2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/task_work.h>
32 #include <trace/events/sched.h>
37 * Targeted preemption latency for CPU-bound tasks:
38 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
40 * NOTE: this latency value is not the same as the concept of
41 * 'timeslice length' - timeslices in CFS are of variable length
42 * and have no persistent notion like in traditional, time-slice
43 * based scheduling concepts.
45 * (to see the precise effective timeslice length of your workload,
46 * run vmstat and monitor the context-switches (cs) field)
48 unsigned int sysctl_sched_latency
= 6000000ULL;
49 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
52 * The initial- and re-scaling of tunables is configurable
53 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
56 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
57 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
58 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
60 enum sched_tunable_scaling sysctl_sched_tunable_scaling
61 = SCHED_TUNABLESCALING_LOG
;
64 * Minimal preemption granularity for CPU-bound tasks:
65 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
67 unsigned int sysctl_sched_min_granularity
= 750000ULL;
68 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
71 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
73 static unsigned int sched_nr_latency
= 8;
76 * After fork, child runs first. If set to 0 (default) then
77 * parent will (try to) run first.
79 unsigned int sysctl_sched_child_runs_first __read_mostly
;
82 * SCHED_OTHER wake-up granularity.
83 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
85 * This option delays the preemption effects of decoupled workloads
86 * and reduces their over-scheduling. Synchronous workloads will still
87 * have immediate wakeup/sleep latencies.
89 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
90 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
92 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
95 * The exponential sliding window over which load is averaged for shares
99 unsigned int __read_mostly sysctl_sched_shares_window
= 10000000UL;
101 #ifdef CONFIG_CFS_BANDWIDTH
103 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
104 * each time a cfs_rq requests quota.
106 * Note: in the case that the slice exceeds the runtime remaining (either due
107 * to consumption or the quota being specified to be smaller than the slice)
108 * we will always only issue the remaining available time.
110 * default: 5 msec, units: microseconds
112 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
116 * Increase the granularity value when there are more CPUs,
117 * because with more CPUs the 'effective latency' as visible
118 * to users decreases. But the relationship is not linear,
119 * so pick a second-best guess by going with the log2 of the
122 * This idea comes from the SD scheduler of Con Kolivas:
124 static int get_update_sysctl_factor(void)
126 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
129 switch (sysctl_sched_tunable_scaling
) {
130 case SCHED_TUNABLESCALING_NONE
:
133 case SCHED_TUNABLESCALING_LINEAR
:
136 case SCHED_TUNABLESCALING_LOG
:
138 factor
= 1 + ilog2(cpus
);
145 static void update_sysctl(void)
147 unsigned int factor
= get_update_sysctl_factor();
149 #define SET_SYSCTL(name) \
150 (sysctl_##name = (factor) * normalized_sysctl_##name)
151 SET_SYSCTL(sched_min_granularity
);
152 SET_SYSCTL(sched_latency
);
153 SET_SYSCTL(sched_wakeup_granularity
);
157 void sched_init_granularity(void)
162 #if BITS_PER_LONG == 32
163 # define WMULT_CONST (~0UL)
165 # define WMULT_CONST (1UL << 32)
168 #define WMULT_SHIFT 32
171 * Shift right and round:
173 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
176 * delta *= weight / lw
179 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
180 struct load_weight
*lw
)
185 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
186 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
187 * 2^SCHED_LOAD_RESOLUTION.
189 if (likely(weight
> (1UL << SCHED_LOAD_RESOLUTION
)))
190 tmp
= (u64
)delta_exec
* scale_load_down(weight
);
192 tmp
= (u64
)delta_exec
;
194 if (!lw
->inv_weight
) {
195 unsigned long w
= scale_load_down(lw
->weight
);
197 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
199 else if (unlikely(!w
))
200 lw
->inv_weight
= WMULT_CONST
;
202 lw
->inv_weight
= WMULT_CONST
/ w
;
206 * Check whether we'd overflow the 64-bit multiplication:
208 if (unlikely(tmp
> WMULT_CONST
))
209 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
212 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
214 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
218 const struct sched_class fair_sched_class
;
220 /**************************************************************
221 * CFS operations on generic schedulable entities:
224 #ifdef CONFIG_FAIR_GROUP_SCHED
226 /* cpu runqueue to which this cfs_rq is attached */
227 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
232 /* An entity is a task if it doesn't "own" a runqueue */
233 #define entity_is_task(se) (!se->my_q)
235 static inline struct task_struct
*task_of(struct sched_entity
*se
)
237 #ifdef CONFIG_SCHED_DEBUG
238 WARN_ON_ONCE(!entity_is_task(se
));
240 return container_of(se
, struct task_struct
, se
);
243 /* Walk up scheduling entities hierarchy */
244 #define for_each_sched_entity(se) \
245 for (; se; se = se->parent)
247 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
252 /* runqueue on which this entity is (to be) queued */
253 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
258 /* runqueue "owned" by this group */
259 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
264 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
266 if (!cfs_rq
->on_list
) {
268 * Ensure we either appear before our parent (if already
269 * enqueued) or force our parent to appear after us when it is
270 * enqueued. The fact that we always enqueue bottom-up
271 * reduces this to two cases.
273 if (cfs_rq
->tg
->parent
&&
274 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
275 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
276 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
278 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
279 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
286 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
288 if (cfs_rq
->on_list
) {
289 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
294 /* Iterate thr' all leaf cfs_rq's on a runqueue */
295 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
296 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
298 /* Do the two (enqueued) entities belong to the same group ? */
300 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
302 if (se
->cfs_rq
== pse
->cfs_rq
)
308 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
313 /* return depth at which a sched entity is present in the hierarchy */
314 static inline int depth_se(struct sched_entity
*se
)
318 for_each_sched_entity(se
)
325 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
327 int se_depth
, pse_depth
;
330 * preemption test can be made between sibling entities who are in the
331 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
332 * both tasks until we find their ancestors who are siblings of common
336 /* First walk up until both entities are at same depth */
337 se_depth
= depth_se(*se
);
338 pse_depth
= depth_se(*pse
);
340 while (se_depth
> pse_depth
) {
342 *se
= parent_entity(*se
);
345 while (pse_depth
> se_depth
) {
347 *pse
= parent_entity(*pse
);
350 while (!is_same_group(*se
, *pse
)) {
351 *se
= parent_entity(*se
);
352 *pse
= parent_entity(*pse
);
356 #else /* !CONFIG_FAIR_GROUP_SCHED */
358 static inline struct task_struct
*task_of(struct sched_entity
*se
)
360 return container_of(se
, struct task_struct
, se
);
363 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
365 return container_of(cfs_rq
, struct rq
, cfs
);
368 #define entity_is_task(se) 1
370 #define for_each_sched_entity(se) \
371 for (; se; se = NULL)
373 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
375 return &task_rq(p
)->cfs
;
378 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
380 struct task_struct
*p
= task_of(se
);
381 struct rq
*rq
= task_rq(p
);
386 /* runqueue "owned" by this group */
387 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
392 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
396 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
400 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
401 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
404 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
409 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
415 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
419 #endif /* CONFIG_FAIR_GROUP_SCHED */
421 static __always_inline
422 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
);
424 /**************************************************************
425 * Scheduling class tree data structure manipulation methods:
428 static inline u64
max_vruntime(u64 min_vruntime
, u64 vruntime
)
430 s64 delta
= (s64
)(vruntime
- min_vruntime
);
432 min_vruntime
= vruntime
;
437 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
439 s64 delta
= (s64
)(vruntime
- min_vruntime
);
441 min_vruntime
= vruntime
;
446 static inline int entity_before(struct sched_entity
*a
,
447 struct sched_entity
*b
)
449 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
452 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
454 u64 vruntime
= cfs_rq
->min_vruntime
;
457 vruntime
= cfs_rq
->curr
->vruntime
;
459 if (cfs_rq
->rb_leftmost
) {
460 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
465 vruntime
= se
->vruntime
;
467 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
470 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
473 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
478 * Enqueue an entity into the rb-tree:
480 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
482 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
483 struct rb_node
*parent
= NULL
;
484 struct sched_entity
*entry
;
488 * Find the right place in the rbtree:
492 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
494 * We dont care about collisions. Nodes with
495 * the same key stay together.
497 if (entity_before(se
, entry
)) {
498 link
= &parent
->rb_left
;
500 link
= &parent
->rb_right
;
506 * Maintain a cache of leftmost tree entries (it is frequently
510 cfs_rq
->rb_leftmost
= &se
->run_node
;
512 rb_link_node(&se
->run_node
, parent
, link
);
513 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
516 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
518 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
519 struct rb_node
*next_node
;
521 next_node
= rb_next(&se
->run_node
);
522 cfs_rq
->rb_leftmost
= next_node
;
525 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
528 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
530 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
535 return rb_entry(left
, struct sched_entity
, run_node
);
538 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
540 struct rb_node
*next
= rb_next(&se
->run_node
);
545 return rb_entry(next
, struct sched_entity
, run_node
);
548 #ifdef CONFIG_SCHED_DEBUG
549 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
551 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
556 return rb_entry(last
, struct sched_entity
, run_node
);
559 /**************************************************************
560 * Scheduling class statistics methods:
563 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
564 void __user
*buffer
, size_t *lenp
,
567 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
568 int factor
= get_update_sysctl_factor();
573 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
574 sysctl_sched_min_granularity
);
576 #define WRT_SYSCTL(name) \
577 (normalized_sysctl_##name = sysctl_##name / (factor))
578 WRT_SYSCTL(sched_min_granularity
);
579 WRT_SYSCTL(sched_latency
);
580 WRT_SYSCTL(sched_wakeup_granularity
);
590 static inline unsigned long
591 calc_delta_fair(unsigned long delta
, struct sched_entity
*se
)
593 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
594 delta
= calc_delta_mine(delta
, NICE_0_LOAD
, &se
->load
);
600 * The idea is to set a period in which each task runs once.
602 * When there are too many tasks (sched_nr_latency) we have to stretch
603 * this period because otherwise the slices get too small.
605 * p = (nr <= nl) ? l : l*nr/nl
607 static u64
__sched_period(unsigned long nr_running
)
609 u64 period
= sysctl_sched_latency
;
610 unsigned long nr_latency
= sched_nr_latency
;
612 if (unlikely(nr_running
> nr_latency
)) {
613 period
= sysctl_sched_min_granularity
;
614 period
*= nr_running
;
621 * We calculate the wall-time slice from the period by taking a part
622 * proportional to the weight.
626 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
628 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
630 for_each_sched_entity(se
) {
631 struct load_weight
*load
;
632 struct load_weight lw
;
634 cfs_rq
= cfs_rq_of(se
);
635 load
= &cfs_rq
->load
;
637 if (unlikely(!se
->on_rq
)) {
640 update_load_add(&lw
, se
->load
.weight
);
643 slice
= calc_delta_mine(slice
, se
->load
.weight
, load
);
649 * We calculate the vruntime slice of a to be inserted task
653 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
655 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
658 static void update_cfs_load(struct cfs_rq
*cfs_rq
, int global_update
);
659 static void update_cfs_shares(struct cfs_rq
*cfs_rq
);
662 * Update the current task's runtime statistics. Skip current tasks that
663 * are not in our scheduling class.
666 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
667 unsigned long delta_exec
)
669 unsigned long delta_exec_weighted
;
671 schedstat_set(curr
->statistics
.exec_max
,
672 max((u64
)delta_exec
, curr
->statistics
.exec_max
));
674 curr
->sum_exec_runtime
+= delta_exec
;
675 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
676 delta_exec_weighted
= calc_delta_fair(delta_exec
, curr
);
678 curr
->vruntime
+= delta_exec_weighted
;
679 update_min_vruntime(cfs_rq
);
681 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
682 cfs_rq
->load_unacc_exec_time
+= delta_exec
;
686 static void update_curr(struct cfs_rq
*cfs_rq
)
688 struct sched_entity
*curr
= cfs_rq
->curr
;
689 u64 now
= rq_of(cfs_rq
)->clock_task
;
690 unsigned long delta_exec
;
696 * Get the amount of time the current task was running
697 * since the last time we changed load (this cannot
698 * overflow on 32 bits):
700 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
704 __update_curr(cfs_rq
, curr
, delta_exec
);
705 curr
->exec_start
= now
;
707 if (entity_is_task(curr
)) {
708 struct task_struct
*curtask
= task_of(curr
);
710 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
711 cpuacct_charge(curtask
, delta_exec
);
712 account_group_exec_runtime(curtask
, delta_exec
);
715 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
719 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
721 schedstat_set(se
->statistics
.wait_start
, rq_of(cfs_rq
)->clock
);
725 * Task is being enqueued - update stats:
727 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
730 * Are we enqueueing a waiting task? (for current tasks
731 * a dequeue/enqueue event is a NOP)
733 if (se
!= cfs_rq
->curr
)
734 update_stats_wait_start(cfs_rq
, se
);
738 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
740 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
741 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
));
742 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
743 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
744 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
);
745 #ifdef CONFIG_SCHEDSTATS
746 if (entity_is_task(se
)) {
747 trace_sched_stat_wait(task_of(se
),
748 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
);
751 schedstat_set(se
->statistics
.wait_start
, 0);
755 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
758 * Mark the end of the wait period if dequeueing a
761 if (se
!= cfs_rq
->curr
)
762 update_stats_wait_end(cfs_rq
, se
);
766 * We are picking a new current task - update its stats:
769 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
772 * We are starting a new run period:
774 se
->exec_start
= rq_of(cfs_rq
)->clock_task
;
777 /**************************************************
778 * Scheduling class queueing methods:
781 #ifdef CONFIG_NUMA_BALANCING
783 * numa task sample period in ms
785 unsigned int sysctl_numa_balancing_scan_period_min
= 100;
786 unsigned int sysctl_numa_balancing_scan_period_max
= 100*16;
788 /* Portion of address space to scan in MB */
789 unsigned int sysctl_numa_balancing_scan_size
= 256;
791 static void task_numa_placement(struct task_struct
*p
)
793 int seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
795 if (p
->numa_scan_seq
== seq
)
797 p
->numa_scan_seq
= seq
;
799 /* FIXME: Scheduling placement policy hints go here */
803 * Got a PROT_NONE fault for a page on @node.
805 void task_numa_fault(int node
, int pages
)
807 struct task_struct
*p
= current
;
809 /* FIXME: Allocate task-specific structure for placement policy here */
811 task_numa_placement(p
);
814 static void reset_ptenuma_scan(struct task_struct
*p
)
816 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
817 p
->mm
->numa_scan_offset
= 0;
821 * The expensive part of numa migration is done from task_work context.
822 * Triggered from task_tick_numa().
824 void task_numa_work(struct callback_head
*work
)
826 unsigned long migrate
, next_scan
, now
= jiffies
;
827 struct task_struct
*p
= current
;
828 struct mm_struct
*mm
= p
->mm
;
829 struct vm_area_struct
*vma
;
830 unsigned long offset
, end
;
833 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
835 work
->next
= work
; /* protect against double add */
837 * Who cares about NUMA placement when they're dying.
839 * NOTE: make sure not to dereference p->mm before this check,
840 * exit_task_work() happens _after_ exit_mm() so we could be called
841 * without p->mm even though we still had it when we enqueued this
844 if (p
->flags
& PF_EXITING
)
848 * Enforce maximal scan/migration frequency..
850 migrate
= mm
->numa_next_scan
;
851 if (time_before(now
, migrate
))
854 if (p
->numa_scan_period
== 0)
855 p
->numa_scan_period
= sysctl_numa_balancing_scan_period_min
;
857 next_scan
= now
+ 2*msecs_to_jiffies(p
->numa_scan_period
);
858 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
861 offset
= mm
->numa_scan_offset
;
862 length
= sysctl_numa_balancing_scan_size
;
865 down_read(&mm
->mmap_sem
);
866 vma
= find_vma(mm
, offset
);
868 reset_ptenuma_scan(p
);
872 for (; vma
&& length
> 0; vma
= vma
->vm_next
) {
873 if (!vma_migratable(vma
))
876 /* Skip small VMAs. They are not likely to be of relevance */
877 if (((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) < HPAGE_PMD_NR
)
880 offset
= max(offset
, vma
->vm_start
);
881 end
= min(ALIGN(offset
+ length
, HPAGE_SIZE
), vma
->vm_end
);
882 length
-= end
- offset
;
884 change_prot_numa(vma
, offset
, end
);
890 * It is possible to reach the end of the VMA list but the last few VMAs are
891 * not guaranteed to the vma_migratable. If they are not, we would find the
892 * !migratable VMA on the next scan but not reset the scanner to the start
896 mm
->numa_scan_offset
= offset
;
898 reset_ptenuma_scan(p
);
899 up_read(&mm
->mmap_sem
);
903 * Drive the periodic memory faults..
905 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
907 struct callback_head
*work
= &curr
->numa_work
;
911 * We don't care about NUMA placement if we don't have memory.
913 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
917 * Using runtime rather than walltime has the dual advantage that
918 * we (mostly) drive the selection from busy threads and that the
919 * task needs to have done some actual work before we bother with
922 now
= curr
->se
.sum_exec_runtime
;
923 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
925 if (now
- curr
->node_stamp
> period
) {
926 curr
->node_stamp
= now
;
928 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
929 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
930 task_work_add(curr
, work
, true);
935 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
938 #endif /* CONFIG_NUMA_BALANCING */
941 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
943 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
944 if (!parent_entity(se
))
945 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
947 if (entity_is_task(se
))
948 list_add(&se
->group_node
, &rq_of(cfs_rq
)->cfs_tasks
);
950 cfs_rq
->nr_running
++;
954 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
956 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
957 if (!parent_entity(se
))
958 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
959 if (entity_is_task(se
))
960 list_del_init(&se
->group_node
);
961 cfs_rq
->nr_running
--;
964 #ifdef CONFIG_FAIR_GROUP_SCHED
965 /* we need this in update_cfs_load and load-balance functions below */
966 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
968 static void update_cfs_rq_load_contribution(struct cfs_rq
*cfs_rq
,
971 struct task_group
*tg
= cfs_rq
->tg
;
974 load_avg
= div64_u64(cfs_rq
->load_avg
, cfs_rq
->load_period
+1);
975 load_avg
-= cfs_rq
->load_contribution
;
977 if (global_update
|| abs(load_avg
) > cfs_rq
->load_contribution
/ 8) {
978 atomic_add(load_avg
, &tg
->load_weight
);
979 cfs_rq
->load_contribution
+= load_avg
;
983 static void update_cfs_load(struct cfs_rq
*cfs_rq
, int global_update
)
985 u64 period
= sysctl_sched_shares_window
;
987 unsigned long load
= cfs_rq
->load
.weight
;
989 if (cfs_rq
->tg
== &root_task_group
|| throttled_hierarchy(cfs_rq
))
992 now
= rq_of(cfs_rq
)->clock_task
;
993 delta
= now
- cfs_rq
->load_stamp
;
995 /* truncate load history at 4 idle periods */
996 if (cfs_rq
->load_stamp
> cfs_rq
->load_last
&&
997 now
- cfs_rq
->load_last
> 4 * period
) {
998 cfs_rq
->load_period
= 0;
999 cfs_rq
->load_avg
= 0;
1003 cfs_rq
->load_stamp
= now
;
1004 cfs_rq
->load_unacc_exec_time
= 0;
1005 cfs_rq
->load_period
+= delta
;
1007 cfs_rq
->load_last
= now
;
1008 cfs_rq
->load_avg
+= delta
* load
;
1011 /* consider updating load contribution on each fold or truncate */
1012 if (global_update
|| cfs_rq
->load_period
> period
1013 || !cfs_rq
->load_period
)
1014 update_cfs_rq_load_contribution(cfs_rq
, global_update
);
1016 while (cfs_rq
->load_period
> period
) {
1018 * Inline assembly required to prevent the compiler
1019 * optimising this loop into a divmod call.
1020 * See __iter_div_u64_rem() for another example of this.
1022 asm("" : "+rm" (cfs_rq
->load_period
));
1023 cfs_rq
->load_period
/= 2;
1024 cfs_rq
->load_avg
/= 2;
1027 if (!cfs_rq
->curr
&& !cfs_rq
->nr_running
&& !cfs_rq
->load_avg
)
1028 list_del_leaf_cfs_rq(cfs_rq
);
1031 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
1036 * Use this CPU's actual weight instead of the last load_contribution
1037 * to gain a more accurate current total weight. See
1038 * update_cfs_rq_load_contribution().
1040 tg_weight
= atomic_read(&tg
->load_weight
);
1041 tg_weight
-= cfs_rq
->load_contribution
;
1042 tg_weight
+= cfs_rq
->load
.weight
;
1047 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1049 long tg_weight
, load
, shares
;
1051 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
1052 load
= cfs_rq
->load
.weight
;
1054 shares
= (tg
->shares
* load
);
1056 shares
/= tg_weight
;
1058 if (shares
< MIN_SHARES
)
1059 shares
= MIN_SHARES
;
1060 if (shares
> tg
->shares
)
1061 shares
= tg
->shares
;
1066 static void update_entity_shares_tick(struct cfs_rq
*cfs_rq
)
1068 if (cfs_rq
->load_unacc_exec_time
> sysctl_sched_shares_window
) {
1069 update_cfs_load(cfs_rq
, 0);
1070 update_cfs_shares(cfs_rq
);
1073 # else /* CONFIG_SMP */
1074 static void update_cfs_load(struct cfs_rq
*cfs_rq
, int global_update
)
1078 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1083 static inline void update_entity_shares_tick(struct cfs_rq
*cfs_rq
)
1086 # endif /* CONFIG_SMP */
1087 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
1088 unsigned long weight
)
1091 /* commit outstanding execution time */
1092 if (cfs_rq
->curr
== se
)
1093 update_curr(cfs_rq
);
1094 account_entity_dequeue(cfs_rq
, se
);
1097 update_load_set(&se
->load
, weight
);
1100 account_entity_enqueue(cfs_rq
, se
);
1103 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
1105 struct task_group
*tg
;
1106 struct sched_entity
*se
;
1110 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
1111 if (!se
|| throttled_hierarchy(cfs_rq
))
1114 if (likely(se
->load
.weight
== tg
->shares
))
1117 shares
= calc_cfs_shares(cfs_rq
, tg
);
1119 reweight_entity(cfs_rq_of(se
), se
, shares
);
1121 #else /* CONFIG_FAIR_GROUP_SCHED */
1122 static void update_cfs_load(struct cfs_rq
*cfs_rq
, int global_update
)
1126 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
1130 static inline void update_entity_shares_tick(struct cfs_rq
*cfs_rq
)
1133 #endif /* CONFIG_FAIR_GROUP_SCHED */
1135 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1137 #ifdef CONFIG_SCHEDSTATS
1138 struct task_struct
*tsk
= NULL
;
1140 if (entity_is_task(se
))
1143 if (se
->statistics
.sleep_start
) {
1144 u64 delta
= rq_of(cfs_rq
)->clock
- se
->statistics
.sleep_start
;
1149 if (unlikely(delta
> se
->statistics
.sleep_max
))
1150 se
->statistics
.sleep_max
= delta
;
1152 se
->statistics
.sleep_start
= 0;
1153 se
->statistics
.sum_sleep_runtime
+= delta
;
1156 account_scheduler_latency(tsk
, delta
>> 10, 1);
1157 trace_sched_stat_sleep(tsk
, delta
);
1160 if (se
->statistics
.block_start
) {
1161 u64 delta
= rq_of(cfs_rq
)->clock
- se
->statistics
.block_start
;
1166 if (unlikely(delta
> se
->statistics
.block_max
))
1167 se
->statistics
.block_max
= delta
;
1169 se
->statistics
.block_start
= 0;
1170 se
->statistics
.sum_sleep_runtime
+= delta
;
1173 if (tsk
->in_iowait
) {
1174 se
->statistics
.iowait_sum
+= delta
;
1175 se
->statistics
.iowait_count
++;
1176 trace_sched_stat_iowait(tsk
, delta
);
1179 trace_sched_stat_blocked(tsk
, delta
);
1182 * Blocking time is in units of nanosecs, so shift by
1183 * 20 to get a milliseconds-range estimation of the
1184 * amount of time that the task spent sleeping:
1186 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
1187 profile_hits(SLEEP_PROFILING
,
1188 (void *)get_wchan(tsk
),
1191 account_scheduler_latency(tsk
, delta
>> 10, 0);
1197 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1199 #ifdef CONFIG_SCHED_DEBUG
1200 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
1205 if (d
> 3*sysctl_sched_latency
)
1206 schedstat_inc(cfs_rq
, nr_spread_over
);
1211 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
1213 u64 vruntime
= cfs_rq
->min_vruntime
;
1216 * The 'current' period is already promised to the current tasks,
1217 * however the extra weight of the new task will slow them down a
1218 * little, place the new task so that it fits in the slot that
1219 * stays open at the end.
1221 if (initial
&& sched_feat(START_DEBIT
))
1222 vruntime
+= sched_vslice(cfs_rq
, se
);
1224 /* sleeps up to a single latency don't count. */
1226 unsigned long thresh
= sysctl_sched_latency
;
1229 * Halve their sleep time's effect, to allow
1230 * for a gentler effect of sleepers:
1232 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
1238 /* ensure we never gain time by being placed backwards. */
1239 vruntime
= max_vruntime(se
->vruntime
, vruntime
);
1241 se
->vruntime
= vruntime
;
1244 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
1247 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1250 * Update the normalized vruntime before updating min_vruntime
1251 * through callig update_curr().
1253 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
1254 se
->vruntime
+= cfs_rq
->min_vruntime
;
1257 * Update run-time statistics of the 'current'.
1259 update_curr(cfs_rq
);
1260 update_cfs_load(cfs_rq
, 0);
1261 account_entity_enqueue(cfs_rq
, se
);
1262 update_cfs_shares(cfs_rq
);
1264 if (flags
& ENQUEUE_WAKEUP
) {
1265 place_entity(cfs_rq
, se
, 0);
1266 enqueue_sleeper(cfs_rq
, se
);
1269 update_stats_enqueue(cfs_rq
, se
);
1270 check_spread(cfs_rq
, se
);
1271 if (se
!= cfs_rq
->curr
)
1272 __enqueue_entity(cfs_rq
, se
);
1275 if (cfs_rq
->nr_running
== 1) {
1276 list_add_leaf_cfs_rq(cfs_rq
);
1277 check_enqueue_throttle(cfs_rq
);
1281 static void __clear_buddies_last(struct sched_entity
*se
)
1283 for_each_sched_entity(se
) {
1284 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1285 if (cfs_rq
->last
== se
)
1286 cfs_rq
->last
= NULL
;
1292 static void __clear_buddies_next(struct sched_entity
*se
)
1294 for_each_sched_entity(se
) {
1295 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1296 if (cfs_rq
->next
== se
)
1297 cfs_rq
->next
= NULL
;
1303 static void __clear_buddies_skip(struct sched_entity
*se
)
1305 for_each_sched_entity(se
) {
1306 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1307 if (cfs_rq
->skip
== se
)
1308 cfs_rq
->skip
= NULL
;
1314 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1316 if (cfs_rq
->last
== se
)
1317 __clear_buddies_last(se
);
1319 if (cfs_rq
->next
== se
)
1320 __clear_buddies_next(se
);
1322 if (cfs_rq
->skip
== se
)
1323 __clear_buddies_skip(se
);
1326 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
1329 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1332 * Update run-time statistics of the 'current'.
1334 update_curr(cfs_rq
);
1336 update_stats_dequeue(cfs_rq
, se
);
1337 if (flags
& DEQUEUE_SLEEP
) {
1338 #ifdef CONFIG_SCHEDSTATS
1339 if (entity_is_task(se
)) {
1340 struct task_struct
*tsk
= task_of(se
);
1342 if (tsk
->state
& TASK_INTERRUPTIBLE
)
1343 se
->statistics
.sleep_start
= rq_of(cfs_rq
)->clock
;
1344 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
1345 se
->statistics
.block_start
= rq_of(cfs_rq
)->clock
;
1350 clear_buddies(cfs_rq
, se
);
1352 if (se
!= cfs_rq
->curr
)
1353 __dequeue_entity(cfs_rq
, se
);
1355 update_cfs_load(cfs_rq
, 0);
1356 account_entity_dequeue(cfs_rq
, se
);
1359 * Normalize the entity after updating the min_vruntime because the
1360 * update can refer to the ->curr item and we need to reflect this
1361 * movement in our normalized position.
1363 if (!(flags
& DEQUEUE_SLEEP
))
1364 se
->vruntime
-= cfs_rq
->min_vruntime
;
1366 /* return excess runtime on last dequeue */
1367 return_cfs_rq_runtime(cfs_rq
);
1369 update_min_vruntime(cfs_rq
);
1370 update_cfs_shares(cfs_rq
);
1374 * Preempt the current task with a newly woken task if needed:
1377 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
1379 unsigned long ideal_runtime
, delta_exec
;
1380 struct sched_entity
*se
;
1383 ideal_runtime
= sched_slice(cfs_rq
, curr
);
1384 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
1385 if (delta_exec
> ideal_runtime
) {
1386 resched_task(rq_of(cfs_rq
)->curr
);
1388 * The current task ran long enough, ensure it doesn't get
1389 * re-elected due to buddy favours.
1391 clear_buddies(cfs_rq
, curr
);
1396 * Ensure that a task that missed wakeup preemption by a
1397 * narrow margin doesn't have to wait for a full slice.
1398 * This also mitigates buddy induced latencies under load.
1400 if (delta_exec
< sysctl_sched_min_granularity
)
1403 se
= __pick_first_entity(cfs_rq
);
1404 delta
= curr
->vruntime
- se
->vruntime
;
1409 if (delta
> ideal_runtime
)
1410 resched_task(rq_of(cfs_rq
)->curr
);
1414 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1416 /* 'current' is not kept within the tree. */
1419 * Any task has to be enqueued before it get to execute on
1420 * a CPU. So account for the time it spent waiting on the
1423 update_stats_wait_end(cfs_rq
, se
);
1424 __dequeue_entity(cfs_rq
, se
);
1427 update_stats_curr_start(cfs_rq
, se
);
1429 #ifdef CONFIG_SCHEDSTATS
1431 * Track our maximum slice length, if the CPU's load is at
1432 * least twice that of our own weight (i.e. dont track it
1433 * when there are only lesser-weight tasks around):
1435 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
1436 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
1437 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
1440 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
1444 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
1447 * Pick the next process, keeping these things in mind, in this order:
1448 * 1) keep things fair between processes/task groups
1449 * 2) pick the "next" process, since someone really wants that to run
1450 * 3) pick the "last" process, for cache locality
1451 * 4) do not run the "skip" process, if something else is available
1453 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
1455 struct sched_entity
*se
= __pick_first_entity(cfs_rq
);
1456 struct sched_entity
*left
= se
;
1459 * Avoid running the skip buddy, if running something else can
1460 * be done without getting too unfair.
1462 if (cfs_rq
->skip
== se
) {
1463 struct sched_entity
*second
= __pick_next_entity(se
);
1464 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
1469 * Prefer last buddy, try to return the CPU to a preempted task.
1471 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
1475 * Someone really wants this to run. If it's not unfair, run it.
1477 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
1480 clear_buddies(cfs_rq
, se
);
1485 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
1487 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
1490 * If still on the runqueue then deactivate_task()
1491 * was not called and update_curr() has to be done:
1494 update_curr(cfs_rq
);
1496 /* throttle cfs_rqs exceeding runtime */
1497 check_cfs_rq_runtime(cfs_rq
);
1499 check_spread(cfs_rq
, prev
);
1501 update_stats_wait_start(cfs_rq
, prev
);
1502 /* Put 'current' back into the tree. */
1503 __enqueue_entity(cfs_rq
, prev
);
1505 cfs_rq
->curr
= NULL
;
1509 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
1512 * Update run-time statistics of the 'current'.
1514 update_curr(cfs_rq
);
1517 * Update share accounting for long-running entities.
1519 update_entity_shares_tick(cfs_rq
);
1521 #ifdef CONFIG_SCHED_HRTICK
1523 * queued ticks are scheduled to match the slice, so don't bother
1524 * validating it and just reschedule.
1527 resched_task(rq_of(cfs_rq
)->curr
);
1531 * don't let the period tick interfere with the hrtick preemption
1533 if (!sched_feat(DOUBLE_TICK
) &&
1534 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
1538 if (cfs_rq
->nr_running
> 1)
1539 check_preempt_tick(cfs_rq
, curr
);
1543 /**************************************************
1544 * CFS bandwidth control machinery
1547 #ifdef CONFIG_CFS_BANDWIDTH
1549 #ifdef HAVE_JUMP_LABEL
1550 static struct static_key __cfs_bandwidth_used
;
1552 static inline bool cfs_bandwidth_used(void)
1554 return static_key_false(&__cfs_bandwidth_used
);
1557 void account_cfs_bandwidth_used(int enabled
, int was_enabled
)
1559 /* only need to count groups transitioning between enabled/!enabled */
1560 if (enabled
&& !was_enabled
)
1561 static_key_slow_inc(&__cfs_bandwidth_used
);
1562 else if (!enabled
&& was_enabled
)
1563 static_key_slow_dec(&__cfs_bandwidth_used
);
1565 #else /* HAVE_JUMP_LABEL */
1566 static bool cfs_bandwidth_used(void)
1571 void account_cfs_bandwidth_used(int enabled
, int was_enabled
) {}
1572 #endif /* HAVE_JUMP_LABEL */
1575 * default period for cfs group bandwidth.
1576 * default: 0.1s, units: nanoseconds
1578 static inline u64
default_cfs_period(void)
1580 return 100000000ULL;
1583 static inline u64
sched_cfs_bandwidth_slice(void)
1585 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
1589 * Replenish runtime according to assigned quota and update expiration time.
1590 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1591 * additional synchronization around rq->lock.
1593 * requires cfs_b->lock
1595 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
1599 if (cfs_b
->quota
== RUNTIME_INF
)
1602 now
= sched_clock_cpu(smp_processor_id());
1603 cfs_b
->runtime
= cfs_b
->quota
;
1604 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
1607 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
1609 return &tg
->cfs_bandwidth
;
1612 /* returns 0 on failure to allocate runtime */
1613 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
1615 struct task_group
*tg
= cfs_rq
->tg
;
1616 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
1617 u64 amount
= 0, min_amount
, expires
;
1619 /* note: this is a positive sum as runtime_remaining <= 0 */
1620 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
1622 raw_spin_lock(&cfs_b
->lock
);
1623 if (cfs_b
->quota
== RUNTIME_INF
)
1624 amount
= min_amount
;
1627 * If the bandwidth pool has become inactive, then at least one
1628 * period must have elapsed since the last consumption.
1629 * Refresh the global state and ensure bandwidth timer becomes
1632 if (!cfs_b
->timer_active
) {
1633 __refill_cfs_bandwidth_runtime(cfs_b
);
1634 __start_cfs_bandwidth(cfs_b
);
1637 if (cfs_b
->runtime
> 0) {
1638 amount
= min(cfs_b
->runtime
, min_amount
);
1639 cfs_b
->runtime
-= amount
;
1643 expires
= cfs_b
->runtime_expires
;
1644 raw_spin_unlock(&cfs_b
->lock
);
1646 cfs_rq
->runtime_remaining
+= amount
;
1648 * we may have advanced our local expiration to account for allowed
1649 * spread between our sched_clock and the one on which runtime was
1652 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
1653 cfs_rq
->runtime_expires
= expires
;
1655 return cfs_rq
->runtime_remaining
> 0;
1659 * Note: This depends on the synchronization provided by sched_clock and the
1660 * fact that rq->clock snapshots this value.
1662 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
1664 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
1665 struct rq
*rq
= rq_of(cfs_rq
);
1667 /* if the deadline is ahead of our clock, nothing to do */
1668 if (likely((s64
)(rq
->clock
- cfs_rq
->runtime_expires
) < 0))
1671 if (cfs_rq
->runtime_remaining
< 0)
1675 * If the local deadline has passed we have to consider the
1676 * possibility that our sched_clock is 'fast' and the global deadline
1677 * has not truly expired.
1679 * Fortunately we can check determine whether this the case by checking
1680 * whether the global deadline has advanced.
1683 if ((s64
)(cfs_rq
->runtime_expires
- cfs_b
->runtime_expires
) >= 0) {
1684 /* extend local deadline, drift is bounded above by 2 ticks */
1685 cfs_rq
->runtime_expires
+= TICK_NSEC
;
1687 /* global deadline is ahead, expiration has passed */
1688 cfs_rq
->runtime_remaining
= 0;
1692 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
,
1693 unsigned long delta_exec
)
1695 /* dock delta_exec before expiring quota (as it could span periods) */
1696 cfs_rq
->runtime_remaining
-= delta_exec
;
1697 expire_cfs_rq_runtime(cfs_rq
);
1699 if (likely(cfs_rq
->runtime_remaining
> 0))
1703 * if we're unable to extend our runtime we resched so that the active
1704 * hierarchy can be throttled
1706 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
1707 resched_task(rq_of(cfs_rq
)->curr
);
1710 static __always_inline
1711 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
)
1713 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
1716 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
1719 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
1721 return cfs_bandwidth_used() && cfs_rq
->throttled
;
1724 /* check whether cfs_rq, or any parent, is throttled */
1725 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
1727 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
1731 * Ensure that neither of the group entities corresponding to src_cpu or
1732 * dest_cpu are members of a throttled hierarchy when performing group
1733 * load-balance operations.
1735 static inline int throttled_lb_pair(struct task_group
*tg
,
1736 int src_cpu
, int dest_cpu
)
1738 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
1740 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
1741 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
1743 return throttled_hierarchy(src_cfs_rq
) ||
1744 throttled_hierarchy(dest_cfs_rq
);
1747 /* updated child weight may affect parent so we have to do this bottom up */
1748 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
1750 struct rq
*rq
= data
;
1751 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
1753 cfs_rq
->throttle_count
--;
1755 if (!cfs_rq
->throttle_count
) {
1756 u64 delta
= rq
->clock_task
- cfs_rq
->load_stamp
;
1758 /* leaving throttled state, advance shares averaging windows */
1759 cfs_rq
->load_stamp
+= delta
;
1760 cfs_rq
->load_last
+= delta
;
1762 /* update entity weight now that we are on_rq again */
1763 update_cfs_shares(cfs_rq
);
1770 static int tg_throttle_down(struct task_group
*tg
, void *data
)
1772 struct rq
*rq
= data
;
1773 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
1775 /* group is entering throttled state, record last load */
1776 if (!cfs_rq
->throttle_count
)
1777 update_cfs_load(cfs_rq
, 0);
1778 cfs_rq
->throttle_count
++;
1783 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
1785 struct rq
*rq
= rq_of(cfs_rq
);
1786 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
1787 struct sched_entity
*se
;
1788 long task_delta
, dequeue
= 1;
1790 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
1792 /* account load preceding throttle */
1794 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
1797 task_delta
= cfs_rq
->h_nr_running
;
1798 for_each_sched_entity(se
) {
1799 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
1800 /* throttled entity or throttle-on-deactivate */
1805 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
1806 qcfs_rq
->h_nr_running
-= task_delta
;
1808 if (qcfs_rq
->load
.weight
)
1813 rq
->nr_running
-= task_delta
;
1815 cfs_rq
->throttled
= 1;
1816 cfs_rq
->throttled_timestamp
= rq
->clock
;
1817 raw_spin_lock(&cfs_b
->lock
);
1818 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
1819 raw_spin_unlock(&cfs_b
->lock
);
1822 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
1824 struct rq
*rq
= rq_of(cfs_rq
);
1825 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
1826 struct sched_entity
*se
;
1830 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
1832 cfs_rq
->throttled
= 0;
1833 raw_spin_lock(&cfs_b
->lock
);
1834 cfs_b
->throttled_time
+= rq
->clock
- cfs_rq
->throttled_timestamp
;
1835 list_del_rcu(&cfs_rq
->throttled_list
);
1836 raw_spin_unlock(&cfs_b
->lock
);
1837 cfs_rq
->throttled_timestamp
= 0;
1839 update_rq_clock(rq
);
1840 /* update hierarchical throttle state */
1841 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
1843 if (!cfs_rq
->load
.weight
)
1846 task_delta
= cfs_rq
->h_nr_running
;
1847 for_each_sched_entity(se
) {
1851 cfs_rq
= cfs_rq_of(se
);
1853 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
1854 cfs_rq
->h_nr_running
+= task_delta
;
1856 if (cfs_rq_throttled(cfs_rq
))
1861 rq
->nr_running
+= task_delta
;
1863 /* determine whether we need to wake up potentially idle cpu */
1864 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
1865 resched_task(rq
->curr
);
1868 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
1869 u64 remaining
, u64 expires
)
1871 struct cfs_rq
*cfs_rq
;
1872 u64 runtime
= remaining
;
1875 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
1877 struct rq
*rq
= rq_of(cfs_rq
);
1879 raw_spin_lock(&rq
->lock
);
1880 if (!cfs_rq_throttled(cfs_rq
))
1883 runtime
= -cfs_rq
->runtime_remaining
+ 1;
1884 if (runtime
> remaining
)
1885 runtime
= remaining
;
1886 remaining
-= runtime
;
1888 cfs_rq
->runtime_remaining
+= runtime
;
1889 cfs_rq
->runtime_expires
= expires
;
1891 /* we check whether we're throttled above */
1892 if (cfs_rq
->runtime_remaining
> 0)
1893 unthrottle_cfs_rq(cfs_rq
);
1896 raw_spin_unlock(&rq
->lock
);
1907 * Responsible for refilling a task_group's bandwidth and unthrottling its
1908 * cfs_rqs as appropriate. If there has been no activity within the last
1909 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1910 * used to track this state.
1912 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
1914 u64 runtime
, runtime_expires
;
1915 int idle
= 1, throttled
;
1917 raw_spin_lock(&cfs_b
->lock
);
1918 /* no need to continue the timer with no bandwidth constraint */
1919 if (cfs_b
->quota
== RUNTIME_INF
)
1922 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
1923 /* idle depends on !throttled (for the case of a large deficit) */
1924 idle
= cfs_b
->idle
&& !throttled
;
1925 cfs_b
->nr_periods
+= overrun
;
1927 /* if we're going inactive then everything else can be deferred */
1931 __refill_cfs_bandwidth_runtime(cfs_b
);
1934 /* mark as potentially idle for the upcoming period */
1939 /* account preceding periods in which throttling occurred */
1940 cfs_b
->nr_throttled
+= overrun
;
1943 * There are throttled entities so we must first use the new bandwidth
1944 * to unthrottle them before making it generally available. This
1945 * ensures that all existing debts will be paid before a new cfs_rq is
1948 runtime
= cfs_b
->runtime
;
1949 runtime_expires
= cfs_b
->runtime_expires
;
1953 * This check is repeated as we are holding onto the new bandwidth
1954 * while we unthrottle. This can potentially race with an unthrottled
1955 * group trying to acquire new bandwidth from the global pool.
1957 while (throttled
&& runtime
> 0) {
1958 raw_spin_unlock(&cfs_b
->lock
);
1959 /* we can't nest cfs_b->lock while distributing bandwidth */
1960 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
1962 raw_spin_lock(&cfs_b
->lock
);
1964 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
1967 /* return (any) remaining runtime */
1968 cfs_b
->runtime
= runtime
;
1970 * While we are ensured activity in the period following an
1971 * unthrottle, this also covers the case in which the new bandwidth is
1972 * insufficient to cover the existing bandwidth deficit. (Forcing the
1973 * timer to remain active while there are any throttled entities.)
1978 cfs_b
->timer_active
= 0;
1979 raw_spin_unlock(&cfs_b
->lock
);
1984 /* a cfs_rq won't donate quota below this amount */
1985 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
1986 /* minimum remaining period time to redistribute slack quota */
1987 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
1988 /* how long we wait to gather additional slack before distributing */
1989 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
1991 /* are we near the end of the current quota period? */
1992 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
1994 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
1997 /* if the call-back is running a quota refresh is already occurring */
1998 if (hrtimer_callback_running(refresh_timer
))
2001 /* is a quota refresh about to occur? */
2002 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
2003 if (remaining
< min_expire
)
2009 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
2011 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
2013 /* if there's a quota refresh soon don't bother with slack */
2014 if (runtime_refresh_within(cfs_b
, min_left
))
2017 start_bandwidth_timer(&cfs_b
->slack_timer
,
2018 ns_to_ktime(cfs_bandwidth_slack_period
));
2021 /* we know any runtime found here is valid as update_curr() precedes return */
2022 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2024 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2025 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
2027 if (slack_runtime
<= 0)
2030 raw_spin_lock(&cfs_b
->lock
);
2031 if (cfs_b
->quota
!= RUNTIME_INF
&&
2032 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
2033 cfs_b
->runtime
+= slack_runtime
;
2035 /* we are under rq->lock, defer unthrottling using a timer */
2036 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
2037 !list_empty(&cfs_b
->throttled_cfs_rq
))
2038 start_cfs_slack_bandwidth(cfs_b
);
2040 raw_spin_unlock(&cfs_b
->lock
);
2042 /* even if it's not valid for return we don't want to try again */
2043 cfs_rq
->runtime_remaining
-= slack_runtime
;
2046 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2048 if (!cfs_bandwidth_used())
2051 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
2054 __return_cfs_rq_runtime(cfs_rq
);
2058 * This is done with a timer (instead of inline with bandwidth return) since
2059 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2061 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
2063 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
2066 /* confirm we're still not at a refresh boundary */
2067 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
))
2070 raw_spin_lock(&cfs_b
->lock
);
2071 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
) {
2072 runtime
= cfs_b
->runtime
;
2075 expires
= cfs_b
->runtime_expires
;
2076 raw_spin_unlock(&cfs_b
->lock
);
2081 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
2083 raw_spin_lock(&cfs_b
->lock
);
2084 if (expires
== cfs_b
->runtime_expires
)
2085 cfs_b
->runtime
= runtime
;
2086 raw_spin_unlock(&cfs_b
->lock
);
2090 * When a group wakes up we want to make sure that its quota is not already
2091 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2092 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2094 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
2096 if (!cfs_bandwidth_used())
2099 /* an active group must be handled by the update_curr()->put() path */
2100 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
2103 /* ensure the group is not already throttled */
2104 if (cfs_rq_throttled(cfs_rq
))
2107 /* update runtime allocation */
2108 account_cfs_rq_runtime(cfs_rq
, 0);
2109 if (cfs_rq
->runtime_remaining
<= 0)
2110 throttle_cfs_rq(cfs_rq
);
2113 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2114 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2116 if (!cfs_bandwidth_used())
2119 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
2123 * it's possible for a throttled entity to be forced into a running
2124 * state (e.g. set_curr_task), in this case we're finished.
2126 if (cfs_rq_throttled(cfs_rq
))
2129 throttle_cfs_rq(cfs_rq
);
2132 static inline u64
default_cfs_period(void);
2133 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
);
2134 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
);
2136 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
2138 struct cfs_bandwidth
*cfs_b
=
2139 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
2140 do_sched_cfs_slack_timer(cfs_b
);
2142 return HRTIMER_NORESTART
;
2145 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
2147 struct cfs_bandwidth
*cfs_b
=
2148 container_of(timer
, struct cfs_bandwidth
, period_timer
);
2154 now
= hrtimer_cb_get_time(timer
);
2155 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
2160 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
2163 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
2166 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2168 raw_spin_lock_init(&cfs_b
->lock
);
2170 cfs_b
->quota
= RUNTIME_INF
;
2171 cfs_b
->period
= ns_to_ktime(default_cfs_period());
2173 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
2174 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2175 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
2176 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2177 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
2180 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2182 cfs_rq
->runtime_enabled
= 0;
2183 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
2186 /* requires cfs_b->lock, may release to reprogram timer */
2187 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2190 * The timer may be active because we're trying to set a new bandwidth
2191 * period or because we're racing with the tear-down path
2192 * (timer_active==0 becomes visible before the hrtimer call-back
2193 * terminates). In either case we ensure that it's re-programmed
2195 while (unlikely(hrtimer_active(&cfs_b
->period_timer
))) {
2196 raw_spin_unlock(&cfs_b
->lock
);
2197 /* ensure cfs_b->lock is available while we wait */
2198 hrtimer_cancel(&cfs_b
->period_timer
);
2200 raw_spin_lock(&cfs_b
->lock
);
2201 /* if someone else restarted the timer then we're done */
2202 if (cfs_b
->timer_active
)
2206 cfs_b
->timer_active
= 1;
2207 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
2210 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2212 hrtimer_cancel(&cfs_b
->period_timer
);
2213 hrtimer_cancel(&cfs_b
->slack_timer
);
2216 static void unthrottle_offline_cfs_rqs(struct rq
*rq
)
2218 struct cfs_rq
*cfs_rq
;
2220 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
2221 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2223 if (!cfs_rq
->runtime_enabled
)
2227 * clock_task is not advancing so we just need to make sure
2228 * there's some valid quota amount
2230 cfs_rq
->runtime_remaining
= cfs_b
->quota
;
2231 if (cfs_rq_throttled(cfs_rq
))
2232 unthrottle_cfs_rq(cfs_rq
);
2236 #else /* CONFIG_CFS_BANDWIDTH */
2237 static __always_inline
2238 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
) {}
2239 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2240 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
2241 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2243 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
2248 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
2253 static inline int throttled_lb_pair(struct task_group
*tg
,
2254 int src_cpu
, int dest_cpu
)
2259 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
2261 #ifdef CONFIG_FAIR_GROUP_SCHED
2262 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2265 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
2269 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
2270 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
2272 #endif /* CONFIG_CFS_BANDWIDTH */
2274 /**************************************************
2275 * CFS operations on tasks:
2278 #ifdef CONFIG_SCHED_HRTICK
2279 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
2281 struct sched_entity
*se
= &p
->se
;
2282 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2284 WARN_ON(task_rq(p
) != rq
);
2286 if (cfs_rq
->nr_running
> 1) {
2287 u64 slice
= sched_slice(cfs_rq
, se
);
2288 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
2289 s64 delta
= slice
- ran
;
2298 * Don't schedule slices shorter than 10000ns, that just
2299 * doesn't make sense. Rely on vruntime for fairness.
2302 delta
= max_t(s64
, 10000LL, delta
);
2304 hrtick_start(rq
, delta
);
2309 * called from enqueue/dequeue and updates the hrtick when the
2310 * current task is from our class and nr_running is low enough
2313 static void hrtick_update(struct rq
*rq
)
2315 struct task_struct
*curr
= rq
->curr
;
2317 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
2320 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
2321 hrtick_start_fair(rq
, curr
);
2323 #else /* !CONFIG_SCHED_HRTICK */
2325 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
2329 static inline void hrtick_update(struct rq
*rq
)
2335 * The enqueue_task method is called before nr_running is
2336 * increased. Here we update the fair scheduling stats and
2337 * then put the task into the rbtree:
2340 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
2342 struct cfs_rq
*cfs_rq
;
2343 struct sched_entity
*se
= &p
->se
;
2345 for_each_sched_entity(se
) {
2348 cfs_rq
= cfs_rq_of(se
);
2349 enqueue_entity(cfs_rq
, se
, flags
);
2352 * end evaluation on encountering a throttled cfs_rq
2354 * note: in the case of encountering a throttled cfs_rq we will
2355 * post the final h_nr_running increment below.
2357 if (cfs_rq_throttled(cfs_rq
))
2359 cfs_rq
->h_nr_running
++;
2361 flags
= ENQUEUE_WAKEUP
;
2364 for_each_sched_entity(se
) {
2365 cfs_rq
= cfs_rq_of(se
);
2366 cfs_rq
->h_nr_running
++;
2368 if (cfs_rq_throttled(cfs_rq
))
2371 update_cfs_load(cfs_rq
, 0);
2372 update_cfs_shares(cfs_rq
);
2380 static void set_next_buddy(struct sched_entity
*se
);
2383 * The dequeue_task method is called before nr_running is
2384 * decreased. We remove the task from the rbtree and
2385 * update the fair scheduling stats:
2387 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
2389 struct cfs_rq
*cfs_rq
;
2390 struct sched_entity
*se
= &p
->se
;
2391 int task_sleep
= flags
& DEQUEUE_SLEEP
;
2393 for_each_sched_entity(se
) {
2394 cfs_rq
= cfs_rq_of(se
);
2395 dequeue_entity(cfs_rq
, se
, flags
);
2398 * end evaluation on encountering a throttled cfs_rq
2400 * note: in the case of encountering a throttled cfs_rq we will
2401 * post the final h_nr_running decrement below.
2403 if (cfs_rq_throttled(cfs_rq
))
2405 cfs_rq
->h_nr_running
--;
2407 /* Don't dequeue parent if it has other entities besides us */
2408 if (cfs_rq
->load
.weight
) {
2410 * Bias pick_next to pick a task from this cfs_rq, as
2411 * p is sleeping when it is within its sched_slice.
2413 if (task_sleep
&& parent_entity(se
))
2414 set_next_buddy(parent_entity(se
));
2416 /* avoid re-evaluating load for this entity */
2417 se
= parent_entity(se
);
2420 flags
|= DEQUEUE_SLEEP
;
2423 for_each_sched_entity(se
) {
2424 cfs_rq
= cfs_rq_of(se
);
2425 cfs_rq
->h_nr_running
--;
2427 if (cfs_rq_throttled(cfs_rq
))
2430 update_cfs_load(cfs_rq
, 0);
2431 update_cfs_shares(cfs_rq
);
2440 /* Used instead of source_load when we know the type == 0 */
2441 static unsigned long weighted_cpuload(const int cpu
)
2443 return cpu_rq(cpu
)->load
.weight
;
2447 * Return a low guess at the load of a migration-source cpu weighted
2448 * according to the scheduling class and "nice" value.
2450 * We want to under-estimate the load of migration sources, to
2451 * balance conservatively.
2453 static unsigned long source_load(int cpu
, int type
)
2455 struct rq
*rq
= cpu_rq(cpu
);
2456 unsigned long total
= weighted_cpuload(cpu
);
2458 if (type
== 0 || !sched_feat(LB_BIAS
))
2461 return min(rq
->cpu_load
[type
-1], total
);
2465 * Return a high guess at the load of a migration-target cpu weighted
2466 * according to the scheduling class and "nice" value.
2468 static unsigned long target_load(int cpu
, int type
)
2470 struct rq
*rq
= cpu_rq(cpu
);
2471 unsigned long total
= weighted_cpuload(cpu
);
2473 if (type
== 0 || !sched_feat(LB_BIAS
))
2476 return max(rq
->cpu_load
[type
-1], total
);
2479 static unsigned long power_of(int cpu
)
2481 return cpu_rq(cpu
)->cpu_power
;
2484 static unsigned long cpu_avg_load_per_task(int cpu
)
2486 struct rq
*rq
= cpu_rq(cpu
);
2487 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
2490 return rq
->load
.weight
/ nr_running
;
2496 static void task_waking_fair(struct task_struct
*p
)
2498 struct sched_entity
*se
= &p
->se
;
2499 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2502 #ifndef CONFIG_64BIT
2503 u64 min_vruntime_copy
;
2506 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
2508 min_vruntime
= cfs_rq
->min_vruntime
;
2509 } while (min_vruntime
!= min_vruntime_copy
);
2511 min_vruntime
= cfs_rq
->min_vruntime
;
2514 se
->vruntime
-= min_vruntime
;
2517 #ifdef CONFIG_FAIR_GROUP_SCHED
2519 * effective_load() calculates the load change as seen from the root_task_group
2521 * Adding load to a group doesn't make a group heavier, but can cause movement
2522 * of group shares between cpus. Assuming the shares were perfectly aligned one
2523 * can calculate the shift in shares.
2525 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2526 * on this @cpu and results in a total addition (subtraction) of @wg to the
2527 * total group weight.
2529 * Given a runqueue weight distribution (rw_i) we can compute a shares
2530 * distribution (s_i) using:
2532 * s_i = rw_i / \Sum rw_j (1)
2534 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2535 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2536 * shares distribution (s_i):
2538 * rw_i = { 2, 4, 1, 0 }
2539 * s_i = { 2/7, 4/7, 1/7, 0 }
2541 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2542 * task used to run on and the CPU the waker is running on), we need to
2543 * compute the effect of waking a task on either CPU and, in case of a sync
2544 * wakeup, compute the effect of the current task going to sleep.
2546 * So for a change of @wl to the local @cpu with an overall group weight change
2547 * of @wl we can compute the new shares distribution (s'_i) using:
2549 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2551 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2552 * differences in waking a task to CPU 0. The additional task changes the
2553 * weight and shares distributions like:
2555 * rw'_i = { 3, 4, 1, 0 }
2556 * s'_i = { 3/8, 4/8, 1/8, 0 }
2558 * We can then compute the difference in effective weight by using:
2560 * dw_i = S * (s'_i - s_i) (3)
2562 * Where 'S' is the group weight as seen by its parent.
2564 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2565 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2566 * 4/7) times the weight of the group.
2568 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
2570 struct sched_entity
*se
= tg
->se
[cpu
];
2572 if (!tg
->parent
) /* the trivial, non-cgroup case */
2575 for_each_sched_entity(se
) {
2581 * W = @wg + \Sum rw_j
2583 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
2588 w
= se
->my_q
->load
.weight
+ wl
;
2591 * wl = S * s'_i; see (2)
2594 wl
= (w
* tg
->shares
) / W
;
2599 * Per the above, wl is the new se->load.weight value; since
2600 * those are clipped to [MIN_SHARES, ...) do so now. See
2601 * calc_cfs_shares().
2603 if (wl
< MIN_SHARES
)
2607 * wl = dw_i = S * (s'_i - s_i); see (3)
2609 wl
-= se
->load
.weight
;
2612 * Recursively apply this logic to all parent groups to compute
2613 * the final effective load change on the root group. Since
2614 * only the @tg group gets extra weight, all parent groups can
2615 * only redistribute existing shares. @wl is the shift in shares
2616 * resulting from this level per the above.
2625 static inline unsigned long effective_load(struct task_group
*tg
, int cpu
,
2626 unsigned long wl
, unsigned long wg
)
2633 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
2635 s64 this_load
, load
;
2636 int idx
, this_cpu
, prev_cpu
;
2637 unsigned long tl_per_task
;
2638 struct task_group
*tg
;
2639 unsigned long weight
;
2643 this_cpu
= smp_processor_id();
2644 prev_cpu
= task_cpu(p
);
2645 load
= source_load(prev_cpu
, idx
);
2646 this_load
= target_load(this_cpu
, idx
);
2649 * If sync wakeup then subtract the (maximum possible)
2650 * effect of the currently running task from the load
2651 * of the current CPU:
2654 tg
= task_group(current
);
2655 weight
= current
->se
.load
.weight
;
2657 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
2658 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
2662 weight
= p
->se
.load
.weight
;
2665 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2666 * due to the sync cause above having dropped this_load to 0, we'll
2667 * always have an imbalance, but there's really nothing you can do
2668 * about that, so that's good too.
2670 * Otherwise check if either cpus are near enough in load to allow this
2671 * task to be woken on this_cpu.
2673 if (this_load
> 0) {
2674 s64 this_eff_load
, prev_eff_load
;
2676 this_eff_load
= 100;
2677 this_eff_load
*= power_of(prev_cpu
);
2678 this_eff_load
*= this_load
+
2679 effective_load(tg
, this_cpu
, weight
, weight
);
2681 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
2682 prev_eff_load
*= power_of(this_cpu
);
2683 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
2685 balanced
= this_eff_load
<= prev_eff_load
;
2690 * If the currently running task will sleep within
2691 * a reasonable amount of time then attract this newly
2694 if (sync
&& balanced
)
2697 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
2698 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
2701 (this_load
<= load
&&
2702 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
2704 * This domain has SD_WAKE_AFFINE and
2705 * p is cache cold in this domain, and
2706 * there is no bad imbalance.
2708 schedstat_inc(sd
, ttwu_move_affine
);
2709 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
2717 * find_idlest_group finds and returns the least busy CPU group within the
2720 static struct sched_group
*
2721 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
2722 int this_cpu
, int load_idx
)
2724 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
2725 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2726 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2729 unsigned long load
, avg_load
;
2733 /* Skip over this group if it has no CPUs allowed */
2734 if (!cpumask_intersects(sched_group_cpus(group
),
2735 tsk_cpus_allowed(p
)))
2738 local_group
= cpumask_test_cpu(this_cpu
,
2739 sched_group_cpus(group
));
2741 /* Tally up the load of all CPUs in the group */
2744 for_each_cpu(i
, sched_group_cpus(group
)) {
2745 /* Bias balancing toward cpus of our domain */
2747 load
= source_load(i
, load_idx
);
2749 load
= target_load(i
, load_idx
);
2754 /* Adjust by relative CPU power of the group */
2755 avg_load
= (avg_load
* SCHED_POWER_SCALE
) / group
->sgp
->power
;
2758 this_load
= avg_load
;
2759 } else if (avg_load
< min_load
) {
2760 min_load
= avg_load
;
2763 } while (group
= group
->next
, group
!= sd
->groups
);
2765 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2771 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2774 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2776 unsigned long load
, min_load
= ULONG_MAX
;
2780 /* Traverse only the allowed CPUs */
2781 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
2782 load
= weighted_cpuload(i
);
2784 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2794 * Try and locate an idle CPU in the sched_domain.
2796 static int select_idle_sibling(struct task_struct
*p
, int target
)
2798 int cpu
= smp_processor_id();
2799 int prev_cpu
= task_cpu(p
);
2800 struct sched_domain
*sd
;
2801 struct sched_group
*sg
;
2805 * If the task is going to be woken-up on this cpu and if it is
2806 * already idle, then it is the right target.
2808 if (target
== cpu
&& idle_cpu(cpu
))
2812 * If the task is going to be woken-up on the cpu where it previously
2813 * ran and if it is currently idle, then it the right target.
2815 if (target
== prev_cpu
&& idle_cpu(prev_cpu
))
2819 * Otherwise, iterate the domains and find an elegible idle cpu.
2821 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
2822 for_each_lower_domain(sd
) {
2825 if (!cpumask_intersects(sched_group_cpus(sg
),
2826 tsk_cpus_allowed(p
)))
2829 for_each_cpu(i
, sched_group_cpus(sg
)) {
2834 target
= cpumask_first_and(sched_group_cpus(sg
),
2835 tsk_cpus_allowed(p
));
2839 } while (sg
!= sd
->groups
);
2846 * sched_balance_self: balance the current task (running on cpu) in domains
2847 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2850 * Balance, ie. select the least loaded group.
2852 * Returns the target CPU number, or the same CPU if no balancing is needed.
2854 * preempt must be disabled.
2857 select_task_rq_fair(struct task_struct
*p
, int sd_flag
, int wake_flags
)
2859 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
2860 int cpu
= smp_processor_id();
2861 int prev_cpu
= task_cpu(p
);
2863 int want_affine
= 0;
2864 int sync
= wake_flags
& WF_SYNC
;
2866 if (p
->nr_cpus_allowed
== 1)
2869 if (sd_flag
& SD_BALANCE_WAKE
) {
2870 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
2876 for_each_domain(cpu
, tmp
) {
2877 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
2881 * If both cpu and prev_cpu are part of this domain,
2882 * cpu is a valid SD_WAKE_AFFINE target.
2884 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
2885 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
2890 if (tmp
->flags
& sd_flag
)
2895 if (cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
2898 new_cpu
= select_idle_sibling(p
, prev_cpu
);
2903 int load_idx
= sd
->forkexec_idx
;
2904 struct sched_group
*group
;
2907 if (!(sd
->flags
& sd_flag
)) {
2912 if (sd_flag
& SD_BALANCE_WAKE
)
2913 load_idx
= sd
->wake_idx
;
2915 group
= find_idlest_group(sd
, p
, cpu
, load_idx
);
2921 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
2922 if (new_cpu
== -1 || new_cpu
== cpu
) {
2923 /* Now try balancing at a lower domain level of cpu */
2928 /* Now try balancing at a lower domain level of new_cpu */
2930 weight
= sd
->span_weight
;
2932 for_each_domain(cpu
, tmp
) {
2933 if (weight
<= tmp
->span_weight
)
2935 if (tmp
->flags
& sd_flag
)
2938 /* while loop will break here if sd == NULL */
2945 #endif /* CONFIG_SMP */
2947 static unsigned long
2948 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
2950 unsigned long gran
= sysctl_sched_wakeup_granularity
;
2953 * Since its curr running now, convert the gran from real-time
2954 * to virtual-time in his units.
2956 * By using 'se' instead of 'curr' we penalize light tasks, so
2957 * they get preempted easier. That is, if 'se' < 'curr' then
2958 * the resulting gran will be larger, therefore penalizing the
2959 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2960 * be smaller, again penalizing the lighter task.
2962 * This is especially important for buddies when the leftmost
2963 * task is higher priority than the buddy.
2965 return calc_delta_fair(gran
, se
);
2969 * Should 'se' preempt 'curr'.
2983 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
2985 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
2990 gran
= wakeup_gran(curr
, se
);
2997 static void set_last_buddy(struct sched_entity
*se
)
2999 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
3002 for_each_sched_entity(se
)
3003 cfs_rq_of(se
)->last
= se
;
3006 static void set_next_buddy(struct sched_entity
*se
)
3008 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
3011 for_each_sched_entity(se
)
3012 cfs_rq_of(se
)->next
= se
;
3015 static void set_skip_buddy(struct sched_entity
*se
)
3017 for_each_sched_entity(se
)
3018 cfs_rq_of(se
)->skip
= se
;
3022 * Preempt the current task with a newly woken task if needed:
3024 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
3026 struct task_struct
*curr
= rq
->curr
;
3027 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
3028 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
3029 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
3030 int next_buddy_marked
= 0;
3032 if (unlikely(se
== pse
))
3036 * This is possible from callers such as move_task(), in which we
3037 * unconditionally check_prempt_curr() after an enqueue (which may have
3038 * lead to a throttle). This both saves work and prevents false
3039 * next-buddy nomination below.
3041 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
3044 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
3045 set_next_buddy(pse
);
3046 next_buddy_marked
= 1;
3050 * We can come here with TIF_NEED_RESCHED already set from new task
3053 * Note: this also catches the edge-case of curr being in a throttled
3054 * group (e.g. via set_curr_task), since update_curr() (in the
3055 * enqueue of curr) will have resulted in resched being set. This
3056 * prevents us from potentially nominating it as a false LAST_BUDDY
3059 if (test_tsk_need_resched(curr
))
3062 /* Idle tasks are by definition preempted by non-idle tasks. */
3063 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
3064 likely(p
->policy
!= SCHED_IDLE
))
3068 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3069 * is driven by the tick):
3071 if (unlikely(p
->policy
!= SCHED_NORMAL
))
3074 find_matching_se(&se
, &pse
);
3075 update_curr(cfs_rq_of(se
));
3077 if (wakeup_preempt_entity(se
, pse
) == 1) {
3079 * Bias pick_next to pick the sched entity that is
3080 * triggering this preemption.
3082 if (!next_buddy_marked
)
3083 set_next_buddy(pse
);
3092 * Only set the backward buddy when the current task is still
3093 * on the rq. This can happen when a wakeup gets interleaved
3094 * with schedule on the ->pre_schedule() or idle_balance()
3095 * point, either of which can * drop the rq lock.
3097 * Also, during early boot the idle thread is in the fair class,
3098 * for obvious reasons its a bad idea to schedule back to it.
3100 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
3103 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
3107 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
3109 struct task_struct
*p
;
3110 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
3111 struct sched_entity
*se
;
3113 if (!cfs_rq
->nr_running
)
3117 se
= pick_next_entity(cfs_rq
);
3118 set_next_entity(cfs_rq
, se
);
3119 cfs_rq
= group_cfs_rq(se
);
3123 if (hrtick_enabled(rq
))
3124 hrtick_start_fair(rq
, p
);
3130 * Account for a descheduled task:
3132 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
3134 struct sched_entity
*se
= &prev
->se
;
3135 struct cfs_rq
*cfs_rq
;
3137 for_each_sched_entity(se
) {
3138 cfs_rq
= cfs_rq_of(se
);
3139 put_prev_entity(cfs_rq
, se
);
3144 * sched_yield() is very simple
3146 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3148 static void yield_task_fair(struct rq
*rq
)
3150 struct task_struct
*curr
= rq
->curr
;
3151 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
3152 struct sched_entity
*se
= &curr
->se
;
3155 * Are we the only task in the tree?
3157 if (unlikely(rq
->nr_running
== 1))
3160 clear_buddies(cfs_rq
, se
);
3162 if (curr
->policy
!= SCHED_BATCH
) {
3163 update_rq_clock(rq
);
3165 * Update run-time statistics of the 'current'.
3167 update_curr(cfs_rq
);
3169 * Tell update_rq_clock() that we've just updated,
3170 * so we don't do microscopic update in schedule()
3171 * and double the fastpath cost.
3173 rq
->skip_clock_update
= 1;
3179 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
3181 struct sched_entity
*se
= &p
->se
;
3183 /* throttled hierarchies are not runnable */
3184 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
3187 /* Tell the scheduler that we'd really like pse to run next. */
3190 yield_task_fair(rq
);
3196 /**************************************************
3197 * Fair scheduling class load-balancing methods:
3200 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
3202 #define LBF_ALL_PINNED 0x01
3203 #define LBF_NEED_BREAK 0x02
3204 #define LBF_SOME_PINNED 0x04
3207 struct sched_domain
*sd
;
3215 struct cpumask
*dst_grpmask
;
3217 enum cpu_idle_type idle
;
3219 /* The set of CPUs under consideration for load-balancing */
3220 struct cpumask
*cpus
;
3225 unsigned int loop_break
;
3226 unsigned int loop_max
;
3230 * move_task - move a task from one runqueue to another runqueue.
3231 * Both runqueues must be locked.
3233 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
3235 deactivate_task(env
->src_rq
, p
, 0);
3236 set_task_cpu(p
, env
->dst_cpu
);
3237 activate_task(env
->dst_rq
, p
, 0);
3238 check_preempt_curr(env
->dst_rq
, p
, 0);
3242 * Is this task likely cache-hot:
3245 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
3249 if (p
->sched_class
!= &fair_sched_class
)
3252 if (unlikely(p
->policy
== SCHED_IDLE
))
3256 * Buddy candidates are cache hot:
3258 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
3259 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
3260 &p
->se
== cfs_rq_of(&p
->se
)->last
))
3263 if (sysctl_sched_migration_cost
== -1)
3265 if (sysctl_sched_migration_cost
== 0)
3268 delta
= now
- p
->se
.exec_start
;
3270 return delta
< (s64
)sysctl_sched_migration_cost
;
3274 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3277 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
3279 int tsk_cache_hot
= 0;
3281 * We do not migrate tasks that are:
3282 * 1) running (obviously), or
3283 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3284 * 3) are cache-hot on their current CPU.
3286 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
3289 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
3292 * Remember if this task can be migrated to any other cpu in
3293 * our sched_group. We may want to revisit it if we couldn't
3294 * meet load balance goals by pulling other tasks on src_cpu.
3296 * Also avoid computing new_dst_cpu if we have already computed
3297 * one in current iteration.
3299 if (!env
->dst_grpmask
|| (env
->flags
& LBF_SOME_PINNED
))
3302 new_dst_cpu
= cpumask_first_and(env
->dst_grpmask
,
3303 tsk_cpus_allowed(p
));
3304 if (new_dst_cpu
< nr_cpu_ids
) {
3305 env
->flags
|= LBF_SOME_PINNED
;
3306 env
->new_dst_cpu
= new_dst_cpu
;
3311 /* Record that we found atleast one task that could run on dst_cpu */
3312 env
->flags
&= ~LBF_ALL_PINNED
;
3314 if (task_running(env
->src_rq
, p
)) {
3315 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
3320 * Aggressive migration if:
3321 * 1) task is cache cold, or
3322 * 2) too many balance attempts have failed.
3325 tsk_cache_hot
= task_hot(p
, env
->src_rq
->clock_task
, env
->sd
);
3326 if (!tsk_cache_hot
||
3327 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
3328 #ifdef CONFIG_SCHEDSTATS
3329 if (tsk_cache_hot
) {
3330 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
3331 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
3337 if (tsk_cache_hot
) {
3338 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
3345 * move_one_task tries to move exactly one task from busiest to this_rq, as
3346 * part of active balancing operations within "domain".
3347 * Returns 1 if successful and 0 otherwise.
3349 * Called with both runqueues locked.
3351 static int move_one_task(struct lb_env
*env
)
3353 struct task_struct
*p
, *n
;
3355 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
3356 if (throttled_lb_pair(task_group(p
), env
->src_rq
->cpu
, env
->dst_cpu
))
3359 if (!can_migrate_task(p
, env
))
3364 * Right now, this is only the second place move_task()
3365 * is called, so we can safely collect move_task()
3366 * stats here rather than inside move_task().
3368 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
3374 static unsigned long task_h_load(struct task_struct
*p
);
3376 static const unsigned int sched_nr_migrate_break
= 32;
3379 * move_tasks tries to move up to imbalance weighted load from busiest to
3380 * this_rq, as part of a balancing operation within domain "sd".
3381 * Returns 1 if successful and 0 otherwise.
3383 * Called with both runqueues locked.
3385 static int move_tasks(struct lb_env
*env
)
3387 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
3388 struct task_struct
*p
;
3392 if (env
->imbalance
<= 0)
3395 while (!list_empty(tasks
)) {
3396 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
3399 /* We've more or less seen every task there is, call it quits */
3400 if (env
->loop
> env
->loop_max
)
3403 /* take a breather every nr_migrate tasks */
3404 if (env
->loop
> env
->loop_break
) {
3405 env
->loop_break
+= sched_nr_migrate_break
;
3406 env
->flags
|= LBF_NEED_BREAK
;
3410 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
3413 load
= task_h_load(p
);
3415 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
3418 if ((load
/ 2) > env
->imbalance
)
3421 if (!can_migrate_task(p
, env
))
3426 env
->imbalance
-= load
;
3428 #ifdef CONFIG_PREEMPT
3430 * NEWIDLE balancing is a source of latency, so preemptible
3431 * kernels will stop after the first task is pulled to minimize
3432 * the critical section.
3434 if (env
->idle
== CPU_NEWLY_IDLE
)
3439 * We only want to steal up to the prescribed amount of
3442 if (env
->imbalance
<= 0)
3447 list_move_tail(&p
->se
.group_node
, tasks
);
3451 * Right now, this is one of only two places move_task() is called,
3452 * so we can safely collect move_task() stats here rather than
3453 * inside move_task().
3455 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
3460 #ifdef CONFIG_FAIR_GROUP_SCHED
3462 * update tg->load_weight by folding this cpu's load_avg
3464 static int update_shares_cpu(struct task_group
*tg
, int cpu
)
3466 struct cfs_rq
*cfs_rq
;
3467 unsigned long flags
;
3474 cfs_rq
= tg
->cfs_rq
[cpu
];
3476 raw_spin_lock_irqsave(&rq
->lock
, flags
);
3478 update_rq_clock(rq
);
3479 update_cfs_load(cfs_rq
, 1);
3482 * We need to update shares after updating tg->load_weight in
3483 * order to adjust the weight of groups with long running tasks.
3485 update_cfs_shares(cfs_rq
);
3487 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
3492 static void update_shares(int cpu
)
3494 struct cfs_rq
*cfs_rq
;
3495 struct rq
*rq
= cpu_rq(cpu
);
3499 * Iterates the task_group tree in a bottom up fashion, see
3500 * list_add_leaf_cfs_rq() for details.
3502 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3503 /* throttled entities do not contribute to load */
3504 if (throttled_hierarchy(cfs_rq
))
3507 update_shares_cpu(cfs_rq
->tg
, cpu
);
3513 * Compute the cpu's hierarchical load factor for each task group.
3514 * This needs to be done in a top-down fashion because the load of a child
3515 * group is a fraction of its parents load.
3517 static int tg_load_down(struct task_group
*tg
, void *data
)
3520 long cpu
= (long)data
;
3523 load
= cpu_rq(cpu
)->load
.weight
;
3525 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
3526 load
*= tg
->se
[cpu
]->load
.weight
;
3527 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
3530 tg
->cfs_rq
[cpu
]->h_load
= load
;
3535 static void update_h_load(long cpu
)
3537 struct rq
*rq
= cpu_rq(cpu
);
3538 unsigned long now
= jiffies
;
3540 if (rq
->h_load_throttle
== now
)
3543 rq
->h_load_throttle
= now
;
3546 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
3550 static unsigned long task_h_load(struct task_struct
*p
)
3552 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
3555 load
= p
->se
.load
.weight
;
3556 load
= div_u64(load
* cfs_rq
->h_load
, cfs_rq
->load
.weight
+ 1);
3561 static inline void update_shares(int cpu
)
3565 static inline void update_h_load(long cpu
)
3569 static unsigned long task_h_load(struct task_struct
*p
)
3571 return p
->se
.load
.weight
;
3575 /********** Helpers for find_busiest_group ************************/
3577 * sd_lb_stats - Structure to store the statistics of a sched_domain
3578 * during load balancing.
3580 struct sd_lb_stats
{
3581 struct sched_group
*busiest
; /* Busiest group in this sd */
3582 struct sched_group
*this; /* Local group in this sd */
3583 unsigned long total_load
; /* Total load of all groups in sd */
3584 unsigned long total_pwr
; /* Total power of all groups in sd */
3585 unsigned long avg_load
; /* Average load across all groups in sd */
3587 /** Statistics of this group */
3588 unsigned long this_load
;
3589 unsigned long this_load_per_task
;
3590 unsigned long this_nr_running
;
3591 unsigned long this_has_capacity
;
3592 unsigned int this_idle_cpus
;
3594 /* Statistics of the busiest group */
3595 unsigned int busiest_idle_cpus
;
3596 unsigned long max_load
;
3597 unsigned long busiest_load_per_task
;
3598 unsigned long busiest_nr_running
;
3599 unsigned long busiest_group_capacity
;
3600 unsigned long busiest_has_capacity
;
3601 unsigned int busiest_group_weight
;
3603 int group_imb
; /* Is there imbalance in this sd */
3607 * sg_lb_stats - stats of a sched_group required for load_balancing
3609 struct sg_lb_stats
{
3610 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3611 unsigned long group_load
; /* Total load over the CPUs of the group */
3612 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3613 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3614 unsigned long group_capacity
;
3615 unsigned long idle_cpus
;
3616 unsigned long group_weight
;
3617 int group_imb
; /* Is there an imbalance in the group ? */
3618 int group_has_capacity
; /* Is there extra capacity in the group? */
3622 * get_sd_load_idx - Obtain the load index for a given sched domain.
3623 * @sd: The sched_domain whose load_idx is to be obtained.
3624 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3626 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3627 enum cpu_idle_type idle
)
3633 load_idx
= sd
->busy_idx
;
3636 case CPU_NEWLY_IDLE
:
3637 load_idx
= sd
->newidle_idx
;
3640 load_idx
= sd
->idle_idx
;
3647 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3649 return SCHED_POWER_SCALE
;
3652 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3654 return default_scale_freq_power(sd
, cpu
);
3657 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3659 unsigned long weight
= sd
->span_weight
;
3660 unsigned long smt_gain
= sd
->smt_gain
;
3667 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3669 return default_scale_smt_power(sd
, cpu
);
3672 unsigned long scale_rt_power(int cpu
)
3674 struct rq
*rq
= cpu_rq(cpu
);
3675 u64 total
, available
, age_stamp
, avg
;
3678 * Since we're reading these variables without serialization make sure
3679 * we read them once before doing sanity checks on them.
3681 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
3682 avg
= ACCESS_ONCE(rq
->rt_avg
);
3684 total
= sched_avg_period() + (rq
->clock
- age_stamp
);
3686 if (unlikely(total
< avg
)) {
3687 /* Ensures that power won't end up being negative */
3690 available
= total
- avg
;
3693 if (unlikely((s64
)total
< SCHED_POWER_SCALE
))
3694 total
= SCHED_POWER_SCALE
;
3696 total
>>= SCHED_POWER_SHIFT
;
3698 return div_u64(available
, total
);
3701 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
3703 unsigned long weight
= sd
->span_weight
;
3704 unsigned long power
= SCHED_POWER_SCALE
;
3705 struct sched_group
*sdg
= sd
->groups
;
3707 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
3708 if (sched_feat(ARCH_POWER
))
3709 power
*= arch_scale_smt_power(sd
, cpu
);
3711 power
*= default_scale_smt_power(sd
, cpu
);
3713 power
>>= SCHED_POWER_SHIFT
;
3716 sdg
->sgp
->power_orig
= power
;
3718 if (sched_feat(ARCH_POWER
))
3719 power
*= arch_scale_freq_power(sd
, cpu
);
3721 power
*= default_scale_freq_power(sd
, cpu
);
3723 power
>>= SCHED_POWER_SHIFT
;
3725 power
*= scale_rt_power(cpu
);
3726 power
>>= SCHED_POWER_SHIFT
;
3731 cpu_rq(cpu
)->cpu_power
= power
;
3732 sdg
->sgp
->power
= power
;
3735 void update_group_power(struct sched_domain
*sd
, int cpu
)
3737 struct sched_domain
*child
= sd
->child
;
3738 struct sched_group
*group
, *sdg
= sd
->groups
;
3739 unsigned long power
;
3740 unsigned long interval
;
3742 interval
= msecs_to_jiffies(sd
->balance_interval
);
3743 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
3744 sdg
->sgp
->next_update
= jiffies
+ interval
;
3747 update_cpu_power(sd
, cpu
);
3753 if (child
->flags
& SD_OVERLAP
) {
3755 * SD_OVERLAP domains cannot assume that child groups
3756 * span the current group.
3759 for_each_cpu(cpu
, sched_group_cpus(sdg
))
3760 power
+= power_of(cpu
);
3763 * !SD_OVERLAP domains can assume that child groups
3764 * span the current group.
3767 group
= child
->groups
;
3769 power
+= group
->sgp
->power
;
3770 group
= group
->next
;
3771 } while (group
!= child
->groups
);
3774 sdg
->sgp
->power_orig
= sdg
->sgp
->power
= power
;
3778 * Try and fix up capacity for tiny siblings, this is needed when
3779 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3780 * which on its own isn't powerful enough.
3782 * See update_sd_pick_busiest() and check_asym_packing().
3785 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
3788 * Only siblings can have significantly less than SCHED_POWER_SCALE
3790 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
3794 * If ~90% of the cpu_power is still there, we're good.
3796 if (group
->sgp
->power
* 32 > group
->sgp
->power_orig
* 29)
3803 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3804 * @env: The load balancing environment.
3805 * @group: sched_group whose statistics are to be updated.
3806 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3807 * @local_group: Does group contain this_cpu.
3808 * @balance: Should we balance.
3809 * @sgs: variable to hold the statistics for this group.
3811 static inline void update_sg_lb_stats(struct lb_env
*env
,
3812 struct sched_group
*group
, int load_idx
,
3813 int local_group
, int *balance
, struct sg_lb_stats
*sgs
)
3815 unsigned long nr_running
, max_nr_running
, min_nr_running
;
3816 unsigned long load
, max_cpu_load
, min_cpu_load
;
3817 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3818 unsigned long avg_load_per_task
= 0;
3822 balance_cpu
= group_balance_cpu(group
);
3824 /* Tally up the load of all CPUs in the group */
3826 min_cpu_load
= ~0UL;
3828 min_nr_running
= ~0UL;
3830 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
3831 struct rq
*rq
= cpu_rq(i
);
3833 nr_running
= rq
->nr_running
;
3835 /* Bias balancing toward cpus of our domain */
3837 if (idle_cpu(i
) && !first_idle_cpu
&&
3838 cpumask_test_cpu(i
, sched_group_mask(group
))) {
3843 load
= target_load(i
, load_idx
);
3845 load
= source_load(i
, load_idx
);
3846 if (load
> max_cpu_load
)
3847 max_cpu_load
= load
;
3848 if (min_cpu_load
> load
)
3849 min_cpu_load
= load
;
3851 if (nr_running
> max_nr_running
)
3852 max_nr_running
= nr_running
;
3853 if (min_nr_running
> nr_running
)
3854 min_nr_running
= nr_running
;
3857 sgs
->group_load
+= load
;
3858 sgs
->sum_nr_running
+= nr_running
;
3859 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3865 * First idle cpu or the first cpu(busiest) in this sched group
3866 * is eligible for doing load balancing at this and above
3867 * domains. In the newly idle case, we will allow all the cpu's
3868 * to do the newly idle load balance.
3871 if (env
->idle
!= CPU_NEWLY_IDLE
) {
3872 if (balance_cpu
!= env
->dst_cpu
) {
3876 update_group_power(env
->sd
, env
->dst_cpu
);
3877 } else if (time_after_eq(jiffies
, group
->sgp
->next_update
))
3878 update_group_power(env
->sd
, env
->dst_cpu
);
3881 /* Adjust by relative CPU power of the group */
3882 sgs
->avg_load
= (sgs
->group_load
*SCHED_POWER_SCALE
) / group
->sgp
->power
;
3885 * Consider the group unbalanced when the imbalance is larger
3886 * than the average weight of a task.
3888 * APZ: with cgroup the avg task weight can vary wildly and
3889 * might not be a suitable number - should we keep a
3890 * normalized nr_running number somewhere that negates
3893 if (sgs
->sum_nr_running
)
3894 avg_load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
3896 if ((max_cpu_load
- min_cpu_load
) >= avg_load_per_task
&&
3897 (max_nr_running
- min_nr_running
) > 1)
3900 sgs
->group_capacity
= DIV_ROUND_CLOSEST(group
->sgp
->power
,
3902 if (!sgs
->group_capacity
)
3903 sgs
->group_capacity
= fix_small_capacity(env
->sd
, group
);
3904 sgs
->group_weight
= group
->group_weight
;
3906 if (sgs
->group_capacity
> sgs
->sum_nr_running
)
3907 sgs
->group_has_capacity
= 1;
3911 * update_sd_pick_busiest - return 1 on busiest group
3912 * @env: The load balancing environment.
3913 * @sds: sched_domain statistics
3914 * @sg: sched_group candidate to be checked for being the busiest
3915 * @sgs: sched_group statistics
3917 * Determine if @sg is a busier group than the previously selected
3920 static bool update_sd_pick_busiest(struct lb_env
*env
,
3921 struct sd_lb_stats
*sds
,
3922 struct sched_group
*sg
,
3923 struct sg_lb_stats
*sgs
)
3925 if (sgs
->avg_load
<= sds
->max_load
)
3928 if (sgs
->sum_nr_running
> sgs
->group_capacity
)
3935 * ASYM_PACKING needs to move all the work to the lowest
3936 * numbered CPUs in the group, therefore mark all groups
3937 * higher than ourself as busy.
3939 if ((env
->sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
3940 env
->dst_cpu
< group_first_cpu(sg
)) {
3944 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
3952 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
3953 * @env: The load balancing environment.
3954 * @balance: Should we balance.
3955 * @sds: variable to hold the statistics for this sched_domain.
3957 static inline void update_sd_lb_stats(struct lb_env
*env
,
3958 int *balance
, struct sd_lb_stats
*sds
)
3960 struct sched_domain
*child
= env
->sd
->child
;
3961 struct sched_group
*sg
= env
->sd
->groups
;
3962 struct sg_lb_stats sgs
;
3963 int load_idx
, prefer_sibling
= 0;
3965 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
3968 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
3973 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
3974 memset(&sgs
, 0, sizeof(sgs
));
3975 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, balance
, &sgs
);
3977 if (local_group
&& !(*balance
))
3980 sds
->total_load
+= sgs
.group_load
;
3981 sds
->total_pwr
+= sg
->sgp
->power
;
3984 * In case the child domain prefers tasks go to siblings
3985 * first, lower the sg capacity to one so that we'll try
3986 * and move all the excess tasks away. We lower the capacity
3987 * of a group only if the local group has the capacity to fit
3988 * these excess tasks, i.e. nr_running < group_capacity. The
3989 * extra check prevents the case where you always pull from the
3990 * heaviest group when it is already under-utilized (possible
3991 * with a large weight task outweighs the tasks on the system).
3993 if (prefer_sibling
&& !local_group
&& sds
->this_has_capacity
)
3994 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
3997 sds
->this_load
= sgs
.avg_load
;
3999 sds
->this_nr_running
= sgs
.sum_nr_running
;
4000 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
4001 sds
->this_has_capacity
= sgs
.group_has_capacity
;
4002 sds
->this_idle_cpus
= sgs
.idle_cpus
;
4003 } else if (update_sd_pick_busiest(env
, sds
, sg
, &sgs
)) {
4004 sds
->max_load
= sgs
.avg_load
;
4006 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
4007 sds
->busiest_idle_cpus
= sgs
.idle_cpus
;
4008 sds
->busiest_group_capacity
= sgs
.group_capacity
;
4009 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
4010 sds
->busiest_has_capacity
= sgs
.group_has_capacity
;
4011 sds
->busiest_group_weight
= sgs
.group_weight
;
4012 sds
->group_imb
= sgs
.group_imb
;
4016 } while (sg
!= env
->sd
->groups
);
4020 * check_asym_packing - Check to see if the group is packed into the
4023 * This is primarily intended to used at the sibling level. Some
4024 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4025 * case of POWER7, it can move to lower SMT modes only when higher
4026 * threads are idle. When in lower SMT modes, the threads will
4027 * perform better since they share less core resources. Hence when we
4028 * have idle threads, we want them to be the higher ones.
4030 * This packing function is run on idle threads. It checks to see if
4031 * the busiest CPU in this domain (core in the P7 case) has a higher
4032 * CPU number than the packing function is being run on. Here we are
4033 * assuming lower CPU number will be equivalent to lower a SMT thread
4036 * Returns 1 when packing is required and a task should be moved to
4037 * this CPU. The amount of the imbalance is returned in *imbalance.
4039 * @env: The load balancing environment.
4040 * @sds: Statistics of the sched_domain which is to be packed
4042 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4046 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
4052 busiest_cpu
= group_first_cpu(sds
->busiest
);
4053 if (env
->dst_cpu
> busiest_cpu
)
4056 env
->imbalance
= DIV_ROUND_CLOSEST(
4057 sds
->max_load
* sds
->busiest
->sgp
->power
, SCHED_POWER_SCALE
);
4063 * fix_small_imbalance - Calculate the minor imbalance that exists
4064 * amongst the groups of a sched_domain, during
4066 * @env: The load balancing environment.
4067 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4070 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4072 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
4073 unsigned int imbn
= 2;
4074 unsigned long scaled_busy_load_per_task
;
4076 if (sds
->this_nr_running
) {
4077 sds
->this_load_per_task
/= sds
->this_nr_running
;
4078 if (sds
->busiest_load_per_task
>
4079 sds
->this_load_per_task
)
4082 sds
->this_load_per_task
=
4083 cpu_avg_load_per_task(env
->dst_cpu
);
4086 scaled_busy_load_per_task
= sds
->busiest_load_per_task
4087 * SCHED_POWER_SCALE
;
4088 scaled_busy_load_per_task
/= sds
->busiest
->sgp
->power
;
4090 if (sds
->max_load
- sds
->this_load
+ scaled_busy_load_per_task
>=
4091 (scaled_busy_load_per_task
* imbn
)) {
4092 env
->imbalance
= sds
->busiest_load_per_task
;
4097 * OK, we don't have enough imbalance to justify moving tasks,
4098 * however we may be able to increase total CPU power used by
4102 pwr_now
+= sds
->busiest
->sgp
->power
*
4103 min(sds
->busiest_load_per_task
, sds
->max_load
);
4104 pwr_now
+= sds
->this->sgp
->power
*
4105 min(sds
->this_load_per_task
, sds
->this_load
);
4106 pwr_now
/= SCHED_POWER_SCALE
;
4108 /* Amount of load we'd subtract */
4109 tmp
= (sds
->busiest_load_per_task
* SCHED_POWER_SCALE
) /
4110 sds
->busiest
->sgp
->power
;
4111 if (sds
->max_load
> tmp
)
4112 pwr_move
+= sds
->busiest
->sgp
->power
*
4113 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
4115 /* Amount of load we'd add */
4116 if (sds
->max_load
* sds
->busiest
->sgp
->power
<
4117 sds
->busiest_load_per_task
* SCHED_POWER_SCALE
)
4118 tmp
= (sds
->max_load
* sds
->busiest
->sgp
->power
) /
4119 sds
->this->sgp
->power
;
4121 tmp
= (sds
->busiest_load_per_task
* SCHED_POWER_SCALE
) /
4122 sds
->this->sgp
->power
;
4123 pwr_move
+= sds
->this->sgp
->power
*
4124 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
4125 pwr_move
/= SCHED_POWER_SCALE
;
4127 /* Move if we gain throughput */
4128 if (pwr_move
> pwr_now
)
4129 env
->imbalance
= sds
->busiest_load_per_task
;
4133 * calculate_imbalance - Calculate the amount of imbalance present within the
4134 * groups of a given sched_domain during load balance.
4135 * @env: load balance environment
4136 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4138 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4140 unsigned long max_pull
, load_above_capacity
= ~0UL;
4142 sds
->busiest_load_per_task
/= sds
->busiest_nr_running
;
4143 if (sds
->group_imb
) {
4144 sds
->busiest_load_per_task
=
4145 min(sds
->busiest_load_per_task
, sds
->avg_load
);
4149 * In the presence of smp nice balancing, certain scenarios can have
4150 * max load less than avg load(as we skip the groups at or below
4151 * its cpu_power, while calculating max_load..)
4153 if (sds
->max_load
< sds
->avg_load
) {
4155 return fix_small_imbalance(env
, sds
);
4158 if (!sds
->group_imb
) {
4160 * Don't want to pull so many tasks that a group would go idle.
4162 load_above_capacity
= (sds
->busiest_nr_running
-
4163 sds
->busiest_group_capacity
);
4165 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_POWER_SCALE
);
4167 load_above_capacity
/= sds
->busiest
->sgp
->power
;
4171 * We're trying to get all the cpus to the average_load, so we don't
4172 * want to push ourselves above the average load, nor do we wish to
4173 * reduce the max loaded cpu below the average load. At the same time,
4174 * we also don't want to reduce the group load below the group capacity
4175 * (so that we can implement power-savings policies etc). Thus we look
4176 * for the minimum possible imbalance.
4177 * Be careful of negative numbers as they'll appear as very large values
4178 * with unsigned longs.
4180 max_pull
= min(sds
->max_load
- sds
->avg_load
, load_above_capacity
);
4182 /* How much load to actually move to equalise the imbalance */
4183 env
->imbalance
= min(max_pull
* sds
->busiest
->sgp
->power
,
4184 (sds
->avg_load
- sds
->this_load
) * sds
->this->sgp
->power
)
4185 / SCHED_POWER_SCALE
;
4188 * if *imbalance is less than the average load per runnable task
4189 * there is no guarantee that any tasks will be moved so we'll have
4190 * a think about bumping its value to force at least one task to be
4193 if (env
->imbalance
< sds
->busiest_load_per_task
)
4194 return fix_small_imbalance(env
, sds
);
4198 /******* find_busiest_group() helpers end here *********************/
4201 * find_busiest_group - Returns the busiest group within the sched_domain
4202 * if there is an imbalance. If there isn't an imbalance, and
4203 * the user has opted for power-savings, it returns a group whose
4204 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4205 * such a group exists.
4207 * Also calculates the amount of weighted load which should be moved
4208 * to restore balance.
4210 * @env: The load balancing environment.
4211 * @balance: Pointer to a variable indicating if this_cpu
4212 * is the appropriate cpu to perform load balancing at this_level.
4214 * Returns: - the busiest group if imbalance exists.
4215 * - If no imbalance and user has opted for power-savings balance,
4216 * return the least loaded group whose CPUs can be
4217 * put to idle by rebalancing its tasks onto our group.
4219 static struct sched_group
*
4220 find_busiest_group(struct lb_env
*env
, int *balance
)
4222 struct sd_lb_stats sds
;
4224 memset(&sds
, 0, sizeof(sds
));
4227 * Compute the various statistics relavent for load balancing at
4230 update_sd_lb_stats(env
, balance
, &sds
);
4233 * this_cpu is not the appropriate cpu to perform load balancing at
4239 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
4240 check_asym_packing(env
, &sds
))
4243 /* There is no busy sibling group to pull tasks from */
4244 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
4247 sds
.avg_load
= (SCHED_POWER_SCALE
* sds
.total_load
) / sds
.total_pwr
;
4250 * If the busiest group is imbalanced the below checks don't
4251 * work because they assumes all things are equal, which typically
4252 * isn't true due to cpus_allowed constraints and the like.
4257 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4258 if (env
->idle
== CPU_NEWLY_IDLE
&& sds
.this_has_capacity
&&
4259 !sds
.busiest_has_capacity
)
4263 * If the local group is more busy than the selected busiest group
4264 * don't try and pull any tasks.
4266 if (sds
.this_load
>= sds
.max_load
)
4270 * Don't pull any tasks if this group is already above the domain
4273 if (sds
.this_load
>= sds
.avg_load
)
4276 if (env
->idle
== CPU_IDLE
) {
4278 * This cpu is idle. If the busiest group load doesn't
4279 * have more tasks than the number of available cpu's and
4280 * there is no imbalance between this and busiest group
4281 * wrt to idle cpu's, it is balanced.
4283 if ((sds
.this_idle_cpus
<= sds
.busiest_idle_cpus
+ 1) &&
4284 sds
.busiest_nr_running
<= sds
.busiest_group_weight
)
4288 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4289 * imbalance_pct to be conservative.
4291 if (100 * sds
.max_load
<= env
->sd
->imbalance_pct
* sds
.this_load
)
4296 /* Looks like there is an imbalance. Compute it */
4297 calculate_imbalance(env
, &sds
);
4307 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4309 static struct rq
*find_busiest_queue(struct lb_env
*env
,
4310 struct sched_group
*group
)
4312 struct rq
*busiest
= NULL
, *rq
;
4313 unsigned long max_load
= 0;
4316 for_each_cpu(i
, sched_group_cpus(group
)) {
4317 unsigned long power
= power_of(i
);
4318 unsigned long capacity
= DIV_ROUND_CLOSEST(power
,
4323 capacity
= fix_small_capacity(env
->sd
, group
);
4325 if (!cpumask_test_cpu(i
, env
->cpus
))
4329 wl
= weighted_cpuload(i
);
4332 * When comparing with imbalance, use weighted_cpuload()
4333 * which is not scaled with the cpu power.
4335 if (capacity
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
4339 * For the load comparisons with the other cpu's, consider
4340 * the weighted_cpuload() scaled with the cpu power, so that
4341 * the load can be moved away from the cpu that is potentially
4342 * running at a lower capacity.
4344 wl
= (wl
* SCHED_POWER_SCALE
) / power
;
4346 if (wl
> max_load
) {
4356 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4357 * so long as it is large enough.
4359 #define MAX_PINNED_INTERVAL 512
4361 /* Working cpumask for load_balance and load_balance_newidle. */
4362 DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4364 static int need_active_balance(struct lb_env
*env
)
4366 struct sched_domain
*sd
= env
->sd
;
4368 if (env
->idle
== CPU_NEWLY_IDLE
) {
4371 * ASYM_PACKING needs to force migrate tasks from busy but
4372 * higher numbered CPUs in order to pack all tasks in the
4373 * lowest numbered CPUs.
4375 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
4379 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
4382 static int active_load_balance_cpu_stop(void *data
);
4385 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4386 * tasks if there is an imbalance.
4388 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4389 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4392 int ld_moved
, cur_ld_moved
, active_balance
= 0;
4393 int lb_iterations
, max_lb_iterations
;
4394 struct sched_group
*group
;
4396 unsigned long flags
;
4397 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4399 struct lb_env env
= {
4401 .dst_cpu
= this_cpu
,
4403 .dst_grpmask
= sched_group_cpus(sd
->groups
),
4405 .loop_break
= sched_nr_migrate_break
,
4409 cpumask_copy(cpus
, cpu_active_mask
);
4410 max_lb_iterations
= cpumask_weight(env
.dst_grpmask
);
4412 schedstat_inc(sd
, lb_count
[idle
]);
4415 group
= find_busiest_group(&env
, balance
);
4421 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4425 busiest
= find_busiest_queue(&env
, group
);
4427 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4431 BUG_ON(busiest
== env
.dst_rq
);
4433 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
4437 if (busiest
->nr_running
> 1) {
4439 * Attempt to move tasks. If find_busiest_group has found
4440 * an imbalance but busiest->nr_running <= 1, the group is
4441 * still unbalanced. ld_moved simply stays zero, so it is
4442 * correctly treated as an imbalance.
4444 env
.flags
|= LBF_ALL_PINNED
;
4445 env
.src_cpu
= busiest
->cpu
;
4446 env
.src_rq
= busiest
;
4447 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
4449 update_h_load(env
.src_cpu
);
4451 local_irq_save(flags
);
4452 double_rq_lock(env
.dst_rq
, busiest
);
4455 * cur_ld_moved - load moved in current iteration
4456 * ld_moved - cumulative load moved across iterations
4458 cur_ld_moved
= move_tasks(&env
);
4459 ld_moved
+= cur_ld_moved
;
4460 double_rq_unlock(env
.dst_rq
, busiest
);
4461 local_irq_restore(flags
);
4463 if (env
.flags
& LBF_NEED_BREAK
) {
4464 env
.flags
&= ~LBF_NEED_BREAK
;
4469 * some other cpu did the load balance for us.
4471 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
4472 resched_cpu(env
.dst_cpu
);
4475 * Revisit (affine) tasks on src_cpu that couldn't be moved to
4476 * us and move them to an alternate dst_cpu in our sched_group
4477 * where they can run. The upper limit on how many times we
4478 * iterate on same src_cpu is dependent on number of cpus in our
4481 * This changes load balance semantics a bit on who can move
4482 * load to a given_cpu. In addition to the given_cpu itself
4483 * (or a ilb_cpu acting on its behalf where given_cpu is
4484 * nohz-idle), we now have balance_cpu in a position to move
4485 * load to given_cpu. In rare situations, this may cause
4486 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
4487 * _independently_ and at _same_ time to move some load to
4488 * given_cpu) causing exceess load to be moved to given_cpu.
4489 * This however should not happen so much in practice and
4490 * moreover subsequent load balance cycles should correct the
4491 * excess load moved.
4493 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0 &&
4494 lb_iterations
++ < max_lb_iterations
) {
4496 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
4497 env
.dst_cpu
= env
.new_dst_cpu
;
4498 env
.flags
&= ~LBF_SOME_PINNED
;
4500 env
.loop_break
= sched_nr_migrate_break
;
4502 * Go back to "more_balance" rather than "redo" since we
4503 * need to continue with same src_cpu.
4508 /* All tasks on this runqueue were pinned by CPU affinity */
4509 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
4510 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4511 if (!cpumask_empty(cpus
)) {
4513 env
.loop_break
= sched_nr_migrate_break
;
4521 schedstat_inc(sd
, lb_failed
[idle
]);
4523 * Increment the failure counter only on periodic balance.
4524 * We do not want newidle balance, which can be very
4525 * frequent, pollute the failure counter causing
4526 * excessive cache_hot migrations and active balances.
4528 if (idle
!= CPU_NEWLY_IDLE
)
4529 sd
->nr_balance_failed
++;
4531 if (need_active_balance(&env
)) {
4532 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
4534 /* don't kick the active_load_balance_cpu_stop,
4535 * if the curr task on busiest cpu can't be
4538 if (!cpumask_test_cpu(this_cpu
,
4539 tsk_cpus_allowed(busiest
->curr
))) {
4540 raw_spin_unlock_irqrestore(&busiest
->lock
,
4542 env
.flags
|= LBF_ALL_PINNED
;
4543 goto out_one_pinned
;
4547 * ->active_balance synchronizes accesses to
4548 * ->active_balance_work. Once set, it's cleared
4549 * only after active load balance is finished.
4551 if (!busiest
->active_balance
) {
4552 busiest
->active_balance
= 1;
4553 busiest
->push_cpu
= this_cpu
;
4556 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
4558 if (active_balance
) {
4559 stop_one_cpu_nowait(cpu_of(busiest
),
4560 active_load_balance_cpu_stop
, busiest
,
4561 &busiest
->active_balance_work
);
4565 * We've kicked active balancing, reset the failure
4568 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4571 sd
->nr_balance_failed
= 0;
4573 if (likely(!active_balance
)) {
4574 /* We were unbalanced, so reset the balancing interval */
4575 sd
->balance_interval
= sd
->min_interval
;
4578 * If we've begun active balancing, start to back off. This
4579 * case may not be covered by the all_pinned logic if there
4580 * is only 1 task on the busy runqueue (because we don't call
4583 if (sd
->balance_interval
< sd
->max_interval
)
4584 sd
->balance_interval
*= 2;
4590 schedstat_inc(sd
, lb_balanced
[idle
]);
4592 sd
->nr_balance_failed
= 0;
4595 /* tune up the balancing interval */
4596 if (((env
.flags
& LBF_ALL_PINNED
) &&
4597 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4598 (sd
->balance_interval
< sd
->max_interval
))
4599 sd
->balance_interval
*= 2;
4607 * idle_balance is called by schedule() if this_cpu is about to become
4608 * idle. Attempts to pull tasks from other CPUs.
4610 void idle_balance(int this_cpu
, struct rq
*this_rq
)
4612 struct sched_domain
*sd
;
4613 int pulled_task
= 0;
4614 unsigned long next_balance
= jiffies
+ HZ
;
4616 this_rq
->idle_stamp
= this_rq
->clock
;
4618 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
4622 * Drop the rq->lock, but keep IRQ/preempt disabled.
4624 raw_spin_unlock(&this_rq
->lock
);
4626 update_shares(this_cpu
);
4628 for_each_domain(this_cpu
, sd
) {
4629 unsigned long interval
;
4632 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4635 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
4636 /* If we've pulled tasks over stop searching: */
4637 pulled_task
= load_balance(this_cpu
, this_rq
,
4638 sd
, CPU_NEWLY_IDLE
, &balance
);
4641 interval
= msecs_to_jiffies(sd
->balance_interval
);
4642 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4643 next_balance
= sd
->last_balance
+ interval
;
4645 this_rq
->idle_stamp
= 0;
4651 raw_spin_lock(&this_rq
->lock
);
4653 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4655 * We are going idle. next_balance may be set based on
4656 * a busy processor. So reset next_balance.
4658 this_rq
->next_balance
= next_balance
;
4663 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4664 * running tasks off the busiest CPU onto idle CPUs. It requires at
4665 * least 1 task to be running on each physical CPU where possible, and
4666 * avoids physical / logical imbalances.
4668 static int active_load_balance_cpu_stop(void *data
)
4670 struct rq
*busiest_rq
= data
;
4671 int busiest_cpu
= cpu_of(busiest_rq
);
4672 int target_cpu
= busiest_rq
->push_cpu
;
4673 struct rq
*target_rq
= cpu_rq(target_cpu
);
4674 struct sched_domain
*sd
;
4676 raw_spin_lock_irq(&busiest_rq
->lock
);
4678 /* make sure the requested cpu hasn't gone down in the meantime */
4679 if (unlikely(busiest_cpu
!= smp_processor_id() ||
4680 !busiest_rq
->active_balance
))
4683 /* Is there any task to move? */
4684 if (busiest_rq
->nr_running
<= 1)
4688 * This condition is "impossible", if it occurs
4689 * we need to fix it. Originally reported by
4690 * Bjorn Helgaas on a 128-cpu setup.
4692 BUG_ON(busiest_rq
== target_rq
);
4694 /* move a task from busiest_rq to target_rq */
4695 double_lock_balance(busiest_rq
, target_rq
);
4697 /* Search for an sd spanning us and the target CPU. */
4699 for_each_domain(target_cpu
, sd
) {
4700 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4701 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4706 struct lb_env env
= {
4708 .dst_cpu
= target_cpu
,
4709 .dst_rq
= target_rq
,
4710 .src_cpu
= busiest_rq
->cpu
,
4711 .src_rq
= busiest_rq
,
4715 schedstat_inc(sd
, alb_count
);
4717 if (move_one_task(&env
))
4718 schedstat_inc(sd
, alb_pushed
);
4720 schedstat_inc(sd
, alb_failed
);
4723 double_unlock_balance(busiest_rq
, target_rq
);
4725 busiest_rq
->active_balance
= 0;
4726 raw_spin_unlock_irq(&busiest_rq
->lock
);
4732 * idle load balancing details
4733 * - When one of the busy CPUs notice that there may be an idle rebalancing
4734 * needed, they will kick the idle load balancer, which then does idle
4735 * load balancing for all the idle CPUs.
4738 cpumask_var_t idle_cpus_mask
;
4740 unsigned long next_balance
; /* in jiffy units */
4741 } nohz ____cacheline_aligned
;
4743 static inline int find_new_ilb(int call_cpu
)
4745 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
4747 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
4754 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4755 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4756 * CPU (if there is one).
4758 static void nohz_balancer_kick(int cpu
)
4762 nohz
.next_balance
++;
4764 ilb_cpu
= find_new_ilb(cpu
);
4766 if (ilb_cpu
>= nr_cpu_ids
)
4769 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
4772 * Use smp_send_reschedule() instead of resched_cpu().
4773 * This way we generate a sched IPI on the target cpu which
4774 * is idle. And the softirq performing nohz idle load balance
4775 * will be run before returning from the IPI.
4777 smp_send_reschedule(ilb_cpu
);
4781 static inline void nohz_balance_exit_idle(int cpu
)
4783 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
4784 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
4785 atomic_dec(&nohz
.nr_cpus
);
4786 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
4790 static inline void set_cpu_sd_state_busy(void)
4792 struct sched_domain
*sd
;
4793 int cpu
= smp_processor_id();
4795 if (!test_bit(NOHZ_IDLE
, nohz_flags(cpu
)))
4797 clear_bit(NOHZ_IDLE
, nohz_flags(cpu
));
4800 for_each_domain(cpu
, sd
)
4801 atomic_inc(&sd
->groups
->sgp
->nr_busy_cpus
);
4805 void set_cpu_sd_state_idle(void)
4807 struct sched_domain
*sd
;
4808 int cpu
= smp_processor_id();
4810 if (test_bit(NOHZ_IDLE
, nohz_flags(cpu
)))
4812 set_bit(NOHZ_IDLE
, nohz_flags(cpu
));
4815 for_each_domain(cpu
, sd
)
4816 atomic_dec(&sd
->groups
->sgp
->nr_busy_cpus
);
4821 * This routine will record that the cpu is going idle with tick stopped.
4822 * This info will be used in performing idle load balancing in the future.
4824 void nohz_balance_enter_idle(int cpu
)
4827 * If this cpu is going down, then nothing needs to be done.
4829 if (!cpu_active(cpu
))
4832 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
4835 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
4836 atomic_inc(&nohz
.nr_cpus
);
4837 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
4840 static int __cpuinit
sched_ilb_notifier(struct notifier_block
*nfb
,
4841 unsigned long action
, void *hcpu
)
4843 switch (action
& ~CPU_TASKS_FROZEN
) {
4845 nohz_balance_exit_idle(smp_processor_id());
4853 static DEFINE_SPINLOCK(balancing
);
4856 * Scale the max load_balance interval with the number of CPUs in the system.
4857 * This trades load-balance latency on larger machines for less cross talk.
4859 void update_max_interval(void)
4861 max_load_balance_interval
= HZ
*num_online_cpus()/10;
4865 * It checks each scheduling domain to see if it is due to be balanced,
4866 * and initiates a balancing operation if so.
4868 * Balancing parameters are set up in arch_init_sched_domains.
4870 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4873 struct rq
*rq
= cpu_rq(cpu
);
4874 unsigned long interval
;
4875 struct sched_domain
*sd
;
4876 /* Earliest time when we have to do rebalance again */
4877 unsigned long next_balance
= jiffies
+ 60*HZ
;
4878 int update_next_balance
= 0;
4884 for_each_domain(cpu
, sd
) {
4885 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4888 interval
= sd
->balance_interval
;
4889 if (idle
!= CPU_IDLE
)
4890 interval
*= sd
->busy_factor
;
4892 /* scale ms to jiffies */
4893 interval
= msecs_to_jiffies(interval
);
4894 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
4896 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4898 if (need_serialize
) {
4899 if (!spin_trylock(&balancing
))
4903 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4904 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4906 * We've pulled tasks over so either we're no
4909 idle
= CPU_NOT_IDLE
;
4911 sd
->last_balance
= jiffies
;
4914 spin_unlock(&balancing
);
4916 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4917 next_balance
= sd
->last_balance
+ interval
;
4918 update_next_balance
= 1;
4922 * Stop the load balance at this level. There is another
4923 * CPU in our sched group which is doing load balancing more
4932 * next_balance will be updated only when there is a need.
4933 * When the cpu is attached to null domain for ex, it will not be
4936 if (likely(update_next_balance
))
4937 rq
->next_balance
= next_balance
;
4942 * In CONFIG_NO_HZ case, the idle balance kickee will do the
4943 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4945 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
)
4947 struct rq
*this_rq
= cpu_rq(this_cpu
);
4951 if (idle
!= CPU_IDLE
||
4952 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
4955 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
4956 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
4960 * If this cpu gets work to do, stop the load balancing
4961 * work being done for other cpus. Next load
4962 * balancing owner will pick it up.
4967 rq
= cpu_rq(balance_cpu
);
4969 raw_spin_lock_irq(&rq
->lock
);
4970 update_rq_clock(rq
);
4971 update_idle_cpu_load(rq
);
4972 raw_spin_unlock_irq(&rq
->lock
);
4974 rebalance_domains(balance_cpu
, CPU_IDLE
);
4976 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4977 this_rq
->next_balance
= rq
->next_balance
;
4979 nohz
.next_balance
= this_rq
->next_balance
;
4981 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
4985 * Current heuristic for kicking the idle load balancer in the presence
4986 * of an idle cpu is the system.
4987 * - This rq has more than one task.
4988 * - At any scheduler domain level, this cpu's scheduler group has multiple
4989 * busy cpu's exceeding the group's power.
4990 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
4991 * domain span are idle.
4993 static inline int nohz_kick_needed(struct rq
*rq
, int cpu
)
4995 unsigned long now
= jiffies
;
4996 struct sched_domain
*sd
;
4998 if (unlikely(idle_cpu(cpu
)))
5002 * We may be recently in ticked or tickless idle mode. At the first
5003 * busy tick after returning from idle, we will update the busy stats.
5005 set_cpu_sd_state_busy();
5006 nohz_balance_exit_idle(cpu
);
5009 * None are in tickless mode and hence no need for NOHZ idle load
5012 if (likely(!atomic_read(&nohz
.nr_cpus
)))
5015 if (time_before(now
, nohz
.next_balance
))
5018 if (rq
->nr_running
>= 2)
5022 for_each_domain(cpu
, sd
) {
5023 struct sched_group
*sg
= sd
->groups
;
5024 struct sched_group_power
*sgp
= sg
->sgp
;
5025 int nr_busy
= atomic_read(&sgp
->nr_busy_cpus
);
5027 if (sd
->flags
& SD_SHARE_PKG_RESOURCES
&& nr_busy
> 1)
5028 goto need_kick_unlock
;
5030 if (sd
->flags
& SD_ASYM_PACKING
&& nr_busy
!= sg
->group_weight
5031 && (cpumask_first_and(nohz
.idle_cpus_mask
,
5032 sched_domain_span(sd
)) < cpu
))
5033 goto need_kick_unlock
;
5035 if (!(sd
->flags
& (SD_SHARE_PKG_RESOURCES
| SD_ASYM_PACKING
)))
5047 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
) { }
5051 * run_rebalance_domains is triggered when needed from the scheduler tick.
5052 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5054 static void run_rebalance_domains(struct softirq_action
*h
)
5056 int this_cpu
= smp_processor_id();
5057 struct rq
*this_rq
= cpu_rq(this_cpu
);
5058 enum cpu_idle_type idle
= this_rq
->idle_balance
?
5059 CPU_IDLE
: CPU_NOT_IDLE
;
5061 rebalance_domains(this_cpu
, idle
);
5064 * If this cpu has a pending nohz_balance_kick, then do the
5065 * balancing on behalf of the other idle cpus whose ticks are
5068 nohz_idle_balance(this_cpu
, idle
);
5071 static inline int on_null_domain(int cpu
)
5073 return !rcu_dereference_sched(cpu_rq(cpu
)->sd
);
5077 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5079 void trigger_load_balance(struct rq
*rq
, int cpu
)
5081 /* Don't need to rebalance while attached to NULL domain */
5082 if (time_after_eq(jiffies
, rq
->next_balance
) &&
5083 likely(!on_null_domain(cpu
)))
5084 raise_softirq(SCHED_SOFTIRQ
);
5086 if (nohz_kick_needed(rq
, cpu
) && likely(!on_null_domain(cpu
)))
5087 nohz_balancer_kick(cpu
);
5091 static void rq_online_fair(struct rq
*rq
)
5096 static void rq_offline_fair(struct rq
*rq
)
5100 /* Ensure any throttled groups are reachable by pick_next_task */
5101 unthrottle_offline_cfs_rqs(rq
);
5104 #endif /* CONFIG_SMP */
5107 * scheduler tick hitting a task of our scheduling class:
5109 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
5111 struct cfs_rq
*cfs_rq
;
5112 struct sched_entity
*se
= &curr
->se
;
5114 for_each_sched_entity(se
) {
5115 cfs_rq
= cfs_rq_of(se
);
5116 entity_tick(cfs_rq
, se
, queued
);
5119 if (sched_feat_numa(NUMA
))
5120 task_tick_numa(rq
, curr
);
5124 * called on fork with the child task as argument from the parent's context
5125 * - child not yet on the tasklist
5126 * - preemption disabled
5128 static void task_fork_fair(struct task_struct
*p
)
5130 struct cfs_rq
*cfs_rq
;
5131 struct sched_entity
*se
= &p
->se
, *curr
;
5132 int this_cpu
= smp_processor_id();
5133 struct rq
*rq
= this_rq();
5134 unsigned long flags
;
5136 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5138 update_rq_clock(rq
);
5140 cfs_rq
= task_cfs_rq(current
);
5141 curr
= cfs_rq
->curr
;
5143 if (unlikely(task_cpu(p
) != this_cpu
)) {
5145 __set_task_cpu(p
, this_cpu
);
5149 update_curr(cfs_rq
);
5152 se
->vruntime
= curr
->vruntime
;
5153 place_entity(cfs_rq
, se
, 1);
5155 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
5157 * Upon rescheduling, sched_class::put_prev_task() will place
5158 * 'current' within the tree based on its new key value.
5160 swap(curr
->vruntime
, se
->vruntime
);
5161 resched_task(rq
->curr
);
5164 se
->vruntime
-= cfs_rq
->min_vruntime
;
5166 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5170 * Priority of the task has changed. Check to see if we preempt
5174 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
5180 * Reschedule if we are currently running on this runqueue and
5181 * our priority decreased, or if we are not currently running on
5182 * this runqueue and our priority is higher than the current's
5184 if (rq
->curr
== p
) {
5185 if (p
->prio
> oldprio
)
5186 resched_task(rq
->curr
);
5188 check_preempt_curr(rq
, p
, 0);
5191 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
5193 struct sched_entity
*se
= &p
->se
;
5194 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5197 * Ensure the task's vruntime is normalized, so that when its
5198 * switched back to the fair class the enqueue_entity(.flags=0) will
5199 * do the right thing.
5201 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5202 * have normalized the vruntime, if it was !on_rq, then only when
5203 * the task is sleeping will it still have non-normalized vruntime.
5205 if (!se
->on_rq
&& p
->state
!= TASK_RUNNING
) {
5207 * Fix up our vruntime so that the current sleep doesn't
5208 * cause 'unlimited' sleep bonus.
5210 place_entity(cfs_rq
, se
, 0);
5211 se
->vruntime
-= cfs_rq
->min_vruntime
;
5216 * We switched to the sched_fair class.
5218 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
5224 * We were most likely switched from sched_rt, so
5225 * kick off the schedule if running, otherwise just see
5226 * if we can still preempt the current task.
5229 resched_task(rq
->curr
);
5231 check_preempt_curr(rq
, p
, 0);
5234 /* Account for a task changing its policy or group.
5236 * This routine is mostly called to set cfs_rq->curr field when a task
5237 * migrates between groups/classes.
5239 static void set_curr_task_fair(struct rq
*rq
)
5241 struct sched_entity
*se
= &rq
->curr
->se
;
5243 for_each_sched_entity(se
) {
5244 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5246 set_next_entity(cfs_rq
, se
);
5247 /* ensure bandwidth has been allocated on our new cfs_rq */
5248 account_cfs_rq_runtime(cfs_rq
, 0);
5252 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
5254 cfs_rq
->tasks_timeline
= RB_ROOT
;
5255 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
5256 #ifndef CONFIG_64BIT
5257 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
5261 #ifdef CONFIG_FAIR_GROUP_SCHED
5262 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
5265 * If the task was not on the rq at the time of this cgroup movement
5266 * it must have been asleep, sleeping tasks keep their ->vruntime
5267 * absolute on their old rq until wakeup (needed for the fair sleeper
5268 * bonus in place_entity()).
5270 * If it was on the rq, we've just 'preempted' it, which does convert
5271 * ->vruntime to a relative base.
5273 * Make sure both cases convert their relative position when migrating
5274 * to another cgroup's rq. This does somewhat interfere with the
5275 * fair sleeper stuff for the first placement, but who cares.
5278 * When !on_rq, vruntime of the task has usually NOT been normalized.
5279 * But there are some cases where it has already been normalized:
5281 * - Moving a forked child which is waiting for being woken up by
5282 * wake_up_new_task().
5283 * - Moving a task which has been woken up by try_to_wake_up() and
5284 * waiting for actually being woken up by sched_ttwu_pending().
5286 * To prevent boost or penalty in the new cfs_rq caused by delta
5287 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5289 if (!on_rq
&& (!p
->se
.sum_exec_runtime
|| p
->state
== TASK_WAKING
))
5293 p
->se
.vruntime
-= cfs_rq_of(&p
->se
)->min_vruntime
;
5294 set_task_rq(p
, task_cpu(p
));
5296 p
->se
.vruntime
+= cfs_rq_of(&p
->se
)->min_vruntime
;
5299 void free_fair_sched_group(struct task_group
*tg
)
5303 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
5305 for_each_possible_cpu(i
) {
5307 kfree(tg
->cfs_rq
[i
]);
5316 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
5318 struct cfs_rq
*cfs_rq
;
5319 struct sched_entity
*se
;
5322 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
5325 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
5329 tg
->shares
= NICE_0_LOAD
;
5331 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
5333 for_each_possible_cpu(i
) {
5334 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
5335 GFP_KERNEL
, cpu_to_node(i
));
5339 se
= kzalloc_node(sizeof(struct sched_entity
),
5340 GFP_KERNEL
, cpu_to_node(i
));
5344 init_cfs_rq(cfs_rq
);
5345 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
5356 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
5358 struct rq
*rq
= cpu_rq(cpu
);
5359 unsigned long flags
;
5362 * Only empty task groups can be destroyed; so we can speculatively
5363 * check on_list without danger of it being re-added.
5365 if (!tg
->cfs_rq
[cpu
]->on_list
)
5368 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5369 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
5370 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5373 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
5374 struct sched_entity
*se
, int cpu
,
5375 struct sched_entity
*parent
)
5377 struct rq
*rq
= cpu_rq(cpu
);
5382 /* allow initial update_cfs_load() to truncate */
5383 cfs_rq
->load_stamp
= 1;
5385 init_cfs_rq_runtime(cfs_rq
);
5387 tg
->cfs_rq
[cpu
] = cfs_rq
;
5390 /* se could be NULL for root_task_group */
5395 se
->cfs_rq
= &rq
->cfs
;
5397 se
->cfs_rq
= parent
->my_q
;
5400 update_load_set(&se
->load
, 0);
5401 se
->parent
= parent
;
5404 static DEFINE_MUTEX(shares_mutex
);
5406 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
5409 unsigned long flags
;
5412 * We can't change the weight of the root cgroup.
5417 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
5419 mutex_lock(&shares_mutex
);
5420 if (tg
->shares
== shares
)
5423 tg
->shares
= shares
;
5424 for_each_possible_cpu(i
) {
5425 struct rq
*rq
= cpu_rq(i
);
5426 struct sched_entity
*se
;
5429 /* Propagate contribution to hierarchy */
5430 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5431 for_each_sched_entity(se
)
5432 update_cfs_shares(group_cfs_rq(se
));
5433 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5437 mutex_unlock(&shares_mutex
);
5440 #else /* CONFIG_FAIR_GROUP_SCHED */
5442 void free_fair_sched_group(struct task_group
*tg
) { }
5444 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
5449 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
5451 #endif /* CONFIG_FAIR_GROUP_SCHED */
5454 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
5456 struct sched_entity
*se
= &task
->se
;
5457 unsigned int rr_interval
= 0;
5460 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5463 if (rq
->cfs
.load
.weight
)
5464 rr_interval
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5470 * All the scheduling class methods:
5472 const struct sched_class fair_sched_class
= {
5473 .next
= &idle_sched_class
,
5474 .enqueue_task
= enqueue_task_fair
,
5475 .dequeue_task
= dequeue_task_fair
,
5476 .yield_task
= yield_task_fair
,
5477 .yield_to_task
= yield_to_task_fair
,
5479 .check_preempt_curr
= check_preempt_wakeup
,
5481 .pick_next_task
= pick_next_task_fair
,
5482 .put_prev_task
= put_prev_task_fair
,
5485 .select_task_rq
= select_task_rq_fair
,
5487 .rq_online
= rq_online_fair
,
5488 .rq_offline
= rq_offline_fair
,
5490 .task_waking
= task_waking_fair
,
5493 .set_curr_task
= set_curr_task_fair
,
5494 .task_tick
= task_tick_fair
,
5495 .task_fork
= task_fork_fair
,
5497 .prio_changed
= prio_changed_fair
,
5498 .switched_from
= switched_from_fair
,
5499 .switched_to
= switched_to_fair
,
5501 .get_rr_interval
= get_rr_interval_fair
,
5503 #ifdef CONFIG_FAIR_GROUP_SCHED
5504 .task_move_group
= task_move_group_fair
,
5508 #ifdef CONFIG_SCHED_DEBUG
5509 void print_cfs_stats(struct seq_file
*m
, int cpu
)
5511 struct cfs_rq
*cfs_rq
;
5514 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
5515 print_cfs_rq(m
, cpu
, cfs_rq
);
5520 __init
void init_sched_fair_class(void)
5523 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
5526 nohz
.next_balance
= jiffies
;
5527 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
5528 cpu_notifier(sched_ilb_notifier
, 0);