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 p
->se
.avg
.decay_count
= 0;
680 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
681 p
->se
.avg
.runnable_avg_sum
= slice
;
682 p
->se
.avg
.runnable_avg_period
= slice
;
683 __update_task_entity_contrib(&p
->se
);
686 void init_task_runnable_average(struct task_struct
*p
)
692 * Update the current task's runtime statistics.
694 static void update_curr(struct cfs_rq
*cfs_rq
)
696 struct sched_entity
*curr
= cfs_rq
->curr
;
697 u64 now
= rq_clock_task(rq_of(cfs_rq
));
703 delta_exec
= now
- curr
->exec_start
;
704 if (unlikely((s64
)delta_exec
<= 0))
707 curr
->exec_start
= now
;
709 schedstat_set(curr
->statistics
.exec_max
,
710 max(delta_exec
, curr
->statistics
.exec_max
));
712 curr
->sum_exec_runtime
+= delta_exec
;
713 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
715 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
716 update_min_vruntime(cfs_rq
);
718 if (entity_is_task(curr
)) {
719 struct task_struct
*curtask
= task_of(curr
);
721 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
722 cpuacct_charge(curtask
, delta_exec
);
723 account_group_exec_runtime(curtask
, delta_exec
);
726 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
730 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
732 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
736 * Task is being enqueued - update stats:
738 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
741 * Are we enqueueing a waiting task? (for current tasks
742 * a dequeue/enqueue event is a NOP)
744 if (se
!= cfs_rq
->curr
)
745 update_stats_wait_start(cfs_rq
, se
);
749 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
751 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
752 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
753 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
754 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
755 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
756 #ifdef CONFIG_SCHEDSTATS
757 if (entity_is_task(se
)) {
758 trace_sched_stat_wait(task_of(se
),
759 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
762 schedstat_set(se
->statistics
.wait_start
, 0);
766 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
769 * Mark the end of the wait period if dequeueing a
772 if (se
!= cfs_rq
->curr
)
773 update_stats_wait_end(cfs_rq
, se
);
777 * We are picking a new current task - update its stats:
780 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
783 * We are starting a new run period:
785 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
788 /**************************************************
789 * Scheduling class queueing methods:
792 #ifdef CONFIG_NUMA_BALANCING
794 * Approximate time to scan a full NUMA task in ms. The task scan period is
795 * calculated based on the tasks virtual memory size and
796 * numa_balancing_scan_size.
798 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
799 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
801 /* Portion of address space to scan in MB */
802 unsigned int sysctl_numa_balancing_scan_size
= 256;
804 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
805 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
807 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
809 unsigned long rss
= 0;
810 unsigned long nr_scan_pages
;
813 * Calculations based on RSS as non-present and empty pages are skipped
814 * by the PTE scanner and NUMA hinting faults should be trapped based
817 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
818 rss
= get_mm_rss(p
->mm
);
822 rss
= round_up(rss
, nr_scan_pages
);
823 return rss
/ nr_scan_pages
;
826 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
827 #define MAX_SCAN_WINDOW 2560
829 static unsigned int task_scan_min(struct task_struct
*p
)
831 unsigned int scan
, floor
;
832 unsigned int windows
= 1;
834 if (sysctl_numa_balancing_scan_size
< MAX_SCAN_WINDOW
)
835 windows
= MAX_SCAN_WINDOW
/ sysctl_numa_balancing_scan_size
;
836 floor
= 1000 / windows
;
838 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
839 return max_t(unsigned int, floor
, scan
);
842 static unsigned int task_scan_max(struct task_struct
*p
)
844 unsigned int smin
= task_scan_min(p
);
847 /* Watch for min being lower than max due to floor calculations */
848 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
849 return max(smin
, smax
);
852 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
854 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
855 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
858 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
860 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
861 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
867 spinlock_t lock
; /* nr_tasks, tasks */
870 struct list_head task_list
;
873 nodemask_t active_nodes
;
874 unsigned long total_faults
;
876 * Faults_cpu is used to decide whether memory should move
877 * towards the CPU. As a consequence, these stats are weighted
878 * more by CPU use than by memory faults.
880 unsigned long *faults_cpu
;
881 unsigned long faults
[0];
884 /* Shared or private faults. */
885 #define NR_NUMA_HINT_FAULT_TYPES 2
887 /* Memory and CPU locality */
888 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
890 /* Averaged statistics, and temporary buffers. */
891 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
893 pid_t
task_numa_group_id(struct task_struct
*p
)
895 return p
->numa_group
? p
->numa_group
->gid
: 0;
898 static inline int task_faults_idx(int nid
, int priv
)
900 return NR_NUMA_HINT_FAULT_TYPES
* nid
+ priv
;
903 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
905 if (!p
->numa_faults_memory
)
908 return p
->numa_faults_memory
[task_faults_idx(nid
, 0)] +
909 p
->numa_faults_memory
[task_faults_idx(nid
, 1)];
912 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
917 return p
->numa_group
->faults
[task_faults_idx(nid
, 0)] +
918 p
->numa_group
->faults
[task_faults_idx(nid
, 1)];
921 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
923 return group
->faults_cpu
[task_faults_idx(nid
, 0)] +
924 group
->faults_cpu
[task_faults_idx(nid
, 1)];
928 * These return the fraction of accesses done by a particular task, or
929 * task group, on a particular numa node. The group weight is given a
930 * larger multiplier, in order to group tasks together that are almost
931 * evenly spread out between numa nodes.
933 static inline unsigned long task_weight(struct task_struct
*p
, int nid
)
935 unsigned long total_faults
;
937 if (!p
->numa_faults_memory
)
940 total_faults
= p
->total_numa_faults
;
945 return 1000 * task_faults(p
, nid
) / total_faults
;
948 static inline unsigned long group_weight(struct task_struct
*p
, int nid
)
950 if (!p
->numa_group
|| !p
->numa_group
->total_faults
)
953 return 1000 * group_faults(p
, nid
) / p
->numa_group
->total_faults
;
956 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
957 int src_nid
, int dst_cpu
)
959 struct numa_group
*ng
= p
->numa_group
;
960 int dst_nid
= cpu_to_node(dst_cpu
);
961 int last_cpupid
, this_cpupid
;
963 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
966 * Multi-stage node selection is used in conjunction with a periodic
967 * migration fault to build a temporal task<->page relation. By using
968 * a two-stage filter we remove short/unlikely relations.
970 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
971 * a task's usage of a particular page (n_p) per total usage of this
972 * page (n_t) (in a given time-span) to a probability.
974 * Our periodic faults will sample this probability and getting the
975 * same result twice in a row, given these samples are fully
976 * independent, is then given by P(n)^2, provided our sample period
977 * is sufficiently short compared to the usage pattern.
979 * This quadric squishes small probabilities, making it less likely we
980 * act on an unlikely task<->page relation.
982 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
983 if (!cpupid_pid_unset(last_cpupid
) &&
984 cpupid_to_nid(last_cpupid
) != dst_nid
)
987 /* Always allow migrate on private faults */
988 if (cpupid_match_pid(p
, last_cpupid
))
991 /* A shared fault, but p->numa_group has not been set up yet. */
996 * Do not migrate if the destination is not a node that
997 * is actively used by this numa group.
999 if (!node_isset(dst_nid
, ng
->active_nodes
))
1003 * Source is a node that is not actively used by this
1004 * numa group, while the destination is. Migrate.
1006 if (!node_isset(src_nid
, ng
->active_nodes
))
1010 * Both source and destination are nodes in active
1011 * use by this numa group. Maximize memory bandwidth
1012 * by migrating from more heavily used groups, to less
1013 * heavily used ones, spreading the load around.
1014 * Use a 1/4 hysteresis to avoid spurious page movement.
1016 return group_faults(p
, dst_nid
) < (group_faults(p
, src_nid
) * 3 / 4);
1019 static unsigned long weighted_cpuload(const int cpu
);
1020 static unsigned long source_load(int cpu
, int type
);
1021 static unsigned long target_load(int cpu
, int type
);
1022 static unsigned long capacity_of(int cpu
);
1023 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
1025 /* Cached statistics for all CPUs within a node */
1027 unsigned long nr_running
;
1030 /* Total compute capacity of CPUs on a node */
1031 unsigned long compute_capacity
;
1033 /* Approximate capacity in terms of runnable tasks on a node */
1034 unsigned long task_capacity
;
1035 int has_free_capacity
;
1039 * XXX borrowed from update_sg_lb_stats
1041 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1043 int smt
, cpu
, cpus
= 0;
1044 unsigned long capacity
;
1046 memset(ns
, 0, sizeof(*ns
));
1047 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1048 struct rq
*rq
= cpu_rq(cpu
);
1050 ns
->nr_running
+= rq
->nr_running
;
1051 ns
->load
+= weighted_cpuload(cpu
);
1052 ns
->compute_capacity
+= capacity_of(cpu
);
1058 * If we raced with hotplug and there are no CPUs left in our mask
1059 * the @ns structure is NULL'ed and task_numa_compare() will
1060 * not find this node attractive.
1062 * We'll either bail at !has_free_capacity, or we'll detect a huge
1063 * imbalance and bail there.
1068 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1069 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, ns
->compute_capacity
);
1070 capacity
= cpus
/ smt
; /* cores */
1072 ns
->task_capacity
= min_t(unsigned, capacity
,
1073 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
));
1074 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1077 struct task_numa_env
{
1078 struct task_struct
*p
;
1080 int src_cpu
, src_nid
;
1081 int dst_cpu
, dst_nid
;
1083 struct numa_stats src_stats
, dst_stats
;
1087 struct task_struct
*best_task
;
1092 static void task_numa_assign(struct task_numa_env
*env
,
1093 struct task_struct
*p
, long imp
)
1096 put_task_struct(env
->best_task
);
1101 env
->best_imp
= imp
;
1102 env
->best_cpu
= env
->dst_cpu
;
1105 static bool load_too_imbalanced(long src_load
, long dst_load
,
1106 struct task_numa_env
*env
)
1109 long orig_src_load
, orig_dst_load
;
1110 long src_capacity
, dst_capacity
;
1113 * The load is corrected for the CPU capacity available on each node.
1116 * ------------ vs ---------
1117 * src_capacity dst_capacity
1119 src_capacity
= env
->src_stats
.compute_capacity
;
1120 dst_capacity
= env
->dst_stats
.compute_capacity
;
1122 /* We care about the slope of the imbalance, not the direction. */
1123 if (dst_load
< src_load
)
1124 swap(dst_load
, src_load
);
1126 /* Is the difference below the threshold? */
1127 imb
= dst_load
* src_capacity
* 100 -
1128 src_load
* dst_capacity
* env
->imbalance_pct
;
1133 * The imbalance is above the allowed threshold.
1134 * Compare it with the old imbalance.
1136 orig_src_load
= env
->src_stats
.load
;
1137 orig_dst_load
= env
->dst_stats
.load
;
1139 if (orig_dst_load
< orig_src_load
)
1140 swap(orig_dst_load
, orig_src_load
);
1142 old_imb
= orig_dst_load
* src_capacity
* 100 -
1143 orig_src_load
* dst_capacity
* env
->imbalance_pct
;
1145 /* Would this change make things worse? */
1146 return (imb
> old_imb
);
1150 * This checks if the overall compute and NUMA accesses of the system would
1151 * be improved if the source tasks was migrated to the target dst_cpu taking
1152 * into account that it might be best if task running on the dst_cpu should
1153 * be exchanged with the source task
1155 static void task_numa_compare(struct task_numa_env
*env
,
1156 long taskimp
, long groupimp
)
1158 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1159 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1160 struct task_struct
*cur
;
1161 long src_load
, dst_load
;
1163 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1168 raw_spin_lock_irq(&dst_rq
->lock
);
1171 * No need to move the exiting task, and this ensures that ->curr
1172 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1173 * is safe under RCU read lock.
1174 * Note that rcu_read_lock() itself can't protect from the final
1175 * put_task_struct() after the last schedule().
1177 if ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
))
1179 raw_spin_unlock_irq(&dst_rq
->lock
);
1182 * "imp" is the fault differential for the source task between the
1183 * source and destination node. Calculate the total differential for
1184 * the source task and potential destination task. The more negative
1185 * the value is, the more rmeote accesses that would be expected to
1186 * be incurred if the tasks were swapped.
1189 /* Skip this swap candidate if cannot move to the source cpu */
1190 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1194 * If dst and source tasks are in the same NUMA group, or not
1195 * in any group then look only at task weights.
1197 if (cur
->numa_group
== env
->p
->numa_group
) {
1198 imp
= taskimp
+ task_weight(cur
, env
->src_nid
) -
1199 task_weight(cur
, env
->dst_nid
);
1201 * Add some hysteresis to prevent swapping the
1202 * tasks within a group over tiny differences.
1204 if (cur
->numa_group
)
1208 * Compare the group weights. If a task is all by
1209 * itself (not part of a group), use the task weight
1212 if (cur
->numa_group
)
1213 imp
+= group_weight(cur
, env
->src_nid
) -
1214 group_weight(cur
, env
->dst_nid
);
1216 imp
+= task_weight(cur
, env
->src_nid
) -
1217 task_weight(cur
, env
->dst_nid
);
1221 if (imp
<= env
->best_imp
&& moveimp
<= env
->best_imp
)
1225 /* Is there capacity at our destination? */
1226 if (env
->src_stats
.nr_running
<= env
->src_stats
.task_capacity
&&
1227 !env
->dst_stats
.has_free_capacity
)
1233 /* Balance doesn't matter much if we're running a task per cpu */
1234 if (imp
> env
->best_imp
&& src_rq
->nr_running
== 1 &&
1235 dst_rq
->nr_running
== 1)
1239 * In the overloaded case, try and keep the load balanced.
1242 load
= task_h_load(env
->p
);
1243 dst_load
= env
->dst_stats
.load
+ load
;
1244 src_load
= env
->src_stats
.load
- load
;
1246 if (moveimp
> imp
&& moveimp
> env
->best_imp
) {
1248 * If the improvement from just moving env->p direction is
1249 * better than swapping tasks around, check if a move is
1250 * possible. Store a slightly smaller score than moveimp,
1251 * so an actually idle CPU will win.
1253 if (!load_too_imbalanced(src_load
, dst_load
, env
)) {
1260 if (imp
<= env
->best_imp
)
1264 load
= task_h_load(cur
);
1269 if (load_too_imbalanced(src_load
, dst_load
, env
))
1273 * One idle CPU per node is evaluated for a task numa move.
1274 * Call select_idle_sibling to maybe find a better one.
1277 env
->dst_cpu
= select_idle_sibling(env
->p
, env
->dst_cpu
);
1280 task_numa_assign(env
, cur
, imp
);
1285 static void task_numa_find_cpu(struct task_numa_env
*env
,
1286 long taskimp
, long groupimp
)
1290 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1291 /* Skip this CPU if the source task cannot migrate */
1292 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1296 task_numa_compare(env
, taskimp
, groupimp
);
1300 static int task_numa_migrate(struct task_struct
*p
)
1302 struct task_numa_env env
= {
1305 .src_cpu
= task_cpu(p
),
1306 .src_nid
= task_node(p
),
1308 .imbalance_pct
= 112,
1314 struct sched_domain
*sd
;
1315 unsigned long taskweight
, groupweight
;
1317 long taskimp
, groupimp
;
1320 * Pick the lowest SD_NUMA domain, as that would have the smallest
1321 * imbalance and would be the first to start moving tasks about.
1323 * And we want to avoid any moving of tasks about, as that would create
1324 * random movement of tasks -- counter the numa conditions we're trying
1328 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1330 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1334 * Cpusets can break the scheduler domain tree into smaller
1335 * balance domains, some of which do not cross NUMA boundaries.
1336 * Tasks that are "trapped" in such domains cannot be migrated
1337 * elsewhere, so there is no point in (re)trying.
1339 if (unlikely(!sd
)) {
1340 p
->numa_preferred_nid
= task_node(p
);
1344 taskweight
= task_weight(p
, env
.src_nid
);
1345 groupweight
= group_weight(p
, env
.src_nid
);
1346 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1347 env
.dst_nid
= p
->numa_preferred_nid
;
1348 taskimp
= task_weight(p
, env
.dst_nid
) - taskweight
;
1349 groupimp
= group_weight(p
, env
.dst_nid
) - groupweight
;
1350 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1352 /* Try to find a spot on the preferred nid. */
1353 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1355 /* No space available on the preferred nid. Look elsewhere. */
1356 if (env
.best_cpu
== -1) {
1357 for_each_online_node(nid
) {
1358 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1361 /* Only consider nodes where both task and groups benefit */
1362 taskimp
= task_weight(p
, nid
) - taskweight
;
1363 groupimp
= group_weight(p
, nid
) - groupweight
;
1364 if (taskimp
< 0 && groupimp
< 0)
1368 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1369 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1374 * If the task is part of a workload that spans multiple NUMA nodes,
1375 * and is migrating into one of the workload's active nodes, remember
1376 * this node as the task's preferred numa node, so the workload can
1378 * A task that migrated to a second choice node will be better off
1379 * trying for a better one later. Do not set the preferred node here.
1381 if (p
->numa_group
) {
1382 if (env
.best_cpu
== -1)
1387 if (node_isset(nid
, p
->numa_group
->active_nodes
))
1388 sched_setnuma(p
, env
.dst_nid
);
1391 /* No better CPU than the current one was found. */
1392 if (env
.best_cpu
== -1)
1396 * Reset the scan period if the task is being rescheduled on an
1397 * alternative node to recheck if the tasks is now properly placed.
1399 p
->numa_scan_period
= task_scan_min(p
);
1401 if (env
.best_task
== NULL
) {
1402 ret
= migrate_task_to(p
, env
.best_cpu
);
1404 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1408 ret
= migrate_swap(p
, env
.best_task
);
1410 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1411 put_task_struct(env
.best_task
);
1415 /* Attempt to migrate a task to a CPU on the preferred node. */
1416 static void numa_migrate_preferred(struct task_struct
*p
)
1418 unsigned long interval
= HZ
;
1420 /* This task has no NUMA fault statistics yet */
1421 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults_memory
))
1424 /* Periodically retry migrating the task to the preferred node */
1425 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1426 p
->numa_migrate_retry
= jiffies
+ interval
;
1428 /* Success if task is already running on preferred CPU */
1429 if (task_node(p
) == p
->numa_preferred_nid
)
1432 /* Otherwise, try migrate to a CPU on the preferred node */
1433 task_numa_migrate(p
);
1437 * Find the nodes on which the workload is actively running. We do this by
1438 * tracking the nodes from which NUMA hinting faults are triggered. This can
1439 * be different from the set of nodes where the workload's memory is currently
1442 * The bitmask is used to make smarter decisions on when to do NUMA page
1443 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1444 * are added when they cause over 6/16 of the maximum number of faults, but
1445 * only removed when they drop below 3/16.
1447 static void update_numa_active_node_mask(struct numa_group
*numa_group
)
1449 unsigned long faults
, max_faults
= 0;
1452 for_each_online_node(nid
) {
1453 faults
= group_faults_cpu(numa_group
, nid
);
1454 if (faults
> max_faults
)
1455 max_faults
= faults
;
1458 for_each_online_node(nid
) {
1459 faults
= group_faults_cpu(numa_group
, nid
);
1460 if (!node_isset(nid
, numa_group
->active_nodes
)) {
1461 if (faults
> max_faults
* 6 / 16)
1462 node_set(nid
, numa_group
->active_nodes
);
1463 } else if (faults
< max_faults
* 3 / 16)
1464 node_clear(nid
, numa_group
->active_nodes
);
1469 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1470 * increments. The more local the fault statistics are, the higher the scan
1471 * period will be for the next scan window. If local/(local+remote) ratio is
1472 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1473 * the scan period will decrease. Aim for 70% local accesses.
1475 #define NUMA_PERIOD_SLOTS 10
1476 #define NUMA_PERIOD_THRESHOLD 7
1479 * Increase the scan period (slow down scanning) if the majority of
1480 * our memory is already on our local node, or if the majority of
1481 * the page accesses are shared with other processes.
1482 * Otherwise, decrease the scan period.
1484 static void update_task_scan_period(struct task_struct
*p
,
1485 unsigned long shared
, unsigned long private)
1487 unsigned int period_slot
;
1491 unsigned long remote
= p
->numa_faults_locality
[0];
1492 unsigned long local
= p
->numa_faults_locality
[1];
1495 * If there were no record hinting faults then either the task is
1496 * completely idle or all activity is areas that are not of interest
1497 * to automatic numa balancing. Scan slower
1499 if (local
+ shared
== 0) {
1500 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1501 p
->numa_scan_period
<< 1);
1503 p
->mm
->numa_next_scan
= jiffies
+
1504 msecs_to_jiffies(p
->numa_scan_period
);
1510 * Prepare to scale scan period relative to the current period.
1511 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1512 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1513 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1515 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1516 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1517 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1518 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1521 diff
= slot
* period_slot
;
1523 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1526 * Scale scan rate increases based on sharing. There is an
1527 * inverse relationship between the degree of sharing and
1528 * the adjustment made to the scanning period. Broadly
1529 * speaking the intent is that there is little point
1530 * scanning faster if shared accesses dominate as it may
1531 * simply bounce migrations uselessly
1533 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
));
1534 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1537 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1538 task_scan_min(p
), task_scan_max(p
));
1539 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1543 * Get the fraction of time the task has been running since the last
1544 * NUMA placement cycle. The scheduler keeps similar statistics, but
1545 * decays those on a 32ms period, which is orders of magnitude off
1546 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1547 * stats only if the task is so new there are no NUMA statistics yet.
1549 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1551 u64 runtime
, delta
, now
;
1552 /* Use the start of this time slice to avoid calculations. */
1553 now
= p
->se
.exec_start
;
1554 runtime
= p
->se
.sum_exec_runtime
;
1556 if (p
->last_task_numa_placement
) {
1557 delta
= runtime
- p
->last_sum_exec_runtime
;
1558 *period
= now
- p
->last_task_numa_placement
;
1560 delta
= p
->se
.avg
.runnable_avg_sum
;
1561 *period
= p
->se
.avg
.runnable_avg_period
;
1564 p
->last_sum_exec_runtime
= runtime
;
1565 p
->last_task_numa_placement
= now
;
1570 static void task_numa_placement(struct task_struct
*p
)
1572 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1573 unsigned long max_faults
= 0, max_group_faults
= 0;
1574 unsigned long fault_types
[2] = { 0, 0 };
1575 unsigned long total_faults
;
1576 u64 runtime
, period
;
1577 spinlock_t
*group_lock
= NULL
;
1579 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1580 if (p
->numa_scan_seq
== seq
)
1582 p
->numa_scan_seq
= seq
;
1583 p
->numa_scan_period_max
= task_scan_max(p
);
1585 total_faults
= p
->numa_faults_locality
[0] +
1586 p
->numa_faults_locality
[1];
1587 runtime
= numa_get_avg_runtime(p
, &period
);
1589 /* If the task is part of a group prevent parallel updates to group stats */
1590 if (p
->numa_group
) {
1591 group_lock
= &p
->numa_group
->lock
;
1592 spin_lock_irq(group_lock
);
1595 /* Find the node with the highest number of faults */
1596 for_each_online_node(nid
) {
1597 unsigned long faults
= 0, group_faults
= 0;
1600 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
1601 long diff
, f_diff
, f_weight
;
1603 i
= task_faults_idx(nid
, priv
);
1605 /* Decay existing window, copy faults since last scan */
1606 diff
= p
->numa_faults_buffer_memory
[i
] - p
->numa_faults_memory
[i
] / 2;
1607 fault_types
[priv
] += p
->numa_faults_buffer_memory
[i
];
1608 p
->numa_faults_buffer_memory
[i
] = 0;
1611 * Normalize the faults_from, so all tasks in a group
1612 * count according to CPU use, instead of by the raw
1613 * number of faults. Tasks with little runtime have
1614 * little over-all impact on throughput, and thus their
1615 * faults are less important.
1617 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
1618 f_weight
= (f_weight
* p
->numa_faults_buffer_cpu
[i
]) /
1620 f_diff
= f_weight
- p
->numa_faults_cpu
[i
] / 2;
1621 p
->numa_faults_buffer_cpu
[i
] = 0;
1623 p
->numa_faults_memory
[i
] += diff
;
1624 p
->numa_faults_cpu
[i
] += f_diff
;
1625 faults
+= p
->numa_faults_memory
[i
];
1626 p
->total_numa_faults
+= diff
;
1627 if (p
->numa_group
) {
1628 /* safe because we can only change our own group */
1629 p
->numa_group
->faults
[i
] += diff
;
1630 p
->numa_group
->faults_cpu
[i
] += f_diff
;
1631 p
->numa_group
->total_faults
+= diff
;
1632 group_faults
+= p
->numa_group
->faults
[i
];
1636 if (faults
> max_faults
) {
1637 max_faults
= faults
;
1641 if (group_faults
> max_group_faults
) {
1642 max_group_faults
= group_faults
;
1643 max_group_nid
= nid
;
1647 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1649 if (p
->numa_group
) {
1650 update_numa_active_node_mask(p
->numa_group
);
1651 spin_unlock_irq(group_lock
);
1652 max_nid
= max_group_nid
;
1656 /* Set the new preferred node */
1657 if (max_nid
!= p
->numa_preferred_nid
)
1658 sched_setnuma(p
, max_nid
);
1660 if (task_node(p
) != p
->numa_preferred_nid
)
1661 numa_migrate_preferred(p
);
1665 static inline int get_numa_group(struct numa_group
*grp
)
1667 return atomic_inc_not_zero(&grp
->refcount
);
1670 static inline void put_numa_group(struct numa_group
*grp
)
1672 if (atomic_dec_and_test(&grp
->refcount
))
1673 kfree_rcu(grp
, rcu
);
1676 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1679 struct numa_group
*grp
, *my_grp
;
1680 struct task_struct
*tsk
;
1682 int cpu
= cpupid_to_cpu(cpupid
);
1685 if (unlikely(!p
->numa_group
)) {
1686 unsigned int size
= sizeof(struct numa_group
) +
1687 4*nr_node_ids
*sizeof(unsigned long);
1689 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1693 atomic_set(&grp
->refcount
, 1);
1694 spin_lock_init(&grp
->lock
);
1695 INIT_LIST_HEAD(&grp
->task_list
);
1697 /* Second half of the array tracks nids where faults happen */
1698 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
1701 node_set(task_node(current
), grp
->active_nodes
);
1703 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1704 grp
->faults
[i
] = p
->numa_faults_memory
[i
];
1706 grp
->total_faults
= p
->total_numa_faults
;
1708 list_add(&p
->numa_entry
, &grp
->task_list
);
1710 rcu_assign_pointer(p
->numa_group
, grp
);
1714 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1716 if (!cpupid_match_pid(tsk
, cpupid
))
1719 grp
= rcu_dereference(tsk
->numa_group
);
1723 my_grp
= p
->numa_group
;
1728 * Only join the other group if its bigger; if we're the bigger group,
1729 * the other task will join us.
1731 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1735 * Tie-break on the grp address.
1737 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1740 /* Always join threads in the same process. */
1741 if (tsk
->mm
== current
->mm
)
1744 /* Simple filter to avoid false positives due to PID collisions */
1745 if (flags
& TNF_SHARED
)
1748 /* Update priv based on whether false sharing was detected */
1751 if (join
&& !get_numa_group(grp
))
1759 BUG_ON(irqs_disabled());
1760 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
1762 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
1763 my_grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1764 grp
->faults
[i
] += p
->numa_faults_memory
[i
];
1766 my_grp
->total_faults
-= p
->total_numa_faults
;
1767 grp
->total_faults
+= p
->total_numa_faults
;
1769 list_move(&p
->numa_entry
, &grp
->task_list
);
1773 spin_unlock(&my_grp
->lock
);
1774 spin_unlock_irq(&grp
->lock
);
1776 rcu_assign_pointer(p
->numa_group
, grp
);
1778 put_numa_group(my_grp
);
1786 void task_numa_free(struct task_struct
*p
)
1788 struct numa_group
*grp
= p
->numa_group
;
1789 void *numa_faults
= p
->numa_faults_memory
;
1790 unsigned long flags
;
1794 spin_lock_irqsave(&grp
->lock
, flags
);
1795 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1796 grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1797 grp
->total_faults
-= p
->total_numa_faults
;
1799 list_del(&p
->numa_entry
);
1801 spin_unlock_irqrestore(&grp
->lock
, flags
);
1802 RCU_INIT_POINTER(p
->numa_group
, NULL
);
1803 put_numa_group(grp
);
1806 p
->numa_faults_memory
= NULL
;
1807 p
->numa_faults_buffer_memory
= NULL
;
1808 p
->numa_faults_cpu
= NULL
;
1809 p
->numa_faults_buffer_cpu
= NULL
;
1814 * Got a PROT_NONE fault for a page on @node.
1816 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
1818 struct task_struct
*p
= current
;
1819 bool migrated
= flags
& TNF_MIGRATED
;
1820 int cpu_node
= task_node(current
);
1821 int local
= !!(flags
& TNF_FAULT_LOCAL
);
1824 if (!numabalancing_enabled
)
1827 /* for example, ksmd faulting in a user's mm */
1831 /* Allocate buffer to track faults on a per-node basis */
1832 if (unlikely(!p
->numa_faults_memory
)) {
1833 int size
= sizeof(*p
->numa_faults_memory
) *
1834 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
1836 p
->numa_faults_memory
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
1837 if (!p
->numa_faults_memory
)
1840 BUG_ON(p
->numa_faults_buffer_memory
);
1842 * The averaged statistics, shared & private, memory & cpu,
1843 * occupy the first half of the array. The second half of the
1844 * array is for current counters, which are averaged into the
1845 * first set by task_numa_placement.
1847 p
->numa_faults_cpu
= p
->numa_faults_memory
+ (2 * nr_node_ids
);
1848 p
->numa_faults_buffer_memory
= p
->numa_faults_memory
+ (4 * nr_node_ids
);
1849 p
->numa_faults_buffer_cpu
= p
->numa_faults_memory
+ (6 * nr_node_ids
);
1850 p
->total_numa_faults
= 0;
1851 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1855 * First accesses are treated as private, otherwise consider accesses
1856 * to be private if the accessing pid has not changed
1858 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
1861 priv
= cpupid_match_pid(p
, last_cpupid
);
1862 if (!priv
&& !(flags
& TNF_NO_GROUP
))
1863 task_numa_group(p
, last_cpupid
, flags
, &priv
);
1867 * If a workload spans multiple NUMA nodes, a shared fault that
1868 * occurs wholly within the set of nodes that the workload is
1869 * actively using should be counted as local. This allows the
1870 * scan rate to slow down when a workload has settled down.
1872 if (!priv
&& !local
&& p
->numa_group
&&
1873 node_isset(cpu_node
, p
->numa_group
->active_nodes
) &&
1874 node_isset(mem_node
, p
->numa_group
->active_nodes
))
1877 task_numa_placement(p
);
1880 * Retry task to preferred node migration periodically, in case it
1881 * case it previously failed, or the scheduler moved us.
1883 if (time_after(jiffies
, p
->numa_migrate_retry
))
1884 numa_migrate_preferred(p
);
1887 p
->numa_pages_migrated
+= pages
;
1889 p
->numa_faults_buffer_memory
[task_faults_idx(mem_node
, priv
)] += pages
;
1890 p
->numa_faults_buffer_cpu
[task_faults_idx(cpu_node
, priv
)] += pages
;
1891 p
->numa_faults_locality
[local
] += pages
;
1894 static void reset_ptenuma_scan(struct task_struct
*p
)
1896 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
1897 p
->mm
->numa_scan_offset
= 0;
1901 * The expensive part of numa migration is done from task_work context.
1902 * Triggered from task_tick_numa().
1904 void task_numa_work(struct callback_head
*work
)
1906 unsigned long migrate
, next_scan
, now
= jiffies
;
1907 struct task_struct
*p
= current
;
1908 struct mm_struct
*mm
= p
->mm
;
1909 struct vm_area_struct
*vma
;
1910 unsigned long start
, end
;
1911 unsigned long nr_pte_updates
= 0;
1914 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
1916 work
->next
= work
; /* protect against double add */
1918 * Who cares about NUMA placement when they're dying.
1920 * NOTE: make sure not to dereference p->mm before this check,
1921 * exit_task_work() happens _after_ exit_mm() so we could be called
1922 * without p->mm even though we still had it when we enqueued this
1925 if (p
->flags
& PF_EXITING
)
1928 if (!mm
->numa_next_scan
) {
1929 mm
->numa_next_scan
= now
+
1930 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1934 * Enforce maximal scan/migration frequency..
1936 migrate
= mm
->numa_next_scan
;
1937 if (time_before(now
, migrate
))
1940 if (p
->numa_scan_period
== 0) {
1941 p
->numa_scan_period_max
= task_scan_max(p
);
1942 p
->numa_scan_period
= task_scan_min(p
);
1945 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1946 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
1950 * Delay this task enough that another task of this mm will likely win
1951 * the next time around.
1953 p
->node_stamp
+= 2 * TICK_NSEC
;
1955 start
= mm
->numa_scan_offset
;
1956 pages
= sysctl_numa_balancing_scan_size
;
1957 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
1961 down_read(&mm
->mmap_sem
);
1962 vma
= find_vma(mm
, start
);
1964 reset_ptenuma_scan(p
);
1968 for (; vma
; vma
= vma
->vm_next
) {
1969 if (!vma_migratable(vma
) || !vma_policy_mof(vma
))
1973 * Shared library pages mapped by multiple processes are not
1974 * migrated as it is expected they are cache replicated. Avoid
1975 * hinting faults in read-only file-backed mappings or the vdso
1976 * as migrating the pages will be of marginal benefit.
1979 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
1983 * Skip inaccessible VMAs to avoid any confusion between
1984 * PROT_NONE and NUMA hinting ptes
1986 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
1990 start
= max(start
, vma
->vm_start
);
1991 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
1992 end
= min(end
, vma
->vm_end
);
1993 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
1996 * Scan sysctl_numa_balancing_scan_size but ensure that
1997 * at least one PTE is updated so that unused virtual
1998 * address space is quickly skipped.
2001 pages
-= (end
- start
) >> PAGE_SHIFT
;
2008 } while (end
!= vma
->vm_end
);
2013 * It is possible to reach the end of the VMA list but the last few
2014 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2015 * would find the !migratable VMA on the next scan but not reset the
2016 * scanner to the start so check it now.
2019 mm
->numa_scan_offset
= start
;
2021 reset_ptenuma_scan(p
);
2022 up_read(&mm
->mmap_sem
);
2026 * Drive the periodic memory faults..
2028 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2030 struct callback_head
*work
= &curr
->numa_work
;
2034 * We don't care about NUMA placement if we don't have memory.
2036 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2040 * Using runtime rather than walltime has the dual advantage that
2041 * we (mostly) drive the selection from busy threads and that the
2042 * task needs to have done some actual work before we bother with
2045 now
= curr
->se
.sum_exec_runtime
;
2046 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2048 if (now
- curr
->node_stamp
> period
) {
2049 if (!curr
->node_stamp
)
2050 curr
->numa_scan_period
= task_scan_min(curr
);
2051 curr
->node_stamp
+= period
;
2053 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2054 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2055 task_work_add(curr
, work
, true);
2060 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2064 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2068 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2071 #endif /* CONFIG_NUMA_BALANCING */
2074 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2076 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2077 if (!parent_entity(se
))
2078 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2080 if (entity_is_task(se
)) {
2081 struct rq
*rq
= rq_of(cfs_rq
);
2083 account_numa_enqueue(rq
, task_of(se
));
2084 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2087 cfs_rq
->nr_running
++;
2091 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2093 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2094 if (!parent_entity(se
))
2095 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2096 if (entity_is_task(se
)) {
2097 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2098 list_del_init(&se
->group_node
);
2100 cfs_rq
->nr_running
--;
2103 #ifdef CONFIG_FAIR_GROUP_SCHED
2105 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
2110 * Use this CPU's actual weight instead of the last load_contribution
2111 * to gain a more accurate current total weight. See
2112 * update_cfs_rq_load_contribution().
2114 tg_weight
= atomic_long_read(&tg
->load_avg
);
2115 tg_weight
-= cfs_rq
->tg_load_contrib
;
2116 tg_weight
+= cfs_rq
->load
.weight
;
2121 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2123 long tg_weight
, load
, shares
;
2125 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
2126 load
= cfs_rq
->load
.weight
;
2128 shares
= (tg
->shares
* load
);
2130 shares
/= tg_weight
;
2132 if (shares
< MIN_SHARES
)
2133 shares
= MIN_SHARES
;
2134 if (shares
> tg
->shares
)
2135 shares
= tg
->shares
;
2139 # else /* CONFIG_SMP */
2140 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2144 # endif /* CONFIG_SMP */
2145 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2146 unsigned long weight
)
2149 /* commit outstanding execution time */
2150 if (cfs_rq
->curr
== se
)
2151 update_curr(cfs_rq
);
2152 account_entity_dequeue(cfs_rq
, se
);
2155 update_load_set(&se
->load
, weight
);
2158 account_entity_enqueue(cfs_rq
, se
);
2161 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2163 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2165 struct task_group
*tg
;
2166 struct sched_entity
*se
;
2170 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2171 if (!se
|| throttled_hierarchy(cfs_rq
))
2174 if (likely(se
->load
.weight
== tg
->shares
))
2177 shares
= calc_cfs_shares(cfs_rq
, tg
);
2179 reweight_entity(cfs_rq_of(se
), se
, shares
);
2181 #else /* CONFIG_FAIR_GROUP_SCHED */
2182 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2185 #endif /* CONFIG_FAIR_GROUP_SCHED */
2189 * We choose a half-life close to 1 scheduling period.
2190 * Note: The tables below are dependent on this value.
2192 #define LOAD_AVG_PERIOD 32
2193 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2194 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2196 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2197 static const u32 runnable_avg_yN_inv
[] = {
2198 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2199 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2200 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2201 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2202 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2203 0x85aac367, 0x82cd8698,
2207 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2208 * over-estimates when re-combining.
2210 static const u32 runnable_avg_yN_sum
[] = {
2211 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2212 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2213 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2218 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2220 static __always_inline u64
decay_load(u64 val
, u64 n
)
2222 unsigned int local_n
;
2226 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2229 /* after bounds checking we can collapse to 32-bit */
2233 * As y^PERIOD = 1/2, we can combine
2234 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2235 * With a look-up table which covers y^n (n<PERIOD)
2237 * To achieve constant time decay_load.
2239 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2240 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2241 local_n
%= LOAD_AVG_PERIOD
;
2244 val
*= runnable_avg_yN_inv
[local_n
];
2245 /* We don't use SRR here since we always want to round down. */
2250 * For updates fully spanning n periods, the contribution to runnable
2251 * average will be: \Sum 1024*y^n
2253 * We can compute this reasonably efficiently by combining:
2254 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2256 static u32
__compute_runnable_contrib(u64 n
)
2260 if (likely(n
<= LOAD_AVG_PERIOD
))
2261 return runnable_avg_yN_sum
[n
];
2262 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2263 return LOAD_AVG_MAX
;
2265 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2267 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2268 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2270 n
-= LOAD_AVG_PERIOD
;
2271 } while (n
> LOAD_AVG_PERIOD
);
2273 contrib
= decay_load(contrib
, n
);
2274 return contrib
+ runnable_avg_yN_sum
[n
];
2278 * We can represent the historical contribution to runnable average as the
2279 * coefficients of a geometric series. To do this we sub-divide our runnable
2280 * history into segments of approximately 1ms (1024us); label the segment that
2281 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2283 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2285 * (now) (~1ms ago) (~2ms ago)
2287 * Let u_i denote the fraction of p_i that the entity was runnable.
2289 * We then designate the fractions u_i as our co-efficients, yielding the
2290 * following representation of historical load:
2291 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2293 * We choose y based on the with of a reasonably scheduling period, fixing:
2296 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2297 * approximately half as much as the contribution to load within the last ms
2300 * When a period "rolls over" and we have new u_0`, multiplying the previous
2301 * sum again by y is sufficient to update:
2302 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2303 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2305 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2306 struct sched_avg
*sa
,
2310 u32 runnable_contrib
;
2311 int delta_w
, decayed
= 0;
2313 delta
= now
- sa
->last_runnable_update
;
2315 * This should only happen when time goes backwards, which it
2316 * unfortunately does during sched clock init when we swap over to TSC.
2318 if ((s64
)delta
< 0) {
2319 sa
->last_runnable_update
= now
;
2324 * Use 1024ns as the unit of measurement since it's a reasonable
2325 * approximation of 1us and fast to compute.
2330 sa
->last_runnable_update
= now
;
2332 /* delta_w is the amount already accumulated against our next period */
2333 delta_w
= sa
->runnable_avg_period
% 1024;
2334 if (delta
+ delta_w
>= 1024) {
2335 /* period roll-over */
2339 * Now that we know we're crossing a period boundary, figure
2340 * out how much from delta we need to complete the current
2341 * period and accrue it.
2343 delta_w
= 1024 - delta_w
;
2345 sa
->runnable_avg_sum
+= delta_w
;
2346 sa
->runnable_avg_period
+= delta_w
;
2350 /* Figure out how many additional periods this update spans */
2351 periods
= delta
/ 1024;
2354 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2356 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2359 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2360 runnable_contrib
= __compute_runnable_contrib(periods
);
2362 sa
->runnable_avg_sum
+= runnable_contrib
;
2363 sa
->runnable_avg_period
+= runnable_contrib
;
2366 /* Remainder of delta accrued against u_0` */
2368 sa
->runnable_avg_sum
+= delta
;
2369 sa
->runnable_avg_period
+= delta
;
2374 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2375 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2377 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2378 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2380 decays
-= se
->avg
.decay_count
;
2384 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2385 se
->avg
.decay_count
= 0;
2390 #ifdef CONFIG_FAIR_GROUP_SCHED
2391 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2394 struct task_group
*tg
= cfs_rq
->tg
;
2397 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2398 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2403 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2404 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2405 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2410 * Aggregate cfs_rq runnable averages into an equivalent task_group
2411 * representation for computing load contributions.
2413 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2414 struct cfs_rq
*cfs_rq
)
2416 struct task_group
*tg
= cfs_rq
->tg
;
2419 /* The fraction of a cpu used by this cfs_rq */
2420 contrib
= div_u64((u64
)sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2421 sa
->runnable_avg_period
+ 1);
2422 contrib
-= cfs_rq
->tg_runnable_contrib
;
2424 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2425 atomic_add(contrib
, &tg
->runnable_avg
);
2426 cfs_rq
->tg_runnable_contrib
+= contrib
;
2430 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2432 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2433 struct task_group
*tg
= cfs_rq
->tg
;
2438 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2439 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2440 atomic_long_read(&tg
->load_avg
) + 1);
2443 * For group entities we need to compute a correction term in the case
2444 * that they are consuming <1 cpu so that we would contribute the same
2445 * load as a task of equal weight.
2447 * Explicitly co-ordinating this measurement would be expensive, but
2448 * fortunately the sum of each cpus contribution forms a usable
2449 * lower-bound on the true value.
2451 * Consider the aggregate of 2 contributions. Either they are disjoint
2452 * (and the sum represents true value) or they are disjoint and we are
2453 * understating by the aggregate of their overlap.
2455 * Extending this to N cpus, for a given overlap, the maximum amount we
2456 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2457 * cpus that overlap for this interval and w_i is the interval width.
2459 * On a small machine; the first term is well-bounded which bounds the
2460 * total error since w_i is a subset of the period. Whereas on a
2461 * larger machine, while this first term can be larger, if w_i is the
2462 * of consequential size guaranteed to see n_i*w_i quickly converge to
2463 * our upper bound of 1-cpu.
2465 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2466 if (runnable_avg
< NICE_0_LOAD
) {
2467 se
->avg
.load_avg_contrib
*= runnable_avg
;
2468 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2472 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2474 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2475 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2477 #else /* CONFIG_FAIR_GROUP_SCHED */
2478 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2479 int force_update
) {}
2480 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2481 struct cfs_rq
*cfs_rq
) {}
2482 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2483 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2484 #endif /* CONFIG_FAIR_GROUP_SCHED */
2486 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2490 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2491 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2492 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2493 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2496 /* Compute the current contribution to load_avg by se, return any delta */
2497 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2499 long old_contrib
= se
->avg
.load_avg_contrib
;
2501 if (entity_is_task(se
)) {
2502 __update_task_entity_contrib(se
);
2504 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2505 __update_group_entity_contrib(se
);
2508 return se
->avg
.load_avg_contrib
- old_contrib
;
2511 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2514 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2515 cfs_rq
->blocked_load_avg
-= load_contrib
;
2517 cfs_rq
->blocked_load_avg
= 0;
2520 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2522 /* Update a sched_entity's runnable average */
2523 static inline void update_entity_load_avg(struct sched_entity
*se
,
2526 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2531 * For a group entity we need to use their owned cfs_rq_clock_task() in
2532 * case they are the parent of a throttled hierarchy.
2534 if (entity_is_task(se
))
2535 now
= cfs_rq_clock_task(cfs_rq
);
2537 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2539 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2542 contrib_delta
= __update_entity_load_avg_contrib(se
);
2548 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2550 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2554 * Decay the load contributed by all blocked children and account this so that
2555 * their contribution may appropriately discounted when they wake up.
2557 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2559 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2562 decays
= now
- cfs_rq
->last_decay
;
2563 if (!decays
&& !force_update
)
2566 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2567 unsigned long removed_load
;
2568 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2569 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2573 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2575 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2576 cfs_rq
->last_decay
= now
;
2579 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2582 /* Add the load generated by se into cfs_rq's child load-average */
2583 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2584 struct sched_entity
*se
,
2588 * We track migrations using entity decay_count <= 0, on a wake-up
2589 * migration we use a negative decay count to track the remote decays
2590 * accumulated while sleeping.
2592 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2593 * are seen by enqueue_entity_load_avg() as a migration with an already
2594 * constructed load_avg_contrib.
2596 if (unlikely(se
->avg
.decay_count
<= 0)) {
2597 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2598 if (se
->avg
.decay_count
) {
2600 * In a wake-up migration we have to approximate the
2601 * time sleeping. This is because we can't synchronize
2602 * clock_task between the two cpus, and it is not
2603 * guaranteed to be read-safe. Instead, we can
2604 * approximate this using our carried decays, which are
2605 * explicitly atomically readable.
2607 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2609 update_entity_load_avg(se
, 0);
2610 /* Indicate that we're now synchronized and on-rq */
2611 se
->avg
.decay_count
= 0;
2615 __synchronize_entity_decay(se
);
2618 /* migrated tasks did not contribute to our blocked load */
2620 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2621 update_entity_load_avg(se
, 0);
2624 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2625 /* we force update consideration on load-balancer moves */
2626 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2630 * Remove se's load from this cfs_rq child load-average, if the entity is
2631 * transitioning to a blocked state we track its projected decay using
2634 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2635 struct sched_entity
*se
,
2638 update_entity_load_avg(se
, 1);
2639 /* we force update consideration on load-balancer moves */
2640 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2642 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2644 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2645 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2646 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2650 * Update the rq's load with the elapsed running time before entering
2651 * idle. if the last scheduled task is not a CFS task, idle_enter will
2652 * be the only way to update the runnable statistic.
2654 void idle_enter_fair(struct rq
*this_rq
)
2656 update_rq_runnable_avg(this_rq
, 1);
2660 * Update the rq's load with the elapsed idle time before a task is
2661 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2662 * be the only way to update the runnable statistic.
2664 void idle_exit_fair(struct rq
*this_rq
)
2666 update_rq_runnable_avg(this_rq
, 0);
2669 static int idle_balance(struct rq
*this_rq
);
2671 #else /* CONFIG_SMP */
2673 static inline void update_entity_load_avg(struct sched_entity
*se
,
2674 int update_cfs_rq
) {}
2675 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2676 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2677 struct sched_entity
*se
,
2679 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2680 struct sched_entity
*se
,
2682 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2683 int force_update
) {}
2685 static inline int idle_balance(struct rq
*rq
)
2690 #endif /* CONFIG_SMP */
2692 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2694 #ifdef CONFIG_SCHEDSTATS
2695 struct task_struct
*tsk
= NULL
;
2697 if (entity_is_task(se
))
2700 if (se
->statistics
.sleep_start
) {
2701 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2706 if (unlikely(delta
> se
->statistics
.sleep_max
))
2707 se
->statistics
.sleep_max
= delta
;
2709 se
->statistics
.sleep_start
= 0;
2710 se
->statistics
.sum_sleep_runtime
+= delta
;
2713 account_scheduler_latency(tsk
, delta
>> 10, 1);
2714 trace_sched_stat_sleep(tsk
, delta
);
2717 if (se
->statistics
.block_start
) {
2718 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2723 if (unlikely(delta
> se
->statistics
.block_max
))
2724 se
->statistics
.block_max
= delta
;
2726 se
->statistics
.block_start
= 0;
2727 se
->statistics
.sum_sleep_runtime
+= delta
;
2730 if (tsk
->in_iowait
) {
2731 se
->statistics
.iowait_sum
+= delta
;
2732 se
->statistics
.iowait_count
++;
2733 trace_sched_stat_iowait(tsk
, delta
);
2736 trace_sched_stat_blocked(tsk
, delta
);
2739 * Blocking time is in units of nanosecs, so shift by
2740 * 20 to get a milliseconds-range estimation of the
2741 * amount of time that the task spent sleeping:
2743 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2744 profile_hits(SLEEP_PROFILING
,
2745 (void *)get_wchan(tsk
),
2748 account_scheduler_latency(tsk
, delta
>> 10, 0);
2754 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2756 #ifdef CONFIG_SCHED_DEBUG
2757 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2762 if (d
> 3*sysctl_sched_latency
)
2763 schedstat_inc(cfs_rq
, nr_spread_over
);
2768 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2770 u64 vruntime
= cfs_rq
->min_vruntime
;
2773 * The 'current' period is already promised to the current tasks,
2774 * however the extra weight of the new task will slow them down a
2775 * little, place the new task so that it fits in the slot that
2776 * stays open at the end.
2778 if (initial
&& sched_feat(START_DEBIT
))
2779 vruntime
+= sched_vslice(cfs_rq
, se
);
2781 /* sleeps up to a single latency don't count. */
2783 unsigned long thresh
= sysctl_sched_latency
;
2786 * Halve their sleep time's effect, to allow
2787 * for a gentler effect of sleepers:
2789 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2795 /* ensure we never gain time by being placed backwards. */
2796 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
2799 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
2802 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2805 * Update the normalized vruntime before updating min_vruntime
2806 * through calling update_curr().
2808 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
2809 se
->vruntime
+= cfs_rq
->min_vruntime
;
2812 * Update run-time statistics of the 'current'.
2814 update_curr(cfs_rq
);
2815 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
2816 account_entity_enqueue(cfs_rq
, se
);
2817 update_cfs_shares(cfs_rq
);
2819 if (flags
& ENQUEUE_WAKEUP
) {
2820 place_entity(cfs_rq
, se
, 0);
2821 enqueue_sleeper(cfs_rq
, se
);
2824 update_stats_enqueue(cfs_rq
, se
);
2825 check_spread(cfs_rq
, se
);
2826 if (se
!= cfs_rq
->curr
)
2827 __enqueue_entity(cfs_rq
, se
);
2830 if (cfs_rq
->nr_running
== 1) {
2831 list_add_leaf_cfs_rq(cfs_rq
);
2832 check_enqueue_throttle(cfs_rq
);
2836 static void __clear_buddies_last(struct sched_entity
*se
)
2838 for_each_sched_entity(se
) {
2839 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2840 if (cfs_rq
->last
!= se
)
2843 cfs_rq
->last
= NULL
;
2847 static void __clear_buddies_next(struct sched_entity
*se
)
2849 for_each_sched_entity(se
) {
2850 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2851 if (cfs_rq
->next
!= se
)
2854 cfs_rq
->next
= NULL
;
2858 static void __clear_buddies_skip(struct sched_entity
*se
)
2860 for_each_sched_entity(se
) {
2861 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2862 if (cfs_rq
->skip
!= se
)
2865 cfs_rq
->skip
= NULL
;
2869 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2871 if (cfs_rq
->last
== se
)
2872 __clear_buddies_last(se
);
2874 if (cfs_rq
->next
== se
)
2875 __clear_buddies_next(se
);
2877 if (cfs_rq
->skip
== se
)
2878 __clear_buddies_skip(se
);
2881 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2884 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2887 * Update run-time statistics of the 'current'.
2889 update_curr(cfs_rq
);
2890 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
2892 update_stats_dequeue(cfs_rq
, se
);
2893 if (flags
& DEQUEUE_SLEEP
) {
2894 #ifdef CONFIG_SCHEDSTATS
2895 if (entity_is_task(se
)) {
2896 struct task_struct
*tsk
= task_of(se
);
2898 if (tsk
->state
& TASK_INTERRUPTIBLE
)
2899 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
2900 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
2901 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
2906 clear_buddies(cfs_rq
, se
);
2908 if (se
!= cfs_rq
->curr
)
2909 __dequeue_entity(cfs_rq
, se
);
2911 account_entity_dequeue(cfs_rq
, se
);
2914 * Normalize the entity after updating the min_vruntime because the
2915 * update can refer to the ->curr item and we need to reflect this
2916 * movement in our normalized position.
2918 if (!(flags
& DEQUEUE_SLEEP
))
2919 se
->vruntime
-= cfs_rq
->min_vruntime
;
2921 /* return excess runtime on last dequeue */
2922 return_cfs_rq_runtime(cfs_rq
);
2924 update_min_vruntime(cfs_rq
);
2925 update_cfs_shares(cfs_rq
);
2929 * Preempt the current task with a newly woken task if needed:
2932 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2934 unsigned long ideal_runtime
, delta_exec
;
2935 struct sched_entity
*se
;
2938 ideal_runtime
= sched_slice(cfs_rq
, curr
);
2939 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
2940 if (delta_exec
> ideal_runtime
) {
2941 resched_curr(rq_of(cfs_rq
));
2943 * The current task ran long enough, ensure it doesn't get
2944 * re-elected due to buddy favours.
2946 clear_buddies(cfs_rq
, curr
);
2951 * Ensure that a task that missed wakeup preemption by a
2952 * narrow margin doesn't have to wait for a full slice.
2953 * This also mitigates buddy induced latencies under load.
2955 if (delta_exec
< sysctl_sched_min_granularity
)
2958 se
= __pick_first_entity(cfs_rq
);
2959 delta
= curr
->vruntime
- se
->vruntime
;
2964 if (delta
> ideal_runtime
)
2965 resched_curr(rq_of(cfs_rq
));
2969 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2971 /* 'current' is not kept within the tree. */
2974 * Any task has to be enqueued before it get to execute on
2975 * a CPU. So account for the time it spent waiting on the
2978 update_stats_wait_end(cfs_rq
, se
);
2979 __dequeue_entity(cfs_rq
, se
);
2982 update_stats_curr_start(cfs_rq
, se
);
2984 #ifdef CONFIG_SCHEDSTATS
2986 * Track our maximum slice length, if the CPU's load is at
2987 * least twice that of our own weight (i.e. dont track it
2988 * when there are only lesser-weight tasks around):
2990 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
2991 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
2992 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
2995 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
2999 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
3002 * Pick the next process, keeping these things in mind, in this order:
3003 * 1) keep things fair between processes/task groups
3004 * 2) pick the "next" process, since someone really wants that to run
3005 * 3) pick the "last" process, for cache locality
3006 * 4) do not run the "skip" process, if something else is available
3008 static struct sched_entity
*
3009 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3011 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
3012 struct sched_entity
*se
;
3015 * If curr is set we have to see if its left of the leftmost entity
3016 * still in the tree, provided there was anything in the tree at all.
3018 if (!left
|| (curr
&& entity_before(curr
, left
)))
3021 se
= left
; /* ideally we run the leftmost entity */
3024 * Avoid running the skip buddy, if running something else can
3025 * be done without getting too unfair.
3027 if (cfs_rq
->skip
== se
) {
3028 struct sched_entity
*second
;
3031 second
= __pick_first_entity(cfs_rq
);
3033 second
= __pick_next_entity(se
);
3034 if (!second
|| (curr
&& entity_before(curr
, second
)))
3038 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3043 * Prefer last buddy, try to return the CPU to a preempted task.
3045 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3049 * Someone really wants this to run. If it's not unfair, run it.
3051 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3054 clear_buddies(cfs_rq
, se
);
3059 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3061 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3064 * If still on the runqueue then deactivate_task()
3065 * was not called and update_curr() has to be done:
3068 update_curr(cfs_rq
);
3070 /* throttle cfs_rqs exceeding runtime */
3071 check_cfs_rq_runtime(cfs_rq
);
3073 check_spread(cfs_rq
, prev
);
3075 update_stats_wait_start(cfs_rq
, prev
);
3076 /* Put 'current' back into the tree. */
3077 __enqueue_entity(cfs_rq
, prev
);
3078 /* in !on_rq case, update occurred at dequeue */
3079 update_entity_load_avg(prev
, 1);
3081 cfs_rq
->curr
= NULL
;
3085 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3088 * Update run-time statistics of the 'current'.
3090 update_curr(cfs_rq
);
3093 * Ensure that runnable average is periodically updated.
3095 update_entity_load_avg(curr
, 1);
3096 update_cfs_rq_blocked_load(cfs_rq
, 1);
3097 update_cfs_shares(cfs_rq
);
3099 #ifdef CONFIG_SCHED_HRTICK
3101 * queued ticks are scheduled to match the slice, so don't bother
3102 * validating it and just reschedule.
3105 resched_curr(rq_of(cfs_rq
));
3109 * don't let the period tick interfere with the hrtick preemption
3111 if (!sched_feat(DOUBLE_TICK
) &&
3112 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3116 if (cfs_rq
->nr_running
> 1)
3117 check_preempt_tick(cfs_rq
, curr
);
3121 /**************************************************
3122 * CFS bandwidth control machinery
3125 #ifdef CONFIG_CFS_BANDWIDTH
3127 #ifdef HAVE_JUMP_LABEL
3128 static struct static_key __cfs_bandwidth_used
;
3130 static inline bool cfs_bandwidth_used(void)
3132 return static_key_false(&__cfs_bandwidth_used
);
3135 void cfs_bandwidth_usage_inc(void)
3137 static_key_slow_inc(&__cfs_bandwidth_used
);
3140 void cfs_bandwidth_usage_dec(void)
3142 static_key_slow_dec(&__cfs_bandwidth_used
);
3144 #else /* HAVE_JUMP_LABEL */
3145 static bool cfs_bandwidth_used(void)
3150 void cfs_bandwidth_usage_inc(void) {}
3151 void cfs_bandwidth_usage_dec(void) {}
3152 #endif /* HAVE_JUMP_LABEL */
3155 * default period for cfs group bandwidth.
3156 * default: 0.1s, units: nanoseconds
3158 static inline u64
default_cfs_period(void)
3160 return 100000000ULL;
3163 static inline u64
sched_cfs_bandwidth_slice(void)
3165 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3169 * Replenish runtime according to assigned quota and update expiration time.
3170 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3171 * additional synchronization around rq->lock.
3173 * requires cfs_b->lock
3175 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3179 if (cfs_b
->quota
== RUNTIME_INF
)
3182 now
= sched_clock_cpu(smp_processor_id());
3183 cfs_b
->runtime
= cfs_b
->quota
;
3184 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3187 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3189 return &tg
->cfs_bandwidth
;
3192 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3193 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3195 if (unlikely(cfs_rq
->throttle_count
))
3196 return cfs_rq
->throttled_clock_task
;
3198 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3201 /* returns 0 on failure to allocate runtime */
3202 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3204 struct task_group
*tg
= cfs_rq
->tg
;
3205 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3206 u64 amount
= 0, min_amount
, expires
;
3208 /* note: this is a positive sum as runtime_remaining <= 0 */
3209 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3211 raw_spin_lock(&cfs_b
->lock
);
3212 if (cfs_b
->quota
== RUNTIME_INF
)
3213 amount
= min_amount
;
3216 * If the bandwidth pool has become inactive, then at least one
3217 * period must have elapsed since the last consumption.
3218 * Refresh the global state and ensure bandwidth timer becomes
3221 if (!cfs_b
->timer_active
) {
3222 __refill_cfs_bandwidth_runtime(cfs_b
);
3223 __start_cfs_bandwidth(cfs_b
, false);
3226 if (cfs_b
->runtime
> 0) {
3227 amount
= min(cfs_b
->runtime
, min_amount
);
3228 cfs_b
->runtime
-= amount
;
3232 expires
= cfs_b
->runtime_expires
;
3233 raw_spin_unlock(&cfs_b
->lock
);
3235 cfs_rq
->runtime_remaining
+= amount
;
3237 * we may have advanced our local expiration to account for allowed
3238 * spread between our sched_clock and the one on which runtime was
3241 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3242 cfs_rq
->runtime_expires
= expires
;
3244 return cfs_rq
->runtime_remaining
> 0;
3248 * Note: This depends on the synchronization provided by sched_clock and the
3249 * fact that rq->clock snapshots this value.
3251 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3253 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3255 /* if the deadline is ahead of our clock, nothing to do */
3256 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3259 if (cfs_rq
->runtime_remaining
< 0)
3263 * If the local deadline has passed we have to consider the
3264 * possibility that our sched_clock is 'fast' and the global deadline
3265 * has not truly expired.
3267 * Fortunately we can check determine whether this the case by checking
3268 * whether the global deadline has advanced. It is valid to compare
3269 * cfs_b->runtime_expires without any locks since we only care about
3270 * exact equality, so a partial write will still work.
3273 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
3274 /* extend local deadline, drift is bounded above by 2 ticks */
3275 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3277 /* global deadline is ahead, expiration has passed */
3278 cfs_rq
->runtime_remaining
= 0;
3282 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3284 /* dock delta_exec before expiring quota (as it could span periods) */
3285 cfs_rq
->runtime_remaining
-= delta_exec
;
3286 expire_cfs_rq_runtime(cfs_rq
);
3288 if (likely(cfs_rq
->runtime_remaining
> 0))
3292 * if we're unable to extend our runtime we resched so that the active
3293 * hierarchy can be throttled
3295 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3296 resched_curr(rq_of(cfs_rq
));
3299 static __always_inline
3300 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3302 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3305 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3308 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3310 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3313 /* check whether cfs_rq, or any parent, is throttled */
3314 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3316 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3320 * Ensure that neither of the group entities corresponding to src_cpu or
3321 * dest_cpu are members of a throttled hierarchy when performing group
3322 * load-balance operations.
3324 static inline int throttled_lb_pair(struct task_group
*tg
,
3325 int src_cpu
, int dest_cpu
)
3327 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3329 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3330 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3332 return throttled_hierarchy(src_cfs_rq
) ||
3333 throttled_hierarchy(dest_cfs_rq
);
3336 /* updated child weight may affect parent so we have to do this bottom up */
3337 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3339 struct rq
*rq
= data
;
3340 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3342 cfs_rq
->throttle_count
--;
3344 if (!cfs_rq
->throttle_count
) {
3345 /* adjust cfs_rq_clock_task() */
3346 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3347 cfs_rq
->throttled_clock_task
;
3354 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3356 struct rq
*rq
= data
;
3357 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3359 /* group is entering throttled state, stop time */
3360 if (!cfs_rq
->throttle_count
)
3361 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3362 cfs_rq
->throttle_count
++;
3367 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3369 struct rq
*rq
= rq_of(cfs_rq
);
3370 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3371 struct sched_entity
*se
;
3372 long task_delta
, dequeue
= 1;
3374 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3376 /* freeze hierarchy runnable averages while throttled */
3378 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3381 task_delta
= cfs_rq
->h_nr_running
;
3382 for_each_sched_entity(se
) {
3383 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3384 /* throttled entity or throttle-on-deactivate */
3389 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3390 qcfs_rq
->h_nr_running
-= task_delta
;
3392 if (qcfs_rq
->load
.weight
)
3397 sub_nr_running(rq
, task_delta
);
3399 cfs_rq
->throttled
= 1;
3400 cfs_rq
->throttled_clock
= rq_clock(rq
);
3401 raw_spin_lock(&cfs_b
->lock
);
3403 * Add to the _head_ of the list, so that an already-started
3404 * distribute_cfs_runtime will not see us
3406 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3407 if (!cfs_b
->timer_active
)
3408 __start_cfs_bandwidth(cfs_b
, false);
3409 raw_spin_unlock(&cfs_b
->lock
);
3412 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3414 struct rq
*rq
= rq_of(cfs_rq
);
3415 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3416 struct sched_entity
*se
;
3420 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3422 cfs_rq
->throttled
= 0;
3424 update_rq_clock(rq
);
3426 raw_spin_lock(&cfs_b
->lock
);
3427 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3428 list_del_rcu(&cfs_rq
->throttled_list
);
3429 raw_spin_unlock(&cfs_b
->lock
);
3431 /* update hierarchical throttle state */
3432 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3434 if (!cfs_rq
->load
.weight
)
3437 task_delta
= cfs_rq
->h_nr_running
;
3438 for_each_sched_entity(se
) {
3442 cfs_rq
= cfs_rq_of(se
);
3444 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3445 cfs_rq
->h_nr_running
+= task_delta
;
3447 if (cfs_rq_throttled(cfs_rq
))
3452 add_nr_running(rq
, task_delta
);
3454 /* determine whether we need to wake up potentially idle cpu */
3455 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3459 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3460 u64 remaining
, u64 expires
)
3462 struct cfs_rq
*cfs_rq
;
3464 u64 starting_runtime
= remaining
;
3467 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3469 struct rq
*rq
= rq_of(cfs_rq
);
3471 raw_spin_lock(&rq
->lock
);
3472 if (!cfs_rq_throttled(cfs_rq
))
3475 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3476 if (runtime
> remaining
)
3477 runtime
= remaining
;
3478 remaining
-= runtime
;
3480 cfs_rq
->runtime_remaining
+= runtime
;
3481 cfs_rq
->runtime_expires
= expires
;
3483 /* we check whether we're throttled above */
3484 if (cfs_rq
->runtime_remaining
> 0)
3485 unthrottle_cfs_rq(cfs_rq
);
3488 raw_spin_unlock(&rq
->lock
);
3495 return starting_runtime
- remaining
;
3499 * Responsible for refilling a task_group's bandwidth and unthrottling its
3500 * cfs_rqs as appropriate. If there has been no activity within the last
3501 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3502 * used to track this state.
3504 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3506 u64 runtime
, runtime_expires
;
3509 /* no need to continue the timer with no bandwidth constraint */
3510 if (cfs_b
->quota
== RUNTIME_INF
)
3511 goto out_deactivate
;
3513 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3514 cfs_b
->nr_periods
+= overrun
;
3517 * idle depends on !throttled (for the case of a large deficit), and if
3518 * we're going inactive then everything else can be deferred
3520 if (cfs_b
->idle
&& !throttled
)
3521 goto out_deactivate
;
3524 * if we have relooped after returning idle once, we need to update our
3525 * status as actually running, so that other cpus doing
3526 * __start_cfs_bandwidth will stop trying to cancel us.
3528 cfs_b
->timer_active
= 1;
3530 __refill_cfs_bandwidth_runtime(cfs_b
);
3533 /* mark as potentially idle for the upcoming period */
3538 /* account preceding periods in which throttling occurred */
3539 cfs_b
->nr_throttled
+= overrun
;
3541 runtime_expires
= cfs_b
->runtime_expires
;
3544 * This check is repeated as we are holding onto the new bandwidth while
3545 * we unthrottle. This can potentially race with an unthrottled group
3546 * trying to acquire new bandwidth from the global pool. This can result
3547 * in us over-using our runtime if it is all used during this loop, but
3548 * only by limited amounts in that extreme case.
3550 while (throttled
&& cfs_b
->runtime
> 0) {
3551 runtime
= cfs_b
->runtime
;
3552 raw_spin_unlock(&cfs_b
->lock
);
3553 /* we can't nest cfs_b->lock while distributing bandwidth */
3554 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3556 raw_spin_lock(&cfs_b
->lock
);
3558 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3560 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3564 * While we are ensured activity in the period following an
3565 * unthrottle, this also covers the case in which the new bandwidth is
3566 * insufficient to cover the existing bandwidth deficit. (Forcing the
3567 * timer to remain active while there are any throttled entities.)
3574 cfs_b
->timer_active
= 0;
3578 /* a cfs_rq won't donate quota below this amount */
3579 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3580 /* minimum remaining period time to redistribute slack quota */
3581 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3582 /* how long we wait to gather additional slack before distributing */
3583 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3586 * Are we near the end of the current quota period?
3588 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3589 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3590 * migrate_hrtimers, base is never cleared, so we are fine.
3592 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3594 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3597 /* if the call-back is running a quota refresh is already occurring */
3598 if (hrtimer_callback_running(refresh_timer
))
3601 /* is a quota refresh about to occur? */
3602 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3603 if (remaining
< min_expire
)
3609 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3611 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3613 /* if there's a quota refresh soon don't bother with slack */
3614 if (runtime_refresh_within(cfs_b
, min_left
))
3617 start_bandwidth_timer(&cfs_b
->slack_timer
,
3618 ns_to_ktime(cfs_bandwidth_slack_period
));
3621 /* we know any runtime found here is valid as update_curr() precedes return */
3622 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3624 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3625 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3627 if (slack_runtime
<= 0)
3630 raw_spin_lock(&cfs_b
->lock
);
3631 if (cfs_b
->quota
!= RUNTIME_INF
&&
3632 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3633 cfs_b
->runtime
+= slack_runtime
;
3635 /* we are under rq->lock, defer unthrottling using a timer */
3636 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3637 !list_empty(&cfs_b
->throttled_cfs_rq
))
3638 start_cfs_slack_bandwidth(cfs_b
);
3640 raw_spin_unlock(&cfs_b
->lock
);
3642 /* even if it's not valid for return we don't want to try again */
3643 cfs_rq
->runtime_remaining
-= slack_runtime
;
3646 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3648 if (!cfs_bandwidth_used())
3651 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3654 __return_cfs_rq_runtime(cfs_rq
);
3658 * This is done with a timer (instead of inline with bandwidth return) since
3659 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3661 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3663 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3666 /* confirm we're still not at a refresh boundary */
3667 raw_spin_lock(&cfs_b
->lock
);
3668 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
3669 raw_spin_unlock(&cfs_b
->lock
);
3673 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
3674 runtime
= cfs_b
->runtime
;
3676 expires
= cfs_b
->runtime_expires
;
3677 raw_spin_unlock(&cfs_b
->lock
);
3682 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3684 raw_spin_lock(&cfs_b
->lock
);
3685 if (expires
== cfs_b
->runtime_expires
)
3686 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3687 raw_spin_unlock(&cfs_b
->lock
);
3691 * When a group wakes up we want to make sure that its quota is not already
3692 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3693 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3695 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3697 if (!cfs_bandwidth_used())
3700 /* an active group must be handled by the update_curr()->put() path */
3701 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3704 /* ensure the group is not already throttled */
3705 if (cfs_rq_throttled(cfs_rq
))
3708 /* update runtime allocation */
3709 account_cfs_rq_runtime(cfs_rq
, 0);
3710 if (cfs_rq
->runtime_remaining
<= 0)
3711 throttle_cfs_rq(cfs_rq
);
3714 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3715 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3717 if (!cfs_bandwidth_used())
3720 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3724 * it's possible for a throttled entity to be forced into a running
3725 * state (e.g. set_curr_task), in this case we're finished.
3727 if (cfs_rq_throttled(cfs_rq
))
3730 throttle_cfs_rq(cfs_rq
);
3734 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3736 struct cfs_bandwidth
*cfs_b
=
3737 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3738 do_sched_cfs_slack_timer(cfs_b
);
3740 return HRTIMER_NORESTART
;
3743 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3745 struct cfs_bandwidth
*cfs_b
=
3746 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3751 raw_spin_lock(&cfs_b
->lock
);
3753 now
= hrtimer_cb_get_time(timer
);
3754 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3759 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3761 raw_spin_unlock(&cfs_b
->lock
);
3763 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3766 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3768 raw_spin_lock_init(&cfs_b
->lock
);
3770 cfs_b
->quota
= RUNTIME_INF
;
3771 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3773 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3774 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3775 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3776 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3777 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3780 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3782 cfs_rq
->runtime_enabled
= 0;
3783 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3786 /* requires cfs_b->lock, may release to reprogram timer */
3787 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
, bool force
)
3790 * The timer may be active because we're trying to set a new bandwidth
3791 * period or because we're racing with the tear-down path
3792 * (timer_active==0 becomes visible before the hrtimer call-back
3793 * terminates). In either case we ensure that it's re-programmed
3795 while (unlikely(hrtimer_active(&cfs_b
->period_timer
)) &&
3796 hrtimer_try_to_cancel(&cfs_b
->period_timer
) < 0) {
3797 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3798 raw_spin_unlock(&cfs_b
->lock
);
3800 raw_spin_lock(&cfs_b
->lock
);
3801 /* if someone else restarted the timer then we're done */
3802 if (!force
&& cfs_b
->timer_active
)
3806 cfs_b
->timer_active
= 1;
3807 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
3810 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3812 hrtimer_cancel(&cfs_b
->period_timer
);
3813 hrtimer_cancel(&cfs_b
->slack_timer
);
3816 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
3818 struct cfs_rq
*cfs_rq
;
3820 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3821 struct cfs_bandwidth
*cfs_b
= &cfs_rq
->tg
->cfs_bandwidth
;
3823 raw_spin_lock(&cfs_b
->lock
);
3824 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
3825 raw_spin_unlock(&cfs_b
->lock
);
3829 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
3831 struct cfs_rq
*cfs_rq
;
3833 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3834 if (!cfs_rq
->runtime_enabled
)
3838 * clock_task is not advancing so we just need to make sure
3839 * there's some valid quota amount
3841 cfs_rq
->runtime_remaining
= 1;
3843 * Offline rq is schedulable till cpu is completely disabled
3844 * in take_cpu_down(), so we prevent new cfs throttling here.
3846 cfs_rq
->runtime_enabled
= 0;
3848 if (cfs_rq_throttled(cfs_rq
))
3849 unthrottle_cfs_rq(cfs_rq
);
3853 #else /* CONFIG_CFS_BANDWIDTH */
3854 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3856 return rq_clock_task(rq_of(cfs_rq
));
3859 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
3860 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
3861 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
3862 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3864 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3869 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3874 static inline int throttled_lb_pair(struct task_group
*tg
,
3875 int src_cpu
, int dest_cpu
)
3880 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3882 #ifdef CONFIG_FAIR_GROUP_SCHED
3883 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3886 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3890 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3891 static inline void update_runtime_enabled(struct rq
*rq
) {}
3892 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
3894 #endif /* CONFIG_CFS_BANDWIDTH */
3896 /**************************************************
3897 * CFS operations on tasks:
3900 #ifdef CONFIG_SCHED_HRTICK
3901 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3903 struct sched_entity
*se
= &p
->se
;
3904 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3906 WARN_ON(task_rq(p
) != rq
);
3908 if (cfs_rq
->nr_running
> 1) {
3909 u64 slice
= sched_slice(cfs_rq
, se
);
3910 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
3911 s64 delta
= slice
- ran
;
3918 hrtick_start(rq
, delta
);
3923 * called from enqueue/dequeue and updates the hrtick when the
3924 * current task is from our class and nr_running is low enough
3927 static void hrtick_update(struct rq
*rq
)
3929 struct task_struct
*curr
= rq
->curr
;
3931 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
3934 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
3935 hrtick_start_fair(rq
, curr
);
3937 #else /* !CONFIG_SCHED_HRTICK */
3939 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3943 static inline void hrtick_update(struct rq
*rq
)
3949 * The enqueue_task method is called before nr_running is
3950 * increased. Here we update the fair scheduling stats and
3951 * then put the task into the rbtree:
3954 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3956 struct cfs_rq
*cfs_rq
;
3957 struct sched_entity
*se
= &p
->se
;
3959 for_each_sched_entity(se
) {
3962 cfs_rq
= cfs_rq_of(se
);
3963 enqueue_entity(cfs_rq
, se
, flags
);
3966 * end evaluation on encountering a throttled cfs_rq
3968 * note: in the case of encountering a throttled cfs_rq we will
3969 * post the final h_nr_running increment below.
3971 if (cfs_rq_throttled(cfs_rq
))
3973 cfs_rq
->h_nr_running
++;
3975 flags
= ENQUEUE_WAKEUP
;
3978 for_each_sched_entity(se
) {
3979 cfs_rq
= cfs_rq_of(se
);
3980 cfs_rq
->h_nr_running
++;
3982 if (cfs_rq_throttled(cfs_rq
))
3985 update_cfs_shares(cfs_rq
);
3986 update_entity_load_avg(se
, 1);
3990 update_rq_runnable_avg(rq
, rq
->nr_running
);
3991 add_nr_running(rq
, 1);
3996 static void set_next_buddy(struct sched_entity
*se
);
3999 * The dequeue_task method is called before nr_running is
4000 * decreased. We remove the task from the rbtree and
4001 * update the fair scheduling stats:
4003 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4005 struct cfs_rq
*cfs_rq
;
4006 struct sched_entity
*se
= &p
->se
;
4007 int task_sleep
= flags
& DEQUEUE_SLEEP
;
4009 for_each_sched_entity(se
) {
4010 cfs_rq
= cfs_rq_of(se
);
4011 dequeue_entity(cfs_rq
, se
, flags
);
4014 * end evaluation on encountering a throttled cfs_rq
4016 * note: in the case of encountering a throttled cfs_rq we will
4017 * post the final h_nr_running decrement below.
4019 if (cfs_rq_throttled(cfs_rq
))
4021 cfs_rq
->h_nr_running
--;
4023 /* Don't dequeue parent if it has other entities besides us */
4024 if (cfs_rq
->load
.weight
) {
4026 * Bias pick_next to pick a task from this cfs_rq, as
4027 * p is sleeping when it is within its sched_slice.
4029 if (task_sleep
&& parent_entity(se
))
4030 set_next_buddy(parent_entity(se
));
4032 /* avoid re-evaluating load for this entity */
4033 se
= parent_entity(se
);
4036 flags
|= DEQUEUE_SLEEP
;
4039 for_each_sched_entity(se
) {
4040 cfs_rq
= cfs_rq_of(se
);
4041 cfs_rq
->h_nr_running
--;
4043 if (cfs_rq_throttled(cfs_rq
))
4046 update_cfs_shares(cfs_rq
);
4047 update_entity_load_avg(se
, 1);
4051 sub_nr_running(rq
, 1);
4052 update_rq_runnable_avg(rq
, 1);
4058 /* Used instead of source_load when we know the type == 0 */
4059 static unsigned long weighted_cpuload(const int cpu
)
4061 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
4065 * Return a low guess at the load of a migration-source cpu weighted
4066 * according to the scheduling class and "nice" value.
4068 * We want to under-estimate the load of migration sources, to
4069 * balance conservatively.
4071 static unsigned long source_load(int cpu
, int type
)
4073 struct rq
*rq
= cpu_rq(cpu
);
4074 unsigned long total
= weighted_cpuload(cpu
);
4076 if (type
== 0 || !sched_feat(LB_BIAS
))
4079 return min(rq
->cpu_load
[type
-1], total
);
4083 * Return a high guess at the load of a migration-target cpu weighted
4084 * according to the scheduling class and "nice" value.
4086 static unsigned long target_load(int cpu
, int type
)
4088 struct rq
*rq
= cpu_rq(cpu
);
4089 unsigned long total
= weighted_cpuload(cpu
);
4091 if (type
== 0 || !sched_feat(LB_BIAS
))
4094 return max(rq
->cpu_load
[type
-1], total
);
4097 static unsigned long capacity_of(int cpu
)
4099 return cpu_rq(cpu
)->cpu_capacity
;
4102 static unsigned long cpu_avg_load_per_task(int cpu
)
4104 struct rq
*rq
= cpu_rq(cpu
);
4105 unsigned long nr_running
= ACCESS_ONCE(rq
->cfs
.h_nr_running
);
4106 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
4109 return load_avg
/ nr_running
;
4114 static void record_wakee(struct task_struct
*p
)
4117 * Rough decay (wiping) for cost saving, don't worry
4118 * about the boundary, really active task won't care
4121 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
4122 current
->wakee_flips
>>= 1;
4123 current
->wakee_flip_decay_ts
= jiffies
;
4126 if (current
->last_wakee
!= p
) {
4127 current
->last_wakee
= p
;
4128 current
->wakee_flips
++;
4132 static void task_waking_fair(struct task_struct
*p
)
4134 struct sched_entity
*se
= &p
->se
;
4135 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4138 #ifndef CONFIG_64BIT
4139 u64 min_vruntime_copy
;
4142 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
4144 min_vruntime
= cfs_rq
->min_vruntime
;
4145 } while (min_vruntime
!= min_vruntime_copy
);
4147 min_vruntime
= cfs_rq
->min_vruntime
;
4150 se
->vruntime
-= min_vruntime
;
4154 #ifdef CONFIG_FAIR_GROUP_SCHED
4156 * effective_load() calculates the load change as seen from the root_task_group
4158 * Adding load to a group doesn't make a group heavier, but can cause movement
4159 * of group shares between cpus. Assuming the shares were perfectly aligned one
4160 * can calculate the shift in shares.
4162 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4163 * on this @cpu and results in a total addition (subtraction) of @wg to the
4164 * total group weight.
4166 * Given a runqueue weight distribution (rw_i) we can compute a shares
4167 * distribution (s_i) using:
4169 * s_i = rw_i / \Sum rw_j (1)
4171 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4172 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4173 * shares distribution (s_i):
4175 * rw_i = { 2, 4, 1, 0 }
4176 * s_i = { 2/7, 4/7, 1/7, 0 }
4178 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4179 * task used to run on and the CPU the waker is running on), we need to
4180 * compute the effect of waking a task on either CPU and, in case of a sync
4181 * wakeup, compute the effect of the current task going to sleep.
4183 * So for a change of @wl to the local @cpu with an overall group weight change
4184 * of @wl we can compute the new shares distribution (s'_i) using:
4186 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4188 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4189 * differences in waking a task to CPU 0. The additional task changes the
4190 * weight and shares distributions like:
4192 * rw'_i = { 3, 4, 1, 0 }
4193 * s'_i = { 3/8, 4/8, 1/8, 0 }
4195 * We can then compute the difference in effective weight by using:
4197 * dw_i = S * (s'_i - s_i) (3)
4199 * Where 'S' is the group weight as seen by its parent.
4201 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4202 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4203 * 4/7) times the weight of the group.
4205 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4207 struct sched_entity
*se
= tg
->se
[cpu
];
4209 if (!tg
->parent
) /* the trivial, non-cgroup case */
4212 for_each_sched_entity(se
) {
4218 * W = @wg + \Sum rw_j
4220 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4225 w
= se
->my_q
->load
.weight
+ wl
;
4228 * wl = S * s'_i; see (2)
4231 wl
= (w
* tg
->shares
) / W
;
4236 * Per the above, wl is the new se->load.weight value; since
4237 * those are clipped to [MIN_SHARES, ...) do so now. See
4238 * calc_cfs_shares().
4240 if (wl
< MIN_SHARES
)
4244 * wl = dw_i = S * (s'_i - s_i); see (3)
4246 wl
-= se
->load
.weight
;
4249 * Recursively apply this logic to all parent groups to compute
4250 * the final effective load change on the root group. Since
4251 * only the @tg group gets extra weight, all parent groups can
4252 * only redistribute existing shares. @wl is the shift in shares
4253 * resulting from this level per the above.
4262 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4269 static int wake_wide(struct task_struct
*p
)
4271 int factor
= this_cpu_read(sd_llc_size
);
4274 * Yeah, it's the switching-frequency, could means many wakee or
4275 * rapidly switch, use factor here will just help to automatically
4276 * adjust the loose-degree, so bigger node will lead to more pull.
4278 if (p
->wakee_flips
> factor
) {
4280 * wakee is somewhat hot, it needs certain amount of cpu
4281 * resource, so if waker is far more hot, prefer to leave
4284 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
4291 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4293 s64 this_load
, load
;
4294 s64 this_eff_load
, prev_eff_load
;
4295 int idx
, this_cpu
, prev_cpu
;
4296 struct task_group
*tg
;
4297 unsigned long weight
;
4301 * If we wake multiple tasks be careful to not bounce
4302 * ourselves around too much.
4308 this_cpu
= smp_processor_id();
4309 prev_cpu
= task_cpu(p
);
4310 load
= source_load(prev_cpu
, idx
);
4311 this_load
= target_load(this_cpu
, idx
);
4314 * If sync wakeup then subtract the (maximum possible)
4315 * effect of the currently running task from the load
4316 * of the current CPU:
4319 tg
= task_group(current
);
4320 weight
= current
->se
.load
.weight
;
4322 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4323 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4327 weight
= p
->se
.load
.weight
;
4330 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4331 * due to the sync cause above having dropped this_load to 0, we'll
4332 * always have an imbalance, but there's really nothing you can do
4333 * about that, so that's good too.
4335 * Otherwise check if either cpus are near enough in load to allow this
4336 * task to be woken on this_cpu.
4338 this_eff_load
= 100;
4339 this_eff_load
*= capacity_of(prev_cpu
);
4341 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4342 prev_eff_load
*= capacity_of(this_cpu
);
4344 if (this_load
> 0) {
4345 this_eff_load
*= this_load
+
4346 effective_load(tg
, this_cpu
, weight
, weight
);
4348 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4351 balanced
= this_eff_load
<= prev_eff_load
;
4353 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4358 schedstat_inc(sd
, ttwu_move_affine
);
4359 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4365 * find_idlest_group finds and returns the least busy CPU group within the
4368 static struct sched_group
*
4369 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4370 int this_cpu
, int sd_flag
)
4372 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4373 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4374 int load_idx
= sd
->forkexec_idx
;
4375 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4377 if (sd_flag
& SD_BALANCE_WAKE
)
4378 load_idx
= sd
->wake_idx
;
4381 unsigned long load
, avg_load
;
4385 /* Skip over this group if it has no CPUs allowed */
4386 if (!cpumask_intersects(sched_group_cpus(group
),
4387 tsk_cpus_allowed(p
)))
4390 local_group
= cpumask_test_cpu(this_cpu
,
4391 sched_group_cpus(group
));
4393 /* Tally up the load of all CPUs in the group */
4396 for_each_cpu(i
, sched_group_cpus(group
)) {
4397 /* Bias balancing toward cpus of our domain */
4399 load
= source_load(i
, load_idx
);
4401 load
= target_load(i
, load_idx
);
4406 /* Adjust by relative CPU capacity of the group */
4407 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) / group
->sgc
->capacity
;
4410 this_load
= avg_load
;
4411 } else if (avg_load
< min_load
) {
4412 min_load
= avg_load
;
4415 } while (group
= group
->next
, group
!= sd
->groups
);
4417 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4423 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4426 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4428 unsigned long load
, min_load
= ULONG_MAX
;
4429 unsigned int min_exit_latency
= UINT_MAX
;
4430 u64 latest_idle_timestamp
= 0;
4431 int least_loaded_cpu
= this_cpu
;
4432 int shallowest_idle_cpu
= -1;
4435 /* Traverse only the allowed CPUs */
4436 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4438 struct rq
*rq
= cpu_rq(i
);
4439 struct cpuidle_state
*idle
= idle_get_state(rq
);
4440 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
4442 * We give priority to a CPU whose idle state
4443 * has the smallest exit latency irrespective
4444 * of any idle timestamp.
4446 min_exit_latency
= idle
->exit_latency
;
4447 latest_idle_timestamp
= rq
->idle_stamp
;
4448 shallowest_idle_cpu
= i
;
4449 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
4450 rq
->idle_stamp
> latest_idle_timestamp
) {
4452 * If equal or no active idle state, then
4453 * the most recently idled CPU might have
4456 latest_idle_timestamp
= rq
->idle_stamp
;
4457 shallowest_idle_cpu
= i
;
4460 load
= weighted_cpuload(i
);
4461 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4463 least_loaded_cpu
= i
;
4468 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
4472 * Try and locate an idle CPU in the sched_domain.
4474 static int select_idle_sibling(struct task_struct
*p
, int target
)
4476 struct sched_domain
*sd
;
4477 struct sched_group
*sg
;
4478 int i
= task_cpu(p
);
4480 if (idle_cpu(target
))
4484 * If the prevous cpu is cache affine and idle, don't be stupid.
4486 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4490 * Otherwise, iterate the domains and find an elegible idle cpu.
4492 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4493 for_each_lower_domain(sd
) {
4496 if (!cpumask_intersects(sched_group_cpus(sg
),
4497 tsk_cpus_allowed(p
)))
4500 for_each_cpu(i
, sched_group_cpus(sg
)) {
4501 if (i
== target
|| !idle_cpu(i
))
4505 target
= cpumask_first_and(sched_group_cpus(sg
),
4506 tsk_cpus_allowed(p
));
4510 } while (sg
!= sd
->groups
);
4517 * select_task_rq_fair: Select target runqueue for the waking task in domains
4518 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4519 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4521 * Balances load by selecting the idlest cpu in the idlest group, or under
4522 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4524 * Returns the target cpu number.
4526 * preempt must be disabled.
4529 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4531 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4532 int cpu
= smp_processor_id();
4534 int want_affine
= 0;
4535 int sync
= wake_flags
& WF_SYNC
;
4537 if (p
->nr_cpus_allowed
== 1)
4540 if (sd_flag
& SD_BALANCE_WAKE
)
4541 want_affine
= cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
));
4544 for_each_domain(cpu
, tmp
) {
4545 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4549 * If both cpu and prev_cpu are part of this domain,
4550 * cpu is a valid SD_WAKE_AFFINE target.
4552 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4553 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4558 if (tmp
->flags
& sd_flag
)
4562 if (affine_sd
&& cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4565 if (sd_flag
& SD_BALANCE_WAKE
) {
4566 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4571 struct sched_group
*group
;
4574 if (!(sd
->flags
& sd_flag
)) {
4579 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
4585 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4586 if (new_cpu
== -1 || new_cpu
== cpu
) {
4587 /* Now try balancing at a lower domain level of cpu */
4592 /* Now try balancing at a lower domain level of new_cpu */
4594 weight
= sd
->span_weight
;
4596 for_each_domain(cpu
, tmp
) {
4597 if (weight
<= tmp
->span_weight
)
4599 if (tmp
->flags
& sd_flag
)
4602 /* while loop will break here if sd == NULL */
4611 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4612 * cfs_rq_of(p) references at time of call are still valid and identify the
4613 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4614 * other assumptions, including the state of rq->lock, should be made.
4617 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4619 struct sched_entity
*se
= &p
->se
;
4620 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4623 * Load tracking: accumulate removed load so that it can be processed
4624 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4625 * to blocked load iff they have a positive decay-count. It can never
4626 * be negative here since on-rq tasks have decay-count == 0.
4628 if (se
->avg
.decay_count
) {
4629 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4630 atomic_long_add(se
->avg
.load_avg_contrib
,
4631 &cfs_rq
->removed_load
);
4634 /* We have migrated, no longer consider this task hot */
4637 #endif /* CONFIG_SMP */
4639 static unsigned long
4640 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4642 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4645 * Since its curr running now, convert the gran from real-time
4646 * to virtual-time in his units.
4648 * By using 'se' instead of 'curr' we penalize light tasks, so
4649 * they get preempted easier. That is, if 'se' < 'curr' then
4650 * the resulting gran will be larger, therefore penalizing the
4651 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4652 * be smaller, again penalizing the lighter task.
4654 * This is especially important for buddies when the leftmost
4655 * task is higher priority than the buddy.
4657 return calc_delta_fair(gran
, se
);
4661 * Should 'se' preempt 'curr'.
4675 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4677 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4682 gran
= wakeup_gran(curr
, se
);
4689 static void set_last_buddy(struct sched_entity
*se
)
4691 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4694 for_each_sched_entity(se
)
4695 cfs_rq_of(se
)->last
= se
;
4698 static void set_next_buddy(struct sched_entity
*se
)
4700 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4703 for_each_sched_entity(se
)
4704 cfs_rq_of(se
)->next
= se
;
4707 static void set_skip_buddy(struct sched_entity
*se
)
4709 for_each_sched_entity(se
)
4710 cfs_rq_of(se
)->skip
= se
;
4714 * Preempt the current task with a newly woken task if needed:
4716 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4718 struct task_struct
*curr
= rq
->curr
;
4719 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4720 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4721 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4722 int next_buddy_marked
= 0;
4724 if (unlikely(se
== pse
))
4728 * This is possible from callers such as attach_tasks(), in which we
4729 * unconditionally check_prempt_curr() after an enqueue (which may have
4730 * lead to a throttle). This both saves work and prevents false
4731 * next-buddy nomination below.
4733 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4736 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4737 set_next_buddy(pse
);
4738 next_buddy_marked
= 1;
4742 * We can come here with TIF_NEED_RESCHED already set from new task
4745 * Note: this also catches the edge-case of curr being in a throttled
4746 * group (e.g. via set_curr_task), since update_curr() (in the
4747 * enqueue of curr) will have resulted in resched being set. This
4748 * prevents us from potentially nominating it as a false LAST_BUDDY
4751 if (test_tsk_need_resched(curr
))
4754 /* Idle tasks are by definition preempted by non-idle tasks. */
4755 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4756 likely(p
->policy
!= SCHED_IDLE
))
4760 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4761 * is driven by the tick):
4763 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4766 find_matching_se(&se
, &pse
);
4767 update_curr(cfs_rq_of(se
));
4769 if (wakeup_preempt_entity(se
, pse
) == 1) {
4771 * Bias pick_next to pick the sched entity that is
4772 * triggering this preemption.
4774 if (!next_buddy_marked
)
4775 set_next_buddy(pse
);
4784 * Only set the backward buddy when the current task is still
4785 * on the rq. This can happen when a wakeup gets interleaved
4786 * with schedule on the ->pre_schedule() or idle_balance()
4787 * point, either of which can * drop the rq lock.
4789 * Also, during early boot the idle thread is in the fair class,
4790 * for obvious reasons its a bad idea to schedule back to it.
4792 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4795 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
4799 static struct task_struct
*
4800 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4802 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
4803 struct sched_entity
*se
;
4804 struct task_struct
*p
;
4808 #ifdef CONFIG_FAIR_GROUP_SCHED
4809 if (!cfs_rq
->nr_running
)
4812 if (prev
->sched_class
!= &fair_sched_class
)
4816 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4817 * likely that a next task is from the same cgroup as the current.
4819 * Therefore attempt to avoid putting and setting the entire cgroup
4820 * hierarchy, only change the part that actually changes.
4824 struct sched_entity
*curr
= cfs_rq
->curr
;
4827 * Since we got here without doing put_prev_entity() we also
4828 * have to consider cfs_rq->curr. If it is still a runnable
4829 * entity, update_curr() will update its vruntime, otherwise
4830 * forget we've ever seen it.
4832 if (curr
&& curr
->on_rq
)
4833 update_curr(cfs_rq
);
4838 * This call to check_cfs_rq_runtime() will do the throttle and
4839 * dequeue its entity in the parent(s). Therefore the 'simple'
4840 * nr_running test will indeed be correct.
4842 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
4845 se
= pick_next_entity(cfs_rq
, curr
);
4846 cfs_rq
= group_cfs_rq(se
);
4852 * Since we haven't yet done put_prev_entity and if the selected task
4853 * is a different task than we started out with, try and touch the
4854 * least amount of cfs_rqs.
4857 struct sched_entity
*pse
= &prev
->se
;
4859 while (!(cfs_rq
= is_same_group(se
, pse
))) {
4860 int se_depth
= se
->depth
;
4861 int pse_depth
= pse
->depth
;
4863 if (se_depth
<= pse_depth
) {
4864 put_prev_entity(cfs_rq_of(pse
), pse
);
4865 pse
= parent_entity(pse
);
4867 if (se_depth
>= pse_depth
) {
4868 set_next_entity(cfs_rq_of(se
), se
);
4869 se
= parent_entity(se
);
4873 put_prev_entity(cfs_rq
, pse
);
4874 set_next_entity(cfs_rq
, se
);
4877 if (hrtick_enabled(rq
))
4878 hrtick_start_fair(rq
, p
);
4885 if (!cfs_rq
->nr_running
)
4888 put_prev_task(rq
, prev
);
4891 se
= pick_next_entity(cfs_rq
, NULL
);
4892 set_next_entity(cfs_rq
, se
);
4893 cfs_rq
= group_cfs_rq(se
);
4898 if (hrtick_enabled(rq
))
4899 hrtick_start_fair(rq
, p
);
4904 new_tasks
= idle_balance(rq
);
4906 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4907 * possible for any higher priority task to appear. In that case we
4908 * must re-start the pick_next_entity() loop.
4920 * Account for a descheduled task:
4922 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4924 struct sched_entity
*se
= &prev
->se
;
4925 struct cfs_rq
*cfs_rq
;
4927 for_each_sched_entity(se
) {
4928 cfs_rq
= cfs_rq_of(se
);
4929 put_prev_entity(cfs_rq
, se
);
4934 * sched_yield() is very simple
4936 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4938 static void yield_task_fair(struct rq
*rq
)
4940 struct task_struct
*curr
= rq
->curr
;
4941 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4942 struct sched_entity
*se
= &curr
->se
;
4945 * Are we the only task in the tree?
4947 if (unlikely(rq
->nr_running
== 1))
4950 clear_buddies(cfs_rq
, se
);
4952 if (curr
->policy
!= SCHED_BATCH
) {
4953 update_rq_clock(rq
);
4955 * Update run-time statistics of the 'current'.
4957 update_curr(cfs_rq
);
4959 * Tell update_rq_clock() that we've just updated,
4960 * so we don't do microscopic update in schedule()
4961 * and double the fastpath cost.
4963 rq
->skip_clock_update
= 1;
4969 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
4971 struct sched_entity
*se
= &p
->se
;
4973 /* throttled hierarchies are not runnable */
4974 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
4977 /* Tell the scheduler that we'd really like pse to run next. */
4980 yield_task_fair(rq
);
4986 /**************************************************
4987 * Fair scheduling class load-balancing methods.
4991 * The purpose of load-balancing is to achieve the same basic fairness the
4992 * per-cpu scheduler provides, namely provide a proportional amount of compute
4993 * time to each task. This is expressed in the following equation:
4995 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4997 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4998 * W_i,0 is defined as:
5000 * W_i,0 = \Sum_j w_i,j (2)
5002 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5003 * is derived from the nice value as per prio_to_weight[].
5005 * The weight average is an exponential decay average of the instantaneous
5008 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5010 * C_i is the compute capacity of cpu i, typically it is the
5011 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5012 * can also include other factors [XXX].
5014 * To achieve this balance we define a measure of imbalance which follows
5015 * directly from (1):
5017 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5019 * We them move tasks around to minimize the imbalance. In the continuous
5020 * function space it is obvious this converges, in the discrete case we get
5021 * a few fun cases generally called infeasible weight scenarios.
5024 * - infeasible weights;
5025 * - local vs global optima in the discrete case. ]
5030 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5031 * for all i,j solution, we create a tree of cpus that follows the hardware
5032 * topology where each level pairs two lower groups (or better). This results
5033 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5034 * tree to only the first of the previous level and we decrease the frequency
5035 * of load-balance at each level inv. proportional to the number of cpus in
5041 * \Sum { --- * --- * 2^i } = O(n) (5)
5043 * `- size of each group
5044 * | | `- number of cpus doing load-balance
5046 * `- sum over all levels
5048 * Coupled with a limit on how many tasks we can migrate every balance pass,
5049 * this makes (5) the runtime complexity of the balancer.
5051 * An important property here is that each CPU is still (indirectly) connected
5052 * to every other cpu in at most O(log n) steps:
5054 * The adjacency matrix of the resulting graph is given by:
5057 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5060 * And you'll find that:
5062 * A^(log_2 n)_i,j != 0 for all i,j (7)
5064 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5065 * The task movement gives a factor of O(m), giving a convergence complexity
5068 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5073 * In order to avoid CPUs going idle while there's still work to do, new idle
5074 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5075 * tree itself instead of relying on other CPUs to bring it work.
5077 * This adds some complexity to both (5) and (8) but it reduces the total idle
5085 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5088 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5093 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5095 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5097 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5100 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5101 * rewrite all of this once again.]
5104 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
5106 enum fbq_type
{ regular
, remote
, all
};
5108 #define LBF_ALL_PINNED 0x01
5109 #define LBF_NEED_BREAK 0x02
5110 #define LBF_DST_PINNED 0x04
5111 #define LBF_SOME_PINNED 0x08
5114 struct sched_domain
*sd
;
5122 struct cpumask
*dst_grpmask
;
5124 enum cpu_idle_type idle
;
5126 /* The set of CPUs under consideration for load-balancing */
5127 struct cpumask
*cpus
;
5132 unsigned int loop_break
;
5133 unsigned int loop_max
;
5135 enum fbq_type fbq_type
;
5136 struct list_head tasks
;
5140 * Is this task likely cache-hot:
5142 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
5146 lockdep_assert_held(&env
->src_rq
->lock
);
5148 if (p
->sched_class
!= &fair_sched_class
)
5151 if (unlikely(p
->policy
== SCHED_IDLE
))
5155 * Buddy candidates are cache hot:
5157 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
5158 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
5159 &p
->se
== cfs_rq_of(&p
->se
)->last
))
5162 if (sysctl_sched_migration_cost
== -1)
5164 if (sysctl_sched_migration_cost
== 0)
5167 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
5169 return delta
< (s64
)sysctl_sched_migration_cost
;
5172 #ifdef CONFIG_NUMA_BALANCING
5173 /* Returns true if the destination node has incurred more faults */
5174 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
5176 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5177 int src_nid
, dst_nid
;
5179 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults_memory
||
5180 !(env
->sd
->flags
& SD_NUMA
)) {
5184 src_nid
= cpu_to_node(env
->src_cpu
);
5185 dst_nid
= cpu_to_node(env
->dst_cpu
);
5187 if (src_nid
== dst_nid
)
5191 /* Task is already in the group's interleave set. */
5192 if (node_isset(src_nid
, numa_group
->active_nodes
))
5195 /* Task is moving into the group's interleave set. */
5196 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5199 return group_faults(p
, dst_nid
) > group_faults(p
, src_nid
);
5202 /* Encourage migration to the preferred node. */
5203 if (dst_nid
== p
->numa_preferred_nid
)
5206 return task_faults(p
, dst_nid
) > task_faults(p
, src_nid
);
5210 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
5212 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5213 int src_nid
, dst_nid
;
5215 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
5218 if (!p
->numa_faults_memory
|| !(env
->sd
->flags
& SD_NUMA
))
5221 src_nid
= cpu_to_node(env
->src_cpu
);
5222 dst_nid
= cpu_to_node(env
->dst_cpu
);
5224 if (src_nid
== dst_nid
)
5228 /* Task is moving within/into the group's interleave set. */
5229 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5232 /* Task is moving out of the group's interleave set. */
5233 if (node_isset(src_nid
, numa_group
->active_nodes
))
5236 return group_faults(p
, dst_nid
) < group_faults(p
, src_nid
);
5239 /* Migrating away from the preferred node is always bad. */
5240 if (src_nid
== p
->numa_preferred_nid
)
5243 return task_faults(p
, dst_nid
) < task_faults(p
, src_nid
);
5247 static inline bool migrate_improves_locality(struct task_struct
*p
,
5253 static inline bool migrate_degrades_locality(struct task_struct
*p
,
5261 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5264 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
5266 int tsk_cache_hot
= 0;
5268 lockdep_assert_held(&env
->src_rq
->lock
);
5271 * We do not migrate tasks that are:
5272 * 1) throttled_lb_pair, or
5273 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5274 * 3) running (obviously), or
5275 * 4) are cache-hot on their current CPU.
5277 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
5280 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
5283 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
5285 env
->flags
|= LBF_SOME_PINNED
;
5288 * Remember if this task can be migrated to any other cpu in
5289 * our sched_group. We may want to revisit it if we couldn't
5290 * meet load balance goals by pulling other tasks on src_cpu.
5292 * Also avoid computing new_dst_cpu if we have already computed
5293 * one in current iteration.
5295 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
5298 /* Prevent to re-select dst_cpu via env's cpus */
5299 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
5300 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
5301 env
->flags
|= LBF_DST_PINNED
;
5302 env
->new_dst_cpu
= cpu
;
5310 /* Record that we found atleast one task that could run on dst_cpu */
5311 env
->flags
&= ~LBF_ALL_PINNED
;
5313 if (task_running(env
->src_rq
, p
)) {
5314 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
5319 * Aggressive migration if:
5320 * 1) destination numa is preferred
5321 * 2) task is cache cold, or
5322 * 3) too many balance attempts have failed.
5324 tsk_cache_hot
= task_hot(p
, env
);
5326 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
5328 if (migrate_improves_locality(p
, env
) || !tsk_cache_hot
||
5329 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
5330 if (tsk_cache_hot
) {
5331 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5332 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5337 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
5342 * detach_task() -- detach the task for the migration specified in env
5344 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
5346 lockdep_assert_held(&env
->src_rq
->lock
);
5348 deactivate_task(env
->src_rq
, p
, 0);
5349 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
5350 set_task_cpu(p
, env
->dst_cpu
);
5354 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5355 * part of active balancing operations within "domain".
5357 * Returns a task if successful and NULL otherwise.
5359 static struct task_struct
*detach_one_task(struct lb_env
*env
)
5361 struct task_struct
*p
, *n
;
5363 lockdep_assert_held(&env
->src_rq
->lock
);
5365 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
5366 if (!can_migrate_task(p
, env
))
5369 detach_task(p
, env
);
5372 * Right now, this is only the second place where
5373 * lb_gained[env->idle] is updated (other is detach_tasks)
5374 * so we can safely collect stats here rather than
5375 * inside detach_tasks().
5377 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
5383 static const unsigned int sched_nr_migrate_break
= 32;
5386 * detach_tasks() -- tries to detach up to imbalance weighted load from
5387 * busiest_rq, as part of a balancing operation within domain "sd".
5389 * Returns number of detached tasks if successful and 0 otherwise.
5391 static int detach_tasks(struct lb_env
*env
)
5393 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
5394 struct task_struct
*p
;
5398 lockdep_assert_held(&env
->src_rq
->lock
);
5400 if (env
->imbalance
<= 0)
5403 while (!list_empty(tasks
)) {
5404 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5407 /* We've more or less seen every task there is, call it quits */
5408 if (env
->loop
> env
->loop_max
)
5411 /* take a breather every nr_migrate tasks */
5412 if (env
->loop
> env
->loop_break
) {
5413 env
->loop_break
+= sched_nr_migrate_break
;
5414 env
->flags
|= LBF_NEED_BREAK
;
5418 if (!can_migrate_task(p
, env
))
5421 load
= task_h_load(p
);
5423 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5426 if ((load
/ 2) > env
->imbalance
)
5429 detach_task(p
, env
);
5430 list_add(&p
->se
.group_node
, &env
->tasks
);
5433 env
->imbalance
-= load
;
5435 #ifdef CONFIG_PREEMPT
5437 * NEWIDLE balancing is a source of latency, so preemptible
5438 * kernels will stop after the first task is detached to minimize
5439 * the critical section.
5441 if (env
->idle
== CPU_NEWLY_IDLE
)
5446 * We only want to steal up to the prescribed amount of
5449 if (env
->imbalance
<= 0)
5454 list_move_tail(&p
->se
.group_node
, tasks
);
5458 * Right now, this is one of only two places we collect this stat
5459 * so we can safely collect detach_one_task() stats here rather
5460 * than inside detach_one_task().
5462 schedstat_add(env
->sd
, lb_gained
[env
->idle
], detached
);
5468 * attach_task() -- attach the task detached by detach_task() to its new rq.
5470 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
5472 lockdep_assert_held(&rq
->lock
);
5474 BUG_ON(task_rq(p
) != rq
);
5475 p
->on_rq
= TASK_ON_RQ_QUEUED
;
5476 activate_task(rq
, p
, 0);
5477 check_preempt_curr(rq
, p
, 0);
5481 * attach_one_task() -- attaches the task returned from detach_one_task() to
5484 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
5486 raw_spin_lock(&rq
->lock
);
5488 raw_spin_unlock(&rq
->lock
);
5492 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5495 static void attach_tasks(struct lb_env
*env
)
5497 struct list_head
*tasks
= &env
->tasks
;
5498 struct task_struct
*p
;
5500 raw_spin_lock(&env
->dst_rq
->lock
);
5502 while (!list_empty(tasks
)) {
5503 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5504 list_del_init(&p
->se
.group_node
);
5506 attach_task(env
->dst_rq
, p
);
5509 raw_spin_unlock(&env
->dst_rq
->lock
);
5512 #ifdef CONFIG_FAIR_GROUP_SCHED
5514 * update tg->load_weight by folding this cpu's load_avg
5516 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5518 struct sched_entity
*se
= tg
->se
[cpu
];
5519 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5521 /* throttled entities do not contribute to load */
5522 if (throttled_hierarchy(cfs_rq
))
5525 update_cfs_rq_blocked_load(cfs_rq
, 1);
5528 update_entity_load_avg(se
, 1);
5530 * We pivot on our runnable average having decayed to zero for
5531 * list removal. This generally implies that all our children
5532 * have also been removed (modulo rounding error or bandwidth
5533 * control); however, such cases are rare and we can fix these
5536 * TODO: fix up out-of-order children on enqueue.
5538 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5539 list_del_leaf_cfs_rq(cfs_rq
);
5541 struct rq
*rq
= rq_of(cfs_rq
);
5542 update_rq_runnable_avg(rq
, rq
->nr_running
);
5546 static void update_blocked_averages(int cpu
)
5548 struct rq
*rq
= cpu_rq(cpu
);
5549 struct cfs_rq
*cfs_rq
;
5550 unsigned long flags
;
5552 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5553 update_rq_clock(rq
);
5555 * Iterates the task_group tree in a bottom up fashion, see
5556 * list_add_leaf_cfs_rq() for details.
5558 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5560 * Note: We may want to consider periodically releasing
5561 * rq->lock about these updates so that creating many task
5562 * groups does not result in continually extending hold time.
5564 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5567 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5571 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5572 * This needs to be done in a top-down fashion because the load of a child
5573 * group is a fraction of its parents load.
5575 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5577 struct rq
*rq
= rq_of(cfs_rq
);
5578 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5579 unsigned long now
= jiffies
;
5582 if (cfs_rq
->last_h_load_update
== now
)
5585 cfs_rq
->h_load_next
= NULL
;
5586 for_each_sched_entity(se
) {
5587 cfs_rq
= cfs_rq_of(se
);
5588 cfs_rq
->h_load_next
= se
;
5589 if (cfs_rq
->last_h_load_update
== now
)
5594 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5595 cfs_rq
->last_h_load_update
= now
;
5598 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5599 load
= cfs_rq
->h_load
;
5600 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5601 cfs_rq
->runnable_load_avg
+ 1);
5602 cfs_rq
= group_cfs_rq(se
);
5603 cfs_rq
->h_load
= load
;
5604 cfs_rq
->last_h_load_update
= now
;
5608 static unsigned long task_h_load(struct task_struct
*p
)
5610 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5612 update_cfs_rq_h_load(cfs_rq
);
5613 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5614 cfs_rq
->runnable_load_avg
+ 1);
5617 static inline void update_blocked_averages(int cpu
)
5621 static unsigned long task_h_load(struct task_struct
*p
)
5623 return p
->se
.avg
.load_avg_contrib
;
5627 /********** Helpers for find_busiest_group ************************/
5636 * sg_lb_stats - stats of a sched_group required for load_balancing
5638 struct sg_lb_stats
{
5639 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5640 unsigned long group_load
; /* Total load over the CPUs of the group */
5641 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5642 unsigned long load_per_task
;
5643 unsigned long group_capacity
;
5644 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5645 unsigned int group_capacity_factor
;
5646 unsigned int idle_cpus
;
5647 unsigned int group_weight
;
5648 enum group_type group_type
;
5649 int group_has_free_capacity
;
5650 #ifdef CONFIG_NUMA_BALANCING
5651 unsigned int nr_numa_running
;
5652 unsigned int nr_preferred_running
;
5657 * sd_lb_stats - Structure to store the statistics of a sched_domain
5658 * during load balancing.
5660 struct sd_lb_stats
{
5661 struct sched_group
*busiest
; /* Busiest group in this sd */
5662 struct sched_group
*local
; /* Local group in this sd */
5663 unsigned long total_load
; /* Total load of all groups in sd */
5664 unsigned long total_capacity
; /* Total capacity of all groups in sd */
5665 unsigned long avg_load
; /* Average load across all groups in sd */
5667 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5668 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5671 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5674 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5675 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5676 * We must however clear busiest_stat::avg_load because
5677 * update_sd_pick_busiest() reads this before assignment.
5679 *sds
= (struct sd_lb_stats
){
5683 .total_capacity
= 0UL,
5686 .sum_nr_running
= 0,
5687 .group_type
= group_other
,
5693 * get_sd_load_idx - Obtain the load index for a given sched domain.
5694 * @sd: The sched_domain whose load_idx is to be obtained.
5695 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5697 * Return: The load index.
5699 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5700 enum cpu_idle_type idle
)
5706 load_idx
= sd
->busy_idx
;
5709 case CPU_NEWLY_IDLE
:
5710 load_idx
= sd
->newidle_idx
;
5713 load_idx
= sd
->idle_idx
;
5720 static unsigned long default_scale_capacity(struct sched_domain
*sd
, int cpu
)
5722 return SCHED_CAPACITY_SCALE
;
5725 unsigned long __weak
arch_scale_freq_capacity(struct sched_domain
*sd
, int cpu
)
5727 return default_scale_capacity(sd
, cpu
);
5730 static unsigned long default_scale_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5732 if ((sd
->flags
& SD_SHARE_CPUCAPACITY
) && (sd
->span_weight
> 1))
5733 return sd
->smt_gain
/ sd
->span_weight
;
5735 return SCHED_CAPACITY_SCALE
;
5738 unsigned long __weak
arch_scale_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5740 return default_scale_cpu_capacity(sd
, cpu
);
5743 static unsigned long scale_rt_capacity(int cpu
)
5745 struct rq
*rq
= cpu_rq(cpu
);
5746 u64 total
, available
, age_stamp
, avg
;
5750 * Since we're reading these variables without serialization make sure
5751 * we read them once before doing sanity checks on them.
5753 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5754 avg
= ACCESS_ONCE(rq
->rt_avg
);
5756 delta
= rq_clock(rq
) - age_stamp
;
5757 if (unlikely(delta
< 0))
5760 total
= sched_avg_period() + delta
;
5762 if (unlikely(total
< avg
)) {
5763 /* Ensures that capacity won't end up being negative */
5766 available
= total
- avg
;
5769 if (unlikely((s64
)total
< SCHED_CAPACITY_SCALE
))
5770 total
= SCHED_CAPACITY_SCALE
;
5772 total
>>= SCHED_CAPACITY_SHIFT
;
5774 return div_u64(available
, total
);
5777 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5779 unsigned long capacity
= SCHED_CAPACITY_SCALE
;
5780 struct sched_group
*sdg
= sd
->groups
;
5782 if (sched_feat(ARCH_CAPACITY
))
5783 capacity
*= arch_scale_cpu_capacity(sd
, cpu
);
5785 capacity
*= default_scale_cpu_capacity(sd
, cpu
);
5787 capacity
>>= SCHED_CAPACITY_SHIFT
;
5789 sdg
->sgc
->capacity_orig
= capacity
;
5791 if (sched_feat(ARCH_CAPACITY
))
5792 capacity
*= arch_scale_freq_capacity(sd
, cpu
);
5794 capacity
*= default_scale_capacity(sd
, cpu
);
5796 capacity
>>= SCHED_CAPACITY_SHIFT
;
5798 capacity
*= scale_rt_capacity(cpu
);
5799 capacity
>>= SCHED_CAPACITY_SHIFT
;
5804 cpu_rq(cpu
)->cpu_capacity
= capacity
;
5805 sdg
->sgc
->capacity
= capacity
;
5808 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
5810 struct sched_domain
*child
= sd
->child
;
5811 struct sched_group
*group
, *sdg
= sd
->groups
;
5812 unsigned long capacity
, capacity_orig
;
5813 unsigned long interval
;
5815 interval
= msecs_to_jiffies(sd
->balance_interval
);
5816 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5817 sdg
->sgc
->next_update
= jiffies
+ interval
;
5820 update_cpu_capacity(sd
, cpu
);
5824 capacity_orig
= capacity
= 0;
5826 if (child
->flags
& SD_OVERLAP
) {
5828 * SD_OVERLAP domains cannot assume that child groups
5829 * span the current group.
5832 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
5833 struct sched_group_capacity
*sgc
;
5834 struct rq
*rq
= cpu_rq(cpu
);
5837 * build_sched_domains() -> init_sched_groups_capacity()
5838 * gets here before we've attached the domains to the
5841 * Use capacity_of(), which is set irrespective of domains
5842 * in update_cpu_capacity().
5844 * This avoids capacity/capacity_orig from being 0 and
5845 * causing divide-by-zero issues on boot.
5847 * Runtime updates will correct capacity_orig.
5849 if (unlikely(!rq
->sd
)) {
5850 capacity_orig
+= capacity_of(cpu
);
5851 capacity
+= capacity_of(cpu
);
5855 sgc
= rq
->sd
->groups
->sgc
;
5856 capacity_orig
+= sgc
->capacity_orig
;
5857 capacity
+= sgc
->capacity
;
5861 * !SD_OVERLAP domains can assume that child groups
5862 * span the current group.
5865 group
= child
->groups
;
5867 capacity_orig
+= group
->sgc
->capacity_orig
;
5868 capacity
+= group
->sgc
->capacity
;
5869 group
= group
->next
;
5870 } while (group
!= child
->groups
);
5873 sdg
->sgc
->capacity_orig
= capacity_orig
;
5874 sdg
->sgc
->capacity
= capacity
;
5878 * Try and fix up capacity for tiny siblings, this is needed when
5879 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5880 * which on its own isn't powerful enough.
5882 * See update_sd_pick_busiest() and check_asym_packing().
5885 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
5888 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
5890 if (!(sd
->flags
& SD_SHARE_CPUCAPACITY
))
5894 * If ~90% of the cpu_capacity is still there, we're good.
5896 if (group
->sgc
->capacity
* 32 > group
->sgc
->capacity_orig
* 29)
5903 * Group imbalance indicates (and tries to solve) the problem where balancing
5904 * groups is inadequate due to tsk_cpus_allowed() constraints.
5906 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5907 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5910 * { 0 1 2 3 } { 4 5 6 7 }
5913 * If we were to balance group-wise we'd place two tasks in the first group and
5914 * two tasks in the second group. Clearly this is undesired as it will overload
5915 * cpu 3 and leave one of the cpus in the second group unused.
5917 * The current solution to this issue is detecting the skew in the first group
5918 * by noticing the lower domain failed to reach balance and had difficulty
5919 * moving tasks due to affinity constraints.
5921 * When this is so detected; this group becomes a candidate for busiest; see
5922 * update_sd_pick_busiest(). And calculate_imbalance() and
5923 * find_busiest_group() avoid some of the usual balance conditions to allow it
5924 * to create an effective group imbalance.
5926 * This is a somewhat tricky proposition since the next run might not find the
5927 * group imbalance and decide the groups need to be balanced again. A most
5928 * subtle and fragile situation.
5931 static inline int sg_imbalanced(struct sched_group
*group
)
5933 return group
->sgc
->imbalance
;
5937 * Compute the group capacity factor.
5939 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
5940 * first dividing out the smt factor and computing the actual number of cores
5941 * and limit unit capacity with that.
5943 static inline int sg_capacity_factor(struct lb_env
*env
, struct sched_group
*group
)
5945 unsigned int capacity_factor
, smt
, cpus
;
5946 unsigned int capacity
, capacity_orig
;
5948 capacity
= group
->sgc
->capacity
;
5949 capacity_orig
= group
->sgc
->capacity_orig
;
5950 cpus
= group
->group_weight
;
5952 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
5953 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, capacity_orig
);
5954 capacity_factor
= cpus
/ smt
; /* cores */
5956 capacity_factor
= min_t(unsigned,
5957 capacity_factor
, DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
));
5958 if (!capacity_factor
)
5959 capacity_factor
= fix_small_capacity(env
->sd
, group
);
5961 return capacity_factor
;
5964 static enum group_type
5965 group_classify(struct sched_group
*group
, struct sg_lb_stats
*sgs
)
5967 if (sgs
->sum_nr_running
> sgs
->group_capacity_factor
)
5968 return group_overloaded
;
5970 if (sg_imbalanced(group
))
5971 return group_imbalanced
;
5977 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5978 * @env: The load balancing environment.
5979 * @group: sched_group whose statistics are to be updated.
5980 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5981 * @local_group: Does group contain this_cpu.
5982 * @sgs: variable to hold the statistics for this group.
5983 * @overload: Indicate more than one runnable task for any CPU.
5985 static inline void update_sg_lb_stats(struct lb_env
*env
,
5986 struct sched_group
*group
, int load_idx
,
5987 int local_group
, struct sg_lb_stats
*sgs
,
5993 memset(sgs
, 0, sizeof(*sgs
));
5995 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5996 struct rq
*rq
= cpu_rq(i
);
5998 /* Bias balancing toward cpus of our domain */
6000 load
= target_load(i
, load_idx
);
6002 load
= source_load(i
, load_idx
);
6004 sgs
->group_load
+= load
;
6005 sgs
->sum_nr_running
+= rq
->cfs
.h_nr_running
;
6007 if (rq
->nr_running
> 1)
6010 #ifdef CONFIG_NUMA_BALANCING
6011 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
6012 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
6014 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
6019 /* Adjust by relative CPU capacity of the group */
6020 sgs
->group_capacity
= group
->sgc
->capacity
;
6021 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
6023 if (sgs
->sum_nr_running
)
6024 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
6026 sgs
->group_weight
= group
->group_weight
;
6027 sgs
->group_capacity_factor
= sg_capacity_factor(env
, group
);
6028 sgs
->group_type
= group_classify(group
, sgs
);
6030 if (sgs
->group_capacity_factor
> sgs
->sum_nr_running
)
6031 sgs
->group_has_free_capacity
= 1;
6035 * update_sd_pick_busiest - return 1 on busiest group
6036 * @env: The load balancing environment.
6037 * @sds: sched_domain statistics
6038 * @sg: sched_group candidate to be checked for being the busiest
6039 * @sgs: sched_group statistics
6041 * Determine if @sg is a busier group than the previously selected
6044 * Return: %true if @sg is a busier group than the previously selected
6045 * busiest group. %false otherwise.
6047 static bool update_sd_pick_busiest(struct lb_env
*env
,
6048 struct sd_lb_stats
*sds
,
6049 struct sched_group
*sg
,
6050 struct sg_lb_stats
*sgs
)
6052 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
6054 if (sgs
->group_type
> busiest
->group_type
)
6057 if (sgs
->group_type
< busiest
->group_type
)
6060 if (sgs
->avg_load
<= busiest
->avg_load
)
6063 /* This is the busiest node in its class. */
6064 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6068 * ASYM_PACKING needs to move all the work to the lowest
6069 * numbered CPUs in the group, therefore mark all groups
6070 * higher than ourself as busy.
6072 if (sgs
->sum_nr_running
&& env
->dst_cpu
< group_first_cpu(sg
)) {
6076 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
6083 #ifdef CONFIG_NUMA_BALANCING
6084 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6086 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
6088 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
6093 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6095 if (rq
->nr_running
> rq
->nr_numa_running
)
6097 if (rq
->nr_running
> rq
->nr_preferred_running
)
6102 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6107 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6111 #endif /* CONFIG_NUMA_BALANCING */
6114 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6115 * @env: The load balancing environment.
6116 * @sds: variable to hold the statistics for this sched_domain.
6118 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6120 struct sched_domain
*child
= env
->sd
->child
;
6121 struct sched_group
*sg
= env
->sd
->groups
;
6122 struct sg_lb_stats tmp_sgs
;
6123 int load_idx
, prefer_sibling
= 0;
6124 bool overload
= false;
6126 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
6129 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
6132 struct sg_lb_stats
*sgs
= &tmp_sgs
;
6135 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
6138 sgs
= &sds
->local_stat
;
6140 if (env
->idle
!= CPU_NEWLY_IDLE
||
6141 time_after_eq(jiffies
, sg
->sgc
->next_update
))
6142 update_group_capacity(env
->sd
, env
->dst_cpu
);
6145 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
,
6152 * In case the child domain prefers tasks go to siblings
6153 * first, lower the sg capacity factor to one so that we'll try
6154 * and move all the excess tasks away. We lower the capacity
6155 * of a group only if the local group has the capacity to fit
6156 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6157 * extra check prevents the case where you always pull from the
6158 * heaviest group when it is already under-utilized (possible
6159 * with a large weight task outweighs the tasks on the system).
6161 if (prefer_sibling
&& sds
->local
&&
6162 sds
->local_stat
.group_has_free_capacity
)
6163 sgs
->group_capacity_factor
= min(sgs
->group_capacity_factor
, 1U);
6165 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
6167 sds
->busiest_stat
= *sgs
;
6171 /* Now, start updating sd_lb_stats */
6172 sds
->total_load
+= sgs
->group_load
;
6173 sds
->total_capacity
+= sgs
->group_capacity
;
6176 } while (sg
!= env
->sd
->groups
);
6178 if (env
->sd
->flags
& SD_NUMA
)
6179 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
6181 if (!env
->sd
->parent
) {
6182 /* update overload indicator if we are at root domain */
6183 if (env
->dst_rq
->rd
->overload
!= overload
)
6184 env
->dst_rq
->rd
->overload
= overload
;
6190 * check_asym_packing - Check to see if the group is packed into the
6193 * This is primarily intended to used at the sibling level. Some
6194 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6195 * case of POWER7, it can move to lower SMT modes only when higher
6196 * threads are idle. When in lower SMT modes, the threads will
6197 * perform better since they share less core resources. Hence when we
6198 * have idle threads, we want them to be the higher ones.
6200 * This packing function is run on idle threads. It checks to see if
6201 * the busiest CPU in this domain (core in the P7 case) has a higher
6202 * CPU number than the packing function is being run on. Here we are
6203 * assuming lower CPU number will be equivalent to lower a SMT thread
6206 * Return: 1 when packing is required and a task should be moved to
6207 * this CPU. The amount of the imbalance is returned in *imbalance.
6209 * @env: The load balancing environment.
6210 * @sds: Statistics of the sched_domain which is to be packed
6212 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6216 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6222 busiest_cpu
= group_first_cpu(sds
->busiest
);
6223 if (env
->dst_cpu
> busiest_cpu
)
6226 env
->imbalance
= DIV_ROUND_CLOSEST(
6227 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
6228 SCHED_CAPACITY_SCALE
);
6234 * fix_small_imbalance - Calculate the minor imbalance that exists
6235 * amongst the groups of a sched_domain, during
6237 * @env: The load balancing environment.
6238 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6241 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6243 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
6244 unsigned int imbn
= 2;
6245 unsigned long scaled_busy_load_per_task
;
6246 struct sg_lb_stats
*local
, *busiest
;
6248 local
= &sds
->local_stat
;
6249 busiest
= &sds
->busiest_stat
;
6251 if (!local
->sum_nr_running
)
6252 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
6253 else if (busiest
->load_per_task
> local
->load_per_task
)
6256 scaled_busy_load_per_task
=
6257 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6258 busiest
->group_capacity
;
6260 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
6261 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
6262 env
->imbalance
= busiest
->load_per_task
;
6267 * OK, we don't have enough imbalance to justify moving tasks,
6268 * however we may be able to increase total CPU capacity used by
6272 capa_now
+= busiest
->group_capacity
*
6273 min(busiest
->load_per_task
, busiest
->avg_load
);
6274 capa_now
+= local
->group_capacity
*
6275 min(local
->load_per_task
, local
->avg_load
);
6276 capa_now
/= SCHED_CAPACITY_SCALE
;
6278 /* Amount of load we'd subtract */
6279 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
6280 capa_move
+= busiest
->group_capacity
*
6281 min(busiest
->load_per_task
,
6282 busiest
->avg_load
- scaled_busy_load_per_task
);
6285 /* Amount of load we'd add */
6286 if (busiest
->avg_load
* busiest
->group_capacity
<
6287 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
6288 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
6289 local
->group_capacity
;
6291 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6292 local
->group_capacity
;
6294 capa_move
+= local
->group_capacity
*
6295 min(local
->load_per_task
, local
->avg_load
+ tmp
);
6296 capa_move
/= SCHED_CAPACITY_SCALE
;
6298 /* Move if we gain throughput */
6299 if (capa_move
> capa_now
)
6300 env
->imbalance
= busiest
->load_per_task
;
6304 * calculate_imbalance - Calculate the amount of imbalance present within the
6305 * groups of a given sched_domain during load balance.
6306 * @env: load balance environment
6307 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6309 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6311 unsigned long max_pull
, load_above_capacity
= ~0UL;
6312 struct sg_lb_stats
*local
, *busiest
;
6314 local
= &sds
->local_stat
;
6315 busiest
= &sds
->busiest_stat
;
6317 if (busiest
->group_type
== group_imbalanced
) {
6319 * In the group_imb case we cannot rely on group-wide averages
6320 * to ensure cpu-load equilibrium, look at wider averages. XXX
6322 busiest
->load_per_task
=
6323 min(busiest
->load_per_task
, sds
->avg_load
);
6327 * In the presence of smp nice balancing, certain scenarios can have
6328 * max load less than avg load(as we skip the groups at or below
6329 * its cpu_capacity, while calculating max_load..)
6331 if (busiest
->avg_load
<= sds
->avg_load
||
6332 local
->avg_load
>= sds
->avg_load
) {
6334 return fix_small_imbalance(env
, sds
);
6338 * If there aren't any idle cpus, avoid creating some.
6340 if (busiest
->group_type
== group_overloaded
&&
6341 local
->group_type
== group_overloaded
) {
6342 load_above_capacity
=
6343 (busiest
->sum_nr_running
- busiest
->group_capacity_factor
);
6345 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_CAPACITY_SCALE
);
6346 load_above_capacity
/= busiest
->group_capacity
;
6350 * We're trying to get all the cpus to the average_load, so we don't
6351 * want to push ourselves above the average load, nor do we wish to
6352 * reduce the max loaded cpu below the average load. At the same time,
6353 * we also don't want to reduce the group load below the group capacity
6354 * (so that we can implement power-savings policies etc). Thus we look
6355 * for the minimum possible imbalance.
6357 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
6359 /* How much load to actually move to equalise the imbalance */
6360 env
->imbalance
= min(
6361 max_pull
* busiest
->group_capacity
,
6362 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
6363 ) / SCHED_CAPACITY_SCALE
;
6366 * if *imbalance is less than the average load per runnable task
6367 * there is no guarantee that any tasks will be moved so we'll have
6368 * a think about bumping its value to force at least one task to be
6371 if (env
->imbalance
< busiest
->load_per_task
)
6372 return fix_small_imbalance(env
, sds
);
6375 /******* find_busiest_group() helpers end here *********************/
6378 * find_busiest_group - Returns the busiest group within the sched_domain
6379 * if there is an imbalance. If there isn't an imbalance, and
6380 * the user has opted for power-savings, it returns a group whose
6381 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6382 * such a group exists.
6384 * Also calculates the amount of weighted load which should be moved
6385 * to restore balance.
6387 * @env: The load balancing environment.
6389 * Return: - The busiest group if imbalance exists.
6390 * - If no imbalance and user has opted for power-savings balance,
6391 * return the least loaded group whose CPUs can be
6392 * put to idle by rebalancing its tasks onto our group.
6394 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
6396 struct sg_lb_stats
*local
, *busiest
;
6397 struct sd_lb_stats sds
;
6399 init_sd_lb_stats(&sds
);
6402 * Compute the various statistics relavent for load balancing at
6405 update_sd_lb_stats(env
, &sds
);
6406 local
= &sds
.local_stat
;
6407 busiest
= &sds
.busiest_stat
;
6409 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
6410 check_asym_packing(env
, &sds
))
6413 /* There is no busy sibling group to pull tasks from */
6414 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
6417 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
6418 / sds
.total_capacity
;
6421 * If the busiest group is imbalanced the below checks don't
6422 * work because they assume all things are equal, which typically
6423 * isn't true due to cpus_allowed constraints and the like.
6425 if (busiest
->group_type
== group_imbalanced
)
6428 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6429 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_free_capacity
&&
6430 !busiest
->group_has_free_capacity
)
6434 * If the local group is busier than the selected busiest group
6435 * don't try and pull any tasks.
6437 if (local
->avg_load
>= busiest
->avg_load
)
6441 * Don't pull any tasks if this group is already above the domain
6444 if (local
->avg_load
>= sds
.avg_load
)
6447 if (env
->idle
== CPU_IDLE
) {
6449 * This cpu is idle. If the busiest group is not overloaded
6450 * and there is no imbalance between this and busiest group
6451 * wrt idle cpus, it is balanced. The imbalance becomes
6452 * significant if the diff is greater than 1 otherwise we
6453 * might end up to just move the imbalance on another group
6455 if ((busiest
->group_type
!= group_overloaded
) &&
6456 (local
->idle_cpus
<= (busiest
->idle_cpus
+ 1)))
6460 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6461 * imbalance_pct to be conservative.
6463 if (100 * busiest
->avg_load
<=
6464 env
->sd
->imbalance_pct
* local
->avg_load
)
6469 /* Looks like there is an imbalance. Compute it */
6470 calculate_imbalance(env
, &sds
);
6479 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6481 static struct rq
*find_busiest_queue(struct lb_env
*env
,
6482 struct sched_group
*group
)
6484 struct rq
*busiest
= NULL
, *rq
;
6485 unsigned long busiest_load
= 0, busiest_capacity
= 1;
6488 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6489 unsigned long capacity
, capacity_factor
, wl
;
6493 rt
= fbq_classify_rq(rq
);
6496 * We classify groups/runqueues into three groups:
6497 * - regular: there are !numa tasks
6498 * - remote: there are numa tasks that run on the 'wrong' node
6499 * - all: there is no distinction
6501 * In order to avoid migrating ideally placed numa tasks,
6502 * ignore those when there's better options.
6504 * If we ignore the actual busiest queue to migrate another
6505 * task, the next balance pass can still reduce the busiest
6506 * queue by moving tasks around inside the node.
6508 * If we cannot move enough load due to this classification
6509 * the next pass will adjust the group classification and
6510 * allow migration of more tasks.
6512 * Both cases only affect the total convergence complexity.
6514 if (rt
> env
->fbq_type
)
6517 capacity
= capacity_of(i
);
6518 capacity_factor
= DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
);
6519 if (!capacity_factor
)
6520 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6522 wl
= weighted_cpuload(i
);
6525 * When comparing with imbalance, use weighted_cpuload()
6526 * which is not scaled with the cpu capacity.
6528 if (capacity_factor
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6532 * For the load comparisons with the other cpu's, consider
6533 * the weighted_cpuload() scaled with the cpu capacity, so
6534 * that the load can be moved away from the cpu that is
6535 * potentially running at a lower capacity.
6537 * Thus we're looking for max(wl_i / capacity_i), crosswise
6538 * multiplication to rid ourselves of the division works out
6539 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6540 * our previous maximum.
6542 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
6544 busiest_capacity
= capacity
;
6553 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6554 * so long as it is large enough.
6556 #define MAX_PINNED_INTERVAL 512
6558 /* Working cpumask for load_balance and load_balance_newidle. */
6559 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6561 static int need_active_balance(struct lb_env
*env
)
6563 struct sched_domain
*sd
= env
->sd
;
6565 if (env
->idle
== CPU_NEWLY_IDLE
) {
6568 * ASYM_PACKING needs to force migrate tasks from busy but
6569 * higher numbered CPUs in order to pack all tasks in the
6570 * lowest numbered CPUs.
6572 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6576 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6579 static int active_load_balance_cpu_stop(void *data
);
6581 static int should_we_balance(struct lb_env
*env
)
6583 struct sched_group
*sg
= env
->sd
->groups
;
6584 struct cpumask
*sg_cpus
, *sg_mask
;
6585 int cpu
, balance_cpu
= -1;
6588 * In the newly idle case, we will allow all the cpu's
6589 * to do the newly idle load balance.
6591 if (env
->idle
== CPU_NEWLY_IDLE
)
6594 sg_cpus
= sched_group_cpus(sg
);
6595 sg_mask
= sched_group_mask(sg
);
6596 /* Try to find first idle cpu */
6597 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6598 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6605 if (balance_cpu
== -1)
6606 balance_cpu
= group_balance_cpu(sg
);
6609 * First idle cpu or the first cpu(busiest) in this sched group
6610 * is eligible for doing load balancing at this and above domains.
6612 return balance_cpu
== env
->dst_cpu
;
6616 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6617 * tasks if there is an imbalance.
6619 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6620 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6621 int *continue_balancing
)
6623 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6624 struct sched_domain
*sd_parent
= sd
->parent
;
6625 struct sched_group
*group
;
6627 unsigned long flags
;
6628 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
6630 struct lb_env env
= {
6632 .dst_cpu
= this_cpu
,
6634 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6636 .loop_break
= sched_nr_migrate_break
,
6639 .tasks
= LIST_HEAD_INIT(env
.tasks
),
6643 * For NEWLY_IDLE load_balancing, we don't need to consider
6644 * other cpus in our group
6646 if (idle
== CPU_NEWLY_IDLE
)
6647 env
.dst_grpmask
= NULL
;
6649 cpumask_copy(cpus
, cpu_active_mask
);
6651 schedstat_inc(sd
, lb_count
[idle
]);
6654 if (!should_we_balance(&env
)) {
6655 *continue_balancing
= 0;
6659 group
= find_busiest_group(&env
);
6661 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6665 busiest
= find_busiest_queue(&env
, group
);
6667 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6671 BUG_ON(busiest
== env
.dst_rq
);
6673 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6676 if (busiest
->nr_running
> 1) {
6678 * Attempt to move tasks. If find_busiest_group has found
6679 * an imbalance but busiest->nr_running <= 1, the group is
6680 * still unbalanced. ld_moved simply stays zero, so it is
6681 * correctly treated as an imbalance.
6683 env
.flags
|= LBF_ALL_PINNED
;
6684 env
.src_cpu
= busiest
->cpu
;
6685 env
.src_rq
= busiest
;
6686 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6689 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6692 * cur_ld_moved - load moved in current iteration
6693 * ld_moved - cumulative load moved across iterations
6695 cur_ld_moved
= detach_tasks(&env
);
6698 * We've detached some tasks from busiest_rq. Every
6699 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6700 * unlock busiest->lock, and we are able to be sure
6701 * that nobody can manipulate the tasks in parallel.
6702 * See task_rq_lock() family for the details.
6705 raw_spin_unlock(&busiest
->lock
);
6709 ld_moved
+= cur_ld_moved
;
6712 local_irq_restore(flags
);
6714 if (env
.flags
& LBF_NEED_BREAK
) {
6715 env
.flags
&= ~LBF_NEED_BREAK
;
6720 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6721 * us and move them to an alternate dst_cpu in our sched_group
6722 * where they can run. The upper limit on how many times we
6723 * iterate on same src_cpu is dependent on number of cpus in our
6726 * This changes load balance semantics a bit on who can move
6727 * load to a given_cpu. In addition to the given_cpu itself
6728 * (or a ilb_cpu acting on its behalf where given_cpu is
6729 * nohz-idle), we now have balance_cpu in a position to move
6730 * load to given_cpu. In rare situations, this may cause
6731 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6732 * _independently_ and at _same_ time to move some load to
6733 * given_cpu) causing exceess load to be moved to given_cpu.
6734 * This however should not happen so much in practice and
6735 * moreover subsequent load balance cycles should correct the
6736 * excess load moved.
6738 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6740 /* Prevent to re-select dst_cpu via env's cpus */
6741 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6743 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6744 env
.dst_cpu
= env
.new_dst_cpu
;
6745 env
.flags
&= ~LBF_DST_PINNED
;
6747 env
.loop_break
= sched_nr_migrate_break
;
6750 * Go back to "more_balance" rather than "redo" since we
6751 * need to continue with same src_cpu.
6757 * We failed to reach balance because of affinity.
6760 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
6762 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
6763 *group_imbalance
= 1;
6766 /* All tasks on this runqueue were pinned by CPU affinity */
6767 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6768 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6769 if (!cpumask_empty(cpus
)) {
6771 env
.loop_break
= sched_nr_migrate_break
;
6774 goto out_all_pinned
;
6779 schedstat_inc(sd
, lb_failed
[idle
]);
6781 * Increment the failure counter only on periodic balance.
6782 * We do not want newidle balance, which can be very
6783 * frequent, pollute the failure counter causing
6784 * excessive cache_hot migrations and active balances.
6786 if (idle
!= CPU_NEWLY_IDLE
)
6787 sd
->nr_balance_failed
++;
6789 if (need_active_balance(&env
)) {
6790 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6792 /* don't kick the active_load_balance_cpu_stop,
6793 * if the curr task on busiest cpu can't be
6796 if (!cpumask_test_cpu(this_cpu
,
6797 tsk_cpus_allowed(busiest
->curr
))) {
6798 raw_spin_unlock_irqrestore(&busiest
->lock
,
6800 env
.flags
|= LBF_ALL_PINNED
;
6801 goto out_one_pinned
;
6805 * ->active_balance synchronizes accesses to
6806 * ->active_balance_work. Once set, it's cleared
6807 * only after active load balance is finished.
6809 if (!busiest
->active_balance
) {
6810 busiest
->active_balance
= 1;
6811 busiest
->push_cpu
= this_cpu
;
6814 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
6816 if (active_balance
) {
6817 stop_one_cpu_nowait(cpu_of(busiest
),
6818 active_load_balance_cpu_stop
, busiest
,
6819 &busiest
->active_balance_work
);
6823 * We've kicked active balancing, reset the failure
6826 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
6829 sd
->nr_balance_failed
= 0;
6831 if (likely(!active_balance
)) {
6832 /* We were unbalanced, so reset the balancing interval */
6833 sd
->balance_interval
= sd
->min_interval
;
6836 * If we've begun active balancing, start to back off. This
6837 * case may not be covered by the all_pinned logic if there
6838 * is only 1 task on the busy runqueue (because we don't call
6841 if (sd
->balance_interval
< sd
->max_interval
)
6842 sd
->balance_interval
*= 2;
6849 * We reach balance although we may have faced some affinity
6850 * constraints. Clear the imbalance flag if it was set.
6853 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
6855 if (*group_imbalance
)
6856 *group_imbalance
= 0;
6861 * We reach balance because all tasks are pinned at this level so
6862 * we can't migrate them. Let the imbalance flag set so parent level
6863 * can try to migrate them.
6865 schedstat_inc(sd
, lb_balanced
[idle
]);
6867 sd
->nr_balance_failed
= 0;
6870 /* tune up the balancing interval */
6871 if (((env
.flags
& LBF_ALL_PINNED
) &&
6872 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
6873 (sd
->balance_interval
< sd
->max_interval
))
6874 sd
->balance_interval
*= 2;
6881 static inline unsigned long
6882 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
6884 unsigned long interval
= sd
->balance_interval
;
6887 interval
*= sd
->busy_factor
;
6889 /* scale ms to jiffies */
6890 interval
= msecs_to_jiffies(interval
);
6891 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6897 update_next_balance(struct sched_domain
*sd
, int cpu_busy
, unsigned long *next_balance
)
6899 unsigned long interval
, next
;
6901 interval
= get_sd_balance_interval(sd
, cpu_busy
);
6902 next
= sd
->last_balance
+ interval
;
6904 if (time_after(*next_balance
, next
))
6905 *next_balance
= next
;
6909 * idle_balance is called by schedule() if this_cpu is about to become
6910 * idle. Attempts to pull tasks from other CPUs.
6912 static int idle_balance(struct rq
*this_rq
)
6914 unsigned long next_balance
= jiffies
+ HZ
;
6915 int this_cpu
= this_rq
->cpu
;
6916 struct sched_domain
*sd
;
6917 int pulled_task
= 0;
6920 idle_enter_fair(this_rq
);
6923 * We must set idle_stamp _before_ calling idle_balance(), such that we
6924 * measure the duration of idle_balance() as idle time.
6926 this_rq
->idle_stamp
= rq_clock(this_rq
);
6928 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
6929 !this_rq
->rd
->overload
) {
6931 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
6933 update_next_balance(sd
, 0, &next_balance
);
6940 * Drop the rq->lock, but keep IRQ/preempt disabled.
6942 raw_spin_unlock(&this_rq
->lock
);
6944 update_blocked_averages(this_cpu
);
6946 for_each_domain(this_cpu
, sd
) {
6947 int continue_balancing
= 1;
6948 u64 t0
, domain_cost
;
6950 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6953 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
6954 update_next_balance(sd
, 0, &next_balance
);
6958 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
6959 t0
= sched_clock_cpu(this_cpu
);
6961 pulled_task
= load_balance(this_cpu
, this_rq
,
6963 &continue_balancing
);
6965 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
6966 if (domain_cost
> sd
->max_newidle_lb_cost
)
6967 sd
->max_newidle_lb_cost
= domain_cost
;
6969 curr_cost
+= domain_cost
;
6972 update_next_balance(sd
, 0, &next_balance
);
6975 * Stop searching for tasks to pull if there are
6976 * now runnable tasks on this rq.
6978 if (pulled_task
|| this_rq
->nr_running
> 0)
6983 raw_spin_lock(&this_rq
->lock
);
6985 if (curr_cost
> this_rq
->max_idle_balance_cost
)
6986 this_rq
->max_idle_balance_cost
= curr_cost
;
6989 * While browsing the domains, we released the rq lock, a task could
6990 * have been enqueued in the meantime. Since we're not going idle,
6991 * pretend we pulled a task.
6993 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
6997 /* Move the next balance forward */
6998 if (time_after(this_rq
->next_balance
, next_balance
))
6999 this_rq
->next_balance
= next_balance
;
7001 /* Is there a task of a high priority class? */
7002 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
7006 idle_exit_fair(this_rq
);
7007 this_rq
->idle_stamp
= 0;
7014 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7015 * running tasks off the busiest CPU onto idle CPUs. It requires at
7016 * least 1 task to be running on each physical CPU where possible, and
7017 * avoids physical / logical imbalances.
7019 static int active_load_balance_cpu_stop(void *data
)
7021 struct rq
*busiest_rq
= data
;
7022 int busiest_cpu
= cpu_of(busiest_rq
);
7023 int target_cpu
= busiest_rq
->push_cpu
;
7024 struct rq
*target_rq
= cpu_rq(target_cpu
);
7025 struct sched_domain
*sd
;
7026 struct task_struct
*p
= NULL
;
7028 raw_spin_lock_irq(&busiest_rq
->lock
);
7030 /* make sure the requested cpu hasn't gone down in the meantime */
7031 if (unlikely(busiest_cpu
!= smp_processor_id() ||
7032 !busiest_rq
->active_balance
))
7035 /* Is there any task to move? */
7036 if (busiest_rq
->nr_running
<= 1)
7040 * This condition is "impossible", if it occurs
7041 * we need to fix it. Originally reported by
7042 * Bjorn Helgaas on a 128-cpu setup.
7044 BUG_ON(busiest_rq
== target_rq
);
7046 /* Search for an sd spanning us and the target CPU. */
7048 for_each_domain(target_cpu
, sd
) {
7049 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
7050 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
7055 struct lb_env env
= {
7057 .dst_cpu
= target_cpu
,
7058 .dst_rq
= target_rq
,
7059 .src_cpu
= busiest_rq
->cpu
,
7060 .src_rq
= busiest_rq
,
7064 schedstat_inc(sd
, alb_count
);
7066 p
= detach_one_task(&env
);
7068 schedstat_inc(sd
, alb_pushed
);
7070 schedstat_inc(sd
, alb_failed
);
7074 busiest_rq
->active_balance
= 0;
7075 raw_spin_unlock(&busiest_rq
->lock
);
7078 attach_one_task(target_rq
, p
);
7085 static inline int on_null_domain(struct rq
*rq
)
7087 return unlikely(!rcu_dereference_sched(rq
->sd
));
7090 #ifdef CONFIG_NO_HZ_COMMON
7092 * idle load balancing details
7093 * - When one of the busy CPUs notice that there may be an idle rebalancing
7094 * needed, they will kick the idle load balancer, which then does idle
7095 * load balancing for all the idle CPUs.
7098 cpumask_var_t idle_cpus_mask
;
7100 unsigned long next_balance
; /* in jiffy units */
7101 } nohz ____cacheline_aligned
;
7103 static inline int find_new_ilb(void)
7105 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
7107 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
7114 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7115 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7116 * CPU (if there is one).
7118 static void nohz_balancer_kick(void)
7122 nohz
.next_balance
++;
7124 ilb_cpu
= find_new_ilb();
7126 if (ilb_cpu
>= nr_cpu_ids
)
7129 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
7132 * Use smp_send_reschedule() instead of resched_cpu().
7133 * This way we generate a sched IPI on the target cpu which
7134 * is idle. And the softirq performing nohz idle load balance
7135 * will be run before returning from the IPI.
7137 smp_send_reschedule(ilb_cpu
);
7141 static inline void nohz_balance_exit_idle(int cpu
)
7143 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
7145 * Completely isolated CPUs don't ever set, so we must test.
7147 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
7148 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
7149 atomic_dec(&nohz
.nr_cpus
);
7151 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7155 static inline void set_cpu_sd_state_busy(void)
7157 struct sched_domain
*sd
;
7158 int cpu
= smp_processor_id();
7161 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7163 if (!sd
|| !sd
->nohz_idle
)
7167 atomic_inc(&sd
->groups
->sgc
->nr_busy_cpus
);
7172 void set_cpu_sd_state_idle(void)
7174 struct sched_domain
*sd
;
7175 int cpu
= smp_processor_id();
7178 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7180 if (!sd
|| sd
->nohz_idle
)
7184 atomic_dec(&sd
->groups
->sgc
->nr_busy_cpus
);
7190 * This routine will record that the cpu is going idle with tick stopped.
7191 * This info will be used in performing idle load balancing in the future.
7193 void nohz_balance_enter_idle(int cpu
)
7196 * If this cpu is going down, then nothing needs to be done.
7198 if (!cpu_active(cpu
))
7201 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
7205 * If we're a completely isolated CPU, we don't play.
7207 if (on_null_domain(cpu_rq(cpu
)))
7210 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
7211 atomic_inc(&nohz
.nr_cpus
);
7212 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7215 static int sched_ilb_notifier(struct notifier_block
*nfb
,
7216 unsigned long action
, void *hcpu
)
7218 switch (action
& ~CPU_TASKS_FROZEN
) {
7220 nohz_balance_exit_idle(smp_processor_id());
7228 static DEFINE_SPINLOCK(balancing
);
7231 * Scale the max load_balance interval with the number of CPUs in the system.
7232 * This trades load-balance latency on larger machines for less cross talk.
7234 void update_max_interval(void)
7236 max_load_balance_interval
= HZ
*num_online_cpus()/10;
7240 * It checks each scheduling domain to see if it is due to be balanced,
7241 * and initiates a balancing operation if so.
7243 * Balancing parameters are set up in init_sched_domains.
7245 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
7247 int continue_balancing
= 1;
7249 unsigned long interval
;
7250 struct sched_domain
*sd
;
7251 /* Earliest time when we have to do rebalance again */
7252 unsigned long next_balance
= jiffies
+ 60*HZ
;
7253 int update_next_balance
= 0;
7254 int need_serialize
, need_decay
= 0;
7257 update_blocked_averages(cpu
);
7260 for_each_domain(cpu
, sd
) {
7262 * Decay the newidle max times here because this is a regular
7263 * visit to all the domains. Decay ~1% per second.
7265 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
7266 sd
->max_newidle_lb_cost
=
7267 (sd
->max_newidle_lb_cost
* 253) / 256;
7268 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
7271 max_cost
+= sd
->max_newidle_lb_cost
;
7273 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7277 * Stop the load balance at this level. There is another
7278 * CPU in our sched group which is doing load balancing more
7281 if (!continue_balancing
) {
7287 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7289 need_serialize
= sd
->flags
& SD_SERIALIZE
;
7290 if (need_serialize
) {
7291 if (!spin_trylock(&balancing
))
7295 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
7296 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
7298 * The LBF_DST_PINNED logic could have changed
7299 * env->dst_cpu, so we can't know our idle
7300 * state even if we migrated tasks. Update it.
7302 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
7304 sd
->last_balance
= jiffies
;
7305 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7308 spin_unlock(&balancing
);
7310 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
7311 next_balance
= sd
->last_balance
+ interval
;
7312 update_next_balance
= 1;
7317 * Ensure the rq-wide value also decays but keep it at a
7318 * reasonable floor to avoid funnies with rq->avg_idle.
7320 rq
->max_idle_balance_cost
=
7321 max((u64
)sysctl_sched_migration_cost
, max_cost
);
7326 * next_balance will be updated only when there is a need.
7327 * When the cpu is attached to null domain for ex, it will not be
7330 if (likely(update_next_balance
))
7331 rq
->next_balance
= next_balance
;
7334 #ifdef CONFIG_NO_HZ_COMMON
7336 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7337 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7339 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
7341 int this_cpu
= this_rq
->cpu
;
7345 if (idle
!= CPU_IDLE
||
7346 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
7349 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
7350 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
7354 * If this cpu gets work to do, stop the load balancing
7355 * work being done for other cpus. Next load
7356 * balancing owner will pick it up.
7361 rq
= cpu_rq(balance_cpu
);
7364 * If time for next balance is due,
7367 if (time_after_eq(jiffies
, rq
->next_balance
)) {
7368 raw_spin_lock_irq(&rq
->lock
);
7369 update_rq_clock(rq
);
7370 update_idle_cpu_load(rq
);
7371 raw_spin_unlock_irq(&rq
->lock
);
7372 rebalance_domains(rq
, CPU_IDLE
);
7375 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
7376 this_rq
->next_balance
= rq
->next_balance
;
7378 nohz
.next_balance
= this_rq
->next_balance
;
7380 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
7384 * Current heuristic for kicking the idle load balancer in the presence
7385 * of an idle cpu is the system.
7386 * - This rq has more than one task.
7387 * - At any scheduler domain level, this cpu's scheduler group has multiple
7388 * busy cpu's exceeding the group's capacity.
7389 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7390 * domain span are idle.
7392 static inline int nohz_kick_needed(struct rq
*rq
)
7394 unsigned long now
= jiffies
;
7395 struct sched_domain
*sd
;
7396 struct sched_group_capacity
*sgc
;
7397 int nr_busy
, cpu
= rq
->cpu
;
7399 if (unlikely(rq
->idle_balance
))
7403 * We may be recently in ticked or tickless idle mode. At the first
7404 * busy tick after returning from idle, we will update the busy stats.
7406 set_cpu_sd_state_busy();
7407 nohz_balance_exit_idle(cpu
);
7410 * None are in tickless mode and hence no need for NOHZ idle load
7413 if (likely(!atomic_read(&nohz
.nr_cpus
)))
7416 if (time_before(now
, nohz
.next_balance
))
7419 if (rq
->nr_running
>= 2)
7423 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7426 sgc
= sd
->groups
->sgc
;
7427 nr_busy
= atomic_read(&sgc
->nr_busy_cpus
);
7430 goto need_kick_unlock
;
7433 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
7435 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
7436 sched_domain_span(sd
)) < cpu
))
7437 goto need_kick_unlock
;
7448 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
7452 * run_rebalance_domains is triggered when needed from the scheduler tick.
7453 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7455 static void run_rebalance_domains(struct softirq_action
*h
)
7457 struct rq
*this_rq
= this_rq();
7458 enum cpu_idle_type idle
= this_rq
->idle_balance
?
7459 CPU_IDLE
: CPU_NOT_IDLE
;
7461 rebalance_domains(this_rq
, idle
);
7464 * If this cpu has a pending nohz_balance_kick, then do the
7465 * balancing on behalf of the other idle cpus whose ticks are
7468 nohz_idle_balance(this_rq
, idle
);
7472 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7474 void trigger_load_balance(struct rq
*rq
)
7476 /* Don't need to rebalance while attached to NULL domain */
7477 if (unlikely(on_null_domain(rq
)))
7480 if (time_after_eq(jiffies
, rq
->next_balance
))
7481 raise_softirq(SCHED_SOFTIRQ
);
7482 #ifdef CONFIG_NO_HZ_COMMON
7483 if (nohz_kick_needed(rq
))
7484 nohz_balancer_kick();
7488 static void rq_online_fair(struct rq
*rq
)
7492 update_runtime_enabled(rq
);
7495 static void rq_offline_fair(struct rq
*rq
)
7499 /* Ensure any throttled groups are reachable by pick_next_task */
7500 unthrottle_offline_cfs_rqs(rq
);
7503 #endif /* CONFIG_SMP */
7506 * scheduler tick hitting a task of our scheduling class:
7508 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
7510 struct cfs_rq
*cfs_rq
;
7511 struct sched_entity
*se
= &curr
->se
;
7513 for_each_sched_entity(se
) {
7514 cfs_rq
= cfs_rq_of(se
);
7515 entity_tick(cfs_rq
, se
, queued
);
7518 if (numabalancing_enabled
)
7519 task_tick_numa(rq
, curr
);
7521 update_rq_runnable_avg(rq
, 1);
7525 * called on fork with the child task as argument from the parent's context
7526 * - child not yet on the tasklist
7527 * - preemption disabled
7529 static void task_fork_fair(struct task_struct
*p
)
7531 struct cfs_rq
*cfs_rq
;
7532 struct sched_entity
*se
= &p
->se
, *curr
;
7533 int this_cpu
= smp_processor_id();
7534 struct rq
*rq
= this_rq();
7535 unsigned long flags
;
7537 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7539 update_rq_clock(rq
);
7541 cfs_rq
= task_cfs_rq(current
);
7542 curr
= cfs_rq
->curr
;
7545 * Not only the cpu but also the task_group of the parent might have
7546 * been changed after parent->se.parent,cfs_rq were copied to
7547 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7548 * of child point to valid ones.
7551 __set_task_cpu(p
, this_cpu
);
7554 update_curr(cfs_rq
);
7557 se
->vruntime
= curr
->vruntime
;
7558 place_entity(cfs_rq
, se
, 1);
7560 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
7562 * Upon rescheduling, sched_class::put_prev_task() will place
7563 * 'current' within the tree based on its new key value.
7565 swap(curr
->vruntime
, se
->vruntime
);
7569 se
->vruntime
-= cfs_rq
->min_vruntime
;
7571 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7575 * Priority of the task has changed. Check to see if we preempt
7579 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
7581 if (!task_on_rq_queued(p
))
7585 * Reschedule if we are currently running on this runqueue and
7586 * our priority decreased, or if we are not currently running on
7587 * this runqueue and our priority is higher than the current's
7589 if (rq
->curr
== p
) {
7590 if (p
->prio
> oldprio
)
7593 check_preempt_curr(rq
, p
, 0);
7596 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
7598 struct sched_entity
*se
= &p
->se
;
7599 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7602 * Ensure the task's vruntime is normalized, so that when it's
7603 * switched back to the fair class the enqueue_entity(.flags=0) will
7604 * do the right thing.
7606 * If it's queued, then the dequeue_entity(.flags=0) will already
7607 * have normalized the vruntime, if it's !queued, then only when
7608 * the task is sleeping will it still have non-normalized vruntime.
7610 if (!task_on_rq_queued(p
) && p
->state
!= TASK_RUNNING
) {
7612 * Fix up our vruntime so that the current sleep doesn't
7613 * cause 'unlimited' sleep bonus.
7615 place_entity(cfs_rq
, se
, 0);
7616 se
->vruntime
-= cfs_rq
->min_vruntime
;
7621 * Remove our load from contribution when we leave sched_fair
7622 * and ensure we don't carry in an old decay_count if we
7625 if (se
->avg
.decay_count
) {
7626 __synchronize_entity_decay(se
);
7627 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7633 * We switched to the sched_fair class.
7635 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7637 #ifdef CONFIG_FAIR_GROUP_SCHED
7638 struct sched_entity
*se
= &p
->se
;
7640 * Since the real-depth could have been changed (only FAIR
7641 * class maintain depth value), reset depth properly.
7643 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7645 if (!task_on_rq_queued(p
))
7649 * We were most likely switched from sched_rt, so
7650 * kick off the schedule if running, otherwise just see
7651 * if we can still preempt the current task.
7656 check_preempt_curr(rq
, p
, 0);
7659 /* Account for a task changing its policy or group.
7661 * This routine is mostly called to set cfs_rq->curr field when a task
7662 * migrates between groups/classes.
7664 static void set_curr_task_fair(struct rq
*rq
)
7666 struct sched_entity
*se
= &rq
->curr
->se
;
7668 for_each_sched_entity(se
) {
7669 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7671 set_next_entity(cfs_rq
, se
);
7672 /* ensure bandwidth has been allocated on our new cfs_rq */
7673 account_cfs_rq_runtime(cfs_rq
, 0);
7677 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7679 cfs_rq
->tasks_timeline
= RB_ROOT
;
7680 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7681 #ifndef CONFIG_64BIT
7682 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7685 atomic64_set(&cfs_rq
->decay_counter
, 1);
7686 atomic_long_set(&cfs_rq
->removed_load
, 0);
7690 #ifdef CONFIG_FAIR_GROUP_SCHED
7691 static void task_move_group_fair(struct task_struct
*p
, int queued
)
7693 struct sched_entity
*se
= &p
->se
;
7694 struct cfs_rq
*cfs_rq
;
7697 * If the task was not on the rq at the time of this cgroup movement
7698 * it must have been asleep, sleeping tasks keep their ->vruntime
7699 * absolute on their old rq until wakeup (needed for the fair sleeper
7700 * bonus in place_entity()).
7702 * If it was on the rq, we've just 'preempted' it, which does convert
7703 * ->vruntime to a relative base.
7705 * Make sure both cases convert their relative position when migrating
7706 * to another cgroup's rq. This does somewhat interfere with the
7707 * fair sleeper stuff for the first placement, but who cares.
7710 * When !queued, vruntime of the task has usually NOT been normalized.
7711 * But there are some cases where it has already been normalized:
7713 * - Moving a forked child which is waiting for being woken up by
7714 * wake_up_new_task().
7715 * - Moving a task which has been woken up by try_to_wake_up() and
7716 * waiting for actually being woken up by sched_ttwu_pending().
7718 * To prevent boost or penalty in the new cfs_rq caused by delta
7719 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7721 if (!queued
&& (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7725 se
->vruntime
-= cfs_rq_of(se
)->min_vruntime
;
7726 set_task_rq(p
, task_cpu(p
));
7727 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7729 cfs_rq
= cfs_rq_of(se
);
7730 se
->vruntime
+= cfs_rq
->min_vruntime
;
7733 * migrate_task_rq_fair() will have removed our previous
7734 * contribution, but we must synchronize for ongoing future
7737 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7738 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
7743 void free_fair_sched_group(struct task_group
*tg
)
7747 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7749 for_each_possible_cpu(i
) {
7751 kfree(tg
->cfs_rq
[i
]);
7760 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7762 struct cfs_rq
*cfs_rq
;
7763 struct sched_entity
*se
;
7766 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7769 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7773 tg
->shares
= NICE_0_LOAD
;
7775 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7777 for_each_possible_cpu(i
) {
7778 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7779 GFP_KERNEL
, cpu_to_node(i
));
7783 se
= kzalloc_node(sizeof(struct sched_entity
),
7784 GFP_KERNEL
, cpu_to_node(i
));
7788 init_cfs_rq(cfs_rq
);
7789 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
7800 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7802 struct rq
*rq
= cpu_rq(cpu
);
7803 unsigned long flags
;
7806 * Only empty task groups can be destroyed; so we can speculatively
7807 * check on_list without danger of it being re-added.
7809 if (!tg
->cfs_rq
[cpu
]->on_list
)
7812 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7813 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
7814 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7817 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7818 struct sched_entity
*se
, int cpu
,
7819 struct sched_entity
*parent
)
7821 struct rq
*rq
= cpu_rq(cpu
);
7825 init_cfs_rq_runtime(cfs_rq
);
7827 tg
->cfs_rq
[cpu
] = cfs_rq
;
7830 /* se could be NULL for root_task_group */
7835 se
->cfs_rq
= &rq
->cfs
;
7838 se
->cfs_rq
= parent
->my_q
;
7839 se
->depth
= parent
->depth
+ 1;
7843 /* guarantee group entities always have weight */
7844 update_load_set(&se
->load
, NICE_0_LOAD
);
7845 se
->parent
= parent
;
7848 static DEFINE_MUTEX(shares_mutex
);
7850 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7853 unsigned long flags
;
7856 * We can't change the weight of the root cgroup.
7861 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
7863 mutex_lock(&shares_mutex
);
7864 if (tg
->shares
== shares
)
7867 tg
->shares
= shares
;
7868 for_each_possible_cpu(i
) {
7869 struct rq
*rq
= cpu_rq(i
);
7870 struct sched_entity
*se
;
7873 /* Propagate contribution to hierarchy */
7874 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7876 /* Possible calls to update_curr() need rq clock */
7877 update_rq_clock(rq
);
7878 for_each_sched_entity(se
)
7879 update_cfs_shares(group_cfs_rq(se
));
7880 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7884 mutex_unlock(&shares_mutex
);
7887 #else /* CONFIG_FAIR_GROUP_SCHED */
7889 void free_fair_sched_group(struct task_group
*tg
) { }
7891 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7896 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
7898 #endif /* CONFIG_FAIR_GROUP_SCHED */
7901 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
7903 struct sched_entity
*se
= &task
->se
;
7904 unsigned int rr_interval
= 0;
7907 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7910 if (rq
->cfs
.load
.weight
)
7911 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
7917 * All the scheduling class methods:
7919 const struct sched_class fair_sched_class
= {
7920 .next
= &idle_sched_class
,
7921 .enqueue_task
= enqueue_task_fair
,
7922 .dequeue_task
= dequeue_task_fair
,
7923 .yield_task
= yield_task_fair
,
7924 .yield_to_task
= yield_to_task_fair
,
7926 .check_preempt_curr
= check_preempt_wakeup
,
7928 .pick_next_task
= pick_next_task_fair
,
7929 .put_prev_task
= put_prev_task_fair
,
7932 .select_task_rq
= select_task_rq_fair
,
7933 .migrate_task_rq
= migrate_task_rq_fair
,
7935 .rq_online
= rq_online_fair
,
7936 .rq_offline
= rq_offline_fair
,
7938 .task_waking
= task_waking_fair
,
7941 .set_curr_task
= set_curr_task_fair
,
7942 .task_tick
= task_tick_fair
,
7943 .task_fork
= task_fork_fair
,
7945 .prio_changed
= prio_changed_fair
,
7946 .switched_from
= switched_from_fair
,
7947 .switched_to
= switched_to_fair
,
7949 .get_rr_interval
= get_rr_interval_fair
,
7951 #ifdef CONFIG_FAIR_GROUP_SCHED
7952 .task_move_group
= task_move_group_fair
,
7956 #ifdef CONFIG_SCHED_DEBUG
7957 void print_cfs_stats(struct seq_file
*m
, int cpu
)
7959 struct cfs_rq
*cfs_rq
;
7962 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
7963 print_cfs_rq(m
, cpu
, cfs_rq
);
7968 __init
void init_sched_fair_class(void)
7971 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7973 #ifdef CONFIG_NO_HZ_COMMON
7974 nohz
.next_balance
= jiffies
;
7975 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7976 cpu_notifier(sched_ilb_notifier
, 0);