too.
4540 * Otherwise check if either cpus are near enough in load to allow this
4541 * task to be woken on this_cpu.
4543 this_eff_load
= 100;
4544 this_eff_load
*= capacity_of(prev_cpu
);
4546 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4547 prev_eff_load
*= capacity_of(this_cpu
);
4549 if (this_load
> 0) {
4550 this_eff_load
*= this_load
+
4551 effective_load(tg
, this_cpu
, weight
, weight
);
4553 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4556 balanced
= this_eff_load
<= prev_eff_load
;
4558 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4563 schedstat_inc(sd
, ttwu_move_affine
);
4564 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4570 * find_idlest_group finds and returns the least busy CPU group within the
4573 static struct sched_group
*
4574 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4575 int this_cpu
, int sd_flag
)
4577 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4578 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4579 int load_idx
= sd
->forkexec_idx
;
4580 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4582 if (sd_flag
& SD_BALANCE_WAKE
)
4583 load_idx
= sd
->wake_idx
;
4586 unsigned long load
, avg_load
;
4590 /* Skip over this group if it has no CPUs allowed */
4591 if (!cpumask_intersects(sched_group_cpus(group
),
4592 tsk_cpus_allowed(p
)))
4595 local_group
= cpumask_test_cpu(this_cpu
,
4596 sched_group_cpus(group
));
4598 /* Tally up the load of all CPUs in the group */
4601 for_each_cpu(i
, sched_group_cpus(group
)) {
4602 /* Bias balancing toward cpus of our domain */
4604 load
= source_load(i
, load_idx
);
4606 load
= target_load(i
, load_idx
);
4611 /* Adjust by relative CPU capacity of the group */
4612 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) / group
->sgc
->capacity
;
4615 this_load
= avg_load
;
4616 } else if (avg_load
< min_load
) {
4617 min_load
= avg_load
;
4620 } while (group
= group
->next
, group
!= sd
->groups
);
4622 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4628 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4631 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4633 unsigned long load
, min_load
= ULONG_MAX
;
4634 unsigned int min_exit_latency
= UINT_MAX
;
4635 u64 latest_idle_timestamp
= 0;
4636 int least_loaded_cpu
= this_cpu
;
4637 int shallowest_idle_cpu
= -1;
4640 /* Traverse only the allowed CPUs */
4641 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4643 struct rq
*rq
= cpu_rq(i
);
4644 struct cpuidle_state
*idle
= idle_get_state(rq
);
4645 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
4647 * We give priority to a CPU whose idle state
4648 * has the smallest exit latency irrespective
4649 * of any idle timestamp.
4651 min_exit_latency
= idle
->exit_latency
;
4652 latest_idle_timestamp
= rq
->idle_stamp
;
4653 shallowest_idle_cpu
= i
;
4654 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
4655 rq
->idle_stamp
> latest_idle_timestamp
) {
4657 * If equal or no active idle state, then
4658 * the most recently idled CPU might have
4661 latest_idle_timestamp
= rq
->idle_stamp
;
4662 shallowest_idle_cpu
= i
;
4664 } else if (shallowest_idle_cpu
== -1) {
4665 load
= weighted_cpuload(i
);
4666 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4668 least_loaded_cpu
= i
;
4673 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
4677 * Try and locate an idle CPU in the sched_domain.
4679 static int select_idle_sibling(struct task_struct
*p
, int target
)
4681 struct sched_domain
*sd
;
4682 struct sched_group
*sg
;
4683 int i
= task_cpu(p
);
4685 if (idle_cpu(target
))
4689 * If the prevous cpu is cache affine and idle, don't be stupid.
4691 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4695 * Otherwise, iterate the domains and find an elegible idle cpu.
4697 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4698 for_each_lower_domain(sd
) {
4701 if (!cpumask_intersects(sched_group_cpus(sg
),
4702 tsk_cpus_allowed(p
)))
4705 for_each_cpu(i
, sched_group_cpus(sg
)) {
4706 if (i
== target
|| !idle_cpu(i
))
4710 target
= cpumask_first_and(sched_group_cpus(sg
),
4711 tsk_cpus_allowed(p
));
4715 } while (sg
!= sd
->groups
);
4722 * select_task_rq_fair: Select target runqueue for the waking task in domains
4723 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4724 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4726 * Balances load by selecting the idlest cpu in the idlest group, or under
4727 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4729 * Returns the target cpu number.
4731 * preempt must be disabled.
4734 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4736 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4737 int cpu
= smp_processor_id();
4739 int want_affine
= 0;
4740 int sync
= wake_flags
& WF_SYNC
;
4742 if (sd_flag
& SD_BALANCE_WAKE
)
4743 want_affine
= cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
));
4746 for_each_domain(cpu
, tmp
) {
4747 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4751 * If both cpu and prev_cpu are part of this domain,
4752 * cpu is a valid SD_WAKE_AFFINE target.
4754 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4755 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4760 if (tmp
->flags
& sd_flag
)
4764 if (affine_sd
&& cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4767 if (sd_flag
& SD_BALANCE_WAKE
) {
4768 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4773 struct sched_group
*group
;
4776 if (!(sd
->flags
& sd_flag
)) {
4781 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
4787 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4788 if (new_cpu
== -1 || new_cpu
== cpu
) {
4789 /* Now try balancing at a lower domain level of cpu */
4794 /* Now try balancing at a lower domain level of new_cpu */
4796 weight
= sd
->span_weight
;
4798 for_each_domain(cpu
, tmp
) {
4799 if (weight
<= tmp
->span_weight
)
4801 if (tmp
->flags
& sd_flag
)
4804 /* while loop will break here if sd == NULL */
4813 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4814 * cfs_rq_of(p) references at time of call are still valid and identify the
4815 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4816 * other assumptions, including the state of rq->lock, should be made.
4819 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4821 struct sched_entity
*se
= &p
->se
;
4822 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4825 * Load tracking: accumulate removed load so that it can be processed
4826 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4827 * to blocked load iff they have a positive decay-count. It can never
4828 * be negative here since on-rq tasks have decay-count == 0.
4830 if (se
->avg
.decay_count
) {
4831 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4832 atomic_long_add(se
->avg
.load_avg_contrib
,
4833 &cfs_rq
->removed_load
);
4836 /* We have migrated, no longer consider this task hot */
4839 #endif /* CONFIG_SMP */
4841 static unsigned long
4842 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4844 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4847 * Since its curr running now, convert the gran from real-time
4848 * to virtual-time in his units.
4850 * By using 'se' instead of 'curr' we penalize light tasks, so
4851 * they get preempted easier. That is, if 'se' < 'curr' then
4852 * the resulting gran will be larger, therefore penalizing the
4853 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4854 * be smaller, again penalizing the lighter task.
4856 * This is especially important for buddies when the leftmost
4857 * task is higher priority than the buddy.
4859 return calc_delta_fair(gran
, se
);
4863 * Should 'se' preempt 'curr'.
4877 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4879 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4884 gran
= wakeup_gran(curr
, se
);
4891 static void set_last_buddy(struct sched_entity
*se
)
4893 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4896 for_each_sched_entity(se
)
4897 cfs_rq_of(se
)->last
= se
;
4900 static void set_next_buddy(struct sched_entity
*se
)
4902 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4905 for_each_sched_entity(se
)
4906 cfs_rq_of(se
)->next
= se
;
4909 static void set_skip_buddy(struct sched_entity
*se
)
4911 for_each_sched_entity(se
)
4912 cfs_rq_of(se
)->skip
= se
;
4916 * Preempt the current task with a newly woken task if needed:
4918 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4920 struct task_struct
*curr
= rq
->curr
;
4921 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4922 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4923 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4924 int next_buddy_marked
= 0;
4926 if (unlikely(se
== pse
))
4930 * This is possible from callers such as attach_tasks(), in which we
4931 * unconditionally check_prempt_curr() after an enqueue (which may have
4932 * lead to a throttle). This both saves work and prevents false
4933 * next-buddy nomination below.
4935 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4938 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4939 set_next_buddy(pse
);
4940 next_buddy_marked
= 1;
4944 * We can come here with TIF_NEED_RESCHED already set from new task
4947 * Note: this also catches the edge-case of curr being in a throttled
4948 * group (e.g. via set_curr_task), since update_curr() (in the
4949 * enqueue of curr) will have resulted in resched being set. This
4950 * prevents us from potentially nominating it as a false LAST_BUDDY
4953 if (test_tsk_need_resched(curr
))
4956 /* Idle tasks are by definition preempted by non-idle tasks. */
4957 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4958 likely(p
->policy
!= SCHED_IDLE
))
4962 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4963 * is driven by the tick):
4965 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4968 find_matching_se(&se
, &pse
);
4969 update_curr(cfs_rq_of(se
));
4971 if (wakeup_preempt_entity(se
, pse
) == 1) {
4973 * Bias pick_next to pick the sched entity that is
4974 * triggering this preemption.
4976 if (!next_buddy_marked
)
4977 set_next_buddy(pse
);
4986 * Only set the backward buddy when the current task is still
4987 * on the rq. This can happen when a wakeup gets interleaved
4988 * with schedule on the ->pre_schedule() or idle_balance()
4989 * point, either of which can * drop the rq lock.
4991 * Also, during early boot the idle thread is in the fair class,
4992 * for obvious reasons its a bad idea to schedule back to it.
4994 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4997 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
5001 static struct task_struct
*
5002 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
)
5004 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
5005 struct sched_entity
*se
;
5006 struct task_struct
*p
;
5010 #ifdef CONFIG_FAIR_GROUP_SCHED
5011 if (!cfs_rq
->nr_running
)
5014 if (prev
->sched_class
!= &fair_sched_class
)
5018 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5019 * likely that a next task is from the same cgroup as the current.
5021 * Therefore attempt to avoid putting and setting the entire cgroup
5022 * hierarchy, only change the part that actually changes.
5026 struct sched_entity
*curr
= cfs_rq
->curr
;
5029 * Since we got here without doing put_prev_entity() we also
5030 * have to consider cfs_rq->curr. If it is still a runnable
5031 * entity, update_curr() will update its vruntime, otherwise
5032 * forget we've ever seen it.
5034 if (curr
&& curr
->on_rq
)
5035 update_curr(cfs_rq
);
5040 * This call to check_cfs_rq_runtime() will do the throttle and
5041 * dequeue its entity in the parent(s). Therefore the 'simple'
5042 * nr_running test will indeed be correct.
5044 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
5047 se
= pick_next_entity(cfs_rq
, curr
);
5048 cfs_rq
= group_cfs_rq(se
);
5054 * Since we haven't yet done put_prev_entity and if the selected task
5055 * is a different task than we started out with, try and touch the
5056 * least amount of cfs_rqs.
5059 struct sched_entity
*pse
= &prev
->se
;
5061 while (!(cfs_rq
= is_same_group(se
, pse
))) {
5062 int se_depth
= se
->depth
;
5063 int pse_depth
= pse
->depth
;
5065 if (se_depth
<= pse_depth
) {
5066 put_prev_entity(cfs_rq_of(pse
), pse
);
5067 pse
= parent_entity(pse
);
5069 if (se_depth
>= pse_depth
) {
5070 set_next_entity(cfs_rq_of(se
), se
);
5071 se
= parent_entity(se
);
5075 put_prev_entity(cfs_rq
, pse
);
5076 set_next_entity(cfs_rq
, se
);
5079 if (hrtick_enabled(rq
))
5080 hrtick_start_fair(rq
, p
);
5087 if (!cfs_rq
->nr_running
)
5090 put_prev_task(rq
, prev
);
5093 se
= pick_next_entity(cfs_rq
, NULL
);
5094 set_next_entity(cfs_rq
, se
);
5095 cfs_rq
= group_cfs_rq(se
);
5100 if (hrtick_enabled(rq
))
5101 hrtick_start_fair(rq
, p
);
5106 new_tasks
= idle_balance(rq
);
5108 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5109 * possible for any higher priority task to appear. In that case we
5110 * must re-start the pick_next_entity() loop.
5122 * Account for a descheduled task:
5124 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
5126 struct sched_entity
*se
= &prev
->se
;
5127 struct cfs_rq
*cfs_rq
;
5129 for_each_sched_entity(se
) {
5130 cfs_rq
= cfs_rq_of(se
);
5131 put_prev_entity(cfs_rq
, se
);
5136 * sched_yield() is very simple
5138 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5140 static void yield_task_fair(struct rq
*rq
)
5142 struct task_struct
*curr
= rq
->curr
;
5143 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
5144 struct sched_entity
*se
= &curr
->se
;
5147 * Are we the only task in the tree?
5149 if (unlikely(rq
->nr_running
== 1))
5152 clear_buddies(cfs_rq
, se
);
5154 if (curr
->policy
!= SCHED_BATCH
) {
5155 update_rq_clock(rq
);
5157 * Update run-time statistics of the 'current'.
5159 update_curr(cfs_rq
);
5161 * Tell update_rq_clock() that we've just updated,
5162 * so we don't do microscopic update in schedule()
5163 * and double the fastpath cost.
5165 rq_clock_skip_update(rq
, true);
5171 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
5173 struct sched_entity
*se
= &p
->se
;
5175 /* throttled hierarchies are not runnable */
5176 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
5179 /* Tell the scheduler that we'd really like pse to run next. */
5182 yield_task_fair(rq
);
5188 /**************************************************
5189 * Fair scheduling class load-balancing methods.
5193 * The purpose of load-balancing is to achieve the same basic fairness the
5194 * per-cpu scheduler provides, namely provide a proportional amount of compute
5195 * time to each task. This is expressed in the following equation:
5197 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5199 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5200 * W_i,0 is defined as:
5202 * W_i,0 = \Sum_j w_i,j (2)
5204 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5205 * is derived from the nice value as per prio_to_weight[].
5207 * The weight average is an exponential decay average of the instantaneous
5210 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5212 * C_i is the compute capacity of cpu i, typically it is the
5213 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5214 * can also include other factors [XXX].
5216 * To achieve this balance we define a measure of imbalance which follows
5217 * directly from (1):
5219 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5221 * We them move tasks around to minimize the imbalance. In the continuous
5222 * function space it is obvious this converges, in the discrete case we get
5223 * a few fun cases generally called infeasible weight scenarios.
5226 * - infeasible weights;
5227 * - local vs global optima in the discrete case. ]
5232 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5233 * for all i,j solution, we create a tree of cpus that follows the hardware
5234 * topology where each level pairs two lower groups (or better). This results
5235 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5236 * tree to only the first of the previous level and we decrease the frequency
5237 * of load-balance at each level inv. proportional to the number of cpus in
5243 * \Sum { --- * --- * 2^i } = O(n) (5)
5245 * `- size of each group
5246 * | | `- number of cpus doing load-balance
5248 * `- sum over all levels
5250 * Coupled with a limit on how many tasks we can migrate every balance pass,
5251 * this makes (5) the runtime complexity of the balancer.
5253 * An important property here is that each CPU is still (indirectly) connected
5254 * to every other cpu in at most O(log n) steps:
5256 * The adjacency matrix of the resulting graph is given by:
5259 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5262 * And you'll find that:
5264 * A^(log_2 n)_i,j != 0 for all i,j (7)
5266 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5267 * The task movement gives a factor of O(m), giving a convergence complexity
5270 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5275 * In order to avoid CPUs going idle while there's still work to do, new idle
5276 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5277 * tree itself instead of relying on other CPUs to bring it work.
5279 * This adds some complexity to both (5) and (8) but it reduces the total idle
5287 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5290 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5295 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5297 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5299 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5302 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5303 * rewrite all of this once again.]
5306 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
5308 enum fbq_type
{ regular
, remote
, all
};
5310 #define LBF_ALL_PINNED 0x01
5311 #define LBF_NEED_BREAK 0x02
5312 #define LBF_DST_PINNED 0x04
5313 #define LBF_SOME_PINNED 0x08
5316 struct sched_domain
*sd
;
5324 struct cpumask
*dst_grpmask
;
5326 enum cpu_idle_type idle
;
5328 /* The set of CPUs under consideration for load-balancing */
5329 struct cpumask
*cpus
;
5334 unsigned int loop_break
;
5335 unsigned int loop_max
;
5337 enum fbq_type fbq_type
;
5338 struct list_head tasks
;
5342 * Is this task likely cache-hot:
5344 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
5348 lockdep_assert_held(&env
->src_rq
->lock
);
5350 if (p
->sched_class
!= &fair_sched_class
)
5353 if (unlikely(p
->policy
== SCHED_IDLE
))
5357 * Buddy candidates are cache hot:
5359 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
5360 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
5361 &p
->se
== cfs_rq_of(&p
->se
)->last
))
5364 if (sysctl_sched_migration_cost
== -1)
5366 if (sysctl_sched_migration_cost
== 0)
5369 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
5371 return delta
< (s64
)sysctl_sched_migration_cost
;
5374 #ifdef CONFIG_NUMA_BALANCING
5375 /* Returns true if the destination node has incurred more faults */
5376 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
5378 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5379 int src_nid
, dst_nid
;
5381 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults
||
5382 !(env
->sd
->flags
& SD_NUMA
)) {
5386 src_nid
= cpu_to_node(env
->src_cpu
);
5387 dst_nid
= cpu_to_node(env
->dst_cpu
);
5389 if (src_nid
== dst_nid
)
5393 /* Task is already in the group's interleave set. */
5394 if (node_isset(src_nid
, numa_group
->active_nodes
))
5397 /* Task is moving into the group's interleave set. */
5398 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5401 return group_faults(p
, dst_nid
) > group_faults(p
, src_nid
);
5404 /* Encourage migration to the preferred node. */
5405 if (dst_nid
== p
->numa_preferred_nid
)
5408 return task_faults(p
, dst_nid
) > task_faults(p
, src_nid
);
5412 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
5414 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5415 int src_nid
, dst_nid
;
5417 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
5420 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
5423 src_nid
= cpu_to_node(env
->src_cpu
);
5424 dst_nid
= cpu_to_node(env
->dst_cpu
);
5426 if (src_nid
== dst_nid
)
5430 /* Task is moving within/into the group's interleave set. */
5431 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5434 /* Task is moving out of the group's interleave set. */
5435 if (node_isset(src_nid
, numa_group
->active_nodes
))
5438 return group_faults(p
, dst_nid
) < group_faults(p
, src_nid
);
5441 /* Migrating away from the preferred node is always bad. */
5442 if (src_nid
== p
->numa_preferred_nid
)
5445 return task_faults(p
, dst_nid
) < task_faults(p
, src_nid
);
5449 static inline bool migrate_improves_locality(struct task_struct
*p
,
5455 static inline bool migrate_degrades_locality(struct task_struct
*p
,
5463 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5466 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
5468 int tsk_cache_hot
= 0;
5470 lockdep_assert_held(&env
->src_rq
->lock
);
5473 * We do not migrate tasks that are:
5474 * 1) throttled_lb_pair, or
5475 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5476 * 3) running (obviously), or
5477 * 4) are cache-hot on their current CPU.
5479 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
5482 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
5485 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
5487 env
->flags
|= LBF_SOME_PINNED
;
5490 * Remember if this task can be migrated to any other cpu in
5491 * our sched_group. We may want to revisit it if we couldn't
5492 * meet load balance goals by pulling other tasks on src_cpu.
5494 * Also avoid computing new_dst_cpu if we have already computed
5495 * one in current iteration.
5497 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
5500 /* Prevent to re-select dst_cpu via env's cpus */
5501 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
5502 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
5503 env
->flags
|= LBF_DST_PINNED
;
5504 env
->new_dst_cpu
= cpu
;
5512 /* Record that we found atleast one task that could run on dst_cpu */
5513 env
->flags
&= ~LBF_ALL_PINNED
;
5515 if (task_running(env
->src_rq
, p
)) {
5516 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
5521 * Aggressive migration if:
5522 * 1) destination numa is preferred
5523 * 2) task is cache cold, or
5524 * 3) too many balance attempts have failed.
5526 tsk_cache_hot
= task_hot(p
, env
);
5528 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
5530 if (migrate_improves_locality(p
, env
) || !tsk_cache_hot
||
5531 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
5532 if (tsk_cache_hot
) {
5533 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5534 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5539 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
5544 * detach_task() -- detach the task for the migration specified in env
5546 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
5548 lockdep_assert_held(&env
->src_rq
->lock
);
5550 deactivate_task(env
->src_rq
, p
, 0);
5551 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
5552 set_task_cpu(p
, env
->dst_cpu
);
5556 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5557 * part of active balancing operations within "domain".
5559 * Returns a task if successful and NULL otherwise.
5561 static struct task_struct
*detach_one_task(struct lb_env
*env
)
5563 struct task_struct
*p
, *n
;
5565 lockdep_assert_held(&env
->src_rq
->lock
);
5567 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
5568 if (!can_migrate_task(p
, env
))
5571 detach_task(p
, env
);
5574 * Right now, this is only the second place where
5575 * lb_gained[env->idle] is updated (other is detach_tasks)
5576 * so we can safely collect stats here rather than
5577 * inside detach_tasks().
5579 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
5585 static const unsigned int sched_nr_migrate_break
= 32;
5588 * detach_tasks() -- tries to detach up to imbalance weighted load from
5589 * busiest_rq, as part of a balancing operation within domain "sd".
5591 * Returns number of detached tasks if successful and 0 otherwise.
5593 static int detach_tasks(struct lb_env
*env
)
5595 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
5596 struct task_struct
*p
;
5600 lockdep_assert_held(&env
->src_rq
->lock
);
5602 if (env
->imbalance
<= 0)
5605 while (!list_empty(tasks
)) {
5606 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5609 /* We've more or less seen every task there is, call it quits */
5610 if (env
->loop
> env
->loop_max
)
5613 /* take a breather every nr_migrate tasks */
5614 if (env
->loop
> env
->loop_break
) {
5615 env
->loop_break
+= sched_nr_migrate_break
;
5616 env
->flags
|= LBF_NEED_BREAK
;
5620 if (!can_migrate_task(p
, env
))
5623 load
= task_h_load(p
);
5625 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5628 if ((load
/ 2) > env
->imbalance
)
5631 detach_task(p
, env
);
5632 list_add(&p
->se
.group_node
, &env
->tasks
);
5635 env
->imbalance
-= load
;
5637 #ifdef CONFIG_PREEMPT
5639 * NEWIDLE balancing is a source of latency, so preemptible
5640 * kernels will stop after the first task is detached to minimize
5641 * the critical section.
5643 if (env
->idle
== CPU_NEWLY_IDLE
)
5648 * We only want to steal up to the prescribed amount of
5651 if (env
->imbalance
<= 0)
5656 list_move_tail(&p
->se
.group_node
, tasks
);
5660 * Right now, this is one of only two places we collect this stat
5661 * so we can safely collect detach_one_task() stats here rather
5662 * than inside detach_one_task().
5664 schedstat_add(env
->sd
, lb_gained
[env
->idle
], detached
);
5670 * attach_task() -- attach the task detached by detach_task() to its new rq.
5672 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
5674 lockdep_assert_held(&rq
->lock
);
5676 BUG_ON(task_rq(p
) != rq
);
5677 p
->on_rq
= TASK_ON_RQ_QUEUED
;
5678 activate_task(rq
, p
, 0);
5679 check_preempt_curr(rq
, p
, 0);
5683 * attach_one_task() -- attaches the task returned from detach_one_task() to
5686 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
5688 raw_spin_lock(&rq
->lock
);
5690 raw_spin_unlock(&rq
->lock
);
5694 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5697 static void attach_tasks(struct lb_env
*env
)
5699 struct list_head
*tasks
= &env
->tasks
;
5700 struct task_struct
*p
;
5702 raw_spin_lock(&env
->dst_rq
->lock
);
5704 while (!list_empty(tasks
)) {
5705 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5706 list_del_init(&p
->se
.group_node
);
5708 attach_task(env
->dst_rq
, p
);
5711 raw_spin_unlock(&env
->dst_rq
->lock
);
5714 #ifdef CONFIG_FAIR_GROUP_SCHED
5716 * update tg->load_weight by folding this cpu's load_avg
5718 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5720 struct sched_entity
*se
= tg
->se
[cpu
];
5721 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5723 /* throttled entities do not contribute to load */
5724 if (throttled_hierarchy(cfs_rq
))
5727 update_cfs_rq_blocked_load(cfs_rq
, 1);
5730 update_entity_load_avg(se
, 1);
5732 * We pivot on our runnable average having decayed to zero for
5733 * list removal. This generally implies that all our children
5734 * have also been removed (modulo rounding error or bandwidth
5735 * control); however, such cases are rare and we can fix these
5738 * TODO: fix up out-of-order children on enqueue.
5740 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5741 list_del_leaf_cfs_rq(cfs_rq
);
5743 struct rq
*rq
= rq_of(cfs_rq
);
5744 update_rq_runnable_avg(rq
, rq
->nr_running
);
5748 static void update_blocked_averages(int cpu
)
5750 struct rq
*rq
= cpu_rq(cpu
);
5751 struct cfs_rq
*cfs_rq
;
5752 unsigned long flags
;
5754 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5755 update_rq_clock(rq
);
5757 * Iterates the task_group tree in a bottom up fashion, see
5758 * list_add_leaf_cfs_rq() for details.
5760 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5762 * Note: We may want to consider periodically releasing
5763 * rq->lock about these updates so that creating many task
5764 * groups does not result in continually extending hold time.
5766 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5769 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5773 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5774 * This needs to be done in a top-down fashion because the load of a child
5775 * group is a fraction of its parents load.
5777 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5779 struct rq
*rq
= rq_of(cfs_rq
);
5780 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5781 unsigned long now
= jiffies
;
5784 if (cfs_rq
->last_h_load_update
== now
)
5787 cfs_rq
->h_load_next
= NULL
;
5788 for_each_sched_entity(se
) {
5789 cfs_rq
= cfs_rq_of(se
);
5790 cfs_rq
->h_load_next
= se
;
5791 if (cfs_rq
->last_h_load_update
== now
)
5796 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5797 cfs_rq
->last_h_load_update
= now
;
5800 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5801 load
= cfs_rq
->h_load
;
5802 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5803 cfs_rq
->runnable_load_avg
+ 1);
5804 cfs_rq
= group_cfs_rq(se
);
5805 cfs_rq
->h_load
= load
;
5806 cfs_rq
->last_h_load_update
= now
;
5810 static unsigned long task_h_load(struct task_struct
*p
)
5812 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5814 update_cfs_rq_h_load(cfs_rq
);
5815 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5816 cfs_rq
->runnable_load_avg
+ 1);
5819 static inline void update_blocked_averages(int cpu
)
5823 static unsigned long task_h_load(struct task_struct
*p
)
5825 return p
->se
.avg
.load_avg_contrib
;
5829 /********** Helpers for find_busiest_group ************************/
5838 * sg_lb_stats - stats of a sched_group required for load_balancing
5840 struct sg_lb_stats
{
5841 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5842 unsigned long group_load
; /* Total load over the CPUs of the group */
5843 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5844 unsigned long load_per_task
;
5845 unsigned long group_capacity
;
5846 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5847 unsigned int group_capacity_factor
;
5848 unsigned int idle_cpus
;
5849 unsigned int group_weight
;
5850 enum group_type group_type
;
5851 int group_has_free_capacity
;
5852 #ifdef CONFIG_NUMA_BALANCING
5853 unsigned int nr_numa_running
;
5854 unsigned int nr_preferred_running
;
5859 * sd_lb_stats - Structure to store the statistics of a sched_domain
5860 * during load balancing.
5862 struct sd_lb_stats
{
5863 struct sched_group
*busiest
; /* Busiest group in this sd */
5864 struct sched_group
*local
; /* Local group in this sd */
5865 unsigned long total_load
; /* Total load of all groups in sd */
5866 unsigned long total_capacity
; /* Total capacity of all groups in sd */
5867 unsigned long avg_load
; /* Average load across all groups in sd */
5869 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5870 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5873 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5876 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5877 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5878 * We must however clear busiest_stat::avg_load because
5879 * update_sd_pick_busiest() reads this before assignment.
5881 *sds
= (struct sd_lb_stats
){
5885 .total_capacity
= 0UL,
5888 .sum_nr_running
= 0,
5889 .group_type
= group_other
,
5895 * get_sd_load_idx - Obtain the load index for a given sched domain.
5896 * @sd: The sched_domain whose load_idx is to be obtained.
5897 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5899 * Return: The load index.
5901 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5902 enum cpu_idle_type idle
)
5908 load_idx
= sd
->busy_idx
;
5911 case CPU_NEWLY_IDLE
:
5912 load_idx
= sd
->newidle_idx
;
5915 load_idx
= sd
->idle_idx
;
5922 static unsigned long default_scale_capacity(struct sched_domain
*sd
, int cpu
)
5924 return SCHED_CAPACITY_SCALE
;
5927 unsigned long __weak
arch_scale_freq_capacity(struct sched_domain
*sd
, int cpu
)
5929 return default_scale_capacity(sd
, cpu
);
5932 static unsigned long default_scale_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5934 if ((sd
->flags
& SD_SHARE_CPUCAPACITY
) && (sd
->span_weight
> 1))
5935 return sd
->smt_gain
/ sd
->span_weight
;
5937 return SCHED_CAPACITY_SCALE
;
5940 unsigned long __weak
arch_scale_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5942 return default_scale_cpu_capacity(sd
, cpu
);
5945 static unsigned long scale_rt_capacity(int cpu
)
5947 struct rq
*rq
= cpu_rq(cpu
);
5948 u64 total
, available
, age_stamp
, avg
;
5952 * Since we're reading these variables without serialization make sure
5953 * we read them once before doing sanity checks on them.
5955 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5956 avg
= ACCESS_ONCE(rq
->rt_avg
);
5957 delta
= __rq_clock_broken(rq
) - age_stamp
;
5959 if (unlikely(delta
< 0))
5962 total
= sched_avg_period() + delta
;
5964 if (unlikely(total
< avg
)) {
5965 /* Ensures that capacity won't end up being negative */
5968 available
= total
- avg
;
5971 if (unlikely((s64
)total
< SCHED_CAPACITY_SCALE
))
5972 total
= SCHED_CAPACITY_SCALE
;
5974 total
>>= SCHED_CAPACITY_SHIFT
;
5976 return div_u64(available
, total
);
5979 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5981 unsigned long capacity
= SCHED_CAPACITY_SCALE
;
5982 struct sched_group
*sdg
= sd
->groups
;
5984 if (sched_feat(ARCH_CAPACITY
))
5985 capacity
*= arch_scale_cpu_capacity(sd
, cpu
);
5987 capacity
*= default_scale_cpu_capacity(sd
, cpu
);
5989 capacity
>>= SCHED_CAPACITY_SHIFT
;
5991 sdg
->sgc
->capacity_orig
= capacity
;
5993 if (sched_feat(ARCH_CAPACITY
))
5994 capacity
*= arch_scale_freq_capacity(sd
, cpu
);
5996 capacity
*= default_scale_capacity(sd
, cpu
);
5998 capacity
>>= SCHED_CAPACITY_SHIFT
;
6000 capacity
*= scale_rt_capacity(cpu
);
6001 capacity
>>= SCHED_CAPACITY_SHIFT
;
6006 cpu_rq(cpu
)->cpu_capacity
= capacity
;
6007 sdg
->sgc
->capacity
= capacity
;
6010 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
6012 struct sched_domain
*child
= sd
->child
;
6013 struct sched_group
*group
, *sdg
= sd
->groups
;
6014 unsigned long capacity
, capacity_orig
;
6015 unsigned long interval
;
6017 interval
= msecs_to_jiffies(sd
->balance_interval
);
6018 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6019 sdg
->sgc
->next_update
= jiffies
+ interval
;
6022 update_cpu_capacity(sd
, cpu
);
6026 capacity_orig
= capacity
= 0;
6028 if (child
->flags
& SD_OVERLAP
) {
6030 * SD_OVERLAP domains cannot assume that child groups
6031 * span the current group.
6034 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
6035 struct sched_group_capacity
*sgc
;
6036 struct rq
*rq
= cpu_rq(cpu
);
6039 * build_sched_domains() -> init_sched_groups_capacity()
6040 * gets here before we've attached the domains to the
6043 * Use capacity_of(), which is set irrespective of domains
6044 * in update_cpu_capacity().
6046 * This avoids capacity/capacity_orig from being 0 and
6047 * causing divide-by-zero issues on boot.
6049 * Runtime updates will correct capacity_orig.
6051 if (unlikely(!rq
->sd
)) {
6052 capacity_orig
+= capacity_of(cpu
);
6053 capacity
+= capacity_of(cpu
);
6057 sgc
= rq
->sd
->groups
->sgc
;
6058 capacity_orig
+= sgc
->capacity_orig
;
6059 capacity
+= sgc
->capacity
;
6063 * !SD_OVERLAP domains can assume that child groups
6064 * span the current group.
6067 group
= child
->groups
;
6069 capacity_orig
+= group
->sgc
->capacity_orig
;
6070 capacity
+= group
->sgc
->capacity
;
6071 group
= group
->next
;
6072 } while (group
!= child
->groups
);
6075 sdg
->sgc
->capacity_orig
= capacity_orig
;
6076 sdg
->sgc
->capacity
= capacity
;
6080 * Try and fix up capacity for tiny siblings, this is needed when
6081 * things like SD_ASYM_PACKING need f_b_g to select another sibling
6082 * which on its own isn't powerful enough.
6084 * See update_sd_pick_busiest() and check_asym_packing().
6087 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
6090 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
6092 if (!(sd
->flags
& SD_SHARE_CPUCAPACITY
))
6096 * If ~90% of the cpu_capacity is still there, we're good.
6098 if (group
->sgc
->capacity
* 32 > group
->sgc
->capacity_orig
* 29)
6105 * Group imbalance indicates (and tries to solve) the problem where balancing
6106 * groups is inadequate due to tsk_cpus_allowed() constraints.
6108 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6109 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6112 * { 0 1 2 3 } { 4 5 6 7 }
6115 * If we were to balance group-wise we'd place two tasks in the first group and
6116 * two tasks in the second group. Clearly this is undesired as it will overload
6117 * cpu 3 and leave one of the cpus in the second group unused.
6119 * The current solution to this issue is detecting the skew in the first group
6120 * by noticing the lower domain failed to reach balance and had difficulty
6121 * moving tasks due to affinity constraints.
6123 * When this is so detected; this group becomes a candidate for busiest; see
6124 * update_sd_pick_busiest(). And calculate_imbalance() and
6125 * find_busiest_group() avoid some of the usual balance conditions to allow it
6126 * to create an effective group imbalance.
6128 * This is a somewhat tricky proposition since the next run might not find the
6129 * group imbalance and decide the groups need to be balanced again. A most
6130 * subtle and fragile situation.
6133 static inline int sg_imbalanced(struct sched_group
*group
)
6135 return group
->sgc
->imbalance
;
6139 * Compute the group capacity factor.
6141 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
6142 * first dividing out the smt factor and computing the actual number of cores
6143 * and limit unit capacity with that.
6145 static inline int sg_capacity_factor(struct lb_env
*env
, struct sched_group
*group
)
6147 unsigned int capacity_factor
, smt
, cpus
;
6148 unsigned int capacity
, capacity_orig
;
6150 capacity
= group
->sgc
->capacity
;
6151 capacity_orig
= group
->sgc
->capacity_orig
;
6152 cpus
= group
->group_weight
;
6154 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
6155 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, capacity_orig
);
6156 capacity_factor
= cpus
/ smt
; /* cores */
6158 capacity_factor
= min_t(unsigned,
6159 capacity_factor
, DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
));
6160 if (!capacity_factor
)
6161 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6163 return capacity_factor
;
6166 static enum group_type
6167 group_classify(struct sched_group
*group
, struct sg_lb_stats
*sgs
)
6169 if (sgs
->sum_nr_running
> sgs
->group_capacity_factor
)
6170 return group_overloaded
;
6172 if (sg_imbalanced(group
))
6173 return group_imbalanced
;
6179 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6180 * @env: The load balancing environment.
6181 * @group: sched_group whose statistics are to be updated.
6182 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6183 * @local_group: Does group contain this_cpu.
6184 * @sgs: variable to hold the statistics for this group.
6185 * @overload: Indicate more than one runnable task for any CPU.
6187 static inline void update_sg_lb_stats(struct lb_env
*env
,
6188 struct sched_group
*group
, int load_idx
,
6189 int local_group
, struct sg_lb_stats
*sgs
,
6195 memset(sgs
, 0, sizeof(*sgs
));
6197 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6198 struct rq
*rq
= cpu_rq(i
);
6200 /* Bias balancing toward cpus of our domain */
6202 load
= target_load(i
, load_idx
);
6204 load
= source_load(i
, load_idx
);
6206 sgs
->group_load
+= load
;
6207 sgs
->sum_nr_running
+= rq
->cfs
.h_nr_running
;
6209 if (rq
->nr_running
> 1)
6212 #ifdef CONFIG_NUMA_BALANCING
6213 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
6214 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
6216 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
6221 /* Adjust by relative CPU capacity of the group */
6222 sgs
->group_capacity
= group
->sgc
->capacity
;
6223 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
6225 if (sgs
->sum_nr_running
)
6226 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
6228 sgs
->group_weight
= group
->group_weight
;
6229 sgs
->group_capacity_factor
= sg_capacity_factor(env
, group
);
6230 sgs
->group_type
= group_classify(group
, sgs
);
6232 if (sgs
->group_capacity_factor
> sgs
->sum_nr_running
)
6233 sgs
->group_has_free_capacity
= 1;
6237 * update_sd_pick_busiest - return 1 on busiest group
6238 * @env: The load balancing environment.
6239 * @sds: sched_domain statistics
6240 * @sg: sched_group candidate to be checked for being the busiest
6241 * @sgs: sched_group statistics
6243 * Determine if @sg is a busier group than the previously selected
6246 * Return: %true if @sg is a busier group than the previously selected
6247 * busiest group. %false otherwise.
6249 static bool update_sd_pick_busiest(struct lb_env
*env
,
6250 struct sd_lb_stats
*sds
,
6251 struct sched_group
*sg
,
6252 struct sg_lb_stats
*sgs
)
6254 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
6256 if (sgs
->group_type
> busiest
->group_type
)
6259 if (sgs
->group_type
< busiest
->group_type
)
6262 if (sgs
->avg_load
<= busiest
->avg_load
)
6265 /* This is the busiest node in its class. */
6266 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6270 * ASYM_PACKING needs to move all the work to the lowest
6271 * numbered CPUs in the group, therefore mark all groups
6272 * higher than ourself as busy.
6274 if (sgs
->sum_nr_running
&& env
->dst_cpu
< group_first_cpu(sg
)) {
6278 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
6285 #ifdef CONFIG_NUMA_BALANCING
6286 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6288 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
6290 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
6295 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6297 if (rq
->nr_running
> rq
->nr_numa_running
)
6299 if (rq
->nr_running
> rq
->nr_preferred_running
)
6304 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6309 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6313 #endif /* CONFIG_NUMA_BALANCING */
6316 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6317 * @env: The load balancing environment.
6318 * @sds: variable to hold the statistics for this sched_domain.
6320 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6322 struct sched_domain
*child
= env
->sd
->child
;
6323 struct sched_group
*sg
= env
->sd
->groups
;
6324 struct sg_lb_stats tmp_sgs
;
6325 int load_idx
, prefer_sibling
= 0;
6326 bool overload
= false;
6328 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
6331 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
6334 struct sg_lb_stats
*sgs
= &tmp_sgs
;
6337 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
6340 sgs
= &sds
->local_stat
;
6342 if (env
->idle
!= CPU_NEWLY_IDLE
||
6343 time_after_eq(jiffies
, sg
->sgc
->next_update
))
6344 update_group_capacity(env
->sd
, env
->dst_cpu
);
6347 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
,
6354 * In case the child domain prefers tasks go to siblings
6355 * first, lower the sg capacity factor to one so that we'll try
6356 * and move all the excess tasks away. We lower the capacity
6357 * of a group only if the local group has the capacity to fit
6358 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6359 * extra check prevents the case where you always pull from the
6360 * heaviest group when it is already under-utilized (possible
6361 * with a large weight task outweighs the tasks on the system).
6363 if (prefer_sibling
&& sds
->local
&&
6364 sds
->local_stat
.group_has_free_capacity
) {
6365 sgs
->group_capacity_factor
= min(sgs
->group_capacity_factor
, 1U);
6366 sgs
->group_type
= group_classify(sg
, sgs
);
6369 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
6371 sds
->busiest_stat
= *sgs
;
6375 /* Now, start updating sd_lb_stats */
6376 sds
->total_load
+= sgs
->group_load
;
6377 sds
->total_capacity
+= sgs
->group_capacity
;
6380 } while (sg
!= env
->sd
->groups
);
6382 if (env
->sd
->flags
& SD_NUMA
)
6383 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
6385 if (!env
->sd
->parent
) {
6386 /* update overload indicator if we are at root domain */
6387 if (env
->dst_rq
->rd
->overload
!= overload
)
6388 env
->dst_rq
->rd
->overload
= overload
;
6394 * check_asym_packing - Check to see if the group is packed into the
6397 * This is primarily intended to used at the sibling level. Some
6398 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6399 * case of POWER7, it can move to lower SMT modes only when higher
6400 * threads are idle. When in lower SMT modes, the threads will
6401 * perform better since they share less core resources. Hence when we
6402 * have idle threads, we want them to be the higher ones.
6404 * This packing function is run on idle threads. It checks to see if
6405 * the busiest CPU in this domain (core in the P7 case) has a higher
6406 * CPU number than the packing function is being run on. Here we are
6407 * assuming lower CPU number will be equivalent to lower a SMT thread
6410 * Return: 1 when packing is required and a task should be moved to
6411 * this CPU. The amount of the imbalance is returned in *imbalance.
6413 * @env: The load balancing environment.
6414 * @sds: Statistics of the sched_domain which is to be packed
6416 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6420 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6426 busiest_cpu
= group_first_cpu(sds
->busiest
);
6427 if (env
->dst_cpu
> busiest_cpu
)
6430 env
->imbalance
= DIV_ROUND_CLOSEST(
6431 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
6432 SCHED_CAPACITY_SCALE
);
6438 * fix_small_imbalance - Calculate the minor imbalance that exists
6439 * amongst the groups of a sched_domain, during
6441 * @env: The load balancing environment.
6442 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6445 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6447 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
6448 unsigned int imbn
= 2;
6449 unsigned long scaled_busy_load_per_task
;
6450 struct sg_lb_stats
*local
, *busiest
;
6452 local
= &sds
->local_stat
;
6453 busiest
= &sds
->busiest_stat
;
6455 if (!local
->sum_nr_running
)
6456 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
6457 else if (busiest
->load_per_task
> local
->load_per_task
)
6460 scaled_busy_load_per_task
=
6461 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6462 busiest
->group_capacity
;
6464 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
6465 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
6466 env
->imbalance
= busiest
->load_per_task
;
6471 * OK, we don't have enough imbalance to justify moving tasks,
6472 * however we may be able to increase total CPU capacity used by
6476 capa_now
+= busiest
->group_capacity
*
6477 min(busiest
->load_per_task
, busiest
->avg_load
);
6478 capa_now
+= local
->group_capacity
*
6479 min(local
->load_per_task
, local
->avg_load
);
6480 capa_now
/= SCHED_CAPACITY_SCALE
;
6482 /* Amount of load we'd subtract */
6483 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
6484 capa_move
+= busiest
->group_capacity
*
6485 min(busiest
->load_per_task
,
6486 busiest
->avg_load
- scaled_busy_load_per_task
);
6489 /* Amount of load we'd add */
6490 if (busiest
->avg_load
* busiest
->group_capacity
<
6491 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
6492 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
6493 local
->group_capacity
;
6495 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6496 local
->group_capacity
;
6498 capa_move
+= local
->group_capacity
*
6499 min(local
->load_per_task
, local
->avg_load
+ tmp
);
6500 capa_move
/= SCHED_CAPACITY_SCALE
;
6502 /* Move if we gain throughput */
6503 if (capa_move
> capa_now
)
6504 env
->imbalance
= busiest
->load_per_task
;
6508 * calculate_imbalance - Calculate the amount of imbalance present within the
6509 * groups of a given sched_domain during load balance.
6510 * @env: load balance environment
6511 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6513 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6515 unsigned long max_pull
, load_above_capacity
= ~0UL;
6516 struct sg_lb_stats
*local
, *busiest
;
6518 local
= &sds
->local_stat
;
6519 busiest
= &sds
->busiest_stat
;
6521 if (busiest
->group_type
== group_imbalanced
) {
6523 * In the group_imb case we cannot rely on group-wide averages
6524 * to ensure cpu-load equilibrium, look at wider averages. XXX
6526 busiest
->load_per_task
=
6527 min(busiest
->load_per_task
, sds
->avg_load
);
6531 * In the presence of smp nice balancing, certain scenarios can have
6532 * max load less than avg load(as we skip the groups at or below
6533 * its cpu_capacity, while calculating max_load..)
6535 if (busiest
->avg_load
<= sds
->avg_load
||
6536 local
->avg_load
>= sds
->avg_load
) {
6538 return fix_small_imbalance(env
, sds
);
6542 * If there aren't any idle cpus, avoid creating some.
6544 if (busiest
->group_type
== group_overloaded
&&
6545 local
->group_type
== group_overloaded
) {
6546 load_above_capacity
=
6547 (busiest
->sum_nr_running
- busiest
->group_capacity_factor
);
6549 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_CAPACITY_SCALE
);
6550 load_above_capacity
/= busiest
->group_capacity
;
6554 * We're trying to get all the cpus to the average_load, so we don't
6555 * want to push ourselves above the average load, nor do we wish to
6556 * reduce the max loaded cpu below the average load. At the same time,
6557 * we also don't want to reduce the group load below the group capacity
6558 * (so that we can implement power-savings policies etc). Thus we look
6559 * for the minimum possible imbalance.
6561 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
6563 /* How much load to actually move to equalise the imbalance */
6564 env
->imbalance
= min(
6565 max_pull
* busiest
->group_capacity
,
6566 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
6567 ) / SCHED_CAPACITY_SCALE
;
6570 * if *imbalance is less than the average load per runnable task
6571 * there is no guarantee that any tasks will be moved so we'll have
6572 * a think about bumping its value to force at least one task to be
6575 if (env
->imbalance
< busiest
->load_per_task
)
6576 return fix_small_imbalance(env
, sds
);
6579 /******* find_busiest_group() helpers end here *********************/
6582 * find_busiest_group - Returns the busiest group within the sched_domain
6583 * if there is an imbalance. If there isn't an imbalance, and
6584 * the user has opted for power-savings, it returns a group whose
6585 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6586 * such a group exists.
6588 * Also calculates the amount of weighted load which should be moved
6589 * to restore balance.
6591 * @env: The load balancing environment.
6593 * Return: - The busiest group if imbalance exists.
6594 * - If no imbalance and user has opted for power-savings balance,
6595 * return the least loaded group whose CPUs can be
6596 * put to idle by rebalancing its tasks onto our group.
6598 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
6600 struct sg_lb_stats
*local
, *busiest
;
6601 struct sd_lb_stats sds
;
6603 init_sd_lb_stats(&sds
);
6606 * Compute the various statistics relavent for load balancing at
6609 update_sd_lb_stats(env
, &sds
);
6610 local
= &sds
.local_stat
;
6611 busiest
= &sds
.busiest_stat
;
6613 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
6614 check_asym_packing(env
, &sds
))
6617 /* There is no busy sibling group to pull tasks from */
6618 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
6621 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
6622 / sds
.total_capacity
;
6625 * If the busiest group is imbalanced the below checks don't
6626 * work because they assume all things are equal, which typically
6627 * isn't true due to cpus_allowed constraints and the like.
6629 if (busiest
->group_type
== group_imbalanced
)
6632 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6633 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_free_capacity
&&
6634 !busiest
->group_has_free_capacity
)
6638 * If the local group is busier than the selected busiest group
6639 * don't try and pull any tasks.
6641 if (local
->avg_load
>= busiest
->avg_load
)
6645 * Don't pull any tasks if this group is already above the domain
6648 if (local
->avg_load
>= sds
.avg_load
)
6651 if (env
->idle
== CPU_IDLE
) {
6653 * This cpu is idle. If the busiest group is not overloaded
6654 * and there is no imbalance between this and busiest group
6655 * wrt idle cpus, it is balanced. The imbalance becomes
6656 * significant if the diff is greater than 1 otherwise we
6657 * might end up to just move the imbalance on another group
6659 if ((busiest
->group_type
!= group_overloaded
) &&
6660 (local
->idle_cpus
<= (busiest
->idle_cpus
+ 1)))
6664 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6665 * imbalance_pct to be conservative.
6667 if (100 * busiest
->avg_load
<=
6668 env
->sd
->imbalance_pct
* local
->avg_load
)
6673 /* Looks like there is an imbalance. Compute it */
6674 calculate_imbalance(env
, &sds
);
6683 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6685 static struct rq
*find_busiest_queue(struct lb_env
*env
,
6686 struct sched_group
*group
)
6688 struct rq
*busiest
= NULL
, *rq
;
6689 unsigned long busiest_load
= 0, busiest_capacity
= 1;
6692 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6693 unsigned long capacity
, capacity_factor
, wl
;
6697 rt
= fbq_classify_rq(rq
);
6700 * We classify groups/runqueues into three groups:
6701 * - regular: there are !numa tasks
6702 * - remote: there are numa tasks that run on the 'wrong' node
6703 * - all: there is no distinction
6705 * In order to avoid migrating ideally placed numa tasks,
6706 * ignore those when there's better options.
6708 * If we ignore the actual busiest queue to migrate another
6709 * task, the next balance pass can still reduce the busiest
6710 * queue by moving tasks around inside the node.
6712 * If we cannot move enough load due to this classification
6713 * the next pass will adjust the group classification and
6714 * allow migration of more tasks.
6716 * Both cases only affect the total convergence complexity.
6718 if (rt
> env
->fbq_type
)
6721 capacity
= capacity_of(i
);
6722 capacity_factor
= DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
);
6723 if (!capacity_factor
)
6724 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6726 wl
= weighted_cpuload(i
);
6729 * When comparing with imbalance, use weighted_cpuload()
6730 * which is not scaled with the cpu capacity.
6732 if (capacity_factor
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6736 * For the load comparisons with the other cpu's, consider
6737 * the weighted_cpuload() scaled with the cpu capacity, so
6738 * that the load can be moved away from the cpu that is
6739 * potentially running at a lower capacity.
6741 * Thus we're looking for max(wl_i / capacity_i), crosswise
6742 * multiplication to rid ourselves of the division works out
6743 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6744 * our previous maximum.
6746 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
6748 busiest_capacity
= capacity
;
6757 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6758 * so long as it is large enough.
6760 #define MAX_PINNED_INTERVAL 512
6762 /* Working cpumask for load_balance and load_balance_newidle. */
6763 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6765 static int need_active_balance(struct lb_env
*env
)
6767 struct sched_domain
*sd
= env
->sd
;
6769 if (env
->idle
== CPU_NEWLY_IDLE
) {
6772 * ASYM_PACKING needs to force migrate tasks from busy but
6773 * higher numbered CPUs in order to pack all tasks in the
6774 * lowest numbered CPUs.
6776 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6780 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6783 static int active_load_balance_cpu_stop(void *data
);
6785 static int should_we_balance(struct lb_env
*env
)
6787 struct sched_group
*sg
= env
->sd
->groups
;
6788 struct cpumask
*sg_cpus
, *sg_mask
;
6789 int cpu
, balance_cpu
= -1;
6792 * In the newly idle case, we will allow all the cpu's
6793 * to do the newly idle load balance.
6795 if (env
->idle
== CPU_NEWLY_IDLE
)
6798 sg_cpus
= sched_group_cpus(sg
);
6799 sg_mask
= sched_group_mask(sg
);
6800 /* Try to find first idle cpu */
6801 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6802 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6809 if (balance_cpu
== -1)
6810 balance_cpu
= group_balance_cpu(sg
);
6813 * First idle cpu or the first cpu(busiest) in this sched group
6814 * is eligible for doing load balancing at this and above domains.
6816 return balance_cpu
== env
->dst_cpu
;
6820 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6821 * tasks if there is an imbalance.
6823 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6824 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6825 int *continue_balancing
)
6827 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6828 struct sched_domain
*sd_parent
= sd
->parent
;
6829 struct sched_group
*group
;
6831 unsigned long flags
;
6832 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
6834 struct lb_env env
= {
6836 .dst_cpu
= this_cpu
,
6838 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6840 .loop_break
= sched_nr_migrate_break
,
6843 .tasks
= LIST_HEAD_INIT(env
.tasks
),
6847 * For NEWLY_IDLE load_balancing, we don't need to consider
6848 * other cpus in our group
6850 if (idle
== CPU_NEWLY_IDLE
)
6851 env
.dst_grpmask
= NULL
;
6853 cpumask_copy(cpus
, cpu_active_mask
);
6855 schedstat_inc(sd
, lb_count
[idle
]);
6858 if (!should_we_balance(&env
)) {
6859 *continue_balancing
= 0;
6863 group
= find_busiest_group(&env
);
6865 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6869 busiest
= find_busiest_queue(&env
, group
);
6871 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6875 BUG_ON(busiest
== env
.dst_rq
);
6877 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6880 if (busiest
->nr_running
> 1) {
6882 * Attempt to move tasks. If find_busiest_group has found
6883 * an imbalance but busiest->nr_running <= 1, the group is
6884 * still unbalanced. ld_moved simply stays zero, so it is
6885 * correctly treated as an imbalance.
6887 env
.flags
|= LBF_ALL_PINNED
;
6888 env
.src_cpu
= busiest
->cpu
;
6889 env
.src_rq
= busiest
;
6890 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6893 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6896 * cur_ld_moved - load moved in current iteration
6897 * ld_moved - cumulative load moved across iterations
6899 cur_ld_moved
= detach_tasks(&env
);
6902 * We've detached some tasks from busiest_rq. Every
6903 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6904 * unlock busiest->lock, and we are able to be sure
6905 * that nobody can manipulate the tasks in parallel.
6906 * See task_rq_lock() family for the details.
6909 raw_spin_unlock(&busiest
->lock
);
6913 ld_moved
+= cur_ld_moved
;
6916 local_irq_restore(flags
);
6918 if (env
.flags
& LBF_NEED_BREAK
) {
6919 env
.flags
&= ~LBF_NEED_BREAK
;
6924 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6925 * us and move them to an alternate dst_cpu in our sched_group
6926 * where they can run. The upper limit on how many times we
6927 * iterate on same src_cpu is dependent on number of cpus in our
6930 * This changes load balance semantics a bit on who can move
6931 * load to a given_cpu. In addition to the given_cpu itself
6932 * (or a ilb_cpu acting on its behalf where given_cpu is
6933 * nohz-idle), we now have balance_cpu in a position to move
6934 * load to given_cpu. In rare situations, this may cause
6935 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6936 * _independently_ and at _same_ time to move some load to
6937 * given_cpu) causing exceess load to be moved to given_cpu.
6938 * This however should not happen so much in practice and
6939 * moreover subsequent load balance cycles should correct the
6940 * excess load moved.
6942 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6944 /* Prevent to re-select dst_cpu via env's cpus */
6945 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6947 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6948 env
.dst_cpu
= env
.new_dst_cpu
;
6949 env
.flags
&= ~LBF_DST_PINNED
;
6951 env
.loop_break
= sched_nr_migrate_break
;
6954 * Go back to "more_balance" rather than "redo" since we
6955 * need to continue with same src_cpu.
6961 * We failed to reach balance because of affinity.
6964 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
6966 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
6967 *group_imbalance
= 1;
6970 /* All tasks on this runqueue were pinned by CPU affinity */
6971 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6972 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6973 if (!cpumask_empty(cpus
)) {
6975 env
.loop_break
= sched_nr_migrate_break
;
6978 goto out_all_pinned
;
6983 schedstat_inc(sd
, lb_failed
[idle
]);
6985 * Increment the failure counter only on periodic balance.
6986 * We do not want newidle balance, which can be very
6987 * frequent, pollute the failure counter causing
6988 * excessive cache_hot migrations and active balances.
6990 if (idle
!= CPU_NEWLY_IDLE
)
6991 sd
->nr_balance_failed
++;
6993 if (need_active_balance(&env
)) {
6994 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6996 /* don't kick the active_load_balance_cpu_stop,
6997 * if the curr task on busiest cpu can't be
7000 if (!cpumask_test_cpu(this_cpu
,
7001 tsk_cpus_allowed(busiest
->curr
))) {
7002 raw_spin_unlock_irqrestore(&busiest
->lock
,
7004 env
.flags
|= LBF_ALL_PINNED
;
7005 goto out_one_pinned
;
7009 * ->active_balance synchronizes accesses to
7010 * ->active_balance_work. Once set, it's cleared
7011 * only after active load balance is finished.
7013 if (!busiest
->active_balance
) {
7014 busiest
->active_balance
= 1;
7015 busiest
->push_cpu
= this_cpu
;
7018 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
7020 if (active_balance
) {
7021 stop_one_cpu_nowait(cpu_of(busiest
),
7022 active_load_balance_cpu_stop
, busiest
,
7023 &busiest
->active_balance_work
);
7027 * We've kicked active balancing, reset the failure
7030 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
7033 sd
->nr_balance_failed
= 0;
7035 if (likely(!active_balance
)) {
7036 /* We were unbalanced, so reset the balancing interval */
7037 sd
->balance_interval
= sd
->min_interval
;
7040 * If we've begun active balancing, start to back off. This
7041 * case may not be covered by the all_pinned logic if there
7042 * is only 1 task on the busy runqueue (because we don't call
7045 if (sd
->balance_interval
< sd
->max_interval
)
7046 sd
->balance_interval
*= 2;
7053 * We reach balance although we may have faced some affinity
7054 * constraints. Clear the imbalance flag if it was set.
7057 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
7059 if (*group_imbalance
)
7060 *group_imbalance
= 0;
7065 * We reach balance because all tasks are pinned at this level so
7066 * we can't migrate them. Let the imbalance flag set so parent level
7067 * can try to migrate them.
7069 schedstat_inc(sd
, lb_balanced
[idle
]);
7071 sd
->nr_balance_failed
= 0;
7074 /* tune up the balancing interval */
7075 if (((env
.flags
& LBF_ALL_PINNED
) &&
7076 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
7077 (sd
->balance_interval
< sd
->max_interval
))
7078 sd
->balance_interval
*= 2;
7085 static inline unsigned long
7086 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
7088 unsigned long interval
= sd
->balance_interval
;
7091 interval
*= sd
->busy_factor
;
7093 /* scale ms to jiffies */
7094 interval
= msecs_to_jiffies(interval
);
7095 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
7101 update_next_balance(struct sched_domain
*sd
, int cpu_busy
, unsigned long *next_balance
)
7103 unsigned long interval
, next
;
7105 interval
= get_sd_balance_interval(sd
, cpu_busy
);
7106 next
= sd
->last_balance
+ interval
;
7108 if (time_after(*next_balance
, next
))
7109 *next_balance
= next
;
7113 * idle_balance is called by schedule() if this_cpu is about to become
7114 * idle. Attempts to pull tasks from other CPUs.
7116 static int idle_balance(struct rq
*this_rq
)
7118 unsigned long next_balance
= jiffies
+ HZ
;
7119 int this_cpu
= this_rq
->cpu
;
7120 struct sched_domain
*sd
;
7121 int pulled_task
= 0;
7124 idle_enter_fair(this_rq
);
7127 * We must set idle_stamp _before_ calling idle_balance(), such that we
7128 * measure the duration of idle_balance() as idle time.
7130 this_rq
->idle_stamp
= rq_clock(this_rq
);
7132 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
7133 !this_rq
->rd
->overload
) {
7135 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
7137 update_next_balance(sd
, 0, &next_balance
);
7144 * Drop the rq->lock, but keep IRQ/preempt disabled.
7146 raw_spin_unlock(&this_rq
->lock
);
7148 update_blocked_averages(this_cpu
);
7150 for_each_domain(this_cpu
, sd
) {
7151 int continue_balancing
= 1;
7152 u64 t0
, domain_cost
;
7154 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7157 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
7158 update_next_balance(sd
, 0, &next_balance
);
7162 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
7163 t0
= sched_clock_cpu(this_cpu
);
7165 pulled_task
= load_balance(this_cpu
, this_rq
,
7167 &continue_balancing
);
7169 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
7170 if (domain_cost
> sd
->max_newidle_lb_cost
)
7171 sd
->max_newidle_lb_cost
= domain_cost
;
7173 curr_cost
+= domain_cost
;
7176 update_next_balance(sd
, 0, &next_balance
);
7179 * Stop searching for tasks to pull if there are
7180 * now runnable tasks on this rq.
7182 if (pulled_task
|| this_rq
->nr_running
> 0)
7187 raw_spin_lock(&this_rq
->lock
);
7189 if (curr_cost
> this_rq
->max_idle_balance_cost
)
7190 this_rq
->max_idle_balance_cost
= curr_cost
;
7193 * While browsing the domains, we released the rq lock, a task could
7194 * have been enqueued in the meantime. Since we're not going idle,
7195 * pretend we pulled a task.
7197 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
7201 /* Move the next balance forward */
7202 if (time_after(this_rq
->next_balance
, next_balance
))
7203 this_rq
->next_balance
= next_balance
;
7205 /* Is there a task of a high priority class? */
7206 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
7210 idle_exit_fair(this_rq
);
7211 this_rq
->idle_stamp
= 0;
7218 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7219 * running tasks off the busiest CPU onto idle CPUs. It requires at
7220 * least 1 task to be running on each physical CPU where possible, and
7221 * avoids physical / logical imbalances.
7223 static int active_load_balance_cpu_stop(void *data
)
7225 struct rq
*busiest_rq
= data
;
7226 int busiest_cpu
= cpu_of(busiest_rq
);
7227 int target_cpu
= busiest_rq
->push_cpu
;
7228 struct rq
*target_rq
= cpu_rq(target_cpu
);
7229 struct sched_domain
*sd
;
7230 struct task_struct
*p
= NULL
;
7232 raw_spin_lock_irq(&busiest_rq
->lock
);
7234 /* make sure the requested cpu hasn't gone down in the meantime */
7235 if (unlikely(busiest_cpu
!= smp_processor_id() ||
7236 !busiest_rq
->active_balance
))
7239 /* Is there any task to move? */
7240 if (busiest_rq
->nr_running
<= 1)
7244 * This condition is "impossible", if it occurs
7245 * we need to fix it. Originally reported by
7246 * Bjorn Helgaas on a 128-cpu setup.
7248 BUG_ON(busiest_rq
== target_rq
);
7250 /* Search for an sd spanning us and the target CPU. */
7252 for_each_domain(target_cpu
, sd
) {
7253 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
7254 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
7259 struct lb_env env
= {
7261 .dst_cpu
= target_cpu
,
7262 .dst_rq
= target_rq
,
7263 .src_cpu
= busiest_rq
->cpu
,
7264 .src_rq
= busiest_rq
,
7268 schedstat_inc(sd
, alb_count
);
7270 p
= detach_one_task(&env
);
7272 schedstat_inc(sd
, alb_pushed
);
7274 schedstat_inc(sd
, alb_failed
);
7278 busiest_rq
->active_balance
= 0;
7279 raw_spin_unlock(&busiest_rq
->lock
);
7282 attach_one_task(target_rq
, p
);
7289 static inline int on_null_domain(struct rq
*rq
)
7291 return unlikely(!rcu_dereference_sched(rq
->sd
));
7294 #ifdef CONFIG_NO_HZ_COMMON
7296 * idle load balancing details
7297 * - When one of the busy CPUs notice that there may be an idle rebalancing
7298 * needed, they will kick the idle load balancer, which then does idle
7299 * load balancing for all the idle CPUs.
7302 cpumask_var_t idle_cpus_mask
;
7304 unsigned long next_balance
; /* in jiffy units */
7305 } nohz ____cacheline_aligned
;
7307 static inline int find_new_ilb(void)
7309 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
7311 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
7318 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7319 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7320 * CPU (if there is one).
7322 static void nohz_balancer_kick(void)
7326 nohz
.next_balance
++;
7328 ilb_cpu
= find_new_ilb();
7330 if (ilb_cpu
>= nr_cpu_ids
)
7333 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
7336 * Use smp_send_reschedule() instead of resched_cpu().
7337 * This way we generate a sched IPI on the target cpu which
7338 * is idle. And the softirq performing nohz idle load balance
7339 * will be run before returning from the IPI.
7341 smp_send_reschedule(ilb_cpu
);
7345 static inline void nohz_balance_exit_idle(int cpu
)
7347 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
7349 * Completely isolated CPUs don't ever set, so we must test.
7351 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
7352 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
7353 atomic_dec(&nohz
.nr_cpus
);
7355 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7359 static inline void set_cpu_sd_state_busy(void)
7361 struct sched_domain
*sd
;
7362 int cpu
= smp_processor_id();
7365 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7367 if (!sd
|| !sd
->nohz_idle
)
7371 atomic_inc(&sd
->groups
->sgc
->nr_busy_cpus
);
7376 void set_cpu_sd_state_idle(void)
7378 struct sched_domain
*sd
;
7379 int cpu
= smp_processor_id();
7382 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7384 if (!sd
|| sd
->nohz_idle
)
7388 atomic_dec(&sd
->groups
->sgc
->nr_busy_cpus
);
7394 * This routine will record that the cpu is going idle with tick stopped.
7395 * This info will be used in performing idle load balancing in the future.
7397 void nohz_balance_enter_idle(int cpu
)
7400 * If this cpu is going down, then nothing needs to be done.
7402 if (!cpu_active(cpu
))
7405 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
7409 * If we're a completely isolated CPU, we don't play.
7411 if (on_null_domain(cpu_rq(cpu
)))
7414 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
7415 atomic_inc(&nohz
.nr_cpus
);
7416 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7419 static int sched_ilb_notifier(struct notifier_block
*nfb
,
7420 unsigned long action
, void *hcpu
)
7422 switch (action
& ~CPU_TASKS_FROZEN
) {
7424 nohz_balance_exit_idle(smp_processor_id());
7432 static DEFINE_SPINLOCK(balancing
);
7435 * Scale the max load_balance interval with the number of CPUs in the system.
7436 * This trades load-balance latency on larger machines for less cross talk.
7438 void update_max_interval(void)
7440 max_load_balance_interval
= HZ
*num_online_cpus()/10;
7444 * It checks each scheduling domain to see if it is due to be balanced,
7445 * and initiates a balancing operation if so.
7447 * Balancing parameters are set up in init_sched_domains.
7449 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
7451 int continue_balancing
= 1;
7453 unsigned long interval
;
7454 struct sched_domain
*sd
;
7455 /* Earliest time when we have to do rebalance again */
7456 unsigned long next_balance
= jiffies
+ 60*HZ
;
7457 int update_next_balance
= 0;
7458 int need_serialize
, need_decay
= 0;
7461 update_blocked_averages(cpu
);
7464 for_each_domain(cpu
, sd
) {
7466 * Decay the newidle max times here because this is a regular
7467 * visit to all the domains. Decay ~1% per second.
7469 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
7470 sd
->max_newidle_lb_cost
=
7471 (sd
->max_newidle_lb_cost
* 253) / 256;
7472 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
7475 max_cost
+= sd
->max_newidle_lb_cost
;
7477 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7481 * Stop the load balance at this level. There is another
7482 * CPU in our sched group which is doing load balancing more
7485 if (!continue_balancing
) {
7491 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7493 need_serialize
= sd
->flags
& SD_SERIALIZE
;
7494 if (need_serialize
) {
7495 if (!spin_trylock(&balancing
))
7499 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
7500 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
7502 * The LBF_DST_PINNED logic could have changed
7503 * env->dst_cpu, so we can't know our idle
7504 * state even if we migrated tasks. Update it.
7506 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
7508 sd
->last_balance
= jiffies
;
7509 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7512 spin_unlock(&balancing
);
7514 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
7515 next_balance
= sd
->last_balance
+ interval
;
7516 update_next_balance
= 1;
7521 * Ensure the rq-wide value also decays but keep it at a
7522 * reasonable floor to avoid funnies with rq->avg_idle.
7524 rq
->max_idle_balance_cost
=
7525 max((u64
)sysctl_sched_migration_cost
, max_cost
);
7530 * next_balance will be updated only when there is a need.
7531 * When the cpu is attached to null domain for ex, it will not be
7534 if (likely(update_next_balance
))
7535 rq
->next_balance
= next_balance
;
7538 #ifdef CONFIG_NO_HZ_COMMON
7540 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7541 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7543 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
7545 int this_cpu
= this_rq
->cpu
;
7549 if (idle
!= CPU_IDLE
||
7550 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
7553 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
7554 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
7558 * If this cpu gets work to do, stop the load balancing
7559 * work being done for other cpus. Next load
7560 * balancing owner will pick it up.
7565 rq
= cpu_rq(balance_cpu
);
7568 * If time for next balance is due,
7571 if (time_after_eq(jiffies
, rq
->next_balance
)) {
7572 raw_spin_lock_irq(&rq
->lock
);
7573 update_rq_clock(rq
);
7574 update_idle_cpu_load(rq
);
7575 raw_spin_unlock_irq(&rq
->lock
);
7576 rebalance_domains(rq
, CPU_IDLE
);
7579 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
7580 this_rq
->next_balance
= rq
->next_balance
;
7582 nohz
.next_balance
= this_rq
->next_balance
;
7584 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
7588 * Current heuristic for kicking the idle load balancer in the presence
7589 * of an idle cpu is the system.
7590 * - This rq has more than one task.
7591 * - At any scheduler domain level, this cpu's scheduler group has multiple
7592 * busy cpu's exceeding the group's capacity.
7593 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7594 * domain span are idle.
7596 static inline int nohz_kick_needed(struct rq
*rq
)
7598 unsigned long now
= jiffies
;
7599 struct sched_domain
*sd
;
7600 struct sched_group_capacity
*sgc
;
7601 int nr_busy
, cpu
= rq
->cpu
;
7603 if (unlikely(rq
->idle_balance
))
7607 * We may be recently in ticked or tickless idle mode. At the first
7608 * busy tick after returning from idle, we will update the busy stats.
7610 set_cpu_sd_state_busy();
7611 nohz_balance_exit_idle(cpu
);
7614 * None are in tickless mode and hence no need for NOHZ idle load
7617 if (likely(!atomic_read(&nohz
.nr_cpus
)))
7620 if (time_before(now
, nohz
.next_balance
))
7623 if (rq
->nr_running
>= 2)
7627 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7630 sgc
= sd
->groups
->sgc
;
7631 nr_busy
= atomic_read(&sgc
->nr_busy_cpus
);
7634 goto need_kick_unlock
;
7637 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
7639 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
7640 sched_domain_span(sd
)) < cpu
))
7641 goto need_kick_unlock
;
7652 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
7656 * run_rebalance_domains is triggered when needed from the scheduler tick.
7657 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7659 static void run_rebalance_domains(struct softirq_action
*h
)
7661 struct rq
*this_rq
= this_rq();
7662 enum cpu_idle_type idle
= this_rq
->idle_balance
?
7663 CPU_IDLE
: CPU_NOT_IDLE
;
7665 rebalance_domains(this_rq
, idle
);
7668 * If this cpu has a pending nohz_balance_kick, then do the
7669 * balancing on behalf of the other idle cpus whose ticks are
7672 nohz_idle_balance(this_rq
, idle
);
7676 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7678 void trigger_load_balance(struct rq
*rq
)
7680 /* Don't need to rebalance while attached to NULL domain */
7681 if (unlikely(on_null_domain(rq
)))
7684 if (time_after_eq(jiffies
, rq
->next_balance
))
7685 raise_softirq(SCHED_SOFTIRQ
);
7686 #ifdef CONFIG_NO_HZ_COMMON
7687 if (nohz_kick_needed(rq
))
7688 nohz_balancer_kick();
7692 static void rq_online_fair(struct rq
*rq
)
7696 update_runtime_enabled(rq
);
7699 static void rq_offline_fair(struct rq
*rq
)
7703 /* Ensure any throttled groups are reachable by pick_next_task */
7704 unthrottle_offline_cfs_rqs(rq
);
7707 #endif /* CONFIG_SMP */
7710 * scheduler tick hitting a task of our scheduling class:
7712 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
7714 struct cfs_rq
*cfs_rq
;
7715 struct sched_entity
*se
= &curr
->se
;
7717 for_each_sched_entity(se
) {
7718 cfs_rq
= cfs_rq_of(se
);
7719 entity_tick(cfs_rq
, se
, queued
);
7722 if (numabalancing_enabled
)
7723 task_tick_numa(rq
, curr
);
7725 update_rq_runnable_avg(rq
, 1);
7729 * called on fork with the child task as argument from the parent's context
7730 * - child not yet on the tasklist
7731 * - preemption disabled
7733 static void task_fork_fair(struct task_struct
*p
)
7735 struct cfs_rq
*cfs_rq
;
7736 struct sched_entity
*se
= &p
->se
, *curr
;
7737 int this_cpu
= smp_processor_id();
7738 struct rq
*rq
= this_rq();
7739 unsigned long flags
;
7741 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7743 update_rq_clock(rq
);
7745 cfs_rq
= task_cfs_rq(current
);
7746 curr
= cfs_rq
->curr
;
7749 * Not only the cpu but also the task_group of the parent might have
7750 * been changed after parent->se.parent,cfs_rq were copied to
7751 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7752 * of child point to valid ones.
7755 __set_task_cpu(p
, this_cpu
);
7758 update_curr(cfs_rq
);
7761 se
->vruntime
= curr
->vruntime
;
7762 place_entity(cfs_rq
, se
, 1);
7764 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
7766 * Upon rescheduling, sched_class::put_prev_task() will place
7767 * 'current' within the tree based on its new key value.
7769 swap(curr
->vruntime
, se
->vruntime
);
7773 se
->vruntime
-= cfs_rq
->min_vruntime
;
7775 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7779 * Priority of the task has changed. Check to see if we preempt
7783 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
7785 if (!task_on_rq_queued(p
))
7789 * Reschedule if we are currently running on this runqueue and
7790 * our priority decreased, or if we are not currently running on
7791 * this runqueue and our priority is higher than the current's
7793 if (rq
->curr
== p
) {
7794 if (p
->prio
> oldprio
)
7797 check_preempt_curr(rq
, p
, 0);
7800 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
7802 struct sched_entity
*se
= &p
->se
;
7803 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7806 * Ensure the task's vruntime is normalized, so that when it's
7807 * switched back to the fair class the enqueue_entity(.flags=0) will
7808 * do the right thing.
7810 * If it's queued, then the dequeue_entity(.flags=0) will already
7811 * have normalized the vruntime, if it's !queued, then only when
7812 * the task is sleeping will it still have non-normalized vruntime.
7814 if (!task_on_rq_queued(p
) && p
->state
!= TASK_RUNNING
) {
7816 * Fix up our vruntime so that the current sleep doesn't
7817 * cause 'unlimited' sleep bonus.
7819 place_entity(cfs_rq
, se
, 0);
7820 se
->vruntime
-= cfs_rq
->min_vruntime
;
7825 * Remove our load from contribution when we leave sched_fair
7826 * and ensure we don't carry in an old decay_count if we
7829 if (se
->avg
.decay_count
) {
7830 __synchronize_entity_decay(se
);
7831 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7837 * We switched to the sched_fair class.
7839 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7841 #ifdef CONFIG_FAIR_GROUP_SCHED
7842 struct sched_entity
*se
= &p
->se
;
7844 * Since the real-depth could have been changed (only FAIR
7845 * class maintain depth value), reset depth properly.
7847 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7849 if (!task_on_rq_queued(p
))
7853 * We were most likely switched from sched_rt, so
7854 * kick off the schedule if running, otherwise just see
7855 * if we can still preempt the current task.
7860 check_preempt_curr(rq
, p
, 0);
7863 /* Account for a task changing its policy or group.
7865 * This routine is mostly called to set cfs_rq->curr field when a task
7866 * migrates between groups/classes.
7868 static void set_curr_task_fair(struct rq
*rq
)
7870 struct sched_entity
*se
= &rq
->curr
->se
;
7872 for_each_sched_entity(se
) {
7873 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7875 set_next_entity(cfs_rq
, se
);
7876 /* ensure bandwidth has been allocated on our new cfs_rq */
7877 account_cfs_rq_runtime(cfs_rq
, 0);
7881 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7883 cfs_rq
->tasks_timeline
= RB_ROOT
;
7884 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7885 #ifndef CONFIG_64BIT
7886 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7889 atomic64_set(&cfs_rq
->decay_counter
, 1);
7890 atomic_long_set(&cfs_rq
->removed_load
, 0);
7894 #ifdef CONFIG_FAIR_GROUP_SCHED
7895 static void task_move_group_fair(struct task_struct
*p
, int queued
)
7897 struct sched_entity
*se
= &p
->se
;
7898 struct cfs_rq
*cfs_rq
;
7901 * If the task was not on the rq at the time of this cgroup movement
7902 * it must have been asleep, sleeping tasks keep their ->vruntime
7903 * absolute on their old rq until wakeup (needed for the fair sleeper
7904 * bonus in place_entity()).
7906 * If it was on the rq, we've just 'preempted' it, which does convert
7907 * ->vruntime to a relative base.
7909 * Make sure both cases convert their relative position when migrating
7910 * to another cgroup's rq. This does somewhat interfere with the
7911 * fair sleeper stuff for the first placement, but who cares.
7914 * When !queued, vruntime of the task has usually NOT been normalized.
7915 * But there are some cases where it has already been normalized:
7917 * - Moving a forked child which is waiting for being woken up by
7918 * wake_up_new_task().
7919 * - Moving a task which has been woken up by try_to_wake_up() and
7920 * waiting for actually being woken up by sched_ttwu_pending().
7922 * To prevent boost or penalty in the new cfs_rq caused by delta
7923 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7925 if (!queued
&& (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7929 se
->vruntime
-= cfs_rq_of(se
)->min_vruntime
;
7930 set_task_rq(p
, task_cpu(p
));
7931 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7933 cfs_rq
= cfs_rq_of(se
);
7934 se
->vruntime
+= cfs_rq
->min_vruntime
;
7937 * migrate_task_rq_fair() will have removed our previous
7938 * contribution, but we must synchronize for ongoing future
7941 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7942 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
7947 void free_fair_sched_group(struct task_group
*tg
)
7951 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7953 for_each_possible_cpu(i
) {
7955 kfree(tg
->cfs_rq
[i
]);
7964 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7966 struct cfs_rq
*cfs_rq
;
7967 struct sched_entity
*se
;
7970 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7973 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7977 tg
->shares
= NICE_0_LOAD
;
7979 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7981 for_each_possible_cpu(i
) {
7982 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7983 GFP_KERNEL
, cpu_to_node(i
));
7987 se
= kzalloc_node(sizeof(struct sched_entity
),
7988 GFP_KERNEL
, cpu_to_node(i
));
7992 init_cfs_rq(cfs_rq
);
7993 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
8004 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8006 struct rq
*rq
= cpu_rq(cpu
);
8007 unsigned long flags
;
8010 * Only empty task groups can be destroyed; so we can speculatively
8011 * check on_list without danger of it being re-added.
8013 if (!tg
->cfs_rq
[cpu
]->on_list
)
8016 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8017 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
8018 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8021 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8022 struct sched_entity
*se
, int cpu
,
8023 struct sched_entity
*parent
)
8025 struct rq
*rq
= cpu_rq(cpu
);
8029 init_cfs_rq_runtime(cfs_rq
);
8031 tg
->cfs_rq
[cpu
] = cfs_rq
;
8034 /* se could be NULL for root_task_group */
8039 se
->cfs_rq
= &rq
->cfs
;
8042 se
->cfs_rq
= parent
->my_q
;
8043 se
->depth
= parent
->depth
+ 1;
8047 /* guarantee group entities always have weight */
8048 update_load_set(&se
->load
, NICE_0_LOAD
);
8049 se
->parent
= parent
;
8052 static DEFINE_MUTEX(shares_mutex
);
8054 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8057 unsigned long flags
;
8060 * We can't change the weight of the root cgroup.
8065 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
8067 mutex_lock(&shares_mutex
);
8068 if (tg
->shares
== shares
)
8071 tg
->shares
= shares
;
8072 for_each_possible_cpu(i
) {
8073 struct rq
*rq
= cpu_rq(i
);
8074 struct sched_entity
*se
;
8077 /* Propagate contribution to hierarchy */
8078 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8080 /* Possible calls to update_curr() need rq clock */
8081 update_rq_clock(rq
);
8082 for_each_sched_entity(se
)
8083 update_cfs_shares(group_cfs_rq(se
));
8084 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8088 mutex_unlock(&shares_mutex
);
8091 #else /* CONFIG_FAIR_GROUP_SCHED */
8093 void free_fair_sched_group(struct task_group
*tg
) { }
8095 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8100 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
8102 #endif /* CONFIG_FAIR_GROUP_SCHED */
8105 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
8107 struct sched_entity
*se
= &task
->se
;
8108 unsigned int rr_interval
= 0;
8111 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8114 if (rq
->cfs
.load
.weight
)
8115 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
8121 * All the scheduling class methods:
8123 const struct sched_class fair_sched_class
= {
8124 .next
= &idle_sched_class
,
8125 .enqueue_task
= enqueue_task_fair
,
8126 .dequeue_task
= dequeue_task_fair
,
8127 .yield_task
= yield_task_fair
,
8128 .yield_to_task
= yield_to_task_fair
,
8130 .check_preempt_curr
= check_preempt_wakeup
,
8132 .pick_next_task
= pick_next_task_fair
,
8133 .put_prev_task
= put_prev_task_fair
,
8136 .select_task_rq
= select_task_rq_fair
,
8137 .migrate_task_rq
= migrate_task_rq_fair
,
8139 .rq_online
= rq_online_fair
,
8140 .rq_offline
= rq_offline_fair
,
8142 .task_waking
= task_waking_fair
,
8145 .set_curr_task
= set_curr_task_fair
,
8146 .task_tick
= task_tick_fair
,
8147 .task_fork
= task_fork_fair
,
8149 .prio_changed
= prio_changed_fair
,
8150 .switched_from
= switched_from_fair
,
8151 .switched_to
= switched_to_fair
,
8153 .get_rr_interval
= get_rr_interval_fair
,
8155 .update_curr
= update_curr_fair
,
8157 #ifdef CONFIG_FAIR_GROUP_SCHED
8158 .task_move_group
= task_move_group_fair
,
8162 #ifdef CONFIG_SCHED_DEBUG
8163 void print_cfs_stats(struct seq_file
*m
, int cpu
)
8165 struct cfs_rq
*cfs_rq
;
8168 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
8169 print_cfs_rq(m
, cpu
, cfs_rq
);
8174 __init
void init_sched_fair_class(void)
8177 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8179 #ifdef CONFIG_NO_HZ_COMMON
8180 nohz
.next_balance
= jiffies
;
8181 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
8182 cpu_notifier(sched_ilb_notifier
, 0);