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
)
1199 long src_capacity
, dst_capacity
;
1201 long load_a
, load_b
;
1206 * The load is corrected for the CPU capacity available on each node.
1209 * ------------ vs ---------
1210 * src_capacity dst_capacity
1212 src_capacity
= env
->src_stats
.compute_capacity
;
1213 dst_capacity
= env
->dst_stats
.compute_capacity
;
1215 /* We care about the slope of the imbalance, not the direction. */
1218 if (load_a
< load_b
)
1219 swap(load_a
, load_b
);
1221 /* Is the difference below the threshold? */
1222 imb
= load_a
* src_capacity
* 100 -
1223 load_b
* dst_capacity
* env
->imbalance_pct
;
1228 * The imbalance is above the allowed threshold.
1229 * Allow a move that brings us closer to a balanced situation,
1230 * without moving things past the point of balance.
1232 orig_src_load
= env
->src_stats
.load
;
1235 * In a task swap, there will be one load moving from src to dst,
1236 * and another moving back. This is the net sum of both moves.
1237 * A simple task move will always have a positive value.
1238 * Allow the move if it brings the system closer to a balanced
1239 * situation, without crossing over the balance point.
1241 moved_load
= orig_src_load
- src_load
;
1244 /* Moving src -> dst. Did we overshoot balance? */
1245 return src_load
* dst_capacity
< dst_load
* src_capacity
;
1247 /* Moving dst -> src. Did we overshoot balance? */
1248 return dst_load
* src_capacity
< src_load
* dst_capacity
;
1252 * This checks if the overall compute and NUMA accesses of the system would
1253 * be improved if the source tasks was migrated to the target dst_cpu taking
1254 * into account that it might be best if task running on the dst_cpu should
1255 * be exchanged with the source task
1257 static void task_numa_compare(struct task_numa_env
*env
,
1258 long taskimp
, long groupimp
)
1260 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1261 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1262 struct task_struct
*cur
;
1263 long src_load
, dst_load
;
1265 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1267 int dist
= env
->dist
;
1271 raw_spin_lock_irq(&dst_rq
->lock
);
1274 * No need to move the exiting task, and this ensures that ->curr
1275 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1276 * is safe under RCU read lock.
1277 * Note that rcu_read_lock() itself can't protect from the final
1278 * put_task_struct() after the last schedule().
1280 if ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
))
1282 raw_spin_unlock_irq(&dst_rq
->lock
);
1285 * Because we have preemption enabled we can get migrated around and
1286 * end try selecting ourselves (current == env->p) as a swap candidate.
1292 * "imp" is the fault differential for the source task between the
1293 * source and destination node. Calculate the total differential for
1294 * the source task and potential destination task. The more negative
1295 * the value is, the more rmeote accesses that would be expected to
1296 * be incurred if the tasks were swapped.
1299 /* Skip this swap candidate if cannot move to the source cpu */
1300 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1304 * If dst and source tasks are in the same NUMA group, or not
1305 * in any group then look only at task weights.
1307 if (cur
->numa_group
== env
->p
->numa_group
) {
1308 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1309 task_weight(cur
, env
->dst_nid
, dist
);
1311 * Add some hysteresis to prevent swapping the
1312 * tasks within a group over tiny differences.
1314 if (cur
->numa_group
)
1318 * Compare the group weights. If a task is all by
1319 * itself (not part of a group), use the task weight
1322 if (cur
->numa_group
)
1323 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1324 group_weight(cur
, env
->dst_nid
, dist
);
1326 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1327 task_weight(cur
, env
->dst_nid
, dist
);
1331 if (imp
<= env
->best_imp
&& moveimp
<= env
->best_imp
)
1335 /* Is there capacity at our destination? */
1336 if (env
->src_stats
.nr_running
<= env
->src_stats
.task_capacity
&&
1337 !env
->dst_stats
.has_free_capacity
)
1343 /* Balance doesn't matter much if we're running a task per cpu */
1344 if (imp
> env
->best_imp
&& src_rq
->nr_running
== 1 &&
1345 dst_rq
->nr_running
== 1)
1349 * In the overloaded case, try and keep the load balanced.
1352 load
= task_h_load(env
->p
);
1353 dst_load
= env
->dst_stats
.load
+ load
;
1354 src_load
= env
->src_stats
.load
- load
;
1356 if (moveimp
> imp
&& moveimp
> env
->best_imp
) {
1358 * If the improvement from just moving env->p direction is
1359 * better than swapping tasks around, check if a move is
1360 * possible. Store a slightly smaller score than moveimp,
1361 * so an actually idle CPU will win.
1363 if (!load_too_imbalanced(src_load
, dst_load
, env
)) {
1370 if (imp
<= env
->best_imp
)
1374 load
= task_h_load(cur
);
1379 if (load_too_imbalanced(src_load
, dst_load
, env
))
1383 * One idle CPU per node is evaluated for a task numa move.
1384 * Call select_idle_sibling to maybe find a better one.
1387 env
->dst_cpu
= select_idle_sibling(env
->p
, env
->dst_cpu
);
1390 task_numa_assign(env
, cur
, imp
);
1395 static void task_numa_find_cpu(struct task_numa_env
*env
,
1396 long taskimp
, long groupimp
)
1400 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1401 /* Skip this CPU if the source task cannot migrate */
1402 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1406 task_numa_compare(env
, taskimp
, groupimp
);
1410 static int task_numa_migrate(struct task_struct
*p
)
1412 struct task_numa_env env
= {
1415 .src_cpu
= task_cpu(p
),
1416 .src_nid
= task_node(p
),
1418 .imbalance_pct
= 112,
1424 struct sched_domain
*sd
;
1425 unsigned long taskweight
, groupweight
;
1427 long taskimp
, groupimp
;
1430 * Pick the lowest SD_NUMA domain, as that would have the smallest
1431 * imbalance and would be the first to start moving tasks about.
1433 * And we want to avoid any moving of tasks about, as that would create
1434 * random movement of tasks -- counter the numa conditions we're trying
1438 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1440 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1444 * Cpusets can break the scheduler domain tree into smaller
1445 * balance domains, some of which do not cross NUMA boundaries.
1446 * Tasks that are "trapped" in such domains cannot be migrated
1447 * elsewhere, so there is no point in (re)trying.
1449 if (unlikely(!sd
)) {
1450 p
->numa_preferred_nid
= task_node(p
);
1454 env
.dst_nid
= p
->numa_preferred_nid
;
1455 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1456 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1457 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1458 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1459 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
1460 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
1461 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1463 /* Try to find a spot on the preferred nid. */
1464 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1467 * Look at other nodes in these cases:
1468 * - there is no space available on the preferred_nid
1469 * - the task is part of a numa_group that is interleaved across
1470 * multiple NUMA nodes; in order to better consolidate the group,
1471 * we need to check other locations.
1473 if (env
.best_cpu
== -1 || (p
->numa_group
&&
1474 nodes_weight(p
->numa_group
->active_nodes
) > 1)) {
1475 for_each_online_node(nid
) {
1476 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1479 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1480 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1482 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1483 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1486 /* Only consider nodes where both task and groups benefit */
1487 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
1488 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
1489 if (taskimp
< 0 && groupimp
< 0)
1494 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1495 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1500 * If the task is part of a workload that spans multiple NUMA nodes,
1501 * and is migrating into one of the workload's active nodes, remember
1502 * this node as the task's preferred numa node, so the workload can
1504 * A task that migrated to a second choice node will be better off
1505 * trying for a better one later. Do not set the preferred node here.
1507 if (p
->numa_group
) {
1508 if (env
.best_cpu
== -1)
1513 if (node_isset(nid
, p
->numa_group
->active_nodes
))
1514 sched_setnuma(p
, env
.dst_nid
);
1517 /* No better CPU than the current one was found. */
1518 if (env
.best_cpu
== -1)
1522 * Reset the scan period if the task is being rescheduled on an
1523 * alternative node to recheck if the tasks is now properly placed.
1525 p
->numa_scan_period
= task_scan_min(p
);
1527 if (env
.best_task
== NULL
) {
1528 ret
= migrate_task_to(p
, env
.best_cpu
);
1530 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1534 ret
= migrate_swap(p
, env
.best_task
);
1536 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1537 put_task_struct(env
.best_task
);
1541 /* Attempt to migrate a task to a CPU on the preferred node. */
1542 static void numa_migrate_preferred(struct task_struct
*p
)
1544 unsigned long interval
= HZ
;
1546 /* This task has no NUMA fault statistics yet */
1547 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults
))
1550 /* Periodically retry migrating the task to the preferred node */
1551 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1552 p
->numa_migrate_retry
= jiffies
+ interval
;
1554 /* Success if task is already running on preferred CPU */
1555 if (task_node(p
) == p
->numa_preferred_nid
)
1558 /* Otherwise, try migrate to a CPU on the preferred node */
1559 task_numa_migrate(p
);
1563 * Find the nodes on which the workload is actively running. We do this by
1564 * tracking the nodes from which NUMA hinting faults are triggered. This can
1565 * be different from the set of nodes where the workload's memory is currently
1568 * The bitmask is used to make smarter decisions on when to do NUMA page
1569 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1570 * are added when they cause over 6/16 of the maximum number of faults, but
1571 * only removed when they drop below 3/16.
1573 static void update_numa_active_node_mask(struct numa_group
*numa_group
)
1575 unsigned long faults
, max_faults
= 0;
1578 for_each_online_node(nid
) {
1579 faults
= group_faults_cpu(numa_group
, nid
);
1580 if (faults
> max_faults
)
1581 max_faults
= faults
;
1584 for_each_online_node(nid
) {
1585 faults
= group_faults_cpu(numa_group
, nid
);
1586 if (!node_isset(nid
, numa_group
->active_nodes
)) {
1587 if (faults
> max_faults
* 6 / 16)
1588 node_set(nid
, numa_group
->active_nodes
);
1589 } else if (faults
< max_faults
* 3 / 16)
1590 node_clear(nid
, numa_group
->active_nodes
);
1595 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1596 * increments. The more local the fault statistics are, the higher the scan
1597 * period will be for the next scan window. If local/(local+remote) ratio is
1598 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1599 * the scan period will decrease. Aim for 70% local accesses.
1601 #define NUMA_PERIOD_SLOTS 10
1602 #define NUMA_PERIOD_THRESHOLD 7
1605 * Increase the scan period (slow down scanning) if the majority of
1606 * our memory is already on our local node, or if the majority of
1607 * the page accesses are shared with other processes.
1608 * Otherwise, decrease the scan period.
1610 static void update_task_scan_period(struct task_struct
*p
,
1611 unsigned long shared
, unsigned long private)
1613 unsigned int period_slot
;
1617 unsigned long remote
= p
->numa_faults_locality
[0];
1618 unsigned long local
= p
->numa_faults_locality
[1];
1621 * If there were no record hinting faults then either the task is
1622 * completely idle or all activity is areas that are not of interest
1623 * to automatic numa balancing. Scan slower
1625 if (local
+ shared
== 0) {
1626 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1627 p
->numa_scan_period
<< 1);
1629 p
->mm
->numa_next_scan
= jiffies
+
1630 msecs_to_jiffies(p
->numa_scan_period
);
1636 * Prepare to scale scan period relative to the current period.
1637 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1638 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1639 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1641 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1642 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1643 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1644 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1647 diff
= slot
* period_slot
;
1649 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1652 * Scale scan rate increases based on sharing. There is an
1653 * inverse relationship between the degree of sharing and
1654 * the adjustment made to the scanning period. Broadly
1655 * speaking the intent is that there is little point
1656 * scanning faster if shared accesses dominate as it may
1657 * simply bounce migrations uselessly
1659 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
+ 1));
1660 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1663 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1664 task_scan_min(p
), task_scan_max(p
));
1665 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1669 * Get the fraction of time the task has been running since the last
1670 * NUMA placement cycle. The scheduler keeps similar statistics, but
1671 * decays those on a 32ms period, which is orders of magnitude off
1672 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1673 * stats only if the task is so new there are no NUMA statistics yet.
1675 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1677 u64 runtime
, delta
, now
;
1678 /* Use the start of this time slice to avoid calculations. */
1679 now
= p
->se
.exec_start
;
1680 runtime
= p
->se
.sum_exec_runtime
;
1682 if (p
->last_task_numa_placement
) {
1683 delta
= runtime
- p
->last_sum_exec_runtime
;
1684 *period
= now
- p
->last_task_numa_placement
;
1686 delta
= p
->se
.avg
.runnable_avg_sum
;
1687 *period
= p
->se
.avg
.runnable_avg_period
;
1690 p
->last_sum_exec_runtime
= runtime
;
1691 p
->last_task_numa_placement
= now
;
1697 * Determine the preferred nid for a task in a numa_group. This needs to
1698 * be done in a way that produces consistent results with group_weight,
1699 * otherwise workloads might not converge.
1701 static int preferred_group_nid(struct task_struct
*p
, int nid
)
1706 /* Direct connections between all NUMA nodes. */
1707 if (sched_numa_topology_type
== NUMA_DIRECT
)
1711 * On a system with glueless mesh NUMA topology, group_weight
1712 * scores nodes according to the number of NUMA hinting faults on
1713 * both the node itself, and on nearby nodes.
1715 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1716 unsigned long score
, max_score
= 0;
1717 int node
, max_node
= nid
;
1719 dist
= sched_max_numa_distance
;
1721 for_each_online_node(node
) {
1722 score
= group_weight(p
, node
, dist
);
1723 if (score
> max_score
) {
1732 * Finding the preferred nid in a system with NUMA backplane
1733 * interconnect topology is more involved. The goal is to locate
1734 * tasks from numa_groups near each other in the system, and
1735 * untangle workloads from different sides of the system. This requires
1736 * searching down the hierarchy of node groups, recursively searching
1737 * inside the highest scoring group of nodes. The nodemask tricks
1738 * keep the complexity of the search down.
1740 nodes
= node_online_map
;
1741 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
1742 unsigned long max_faults
= 0;
1743 nodemask_t max_group
= NODE_MASK_NONE
;
1746 /* Are there nodes at this distance from each other? */
1747 if (!find_numa_distance(dist
))
1750 for_each_node_mask(a
, nodes
) {
1751 unsigned long faults
= 0;
1752 nodemask_t this_group
;
1753 nodes_clear(this_group
);
1755 /* Sum group's NUMA faults; includes a==b case. */
1756 for_each_node_mask(b
, nodes
) {
1757 if (node_distance(a
, b
) < dist
) {
1758 faults
+= group_faults(p
, b
);
1759 node_set(b
, this_group
);
1760 node_clear(b
, nodes
);
1764 /* Remember the top group. */
1765 if (faults
> max_faults
) {
1766 max_faults
= faults
;
1767 max_group
= this_group
;
1769 * subtle: at the smallest distance there is
1770 * just one node left in each "group", the
1771 * winner is the preferred nid.
1776 /* Next round, evaluate the nodes within max_group. */
1782 static void task_numa_placement(struct task_struct
*p
)
1784 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1785 unsigned long max_faults
= 0, max_group_faults
= 0;
1786 unsigned long fault_types
[2] = { 0, 0 };
1787 unsigned long total_faults
;
1788 u64 runtime
, period
;
1789 spinlock_t
*group_lock
= NULL
;
1791 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1792 if (p
->numa_scan_seq
== seq
)
1794 p
->numa_scan_seq
= seq
;
1795 p
->numa_scan_period_max
= task_scan_max(p
);
1797 total_faults
= p
->numa_faults_locality
[0] +
1798 p
->numa_faults_locality
[1];
1799 runtime
= numa_get_avg_runtime(p
, &period
);
1801 /* If the task is part of a group prevent parallel updates to group stats */
1802 if (p
->numa_group
) {
1803 group_lock
= &p
->numa_group
->lock
;
1804 spin_lock_irq(group_lock
);
1807 /* Find the node with the highest number of faults */
1808 for_each_online_node(nid
) {
1809 /* Keep track of the offsets in numa_faults array */
1810 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
1811 unsigned long faults
= 0, group_faults
= 0;
1814 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
1815 long diff
, f_diff
, f_weight
;
1817 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
1818 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
1819 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
1820 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
1822 /* Decay existing window, copy faults since last scan */
1823 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
1824 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
1825 p
->numa_faults
[membuf_idx
] = 0;
1828 * Normalize the faults_from, so all tasks in a group
1829 * count according to CPU use, instead of by the raw
1830 * number of faults. Tasks with little runtime have
1831 * little over-all impact on throughput, and thus their
1832 * faults are less important.
1834 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
1835 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
1837 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
1838 p
->numa_faults
[cpubuf_idx
] = 0;
1840 p
->numa_faults
[mem_idx
] += diff
;
1841 p
->numa_faults
[cpu_idx
] += f_diff
;
1842 faults
+= p
->numa_faults
[mem_idx
];
1843 p
->total_numa_faults
+= diff
;
1844 if (p
->numa_group
) {
1846 * safe because we can only change our own group
1848 * mem_idx represents the offset for a given
1849 * nid and priv in a specific region because it
1850 * is at the beginning of the numa_faults array.
1852 p
->numa_group
->faults
[mem_idx
] += diff
;
1853 p
->numa_group
->faults_cpu
[mem_idx
] += f_diff
;
1854 p
->numa_group
->total_faults
+= diff
;
1855 group_faults
+= p
->numa_group
->faults
[mem_idx
];
1859 if (faults
> max_faults
) {
1860 max_faults
= faults
;
1864 if (group_faults
> max_group_faults
) {
1865 max_group_faults
= group_faults
;
1866 max_group_nid
= nid
;
1870 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1872 if (p
->numa_group
) {
1873 update_numa_active_node_mask(p
->numa_group
);
1874 spin_unlock_irq(group_lock
);
1875 max_nid
= preferred_group_nid(p
, max_group_nid
);
1879 /* Set the new preferred node */
1880 if (max_nid
!= p
->numa_preferred_nid
)
1881 sched_setnuma(p
, max_nid
);
1883 if (task_node(p
) != p
->numa_preferred_nid
)
1884 numa_migrate_preferred(p
);
1888 static inline int get_numa_group(struct numa_group
*grp
)
1890 return atomic_inc_not_zero(&grp
->refcount
);
1893 static inline void put_numa_group(struct numa_group
*grp
)
1895 if (atomic_dec_and_test(&grp
->refcount
))
1896 kfree_rcu(grp
, rcu
);
1899 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1902 struct numa_group
*grp
, *my_grp
;
1903 struct task_struct
*tsk
;
1905 int cpu
= cpupid_to_cpu(cpupid
);
1908 if (unlikely(!p
->numa_group
)) {
1909 unsigned int size
= sizeof(struct numa_group
) +
1910 4*nr_node_ids
*sizeof(unsigned long);
1912 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1916 atomic_set(&grp
->refcount
, 1);
1917 spin_lock_init(&grp
->lock
);
1919 /* Second half of the array tracks nids where faults happen */
1920 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
1923 node_set(task_node(current
), grp
->active_nodes
);
1925 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1926 grp
->faults
[i
] = p
->numa_faults
[i
];
1928 grp
->total_faults
= p
->total_numa_faults
;
1931 rcu_assign_pointer(p
->numa_group
, grp
);
1935 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1937 if (!cpupid_match_pid(tsk
, cpupid
))
1940 grp
= rcu_dereference(tsk
->numa_group
);
1944 my_grp
= p
->numa_group
;
1949 * Only join the other group if its bigger; if we're the bigger group,
1950 * the other task will join us.
1952 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1956 * Tie-break on the grp address.
1958 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1961 /* Always join threads in the same process. */
1962 if (tsk
->mm
== current
->mm
)
1965 /* Simple filter to avoid false positives due to PID collisions */
1966 if (flags
& TNF_SHARED
)
1969 /* Update priv based on whether false sharing was detected */
1972 if (join
&& !get_numa_group(grp
))
1980 BUG_ON(irqs_disabled());
1981 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
1983 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
1984 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
1985 grp
->faults
[i
] += p
->numa_faults
[i
];
1987 my_grp
->total_faults
-= p
->total_numa_faults
;
1988 grp
->total_faults
+= p
->total_numa_faults
;
1993 spin_unlock(&my_grp
->lock
);
1994 spin_unlock_irq(&grp
->lock
);
1996 rcu_assign_pointer(p
->numa_group
, grp
);
1998 put_numa_group(my_grp
);
2006 void task_numa_free(struct task_struct
*p
)
2008 struct numa_group
*grp
= p
->numa_group
;
2009 void *numa_faults
= p
->numa_faults
;
2010 unsigned long flags
;
2014 spin_lock_irqsave(&grp
->lock
, flags
);
2015 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2016 grp
->faults
[i
] -= p
->numa_faults
[i
];
2017 grp
->total_faults
-= p
->total_numa_faults
;
2020 spin_unlock_irqrestore(&grp
->lock
, flags
);
2021 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2022 put_numa_group(grp
);
2025 p
->numa_faults
= NULL
;
2030 * Got a PROT_NONE fault for a page on @node.
2032 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2034 struct task_struct
*p
= current
;
2035 bool migrated
= flags
& TNF_MIGRATED
;
2036 int cpu_node
= task_node(current
);
2037 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2040 if (!numabalancing_enabled
)
2043 /* for example, ksmd faulting in a user's mm */
2047 /* Allocate buffer to track faults on a per-node basis */
2048 if (unlikely(!p
->numa_faults
)) {
2049 int size
= sizeof(*p
->numa_faults
) *
2050 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2052 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2053 if (!p
->numa_faults
)
2056 p
->total_numa_faults
= 0;
2057 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2061 * First accesses are treated as private, otherwise consider accesses
2062 * to be private if the accessing pid has not changed
2064 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2067 priv
= cpupid_match_pid(p
, last_cpupid
);
2068 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2069 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2073 * If a workload spans multiple NUMA nodes, a shared fault that
2074 * occurs wholly within the set of nodes that the workload is
2075 * actively using should be counted as local. This allows the
2076 * scan rate to slow down when a workload has settled down.
2078 if (!priv
&& !local
&& p
->numa_group
&&
2079 node_isset(cpu_node
, p
->numa_group
->active_nodes
) &&
2080 node_isset(mem_node
, p
->numa_group
->active_nodes
))
2083 task_numa_placement(p
);
2086 * Retry task to preferred node migration periodically, in case it
2087 * case it previously failed, or the scheduler moved us.
2089 if (time_after(jiffies
, p
->numa_migrate_retry
))
2090 numa_migrate_preferred(p
);
2093 p
->numa_pages_migrated
+= pages
;
2095 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2096 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2097 p
->numa_faults_locality
[local
] += pages
;
2100 static void reset_ptenuma_scan(struct task_struct
*p
)
2102 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
2103 p
->mm
->numa_scan_offset
= 0;
2107 * The expensive part of numa migration is done from task_work context.
2108 * Triggered from task_tick_numa().
2110 void task_numa_work(struct callback_head
*work
)
2112 unsigned long migrate
, next_scan
, now
= jiffies
;
2113 struct task_struct
*p
= current
;
2114 struct mm_struct
*mm
= p
->mm
;
2115 struct vm_area_struct
*vma
;
2116 unsigned long start
, end
;
2117 unsigned long nr_pte_updates
= 0;
2120 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
2122 work
->next
= work
; /* protect against double add */
2124 * Who cares about NUMA placement when they're dying.
2126 * NOTE: make sure not to dereference p->mm before this check,
2127 * exit_task_work() happens _after_ exit_mm() so we could be called
2128 * without p->mm even though we still had it when we enqueued this
2131 if (p
->flags
& PF_EXITING
)
2134 if (!mm
->numa_next_scan
) {
2135 mm
->numa_next_scan
= now
+
2136 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2140 * Enforce maximal scan/migration frequency..
2142 migrate
= mm
->numa_next_scan
;
2143 if (time_before(now
, migrate
))
2146 if (p
->numa_scan_period
== 0) {
2147 p
->numa_scan_period_max
= task_scan_max(p
);
2148 p
->numa_scan_period
= task_scan_min(p
);
2151 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2152 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2156 * Delay this task enough that another task of this mm will likely win
2157 * the next time around.
2159 p
->node_stamp
+= 2 * TICK_NSEC
;
2161 start
= mm
->numa_scan_offset
;
2162 pages
= sysctl_numa_balancing_scan_size
;
2163 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2167 down_read(&mm
->mmap_sem
);
2168 vma
= find_vma(mm
, start
);
2170 reset_ptenuma_scan(p
);
2174 for (; vma
; vma
= vma
->vm_next
) {
2175 if (!vma_migratable(vma
) || !vma_policy_mof(vma
))
2179 * Shared library pages mapped by multiple processes are not
2180 * migrated as it is expected they are cache replicated. Avoid
2181 * hinting faults in read-only file-backed mappings or the vdso
2182 * as migrating the pages will be of marginal benefit.
2185 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2189 * Skip inaccessible VMAs to avoid any confusion between
2190 * PROT_NONE and NUMA hinting ptes
2192 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
2196 start
= max(start
, vma
->vm_start
);
2197 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2198 end
= min(end
, vma
->vm_end
);
2199 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
2202 * Scan sysctl_numa_balancing_scan_size but ensure that
2203 * at least one PTE is updated so that unused virtual
2204 * address space is quickly skipped.
2207 pages
-= (end
- start
) >> PAGE_SHIFT
;
2214 } while (end
!= vma
->vm_end
);
2219 * It is possible to reach the end of the VMA list but the last few
2220 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2221 * would find the !migratable VMA on the next scan but not reset the
2222 * scanner to the start so check it now.
2225 mm
->numa_scan_offset
= start
;
2227 reset_ptenuma_scan(p
);
2228 up_read(&mm
->mmap_sem
);
2232 * Drive the periodic memory faults..
2234 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2236 struct callback_head
*work
= &curr
->numa_work
;
2240 * We don't care about NUMA placement if we don't have memory.
2242 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2246 * Using runtime rather than walltime has the dual advantage that
2247 * we (mostly) drive the selection from busy threads and that the
2248 * task needs to have done some actual work before we bother with
2251 now
= curr
->se
.sum_exec_runtime
;
2252 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2254 if (now
- curr
->node_stamp
> period
) {
2255 if (!curr
->node_stamp
)
2256 curr
->numa_scan_period
= task_scan_min(curr
);
2257 curr
->node_stamp
+= period
;
2259 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2260 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2261 task_work_add(curr
, work
, true);
2266 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2270 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2274 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2277 #endif /* CONFIG_NUMA_BALANCING */
2280 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2282 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2283 if (!parent_entity(se
))
2284 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2286 if (entity_is_task(se
)) {
2287 struct rq
*rq
= rq_of(cfs_rq
);
2289 account_numa_enqueue(rq
, task_of(se
));
2290 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2293 cfs_rq
->nr_running
++;
2297 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2299 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2300 if (!parent_entity(se
))
2301 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2302 if (entity_is_task(se
)) {
2303 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2304 list_del_init(&se
->group_node
);
2306 cfs_rq
->nr_running
--;
2309 #ifdef CONFIG_FAIR_GROUP_SCHED
2311 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
2316 * Use this CPU's actual weight instead of the last load_contribution
2317 * to gain a more accurate current total weight. See
2318 * update_cfs_rq_load_contribution().
2320 tg_weight
= atomic_long_read(&tg
->load_avg
);
2321 tg_weight
-= cfs_rq
->tg_load_contrib
;
2322 tg_weight
+= cfs_rq
->load
.weight
;
2327 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2329 long tg_weight
, load
, shares
;
2331 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
2332 load
= cfs_rq
->load
.weight
;
2334 shares
= (tg
->shares
* load
);
2336 shares
/= tg_weight
;
2338 if (shares
< MIN_SHARES
)
2339 shares
= MIN_SHARES
;
2340 if (shares
> tg
->shares
)
2341 shares
= tg
->shares
;
2345 # else /* CONFIG_SMP */
2346 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2350 # endif /* CONFIG_SMP */
2351 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2352 unsigned long weight
)
2355 /* commit outstanding execution time */
2356 if (cfs_rq
->curr
== se
)
2357 update_curr(cfs_rq
);
2358 account_entity_dequeue(cfs_rq
, se
);
2361 update_load_set(&se
->load
, weight
);
2364 account_entity_enqueue(cfs_rq
, se
);
2367 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2369 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2371 struct task_group
*tg
;
2372 struct sched_entity
*se
;
2376 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2377 if (!se
|| throttled_hierarchy(cfs_rq
))
2380 if (likely(se
->load
.weight
== tg
->shares
))
2383 shares
= calc_cfs_shares(cfs_rq
, tg
);
2385 reweight_entity(cfs_rq_of(se
), se
, shares
);
2387 #else /* CONFIG_FAIR_GROUP_SCHED */
2388 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2391 #endif /* CONFIG_FAIR_GROUP_SCHED */
2395 * We choose a half-life close to 1 scheduling period.
2396 * Note: The tables below are dependent on this value.
2398 #define LOAD_AVG_PERIOD 32
2399 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2400 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2402 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2403 static const u32 runnable_avg_yN_inv
[] = {
2404 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2405 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2406 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2407 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2408 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2409 0x85aac367, 0x82cd8698,
2413 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2414 * over-estimates when re-combining.
2416 static const u32 runnable_avg_yN_sum
[] = {
2417 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2418 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2419 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2424 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2426 static __always_inline u64
decay_load(u64 val
, u64 n
)
2428 unsigned int local_n
;
2432 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2435 /* after bounds checking we can collapse to 32-bit */
2439 * As y^PERIOD = 1/2, we can combine
2440 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2441 * With a look-up table which covers y^n (n<PERIOD)
2443 * To achieve constant time decay_load.
2445 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2446 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2447 local_n
%= LOAD_AVG_PERIOD
;
2450 val
*= runnable_avg_yN_inv
[local_n
];
2451 /* We don't use SRR here since we always want to round down. */
2456 * For updates fully spanning n periods, the contribution to runnable
2457 * average will be: \Sum 1024*y^n
2459 * We can compute this reasonably efficiently by combining:
2460 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2462 static u32
__compute_runnable_contrib(u64 n
)
2466 if (likely(n
<= LOAD_AVG_PERIOD
))
2467 return runnable_avg_yN_sum
[n
];
2468 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2469 return LOAD_AVG_MAX
;
2471 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2473 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2474 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2476 n
-= LOAD_AVG_PERIOD
;
2477 } while (n
> LOAD_AVG_PERIOD
);
2479 contrib
= decay_load(contrib
, n
);
2480 return contrib
+ runnable_avg_yN_sum
[n
];
2484 * We can represent the historical contribution to runnable average as the
2485 * coefficients of a geometric series. To do this we sub-divide our runnable
2486 * history into segments of approximately 1ms (1024us); label the segment that
2487 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2489 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2491 * (now) (~1ms ago) (~2ms ago)
2493 * Let u_i denote the fraction of p_i that the entity was runnable.
2495 * We then designate the fractions u_i as our co-efficients, yielding the
2496 * following representation of historical load:
2497 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2499 * We choose y based on the with of a reasonably scheduling period, fixing:
2502 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2503 * approximately half as much as the contribution to load within the last ms
2506 * When a period "rolls over" and we have new u_0`, multiplying the previous
2507 * sum again by y is sufficient to update:
2508 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2509 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2511 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2512 struct sched_avg
*sa
,
2516 u32 runnable_contrib
;
2517 int delta_w
, decayed
= 0;
2519 delta
= now
- sa
->last_runnable_update
;
2521 * This should only happen when time goes backwards, which it
2522 * unfortunately does during sched clock init when we swap over to TSC.
2524 if ((s64
)delta
< 0) {
2525 sa
->last_runnable_update
= now
;
2530 * Use 1024ns as the unit of measurement since it's a reasonable
2531 * approximation of 1us and fast to compute.
2536 sa
->last_runnable_update
= now
;
2538 /* delta_w is the amount already accumulated against our next period */
2539 delta_w
= sa
->runnable_avg_period
% 1024;
2540 if (delta
+ delta_w
>= 1024) {
2541 /* period roll-over */
2545 * Now that we know we're crossing a period boundary, figure
2546 * out how much from delta we need to complete the current
2547 * period and accrue it.
2549 delta_w
= 1024 - delta_w
;
2551 sa
->runnable_avg_sum
+= delta_w
;
2552 sa
->runnable_avg_period
+= delta_w
;
2556 /* Figure out how many additional periods this update spans */
2557 periods
= delta
/ 1024;
2560 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2562 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2565 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2566 runnable_contrib
= __compute_runnable_contrib(periods
);
2568 sa
->runnable_avg_sum
+= runnable_contrib
;
2569 sa
->runnable_avg_period
+= runnable_contrib
;
2572 /* Remainder of delta accrued against u_0` */
2574 sa
->runnable_avg_sum
+= delta
;
2575 sa
->runnable_avg_period
+= delta
;
2580 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2581 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2583 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2584 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2586 decays
-= se
->avg
.decay_count
;
2587 se
->avg
.decay_count
= 0;
2591 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2596 #ifdef CONFIG_FAIR_GROUP_SCHED
2597 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2600 struct task_group
*tg
= cfs_rq
->tg
;
2603 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2604 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2609 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2610 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2611 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2616 * Aggregate cfs_rq runnable averages into an equivalent task_group
2617 * representation for computing load contributions.
2619 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2620 struct cfs_rq
*cfs_rq
)
2622 struct task_group
*tg
= cfs_rq
->tg
;
2625 /* The fraction of a cpu used by this cfs_rq */
2626 contrib
= div_u64((u64
)sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2627 sa
->runnable_avg_period
+ 1);
2628 contrib
-= cfs_rq
->tg_runnable_contrib
;
2630 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2631 atomic_add(contrib
, &tg
->runnable_avg
);
2632 cfs_rq
->tg_runnable_contrib
+= contrib
;
2636 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2638 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2639 struct task_group
*tg
= cfs_rq
->tg
;
2644 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2645 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2646 atomic_long_read(&tg
->load_avg
) + 1);
2649 * For group entities we need to compute a correction term in the case
2650 * that they are consuming <1 cpu so that we would contribute the same
2651 * load as a task of equal weight.
2653 * Explicitly co-ordinating this measurement would be expensive, but
2654 * fortunately the sum of each cpus contribution forms a usable
2655 * lower-bound on the true value.
2657 * Consider the aggregate of 2 contributions. Either they are disjoint
2658 * (and the sum represents true value) or they are disjoint and we are
2659 * understating by the aggregate of their overlap.
2661 * Extending this to N cpus, for a given overlap, the maximum amount we
2662 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2663 * cpus that overlap for this interval and w_i is the interval width.
2665 * On a small machine; the first term is well-bounded which bounds the
2666 * total error since w_i is a subset of the period. Whereas on a
2667 * larger machine, while this first term can be larger, if w_i is the
2668 * of consequential size guaranteed to see n_i*w_i quickly converge to
2669 * our upper bound of 1-cpu.
2671 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2672 if (runnable_avg
< NICE_0_LOAD
) {
2673 se
->avg
.load_avg_contrib
*= runnable_avg
;
2674 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2678 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2680 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2681 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2683 #else /* CONFIG_FAIR_GROUP_SCHED */
2684 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2685 int force_update
) {}
2686 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2687 struct cfs_rq
*cfs_rq
) {}
2688 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2689 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2690 #endif /* CONFIG_FAIR_GROUP_SCHED */
2692 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2696 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2697 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2698 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2699 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2702 /* Compute the current contribution to load_avg by se, return any delta */
2703 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2705 long old_contrib
= se
->avg
.load_avg_contrib
;
2707 if (entity_is_task(se
)) {
2708 __update_task_entity_contrib(se
);
2710 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2711 __update_group_entity_contrib(se
);
2714 return se
->avg
.load_avg_contrib
- old_contrib
;
2717 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2720 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2721 cfs_rq
->blocked_load_avg
-= load_contrib
;
2723 cfs_rq
->blocked_load_avg
= 0;
2726 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2728 /* Update a sched_entity's runnable average */
2729 static inline void update_entity_load_avg(struct sched_entity
*se
,
2732 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2737 * For a group entity we need to use their owned cfs_rq_clock_task() in
2738 * case they are the parent of a throttled hierarchy.
2740 if (entity_is_task(se
))
2741 now
= cfs_rq_clock_task(cfs_rq
);
2743 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2745 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2748 contrib_delta
= __update_entity_load_avg_contrib(se
);
2754 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2756 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2760 * Decay the load contributed by all blocked children and account this so that
2761 * their contribution may appropriately discounted when they wake up.
2763 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2765 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2768 decays
= now
- cfs_rq
->last_decay
;
2769 if (!decays
&& !force_update
)
2772 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2773 unsigned long removed_load
;
2774 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2775 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2779 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2781 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2782 cfs_rq
->last_decay
= now
;
2785 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2788 /* Add the load generated by se into cfs_rq's child load-average */
2789 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2790 struct sched_entity
*se
,
2794 * We track migrations using entity decay_count <= 0, on a wake-up
2795 * migration we use a negative decay count to track the remote decays
2796 * accumulated while sleeping.
2798 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2799 * are seen by enqueue_entity_load_avg() as a migration with an already
2800 * constructed load_avg_contrib.
2802 if (unlikely(se
->avg
.decay_count
<= 0)) {
2803 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2804 if (se
->avg
.decay_count
) {
2806 * In a wake-up migration we have to approximate the
2807 * time sleeping. This is because we can't synchronize
2808 * clock_task between the two cpus, and it is not
2809 * guaranteed to be read-safe. Instead, we can
2810 * approximate this using our carried decays, which are
2811 * explicitly atomically readable.
2813 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2815 update_entity_load_avg(se
, 0);
2816 /* Indicate that we're now synchronized and on-rq */
2817 se
->avg
.decay_count
= 0;
2821 __synchronize_entity_decay(se
);
2824 /* migrated tasks did not contribute to our blocked load */
2826 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2827 update_entity_load_avg(se
, 0);
2830 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2831 /* we force update consideration on load-balancer moves */
2832 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2836 * Remove se's load from this cfs_rq child load-average, if the entity is
2837 * transitioning to a blocked state we track its projected decay using
2840 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2841 struct sched_entity
*se
,
2844 update_entity_load_avg(se
, 1);
2845 /* we force update consideration on load-balancer moves */
2846 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2848 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2850 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2851 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2852 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2856 * Update the rq's load with the elapsed running time before entering
2857 * idle. if the last scheduled task is not a CFS task, idle_enter will
2858 * be the only way to update the runnable statistic.
2860 void idle_enter_fair(struct rq
*this_rq
)
2862 update_rq_runnable_avg(this_rq
, 1);
2866 * Update the rq's load with the elapsed idle time before a task is
2867 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2868 * be the only way to update the runnable statistic.
2870 void idle_exit_fair(struct rq
*this_rq
)
2872 update_rq_runnable_avg(this_rq
, 0);
2875 static int idle_balance(struct rq
*this_rq
);
2877 #else /* CONFIG_SMP */
2879 static inline void update_entity_load_avg(struct sched_entity
*se
,
2880 int update_cfs_rq
) {}
2881 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2882 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2883 struct sched_entity
*se
,
2885 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2886 struct sched_entity
*se
,
2888 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2889 int force_update
) {}
2891 static inline int idle_balance(struct rq
*rq
)
2896 #endif /* CONFIG_SMP */
2898 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2900 #ifdef CONFIG_SCHEDSTATS
2901 struct task_struct
*tsk
= NULL
;
2903 if (entity_is_task(se
))
2906 if (se
->statistics
.sleep_start
) {
2907 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2912 if (unlikely(delta
> se
->statistics
.sleep_max
))
2913 se
->statistics
.sleep_max
= delta
;
2915 se
->statistics
.sleep_start
= 0;
2916 se
->statistics
.sum_sleep_runtime
+= delta
;
2919 account_scheduler_latency(tsk
, delta
>> 10, 1);
2920 trace_sched_stat_sleep(tsk
, delta
);
2923 if (se
->statistics
.block_start
) {
2924 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2929 if (unlikely(delta
> se
->statistics
.block_max
))
2930 se
->statistics
.block_max
= delta
;
2932 se
->statistics
.block_start
= 0;
2933 se
->statistics
.sum_sleep_runtime
+= delta
;
2936 if (tsk
->in_iowait
) {
2937 se
->statistics
.iowait_sum
+= delta
;
2938 se
->statistics
.iowait_count
++;
2939 trace_sched_stat_iowait(tsk
, delta
);
2942 trace_sched_stat_blocked(tsk
, delta
);
2945 * Blocking time is in units of nanosecs, so shift by
2946 * 20 to get a milliseconds-range estimation of the
2947 * amount of time that the task spent sleeping:
2949 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2950 profile_hits(SLEEP_PROFILING
,
2951 (void *)get_wchan(tsk
),
2954 account_scheduler_latency(tsk
, delta
>> 10, 0);
2960 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2962 #ifdef CONFIG_SCHED_DEBUG
2963 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2968 if (d
> 3*sysctl_sched_latency
)
2969 schedstat_inc(cfs_rq
, nr_spread_over
);
2974 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2976 u64 vruntime
= cfs_rq
->min_vruntime
;
2979 * The 'current' period is already promised to the current tasks,
2980 * however the extra weight of the new task will slow them down a
2981 * little, place the new task so that it fits in the slot that
2982 * stays open at the end.
2984 if (initial
&& sched_feat(START_DEBIT
))
2985 vruntime
+= sched_vslice(cfs_rq
, se
);
2987 /* sleeps up to a single latency don't count. */
2989 unsigned long thresh
= sysctl_sched_latency
;
2992 * Halve their sleep time's effect, to allow
2993 * for a gentler effect of sleepers:
2995 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
3001 /* ensure we never gain time by being placed backwards. */
3002 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
3005 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
3008 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3011 * Update the normalized vruntime before updating min_vruntime
3012 * through calling update_curr().
3014 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
3015 se
->vruntime
+= cfs_rq
->min_vruntime
;
3018 * Update run-time statistics of the 'current'.
3020 update_curr(cfs_rq
);
3021 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
3022 account_entity_enqueue(cfs_rq
, se
);
3023 update_cfs_shares(cfs_rq
);
3025 if (flags
& ENQUEUE_WAKEUP
) {
3026 place_entity(cfs_rq
, se
, 0);
3027 enqueue_sleeper(cfs_rq
, se
);
3030 update_stats_enqueue(cfs_rq
, se
);
3031 check_spread(cfs_rq
, se
);
3032 if (se
!= cfs_rq
->curr
)
3033 __enqueue_entity(cfs_rq
, se
);
3036 if (cfs_rq
->nr_running
== 1) {
3037 list_add_leaf_cfs_rq(cfs_rq
);
3038 check_enqueue_throttle(cfs_rq
);
3042 static void __clear_buddies_last(struct sched_entity
*se
)
3044 for_each_sched_entity(se
) {
3045 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3046 if (cfs_rq
->last
!= se
)
3049 cfs_rq
->last
= NULL
;
3053 static void __clear_buddies_next(struct sched_entity
*se
)
3055 for_each_sched_entity(se
) {
3056 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3057 if (cfs_rq
->next
!= se
)
3060 cfs_rq
->next
= NULL
;
3064 static void __clear_buddies_skip(struct sched_entity
*se
)
3066 for_each_sched_entity(se
) {
3067 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3068 if (cfs_rq
->skip
!= se
)
3071 cfs_rq
->skip
= NULL
;
3075 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3077 if (cfs_rq
->last
== se
)
3078 __clear_buddies_last(se
);
3080 if (cfs_rq
->next
== se
)
3081 __clear_buddies_next(se
);
3083 if (cfs_rq
->skip
== se
)
3084 __clear_buddies_skip(se
);
3087 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3090 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3093 * Update run-time statistics of the 'current'.
3095 update_curr(cfs_rq
);
3096 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
3098 update_stats_dequeue(cfs_rq
, se
);
3099 if (flags
& DEQUEUE_SLEEP
) {
3100 #ifdef CONFIG_SCHEDSTATS
3101 if (entity_is_task(se
)) {
3102 struct task_struct
*tsk
= task_of(se
);
3104 if (tsk
->state
& TASK_INTERRUPTIBLE
)
3105 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
3106 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
3107 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
3112 clear_buddies(cfs_rq
, se
);
3114 if (se
!= cfs_rq
->curr
)
3115 __dequeue_entity(cfs_rq
, se
);
3117 account_entity_dequeue(cfs_rq
, se
);
3120 * Normalize the entity after updating the min_vruntime because the
3121 * update can refer to the ->curr item and we need to reflect this
3122 * movement in our normalized position.
3124 if (!(flags
& DEQUEUE_SLEEP
))
3125 se
->vruntime
-= cfs_rq
->min_vruntime
;
3127 /* return excess runtime on last dequeue */
3128 return_cfs_rq_runtime(cfs_rq
);
3130 update_min_vruntime(cfs_rq
);
3131 update_cfs_shares(cfs_rq
);
3135 * Preempt the current task with a newly woken task if needed:
3138 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3140 unsigned long ideal_runtime
, delta_exec
;
3141 struct sched_entity
*se
;
3144 ideal_runtime
= sched_slice(cfs_rq
, curr
);
3145 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
3146 if (delta_exec
> ideal_runtime
) {
3147 resched_curr(rq_of(cfs_rq
));
3149 * The current task ran long enough, ensure it doesn't get
3150 * re-elected due to buddy favours.
3152 clear_buddies(cfs_rq
, curr
);
3157 * Ensure that a task that missed wakeup preemption by a
3158 * narrow margin doesn't have to wait for a full slice.
3159 * This also mitigates buddy induced latencies under load.
3161 if (delta_exec
< sysctl_sched_min_granularity
)
3164 se
= __pick_first_entity(cfs_rq
);
3165 delta
= curr
->vruntime
- se
->vruntime
;
3170 if (delta
> ideal_runtime
)
3171 resched_curr(rq_of(cfs_rq
));
3175 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3177 /* 'current' is not kept within the tree. */
3180 * Any task has to be enqueued before it get to execute on
3181 * a CPU. So account for the time it spent waiting on the
3184 update_stats_wait_end(cfs_rq
, se
);
3185 __dequeue_entity(cfs_rq
, se
);
3188 update_stats_curr_start(cfs_rq
, se
);
3190 #ifdef CONFIG_SCHEDSTATS
3192 * Track our maximum slice length, if the CPU's load is at
3193 * least twice that of our own weight (i.e. dont track it
3194 * when there are only lesser-weight tasks around):
3196 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
3197 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
3198 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
3201 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
3205 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
3208 * Pick the next process, keeping these things in mind, in this order:
3209 * 1) keep things fair between processes/task groups
3210 * 2) pick the "next" process, since someone really wants that to run
3211 * 3) pick the "last" process, for cache locality
3212 * 4) do not run the "skip" process, if something else is available
3214 static struct sched_entity
*
3215 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3217 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
3218 struct sched_entity
*se
;
3221 * If curr is set we have to see if its left of the leftmost entity
3222 * still in the tree, provided there was anything in the tree at all.
3224 if (!left
|| (curr
&& entity_before(curr
, left
)))
3227 se
= left
; /* ideally we run the leftmost entity */
3230 * Avoid running the skip buddy, if running something else can
3231 * be done without getting too unfair.
3233 if (cfs_rq
->skip
== se
) {
3234 struct sched_entity
*second
;
3237 second
= __pick_first_entity(cfs_rq
);
3239 second
= __pick_next_entity(se
);
3240 if (!second
|| (curr
&& entity_before(curr
, second
)))
3244 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3249 * Prefer last buddy, try to return the CPU to a preempted task.
3251 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3255 * Someone really wants this to run. If it's not unfair, run it.
3257 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3260 clear_buddies(cfs_rq
, se
);
3265 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3267 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3270 * If still on the runqueue then deactivate_task()
3271 * was not called and update_curr() has to be done:
3274 update_curr(cfs_rq
);
3276 /* throttle cfs_rqs exceeding runtime */
3277 check_cfs_rq_runtime(cfs_rq
);
3279 check_spread(cfs_rq
, prev
);
3281 update_stats_wait_start(cfs_rq
, prev
);
3282 /* Put 'current' back into the tree. */
3283 __enqueue_entity(cfs_rq
, prev
);
3284 /* in !on_rq case, update occurred at dequeue */
3285 update_entity_load_avg(prev
, 1);
3287 cfs_rq
->curr
= NULL
;
3291 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3294 * Update run-time statistics of the 'current'.
3296 update_curr(cfs_rq
);
3299 * Ensure that runnable average is periodically updated.
3301 update_entity_load_avg(curr
, 1);
3302 update_cfs_rq_blocked_load(cfs_rq
, 1);
3303 update_cfs_shares(cfs_rq
);
3305 #ifdef CONFIG_SCHED_HRTICK
3307 * queued ticks are scheduled to match the slice, so don't bother
3308 * validating it and just reschedule.
3311 resched_curr(rq_of(cfs_rq
));
3315 * don't let the period tick interfere with the hrtick preemption
3317 if (!sched_feat(DOUBLE_TICK
) &&
3318 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3322 if (cfs_rq
->nr_running
> 1)
3323 check_preempt_tick(cfs_rq
, curr
);
3327 /**************************************************
3328 * CFS bandwidth control machinery
3331 #ifdef CONFIG_CFS_BANDWIDTH
3333 #ifdef HAVE_JUMP_LABEL
3334 static struct static_key __cfs_bandwidth_used
;
3336 static inline bool cfs_bandwidth_used(void)
3338 return static_key_false(&__cfs_bandwidth_used
);
3341 void cfs_bandwidth_usage_inc(void)
3343 static_key_slow_inc(&__cfs_bandwidth_used
);
3346 void cfs_bandwidth_usage_dec(void)
3348 static_key_slow_dec(&__cfs_bandwidth_used
);
3350 #else /* HAVE_JUMP_LABEL */
3351 static bool cfs_bandwidth_used(void)
3356 void cfs_bandwidth_usage_inc(void) {}
3357 void cfs_bandwidth_usage_dec(void) {}
3358 #endif /* HAVE_JUMP_LABEL */
3361 * default period for cfs group bandwidth.
3362 * default: 0.1s, units: nanoseconds
3364 static inline u64
default_cfs_period(void)
3366 return 100000000ULL;
3369 static inline u64
sched_cfs_bandwidth_slice(void)
3371 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3375 * Replenish runtime according to assigned quota and update expiration time.
3376 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3377 * additional synchronization around rq->lock.
3379 * requires cfs_b->lock
3381 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3385 if (cfs_b
->quota
== RUNTIME_INF
)
3388 now
= sched_clock_cpu(smp_processor_id());
3389 cfs_b
->runtime
= cfs_b
->quota
;
3390 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3393 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3395 return &tg
->cfs_bandwidth
;
3398 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3399 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3401 if (unlikely(cfs_rq
->throttle_count
))
3402 return cfs_rq
->throttled_clock_task
;
3404 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3407 /* returns 0 on failure to allocate runtime */
3408 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3410 struct task_group
*tg
= cfs_rq
->tg
;
3411 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3412 u64 amount
= 0, min_amount
, expires
;
3414 /* note: this is a positive sum as runtime_remaining <= 0 */
3415 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3417 raw_spin_lock(&cfs_b
->lock
);
3418 if (cfs_b
->quota
== RUNTIME_INF
)
3419 amount
= min_amount
;
3422 * If the bandwidth pool has become inactive, then at least one
3423 * period must have elapsed since the last consumption.
3424 * Refresh the global state and ensure bandwidth timer becomes
3427 if (!cfs_b
->timer_active
) {
3428 __refill_cfs_bandwidth_runtime(cfs_b
);
3429 __start_cfs_bandwidth(cfs_b
, false);
3432 if (cfs_b
->runtime
> 0) {
3433 amount
= min(cfs_b
->runtime
, min_amount
);
3434 cfs_b
->runtime
-= amount
;
3438 expires
= cfs_b
->runtime_expires
;
3439 raw_spin_unlock(&cfs_b
->lock
);
3441 cfs_rq
->runtime_remaining
+= amount
;
3443 * we may have advanced our local expiration to account for allowed
3444 * spread between our sched_clock and the one on which runtime was
3447 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3448 cfs_rq
->runtime_expires
= expires
;
3450 return cfs_rq
->runtime_remaining
> 0;
3454 * Note: This depends on the synchronization provided by sched_clock and the
3455 * fact that rq->clock snapshots this value.
3457 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3459 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3461 /* if the deadline is ahead of our clock, nothing to do */
3462 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3465 if (cfs_rq
->runtime_remaining
< 0)
3469 * If the local deadline has passed we have to consider the
3470 * possibility that our sched_clock is 'fast' and the global deadline
3471 * has not truly expired.
3473 * Fortunately we can check determine whether this the case by checking
3474 * whether the global deadline has advanced. It is valid to compare
3475 * cfs_b->runtime_expires without any locks since we only care about
3476 * exact equality, so a partial write will still work.
3479 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
3480 /* extend local deadline, drift is bounded above by 2 ticks */
3481 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3483 /* global deadline is ahead, expiration has passed */
3484 cfs_rq
->runtime_remaining
= 0;
3488 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3490 /* dock delta_exec before expiring quota (as it could span periods) */
3491 cfs_rq
->runtime_remaining
-= delta_exec
;
3492 expire_cfs_rq_runtime(cfs_rq
);
3494 if (likely(cfs_rq
->runtime_remaining
> 0))
3498 * if we're unable to extend our runtime we resched so that the active
3499 * hierarchy can be throttled
3501 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3502 resched_curr(rq_of(cfs_rq
));
3505 static __always_inline
3506 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3508 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3511 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3514 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3516 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3519 /* check whether cfs_rq, or any parent, is throttled */
3520 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3522 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3526 * Ensure that neither of the group entities corresponding to src_cpu or
3527 * dest_cpu are members of a throttled hierarchy when performing group
3528 * load-balance operations.
3530 static inline int throttled_lb_pair(struct task_group
*tg
,
3531 int src_cpu
, int dest_cpu
)
3533 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3535 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3536 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3538 return throttled_hierarchy(src_cfs_rq
) ||
3539 throttled_hierarchy(dest_cfs_rq
);
3542 /* updated child weight may affect parent so we have to do this bottom up */
3543 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3545 struct rq
*rq
= data
;
3546 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3548 cfs_rq
->throttle_count
--;
3550 if (!cfs_rq
->throttle_count
) {
3551 /* adjust cfs_rq_clock_task() */
3552 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3553 cfs_rq
->throttled_clock_task
;
3560 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3562 struct rq
*rq
= data
;
3563 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3565 /* group is entering throttled state, stop time */
3566 if (!cfs_rq
->throttle_count
)
3567 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3568 cfs_rq
->throttle_count
++;
3573 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3575 struct rq
*rq
= rq_of(cfs_rq
);
3576 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3577 struct sched_entity
*se
;
3578 long task_delta
, dequeue
= 1;
3580 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3582 /* freeze hierarchy runnable averages while throttled */
3584 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3587 task_delta
= cfs_rq
->h_nr_running
;
3588 for_each_sched_entity(se
) {
3589 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3590 /* throttled entity or throttle-on-deactivate */
3595 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3596 qcfs_rq
->h_nr_running
-= task_delta
;
3598 if (qcfs_rq
->load
.weight
)
3603 sub_nr_running(rq
, task_delta
);
3605 cfs_rq
->throttled
= 1;
3606 cfs_rq
->throttled_clock
= rq_clock(rq
);
3607 raw_spin_lock(&cfs_b
->lock
);
3609 * Add to the _head_ of the list, so that an already-started
3610 * distribute_cfs_runtime will not see us
3612 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3613 if (!cfs_b
->timer_active
)
3614 __start_cfs_bandwidth(cfs_b
, false);
3615 raw_spin_unlock(&cfs_b
->lock
);
3618 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3620 struct rq
*rq
= rq_of(cfs_rq
);
3621 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3622 struct sched_entity
*se
;
3626 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3628 cfs_rq
->throttled
= 0;
3630 update_rq_clock(rq
);
3632 raw_spin_lock(&cfs_b
->lock
);
3633 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3634 list_del_rcu(&cfs_rq
->throttled_list
);
3635 raw_spin_unlock(&cfs_b
->lock
);
3637 /* update hierarchical throttle state */
3638 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3640 if (!cfs_rq
->load
.weight
)
3643 task_delta
= cfs_rq
->h_nr_running
;
3644 for_each_sched_entity(se
) {
3648 cfs_rq
= cfs_rq_of(se
);
3650 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3651 cfs_rq
->h_nr_running
+= task_delta
;
3653 if (cfs_rq_throttled(cfs_rq
))
3658 add_nr_running(rq
, task_delta
);
3660 /* determine whether we need to wake up potentially idle cpu */
3661 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3665 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3666 u64 remaining
, u64 expires
)
3668 struct cfs_rq
*cfs_rq
;
3670 u64 starting_runtime
= remaining
;
3673 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3675 struct rq
*rq
= rq_of(cfs_rq
);
3677 raw_spin_lock(&rq
->lock
);
3678 if (!cfs_rq_throttled(cfs_rq
))
3681 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3682 if (runtime
> remaining
)
3683 runtime
= remaining
;
3684 remaining
-= runtime
;
3686 cfs_rq
->runtime_remaining
+= runtime
;
3687 cfs_rq
->runtime_expires
= expires
;
3689 /* we check whether we're throttled above */
3690 if (cfs_rq
->runtime_remaining
> 0)
3691 unthrottle_cfs_rq(cfs_rq
);
3694 raw_spin_unlock(&rq
->lock
);
3701 return starting_runtime
- remaining
;
3705 * Responsible for refilling a task_group's bandwidth and unthrottling its
3706 * cfs_rqs as appropriate. If there has been no activity within the last
3707 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3708 * used to track this state.
3710 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3712 u64 runtime
, runtime_expires
;
3715 /* no need to continue the timer with no bandwidth constraint */
3716 if (cfs_b
->quota
== RUNTIME_INF
)
3717 goto out_deactivate
;
3719 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3720 cfs_b
->nr_periods
+= overrun
;
3723 * idle depends on !throttled (for the case of a large deficit), and if
3724 * we're going inactive then everything else can be deferred
3726 if (cfs_b
->idle
&& !throttled
)
3727 goto out_deactivate
;
3730 * if we have relooped after returning idle once, we need to update our
3731 * status as actually running, so that other cpus doing
3732 * __start_cfs_bandwidth will stop trying to cancel us.
3734 cfs_b
->timer_active
= 1;
3736 __refill_cfs_bandwidth_runtime(cfs_b
);
3739 /* mark as potentially idle for the upcoming period */
3744 /* account preceding periods in which throttling occurred */
3745 cfs_b
->nr_throttled
+= overrun
;
3747 runtime_expires
= cfs_b
->runtime_expires
;
3750 * This check is repeated as we are holding onto the new bandwidth while
3751 * we unthrottle. This can potentially race with an unthrottled group
3752 * trying to acquire new bandwidth from the global pool. This can result
3753 * in us over-using our runtime if it is all used during this loop, but
3754 * only by limited amounts in that extreme case.
3756 while (throttled
&& cfs_b
->runtime
> 0) {
3757 runtime
= cfs_b
->runtime
;
3758 raw_spin_unlock(&cfs_b
->lock
);
3759 /* we can't nest cfs_b->lock while distributing bandwidth */
3760 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3762 raw_spin_lock(&cfs_b
->lock
);
3764 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3766 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3770 * While we are ensured activity in the period following an
3771 * unthrottle, this also covers the case in which the new bandwidth is
3772 * insufficient to cover the existing bandwidth deficit. (Forcing the
3773 * timer to remain active while there are any throttled entities.)
3780 cfs_b
->timer_active
= 0;
3784 /* a cfs_rq won't donate quota below this amount */
3785 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3786 /* minimum remaining period time to redistribute slack quota */
3787 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3788 /* how long we wait to gather additional slack before distributing */
3789 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3792 * Are we near the end of the current quota period?
3794 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3795 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3796 * migrate_hrtimers, base is never cleared, so we are fine.
3798 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3800 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3803 /* if the call-back is running a quota refresh is already occurring */
3804 if (hrtimer_callback_running(refresh_timer
))
3807 /* is a quota refresh about to occur? */
3808 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3809 if (remaining
< min_expire
)
3815 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3817 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3819 /* if there's a quota refresh soon don't bother with slack */
3820 if (runtime_refresh_within(cfs_b
, min_left
))
3823 start_bandwidth_timer(&cfs_b
->slack_timer
,
3824 ns_to_ktime(cfs_bandwidth_slack_period
));
3827 /* we know any runtime found here is valid as update_curr() precedes return */
3828 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3830 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3831 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3833 if (slack_runtime
<= 0)
3836 raw_spin_lock(&cfs_b
->lock
);
3837 if (cfs_b
->quota
!= RUNTIME_INF
&&
3838 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3839 cfs_b
->runtime
+= slack_runtime
;
3841 /* we are under rq->lock, defer unthrottling using a timer */
3842 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3843 !list_empty(&cfs_b
->throttled_cfs_rq
))
3844 start_cfs_slack_bandwidth(cfs_b
);
3846 raw_spin_unlock(&cfs_b
->lock
);
3848 /* even if it's not valid for return we don't want to try again */
3849 cfs_rq
->runtime_remaining
-= slack_runtime
;
3852 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3854 if (!cfs_bandwidth_used())
3857 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3860 __return_cfs_rq_runtime(cfs_rq
);
3864 * This is done with a timer (instead of inline with bandwidth return) since
3865 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3867 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3869 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3872 /* confirm we're still not at a refresh boundary */
3873 raw_spin_lock(&cfs_b
->lock
);
3874 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
3875 raw_spin_unlock(&cfs_b
->lock
);
3879 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
3880 runtime
= cfs_b
->runtime
;
3882 expires
= cfs_b
->runtime_expires
;
3883 raw_spin_unlock(&cfs_b
->lock
);
3888 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3890 raw_spin_lock(&cfs_b
->lock
);
3891 if (expires
== cfs_b
->runtime_expires
)
3892 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3893 raw_spin_unlock(&cfs_b
->lock
);
3897 * When a group wakes up we want to make sure that its quota is not already
3898 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3899 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3901 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3903 if (!cfs_bandwidth_used())
3906 /* an active group must be handled by the update_curr()->put() path */
3907 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3910 /* ensure the group is not already throttled */
3911 if (cfs_rq_throttled(cfs_rq
))
3914 /* update runtime allocation */
3915 account_cfs_rq_runtime(cfs_rq
, 0);
3916 if (cfs_rq
->runtime_remaining
<= 0)
3917 throttle_cfs_rq(cfs_rq
);
3920 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3921 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3923 if (!cfs_bandwidth_used())
3926 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3930 * it's possible for a throttled entity to be forced into a running
3931 * state (e.g. set_curr_task), in this case we're finished.
3933 if (cfs_rq_throttled(cfs_rq
))
3936 throttle_cfs_rq(cfs_rq
);
3940 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3942 struct cfs_bandwidth
*cfs_b
=
3943 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3944 do_sched_cfs_slack_timer(cfs_b
);
3946 return HRTIMER_NORESTART
;
3949 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3951 struct cfs_bandwidth
*cfs_b
=
3952 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3957 raw_spin_lock(&cfs_b
->lock
);
3959 now
= hrtimer_cb_get_time(timer
);
3960 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3965 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3967 raw_spin_unlock(&cfs_b
->lock
);
3969 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3972 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3974 raw_spin_lock_init(&cfs_b
->lock
);
3976 cfs_b
->quota
= RUNTIME_INF
;
3977 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3979 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3980 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3981 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3982 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3983 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3986 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3988 cfs_rq
->runtime_enabled
= 0;
3989 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3992 /* requires cfs_b->lock, may release to reprogram timer */
3993 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
, bool force
)
3996 * The timer may be active because we're trying to set a new bandwidth
3997 * period or because we're racing with the tear-down path
3998 * (timer_active==0 becomes visible before the hrtimer call-back
3999 * terminates). In either case we ensure that it's re-programmed
4001 while (unlikely(hrtimer_active(&cfs_b
->period_timer
)) &&
4002 hrtimer_try_to_cancel(&cfs_b
->period_timer
) < 0) {
4003 /* bounce the lock to allow do_sched_cfs_period_timer to run */
4004 raw_spin_unlock(&cfs_b
->lock
);
4006 raw_spin_lock(&cfs_b
->lock
);
4007 /* if someone else restarted the timer then we're done */
4008 if (!force
&& cfs_b
->timer_active
)
4012 cfs_b
->timer_active
= 1;
4013 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
4016 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4018 /* init_cfs_bandwidth() was not called */
4019 if (!cfs_b
->throttled_cfs_rq
.next
)
4022 hrtimer_cancel(&cfs_b
->period_timer
);
4023 hrtimer_cancel(&cfs_b
->slack_timer
);
4026 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
4028 struct cfs_rq
*cfs_rq
;
4030 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4031 struct cfs_bandwidth
*cfs_b
= &cfs_rq
->tg
->cfs_bandwidth
;
4033 raw_spin_lock(&cfs_b
->lock
);
4034 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
4035 raw_spin_unlock(&cfs_b
->lock
);
4039 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
4041 struct cfs_rq
*cfs_rq
;
4043 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4044 if (!cfs_rq
->runtime_enabled
)
4048 * clock_task is not advancing so we just need to make sure
4049 * there's some valid quota amount
4051 cfs_rq
->runtime_remaining
= 1;
4053 * Offline rq is schedulable till cpu is completely disabled
4054 * in take_cpu_down(), so we prevent new cfs throttling here.
4056 cfs_rq
->runtime_enabled
= 0;
4058 if (cfs_rq_throttled(cfs_rq
))
4059 unthrottle_cfs_rq(cfs_rq
);
4063 #else /* CONFIG_CFS_BANDWIDTH */
4064 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
4066 return rq_clock_task(rq_of(cfs_rq
));
4069 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
4070 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
4071 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
4072 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4074 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4079 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4084 static inline int throttled_lb_pair(struct task_group
*tg
,
4085 int src_cpu
, int dest_cpu
)
4090 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4092 #ifdef CONFIG_FAIR_GROUP_SCHED
4093 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4096 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4100 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4101 static inline void update_runtime_enabled(struct rq
*rq
) {}
4102 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
4104 #endif /* CONFIG_CFS_BANDWIDTH */
4106 /**************************************************
4107 * CFS operations on tasks:
4110 #ifdef CONFIG_SCHED_HRTICK
4111 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4113 struct sched_entity
*se
= &p
->se
;
4114 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4116 WARN_ON(task_rq(p
) != rq
);
4118 if (cfs_rq
->nr_running
> 1) {
4119 u64 slice
= sched_slice(cfs_rq
, se
);
4120 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
4121 s64 delta
= slice
- ran
;
4128 hrtick_start(rq
, delta
);
4133 * called from enqueue/dequeue and updates the hrtick when the
4134 * current task is from our class and nr_running is low enough
4137 static void hrtick_update(struct rq
*rq
)
4139 struct task_struct
*curr
= rq
->curr
;
4141 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
4144 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
4145 hrtick_start_fair(rq
, curr
);
4147 #else /* !CONFIG_SCHED_HRTICK */
4149 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4153 static inline void hrtick_update(struct rq
*rq
)
4159 * The enqueue_task method is called before nr_running is
4160 * increased. Here we update the fair scheduling stats and
4161 * then put the task into the rbtree:
4164 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4166 struct cfs_rq
*cfs_rq
;
4167 struct sched_entity
*se
= &p
->se
;
4169 for_each_sched_entity(se
) {
4172 cfs_rq
= cfs_rq_of(se
);
4173 enqueue_entity(cfs_rq
, se
, flags
);
4176 * end evaluation on encountering a throttled cfs_rq
4178 * note: in the case of encountering a throttled cfs_rq we will
4179 * post the final h_nr_running increment below.
4181 if (cfs_rq_throttled(cfs_rq
))
4183 cfs_rq
->h_nr_running
++;
4185 flags
= ENQUEUE_WAKEUP
;
4188 for_each_sched_entity(se
) {
4189 cfs_rq
= cfs_rq_of(se
);
4190 cfs_rq
->h_nr_running
++;
4192 if (cfs_rq_throttled(cfs_rq
))
4195 update_cfs_shares(cfs_rq
);
4196 update_entity_load_avg(se
, 1);
4200 update_rq_runnable_avg(rq
, rq
->nr_running
);
4201 add_nr_running(rq
, 1);
4206 static void set_next_buddy(struct sched_entity
*se
);
4209 * The dequeue_task method is called before nr_running is
4210 * decreased. We remove the task from the rbtree and
4211 * update the fair scheduling stats:
4213 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4215 struct cfs_rq
*cfs_rq
;
4216 struct sched_entity
*se
= &p
->se
;
4217 int task_sleep
= flags
& DEQUEUE_SLEEP
;
4219 for_each_sched_entity(se
) {
4220 cfs_rq
= cfs_rq_of(se
);
4221 dequeue_entity(cfs_rq
, se
, flags
);
4224 * end evaluation on encountering a throttled cfs_rq
4226 * note: in the case of encountering a throttled cfs_rq we will
4227 * post the final h_nr_running decrement below.
4229 if (cfs_rq_throttled(cfs_rq
))
4231 cfs_rq
->h_nr_running
--;
4233 /* Don't dequeue parent if it has other entities besides us */
4234 if (cfs_rq
->load
.weight
) {
4236 * Bias pick_next to pick a task from this cfs_rq, as
4237 * p is sleeping when it is within its sched_slice.
4239 if (task_sleep
&& parent_entity(se
))
4240 set_next_buddy(parent_entity(se
));
4242 /* avoid re-evaluating load for this entity */
4243 se
= parent_entity(se
);
4246 flags
|= DEQUEUE_SLEEP
;
4249 for_each_sched_entity(se
) {
4250 cfs_rq
= cfs_rq_of(se
);
4251 cfs_rq
->h_nr_running
--;
4253 if (cfs_rq_throttled(cfs_rq
))
4256 update_cfs_shares(cfs_rq
);
4257 update_entity_load_avg(se
, 1);
4261 sub_nr_running(rq
, 1);
4262 update_rq_runnable_avg(rq
, 1);
4268 /* Used instead of source_load when we know the type == 0 */
4269 static unsigned long weighted_cpuload(const int cpu
)
4271 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
4275 * Return a low guess at the load of a migration-source cpu weighted
4276 * according to the scheduling class and "nice" value.
4278 * We want to under-estimate the load of migration sources, to
4279 * balance conservatively.
4281 static unsigned long source_load(int cpu
, int type
)
4283 struct rq
*rq
= cpu_rq(cpu
);
4284 unsigned long total
= weighted_cpuload(cpu
);
4286 if (type
== 0 || !sched_feat(LB_BIAS
))
4289 return min(rq
->cpu_load
[type
-1], total
);
4293 * Return a high guess at the load of a migration-target cpu weighted
4294 * according to the scheduling class and "nice" value.
4296 static unsigned long target_load(int cpu
, int type
)
4298 struct rq
*rq
= cpu_rq(cpu
);
4299 unsigned long total
= weighted_cpuload(cpu
);
4301 if (type
== 0 || !sched_feat(LB_BIAS
))
4304 return max(rq
->cpu_load
[type
-1], total
);
4307 static unsigned long capacity_of(int cpu
)
4309 return cpu_rq(cpu
)->cpu_capacity
;
4312 static unsigned long cpu_avg_load_per_task(int cpu
)
4314 struct rq
*rq
= cpu_rq(cpu
);
4315 unsigned long nr_running
= ACCESS_ONCE(rq
->cfs
.h_nr_running
);
4316 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
4319 return load_avg
/ nr_running
;
4324 static void record_wakee(struct task_struct
*p
)
4327 * Rough decay (wiping) for cost saving, don't worry
4328 * about the boundary, really active task won't care
4331 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
4332 current
->wakee_flips
>>= 1;
4333 current
->wakee_flip_decay_ts
= jiffies
;
4336 if (current
->last_wakee
!= p
) {
4337 current
->last_wakee
= p
;
4338 current
->wakee_flips
++;
4342 static void task_waking_fair(struct task_struct
*p
)
4344 struct sched_entity
*se
= &p
->se
;
4345 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4348 #ifndef CONFIG_64BIT
4349 u64 min_vruntime_copy
;
4352 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
4354 min_vruntime
= cfs_rq
->min_vruntime
;
4355 } while (min_vruntime
!= min_vruntime_copy
);
4357 min_vruntime
= cfs_rq
->min_vruntime
;
4360 se
->vruntime
-= min_vruntime
;
4364 #ifdef CONFIG_FAIR_GROUP_SCHED
4366 * effective_load() calculates the load change as seen from the root_task_group
4368 * Adding load to a group doesn't make a group heavier, but can cause movement
4369 * of group shares between cpus. Assuming the shares were perfectly aligned one
4370 * can calculate the shift in shares.
4372 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4373 * on this @cpu and results in a total addition (subtraction) of @wg to the
4374 * total group weight.
4376 * Given a runqueue weight distribution (rw_i) we can compute a shares
4377 * distribution (s_i) using:
4379 * s_i = rw_i / \Sum rw_j (1)
4381 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4382 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4383 * shares distribution (s_i):
4385 * rw_i = { 2, 4, 1, 0 }
4386 * s_i = { 2/7, 4/7, 1/7, 0 }
4388 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4389 * task used to run on and the CPU the waker is running on), we need to
4390 * compute the effect of waking a task on either CPU and, in case of a sync
4391 * wakeup, compute the effect of the current task going to sleep.
4393 * So for a change of @wl to the local @cpu with an overall group weight change
4394 * of @wl we can compute the new shares distribution (s'_i) using:
4396 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4398 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4399 * differences in waking a task to CPU 0. The additional task changes the
4400 * weight and shares distributions like:
4402 * rw'_i = { 3, 4, 1, 0 }
4403 * s'_i = { 3/8, 4/8, 1/8, 0 }
4405 * We can then compute the difference in effective weight by using:
4407 * dw_i = S * (s'_i - s_i) (3)
4409 * Where 'S' is the group weight as seen by its parent.
4411 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4412 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4413 * 4/7) times the weight of the group.
4415 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4417 struct sched_entity
*se
= tg
->se
[cpu
];
4419 if (!tg
->parent
) /* the trivial, non-cgroup case */
4422 for_each_sched_entity(se
) {
4428 * W = @wg + \Sum rw_j
4430 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4435 w
= se
->my_q
->load
.weight
+ wl
;
4438 * wl = S * s'_i; see (2)
4441 wl
= (w
* (long)tg
->shares
) / W
;
4446 * Per the above, wl is the new se->load.weight value; since
4447 * those are clipped to [MIN_SHARES, ...) do so now. See
4448 * calc_cfs_shares().
4450 if (wl
< MIN_SHARES
)
4454 * wl = dw_i = S * (s'_i - s_i); see (3)
4456 wl
-= se
->load
.weight
;
4459 * Recursively apply this logic to all parent groups to compute
4460 * the final effective load change on the root group. Since
4461 * only the @tg group gets extra weight, all parent groups can
4462 * only redistribute existing shares. @wl is the shift in shares
4463 * resulting from this level per the above.
4472 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4479 static int wake_wide(struct task_struct
*p
)
4481 int factor
= this_cpu_read(sd_llc_size
);
4484 * Yeah, it's the switching-frequency, could means many wakee or
4485 * rapidly switch, use factor here will just help to automatically
4486 * adjust the loose-degree, so bigger node will lead to more pull.
4488 if (p
->wakee_flips
> factor
) {
4490 * wakee is somewhat hot, it needs certain amount of cpu
4491 * resource, so if waker is far more hot, prefer to leave
4494 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
4501 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4503 s64 this_load
, load
;
4504 s64 this_eff_load
, prev_eff_load
;
4505 int idx
, this_cpu
, prev_cpu
;
4506 struct task_group
*tg
;
4507 unsigned long weight
;
4511 * If we wake multiple tasks be careful to not bounce
4512 * ourselves around too much.
4518 this_cpu
= smp_processor_id();
4519 prev_cpu
= task_cpu(p
);
4520 load
= source_load(prev_cpu
, idx
);
4521 this_load
= target_load(this_cpu
, idx
);
4524 * If sync wakeup then subtract the (maximum possible)
4525 * effect of the currently running task from the load
4526 * of the current CPU:
4529 tg
= task_group(current
);
4530 weight
= current
->se
.load
.weight
;
4532 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4533 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4537 weight
= p
->se
.load
.weight
;
4540 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4541 * due to the sync cause above having dropped this_load to 0, we'll
4542 * always have an imbalance, but there's really nothing you can do
4543 * about that, so that's good too.
4545 * Otherwise check if either cpus are near enough in load to allow this
4546 * task to be woken on this_cpu.
4548 this_eff_load
= 100;
4549 this_eff_load
*= capacity_of(prev_cpu
);
4551 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4552 prev_eff_load
*= capacity_of(this_cpu
);
4554 if (this_load
> 0) {
4555 this_eff_load
*= this_load
+
4556 effective_load(tg
, this_cpu
, weight
, weight
);
4558 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4561 balanced
= this_eff_load
<= prev_eff_load
;
4563 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4568 schedstat_inc(sd
, ttwu_move_affine
);
4569 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4575 * find_idlest_group finds and returns the least busy CPU group within the
4578 static struct sched_group
*
4579 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4580 int this_cpu
, int sd_flag
)
4582 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4583 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4584 int load_idx
= sd
->forkexec_idx
;
4585 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4587 if (sd_flag
& SD_BALANCE_WAKE
)
4588 load_idx
= sd
->wake_idx
;
4591 unsigned long load
, avg_load
;
4595 /* Skip over this group if it has no CPUs allowed */
4596 if (!cpumask_intersects(sched_group_cpus(group
),
4597 tsk_cpus_allowed(p
)))
4600 local_group
= cpumask_test_cpu(this_cpu
,
4601 sched_group_cpus(group
));
4603 /* Tally up the load of all CPUs in the group */
4606 for_each_cpu(i
, sched_group_cpus(group
)) {
4607 /* Bias balancing toward cpus of our domain */
4609 load
= source_load(i
, load_idx
);
4611 load
= target_load(i
, load_idx
);
4616 /* Adjust by relative CPU capacity of the group */
4617 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) / group
->sgc
->capacity
;
4620 this_load
= avg_load
;
4621 } else if (avg_load
< min_load
) {
4622 min_load
= avg_load
;
4625 } while (group
= group
->next
, group
!= sd
->groups
);
4627 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4633 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4636 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4638 unsigned long load
, min_load
= ULONG_MAX
;
4639 unsigned int min_exit_latency
= UINT_MAX
;
4640 u64 latest_idle_timestamp
= 0;
4641 int least_loaded_cpu
= this_cpu
;
4642 int shallowest_idle_cpu
= -1;
4645 /* Traverse only the allowed CPUs */
4646 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4648 struct rq
*rq
= cpu_rq(i
);
4649 struct cpuidle_state
*idle
= idle_get_state(rq
);
4650 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
4652 * We give priority to a CPU whose idle state
4653 * has the smallest exit latency irrespective
4654 * of any idle timestamp.
4656 min_exit_latency
= idle
->exit_latency
;
4657 latest_idle_timestamp
= rq
->idle_stamp
;
4658 shallowest_idle_cpu
= i
;
4659 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
4660 rq
->idle_stamp
> latest_idle_timestamp
) {
4662 * If equal or no active idle state, then
4663 * the most recently idled CPU might have
4666 latest_idle_timestamp
= rq
->idle_stamp
;
4667 shallowest_idle_cpu
= i
;
4669 } else if (shallowest_idle_cpu
== -1) {
4670 load
= weighted_cpuload(i
);
4671 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4673 least_loaded_cpu
= i
;
4678 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
4682 * Try and locate an idle CPU in the sched_domain.
4684 static int select_idle_sibling(struct task_struct
*p
, int target
)
4686 struct sched_domain
*sd
;
4687 struct sched_group
*sg
;
4688 int i
= task_cpu(p
);
4690 if (idle_cpu(target
))
4694 * If the prevous cpu is cache affine and idle, don't be stupid.
4696 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4700 * Otherwise, iterate the domains and find an elegible idle cpu.
4702 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4703 for_each_lower_domain(sd
) {
4706 if (!cpumask_intersects(sched_group_cpus(sg
),
4707 tsk_cpus_allowed(p
)))
4710 for_each_cpu(i
, sched_group_cpus(sg
)) {
4711 if (i
== target
|| !idle_cpu(i
))
4715 target
= cpumask_first_and(sched_group_cpus(sg
),
4716 tsk_cpus_allowed(p
));
4720 } while (sg
!= sd
->groups
);
4727 * select_task_rq_fair: Select target runqueue for the waking task in domains
4728 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4729 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4731 * Balances load by selecting the idlest cpu in the idlest group, or under
4732 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4734 * Returns the target cpu number.
4736 * preempt must be disabled.
4739 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4741 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4742 int cpu
= smp_processor_id();
4744 int want_affine
= 0;
4745 int sync
= wake_flags
& WF_SYNC
;
4747 if (sd_flag
& SD_BALANCE_WAKE
)
4748 want_affine
= cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
));
4751 for_each_domain(cpu
, tmp
) {
4752 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4756 * If both cpu and prev_cpu are part of this domain,
4757 * cpu is a valid SD_WAKE_AFFINE target.
4759 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4760 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4765 if (tmp
->flags
& sd_flag
)
4769 if (affine_sd
&& cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4772 if (sd_flag
& SD_BALANCE_WAKE
) {
4773 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4778 struct sched_group
*group
;
4781 if (!(sd
->flags
& sd_flag
)) {
4786 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
4792 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4793 if (new_cpu
== -1 || new_cpu
== cpu
) {
4794 /* Now try balancing at a lower domain level of cpu */
4799 /* Now try balancing at a lower domain level of new_cpu */
4801 weight
= sd
->span_weight
;
4803 for_each_domain(cpu
, tmp
) {
4804 if (weight
<= tmp
->span_weight
)
4806 if (tmp
->flags
& sd_flag
)
4809 /* while loop will break here if sd == NULL */
4818 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4819 * cfs_rq_of(p) references at time of call are still valid and identify the
4820 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4821 * other assumptions, including the state of rq->lock, should be made.
4824 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4826 struct sched_entity
*se
= &p
->se
;
4827 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4830 * Load tracking: accumulate removed load so that it can be processed
4831 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4832 * to blocked load iff they have a positive decay-count. It can never
4833 * be negative here since on-rq tasks have decay-count == 0.
4835 if (se
->avg
.decay_count
) {
4836 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4837 atomic_long_add(se
->avg
.load_avg_contrib
,
4838 &cfs_rq
->removed_load
);
4841 /* We have migrated, no longer consider this task hot */
4844 #endif /* CONFIG_SMP */
4846 static unsigned long
4847 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4849 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4852 * Since its curr running now, convert the gran from real-time
4853 * to virtual-time in his units.
4855 * By using 'se' instead of 'curr' we penalize light tasks, so
4856 * they get preempted easier. That is, if 'se' < 'curr' then
4857 * the resulting gran will be larger, therefore penalizing the
4858 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4859 * be smaller, again penalizing the lighter task.
4861 * This is especially important for buddies when the leftmost
4862 * task is higher priority than the buddy.
4864 return calc_delta_fair(gran
, se
);
4868 * Should 'se' preempt 'curr'.
4882 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4884 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4889 gran
= wakeup_gran(curr
, se
);
4896 static void set_last_buddy(struct sched_entity
*se
)
4898 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4901 for_each_sched_entity(se
)
4902 cfs_rq_of(se
)->last
= se
;
4905 static void set_next_buddy(struct sched_entity
*se
)
4907 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4910 for_each_sched_entity(se
)
4911 cfs_rq_of(se
)->next
= se
;
4914 static void set_skip_buddy(struct sched_entity
*se
)
4916 for_each_sched_entity(se
)
4917 cfs_rq_of(se
)->skip
= se
;
4921 * Preempt the current task with a newly woken task if needed:
4923 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4925 struct task_struct
*curr
= rq
->curr
;
4926 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4927 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4928 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4929 int next_buddy_marked
= 0;
4931 if (unlikely(se
== pse
))
4935 * This is possible from callers such as attach_tasks(), in which we
4936 * unconditionally check_prempt_curr() after an enqueue (which may have
4937 * lead to a throttle). This both saves work and prevents false
4938 * next-buddy nomination below.
4940 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4943 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4944 set_next_buddy(pse
);
4945 next_buddy_marked
= 1;
4949 * We can come here with TIF_NEED_RESCHED already set from new task
4952 * Note: this also catches the edge-case of curr being in a throttled
4953 * group (e.g. via set_curr_task), since update_curr() (in the
4954 * enqueue of curr) will have resulted in resched being set. This
4955 * prevents us from potentially nominating it as a false LAST_BUDDY
4958 if (test_tsk_need_resched(curr
))
4961 /* Idle tasks are by definition preempted by non-idle tasks. */
4962 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4963 likely(p
->policy
!= SCHED_IDLE
))
4967 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4968 * is driven by the tick):
4970 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4973 find_matching_se(&se
, &pse
);
4974 update_curr(cfs_rq_of(se
));
4976 if (wakeup_preempt_entity(se
, pse
) == 1) {
4978 * Bias pick_next to pick the sched entity that is
4979 * triggering this preemption.
4981 if (!next_buddy_marked
)
4982 set_next_buddy(pse
);
4991 * Only set the backward buddy when the current task is still
4992 * on the rq. This can happen when a wakeup gets interleaved
4993 * with schedule on the ->pre_schedule() or idle_balance()
4994 * point, either of which can * drop the rq lock.
4996 * Also, during early boot the idle thread is in the fair class,
4997 * for obvious reasons its a bad idea to schedule back to it.
4999 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
5002 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
5006 static struct task_struct
*
5007 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
)
5009 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
5010 struct sched_entity
*se
;
5011 struct task_struct
*p
;
5015 #ifdef CONFIG_FAIR_GROUP_SCHED
5016 if (!cfs_rq
->nr_running
)
5019 if (prev
->sched_class
!= &fair_sched_class
)
5023 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5024 * likely that a next task is from the same cgroup as the current.
5026 * Therefore attempt to avoid putting and setting the entire cgroup
5027 * hierarchy, only change the part that actually changes.
5031 struct sched_entity
*curr
= cfs_rq
->curr
;
5034 * Since we got here without doing put_prev_entity() we also
5035 * have to consider cfs_rq->curr. If it is still a runnable
5036 * entity, update_curr() will update its vruntime, otherwise
5037 * forget we've ever seen it.
5039 if (curr
&& curr
->on_rq
)
5040 update_curr(cfs_rq
);
5045 * This call to check_cfs_rq_runtime() will do the throttle and
5046 * dequeue its entity in the parent(s). Therefore the 'simple'
5047 * nr_running test will indeed be correct.
5049 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
5052 se
= pick_next_entity(cfs_rq
, curr
);
5053 cfs_rq
= group_cfs_rq(se
);
5059 * Since we haven't yet done put_prev_entity and if the selected task
5060 * is a different task than we started out with, try and touch the
5061 * least amount of cfs_rqs.
5064 struct sched_entity
*pse
= &prev
->se
;
5066 while (!(cfs_rq
= is_same_group(se
, pse
))) {
5067 int se_depth
= se
->depth
;
5068 int pse_depth
= pse
->depth
;
5070 if (se_depth
<= pse_depth
) {
5071 put_prev_entity(cfs_rq_of(pse
), pse
);
5072 pse
= parent_entity(pse
);
5074 if (se_depth
>= pse_depth
) {
5075 set_next_entity(cfs_rq_of(se
), se
);
5076 se
= parent_entity(se
);
5080 put_prev_entity(cfs_rq
, pse
);
5081 set_next_entity(cfs_rq
, se
);
5084 if (hrtick_enabled(rq
))
5085 hrtick_start_fair(rq
, p
);
5092 if (!cfs_rq
->nr_running
)
5095 put_prev_task(rq
, prev
);
5098 se
= pick_next_entity(cfs_rq
, NULL
);
5099 set_next_entity(cfs_rq
, se
);
5100 cfs_rq
= group_cfs_rq(se
);
5105 if (hrtick_enabled(rq
))
5106 hrtick_start_fair(rq
, p
);
5111 new_tasks
= idle_balance(rq
);
5113 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5114 * possible for any higher priority task to appear. In that case we
5115 * must re-start the pick_next_entity() loop.
5127 * Account for a descheduled task:
5129 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
5131 struct sched_entity
*se
= &prev
->se
;
5132 struct cfs_rq
*cfs_rq
;
5134 for_each_sched_entity(se
) {
5135 cfs_rq
= cfs_rq_of(se
);
5136 put_prev_entity(cfs_rq
, se
);
5141 * sched_yield() is very simple
5143 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5145 static void yield_task_fair(struct rq
*rq
)
5147 struct task_struct
*curr
= rq
->curr
;
5148 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
5149 struct sched_entity
*se
= &curr
->se
;
5152 * Are we the only task in the tree?
5154 if (unlikely(rq
->nr_running
== 1))
5157 clear_buddies(cfs_rq
, se
);
5159 if (curr
->policy
!= SCHED_BATCH
) {
5160 update_rq_clock(rq
);
5162 * Update run-time statistics of the 'current'.
5164 update_curr(cfs_rq
);
5166 * Tell update_rq_clock() that we've just updated,
5167 * so we don't do microscopic update in schedule()
5168 * and double the fastpath cost.
5170 rq_clock_skip_update(rq
, true);
5176 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
5178 struct sched_entity
*se
= &p
->se
;
5180 /* throttled hierarchies are not runnable */
5181 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
5184 /* Tell the scheduler that we'd really like pse to run next. */
5187 yield_task_fair(rq
);
5193 /**************************************************
5194 * Fair scheduling class load-balancing methods.
5198 * The purpose of load-balancing is to achieve the same basic fairness the
5199 * per-cpu scheduler provides, namely provide a proportional amount of compute
5200 * time to each task. This is expressed in the following equation:
5202 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5204 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5205 * W_i,0 is defined as:
5207 * W_i,0 = \Sum_j w_i,j (2)
5209 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5210 * is derived from the nice value as per prio_to_weight[].
5212 * The weight average is an exponential decay average of the instantaneous
5215 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5217 * C_i is the compute capacity of cpu i, typically it is the
5218 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5219 * can also include other factors [XXX].
5221 * To achieve this balance we define a measure of imbalance which follows
5222 * directly from (1):
5224 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5226 * We them move tasks around to minimize the imbalance. In the continuous
5227 * function space it is obvious this converges, in the discrete case we get
5228 * a few fun cases generally called infeasible weight scenarios.
5231 * - infeasible weights;
5232 * - local vs global optima in the discrete case. ]
5237 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5238 * for all i,j solution, we create a tree of cpus that follows the hardware
5239 * topology where each level pairs two lower groups (or better). This results
5240 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5241 * tree to only the first of the previous level and we decrease the frequency
5242 * of load-balance at each level inv. proportional to the number of cpus in
5248 * \Sum { --- * --- * 2^i } = O(n) (5)
5250 * `- size of each group
5251 * | | `- number of cpus doing load-balance
5253 * `- sum over all levels
5255 * Coupled with a limit on how many tasks we can migrate every balance pass,
5256 * this makes (5) the runtime complexity of the balancer.
5258 * An important property here is that each CPU is still (indirectly) connected
5259 * to every other cpu in at most O(log n) steps:
5261 * The adjacency matrix of the resulting graph is given by:
5264 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5267 * And you'll find that:
5269 * A^(log_2 n)_i,j != 0 for all i,j (7)
5271 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5272 * The task movement gives a factor of O(m), giving a convergence complexity
5275 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5280 * In order to avoid CPUs going idle while there's still work to do, new idle
5281 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5282 * tree itself instead of relying on other CPUs to bring it work.
5284 * This adds some complexity to both (5) and (8) but it reduces the total idle
5292 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5295 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5300 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5302 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5304 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5307 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5308 * rewrite all of this once again.]
5311 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
5313 enum fbq_type
{ regular
, remote
, all
};
5315 #define LBF_ALL_PINNED 0x01
5316 #define LBF_NEED_BREAK 0x02
5317 #define LBF_DST_PINNED 0x04
5318 #define LBF_SOME_PINNED 0x08
5321 struct sched_domain
*sd
;
5329 struct cpumask
*dst_grpmask
;
5331 enum cpu_idle_type idle
;
5333 /* The set of CPUs under consideration for load-balancing */
5334 struct cpumask
*cpus
;
5339 unsigned int loop_break
;
5340 unsigned int loop_max
;
5342 enum fbq_type fbq_type
;
5343 struct list_head tasks
;
5347 * Is this task likely cache-hot:
5349 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
5353 lockdep_assert_held(&env
->src_rq
->lock
);
5355 if (p
->sched_class
!= &fair_sched_class
)
5358 if (unlikely(p
->policy
== SCHED_IDLE
))
5362 * Buddy candidates are cache hot:
5364 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
5365 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
5366 &p
->se
== cfs_rq_of(&p
->se
)->last
))
5369 if (sysctl_sched_migration_cost
== -1)
5371 if (sysctl_sched_migration_cost
== 0)
5374 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
5376 return delta
< (s64
)sysctl_sched_migration_cost
;
5379 #ifdef CONFIG_NUMA_BALANCING
5380 /* Returns true if the destination node has incurred more faults */
5381 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
5383 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5384 int src_nid
, dst_nid
;
5386 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults
||
5387 !(env
->sd
->flags
& SD_NUMA
)) {
5391 src_nid
= cpu_to_node(env
->src_cpu
);
5392 dst_nid
= cpu_to_node(env
->dst_cpu
);
5394 if (src_nid
== dst_nid
)
5398 /* Task is already in the group's interleave set. */
5399 if (node_isset(src_nid
, numa_group
->active_nodes
))
5402 /* Task is moving into the group's interleave set. */
5403 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5406 return group_faults(p
, dst_nid
) > group_faults(p
, src_nid
);
5409 /* Encourage migration to the preferred node. */
5410 if (dst_nid
== p
->numa_preferred_nid
)
5413 return task_faults(p
, dst_nid
) > task_faults(p
, src_nid
);
5417 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
5419 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5420 int src_nid
, dst_nid
;
5422 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
5425 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
5428 src_nid
= cpu_to_node(env
->src_cpu
);
5429 dst_nid
= cpu_to_node(env
->dst_cpu
);
5431 if (src_nid
== dst_nid
)
5435 /* Task is moving within/into the group's interleave set. */
5436 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5439 /* Task is moving out of the group's interleave set. */
5440 if (node_isset(src_nid
, numa_group
->active_nodes
))
5443 return group_faults(p
, dst_nid
) < group_faults(p
, src_nid
);
5446 /* Migrating away from the preferred node is always bad. */
5447 if (src_nid
== p
->numa_preferred_nid
)
5450 return task_faults(p
, dst_nid
) < task_faults(p
, src_nid
);
5454 static inline bool migrate_improves_locality(struct task_struct
*p
,
5460 static inline bool migrate_degrades_locality(struct task_struct
*p
,
5468 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5471 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
5473 int tsk_cache_hot
= 0;
5475 lockdep_assert_held(&env
->src_rq
->lock
);
5478 * We do not migrate tasks that are:
5479 * 1) throttled_lb_pair, or
5480 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5481 * 3) running (obviously), or
5482 * 4) are cache-hot on their current CPU.
5484 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
5487 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
5490 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
5492 env
->flags
|= LBF_SOME_PINNED
;
5495 * Remember if this task can be migrated to any other cpu in
5496 * our sched_group. We may want to revisit it if we couldn't
5497 * meet load balance goals by pulling other tasks on src_cpu.
5499 * Also avoid computing new_dst_cpu if we have already computed
5500 * one in current iteration.
5502 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
5505 /* Prevent to re-select dst_cpu via env's cpus */
5506 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
5507 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
5508 env
->flags
|= LBF_DST_PINNED
;
5509 env
->new_dst_cpu
= cpu
;
5517 /* Record that we found atleast one task that could run on dst_cpu */
5518 env
->flags
&= ~LBF_ALL_PINNED
;
5520 if (task_running(env
->src_rq
, p
)) {
5521 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
5526 * Aggressive migration if:
5527 * 1) destination numa is preferred
5528 * 2) task is cache cold, or
5529 * 3) too many balance attempts have failed.
5531 tsk_cache_hot
= task_hot(p
, env
);
5533 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
5535 if (migrate_improves_locality(p
, env
) || !tsk_cache_hot
||
5536 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
5537 if (tsk_cache_hot
) {
5538 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5539 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5544 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
5549 * detach_task() -- detach the task for the migration specified in env
5551 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
5553 lockdep_assert_held(&env
->src_rq
->lock
);
5555 deactivate_task(env
->src_rq
, p
, 0);
5556 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
5557 set_task_cpu(p
, env
->dst_cpu
);
5561 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5562 * part of active balancing operations within "domain".
5564 * Returns a task if successful and NULL otherwise.
5566 static struct task_struct
*detach_one_task(struct lb_env
*env
)
5568 struct task_struct
*p
, *n
;
5570 lockdep_assert_held(&env
->src_rq
->lock
);
5572 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
5573 if (!can_migrate_task(p
, env
))
5576 detach_task(p
, env
);
5579 * Right now, this is only the second place where
5580 * lb_gained[env->idle] is updated (other is detach_tasks)
5581 * so we can safely collect stats here rather than
5582 * inside detach_tasks().
5584 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
5590 static const unsigned int sched_nr_migrate_break
= 32;
5593 * detach_tasks() -- tries to detach up to imbalance weighted load from
5594 * busiest_rq, as part of a balancing operation within domain "sd".
5596 * Returns number of detached tasks if successful and 0 otherwise.
5598 static int detach_tasks(struct lb_env
*env
)
5600 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
5601 struct task_struct
*p
;
5605 lockdep_assert_held(&env
->src_rq
->lock
);
5607 if (env
->imbalance
<= 0)
5610 while (!list_empty(tasks
)) {
5611 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5614 /* We've more or less seen every task there is, call it quits */
5615 if (env
->loop
> env
->loop_max
)
5618 /* take a breather every nr_migrate tasks */
5619 if (env
->loop
> env
->loop_break
) {
5620 env
->loop_break
+= sched_nr_migrate_break
;
5621 env
->flags
|= LBF_NEED_BREAK
;
5625 if (!can_migrate_task(p
, env
))
5628 load
= task_h_load(p
);
5630 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5633 if ((load
/ 2) > env
->imbalance
)
5636 detach_task(p
, env
);
5637 list_add(&p
->se
.group_node
, &env
->tasks
);
5640 env
->imbalance
-= load
;
5642 #ifdef CONFIG_PREEMPT
5644 * NEWIDLE balancing is a source of latency, so preemptible
5645 * kernels will stop after the first task is detached to minimize
5646 * the critical section.
5648 if (env
->idle
== CPU_NEWLY_IDLE
)
5653 * We only want to steal up to the prescribed amount of
5656 if (env
->imbalance
<= 0)
5661 list_move_tail(&p
->se
.group_node
, tasks
);
5665 * Right now, this is one of only two places we collect this stat
5666 * so we can safely collect detach_one_task() stats here rather
5667 * than inside detach_one_task().
5669 schedstat_add(env
->sd
, lb_gained
[env
->idle
], detached
);
5675 * attach_task() -- attach the task detached by detach_task() to its new rq.
5677 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
5679 lockdep_assert_held(&rq
->lock
);
5681 BUG_ON(task_rq(p
) != rq
);
5682 p
->on_rq
= TASK_ON_RQ_QUEUED
;
5683 activate_task(rq
, p
, 0);
5684 check_preempt_curr(rq
, p
, 0);
5688 * attach_one_task() -- attaches the task returned from detach_one_task() to
5691 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
5693 raw_spin_lock(&rq
->lock
);
5695 raw_spin_unlock(&rq
->lock
);
5699 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5702 static void attach_tasks(struct lb_env
*env
)
5704 struct list_head
*tasks
= &env
->tasks
;
5705 struct task_struct
*p
;
5707 raw_spin_lock(&env
->dst_rq
->lock
);
5709 while (!list_empty(tasks
)) {
5710 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5711 list_del_init(&p
->se
.group_node
);
5713 attach_task(env
->dst_rq
, p
);
5716 raw_spin_unlock(&env
->dst_rq
->lock
);
5719 #ifdef CONFIG_FAIR_GROUP_SCHED
5721 * update tg->load_weight by folding this cpu's load_avg
5723 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5725 struct sched_entity
*se
= tg
->se
[cpu
];
5726 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5728 /* throttled entities do not contribute to load */
5729 if (throttled_hierarchy(cfs_rq
))
5732 update_cfs_rq_blocked_load(cfs_rq
, 1);
5735 update_entity_load_avg(se
, 1);
5737 * We pivot on our runnable average having decayed to zero for
5738 * list removal. This generally implies that all our children
5739 * have also been removed (modulo rounding error or bandwidth
5740 * control); however, such cases are rare and we can fix these
5743 * TODO: fix up out-of-order children on enqueue.
5745 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5746 list_del_leaf_cfs_rq(cfs_rq
);
5748 struct rq
*rq
= rq_of(cfs_rq
);
5749 update_rq_runnable_avg(rq
, rq
->nr_running
);
5753 static void update_blocked_averages(int cpu
)
5755 struct rq
*rq
= cpu_rq(cpu
);
5756 struct cfs_rq
*cfs_rq
;
5757 unsigned long flags
;
5759 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5760 update_rq_clock(rq
);
5762 * Iterates the task_group tree in a bottom up fashion, see
5763 * list_add_leaf_cfs_rq() for details.
5765 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5767 * Note: We may want to consider periodically releasing
5768 * rq->lock about these updates so that creating many task
5769 * groups does not result in continually extending hold time.
5771 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5774 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5778 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5779 * This needs to be done in a top-down fashion because the load of a child
5780 * group is a fraction of its parents load.
5782 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5784 struct rq
*rq
= rq_of(cfs_rq
);
5785 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5786 unsigned long now
= jiffies
;
5789 if (cfs_rq
->last_h_load_update
== now
)
5792 cfs_rq
->h_load_next
= NULL
;
5793 for_each_sched_entity(se
) {
5794 cfs_rq
= cfs_rq_of(se
);
5795 cfs_rq
->h_load_next
= se
;
5796 if (cfs_rq
->last_h_load_update
== now
)
5801 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5802 cfs_rq
->last_h_load_update
= now
;
5805 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5806 load
= cfs_rq
->h_load
;
5807 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5808 cfs_rq
->runnable_load_avg
+ 1);
5809 cfs_rq
= group_cfs_rq(se
);
5810 cfs_rq
->h_load
= load
;
5811 cfs_rq
->last_h_load_update
= now
;
5815 static unsigned long task_h_load(struct task_struct
*p
)
5817 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5819 update_cfs_rq_h_load(cfs_rq
);
5820 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5821 cfs_rq
->runnable_load_avg
+ 1);
5824 static inline void update_blocked_averages(int cpu
)
5828 static unsigned long task_h_load(struct task_struct
*p
)
5830 return p
->se
.avg
.load_avg_contrib
;
5834 /********** Helpers for find_busiest_group ************************/
5843 * sg_lb_stats - stats of a sched_group required for load_balancing
5845 struct sg_lb_stats
{
5846 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5847 unsigned long group_load
; /* Total load over the CPUs of the group */
5848 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5849 unsigned long load_per_task
;
5850 unsigned long group_capacity
;
5851 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5852 unsigned int group_capacity_factor
;
5853 unsigned int idle_cpus
;
5854 unsigned int group_weight
;
5855 enum group_type group_type
;
5856 int group_has_free_capacity
;
5857 #ifdef CONFIG_NUMA_BALANCING
5858 unsigned int nr_numa_running
;
5859 unsigned int nr_preferred_running
;
5864 * sd_lb_stats - Structure to store the statistics of a sched_domain
5865 * during load balancing.
5867 struct sd_lb_stats
{
5868 struct sched_group
*busiest
; /* Busiest group in this sd */
5869 struct sched_group
*local
; /* Local group in this sd */
5870 unsigned long total_load
; /* Total load of all groups in sd */
5871 unsigned long total_capacity
; /* Total capacity of all groups in sd */
5872 unsigned long avg_load
; /* Average load across all groups in sd */
5874 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5875 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5878 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5881 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5882 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5883 * We must however clear busiest_stat::avg_load because
5884 * update_sd_pick_busiest() reads this before assignment.
5886 *sds
= (struct sd_lb_stats
){
5890 .total_capacity
= 0UL,
5893 .sum_nr_running
= 0,
5894 .group_type
= group_other
,
5900 * get_sd_load_idx - Obtain the load index for a given sched domain.
5901 * @sd: The sched_domain whose load_idx is to be obtained.
5902 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5904 * Return: The load index.
5906 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5907 enum cpu_idle_type idle
)
5913 load_idx
= sd
->busy_idx
;
5916 case CPU_NEWLY_IDLE
:
5917 load_idx
= sd
->newidle_idx
;
5920 load_idx
= sd
->idle_idx
;
5927 static unsigned long default_scale_capacity(struct sched_domain
*sd
, int cpu
)
5929 return SCHED_CAPACITY_SCALE
;
5932 unsigned long __weak
arch_scale_freq_capacity(struct sched_domain
*sd
, int cpu
)
5934 return default_scale_capacity(sd
, cpu
);
5937 static unsigned long default_scale_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5939 if ((sd
->flags
& SD_SHARE_CPUCAPACITY
) && (sd
->span_weight
> 1))
5940 return sd
->smt_gain
/ sd
->span_weight
;
5942 return SCHED_CAPACITY_SCALE
;
5945 unsigned long __weak
arch_scale_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5947 return default_scale_cpu_capacity(sd
, cpu
);
5950 static unsigned long scale_rt_capacity(int cpu
)
5952 struct rq
*rq
= cpu_rq(cpu
);
5953 u64 total
, available
, age_stamp
, avg
;
5957 * Since we're reading these variables without serialization make sure
5958 * we read them once before doing sanity checks on them.
5960 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5961 avg
= ACCESS_ONCE(rq
->rt_avg
);
5962 delta
= __rq_clock_broken(rq
) - age_stamp
;
5964 if (unlikely(delta
< 0))
5967 total
= sched_avg_period() + delta
;
5969 if (unlikely(total
< avg
)) {
5970 /* Ensures that capacity won't end up being negative */
5973 available
= total
- avg
;
5976 if (unlikely((s64
)total
< SCHED_CAPACITY_SCALE
))
5977 total
= SCHED_CAPACITY_SCALE
;
5979 total
>>= SCHED_CAPACITY_SHIFT
;
5981 return div_u64(available
, total
);
5984 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5986 unsigned long capacity
= SCHED_CAPACITY_SCALE
;
5987 struct sched_group
*sdg
= sd
->groups
;
5989 if (sched_feat(ARCH_CAPACITY
))
5990 capacity
*= arch_scale_cpu_capacity(sd
, cpu
);
5992 capacity
*= default_scale_cpu_capacity(sd
, cpu
);
5994 capacity
>>= SCHED_CAPACITY_SHIFT
;
5996 sdg
->sgc
->capacity_orig
= capacity
;
5998 if (sched_feat(ARCH_CAPACITY
))
5999 capacity
*= arch_scale_freq_capacity(sd
, cpu
);
6001 capacity
*= default_scale_capacity(sd
, cpu
);
6003 capacity
>>= SCHED_CAPACITY_SHIFT
;
6005 capacity
*= scale_rt_capacity(cpu
);
6006 capacity
>>= SCHED_CAPACITY_SHIFT
;
6011 cpu_rq(cpu
)->cpu_capacity
= capacity
;
6012 sdg
->sgc
->capacity
= capacity
;
6015 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
6017 struct sched_domain
*child
= sd
->child
;
6018 struct sched_group
*group
, *sdg
= sd
->groups
;
6019 unsigned long capacity
, capacity_orig
;
6020 unsigned long interval
;
6022 interval
= msecs_to_jiffies(sd
->balance_interval
);
6023 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6024 sdg
->sgc
->next_update
= jiffies
+ interval
;
6027 update_cpu_capacity(sd
, cpu
);
6031 capacity_orig
= capacity
= 0;
6033 if (child
->flags
& SD_OVERLAP
) {
6035 * SD_OVERLAP domains cannot assume that child groups
6036 * span the current group.
6039 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
6040 struct sched_group_capacity
*sgc
;
6041 struct rq
*rq
= cpu_rq(cpu
);
6044 * build_sched_domains() -> init_sched_groups_capacity()
6045 * gets here before we've attached the domains to the
6048 * Use capacity_of(), which is set irrespective of domains
6049 * in update_cpu_capacity().
6051 * This avoids capacity/capacity_orig from being 0 and
6052 * causing divide-by-zero issues on boot.
6054 * Runtime updates will correct capacity_orig.
6056 if (unlikely(!rq
->sd
)) {
6057 capacity_orig
+= capacity_of(cpu
);
6058 capacity
+= capacity_of(cpu
);
6062 sgc
= rq
->sd
->groups
->sgc
;
6063 capacity_orig
+= sgc
->capacity_orig
;
6064 capacity
+= sgc
->capacity
;
6068 * !SD_OVERLAP domains can assume that child groups
6069 * span the current group.
6072 group
= child
->groups
;
6074 capacity_orig
+= group
->sgc
->capacity_orig
;
6075 capacity
+= group
->sgc
->capacity
;
6076 group
= group
->next
;
6077 } while (group
!= child
->groups
);
6080 sdg
->sgc
->capacity_orig
= capacity_orig
;
6081 sdg
->sgc
->capacity
= capacity
;
6085 * Try and fix up capacity for tiny siblings, this is needed when
6086 * things like SD_ASYM_PACKING need f_b_g to select another sibling
6087 * which on its own isn't powerful enough.
6089 * See update_sd_pick_busiest() and check_asym_packing().
6092 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
6095 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
6097 if (!(sd
->flags
& SD_SHARE_CPUCAPACITY
))
6101 * If ~90% of the cpu_capacity is still there, we're good.
6103 if (group
->sgc
->capacity
* 32 > group
->sgc
->capacity_orig
* 29)
6110 * Group imbalance indicates (and tries to solve) the problem where balancing
6111 * groups is inadequate due to tsk_cpus_allowed() constraints.
6113 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6114 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6117 * { 0 1 2 3 } { 4 5 6 7 }
6120 * If we were to balance group-wise we'd place two tasks in the first group and
6121 * two tasks in the second group. Clearly this is undesired as it will overload
6122 * cpu 3 and leave one of the cpus in the second group unused.
6124 * The current solution to this issue is detecting the skew in the first group
6125 * by noticing the lower domain failed to reach balance and had difficulty
6126 * moving tasks due to affinity constraints.
6128 * When this is so detected; this group becomes a candidate for busiest; see
6129 * update_sd_pick_busiest(). And calculate_imbalance() and
6130 * find_busiest_group() avoid some of the usual balance conditions to allow it
6131 * to create an effective group imbalance.
6133 * This is a somewhat tricky proposition since the next run might not find the
6134 * group imbalance and decide the groups need to be balanced again. A most
6135 * subtle and fragile situation.
6138 static inline int sg_imbalanced(struct sched_group
*group
)
6140 return group
->sgc
->imbalance
;
6144 * Compute the group capacity factor.
6146 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
6147 * first dividing out the smt factor and computing the actual number of cores
6148 * and limit unit capacity with that.
6150 static inline int sg_capacity_factor(struct lb_env
*env
, struct sched_group
*group
)
6152 unsigned int capacity_factor
, smt
, cpus
;
6153 unsigned int capacity
, capacity_orig
;
6155 capacity
= group
->sgc
->capacity
;
6156 capacity_orig
= group
->sgc
->capacity_orig
;
6157 cpus
= group
->group_weight
;
6159 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
6160 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, capacity_orig
);
6161 capacity_factor
= cpus
/ smt
; /* cores */
6163 capacity_factor
= min_t(unsigned,
6164 capacity_factor
, DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
));
6165 if (!capacity_factor
)
6166 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6168 return capacity_factor
;
6171 static enum group_type
6172 group_classify(struct sched_group
*group
, struct sg_lb_stats
*sgs
)
6174 if (sgs
->sum_nr_running
> sgs
->group_capacity_factor
)
6175 return group_overloaded
;
6177 if (sg_imbalanced(group
))
6178 return group_imbalanced
;
6184 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6185 * @env: The load balancing environment.
6186 * @group: sched_group whose statistics are to be updated.
6187 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6188 * @local_group: Does group contain this_cpu.
6189 * @sgs: variable to hold the statistics for this group.
6190 * @overload: Indicate more than one runnable task for any CPU.
6192 static inline void update_sg_lb_stats(struct lb_env
*env
,
6193 struct sched_group
*group
, int load_idx
,
6194 int local_group
, struct sg_lb_stats
*sgs
,
6200 memset(sgs
, 0, sizeof(*sgs
));
6202 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6203 struct rq
*rq
= cpu_rq(i
);
6205 /* Bias balancing toward cpus of our domain */
6207 load
= target_load(i
, load_idx
);
6209 load
= source_load(i
, load_idx
);
6211 sgs
->group_load
+= load
;
6212 sgs
->sum_nr_running
+= rq
->cfs
.h_nr_running
;
6214 if (rq
->nr_running
> 1)
6217 #ifdef CONFIG_NUMA_BALANCING
6218 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
6219 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
6221 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
6226 /* Adjust by relative CPU capacity of the group */
6227 sgs
->group_capacity
= group
->sgc
->capacity
;
6228 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
6230 if (sgs
->sum_nr_running
)
6231 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
6233 sgs
->group_weight
= group
->group_weight
;
6234 sgs
->group_capacity_factor
= sg_capacity_factor(env
, group
);
6235 sgs
->group_type
= group_classify(group
, sgs
);
6237 if (sgs
->group_capacity_factor
> sgs
->sum_nr_running
)
6238 sgs
->group_has_free_capacity
= 1;
6242 * update_sd_pick_busiest - return 1 on busiest group
6243 * @env: The load balancing environment.
6244 * @sds: sched_domain statistics
6245 * @sg: sched_group candidate to be checked for being the busiest
6246 * @sgs: sched_group statistics
6248 * Determine if @sg is a busier group than the previously selected
6251 * Return: %true if @sg is a busier group than the previously selected
6252 * busiest group. %false otherwise.
6254 static bool update_sd_pick_busiest(struct lb_env
*env
,
6255 struct sd_lb_stats
*sds
,
6256 struct sched_group
*sg
,
6257 struct sg_lb_stats
*sgs
)
6259 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
6261 if (sgs
->group_type
> busiest
->group_type
)
6264 if (sgs
->group_type
< busiest
->group_type
)
6267 if (sgs
->avg_load
<= busiest
->avg_load
)
6270 /* This is the busiest node in its class. */
6271 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6275 * ASYM_PACKING needs to move all the work to the lowest
6276 * numbered CPUs in the group, therefore mark all groups
6277 * higher than ourself as busy.
6279 if (sgs
->sum_nr_running
&& env
->dst_cpu
< group_first_cpu(sg
)) {
6283 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
6290 #ifdef CONFIG_NUMA_BALANCING
6291 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6293 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
6295 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
6300 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6302 if (rq
->nr_running
> rq
->nr_numa_running
)
6304 if (rq
->nr_running
> rq
->nr_preferred_running
)
6309 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6314 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6318 #endif /* CONFIG_NUMA_BALANCING */
6321 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6322 * @env: The load balancing environment.
6323 * @sds: variable to hold the statistics for this sched_domain.
6325 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6327 struct sched_domain
*child
= env
->sd
->child
;
6328 struct sched_group
*sg
= env
->sd
->groups
;
6329 struct sg_lb_stats tmp_sgs
;
6330 int load_idx
, prefer_sibling
= 0;
6331 bool overload
= false;
6333 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
6336 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
6339 struct sg_lb_stats
*sgs
= &tmp_sgs
;
6342 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
6345 sgs
= &sds
->local_stat
;
6347 if (env
->idle
!= CPU_NEWLY_IDLE
||
6348 time_after_eq(jiffies
, sg
->sgc
->next_update
))
6349 update_group_capacity(env
->sd
, env
->dst_cpu
);
6352 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
,
6359 * In case the child domain prefers tasks go to siblings
6360 * first, lower the sg capacity factor to one so that we'll try
6361 * and move all the excess tasks away. We lower the capacity
6362 * of a group only if the local group has the capacity to fit
6363 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6364 * extra check prevents the case where you always pull from the
6365 * heaviest group when it is already under-utilized (possible
6366 * with a large weight task outweighs the tasks on the system).
6368 if (prefer_sibling
&& sds
->local
&&
6369 sds
->local_stat
.group_has_free_capacity
) {
6370 sgs
->group_capacity_factor
= min(sgs
->group_capacity_factor
, 1U);
6371 sgs
->group_type
= group_classify(sg
, sgs
);
6374 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
6376 sds
->busiest_stat
= *sgs
;
6380 /* Now, start updating sd_lb_stats */
6381 sds
->total_load
+= sgs
->group_load
;
6382 sds
->total_capacity
+= sgs
->group_capacity
;
6385 } while (sg
!= env
->sd
->groups
);
6387 if (env
->sd
->flags
& SD_NUMA
)
6388 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
6390 if (!env
->sd
->parent
) {
6391 /* update overload indicator if we are at root domain */
6392 if (env
->dst_rq
->rd
->overload
!= overload
)
6393 env
->dst_rq
->rd
->overload
= overload
;
6399 * check_asym_packing - Check to see if the group is packed into the
6402 * This is primarily intended to used at the sibling level. Some
6403 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6404 * case of POWER7, it can move to lower SMT modes only when higher
6405 * threads are idle. When in lower SMT modes, the threads will
6406 * perform better since they share less core resources. Hence when we
6407 * have idle threads, we want them to be the higher ones.
6409 * This packing function is run on idle threads. It checks to see if
6410 * the busiest CPU in this domain (core in the P7 case) has a higher
6411 * CPU number than the packing function is being run on. Here we are
6412 * assuming lower CPU number will be equivalent to lower a SMT thread
6415 * Return: 1 when packing is required and a task should be moved to
6416 * this CPU. The amount of the imbalance is returned in *imbalance.
6418 * @env: The load balancing environment.
6419 * @sds: Statistics of the sched_domain which is to be packed
6421 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6425 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6431 busiest_cpu
= group_first_cpu(sds
->busiest
);
6432 if (env
->dst_cpu
> busiest_cpu
)
6435 env
->imbalance
= DIV_ROUND_CLOSEST(
6436 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
6437 SCHED_CAPACITY_SCALE
);
6443 * fix_small_imbalance - Calculate the minor imbalance that exists
6444 * amongst the groups of a sched_domain, during
6446 * @env: The load balancing environment.
6447 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6450 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6452 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
6453 unsigned int imbn
= 2;
6454 unsigned long scaled_busy_load_per_task
;
6455 struct sg_lb_stats
*local
, *busiest
;
6457 local
= &sds
->local_stat
;
6458 busiest
= &sds
->busiest_stat
;
6460 if (!local
->sum_nr_running
)
6461 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
6462 else if (busiest
->load_per_task
> local
->load_per_task
)
6465 scaled_busy_load_per_task
=
6466 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6467 busiest
->group_capacity
;
6469 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
6470 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
6471 env
->imbalance
= busiest
->load_per_task
;
6476 * OK, we don't have enough imbalance to justify moving tasks,
6477 * however we may be able to increase total CPU capacity used by
6481 capa_now
+= busiest
->group_capacity
*
6482 min(busiest
->load_per_task
, busiest
->avg_load
);
6483 capa_now
+= local
->group_capacity
*
6484 min(local
->load_per_task
, local
->avg_load
);
6485 capa_now
/= SCHED_CAPACITY_SCALE
;
6487 /* Amount of load we'd subtract */
6488 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
6489 capa_move
+= busiest
->group_capacity
*
6490 min(busiest
->load_per_task
,
6491 busiest
->avg_load
- scaled_busy_load_per_task
);
6494 /* Amount of load we'd add */
6495 if (busiest
->avg_load
* busiest
->group_capacity
<
6496 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
6497 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
6498 local
->group_capacity
;
6500 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6501 local
->group_capacity
;
6503 capa_move
+= local
->group_capacity
*
6504 min(local
->load_per_task
, local
->avg_load
+ tmp
);
6505 capa_move
/= SCHED_CAPACITY_SCALE
;
6507 /* Move if we gain throughput */
6508 if (capa_move
> capa_now
)
6509 env
->imbalance
= busiest
->load_per_task
;
6513 * calculate_imbalance - Calculate the amount of imbalance present within the
6514 * groups of a given sched_domain during load balance.
6515 * @env: load balance environment
6516 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6518 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6520 unsigned long max_pull
, load_above_capacity
= ~0UL;
6521 struct sg_lb_stats
*local
, *busiest
;
6523 local
= &sds
->local_stat
;
6524 busiest
= &sds
->busiest_stat
;
6526 if (busiest
->group_type
== group_imbalanced
) {
6528 * In the group_imb case we cannot rely on group-wide averages
6529 * to ensure cpu-load equilibrium, look at wider averages. XXX
6531 busiest
->load_per_task
=
6532 min(busiest
->load_per_task
, sds
->avg_load
);
6536 * In the presence of smp nice balancing, certain scenarios can have
6537 * max load less than avg load(as we skip the groups at or below
6538 * its cpu_capacity, while calculating max_load..)
6540 if (busiest
->avg_load
<= sds
->avg_load
||
6541 local
->avg_load
>= sds
->avg_load
) {
6543 return fix_small_imbalance(env
, sds
);
6547 * If there aren't any idle cpus, avoid creating some.
6549 if (busiest
->group_type
== group_overloaded
&&
6550 local
->group_type
== group_overloaded
) {
6551 load_above_capacity
=
6552 (busiest
->sum_nr_running
- busiest
->group_capacity_factor
);
6554 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_CAPACITY_SCALE
);
6555 load_above_capacity
/= busiest
->group_capacity
;
6559 * We're trying to get all the cpus to the average_load, so we don't
6560 * want to push ourselves above the average load, nor do we wish to
6561 * reduce the max loaded cpu below the average load. At the same time,
6562 * we also don't want to reduce the group load below the group capacity
6563 * (so that we can implement power-savings policies etc). Thus we look
6564 * for the minimum possible imbalance.
6566 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
6568 /* How much load to actually move to equalise the imbalance */
6569 env
->imbalance
= min(
6570 max_pull
* busiest
->group_capacity
,
6571 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
6572 ) / SCHED_CAPACITY_SCALE
;
6575 * if *imbalance is less than the average load per runnable task
6576 * there is no guarantee that any tasks will be moved so we'll have
6577 * a think about bumping its value to force at least one task to be
6580 if (env
->imbalance
< busiest
->load_per_task
)
6581 return fix_small_imbalance(env
, sds
);
6584 /******* find_busiest_group() helpers end here *********************/
6587 * find_busiest_group - Returns the busiest group within the sched_domain
6588 * if there is an imbalance. If there isn't an imbalance, and
6589 * the user has opted for power-savings, it returns a group whose
6590 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6591 * such a group exists.
6593 * Also calculates the amount of weighted load which should be moved
6594 * to restore balance.
6596 * @env: The load balancing environment.
6598 * Return: - The busiest group if imbalance exists.
6599 * - If no imbalance and user has opted for power-savings balance,
6600 * return the least loaded group whose CPUs can be
6601 * put to idle by rebalancing its tasks onto our group.
6603 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
6605 struct sg_lb_stats
*local
, *busiest
;
6606 struct sd_lb_stats sds
;
6608 init_sd_lb_stats(&sds
);
6611 * Compute the various statistics relavent for load balancing at
6614 update_sd_lb_stats(env
, &sds
);
6615 local
= &sds
.local_stat
;
6616 busiest
= &sds
.busiest_stat
;
6618 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
6619 check_asym_packing(env
, &sds
))
6622 /* There is no busy sibling group to pull tasks from */
6623 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
6626 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
6627 / sds
.total_capacity
;
6630 * If the busiest group is imbalanced the below checks don't
6631 * work because they assume all things are equal, which typically
6632 * isn't true due to cpus_allowed constraints and the like.
6634 if (busiest
->group_type
== group_imbalanced
)
6637 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6638 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_free_capacity
&&
6639 !busiest
->group_has_free_capacity
)
6643 * If the local group is busier than the selected busiest group
6644 * don't try and pull any tasks.
6646 if (local
->avg_load
>= busiest
->avg_load
)
6650 * Don't pull any tasks if this group is already above the domain
6653 if (local
->avg_load
>= sds
.avg_load
)
6656 if (env
->idle
== CPU_IDLE
) {
6658 * This cpu is idle. If the busiest group is not overloaded
6659 * and there is no imbalance between this and busiest group
6660 * wrt idle cpus, it is balanced. The imbalance becomes
6661 * significant if the diff is greater than 1 otherwise we
6662 * might end up to just move the imbalance on another group
6664 if ((busiest
->group_type
!= group_overloaded
) &&
6665 (local
->idle_cpus
<= (busiest
->idle_cpus
+ 1)))
6669 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6670 * imbalance_pct to be conservative.
6672 if (100 * busiest
->avg_load
<=
6673 env
->sd
->imbalance_pct
* local
->avg_load
)
6678 /* Looks like there is an imbalance. Compute it */
6679 calculate_imbalance(env
, &sds
);
6688 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6690 static struct rq
*find_busiest_queue(struct lb_env
*env
,
6691 struct sched_group
*group
)
6693 struct rq
*busiest
= NULL
, *rq
;
6694 unsigned long busiest_load
= 0, busiest_capacity
= 1;
6697 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6698 unsigned long capacity
, capacity_factor
, wl
;
6702 rt
= fbq_classify_rq(rq
);
6705 * We classify groups/runqueues into three groups:
6706 * - regular: there are !numa tasks
6707 * - remote: there are numa tasks that run on the 'wrong' node
6708 * - all: there is no distinction
6710 * In order to avoid migrating ideally placed numa tasks,
6711 * ignore those when there's better options.
6713 * If we ignore the actual busiest queue to migrate another
6714 * task, the next balance pass can still reduce the busiest
6715 * queue by moving tasks around inside the node.
6717 * If we cannot move enough load due to this classification
6718 * the next pass will adjust the group classification and
6719 * allow migration of more tasks.
6721 * Both cases only affect the total convergence complexity.
6723 if (rt
> env
->fbq_type
)
6726 capacity
= capacity_of(i
);
6727 capacity_factor
= DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
);
6728 if (!capacity_factor
)
6729 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6731 wl
= weighted_cpuload(i
);
6734 * When comparing with imbalance, use weighted_cpuload()
6735 * which is not scaled with the cpu capacity.
6737 if (capacity_factor
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6741 * For the load comparisons with the other cpu's, consider
6742 * the weighted_cpuload() scaled with the cpu capacity, so
6743 * that the load can be moved away from the cpu that is
6744 * potentially running at a lower capacity.
6746 * Thus we're looking for max(wl_i / capacity_i), crosswise
6747 * multiplication to rid ourselves of the division works out
6748 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6749 * our previous maximum.
6751 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
6753 busiest_capacity
= capacity
;
6762 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6763 * so long as it is large enough.
6765 #define MAX_PINNED_INTERVAL 512
6767 /* Working cpumask for load_balance and load_balance_newidle. */
6768 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6770 static int need_active_balance(struct lb_env
*env
)
6772 struct sched_domain
*sd
= env
->sd
;
6774 if (env
->idle
== CPU_NEWLY_IDLE
) {
6777 * ASYM_PACKING needs to force migrate tasks from busy but
6778 * higher numbered CPUs in order to pack all tasks in the
6779 * lowest numbered CPUs.
6781 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6785 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6788 static int active_load_balance_cpu_stop(void *data
);
6790 static int should_we_balance(struct lb_env
*env
)
6792 struct sched_group
*sg
= env
->sd
->groups
;
6793 struct cpumask
*sg_cpus
, *sg_mask
;
6794 int cpu
, balance_cpu
= -1;
6797 * In the newly idle case, we will allow all the cpu's
6798 * to do the newly idle load balance.
6800 if (env
->idle
== CPU_NEWLY_IDLE
)
6803 sg_cpus
= sched_group_cpus(sg
);
6804 sg_mask
= sched_group_mask(sg
);
6805 /* Try to find first idle cpu */
6806 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6807 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6814 if (balance_cpu
== -1)
6815 balance_cpu
= group_balance_cpu(sg
);
6818 * First idle cpu or the first cpu(busiest) in this sched group
6819 * is eligible for doing load balancing at this and above domains.
6821 return balance_cpu
== env
->dst_cpu
;
6825 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6826 * tasks if there is an imbalance.
6828 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6829 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6830 int *continue_balancing
)
6832 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6833 struct sched_domain
*sd_parent
= sd
->parent
;
6834 struct sched_group
*group
;
6836 unsigned long flags
;
6837 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
6839 struct lb_env env
= {
6841 .dst_cpu
= this_cpu
,
6843 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6845 .loop_break
= sched_nr_migrate_break
,
6848 .tasks
= LIST_HEAD_INIT(env
.tasks
),
6852 * For NEWLY_IDLE load_balancing, we don't need to consider
6853 * other cpus in our group
6855 if (idle
== CPU_NEWLY_IDLE
)
6856 env
.dst_grpmask
= NULL
;
6858 cpumask_copy(cpus
, cpu_active_mask
);
6860 schedstat_inc(sd
, lb_count
[idle
]);
6863 if (!should_we_balance(&env
)) {
6864 *continue_balancing
= 0;
6868 group
= find_busiest_group(&env
);
6870 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6874 busiest
= find_busiest_queue(&env
, group
);
6876 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6880 BUG_ON(busiest
== env
.dst_rq
);
6882 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6885 if (busiest
->nr_running
> 1) {
6887 * Attempt to move tasks. If find_busiest_group has found
6888 * an imbalance but busiest->nr_running <= 1, the group is
6889 * still unbalanced. ld_moved simply stays zero, so it is
6890 * correctly treated as an imbalance.
6892 env
.flags
|= LBF_ALL_PINNED
;
6893 env
.src_cpu
= busiest
->cpu
;
6894 env
.src_rq
= busiest
;
6895 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6898 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6901 * cur_ld_moved - load moved in current iteration
6902 * ld_moved - cumulative load moved across iterations
6904 cur_ld_moved
= detach_tasks(&env
);
6907 * We've detached some tasks from busiest_rq. Every
6908 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6909 * unlock busiest->lock, and we are able to be sure
6910 * that nobody can manipulate the tasks in parallel.
6911 * See task_rq_lock() family for the details.
6914 raw_spin_unlock(&busiest
->lock
);
6918 ld_moved
+= cur_ld_moved
;
6921 local_irq_restore(flags
);
6923 if (env
.flags
& LBF_NEED_BREAK
) {
6924 env
.flags
&= ~LBF_NEED_BREAK
;
6929 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6930 * us and move them to an alternate dst_cpu in our sched_group
6931 * where they can run. The upper limit on how many times we
6932 * iterate on same src_cpu is dependent on number of cpus in our
6935 * This changes load balance semantics a bit on who can move
6936 * load to a given_cpu. In addition to the given_cpu itself
6937 * (or a ilb_cpu acting on its behalf where given_cpu is
6938 * nohz-idle), we now have balance_cpu in a position to move
6939 * load to given_cpu. In rare situations, this may cause
6940 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6941 * _independently_ and at _same_ time to move some load to
6942 * given_cpu) causing exceess load to be moved to given_cpu.
6943 * This however should not happen so much in practice and
6944 * moreover subsequent load balance cycles should correct the
6945 * excess load moved.
6947 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6949 /* Prevent to re-select dst_cpu via env's cpus */
6950 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6952 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6953 env
.dst_cpu
= env
.new_dst_cpu
;
6954 env
.flags
&= ~LBF_DST_PINNED
;
6956 env
.loop_break
= sched_nr_migrate_break
;
6959 * Go back to "more_balance" rather than "redo" since we
6960 * need to continue with same src_cpu.
6966 * We failed to reach balance because of affinity.
6969 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
6971 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
6972 *group_imbalance
= 1;
6975 /* All tasks on this runqueue were pinned by CPU affinity */
6976 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6977 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6978 if (!cpumask_empty(cpus
)) {
6980 env
.loop_break
= sched_nr_migrate_break
;
6983 goto out_all_pinned
;
6988 schedstat_inc(sd
, lb_failed
[idle
]);
6990 * Increment the failure counter only on periodic balance.
6991 * We do not want newidle balance, which can be very
6992 * frequent, pollute the failure counter causing
6993 * excessive cache_hot migrations and active balances.
6995 if (idle
!= CPU_NEWLY_IDLE
)
6996 sd
->nr_balance_failed
++;
6998 if (need_active_balance(&env
)) {
6999 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
7001 /* don't kick the active_load_balance_cpu_stop,
7002 * if the curr task on busiest cpu can't be
7005 if (!cpumask_test_cpu(this_cpu
,
7006 tsk_cpus_allowed(busiest
->curr
))) {
7007 raw_spin_unlock_irqrestore(&busiest
->lock
,
7009 env
.flags
|= LBF_ALL_PINNED
;
7010 goto out_one_pinned
;
7014 * ->active_balance synchronizes accesses to
7015 * ->active_balance_work. Once set, it's cleared
7016 * only after active load balance is finished.
7018 if (!busiest
->active_balance
) {
7019 busiest
->active_balance
= 1;
7020 busiest
->push_cpu
= this_cpu
;
7023 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
7025 if (active_balance
) {
7026 stop_one_cpu_nowait(cpu_of(busiest
),
7027 active_load_balance_cpu_stop
, busiest
,
7028 &busiest
->active_balance_work
);
7032 * We've kicked active balancing, reset the failure
7035 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
7038 sd
->nr_balance_failed
= 0;
7040 if (likely(!active_balance
)) {
7041 /* We were unbalanced, so reset the balancing interval */
7042 sd
->balance_interval
= sd
->min_interval
;
7045 * If we've begun active balancing, start to back off. This
7046 * case may not be covered by the all_pinned logic if there
7047 * is only 1 task on the busy runqueue (because we don't call
7050 if (sd
->balance_interval
< sd
->max_interval
)
7051 sd
->balance_interval
*= 2;
7058 * We reach balance although we may have faced some affinity
7059 * constraints. Clear the imbalance flag if it was set.
7062 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
7064 if (*group_imbalance
)
7065 *group_imbalance
= 0;
7070 * We reach balance because all tasks are pinned at this level so
7071 * we can't migrate them. Let the imbalance flag set so parent level
7072 * can try to migrate them.
7074 schedstat_inc(sd
, lb_balanced
[idle
]);
7076 sd
->nr_balance_failed
= 0;
7079 /* tune up the balancing interval */
7080 if (((env
.flags
& LBF_ALL_PINNED
) &&
7081 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
7082 (sd
->balance_interval
< sd
->max_interval
))
7083 sd
->balance_interval
*= 2;
7090 static inline unsigned long
7091 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
7093 unsigned long interval
= sd
->balance_interval
;
7096 interval
*= sd
->busy_factor
;
7098 /* scale ms to jiffies */
7099 interval
= msecs_to_jiffies(interval
);
7100 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
7106 update_next_balance(struct sched_domain
*sd
, int cpu_busy
, unsigned long *next_balance
)
7108 unsigned long interval
, next
;
7110 interval
= get_sd_balance_interval(sd
, cpu_busy
);
7111 next
= sd
->last_balance
+ interval
;
7113 if (time_after(*next_balance
, next
))
7114 *next_balance
= next
;
7118 * idle_balance is called by schedule() if this_cpu is about to become
7119 * idle. Attempts to pull tasks from other CPUs.
7121 static int idle_balance(struct rq
*this_rq
)
7123 unsigned long next_balance
= jiffies
+ HZ
;
7124 int this_cpu
= this_rq
->cpu
;
7125 struct sched_domain
*sd
;
7126 int pulled_task
= 0;
7129 idle_enter_fair(this_rq
);
7132 * We must set idle_stamp _before_ calling idle_balance(), such that we
7133 * measure the duration of idle_balance() as idle time.
7135 this_rq
->idle_stamp
= rq_clock(this_rq
);
7137 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
7138 !this_rq
->rd
->overload
) {
7140 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
7142 update_next_balance(sd
, 0, &next_balance
);
7149 * Drop the rq->lock, but keep IRQ/preempt disabled.
7151 raw_spin_unlock(&this_rq
->lock
);
7153 update_blocked_averages(this_cpu
);
7155 for_each_domain(this_cpu
, sd
) {
7156 int continue_balancing
= 1;
7157 u64 t0
, domain_cost
;
7159 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7162 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
7163 update_next_balance(sd
, 0, &next_balance
);
7167 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
7168 t0
= sched_clock_cpu(this_cpu
);
7170 pulled_task
= load_balance(this_cpu
, this_rq
,
7172 &continue_balancing
);
7174 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
7175 if (domain_cost
> sd
->max_newidle_lb_cost
)
7176 sd
->max_newidle_lb_cost
= domain_cost
;
7178 curr_cost
+= domain_cost
;
7181 update_next_balance(sd
, 0, &next_balance
);
7184 * Stop searching for tasks to pull if there are
7185 * now runnable tasks on this rq.
7187 if (pulled_task
|| this_rq
->nr_running
> 0)
7192 raw_spin_lock(&this_rq
->lock
);
7194 if (curr_cost
> this_rq
->max_idle_balance_cost
)
7195 this_rq
->max_idle_balance_cost
= curr_cost
;
7198 * While browsing the domains, we released the rq lock, a task could
7199 * have been enqueued in the meantime. Since we're not going idle,
7200 * pretend we pulled a task.
7202 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
7206 /* Move the next balance forward */
7207 if (time_after(this_rq
->next_balance
, next_balance
))
7208 this_rq
->next_balance
= next_balance
;
7210 /* Is there a task of a high priority class? */
7211 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
7215 idle_exit_fair(this_rq
);
7216 this_rq
->idle_stamp
= 0;
7223 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7224 * running tasks off the busiest CPU onto idle CPUs. It requires at
7225 * least 1 task to be running on each physical CPU where possible, and
7226 * avoids physical / logical imbalances.
7228 static int active_load_balance_cpu_stop(void *data
)
7230 struct rq
*busiest_rq
= data
;
7231 int busiest_cpu
= cpu_of(busiest_rq
);
7232 int target_cpu
= busiest_rq
->push_cpu
;
7233 struct rq
*target_rq
= cpu_rq(target_cpu
);
7234 struct sched_domain
*sd
;
7235 struct task_struct
*p
= NULL
;
7237 raw_spin_lock_irq(&busiest_rq
->lock
);
7239 /* make sure the requested cpu hasn't gone down in the meantime */
7240 if (unlikely(busiest_cpu
!= smp_processor_id() ||
7241 !busiest_rq
->active_balance
))
7244 /* Is there any task to move? */
7245 if (busiest_rq
->nr_running
<= 1)
7249 * This condition is "impossible", if it occurs
7250 * we need to fix it. Originally reported by
7251 * Bjorn Helgaas on a 128-cpu setup.
7253 BUG_ON(busiest_rq
== target_rq
);
7255 /* Search for an sd spanning us and the target CPU. */
7257 for_each_domain(target_cpu
, sd
) {
7258 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
7259 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
7264 struct lb_env env
= {
7266 .dst_cpu
= target_cpu
,
7267 .dst_rq
= target_rq
,
7268 .src_cpu
= busiest_rq
->cpu
,
7269 .src_rq
= busiest_rq
,
7273 schedstat_inc(sd
, alb_count
);
7275 p
= detach_one_task(&env
);
7277 schedstat_inc(sd
, alb_pushed
);
7279 schedstat_inc(sd
, alb_failed
);
7283 busiest_rq
->active_balance
= 0;
7284 raw_spin_unlock(&busiest_rq
->lock
);
7287 attach_one_task(target_rq
, p
);
7294 static inline int on_null_domain(struct rq
*rq
)
7296 return unlikely(!rcu_dereference_sched(rq
->sd
));
7299 #ifdef CONFIG_NO_HZ_COMMON
7301 * idle load balancing details
7302 * - When one of the busy CPUs notice that there may be an idle rebalancing
7303 * needed, they will kick the idle load balancer, which then does idle
7304 * load balancing for all the idle CPUs.
7307 cpumask_var_t idle_cpus_mask
;
7309 unsigned long next_balance
; /* in jiffy units */
7310 } nohz ____cacheline_aligned
;
7312 static inline int find_new_ilb(void)
7314 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
7316 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
7323 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7324 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7325 * CPU (if there is one).
7327 static void nohz_balancer_kick(void)
7331 nohz
.next_balance
++;
7333 ilb_cpu
= find_new_ilb();
7335 if (ilb_cpu
>= nr_cpu_ids
)
7338 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
7341 * Use smp_send_reschedule() instead of resched_cpu().
7342 * This way we generate a sched IPI on the target cpu which
7343 * is idle. And the softirq performing nohz idle load balance
7344 * will be run before returning from the IPI.
7346 smp_send_reschedule(ilb_cpu
);
7350 static inline void nohz_balance_exit_idle(int cpu
)
7352 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
7354 * Completely isolated CPUs don't ever set, so we must test.
7356 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
7357 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
7358 atomic_dec(&nohz
.nr_cpus
);
7360 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7364 static inline void set_cpu_sd_state_busy(void)
7366 struct sched_domain
*sd
;
7367 int cpu
= smp_processor_id();
7370 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7372 if (!sd
|| !sd
->nohz_idle
)
7376 atomic_inc(&sd
->groups
->sgc
->nr_busy_cpus
);
7381 void set_cpu_sd_state_idle(void)
7383 struct sched_domain
*sd
;
7384 int cpu
= smp_processor_id();
7387 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7389 if (!sd
|| sd
->nohz_idle
)
7393 atomic_dec(&sd
->groups
->sgc
->nr_busy_cpus
);
7399 * This routine will record that the cpu is going idle with tick stopped.
7400 * This info will be used in performing idle load balancing in the future.
7402 void nohz_balance_enter_idle(int cpu
)
7405 * If this cpu is going down, then nothing needs to be done.
7407 if (!cpu_active(cpu
))
7410 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
7414 * If we're a completely isolated CPU, we don't play.
7416 if (on_null_domain(cpu_rq(cpu
)))
7419 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
7420 atomic_inc(&nohz
.nr_cpus
);
7421 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7424 static int sched_ilb_notifier(struct notifier_block
*nfb
,
7425 unsigned long action
, void *hcpu
)
7427 switch (action
& ~CPU_TASKS_FROZEN
) {
7429 nohz_balance_exit_idle(smp_processor_id());
7437 static DEFINE_SPINLOCK(balancing
);
7440 * Scale the max load_balance interval with the number of CPUs in the system.
7441 * This trades load-balance latency on larger machines for less cross talk.
7443 void update_max_interval(void)
7445 max_load_balance_interval
= HZ
*num_online_cpus()/10;
7449 * It checks each scheduling domain to see if it is due to be balanced,
7450 * and initiates a balancing operation if so.
7452 * Balancing parameters are set up in init_sched_domains.
7454 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
7456 int continue_balancing
= 1;
7458 unsigned long interval
;
7459 struct sched_domain
*sd
;
7460 /* Earliest time when we have to do rebalance again */
7461 unsigned long next_balance
= jiffies
+ 60*HZ
;
7462 int update_next_balance
= 0;
7463 int need_serialize
, need_decay
= 0;
7466 update_blocked_averages(cpu
);
7469 for_each_domain(cpu
, sd
) {
7471 * Decay the newidle max times here because this is a regular
7472 * visit to all the domains. Decay ~1% per second.
7474 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
7475 sd
->max_newidle_lb_cost
=
7476 (sd
->max_newidle_lb_cost
* 253) / 256;
7477 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
7480 max_cost
+= sd
->max_newidle_lb_cost
;
7482 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7486 * Stop the load balance at this level. There is another
7487 * CPU in our sched group which is doing load balancing more
7490 if (!continue_balancing
) {
7496 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7498 need_serialize
= sd
->flags
& SD_SERIALIZE
;
7499 if (need_serialize
) {
7500 if (!spin_trylock(&balancing
))
7504 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
7505 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
7507 * The LBF_DST_PINNED logic could have changed
7508 * env->dst_cpu, so we can't know our idle
7509 * state even if we migrated tasks. Update it.
7511 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
7513 sd
->last_balance
= jiffies
;
7514 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7517 spin_unlock(&balancing
);
7519 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
7520 next_balance
= sd
->last_balance
+ interval
;
7521 update_next_balance
= 1;
7526 * Ensure the rq-wide value also decays but keep it at a
7527 * reasonable floor to avoid funnies with rq->avg_idle.
7529 rq
->max_idle_balance_cost
=
7530 max((u64
)sysctl_sched_migration_cost
, max_cost
);
7535 * next_balance will be updated only when there is a need.
7536 * When the cpu is attached to null domain for ex, it will not be
7539 if (likely(update_next_balance
))
7540 rq
->next_balance
= next_balance
;
7543 #ifdef CONFIG_NO_HZ_COMMON
7545 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7546 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7548 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
7550 int this_cpu
= this_rq
->cpu
;
7554 if (idle
!= CPU_IDLE
||
7555 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
7558 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
7559 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
7563 * If this cpu gets work to do, stop the load balancing
7564 * work being done for other cpus. Next load
7565 * balancing owner will pick it up.
7570 rq
= cpu_rq(balance_cpu
);
7573 * If time for next balance is due,
7576 if (time_after_eq(jiffies
, rq
->next_balance
)) {
7577 raw_spin_lock_irq(&rq
->lock
);
7578 update_rq_clock(rq
);
7579 update_idle_cpu_load(rq
);
7580 raw_spin_unlock_irq(&rq
->lock
);
7581 rebalance_domains(rq
, CPU_IDLE
);
7584 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
7585 this_rq
->next_balance
= rq
->next_balance
;
7587 nohz
.next_balance
= this_rq
->next_balance
;
7589 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
7593 * Current heuristic for kicking the idle load balancer in the presence
7594 * of an idle cpu is the system.
7595 * - This rq has more than one task.
7596 * - At any scheduler domain level, this cpu's scheduler group has multiple
7597 * busy cpu's exceeding the group's capacity.
7598 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7599 * domain span are idle.
7601 static inline int nohz_kick_needed(struct rq
*rq
)
7603 unsigned long now
= jiffies
;
7604 struct sched_domain
*sd
;
7605 struct sched_group_capacity
*sgc
;
7606 int nr_busy
, cpu
= rq
->cpu
;
7608 if (unlikely(rq
->idle_balance
))
7612 * We may be recently in ticked or tickless idle mode. At the first
7613 * busy tick after returning from idle, we will update the busy stats.
7615 set_cpu_sd_state_busy();
7616 nohz_balance_exit_idle(cpu
);
7619 * None are in tickless mode and hence no need for NOHZ idle load
7622 if (likely(!atomic_read(&nohz
.nr_cpus
)))
7625 if (time_before(now
, nohz
.next_balance
))
7628 if (rq
->nr_running
>= 2)
7632 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7635 sgc
= sd
->groups
->sgc
;
7636 nr_busy
= atomic_read(&sgc
->nr_busy_cpus
);
7639 goto need_kick_unlock
;
7642 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
7644 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
7645 sched_domain_span(sd
)) < cpu
))
7646 goto need_kick_unlock
;
7657 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
7661 * run_rebalance_domains is triggered when needed from the scheduler tick.
7662 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7664 static void run_rebalance_domains(struct softirq_action
*h
)
7666 struct rq
*this_rq
= this_rq();
7667 enum cpu_idle_type idle
= this_rq
->idle_balance
?
7668 CPU_IDLE
: CPU_NOT_IDLE
;
7670 rebalance_domains(this_rq
, idle
);
7673 * If this cpu has a pending nohz_balance_kick, then do the
7674 * balancing on behalf of the other idle cpus whose ticks are
7677 nohz_idle_balance(this_rq
, idle
);
7681 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7683 void trigger_load_balance(struct rq
*rq
)
7685 /* Don't need to rebalance while attached to NULL domain */
7686 if (unlikely(on_null_domain(rq
)))
7689 if (time_after_eq(jiffies
, rq
->next_balance
))
7690 raise_softirq(SCHED_SOFTIRQ
);
7691 #ifdef CONFIG_NO_HZ_COMMON
7692 if (nohz_kick_needed(rq
))
7693 nohz_balancer_kick();
7697 static void rq_online_fair(struct rq
*rq
)
7701 update_runtime_enabled(rq
);
7704 static void rq_offline_fair(struct rq
*rq
)
7708 /* Ensure any throttled groups are reachable by pick_next_task */
7709 unthrottle_offline_cfs_rqs(rq
);
7712 #endif /* CONFIG_SMP */
7715 * scheduler tick hitting a task of our scheduling class:
7717 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
7719 struct cfs_rq
*cfs_rq
;
7720 struct sched_entity
*se
= &curr
->se
;
7722 for_each_sched_entity(se
) {
7723 cfs_rq
= cfs_rq_of(se
);
7724 entity_tick(cfs_rq
, se
, queued
);
7727 if (numabalancing_enabled
)
7728 task_tick_numa(rq
, curr
);
7730 update_rq_runnable_avg(rq
, 1);
7734 * called on fork with the child task as argument from the parent's context
7735 * - child not yet on the tasklist
7736 * - preemption disabled
7738 static void task_fork_fair(struct task_struct
*p
)
7740 struct cfs_rq
*cfs_rq
;
7741 struct sched_entity
*se
= &p
->se
, *curr
;
7742 int this_cpu
= smp_processor_id();
7743 struct rq
*rq
= this_rq();
7744 unsigned long flags
;
7746 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7748 update_rq_clock(rq
);
7750 cfs_rq
= task_cfs_rq(current
);
7751 curr
= cfs_rq
->curr
;
7754 * Not only the cpu but also the task_group of the parent might have
7755 * been changed after parent->se.parent,cfs_rq were copied to
7756 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7757 * of child point to valid ones.
7760 __set_task_cpu(p
, this_cpu
);
7763 update_curr(cfs_rq
);
7766 se
->vruntime
= curr
->vruntime
;
7767 place_entity(cfs_rq
, se
, 1);
7769 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
7771 * Upon rescheduling, sched_class::put_prev_task() will place
7772 * 'current' within the tree based on its new key value.
7774 swap(curr
->vruntime
, se
->vruntime
);
7778 se
->vruntime
-= cfs_rq
->min_vruntime
;
7780 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7784 * Priority of the task has changed. Check to see if we preempt
7788 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
7790 if (!task_on_rq_queued(p
))
7794 * Reschedule if we are currently running on this runqueue and
7795 * our priority decreased, or if we are not currently running on
7796 * this runqueue and our priority is higher than the current's
7798 if (rq
->curr
== p
) {
7799 if (p
->prio
> oldprio
)
7802 check_preempt_curr(rq
, p
, 0);
7805 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
7807 struct sched_entity
*se
= &p
->se
;
7808 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7811 * Ensure the task's vruntime is normalized, so that when it's
7812 * switched back to the fair class the enqueue_entity(.flags=0) will
7813 * do the right thing.
7815 * If it's queued, then the dequeue_entity(.flags=0) will already
7816 * have normalized the vruntime, if it's !queued, then only when
7817 * the task is sleeping will it still have non-normalized vruntime.
7819 if (!task_on_rq_queued(p
) && p
->state
!= TASK_RUNNING
) {
7821 * Fix up our vruntime so that the current sleep doesn't
7822 * cause 'unlimited' sleep bonus.
7824 place_entity(cfs_rq
, se
, 0);
7825 se
->vruntime
-= cfs_rq
->min_vruntime
;
7830 * Remove our load from contribution when we leave sched_fair
7831 * and ensure we don't carry in an old decay_count if we
7834 if (se
->avg
.decay_count
) {
7835 __synchronize_entity_decay(se
);
7836 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7842 * We switched to the sched_fair class.
7844 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7846 #ifdef CONFIG_FAIR_GROUP_SCHED
7847 struct sched_entity
*se
= &p
->se
;
7849 * Since the real-depth could have been changed (only FAIR
7850 * class maintain depth value), reset depth properly.
7852 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7854 if (!task_on_rq_queued(p
))
7858 * We were most likely switched from sched_rt, so
7859 * kick off the schedule if running, otherwise just see
7860 * if we can still preempt the current task.
7865 check_preempt_curr(rq
, p
, 0);
7868 /* Account for a task changing its policy or group.
7870 * This routine is mostly called to set cfs_rq->curr field when a task
7871 * migrates between groups/classes.
7873 static void set_curr_task_fair(struct rq
*rq
)
7875 struct sched_entity
*se
= &rq
->curr
->se
;
7877 for_each_sched_entity(se
) {
7878 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7880 set_next_entity(cfs_rq
, se
);
7881 /* ensure bandwidth has been allocated on our new cfs_rq */
7882 account_cfs_rq_runtime(cfs_rq
, 0);
7886 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7888 cfs_rq
->tasks_timeline
= RB_ROOT
;
7889 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7890 #ifndef CONFIG_64BIT
7891 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7894 atomic64_set(&cfs_rq
->decay_counter
, 1);
7895 atomic_long_set(&cfs_rq
->removed_load
, 0);
7899 #ifdef CONFIG_FAIR_GROUP_SCHED
7900 static void task_move_group_fair(struct task_struct
*p
, int queued
)
7902 struct sched_entity
*se
= &p
->se
;
7903 struct cfs_rq
*cfs_rq
;
7906 * If the task was not on the rq at the time of this cgroup movement
7907 * it must have been asleep, sleeping tasks keep their ->vruntime
7908 * absolute on their old rq until wakeup (needed for the fair sleeper
7909 * bonus in place_entity()).
7911 * If it was on the rq, we've just 'preempted' it, which does convert
7912 * ->vruntime to a relative base.
7914 * Make sure both cases convert their relative position when migrating
7915 * to another cgroup's rq. This does somewhat interfere with the
7916 * fair sleeper stuff for the first placement, but who cares.
7919 * When !queued, vruntime of the task has usually NOT been normalized.
7920 * But there are some cases where it has already been normalized:
7922 * - Moving a forked child which is waiting for being woken up by
7923 * wake_up_new_task().
7924 * - Moving a task which has been woken up by try_to_wake_up() and
7925 * waiting for actually being woken up by sched_ttwu_pending().
7927 * To prevent boost or penalty in the new cfs_rq caused by delta
7928 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7930 if (!queued
&& (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7934 se
->vruntime
-= cfs_rq_of(se
)->min_vruntime
;
7935 set_task_rq(p
, task_cpu(p
));
7936 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7938 cfs_rq
= cfs_rq_of(se
);
7939 se
->vruntime
+= cfs_rq
->min_vruntime
;
7942 * migrate_task_rq_fair() will have removed our previous
7943 * contribution, but we must synchronize for ongoing future
7946 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7947 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
7952 void free_fair_sched_group(struct task_group
*tg
)
7956 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7958 for_each_possible_cpu(i
) {
7960 kfree(tg
->cfs_rq
[i
]);
7969 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7971 struct cfs_rq
*cfs_rq
;
7972 struct sched_entity
*se
;
7975 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7978 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7982 tg
->shares
= NICE_0_LOAD
;
7984 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7986 for_each_possible_cpu(i
) {
7987 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7988 GFP_KERNEL
, cpu_to_node(i
));
7992 se
= kzalloc_node(sizeof(struct sched_entity
),
7993 GFP_KERNEL
, cpu_to_node(i
));
7997 init_cfs_rq(cfs_rq
);
7998 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
8009 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8011 struct rq
*rq
= cpu_rq(cpu
);
8012 unsigned long flags
;
8015 * Only empty task groups can be destroyed; so we can speculatively
8016 * check on_list without danger of it being re-added.
8018 if (!tg
->cfs_rq
[cpu
]->on_list
)
8021 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8022 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
8023 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8026 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8027 struct sched_entity
*se
, int cpu
,
8028 struct sched_entity
*parent
)
8030 struct rq
*rq
= cpu_rq(cpu
);
8034 init_cfs_rq_runtime(cfs_rq
);
8036 tg
->cfs_rq
[cpu
] = cfs_rq
;
8039 /* se could be NULL for root_task_group */
8044 se
->cfs_rq
= &rq
->cfs
;
8047 se
->cfs_rq
= parent
->my_q
;
8048 se
->depth
= parent
->depth
+ 1;
8052 /* guarantee group entities always have weight */
8053 update_load_set(&se
->load
, NICE_0_LOAD
);
8054 se
->parent
= parent
;
8057 static DEFINE_MUTEX(shares_mutex
);
8059 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8062 unsigned long flags
;
8065 * We can't change the weight of the root cgroup.
8070 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
8072 mutex_lock(&shares_mutex
);
8073 if (tg
->shares
== shares
)
8076 tg
->shares
= shares
;
8077 for_each_possible_cpu(i
) {
8078 struct rq
*rq
= cpu_rq(i
);
8079 struct sched_entity
*se
;
8082 /* Propagate contribution to hierarchy */
8083 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8085 /* Possible calls to update_curr() need rq clock */
8086 update_rq_clock(rq
);
8087 for_each_sched_entity(se
)
8088 update_cfs_shares(group_cfs_rq(se
));
8089 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8093 mutex_unlock(&shares_mutex
);
8096 #else /* CONFIG_FAIR_GROUP_SCHED */
8098 void free_fair_sched_group(struct task_group
*tg
) { }
8100 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8105 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
8107 #endif /* CONFIG_FAIR_GROUP_SCHED */
8110 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
8112 struct sched_entity
*se
= &task
->se
;
8113 unsigned int rr_interval
= 0;
8116 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8119 if (rq
->cfs
.load
.weight
)
8120 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
8126 * All the scheduling class methods:
8128 const struct sched_class fair_sched_class
= {
8129 .next
= &idle_sched_class
,
8130 .enqueue_task
= enqueue_task_fair
,
8131 .dequeue_task
= dequeue_task_fair
,
8132 .yield_task
= yield_task_fair
,
8133 .yield_to_task
= yield_to_task_fair
,
8135 .check_preempt_curr
= check_preempt_wakeup
,
8137 .pick_next_task
= pick_next_task_fair
,
8138 .put_prev_task
= put_prev_task_fair
,
8141 .select_task_rq
= select_task_rq_fair
,
8142 .migrate_task_rq
= migrate_task_rq_fair
,
8144 .rq_online
= rq_online_fair
,
8145 .rq_offline
= rq_offline_fair
,
8147 .task_waking
= task_waking_fair
,
8150 .set_curr_task
= set_curr_task_fair
,
8151 .task_tick
= task_tick_fair
,
8152 .task_fork
= task_fork_fair
,
8154 .prio_changed
= prio_changed_fair
,
8155 .switched_from
= switched_from_fair
,
8156 .switched_to
= switched_to_fair
,
8158 .get_rr_interval
= get_rr_interval_fair
,
8160 .update_curr
= update_curr_fair
,
8162 #ifdef CONFIG_FAIR_GROUP_SCHED
8163 .task_move_group
= task_move_group_fair
,
8167 #ifdef CONFIG_SCHED_DEBUG
8168 void print_cfs_stats(struct seq_file
*m
, int cpu
)
8170 struct cfs_rq
*cfs_rq
;
8173 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
8174 print_cfs_rq(m
, cpu
, cfs_rq
);
8179 __init
void init_sched_fair_class(void)
8182 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8184 #ifdef CONFIG_NO_HZ_COMMON
8185 nohz
.next_balance
= jiffies
;
8186 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
8187 cpu_notifier(sched_ilb_notifier
, 0);