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>
28 * Targeted preemption latency for CPU-bound tasks:
29 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
31 * NOTE: this latency value is not the same as the concept of
32 * 'timeslice length' - timeslices in CFS are of variable length
33 * and have no persistent notion like in traditional, time-slice
34 * based scheduling concepts.
36 * (to see the precise effective timeslice length of your workload,
37 * run vmstat and monitor the context-switches (cs) field)
39 unsigned int sysctl_sched_latency
= 6000000ULL;
40 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
43 * The initial- and re-scaling of tunables is configurable
44 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
47 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
48 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
49 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
51 enum sched_tunable_scaling sysctl_sched_tunable_scaling
52 = SCHED_TUNABLESCALING_LOG
;
55 * Minimal preemption granularity for CPU-bound tasks:
56 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
58 unsigned int sysctl_sched_min_granularity
= 750000ULL;
59 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
62 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
64 static unsigned int sched_nr_latency
= 8;
67 * After fork, child runs first. If set to 0 (default) then
68 * parent will (try to) run first.
70 unsigned int sysctl_sched_child_runs_first __read_mostly
;
73 * SCHED_OTHER wake-up granularity.
74 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
76 * This option delays the preemption effects of decoupled workloads
77 * and reduces their over-scheduling. Synchronous workloads will still
78 * have immediate wakeup/sleep latencies.
80 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
81 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
83 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
86 * The exponential sliding window over which load is averaged for shares
90 unsigned int __read_mostly sysctl_sched_shares_window
= 10000000UL;
92 static const struct sched_class fair_sched_class
;
94 /**************************************************************
95 * CFS operations on generic schedulable entities:
98 #ifdef CONFIG_FAIR_GROUP_SCHED
100 /* cpu runqueue to which this cfs_rq is attached */
101 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
106 /* An entity is a task if it doesn't "own" a runqueue */
107 #define entity_is_task(se) (!se->my_q)
109 static inline struct task_struct
*task_of(struct sched_entity
*se
)
111 #ifdef CONFIG_SCHED_DEBUG
112 WARN_ON_ONCE(!entity_is_task(se
));
114 return container_of(se
, struct task_struct
, se
);
117 /* Walk up scheduling entities hierarchy */
118 #define for_each_sched_entity(se) \
119 for (; se; se = se->parent)
121 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
126 /* runqueue on which this entity is (to be) queued */
127 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
132 /* runqueue "owned" by this group */
133 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
138 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
139 * another cpu ('this_cpu')
141 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
143 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
146 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
148 if (!cfs_rq
->on_list
) {
150 * Ensure we either appear before our parent (if already
151 * enqueued) or force our parent to appear after us when it is
152 * enqueued. The fact that we always enqueue bottom-up
153 * reduces this to two cases.
155 if (cfs_rq
->tg
->parent
&&
156 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
157 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
158 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
160 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
161 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
168 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
170 if (cfs_rq
->on_list
) {
171 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
176 /* Iterate thr' all leaf cfs_rq's on a runqueue */
177 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
178 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
180 /* Do the two (enqueued) entities belong to the same group ? */
182 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
184 if (se
->cfs_rq
== pse
->cfs_rq
)
190 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
195 /* return depth at which a sched entity is present in the hierarchy */
196 static inline int depth_se(struct sched_entity
*se
)
200 for_each_sched_entity(se
)
207 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
209 int se_depth
, pse_depth
;
212 * preemption test can be made between sibling entities who are in the
213 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
214 * both tasks until we find their ancestors who are siblings of common
218 /* First walk up until both entities are at same depth */
219 se_depth
= depth_se(*se
);
220 pse_depth
= depth_se(*pse
);
222 while (se_depth
> pse_depth
) {
224 *se
= parent_entity(*se
);
227 while (pse_depth
> se_depth
) {
229 *pse
= parent_entity(*pse
);
232 while (!is_same_group(*se
, *pse
)) {
233 *se
= parent_entity(*se
);
234 *pse
= parent_entity(*pse
);
238 #else /* !CONFIG_FAIR_GROUP_SCHED */
240 static inline struct task_struct
*task_of(struct sched_entity
*se
)
242 return container_of(se
, struct task_struct
, se
);
245 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
247 return container_of(cfs_rq
, struct rq
, cfs
);
250 #define entity_is_task(se) 1
252 #define for_each_sched_entity(se) \
253 for (; se; se = NULL)
255 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
257 return &task_rq(p
)->cfs
;
260 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
262 struct task_struct
*p
= task_of(se
);
263 struct rq
*rq
= task_rq(p
);
268 /* runqueue "owned" by this group */
269 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
274 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
276 return &cpu_rq(this_cpu
)->cfs
;
279 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
283 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
287 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
288 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
291 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
296 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
302 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
306 #endif /* CONFIG_FAIR_GROUP_SCHED */
309 /**************************************************************
310 * Scheduling class tree data structure manipulation methods:
313 static inline u64
max_vruntime(u64 min_vruntime
, u64 vruntime
)
315 s64 delta
= (s64
)(vruntime
- min_vruntime
);
317 min_vruntime
= vruntime
;
322 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
324 s64 delta
= (s64
)(vruntime
- min_vruntime
);
326 min_vruntime
= vruntime
;
331 static inline int entity_before(struct sched_entity
*a
,
332 struct sched_entity
*b
)
334 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
337 static inline s64
entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
339 return se
->vruntime
- cfs_rq
->min_vruntime
;
342 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
344 u64 vruntime
= cfs_rq
->min_vruntime
;
347 vruntime
= cfs_rq
->curr
->vruntime
;
349 if (cfs_rq
->rb_leftmost
) {
350 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
355 vruntime
= se
->vruntime
;
357 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
360 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
363 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
368 * Enqueue an entity into the rb-tree:
370 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
372 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
373 struct rb_node
*parent
= NULL
;
374 struct sched_entity
*entry
;
375 s64 key
= entity_key(cfs_rq
, se
);
379 * Find the right place in the rbtree:
383 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
385 * We dont care about collisions. Nodes with
386 * the same key stay together.
388 if (key
< entity_key(cfs_rq
, entry
)) {
389 link
= &parent
->rb_left
;
391 link
= &parent
->rb_right
;
397 * Maintain a cache of leftmost tree entries (it is frequently
401 cfs_rq
->rb_leftmost
= &se
->run_node
;
403 rb_link_node(&se
->run_node
, parent
, link
);
404 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
407 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
409 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
410 struct rb_node
*next_node
;
412 next_node
= rb_next(&se
->run_node
);
413 cfs_rq
->rb_leftmost
= next_node
;
416 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
419 static struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
421 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
426 return rb_entry(left
, struct sched_entity
, run_node
);
429 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
431 struct rb_node
*next
= rb_next(&se
->run_node
);
436 return rb_entry(next
, struct sched_entity
, run_node
);
439 #ifdef CONFIG_SCHED_DEBUG
440 static struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
442 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
447 return rb_entry(last
, struct sched_entity
, run_node
);
450 /**************************************************************
451 * Scheduling class statistics methods:
454 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
455 void __user
*buffer
, size_t *lenp
,
458 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
459 int factor
= get_update_sysctl_factor();
464 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
465 sysctl_sched_min_granularity
);
467 #define WRT_SYSCTL(name) \
468 (normalized_sysctl_##name = sysctl_##name / (factor))
469 WRT_SYSCTL(sched_min_granularity
);
470 WRT_SYSCTL(sched_latency
);
471 WRT_SYSCTL(sched_wakeup_granularity
);
481 static inline unsigned long
482 calc_delta_fair(unsigned long delta
, struct sched_entity
*se
)
484 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
485 delta
= calc_delta_mine(delta
, NICE_0_LOAD
, &se
->load
);
491 * The idea is to set a period in which each task runs once.
493 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
494 * this period because otherwise the slices get too small.
496 * p = (nr <= nl) ? l : l*nr/nl
498 static u64
__sched_period(unsigned long nr_running
)
500 u64 period
= sysctl_sched_latency
;
501 unsigned long nr_latency
= sched_nr_latency
;
503 if (unlikely(nr_running
> nr_latency
)) {
504 period
= sysctl_sched_min_granularity
;
505 period
*= nr_running
;
512 * We calculate the wall-time slice from the period by taking a part
513 * proportional to the weight.
517 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
519 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
521 for_each_sched_entity(se
) {
522 struct load_weight
*load
;
523 struct load_weight lw
;
525 cfs_rq
= cfs_rq_of(se
);
526 load
= &cfs_rq
->load
;
528 if (unlikely(!se
->on_rq
)) {
531 update_load_add(&lw
, se
->load
.weight
);
534 slice
= calc_delta_mine(slice
, se
->load
.weight
, load
);
540 * We calculate the vruntime slice of a to be inserted task
544 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
546 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
549 static void update_cfs_load(struct cfs_rq
*cfs_rq
, int global_update
);
550 static void update_cfs_shares(struct cfs_rq
*cfs_rq
);
553 * Update the current task's runtime statistics. Skip current tasks that
554 * are not in our scheduling class.
557 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
558 unsigned long delta_exec
)
560 unsigned long delta_exec_weighted
;
562 schedstat_set(curr
->statistics
.exec_max
,
563 max((u64
)delta_exec
, curr
->statistics
.exec_max
));
565 curr
->sum_exec_runtime
+= delta_exec
;
566 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
567 delta_exec_weighted
= calc_delta_fair(delta_exec
, curr
);
569 curr
->vruntime
+= delta_exec_weighted
;
570 update_min_vruntime(cfs_rq
);
572 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
573 cfs_rq
->load_unacc_exec_time
+= delta_exec
;
577 static void update_curr(struct cfs_rq
*cfs_rq
)
579 struct sched_entity
*curr
= cfs_rq
->curr
;
580 u64 now
= rq_of(cfs_rq
)->clock_task
;
581 unsigned long delta_exec
;
587 * Get the amount of time the current task was running
588 * since the last time we changed load (this cannot
589 * overflow on 32 bits):
591 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
595 __update_curr(cfs_rq
, curr
, delta_exec
);
596 curr
->exec_start
= now
;
598 if (entity_is_task(curr
)) {
599 struct task_struct
*curtask
= task_of(curr
);
601 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
602 cpuacct_charge(curtask
, delta_exec
);
603 account_group_exec_runtime(curtask
, delta_exec
);
608 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
610 schedstat_set(se
->statistics
.wait_start
, rq_of(cfs_rq
)->clock
);
614 * Task is being enqueued - update stats:
616 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
619 * Are we enqueueing a waiting task? (for current tasks
620 * a dequeue/enqueue event is a NOP)
622 if (se
!= cfs_rq
->curr
)
623 update_stats_wait_start(cfs_rq
, se
);
627 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
629 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
630 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
));
631 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
632 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
633 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
);
634 #ifdef CONFIG_SCHEDSTATS
635 if (entity_is_task(se
)) {
636 trace_sched_stat_wait(task_of(se
),
637 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
);
640 schedstat_set(se
->statistics
.wait_start
, 0);
644 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
647 * Mark the end of the wait period if dequeueing a
650 if (se
!= cfs_rq
->curr
)
651 update_stats_wait_end(cfs_rq
, se
);
655 * We are picking a new current task - update its stats:
658 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
661 * We are starting a new run period:
663 se
->exec_start
= rq_of(cfs_rq
)->clock_task
;
666 /**************************************************
667 * Scheduling class queueing methods:
670 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
672 add_cfs_task_weight(struct cfs_rq
*cfs_rq
, unsigned long weight
)
674 cfs_rq
->task_weight
+= weight
;
678 add_cfs_task_weight(struct cfs_rq
*cfs_rq
, unsigned long weight
)
684 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
686 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
687 if (!parent_entity(se
))
688 inc_cpu_load(rq_of(cfs_rq
), se
->load
.weight
);
689 if (entity_is_task(se
)) {
690 add_cfs_task_weight(cfs_rq
, se
->load
.weight
);
691 list_add(&se
->group_node
, &cfs_rq
->tasks
);
693 cfs_rq
->nr_running
++;
697 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
699 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
700 if (!parent_entity(se
))
701 dec_cpu_load(rq_of(cfs_rq
), se
->load
.weight
);
702 if (entity_is_task(se
)) {
703 add_cfs_task_weight(cfs_rq
, -se
->load
.weight
);
704 list_del_init(&se
->group_node
);
706 cfs_rq
->nr_running
--;
709 #ifdef CONFIG_FAIR_GROUP_SCHED
711 static void update_cfs_rq_load_contribution(struct cfs_rq
*cfs_rq
,
714 struct task_group
*tg
= cfs_rq
->tg
;
717 load_avg
= div64_u64(cfs_rq
->load_avg
, cfs_rq
->load_period
+1);
718 load_avg
-= cfs_rq
->load_contribution
;
720 if (global_update
|| abs(load_avg
) > cfs_rq
->load_contribution
/ 8) {
721 atomic_add(load_avg
, &tg
->load_weight
);
722 cfs_rq
->load_contribution
+= load_avg
;
726 static void update_cfs_load(struct cfs_rq
*cfs_rq
, int global_update
)
728 u64 period
= sysctl_sched_shares_window
;
730 unsigned long load
= cfs_rq
->load
.weight
;
732 if (cfs_rq
->tg
== &root_task_group
)
735 now
= rq_of(cfs_rq
)->clock_task
;
736 delta
= now
- cfs_rq
->load_stamp
;
738 /* truncate load history at 4 idle periods */
739 if (cfs_rq
->load_stamp
> cfs_rq
->load_last
&&
740 now
- cfs_rq
->load_last
> 4 * period
) {
741 cfs_rq
->load_period
= 0;
742 cfs_rq
->load_avg
= 0;
746 cfs_rq
->load_stamp
= now
;
747 cfs_rq
->load_unacc_exec_time
= 0;
748 cfs_rq
->load_period
+= delta
;
750 cfs_rq
->load_last
= now
;
751 cfs_rq
->load_avg
+= delta
* load
;
754 /* consider updating load contribution on each fold or truncate */
755 if (global_update
|| cfs_rq
->load_period
> period
756 || !cfs_rq
->load_period
)
757 update_cfs_rq_load_contribution(cfs_rq
, global_update
);
759 while (cfs_rq
->load_period
> period
) {
761 * Inline assembly required to prevent the compiler
762 * optimising this loop into a divmod call.
763 * See __iter_div_u64_rem() for another example of this.
765 asm("" : "+rm" (cfs_rq
->load_period
));
766 cfs_rq
->load_period
/= 2;
767 cfs_rq
->load_avg
/= 2;
770 if (!cfs_rq
->curr
&& !cfs_rq
->nr_running
&& !cfs_rq
->load_avg
)
771 list_del_leaf_cfs_rq(cfs_rq
);
774 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
776 long load_weight
, load
, shares
;
778 load
= cfs_rq
->load
.weight
;
780 load_weight
= atomic_read(&tg
->load_weight
);
782 load_weight
-= cfs_rq
->load_contribution
;
784 shares
= (tg
->shares
* load
);
786 shares
/= load_weight
;
788 if (shares
< MIN_SHARES
)
790 if (shares
> tg
->shares
)
796 static void update_entity_shares_tick(struct cfs_rq
*cfs_rq
)
798 if (cfs_rq
->load_unacc_exec_time
> sysctl_sched_shares_window
) {
799 update_cfs_load(cfs_rq
, 0);
800 update_cfs_shares(cfs_rq
);
803 # else /* CONFIG_SMP */
804 static void update_cfs_load(struct cfs_rq
*cfs_rq
, int global_update
)
808 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
813 static inline void update_entity_shares_tick(struct cfs_rq
*cfs_rq
)
816 # endif /* CONFIG_SMP */
817 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
818 unsigned long weight
)
821 /* commit outstanding execution time */
822 if (cfs_rq
->curr
== se
)
824 account_entity_dequeue(cfs_rq
, se
);
827 update_load_set(&se
->load
, weight
);
830 account_entity_enqueue(cfs_rq
, se
);
833 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
835 struct task_group
*tg
;
836 struct sched_entity
*se
;
840 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
844 if (likely(se
->load
.weight
== tg
->shares
))
847 shares
= calc_cfs_shares(cfs_rq
, tg
);
849 reweight_entity(cfs_rq_of(se
), se
, shares
);
851 #else /* CONFIG_FAIR_GROUP_SCHED */
852 static void update_cfs_load(struct cfs_rq
*cfs_rq
, int global_update
)
856 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
860 static inline void update_entity_shares_tick(struct cfs_rq
*cfs_rq
)
863 #endif /* CONFIG_FAIR_GROUP_SCHED */
865 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
867 #ifdef CONFIG_SCHEDSTATS
868 struct task_struct
*tsk
= NULL
;
870 if (entity_is_task(se
))
873 if (se
->statistics
.sleep_start
) {
874 u64 delta
= rq_of(cfs_rq
)->clock
- se
->statistics
.sleep_start
;
879 if (unlikely(delta
> se
->statistics
.sleep_max
))
880 se
->statistics
.sleep_max
= delta
;
882 se
->statistics
.sleep_start
= 0;
883 se
->statistics
.sum_sleep_runtime
+= delta
;
886 account_scheduler_latency(tsk
, delta
>> 10, 1);
887 trace_sched_stat_sleep(tsk
, delta
);
890 if (se
->statistics
.block_start
) {
891 u64 delta
= rq_of(cfs_rq
)->clock
- se
->statistics
.block_start
;
896 if (unlikely(delta
> se
->statistics
.block_max
))
897 se
->statistics
.block_max
= delta
;
899 se
->statistics
.block_start
= 0;
900 se
->statistics
.sum_sleep_runtime
+= delta
;
903 if (tsk
->in_iowait
) {
904 se
->statistics
.iowait_sum
+= delta
;
905 se
->statistics
.iowait_count
++;
906 trace_sched_stat_iowait(tsk
, delta
);
910 * Blocking time is in units of nanosecs, so shift by
911 * 20 to get a milliseconds-range estimation of the
912 * amount of time that the task spent sleeping:
914 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
915 profile_hits(SLEEP_PROFILING
,
916 (void *)get_wchan(tsk
),
919 account_scheduler_latency(tsk
, delta
>> 10, 0);
925 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
927 #ifdef CONFIG_SCHED_DEBUG
928 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
933 if (d
> 3*sysctl_sched_latency
)
934 schedstat_inc(cfs_rq
, nr_spread_over
);
939 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
941 u64 vruntime
= cfs_rq
->min_vruntime
;
944 * The 'current' period is already promised to the current tasks,
945 * however the extra weight of the new task will slow them down a
946 * little, place the new task so that it fits in the slot that
947 * stays open at the end.
949 if (initial
&& sched_feat(START_DEBIT
))
950 vruntime
+= sched_vslice(cfs_rq
, se
);
952 /* sleeps up to a single latency don't count. */
954 unsigned long thresh
= sysctl_sched_latency
;
957 * Halve their sleep time's effect, to allow
958 * for a gentler effect of sleepers:
960 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
966 /* ensure we never gain time by being placed backwards. */
967 vruntime
= max_vruntime(se
->vruntime
, vruntime
);
969 se
->vruntime
= vruntime
;
973 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
976 * Update the normalized vruntime before updating min_vruntime
977 * through callig update_curr().
979 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
980 se
->vruntime
+= cfs_rq
->min_vruntime
;
983 * Update run-time statistics of the 'current'.
986 update_cfs_load(cfs_rq
, 0);
987 account_entity_enqueue(cfs_rq
, se
);
988 update_cfs_shares(cfs_rq
);
990 if (flags
& ENQUEUE_WAKEUP
) {
991 place_entity(cfs_rq
, se
, 0);
992 enqueue_sleeper(cfs_rq
, se
);
995 update_stats_enqueue(cfs_rq
, se
);
996 check_spread(cfs_rq
, se
);
997 if (se
!= cfs_rq
->curr
)
998 __enqueue_entity(cfs_rq
, se
);
1001 if (cfs_rq
->nr_running
== 1)
1002 list_add_leaf_cfs_rq(cfs_rq
);
1005 static void __clear_buddies_last(struct sched_entity
*se
)
1007 for_each_sched_entity(se
) {
1008 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1009 if (cfs_rq
->last
== se
)
1010 cfs_rq
->last
= NULL
;
1016 static void __clear_buddies_next(struct sched_entity
*se
)
1018 for_each_sched_entity(se
) {
1019 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1020 if (cfs_rq
->next
== se
)
1021 cfs_rq
->next
= NULL
;
1027 static void __clear_buddies_skip(struct sched_entity
*se
)
1029 for_each_sched_entity(se
) {
1030 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1031 if (cfs_rq
->skip
== se
)
1032 cfs_rq
->skip
= NULL
;
1038 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1040 if (cfs_rq
->last
== se
)
1041 __clear_buddies_last(se
);
1043 if (cfs_rq
->next
== se
)
1044 __clear_buddies_next(se
);
1046 if (cfs_rq
->skip
== se
)
1047 __clear_buddies_skip(se
);
1051 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1054 * Update run-time statistics of the 'current'.
1056 update_curr(cfs_rq
);
1058 update_stats_dequeue(cfs_rq
, se
);
1059 if (flags
& DEQUEUE_SLEEP
) {
1060 #ifdef CONFIG_SCHEDSTATS
1061 if (entity_is_task(se
)) {
1062 struct task_struct
*tsk
= task_of(se
);
1064 if (tsk
->state
& TASK_INTERRUPTIBLE
)
1065 se
->statistics
.sleep_start
= rq_of(cfs_rq
)->clock
;
1066 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
1067 se
->statistics
.block_start
= rq_of(cfs_rq
)->clock
;
1072 clear_buddies(cfs_rq
, se
);
1074 if (se
!= cfs_rq
->curr
)
1075 __dequeue_entity(cfs_rq
, se
);
1077 update_cfs_load(cfs_rq
, 0);
1078 account_entity_dequeue(cfs_rq
, se
);
1081 * Normalize the entity after updating the min_vruntime because the
1082 * update can refer to the ->curr item and we need to reflect this
1083 * movement in our normalized position.
1085 if (!(flags
& DEQUEUE_SLEEP
))
1086 se
->vruntime
-= cfs_rq
->min_vruntime
;
1088 update_min_vruntime(cfs_rq
);
1089 update_cfs_shares(cfs_rq
);
1093 * Preempt the current task with a newly woken task if needed:
1096 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
1098 unsigned long ideal_runtime
, delta_exec
;
1100 ideal_runtime
= sched_slice(cfs_rq
, curr
);
1101 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
1102 if (delta_exec
> ideal_runtime
) {
1103 resched_task(rq_of(cfs_rq
)->curr
);
1105 * The current task ran long enough, ensure it doesn't get
1106 * re-elected due to buddy favours.
1108 clear_buddies(cfs_rq
, curr
);
1113 * Ensure that a task that missed wakeup preemption by a
1114 * narrow margin doesn't have to wait for a full slice.
1115 * This also mitigates buddy induced latencies under load.
1117 if (!sched_feat(WAKEUP_PREEMPT
))
1120 if (delta_exec
< sysctl_sched_min_granularity
)
1123 if (cfs_rq
->nr_running
> 1) {
1124 struct sched_entity
*se
= __pick_first_entity(cfs_rq
);
1125 s64 delta
= curr
->vruntime
- se
->vruntime
;
1130 if (delta
> ideal_runtime
)
1131 resched_task(rq_of(cfs_rq
)->curr
);
1136 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1138 /* 'current' is not kept within the tree. */
1141 * Any task has to be enqueued before it get to execute on
1142 * a CPU. So account for the time it spent waiting on the
1145 update_stats_wait_end(cfs_rq
, se
);
1146 __dequeue_entity(cfs_rq
, se
);
1149 update_stats_curr_start(cfs_rq
, se
);
1151 #ifdef CONFIG_SCHEDSTATS
1153 * Track our maximum slice length, if the CPU's load is at
1154 * least twice that of our own weight (i.e. dont track it
1155 * when there are only lesser-weight tasks around):
1157 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
1158 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
1159 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
1162 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
1166 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
1169 * Pick the next process, keeping these things in mind, in this order:
1170 * 1) keep things fair between processes/task groups
1171 * 2) pick the "next" process, since someone really wants that to run
1172 * 3) pick the "last" process, for cache locality
1173 * 4) do not run the "skip" process, if something else is available
1175 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
1177 struct sched_entity
*se
= __pick_first_entity(cfs_rq
);
1178 struct sched_entity
*left
= se
;
1181 * Avoid running the skip buddy, if running something else can
1182 * be done without getting too unfair.
1184 if (cfs_rq
->skip
== se
) {
1185 struct sched_entity
*second
= __pick_next_entity(se
);
1186 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
1191 * Prefer last buddy, try to return the CPU to a preempted task.
1193 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
1197 * Someone really wants this to run. If it's not unfair, run it.
1199 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
1202 clear_buddies(cfs_rq
, se
);
1207 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
1210 * If still on the runqueue then deactivate_task()
1211 * was not called and update_curr() has to be done:
1214 update_curr(cfs_rq
);
1216 check_spread(cfs_rq
, prev
);
1218 update_stats_wait_start(cfs_rq
, prev
);
1219 /* Put 'current' back into the tree. */
1220 __enqueue_entity(cfs_rq
, prev
);
1222 cfs_rq
->curr
= NULL
;
1226 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
1229 * Update run-time statistics of the 'current'.
1231 update_curr(cfs_rq
);
1234 * Update share accounting for long-running entities.
1236 update_entity_shares_tick(cfs_rq
);
1238 #ifdef CONFIG_SCHED_HRTICK
1240 * queued ticks are scheduled to match the slice, so don't bother
1241 * validating it and just reschedule.
1244 resched_task(rq_of(cfs_rq
)->curr
);
1248 * don't let the period tick interfere with the hrtick preemption
1250 if (!sched_feat(DOUBLE_TICK
) &&
1251 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
1255 if (cfs_rq
->nr_running
> 1 || !sched_feat(WAKEUP_PREEMPT
))
1256 check_preempt_tick(cfs_rq
, curr
);
1259 /**************************************************
1260 * CFS operations on tasks:
1263 #ifdef CONFIG_SCHED_HRTICK
1264 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
1266 struct sched_entity
*se
= &p
->se
;
1267 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1269 WARN_ON(task_rq(p
) != rq
);
1271 if (hrtick_enabled(rq
) && cfs_rq
->nr_running
> 1) {
1272 u64 slice
= sched_slice(cfs_rq
, se
);
1273 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
1274 s64 delta
= slice
- ran
;
1283 * Don't schedule slices shorter than 10000ns, that just
1284 * doesn't make sense. Rely on vruntime for fairness.
1287 delta
= max_t(s64
, 10000LL, delta
);
1289 hrtick_start(rq
, delta
);
1294 * called from enqueue/dequeue and updates the hrtick when the
1295 * current task is from our class and nr_running is low enough
1298 static void hrtick_update(struct rq
*rq
)
1300 struct task_struct
*curr
= rq
->curr
;
1302 if (curr
->sched_class
!= &fair_sched_class
)
1305 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
1306 hrtick_start_fair(rq
, curr
);
1308 #else /* !CONFIG_SCHED_HRTICK */
1310 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
1314 static inline void hrtick_update(struct rq
*rq
)
1320 * The enqueue_task method is called before nr_running is
1321 * increased. Here we update the fair scheduling stats and
1322 * then put the task into the rbtree:
1325 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
1327 struct cfs_rq
*cfs_rq
;
1328 struct sched_entity
*se
= &p
->se
;
1330 for_each_sched_entity(se
) {
1333 cfs_rq
= cfs_rq_of(se
);
1334 enqueue_entity(cfs_rq
, se
, flags
);
1335 flags
= ENQUEUE_WAKEUP
;
1338 for_each_sched_entity(se
) {
1339 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1341 update_cfs_load(cfs_rq
, 0);
1342 update_cfs_shares(cfs_rq
);
1348 static void set_next_buddy(struct sched_entity
*se
);
1351 * The dequeue_task method is called before nr_running is
1352 * decreased. We remove the task from the rbtree and
1353 * update the fair scheduling stats:
1355 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
1357 struct cfs_rq
*cfs_rq
;
1358 struct sched_entity
*se
= &p
->se
;
1359 int task_sleep
= flags
& DEQUEUE_SLEEP
;
1361 for_each_sched_entity(se
) {
1362 cfs_rq
= cfs_rq_of(se
);
1363 dequeue_entity(cfs_rq
, se
, flags
);
1365 /* Don't dequeue parent if it has other entities besides us */
1366 if (cfs_rq
->load
.weight
) {
1368 * Bias pick_next to pick a task from this cfs_rq, as
1369 * p is sleeping when it is within its sched_slice.
1371 if (task_sleep
&& parent_entity(se
))
1372 set_next_buddy(parent_entity(se
));
1375 flags
|= DEQUEUE_SLEEP
;
1378 for_each_sched_entity(se
) {
1379 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1381 update_cfs_load(cfs_rq
, 0);
1382 update_cfs_shares(cfs_rq
);
1390 static void task_waking_fair(struct task_struct
*p
)
1392 struct sched_entity
*se
= &p
->se
;
1393 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1396 #ifndef CONFIG_64BIT
1397 u64 min_vruntime_copy
;
1400 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
1402 min_vruntime
= cfs_rq
->min_vruntime
;
1403 } while (min_vruntime
!= min_vruntime_copy
);
1405 min_vruntime
= cfs_rq
->min_vruntime
;
1408 se
->vruntime
-= min_vruntime
;
1411 #ifdef CONFIG_FAIR_GROUP_SCHED
1413 * effective_load() calculates the load change as seen from the root_task_group
1415 * Adding load to a group doesn't make a group heavier, but can cause movement
1416 * of group shares between cpus. Assuming the shares were perfectly aligned one
1417 * can calculate the shift in shares.
1419 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
1421 struct sched_entity
*se
= tg
->se
[cpu
];
1426 for_each_sched_entity(se
) {
1430 w
= se
->my_q
->load
.weight
;
1432 /* use this cpu's instantaneous contribution */
1433 lw
= atomic_read(&tg
->load_weight
);
1434 lw
-= se
->my_q
->load_contribution
;
1439 if (lw
> 0 && wl
< lw
)
1440 wl
= (wl
* tg
->shares
) / lw
;
1444 /* zero point is MIN_SHARES */
1445 if (wl
< MIN_SHARES
)
1447 wl
-= se
->load
.weight
;
1456 static inline unsigned long effective_load(struct task_group
*tg
, int cpu
,
1457 unsigned long wl
, unsigned long wg
)
1464 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
1466 s64 this_load
, load
;
1467 int idx
, this_cpu
, prev_cpu
;
1468 unsigned long tl_per_task
;
1469 struct task_group
*tg
;
1470 unsigned long weight
;
1474 this_cpu
= smp_processor_id();
1475 prev_cpu
= task_cpu(p
);
1476 load
= source_load(prev_cpu
, idx
);
1477 this_load
= target_load(this_cpu
, idx
);
1480 * If sync wakeup then subtract the (maximum possible)
1481 * effect of the currently running task from the load
1482 * of the current CPU:
1485 tg
= task_group(current
);
1486 weight
= current
->se
.load
.weight
;
1488 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
1489 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
1493 weight
= p
->se
.load
.weight
;
1496 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1497 * due to the sync cause above having dropped this_load to 0, we'll
1498 * always have an imbalance, but there's really nothing you can do
1499 * about that, so that's good too.
1501 * Otherwise check if either cpus are near enough in load to allow this
1502 * task to be woken on this_cpu.
1504 if (this_load
> 0) {
1505 s64 this_eff_load
, prev_eff_load
;
1507 this_eff_load
= 100;
1508 this_eff_load
*= power_of(prev_cpu
);
1509 this_eff_load
*= this_load
+
1510 effective_load(tg
, this_cpu
, weight
, weight
);
1512 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
1513 prev_eff_load
*= power_of(this_cpu
);
1514 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
1516 balanced
= this_eff_load
<= prev_eff_load
;
1521 * If the currently running task will sleep within
1522 * a reasonable amount of time then attract this newly
1525 if (sync
&& balanced
)
1528 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
1529 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1532 (this_load
<= load
&&
1533 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
1535 * This domain has SD_WAKE_AFFINE and
1536 * p is cache cold in this domain, and
1537 * there is no bad imbalance.
1539 schedstat_inc(sd
, ttwu_move_affine
);
1540 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
1548 * find_idlest_group finds and returns the least busy CPU group within the
1551 static struct sched_group
*
1552 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
1553 int this_cpu
, int load_idx
)
1555 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
1556 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1557 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1560 unsigned long load
, avg_load
;
1564 /* Skip over this group if it has no CPUs allowed */
1565 if (!cpumask_intersects(sched_group_cpus(group
),
1569 local_group
= cpumask_test_cpu(this_cpu
,
1570 sched_group_cpus(group
));
1572 /* Tally up the load of all CPUs in the group */
1575 for_each_cpu(i
, sched_group_cpus(group
)) {
1576 /* Bias balancing toward cpus of our domain */
1578 load
= source_load(i
, load_idx
);
1580 load
= target_load(i
, load_idx
);
1585 /* Adjust by relative CPU power of the group */
1586 avg_load
= (avg_load
* SCHED_POWER_SCALE
) / group
->sgp
->power
;
1589 this_load
= avg_load
;
1590 } else if (avg_load
< min_load
) {
1591 min_load
= avg_load
;
1594 } while (group
= group
->next
, group
!= sd
->groups
);
1596 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1602 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1605 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1607 unsigned long load
, min_load
= ULONG_MAX
;
1611 /* Traverse only the allowed CPUs */
1612 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
1613 load
= weighted_cpuload(i
);
1615 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1625 * Try and locate an idle CPU in the sched_domain.
1627 static int select_idle_sibling(struct task_struct
*p
, int target
)
1629 int cpu
= smp_processor_id();
1630 int prev_cpu
= task_cpu(p
);
1631 struct sched_domain
*sd
;
1635 * If the task is going to be woken-up on this cpu and if it is
1636 * already idle, then it is the right target.
1638 if (target
== cpu
&& idle_cpu(cpu
))
1642 * If the task is going to be woken-up on the cpu where it previously
1643 * ran and if it is currently idle, then it the right target.
1645 if (target
== prev_cpu
&& idle_cpu(prev_cpu
))
1649 * Otherwise, iterate the domains and find an elegible idle cpu.
1652 for_each_domain(target
, sd
) {
1653 if (!(sd
->flags
& SD_SHARE_PKG_RESOURCES
))
1656 for_each_cpu_and(i
, sched_domain_span(sd
), &p
->cpus_allowed
) {
1664 * Lets stop looking for an idle sibling when we reached
1665 * the domain that spans the current cpu and prev_cpu.
1667 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
)) &&
1668 cpumask_test_cpu(prev_cpu
, sched_domain_span(sd
)))
1677 * sched_balance_self: balance the current task (running on cpu) in domains
1678 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1681 * Balance, ie. select the least loaded group.
1683 * Returns the target CPU number, or the same CPU if no balancing is needed.
1685 * preempt must be disabled.
1688 select_task_rq_fair(struct task_struct
*p
, int sd_flag
, int wake_flags
)
1690 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
1691 int cpu
= smp_processor_id();
1692 int prev_cpu
= task_cpu(p
);
1694 int want_affine
= 0;
1696 int sync
= wake_flags
& WF_SYNC
;
1698 if (sd_flag
& SD_BALANCE_WAKE
) {
1699 if (cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
1705 for_each_domain(cpu
, tmp
) {
1706 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
1710 * If power savings logic is enabled for a domain, see if we
1711 * are not overloaded, if so, don't balance wider.
1713 if (tmp
->flags
& (SD_POWERSAVINGS_BALANCE
|SD_PREFER_LOCAL
)) {
1714 unsigned long power
= 0;
1715 unsigned long nr_running
= 0;
1716 unsigned long capacity
;
1719 for_each_cpu(i
, sched_domain_span(tmp
)) {
1720 power
+= power_of(i
);
1721 nr_running
+= cpu_rq(i
)->cfs
.nr_running
;
1724 capacity
= DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
);
1726 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1729 if (nr_running
< capacity
)
1734 * If both cpu and prev_cpu are part of this domain,
1735 * cpu is a valid SD_WAKE_AFFINE target.
1737 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
1738 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
1743 if (!want_sd
&& !want_affine
)
1746 if (!(tmp
->flags
& sd_flag
))
1754 if (cpu
== prev_cpu
|| wake_affine(affine_sd
, p
, sync
))
1757 new_cpu
= select_idle_sibling(p
, prev_cpu
);
1762 int load_idx
= sd
->forkexec_idx
;
1763 struct sched_group
*group
;
1766 if (!(sd
->flags
& sd_flag
)) {
1771 if (sd_flag
& SD_BALANCE_WAKE
)
1772 load_idx
= sd
->wake_idx
;
1774 group
= find_idlest_group(sd
, p
, cpu
, load_idx
);
1780 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
1781 if (new_cpu
== -1 || new_cpu
== cpu
) {
1782 /* Now try balancing at a lower domain level of cpu */
1787 /* Now try balancing at a lower domain level of new_cpu */
1789 weight
= sd
->span_weight
;
1791 for_each_domain(cpu
, tmp
) {
1792 if (weight
<= tmp
->span_weight
)
1794 if (tmp
->flags
& sd_flag
)
1797 /* while loop will break here if sd == NULL */
1804 #endif /* CONFIG_SMP */
1806 static unsigned long
1807 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
1809 unsigned long gran
= sysctl_sched_wakeup_granularity
;
1812 * Since its curr running now, convert the gran from real-time
1813 * to virtual-time in his units.
1815 * By using 'se' instead of 'curr' we penalize light tasks, so
1816 * they get preempted easier. That is, if 'se' < 'curr' then
1817 * the resulting gran will be larger, therefore penalizing the
1818 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1819 * be smaller, again penalizing the lighter task.
1821 * This is especially important for buddies when the leftmost
1822 * task is higher priority than the buddy.
1824 return calc_delta_fair(gran
, se
);
1828 * Should 'se' preempt 'curr'.
1842 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
1844 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
1849 gran
= wakeup_gran(curr
, se
);
1856 static void set_last_buddy(struct sched_entity
*se
)
1858 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
1861 for_each_sched_entity(se
)
1862 cfs_rq_of(se
)->last
= se
;
1865 static void set_next_buddy(struct sched_entity
*se
)
1867 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
1870 for_each_sched_entity(se
)
1871 cfs_rq_of(se
)->next
= se
;
1874 static void set_skip_buddy(struct sched_entity
*se
)
1876 for_each_sched_entity(se
)
1877 cfs_rq_of(se
)->skip
= se
;
1881 * Preempt the current task with a newly woken task if needed:
1883 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1885 struct task_struct
*curr
= rq
->curr
;
1886 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
1887 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
1888 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
1889 int next_buddy_marked
= 0;
1891 if (unlikely(se
== pse
))
1894 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
1895 set_next_buddy(pse
);
1896 next_buddy_marked
= 1;
1900 * We can come here with TIF_NEED_RESCHED already set from new task
1903 if (test_tsk_need_resched(curr
))
1906 /* Idle tasks are by definition preempted by non-idle tasks. */
1907 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
1908 likely(p
->policy
!= SCHED_IDLE
))
1912 * Batch and idle tasks do not preempt non-idle tasks (their preemption
1913 * is driven by the tick):
1915 if (unlikely(p
->policy
!= SCHED_NORMAL
))
1919 if (!sched_feat(WAKEUP_PREEMPT
))
1922 update_curr(cfs_rq
);
1923 find_matching_se(&se
, &pse
);
1925 if (wakeup_preempt_entity(se
, pse
) == 1) {
1927 * Bias pick_next to pick the sched entity that is
1928 * triggering this preemption.
1930 if (!next_buddy_marked
)
1931 set_next_buddy(pse
);
1940 * Only set the backward buddy when the current task is still
1941 * on the rq. This can happen when a wakeup gets interleaved
1942 * with schedule on the ->pre_schedule() or idle_balance()
1943 * point, either of which can * drop the rq lock.
1945 * Also, during early boot the idle thread is in the fair class,
1946 * for obvious reasons its a bad idea to schedule back to it.
1948 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
1951 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
1955 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
1957 struct task_struct
*p
;
1958 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
1959 struct sched_entity
*se
;
1961 if (!cfs_rq
->nr_running
)
1965 se
= pick_next_entity(cfs_rq
);
1966 set_next_entity(cfs_rq
, se
);
1967 cfs_rq
= group_cfs_rq(se
);
1971 hrtick_start_fair(rq
, p
);
1977 * Account for a descheduled task:
1979 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
1981 struct sched_entity
*se
= &prev
->se
;
1982 struct cfs_rq
*cfs_rq
;
1984 for_each_sched_entity(se
) {
1985 cfs_rq
= cfs_rq_of(se
);
1986 put_prev_entity(cfs_rq
, se
);
1991 * sched_yield() is very simple
1993 * The magic of dealing with the ->skip buddy is in pick_next_entity.
1995 static void yield_task_fair(struct rq
*rq
)
1997 struct task_struct
*curr
= rq
->curr
;
1998 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
1999 struct sched_entity
*se
= &curr
->se
;
2002 * Are we the only task in the tree?
2004 if (unlikely(rq
->nr_running
== 1))
2007 clear_buddies(cfs_rq
, se
);
2009 if (curr
->policy
!= SCHED_BATCH
) {
2010 update_rq_clock(rq
);
2012 * Update run-time statistics of the 'current'.
2014 update_curr(cfs_rq
);
2020 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
2022 struct sched_entity
*se
= &p
->se
;
2027 /* Tell the scheduler that we'd really like pse to run next. */
2030 yield_task_fair(rq
);
2036 /**************************************************
2037 * Fair scheduling class load-balancing methods:
2041 * pull_task - move a task from a remote runqueue to the local runqueue.
2042 * Both runqueues must be locked.
2044 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2045 struct rq
*this_rq
, int this_cpu
)
2047 deactivate_task(src_rq
, p
, 0);
2048 set_task_cpu(p
, this_cpu
);
2049 activate_task(this_rq
, p
, 0);
2050 check_preempt_curr(this_rq
, p
, 0);
2054 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2057 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2058 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2061 int tsk_cache_hot
= 0;
2063 * We do not migrate tasks that are:
2064 * 1) running (obviously), or
2065 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2066 * 3) are cache-hot on their current CPU.
2068 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
2069 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
2074 if (task_running(rq
, p
)) {
2075 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
2080 * Aggressive migration if:
2081 * 1) task is cache cold, or
2082 * 2) too many balance attempts have failed.
2085 tsk_cache_hot
= task_hot(p
, rq
->clock_task
, sd
);
2086 if (!tsk_cache_hot
||
2087 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2088 #ifdef CONFIG_SCHEDSTATS
2089 if (tsk_cache_hot
) {
2090 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2091 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
2097 if (tsk_cache_hot
) {
2098 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
2105 * move_one_task tries to move exactly one task from busiest to this_rq, as
2106 * part of active balancing operations within "domain".
2107 * Returns 1 if successful and 0 otherwise.
2109 * Called with both runqueues locked.
2112 move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2113 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2115 struct task_struct
*p
, *n
;
2116 struct cfs_rq
*cfs_rq
;
2119 for_each_leaf_cfs_rq(busiest
, cfs_rq
) {
2120 list_for_each_entry_safe(p
, n
, &cfs_rq
->tasks
, se
.group_node
) {
2122 if (!can_migrate_task(p
, busiest
, this_cpu
,
2126 pull_task(busiest
, p
, this_rq
, this_cpu
);
2128 * Right now, this is only the second place pull_task()
2129 * is called, so we can safely collect pull_task()
2130 * stats here rather than inside pull_task().
2132 schedstat_inc(sd
, lb_gained
[idle
]);
2140 static unsigned long
2141 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2142 unsigned long max_load_move
, struct sched_domain
*sd
,
2143 enum cpu_idle_type idle
, int *all_pinned
,
2144 struct cfs_rq
*busiest_cfs_rq
)
2146 int loops
= 0, pulled
= 0;
2147 long rem_load_move
= max_load_move
;
2148 struct task_struct
*p
, *n
;
2150 if (max_load_move
== 0)
2153 list_for_each_entry_safe(p
, n
, &busiest_cfs_rq
->tasks
, se
.group_node
) {
2154 if (loops
++ > sysctl_sched_nr_migrate
)
2157 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
2158 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
,
2162 pull_task(busiest
, p
, this_rq
, this_cpu
);
2164 rem_load_move
-= p
->se
.load
.weight
;
2166 #ifdef CONFIG_PREEMPT
2168 * NEWIDLE balancing is a source of latency, so preemptible
2169 * kernels will stop after the first task is pulled to minimize
2170 * the critical section.
2172 if (idle
== CPU_NEWLY_IDLE
)
2177 * We only want to steal up to the prescribed amount of
2180 if (rem_load_move
<= 0)
2185 * Right now, this is one of only two places pull_task() is called,
2186 * so we can safely collect pull_task() stats here rather than
2187 * inside pull_task().
2189 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2191 return max_load_move
- rem_load_move
;
2194 #ifdef CONFIG_FAIR_GROUP_SCHED
2196 * update tg->load_weight by folding this cpu's load_avg
2198 static int update_shares_cpu(struct task_group
*tg
, int cpu
)
2200 struct cfs_rq
*cfs_rq
;
2201 unsigned long flags
;
2208 cfs_rq
= tg
->cfs_rq
[cpu
];
2210 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2212 update_rq_clock(rq
);
2213 update_cfs_load(cfs_rq
, 1);
2216 * We need to update shares after updating tg->load_weight in
2217 * order to adjust the weight of groups with long running tasks.
2219 update_cfs_shares(cfs_rq
);
2221 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2226 static void update_shares(int cpu
)
2228 struct cfs_rq
*cfs_rq
;
2229 struct rq
*rq
= cpu_rq(cpu
);
2232 for_each_leaf_cfs_rq(rq
, cfs_rq
)
2233 update_shares_cpu(cfs_rq
->tg
, cpu
);
2237 static unsigned long
2238 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2239 unsigned long max_load_move
,
2240 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2243 long rem_load_move
= max_load_move
;
2244 int busiest_cpu
= cpu_of(busiest
);
2245 struct task_group
*tg
;
2248 update_h_load(busiest_cpu
);
2250 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
2251 struct cfs_rq
*busiest_cfs_rq
= tg
->cfs_rq
[busiest_cpu
];
2252 unsigned long busiest_h_load
= busiest_cfs_rq
->h_load
;
2253 unsigned long busiest_weight
= busiest_cfs_rq
->load
.weight
;
2254 u64 rem_load
, moved_load
;
2259 if (!busiest_cfs_rq
->task_weight
)
2262 rem_load
= (u64
)rem_load_move
* busiest_weight
;
2263 rem_load
= div_u64(rem_load
, busiest_h_load
+ 1);
2265 moved_load
= balance_tasks(this_rq
, this_cpu
, busiest
,
2266 rem_load
, sd
, idle
, all_pinned
,
2272 moved_load
*= busiest_h_load
;
2273 moved_load
= div_u64(moved_load
, busiest_weight
+ 1);
2275 rem_load_move
-= moved_load
;
2276 if (rem_load_move
< 0)
2281 return max_load_move
- rem_load_move
;
2284 static inline void update_shares(int cpu
)
2288 static unsigned long
2289 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2290 unsigned long max_load_move
,
2291 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2294 return balance_tasks(this_rq
, this_cpu
, busiest
,
2295 max_load_move
, sd
, idle
, all_pinned
,
2301 * move_tasks tries to move up to max_load_move weighted load from busiest to
2302 * this_rq, as part of a balancing operation within domain "sd".
2303 * Returns 1 if successful and 0 otherwise.
2305 * Called with both runqueues locked.
2307 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2308 unsigned long max_load_move
,
2309 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2312 unsigned long total_load_moved
= 0, load_moved
;
2315 load_moved
= load_balance_fair(this_rq
, this_cpu
, busiest
,
2316 max_load_move
- total_load_moved
,
2317 sd
, idle
, all_pinned
);
2319 total_load_moved
+= load_moved
;
2321 #ifdef CONFIG_PREEMPT
2323 * NEWIDLE balancing is a source of latency, so preemptible
2324 * kernels will stop after the first task is pulled to minimize
2325 * the critical section.
2327 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
2330 if (raw_spin_is_contended(&this_rq
->lock
) ||
2331 raw_spin_is_contended(&busiest
->lock
))
2334 } while (load_moved
&& max_load_move
> total_load_moved
);
2336 return total_load_moved
> 0;
2339 /********** Helpers for find_busiest_group ************************/
2341 * sd_lb_stats - Structure to store the statistics of a sched_domain
2342 * during load balancing.
2344 struct sd_lb_stats
{
2345 struct sched_group
*busiest
; /* Busiest group in this sd */
2346 struct sched_group
*this; /* Local group in this sd */
2347 unsigned long total_load
; /* Total load of all groups in sd */
2348 unsigned long total_pwr
; /* Total power of all groups in sd */
2349 unsigned long avg_load
; /* Average load across all groups in sd */
2351 /** Statistics of this group */
2352 unsigned long this_load
;
2353 unsigned long this_load_per_task
;
2354 unsigned long this_nr_running
;
2355 unsigned long this_has_capacity
;
2356 unsigned int this_idle_cpus
;
2358 /* Statistics of the busiest group */
2359 unsigned int busiest_idle_cpus
;
2360 unsigned long max_load
;
2361 unsigned long busiest_load_per_task
;
2362 unsigned long busiest_nr_running
;
2363 unsigned long busiest_group_capacity
;
2364 unsigned long busiest_has_capacity
;
2365 unsigned int busiest_group_weight
;
2367 int group_imb
; /* Is there imbalance in this sd */
2368 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2369 int power_savings_balance
; /* Is powersave balance needed for this sd */
2370 struct sched_group
*group_min
; /* Least loaded group in sd */
2371 struct sched_group
*group_leader
; /* Group which relieves group_min */
2372 unsigned long min_load_per_task
; /* load_per_task in group_min */
2373 unsigned long leader_nr_running
; /* Nr running of group_leader */
2374 unsigned long min_nr_running
; /* Nr running of group_min */
2379 * sg_lb_stats - stats of a sched_group required for load_balancing
2381 struct sg_lb_stats
{
2382 unsigned long avg_load
; /*Avg load across the CPUs of the group */
2383 unsigned long group_load
; /* Total load over the CPUs of the group */
2384 unsigned long sum_nr_running
; /* Nr tasks running in the group */
2385 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
2386 unsigned long group_capacity
;
2387 unsigned long idle_cpus
;
2388 unsigned long group_weight
;
2389 int group_imb
; /* Is there an imbalance in the group ? */
2390 int group_has_capacity
; /* Is there extra capacity in the group? */
2394 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2395 * @group: The group whose first cpu is to be returned.
2397 static inline unsigned int group_first_cpu(struct sched_group
*group
)
2399 return cpumask_first(sched_group_cpus(group
));
2403 * get_sd_load_idx - Obtain the load index for a given sched domain.
2404 * @sd: The sched_domain whose load_idx is to be obtained.
2405 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2407 static inline int get_sd_load_idx(struct sched_domain
*sd
,
2408 enum cpu_idle_type idle
)
2414 load_idx
= sd
->busy_idx
;
2417 case CPU_NEWLY_IDLE
:
2418 load_idx
= sd
->newidle_idx
;
2421 load_idx
= sd
->idle_idx
;
2429 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2431 * init_sd_power_savings_stats - Initialize power savings statistics for
2432 * the given sched_domain, during load balancing.
2434 * @sd: Sched domain whose power-savings statistics are to be initialized.
2435 * @sds: Variable containing the statistics for sd.
2436 * @idle: Idle status of the CPU at which we're performing load-balancing.
2438 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
2439 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
2442 * Busy processors will not participate in power savings
2445 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2446 sds
->power_savings_balance
= 0;
2448 sds
->power_savings_balance
= 1;
2449 sds
->min_nr_running
= ULONG_MAX
;
2450 sds
->leader_nr_running
= 0;
2455 * update_sd_power_savings_stats - Update the power saving stats for a
2456 * sched_domain while performing load balancing.
2458 * @group: sched_group belonging to the sched_domain under consideration.
2459 * @sds: Variable containing the statistics of the sched_domain
2460 * @local_group: Does group contain the CPU for which we're performing
2462 * @sgs: Variable containing the statistics of the group.
2464 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
2465 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
2468 if (!sds
->power_savings_balance
)
2472 * If the local group is idle or completely loaded
2473 * no need to do power savings balance at this domain
2475 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
2476 !sds
->this_nr_running
))
2477 sds
->power_savings_balance
= 0;
2480 * If a group is already running at full capacity or idle,
2481 * don't include that group in power savings calculations
2483 if (!sds
->power_savings_balance
||
2484 sgs
->sum_nr_running
>= sgs
->group_capacity
||
2485 !sgs
->sum_nr_running
)
2489 * Calculate the group which has the least non-idle load.
2490 * This is the group from where we need to pick up the load
2493 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
2494 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
2495 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
2496 sds
->group_min
= group
;
2497 sds
->min_nr_running
= sgs
->sum_nr_running
;
2498 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
2499 sgs
->sum_nr_running
;
2503 * Calculate the group which is almost near its
2504 * capacity but still has some space to pick up some load
2505 * from other group and save more power
2507 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
2510 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
2511 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
2512 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
2513 sds
->group_leader
= group
;
2514 sds
->leader_nr_running
= sgs
->sum_nr_running
;
2519 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2520 * @sds: Variable containing the statistics of the sched_domain
2521 * under consideration.
2522 * @this_cpu: Cpu at which we're currently performing load-balancing.
2523 * @imbalance: Variable to store the imbalance.
2526 * Check if we have potential to perform some power-savings balance.
2527 * If yes, set the busiest group to be the least loaded group in the
2528 * sched_domain, so that it's CPUs can be put to idle.
2530 * Returns 1 if there is potential to perform power-savings balance.
2533 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
2534 int this_cpu
, unsigned long *imbalance
)
2536 if (!sds
->power_savings_balance
)
2539 if (sds
->this != sds
->group_leader
||
2540 sds
->group_leader
== sds
->group_min
)
2543 *imbalance
= sds
->min_load_per_task
;
2544 sds
->busiest
= sds
->group_min
;
2549 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2550 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
2551 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
2556 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
2557 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
2562 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
2563 int this_cpu
, unsigned long *imbalance
)
2567 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2570 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
2572 return SCHED_POWER_SCALE
;
2575 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
2577 return default_scale_freq_power(sd
, cpu
);
2580 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
2582 unsigned long weight
= sd
->span_weight
;
2583 unsigned long smt_gain
= sd
->smt_gain
;
2590 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
2592 return default_scale_smt_power(sd
, cpu
);
2595 unsigned long scale_rt_power(int cpu
)
2597 struct rq
*rq
= cpu_rq(cpu
);
2598 u64 total
, available
;
2600 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
2602 if (unlikely(total
< rq
->rt_avg
)) {
2603 /* Ensures that power won't end up being negative */
2606 available
= total
- rq
->rt_avg
;
2609 if (unlikely((s64
)total
< SCHED_POWER_SCALE
))
2610 total
= SCHED_POWER_SCALE
;
2612 total
>>= SCHED_POWER_SHIFT
;
2614 return div_u64(available
, total
);
2617 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
2619 unsigned long weight
= sd
->span_weight
;
2620 unsigned long power
= SCHED_POWER_SCALE
;
2621 struct sched_group
*sdg
= sd
->groups
;
2623 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
2624 if (sched_feat(ARCH_POWER
))
2625 power
*= arch_scale_smt_power(sd
, cpu
);
2627 power
*= default_scale_smt_power(sd
, cpu
);
2629 power
>>= SCHED_POWER_SHIFT
;
2632 sdg
->sgp
->power_orig
= power
;
2634 if (sched_feat(ARCH_POWER
))
2635 power
*= arch_scale_freq_power(sd
, cpu
);
2637 power
*= default_scale_freq_power(sd
, cpu
);
2639 power
>>= SCHED_POWER_SHIFT
;
2641 power
*= scale_rt_power(cpu
);
2642 power
>>= SCHED_POWER_SHIFT
;
2647 cpu_rq(cpu
)->cpu_power
= power
;
2648 sdg
->sgp
->power
= power
;
2651 static void update_group_power(struct sched_domain
*sd
, int cpu
)
2653 struct sched_domain
*child
= sd
->child
;
2654 struct sched_group
*group
, *sdg
= sd
->groups
;
2655 unsigned long power
;
2658 update_cpu_power(sd
, cpu
);
2664 group
= child
->groups
;
2666 power
+= group
->sgp
->power
;
2667 group
= group
->next
;
2668 } while (group
!= child
->groups
);
2670 sdg
->sgp
->power
= power
;
2674 * Try and fix up capacity for tiny siblings, this is needed when
2675 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2676 * which on its own isn't powerful enough.
2678 * See update_sd_pick_busiest() and check_asym_packing().
2681 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
2684 * Only siblings can have significantly less than SCHED_POWER_SCALE
2686 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
2690 * If ~90% of the cpu_power is still there, we're good.
2692 if (group
->sgp
->power
* 32 > group
->sgp
->power_orig
* 29)
2699 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2700 * @sd: The sched_domain whose statistics are to be updated.
2701 * @group: sched_group whose statistics are to be updated.
2702 * @this_cpu: Cpu for which load balance is currently performed.
2703 * @idle: Idle status of this_cpu
2704 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2705 * @local_group: Does group contain this_cpu.
2706 * @cpus: Set of cpus considered for load balancing.
2707 * @balance: Should we balance.
2708 * @sgs: variable to hold the statistics for this group.
2710 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
2711 struct sched_group
*group
, int this_cpu
,
2712 enum cpu_idle_type idle
, int load_idx
,
2713 int local_group
, const struct cpumask
*cpus
,
2714 int *balance
, struct sg_lb_stats
*sgs
)
2716 unsigned long load
, max_cpu_load
, min_cpu_load
, max_nr_running
;
2718 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2719 unsigned long avg_load_per_task
= 0;
2722 balance_cpu
= group_first_cpu(group
);
2724 /* Tally up the load of all CPUs in the group */
2726 min_cpu_load
= ~0UL;
2729 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
2730 struct rq
*rq
= cpu_rq(i
);
2732 /* Bias balancing toward cpus of our domain */
2734 if (idle_cpu(i
) && !first_idle_cpu
) {
2739 load
= target_load(i
, load_idx
);
2741 load
= source_load(i
, load_idx
);
2742 if (load
> max_cpu_load
) {
2743 max_cpu_load
= load
;
2744 max_nr_running
= rq
->nr_running
;
2746 if (min_cpu_load
> load
)
2747 min_cpu_load
= load
;
2750 sgs
->group_load
+= load
;
2751 sgs
->sum_nr_running
+= rq
->nr_running
;
2752 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
2758 * First idle cpu or the first cpu(busiest) in this sched group
2759 * is eligible for doing load balancing at this and above
2760 * domains. In the newly idle case, we will allow all the cpu's
2761 * to do the newly idle load balance.
2763 if (idle
!= CPU_NEWLY_IDLE
&& local_group
) {
2764 if (balance_cpu
!= this_cpu
) {
2768 update_group_power(sd
, this_cpu
);
2771 /* Adjust by relative CPU power of the group */
2772 sgs
->avg_load
= (sgs
->group_load
*SCHED_POWER_SCALE
) / group
->sgp
->power
;
2775 * Consider the group unbalanced when the imbalance is larger
2776 * than the average weight of a task.
2778 * APZ: with cgroup the avg task weight can vary wildly and
2779 * might not be a suitable number - should we keep a
2780 * normalized nr_running number somewhere that negates
2783 if (sgs
->sum_nr_running
)
2784 avg_load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
2786 if ((max_cpu_load
- min_cpu_load
) >= avg_load_per_task
&& max_nr_running
> 1)
2789 sgs
->group_capacity
= DIV_ROUND_CLOSEST(group
->sgp
->power
,
2791 if (!sgs
->group_capacity
)
2792 sgs
->group_capacity
= fix_small_capacity(sd
, group
);
2793 sgs
->group_weight
= group
->group_weight
;
2795 if (sgs
->group_capacity
> sgs
->sum_nr_running
)
2796 sgs
->group_has_capacity
= 1;
2800 * update_sd_pick_busiest - return 1 on busiest group
2801 * @sd: sched_domain whose statistics are to be checked
2802 * @sds: sched_domain statistics
2803 * @sg: sched_group candidate to be checked for being the busiest
2804 * @sgs: sched_group statistics
2805 * @this_cpu: the current cpu
2807 * Determine if @sg is a busier group than the previously selected
2810 static bool update_sd_pick_busiest(struct sched_domain
*sd
,
2811 struct sd_lb_stats
*sds
,
2812 struct sched_group
*sg
,
2813 struct sg_lb_stats
*sgs
,
2816 if (sgs
->avg_load
<= sds
->max_load
)
2819 if (sgs
->sum_nr_running
> sgs
->group_capacity
)
2826 * ASYM_PACKING needs to move all the work to the lowest
2827 * numbered CPUs in the group, therefore mark all groups
2828 * higher than ourself as busy.
2830 if ((sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
2831 this_cpu
< group_first_cpu(sg
)) {
2835 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
2843 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2844 * @sd: sched_domain whose statistics are to be updated.
2845 * @this_cpu: Cpu for which load balance is currently performed.
2846 * @idle: Idle status of this_cpu
2847 * @cpus: Set of cpus considered for load balancing.
2848 * @balance: Should we balance.
2849 * @sds: variable to hold the statistics for this sched_domain.
2851 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
2852 enum cpu_idle_type idle
, const struct cpumask
*cpus
,
2853 int *balance
, struct sd_lb_stats
*sds
)
2855 struct sched_domain
*child
= sd
->child
;
2856 struct sched_group
*sg
= sd
->groups
;
2857 struct sg_lb_stats sgs
;
2858 int load_idx
, prefer_sibling
= 0;
2860 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
2863 init_sd_power_savings_stats(sd
, sds
, idle
);
2864 load_idx
= get_sd_load_idx(sd
, idle
);
2869 local_group
= cpumask_test_cpu(this_cpu
, sched_group_cpus(sg
));
2870 memset(&sgs
, 0, sizeof(sgs
));
2871 update_sg_lb_stats(sd
, sg
, this_cpu
, idle
, load_idx
,
2872 local_group
, cpus
, balance
, &sgs
);
2874 if (local_group
&& !(*balance
))
2877 sds
->total_load
+= sgs
.group_load
;
2878 sds
->total_pwr
+= sg
->sgp
->power
;
2881 * In case the child domain prefers tasks go to siblings
2882 * first, lower the sg capacity to one so that we'll try
2883 * and move all the excess tasks away. We lower the capacity
2884 * of a group only if the local group has the capacity to fit
2885 * these excess tasks, i.e. nr_running < group_capacity. The
2886 * extra check prevents the case where you always pull from the
2887 * heaviest group when it is already under-utilized (possible
2888 * with a large weight task outweighs the tasks on the system).
2890 if (prefer_sibling
&& !local_group
&& sds
->this_has_capacity
)
2891 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
2894 sds
->this_load
= sgs
.avg_load
;
2896 sds
->this_nr_running
= sgs
.sum_nr_running
;
2897 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
2898 sds
->this_has_capacity
= sgs
.group_has_capacity
;
2899 sds
->this_idle_cpus
= sgs
.idle_cpus
;
2900 } else if (update_sd_pick_busiest(sd
, sds
, sg
, &sgs
, this_cpu
)) {
2901 sds
->max_load
= sgs
.avg_load
;
2903 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
2904 sds
->busiest_idle_cpus
= sgs
.idle_cpus
;
2905 sds
->busiest_group_capacity
= sgs
.group_capacity
;
2906 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
2907 sds
->busiest_has_capacity
= sgs
.group_has_capacity
;
2908 sds
->busiest_group_weight
= sgs
.group_weight
;
2909 sds
->group_imb
= sgs
.group_imb
;
2912 update_sd_power_savings_stats(sg
, sds
, local_group
, &sgs
);
2914 } while (sg
!= sd
->groups
);
2917 int __weak
arch_sd_sibling_asym_packing(void)
2919 return 0*SD_ASYM_PACKING
;
2923 * check_asym_packing - Check to see if the group is packed into the
2926 * This is primarily intended to used at the sibling level. Some
2927 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2928 * case of POWER7, it can move to lower SMT modes only when higher
2929 * threads are idle. When in lower SMT modes, the threads will
2930 * perform better since they share less core resources. Hence when we
2931 * have idle threads, we want them to be the higher ones.
2933 * This packing function is run on idle threads. It checks to see if
2934 * the busiest CPU in this domain (core in the P7 case) has a higher
2935 * CPU number than the packing function is being run on. Here we are
2936 * assuming lower CPU number will be equivalent to lower a SMT thread
2939 * Returns 1 when packing is required and a task should be moved to
2940 * this CPU. The amount of the imbalance is returned in *imbalance.
2942 * @sd: The sched_domain whose packing is to be checked.
2943 * @sds: Statistics of the sched_domain which is to be packed
2944 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2945 * @imbalance: returns amount of imbalanced due to packing.
2947 static int check_asym_packing(struct sched_domain
*sd
,
2948 struct sd_lb_stats
*sds
,
2949 int this_cpu
, unsigned long *imbalance
)
2953 if (!(sd
->flags
& SD_ASYM_PACKING
))
2959 busiest_cpu
= group_first_cpu(sds
->busiest
);
2960 if (this_cpu
> busiest_cpu
)
2963 *imbalance
= DIV_ROUND_CLOSEST(sds
->max_load
* sds
->busiest
->sgp
->power
,
2969 * fix_small_imbalance - Calculate the minor imbalance that exists
2970 * amongst the groups of a sched_domain, during
2972 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2973 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2974 * @imbalance: Variable to store the imbalance.
2976 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
2977 int this_cpu
, unsigned long *imbalance
)
2979 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
2980 unsigned int imbn
= 2;
2981 unsigned long scaled_busy_load_per_task
;
2983 if (sds
->this_nr_running
) {
2984 sds
->this_load_per_task
/= sds
->this_nr_running
;
2985 if (sds
->busiest_load_per_task
>
2986 sds
->this_load_per_task
)
2989 sds
->this_load_per_task
=
2990 cpu_avg_load_per_task(this_cpu
);
2992 scaled_busy_load_per_task
= sds
->busiest_load_per_task
2993 * SCHED_POWER_SCALE
;
2994 scaled_busy_load_per_task
/= sds
->busiest
->sgp
->power
;
2996 if (sds
->max_load
- sds
->this_load
+ scaled_busy_load_per_task
>=
2997 (scaled_busy_load_per_task
* imbn
)) {
2998 *imbalance
= sds
->busiest_load_per_task
;
3003 * OK, we don't have enough imbalance to justify moving tasks,
3004 * however we may be able to increase total CPU power used by
3008 pwr_now
+= sds
->busiest
->sgp
->power
*
3009 min(sds
->busiest_load_per_task
, sds
->max_load
);
3010 pwr_now
+= sds
->this->sgp
->power
*
3011 min(sds
->this_load_per_task
, sds
->this_load
);
3012 pwr_now
/= SCHED_POWER_SCALE
;
3014 /* Amount of load we'd subtract */
3015 tmp
= (sds
->busiest_load_per_task
* SCHED_POWER_SCALE
) /
3016 sds
->busiest
->sgp
->power
;
3017 if (sds
->max_load
> tmp
)
3018 pwr_move
+= sds
->busiest
->sgp
->power
*
3019 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3021 /* Amount of load we'd add */
3022 if (sds
->max_load
* sds
->busiest
->sgp
->power
<
3023 sds
->busiest_load_per_task
* SCHED_POWER_SCALE
)
3024 tmp
= (sds
->max_load
* sds
->busiest
->sgp
->power
) /
3025 sds
->this->sgp
->power
;
3027 tmp
= (sds
->busiest_load_per_task
* SCHED_POWER_SCALE
) /
3028 sds
->this->sgp
->power
;
3029 pwr_move
+= sds
->this->sgp
->power
*
3030 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3031 pwr_move
/= SCHED_POWER_SCALE
;
3033 /* Move if we gain throughput */
3034 if (pwr_move
> pwr_now
)
3035 *imbalance
= sds
->busiest_load_per_task
;
3039 * calculate_imbalance - Calculate the amount of imbalance present within the
3040 * groups of a given sched_domain during load balance.
3041 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3042 * @this_cpu: Cpu for which currently load balance is being performed.
3043 * @imbalance: The variable to store the imbalance.
3045 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3046 unsigned long *imbalance
)
3048 unsigned long max_pull
, load_above_capacity
= ~0UL;
3050 sds
->busiest_load_per_task
/= sds
->busiest_nr_running
;
3051 if (sds
->group_imb
) {
3052 sds
->busiest_load_per_task
=
3053 min(sds
->busiest_load_per_task
, sds
->avg_load
);
3057 * In the presence of smp nice balancing, certain scenarios can have
3058 * max load less than avg load(as we skip the groups at or below
3059 * its cpu_power, while calculating max_load..)
3061 if (sds
->max_load
< sds
->avg_load
) {
3063 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3066 if (!sds
->group_imb
) {
3068 * Don't want to pull so many tasks that a group would go idle.
3070 load_above_capacity
= (sds
->busiest_nr_running
-
3071 sds
->busiest_group_capacity
);
3073 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_POWER_SCALE
);
3075 load_above_capacity
/= sds
->busiest
->sgp
->power
;
3079 * We're trying to get all the cpus to the average_load, so we don't
3080 * want to push ourselves above the average load, nor do we wish to
3081 * reduce the max loaded cpu below the average load. At the same time,
3082 * we also don't want to reduce the group load below the group capacity
3083 * (so that we can implement power-savings policies etc). Thus we look
3084 * for the minimum possible imbalance.
3085 * Be careful of negative numbers as they'll appear as very large values
3086 * with unsigned longs.
3088 max_pull
= min(sds
->max_load
- sds
->avg_load
, load_above_capacity
);
3090 /* How much load to actually move to equalise the imbalance */
3091 *imbalance
= min(max_pull
* sds
->busiest
->sgp
->power
,
3092 (sds
->avg_load
- sds
->this_load
) * sds
->this->sgp
->power
)
3093 / SCHED_POWER_SCALE
;
3096 * if *imbalance is less than the average load per runnable task
3097 * there is no guarantee that any tasks will be moved so we'll have
3098 * a think about bumping its value to force at least one task to be
3101 if (*imbalance
< sds
->busiest_load_per_task
)
3102 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3106 /******* find_busiest_group() helpers end here *********************/
3109 * find_busiest_group - Returns the busiest group within the sched_domain
3110 * if there is an imbalance. If there isn't an imbalance, and
3111 * the user has opted for power-savings, it returns a group whose
3112 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3113 * such a group exists.
3115 * Also calculates the amount of weighted load which should be moved
3116 * to restore balance.
3118 * @sd: The sched_domain whose busiest group is to be returned.
3119 * @this_cpu: The cpu for which load balancing is currently being performed.
3120 * @imbalance: Variable which stores amount of weighted load which should
3121 * be moved to restore balance/put a group to idle.
3122 * @idle: The idle status of this_cpu.
3123 * @cpus: The set of CPUs under consideration for load-balancing.
3124 * @balance: Pointer to a variable indicating if this_cpu
3125 * is the appropriate cpu to perform load balancing at this_level.
3127 * Returns: - the busiest group if imbalance exists.
3128 * - If no imbalance and user has opted for power-savings balance,
3129 * return the least loaded group whose CPUs can be
3130 * put to idle by rebalancing its tasks onto our group.
3132 static struct sched_group
*
3133 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3134 unsigned long *imbalance
, enum cpu_idle_type idle
,
3135 const struct cpumask
*cpus
, int *balance
)
3137 struct sd_lb_stats sds
;
3139 memset(&sds
, 0, sizeof(sds
));
3142 * Compute the various statistics relavent for load balancing at
3145 update_sd_lb_stats(sd
, this_cpu
, idle
, cpus
, balance
, &sds
);
3148 * this_cpu is not the appropriate cpu to perform load balancing at
3154 if ((idle
== CPU_IDLE
|| idle
== CPU_NEWLY_IDLE
) &&
3155 check_asym_packing(sd
, &sds
, this_cpu
, imbalance
))
3158 /* There is no busy sibling group to pull tasks from */
3159 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
3162 sds
.avg_load
= (SCHED_POWER_SCALE
* sds
.total_load
) / sds
.total_pwr
;
3165 * If the busiest group is imbalanced the below checks don't
3166 * work because they assumes all things are equal, which typically
3167 * isn't true due to cpus_allowed constraints and the like.
3172 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3173 if (idle
== CPU_NEWLY_IDLE
&& sds
.this_has_capacity
&&
3174 !sds
.busiest_has_capacity
)
3178 * If the local group is more busy than the selected busiest group
3179 * don't try and pull any tasks.
3181 if (sds
.this_load
>= sds
.max_load
)
3185 * Don't pull any tasks if this group is already above the domain
3188 if (sds
.this_load
>= sds
.avg_load
)
3191 if (idle
== CPU_IDLE
) {
3193 * This cpu is idle. If the busiest group load doesn't
3194 * have more tasks than the number of available cpu's and
3195 * there is no imbalance between this and busiest group
3196 * wrt to idle cpu's, it is balanced.
3198 if ((sds
.this_idle_cpus
<= sds
.busiest_idle_cpus
+ 1) &&
3199 sds
.busiest_nr_running
<= sds
.busiest_group_weight
)
3203 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
3204 * imbalance_pct to be conservative.
3206 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
3211 /* Looks like there is an imbalance. Compute it */
3212 calculate_imbalance(&sds
, this_cpu
, imbalance
);
3217 * There is no obvious imbalance. But check if we can do some balancing
3220 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
3228 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3231 find_busiest_queue(struct sched_domain
*sd
, struct sched_group
*group
,
3232 enum cpu_idle_type idle
, unsigned long imbalance
,
3233 const struct cpumask
*cpus
)
3235 struct rq
*busiest
= NULL
, *rq
;
3236 unsigned long max_load
= 0;
3239 for_each_cpu(i
, sched_group_cpus(group
)) {
3240 unsigned long power
= power_of(i
);
3241 unsigned long capacity
= DIV_ROUND_CLOSEST(power
,
3246 capacity
= fix_small_capacity(sd
, group
);
3248 if (!cpumask_test_cpu(i
, cpus
))
3252 wl
= weighted_cpuload(i
);
3255 * When comparing with imbalance, use weighted_cpuload()
3256 * which is not scaled with the cpu power.
3258 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
3262 * For the load comparisons with the other cpu's, consider
3263 * the weighted_cpuload() scaled with the cpu power, so that
3264 * the load can be moved away from the cpu that is potentially
3265 * running at a lower capacity.
3267 wl
= (wl
* SCHED_POWER_SCALE
) / power
;
3269 if (wl
> max_load
) {
3279 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3280 * so long as it is large enough.
3282 #define MAX_PINNED_INTERVAL 512
3284 /* Working cpumask for load_balance and load_balance_newidle. */
3285 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
3287 static int need_active_balance(struct sched_domain
*sd
, int idle
,
3288 int busiest_cpu
, int this_cpu
)
3290 if (idle
== CPU_NEWLY_IDLE
) {
3293 * ASYM_PACKING needs to force migrate tasks from busy but
3294 * higher numbered CPUs in order to pack all tasks in the
3295 * lowest numbered CPUs.
3297 if ((sd
->flags
& SD_ASYM_PACKING
) && busiest_cpu
> this_cpu
)
3301 * The only task running in a non-idle cpu can be moved to this
3302 * cpu in an attempt to completely freeup the other CPU
3305 * The package power saving logic comes from
3306 * find_busiest_group(). If there are no imbalance, then
3307 * f_b_g() will return NULL. However when sched_mc={1,2} then
3308 * f_b_g() will select a group from which a running task may be
3309 * pulled to this cpu in order to make the other package idle.
3310 * If there is no opportunity to make a package idle and if
3311 * there are no imbalance, then f_b_g() will return NULL and no
3312 * action will be taken in load_balance_newidle().
3314 * Under normal task pull operation due to imbalance, there
3315 * will be more than one task in the source run queue and
3316 * move_tasks() will succeed. ld_moved will be true and this
3317 * active balance code will not be triggered.
3319 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
3323 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
3326 static int active_load_balance_cpu_stop(void *data
);
3329 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3330 * tasks if there is an imbalance.
3332 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3333 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3336 int ld_moved
, all_pinned
= 0, active_balance
= 0;
3337 struct sched_group
*group
;
3338 unsigned long imbalance
;
3340 unsigned long flags
;
3341 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
3343 cpumask_copy(cpus
, cpu_active_mask
);
3345 schedstat_inc(sd
, lb_count
[idle
]);
3348 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
,
3355 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3359 busiest
= find_busiest_queue(sd
, group
, idle
, imbalance
, cpus
);
3361 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3365 BUG_ON(busiest
== this_rq
);
3367 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3370 if (busiest
->nr_running
> 1) {
3372 * Attempt to move tasks. If find_busiest_group has found
3373 * an imbalance but busiest->nr_running <= 1, the group is
3374 * still unbalanced. ld_moved simply stays zero, so it is
3375 * correctly treated as an imbalance.
3378 local_irq_save(flags
);
3379 double_rq_lock(this_rq
, busiest
);
3380 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3381 imbalance
, sd
, idle
, &all_pinned
);
3382 double_rq_unlock(this_rq
, busiest
);
3383 local_irq_restore(flags
);
3386 * some other cpu did the load balance for us.
3388 if (ld_moved
&& this_cpu
!= smp_processor_id())
3389 resched_cpu(this_cpu
);
3391 /* All tasks on this runqueue were pinned by CPU affinity */
3392 if (unlikely(all_pinned
)) {
3393 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3394 if (!cpumask_empty(cpus
))
3401 schedstat_inc(sd
, lb_failed
[idle
]);
3403 * Increment the failure counter only on periodic balance.
3404 * We do not want newidle balance, which can be very
3405 * frequent, pollute the failure counter causing
3406 * excessive cache_hot migrations and active balances.
3408 if (idle
!= CPU_NEWLY_IDLE
)
3409 sd
->nr_balance_failed
++;
3411 if (need_active_balance(sd
, idle
, cpu_of(busiest
), this_cpu
)) {
3412 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
3414 /* don't kick the active_load_balance_cpu_stop,
3415 * if the curr task on busiest cpu can't be
3418 if (!cpumask_test_cpu(this_cpu
,
3419 &busiest
->curr
->cpus_allowed
)) {
3420 raw_spin_unlock_irqrestore(&busiest
->lock
,
3423 goto out_one_pinned
;
3427 * ->active_balance synchronizes accesses to
3428 * ->active_balance_work. Once set, it's cleared
3429 * only after active load balance is finished.
3431 if (!busiest
->active_balance
) {
3432 busiest
->active_balance
= 1;
3433 busiest
->push_cpu
= this_cpu
;
3436 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
3439 stop_one_cpu_nowait(cpu_of(busiest
),
3440 active_load_balance_cpu_stop
, busiest
,
3441 &busiest
->active_balance_work
);
3444 * We've kicked active balancing, reset the failure
3447 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3450 sd
->nr_balance_failed
= 0;
3452 if (likely(!active_balance
)) {
3453 /* We were unbalanced, so reset the balancing interval */
3454 sd
->balance_interval
= sd
->min_interval
;
3457 * If we've begun active balancing, start to back off. This
3458 * case may not be covered by the all_pinned logic if there
3459 * is only 1 task on the busy runqueue (because we don't call
3462 if (sd
->balance_interval
< sd
->max_interval
)
3463 sd
->balance_interval
*= 2;
3469 schedstat_inc(sd
, lb_balanced
[idle
]);
3471 sd
->nr_balance_failed
= 0;
3474 /* tune up the balancing interval */
3475 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3476 (sd
->balance_interval
< sd
->max_interval
))
3477 sd
->balance_interval
*= 2;
3485 * idle_balance is called by schedule() if this_cpu is about to become
3486 * idle. Attempts to pull tasks from other CPUs.
3488 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3490 struct sched_domain
*sd
;
3491 int pulled_task
= 0;
3492 unsigned long next_balance
= jiffies
+ HZ
;
3494 this_rq
->idle_stamp
= this_rq
->clock
;
3496 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
3500 * Drop the rq->lock, but keep IRQ/preempt disabled.
3502 raw_spin_unlock(&this_rq
->lock
);
3504 update_shares(this_cpu
);
3506 for_each_domain(this_cpu
, sd
) {
3507 unsigned long interval
;
3510 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3513 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
3514 /* If we've pulled tasks over stop searching: */
3515 pulled_task
= load_balance(this_cpu
, this_rq
,
3516 sd
, CPU_NEWLY_IDLE
, &balance
);
3519 interval
= msecs_to_jiffies(sd
->balance_interval
);
3520 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3521 next_balance
= sd
->last_balance
+ interval
;
3523 this_rq
->idle_stamp
= 0;
3529 raw_spin_lock(&this_rq
->lock
);
3531 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3533 * We are going idle. next_balance may be set based on
3534 * a busy processor. So reset next_balance.
3536 this_rq
->next_balance
= next_balance
;
3541 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3542 * running tasks off the busiest CPU onto idle CPUs. It requires at
3543 * least 1 task to be running on each physical CPU where possible, and
3544 * avoids physical / logical imbalances.
3546 static int active_load_balance_cpu_stop(void *data
)
3548 struct rq
*busiest_rq
= data
;
3549 int busiest_cpu
= cpu_of(busiest_rq
);
3550 int target_cpu
= busiest_rq
->push_cpu
;
3551 struct rq
*target_rq
= cpu_rq(target_cpu
);
3552 struct sched_domain
*sd
;
3554 raw_spin_lock_irq(&busiest_rq
->lock
);
3556 /* make sure the requested cpu hasn't gone down in the meantime */
3557 if (unlikely(busiest_cpu
!= smp_processor_id() ||
3558 !busiest_rq
->active_balance
))
3561 /* Is there any task to move? */
3562 if (busiest_rq
->nr_running
<= 1)
3566 * This condition is "impossible", if it occurs
3567 * we need to fix it. Originally reported by
3568 * Bjorn Helgaas on a 128-cpu setup.
3570 BUG_ON(busiest_rq
== target_rq
);
3572 /* move a task from busiest_rq to target_rq */
3573 double_lock_balance(busiest_rq
, target_rq
);
3575 /* Search for an sd spanning us and the target CPU. */
3577 for_each_domain(target_cpu
, sd
) {
3578 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3579 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
3584 schedstat_inc(sd
, alb_count
);
3586 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3588 schedstat_inc(sd
, alb_pushed
);
3590 schedstat_inc(sd
, alb_failed
);
3593 double_unlock_balance(busiest_rq
, target_rq
);
3595 busiest_rq
->active_balance
= 0;
3596 raw_spin_unlock_irq(&busiest_rq
->lock
);
3602 static DEFINE_PER_CPU(struct call_single_data
, remote_sched_softirq_cb
);
3604 static void trigger_sched_softirq(void *data
)
3606 raise_softirq_irqoff(SCHED_SOFTIRQ
);
3609 static inline void init_sched_softirq_csd(struct call_single_data
*csd
)
3611 csd
->func
= trigger_sched_softirq
;
3618 * idle load balancing details
3619 * - One of the idle CPUs nominates itself as idle load_balancer, while
3621 * - This idle load balancer CPU will also go into tickless mode when
3622 * it is idle, just like all other idle CPUs
3623 * - When one of the busy CPUs notice that there may be an idle rebalancing
3624 * needed, they will kick the idle load balancer, which then does idle
3625 * load balancing for all the idle CPUs.
3628 atomic_t load_balancer
;
3629 atomic_t first_pick_cpu
;
3630 atomic_t second_pick_cpu
;
3631 cpumask_var_t idle_cpus_mask
;
3632 cpumask_var_t grp_idle_mask
;
3633 unsigned long next_balance
; /* in jiffy units */
3634 } nohz ____cacheline_aligned
;
3636 int get_nohz_load_balancer(void)
3638 return atomic_read(&nohz
.load_balancer
);
3641 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3643 * lowest_flag_domain - Return lowest sched_domain containing flag.
3644 * @cpu: The cpu whose lowest level of sched domain is to
3646 * @flag: The flag to check for the lowest sched_domain
3647 * for the given cpu.
3649 * Returns the lowest sched_domain of a cpu which contains the given flag.
3651 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
3653 struct sched_domain
*sd
;
3655 for_each_domain(cpu
, sd
)
3656 if (sd
&& (sd
->flags
& flag
))
3663 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3664 * @cpu: The cpu whose domains we're iterating over.
3665 * @sd: variable holding the value of the power_savings_sd
3667 * @flag: The flag to filter the sched_domains to be iterated.
3669 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3670 * set, starting from the lowest sched_domain to the highest.
3672 #define for_each_flag_domain(cpu, sd, flag) \
3673 for (sd = lowest_flag_domain(cpu, flag); \
3674 (sd && (sd->flags & flag)); sd = sd->parent)
3677 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3678 * @ilb_group: group to be checked for semi-idleness
3680 * Returns: 1 if the group is semi-idle. 0 otherwise.
3682 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3683 * and atleast one non-idle CPU. This helper function checks if the given
3684 * sched_group is semi-idle or not.
3686 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
3688 cpumask_and(nohz
.grp_idle_mask
, nohz
.idle_cpus_mask
,
3689 sched_group_cpus(ilb_group
));
3692 * A sched_group is semi-idle when it has atleast one busy cpu
3693 * and atleast one idle cpu.
3695 if (cpumask_empty(nohz
.grp_idle_mask
))
3698 if (cpumask_equal(nohz
.grp_idle_mask
, sched_group_cpus(ilb_group
)))
3704 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3705 * @cpu: The cpu which is nominating a new idle_load_balancer.
3707 * Returns: Returns the id of the idle load balancer if it exists,
3708 * Else, returns >= nr_cpu_ids.
3710 * This algorithm picks the idle load balancer such that it belongs to a
3711 * semi-idle powersavings sched_domain. The idea is to try and avoid
3712 * completely idle packages/cores just for the purpose of idle load balancing
3713 * when there are other idle cpu's which are better suited for that job.
3715 static int find_new_ilb(int cpu
)
3717 struct sched_domain
*sd
;
3718 struct sched_group
*ilb_group
;
3719 int ilb
= nr_cpu_ids
;
3722 * Have idle load balancer selection from semi-idle packages only
3723 * when power-aware load balancing is enabled
3725 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
3729 * Optimize for the case when we have no idle CPUs or only one
3730 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3732 if (cpumask_weight(nohz
.idle_cpus_mask
) < 2)
3736 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
3737 ilb_group
= sd
->groups
;
3740 if (is_semi_idle_group(ilb_group
)) {
3741 ilb
= cpumask_first(nohz
.grp_idle_mask
);
3745 ilb_group
= ilb_group
->next
;
3747 } while (ilb_group
!= sd
->groups
);
3755 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3756 static inline int find_new_ilb(int call_cpu
)
3763 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3764 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3765 * CPU (if there is one).
3767 static void nohz_balancer_kick(int cpu
)
3771 nohz
.next_balance
++;
3773 ilb_cpu
= get_nohz_load_balancer();
3775 if (ilb_cpu
>= nr_cpu_ids
) {
3776 ilb_cpu
= cpumask_first(nohz
.idle_cpus_mask
);
3777 if (ilb_cpu
>= nr_cpu_ids
)
3781 if (!cpu_rq(ilb_cpu
)->nohz_balance_kick
) {
3782 struct call_single_data
*cp
;
3784 cpu_rq(ilb_cpu
)->nohz_balance_kick
= 1;
3785 cp
= &per_cpu(remote_sched_softirq_cb
, cpu
);
3786 __smp_call_function_single(ilb_cpu
, cp
, 0);
3792 * This routine will try to nominate the ilb (idle load balancing)
3793 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3794 * load balancing on behalf of all those cpus.
3796 * When the ilb owner becomes busy, we will not have new ilb owner until some
3797 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3798 * idle load balancing by kicking one of the idle CPUs.
3800 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3801 * ilb owner CPU in future (when there is a need for idle load balancing on
3802 * behalf of all idle CPUs).
3804 void select_nohz_load_balancer(int stop_tick
)
3806 int cpu
= smp_processor_id();
3809 if (!cpu_active(cpu
)) {
3810 if (atomic_read(&nohz
.load_balancer
) != cpu
)
3814 * If we are going offline and still the leader,
3817 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
,
3824 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
3826 if (atomic_read(&nohz
.first_pick_cpu
) == cpu
)
3827 atomic_cmpxchg(&nohz
.first_pick_cpu
, cpu
, nr_cpu_ids
);
3828 if (atomic_read(&nohz
.second_pick_cpu
) == cpu
)
3829 atomic_cmpxchg(&nohz
.second_pick_cpu
, cpu
, nr_cpu_ids
);
3831 if (atomic_read(&nohz
.load_balancer
) >= nr_cpu_ids
) {
3834 /* make me the ilb owner */
3835 if (atomic_cmpxchg(&nohz
.load_balancer
, nr_cpu_ids
,
3840 * Check to see if there is a more power-efficient
3843 new_ilb
= find_new_ilb(cpu
);
3844 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
3845 atomic_set(&nohz
.load_balancer
, nr_cpu_ids
);
3846 resched_cpu(new_ilb
);
3852 if (!cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))
3855 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
3857 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3858 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
,
3866 static DEFINE_SPINLOCK(balancing
);
3868 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
3871 * Scale the max load_balance interval with the number of CPUs in the system.
3872 * This trades load-balance latency on larger machines for less cross talk.
3874 static void update_max_interval(void)
3876 max_load_balance_interval
= HZ
*num_online_cpus()/10;
3880 * It checks each scheduling domain to see if it is due to be balanced,
3881 * and initiates a balancing operation if so.
3883 * Balancing parameters are set up in arch_init_sched_domains.
3885 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3888 struct rq
*rq
= cpu_rq(cpu
);
3889 unsigned long interval
;
3890 struct sched_domain
*sd
;
3891 /* Earliest time when we have to do rebalance again */
3892 unsigned long next_balance
= jiffies
+ 60*HZ
;
3893 int update_next_balance
= 0;
3899 for_each_domain(cpu
, sd
) {
3900 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3903 interval
= sd
->balance_interval
;
3904 if (idle
!= CPU_IDLE
)
3905 interval
*= sd
->busy_factor
;
3907 /* scale ms to jiffies */
3908 interval
= msecs_to_jiffies(interval
);
3909 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
3911 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3913 if (need_serialize
) {
3914 if (!spin_trylock(&balancing
))
3918 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3919 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3921 * We've pulled tasks over so either we're no
3924 idle
= CPU_NOT_IDLE
;
3926 sd
->last_balance
= jiffies
;
3929 spin_unlock(&balancing
);
3931 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3932 next_balance
= sd
->last_balance
+ interval
;
3933 update_next_balance
= 1;
3937 * Stop the load balance at this level. There is another
3938 * CPU in our sched group which is doing load balancing more
3947 * next_balance will be updated only when there is a need.
3948 * When the cpu is attached to null domain for ex, it will not be
3951 if (likely(update_next_balance
))
3952 rq
->next_balance
= next_balance
;
3957 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3958 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3960 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
)
3962 struct rq
*this_rq
= cpu_rq(this_cpu
);
3966 if (idle
!= CPU_IDLE
|| !this_rq
->nohz_balance_kick
)
3969 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
3970 if (balance_cpu
== this_cpu
)
3974 * If this cpu gets work to do, stop the load balancing
3975 * work being done for other cpus. Next load
3976 * balancing owner will pick it up.
3978 if (need_resched()) {
3979 this_rq
->nohz_balance_kick
= 0;
3983 raw_spin_lock_irq(&this_rq
->lock
);
3984 update_rq_clock(this_rq
);
3985 update_cpu_load(this_rq
);
3986 raw_spin_unlock_irq(&this_rq
->lock
);
3988 rebalance_domains(balance_cpu
, CPU_IDLE
);
3990 rq
= cpu_rq(balance_cpu
);
3991 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3992 this_rq
->next_balance
= rq
->next_balance
;
3994 nohz
.next_balance
= this_rq
->next_balance
;
3995 this_rq
->nohz_balance_kick
= 0;
3999 * Current heuristic for kicking the idle load balancer
4000 * - first_pick_cpu is the one of the busy CPUs. It will kick
4001 * idle load balancer when it has more than one process active. This
4002 * eliminates the need for idle load balancing altogether when we have
4003 * only one running process in the system (common case).
4004 * - If there are more than one busy CPU, idle load balancer may have
4005 * to run for active_load_balance to happen (i.e., two busy CPUs are
4006 * SMT or core siblings and can run better if they move to different
4007 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
4008 * which will kick idle load balancer as soon as it has any load.
4010 static inline int nohz_kick_needed(struct rq
*rq
, int cpu
)
4012 unsigned long now
= jiffies
;
4014 int first_pick_cpu
, second_pick_cpu
;
4016 if (time_before(now
, nohz
.next_balance
))
4019 if (rq
->idle_at_tick
)
4022 first_pick_cpu
= atomic_read(&nohz
.first_pick_cpu
);
4023 second_pick_cpu
= atomic_read(&nohz
.second_pick_cpu
);
4025 if (first_pick_cpu
< nr_cpu_ids
&& first_pick_cpu
!= cpu
&&
4026 second_pick_cpu
< nr_cpu_ids
&& second_pick_cpu
!= cpu
)
4029 ret
= atomic_cmpxchg(&nohz
.first_pick_cpu
, nr_cpu_ids
, cpu
);
4030 if (ret
== nr_cpu_ids
|| ret
== cpu
) {
4031 atomic_cmpxchg(&nohz
.second_pick_cpu
, cpu
, nr_cpu_ids
);
4032 if (rq
->nr_running
> 1)
4035 ret
= atomic_cmpxchg(&nohz
.second_pick_cpu
, nr_cpu_ids
, cpu
);
4036 if (ret
== nr_cpu_ids
|| ret
== cpu
) {
4044 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
) { }
4048 * run_rebalance_domains is triggered when needed from the scheduler tick.
4049 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
4051 static void run_rebalance_domains(struct softirq_action
*h
)
4053 int this_cpu
= smp_processor_id();
4054 struct rq
*this_rq
= cpu_rq(this_cpu
);
4055 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4056 CPU_IDLE
: CPU_NOT_IDLE
;
4058 rebalance_domains(this_cpu
, idle
);
4061 * If this cpu has a pending nohz_balance_kick, then do the
4062 * balancing on behalf of the other idle cpus whose ticks are
4065 nohz_idle_balance(this_cpu
, idle
);
4068 static inline int on_null_domain(int cpu
)
4070 return !rcu_dereference_sched(cpu_rq(cpu
)->sd
);
4074 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4076 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4078 /* Don't need to rebalance while attached to NULL domain */
4079 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4080 likely(!on_null_domain(cpu
)))
4081 raise_softirq(SCHED_SOFTIRQ
);
4083 else if (nohz_kick_needed(rq
, cpu
) && likely(!on_null_domain(cpu
)))
4084 nohz_balancer_kick(cpu
);
4088 static void rq_online_fair(struct rq
*rq
)
4093 static void rq_offline_fair(struct rq
*rq
)
4098 #else /* CONFIG_SMP */
4101 * on UP we do not need to balance between CPUs:
4103 static inline void idle_balance(int cpu
, struct rq
*rq
)
4107 #endif /* CONFIG_SMP */
4110 * scheduler tick hitting a task of our scheduling class:
4112 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
4114 struct cfs_rq
*cfs_rq
;
4115 struct sched_entity
*se
= &curr
->se
;
4117 for_each_sched_entity(se
) {
4118 cfs_rq
= cfs_rq_of(se
);
4119 entity_tick(cfs_rq
, se
, queued
);
4124 * called on fork with the child task as argument from the parent's context
4125 * - child not yet on the tasklist
4126 * - preemption disabled
4128 static void task_fork_fair(struct task_struct
*p
)
4130 struct cfs_rq
*cfs_rq
= task_cfs_rq(current
);
4131 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
4132 int this_cpu
= smp_processor_id();
4133 struct rq
*rq
= this_rq();
4134 unsigned long flags
;
4136 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4138 update_rq_clock(rq
);
4140 if (unlikely(task_cpu(p
) != this_cpu
)) {
4142 __set_task_cpu(p
, this_cpu
);
4146 update_curr(cfs_rq
);
4149 se
->vruntime
= curr
->vruntime
;
4150 place_entity(cfs_rq
, se
, 1);
4152 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
4154 * Upon rescheduling, sched_class::put_prev_task() will place
4155 * 'current' within the tree based on its new key value.
4157 swap(curr
->vruntime
, se
->vruntime
);
4158 resched_task(rq
->curr
);
4161 se
->vruntime
-= cfs_rq
->min_vruntime
;
4163 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4167 * Priority of the task has changed. Check to see if we preempt
4171 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
4177 * Reschedule if we are currently running on this runqueue and
4178 * our priority decreased, or if we are not currently running on
4179 * this runqueue and our priority is higher than the current's
4181 if (rq
->curr
== p
) {
4182 if (p
->prio
> oldprio
)
4183 resched_task(rq
->curr
);
4185 check_preempt_curr(rq
, p
, 0);
4188 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
4190 struct sched_entity
*se
= &p
->se
;
4191 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4194 * Ensure the task's vruntime is normalized, so that when its
4195 * switched back to the fair class the enqueue_entity(.flags=0) will
4196 * do the right thing.
4198 * If it was on_rq, then the dequeue_entity(.flags=0) will already
4199 * have normalized the vruntime, if it was !on_rq, then only when
4200 * the task is sleeping will it still have non-normalized vruntime.
4202 if (!se
->on_rq
&& p
->state
!= TASK_RUNNING
) {
4204 * Fix up our vruntime so that the current sleep doesn't
4205 * cause 'unlimited' sleep bonus.
4207 place_entity(cfs_rq
, se
, 0);
4208 se
->vruntime
-= cfs_rq
->min_vruntime
;
4213 * We switched to the sched_fair class.
4215 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
4221 * We were most likely switched from sched_rt, so
4222 * kick off the schedule if running, otherwise just see
4223 * if we can still preempt the current task.
4226 resched_task(rq
->curr
);
4228 check_preempt_curr(rq
, p
, 0);
4231 /* Account for a task changing its policy or group.
4233 * This routine is mostly called to set cfs_rq->curr field when a task
4234 * migrates between groups/classes.
4236 static void set_curr_task_fair(struct rq
*rq
)
4238 struct sched_entity
*se
= &rq
->curr
->se
;
4240 for_each_sched_entity(se
)
4241 set_next_entity(cfs_rq_of(se
), se
);
4244 #ifdef CONFIG_FAIR_GROUP_SCHED
4245 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
4248 * If the task was not on the rq at the time of this cgroup movement
4249 * it must have been asleep, sleeping tasks keep their ->vruntime
4250 * absolute on their old rq until wakeup (needed for the fair sleeper
4251 * bonus in place_entity()).
4253 * If it was on the rq, we've just 'preempted' it, which does convert
4254 * ->vruntime to a relative base.
4256 * Make sure both cases convert their relative position when migrating
4257 * to another cgroup's rq. This does somewhat interfere with the
4258 * fair sleeper stuff for the first placement, but who cares.
4261 p
->se
.vruntime
-= cfs_rq_of(&p
->se
)->min_vruntime
;
4262 set_task_rq(p
, task_cpu(p
));
4264 p
->se
.vruntime
+= cfs_rq_of(&p
->se
)->min_vruntime
;
4268 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
4270 struct sched_entity
*se
= &task
->se
;
4271 unsigned int rr_interval
= 0;
4274 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4277 if (rq
->cfs
.load
.weight
)
4278 rr_interval
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
4284 * All the scheduling class methods:
4286 static const struct sched_class fair_sched_class
= {
4287 .next
= &idle_sched_class
,
4288 .enqueue_task
= enqueue_task_fair
,
4289 .dequeue_task
= dequeue_task_fair
,
4290 .yield_task
= yield_task_fair
,
4291 .yield_to_task
= yield_to_task_fair
,
4293 .check_preempt_curr
= check_preempt_wakeup
,
4295 .pick_next_task
= pick_next_task_fair
,
4296 .put_prev_task
= put_prev_task_fair
,
4299 .select_task_rq
= select_task_rq_fair
,
4301 .rq_online
= rq_online_fair
,
4302 .rq_offline
= rq_offline_fair
,
4304 .task_waking
= task_waking_fair
,
4307 .set_curr_task
= set_curr_task_fair
,
4308 .task_tick
= task_tick_fair
,
4309 .task_fork
= task_fork_fair
,
4311 .prio_changed
= prio_changed_fair
,
4312 .switched_from
= switched_from_fair
,
4313 .switched_to
= switched_to_fair
,
4315 .get_rr_interval
= get_rr_interval_fair
,
4317 #ifdef CONFIG_FAIR_GROUP_SCHED
4318 .task_move_group
= task_move_group_fair
,
4322 #ifdef CONFIG_SCHED_DEBUG
4323 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
4325 struct cfs_rq
*cfs_rq
;
4328 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
4329 print_cfs_rq(m
, cpu
, cfs_rq
);