2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
30 #include <trace/events/sched.h>
35 * Targeted preemption latency for CPU-bound tasks:
36 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
38 * NOTE: this latency value is not the same as the concept of
39 * 'timeslice length' - timeslices in CFS are of variable length
40 * and have no persistent notion like in traditional, time-slice
41 * based scheduling concepts.
43 * (to see the precise effective timeslice length of your workload,
44 * run vmstat and monitor the context-switches (cs) field)
46 unsigned int sysctl_sched_latency
= 6000000ULL;
47 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
50 * The initial- and re-scaling of tunables is configurable
51 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
54 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
55 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
56 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
58 enum sched_tunable_scaling sysctl_sched_tunable_scaling
59 = SCHED_TUNABLESCALING_LOG
;
62 * Minimal preemption granularity for CPU-bound tasks:
63 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
65 unsigned int sysctl_sched_min_granularity
= 750000ULL;
66 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
69 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
71 static unsigned int sched_nr_latency
= 8;
74 * After fork, child runs first. If set to 0 (default) then
75 * parent will (try to) run first.
77 unsigned int sysctl_sched_child_runs_first __read_mostly
;
80 * SCHED_OTHER wake-up granularity.
81 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 * This option delays the preemption effects of decoupled workloads
84 * and reduces their over-scheduling. Synchronous workloads will still
85 * have immediate wakeup/sleep latencies.
87 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
88 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
90 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
93 * The exponential sliding window over which load is averaged for shares
97 unsigned int __read_mostly sysctl_sched_shares_window
= 10000000UL;
99 #ifdef CONFIG_CFS_BANDWIDTH
101 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
102 * each time a cfs_rq requests quota.
104 * Note: in the case that the slice exceeds the runtime remaining (either due
105 * to consumption or the quota being specified to be smaller than the slice)
106 * we will always only issue the remaining available time.
108 * default: 5 msec, units: microseconds
110 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
114 * Increase the granularity value when there are more CPUs,
115 * because with more CPUs the 'effective latency' as visible
116 * to users decreases. But the relationship is not linear,
117 * so pick a second-best guess by going with the log2 of the
120 * This idea comes from the SD scheduler of Con Kolivas:
122 static int get_update_sysctl_factor(void)
124 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
127 switch (sysctl_sched_tunable_scaling
) {
128 case SCHED_TUNABLESCALING_NONE
:
131 case SCHED_TUNABLESCALING_LINEAR
:
134 case SCHED_TUNABLESCALING_LOG
:
136 factor
= 1 + ilog2(cpus
);
143 static void update_sysctl(void)
145 unsigned int factor
= get_update_sysctl_factor();
147 #define SET_SYSCTL(name) \
148 (sysctl_##name = (factor) * normalized_sysctl_##name)
149 SET_SYSCTL(sched_min_granularity
);
150 SET_SYSCTL(sched_latency
);
151 SET_SYSCTL(sched_wakeup_granularity
);
155 void sched_init_granularity(void)
160 #if BITS_PER_LONG == 32
161 # define WMULT_CONST (~0UL)
163 # define WMULT_CONST (1UL << 32)
166 #define WMULT_SHIFT 32
169 * Shift right and round:
171 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
174 * delta *= weight / lw
177 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
178 struct load_weight
*lw
)
183 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
184 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
185 * 2^SCHED_LOAD_RESOLUTION.
187 if (likely(weight
> (1UL << SCHED_LOAD_RESOLUTION
)))
188 tmp
= (u64
)delta_exec
* scale_load_down(weight
);
190 tmp
= (u64
)delta_exec
;
192 if (!lw
->inv_weight
) {
193 unsigned long w
= scale_load_down(lw
->weight
);
195 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
197 else if (unlikely(!w
))
198 lw
->inv_weight
= WMULT_CONST
;
200 lw
->inv_weight
= WMULT_CONST
/ w
;
204 * Check whether we'd overflow the 64-bit multiplication:
206 if (unlikely(tmp
> WMULT_CONST
))
207 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
210 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
212 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
216 const struct sched_class fair_sched_class
;
218 /**************************************************************
219 * CFS operations on generic schedulable entities:
222 #ifdef CONFIG_FAIR_GROUP_SCHED
224 /* cpu runqueue to which this cfs_rq is attached */
225 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
230 /* An entity is a task if it doesn't "own" a runqueue */
231 #define entity_is_task(se) (!se->my_q)
233 static inline struct task_struct
*task_of(struct sched_entity
*se
)
235 #ifdef CONFIG_SCHED_DEBUG
236 WARN_ON_ONCE(!entity_is_task(se
));
238 return container_of(se
, struct task_struct
, se
);
241 /* Walk up scheduling entities hierarchy */
242 #define for_each_sched_entity(se) \
243 for (; se; se = se->parent)
245 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
250 /* runqueue on which this entity is (to be) queued */
251 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
256 /* runqueue "owned" by this group */
257 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
262 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
264 if (!cfs_rq
->on_list
) {
266 * Ensure we either appear before our parent (if already
267 * enqueued) or force our parent to appear after us when it is
268 * enqueued. The fact that we always enqueue bottom-up
269 * reduces this to two cases.
271 if (cfs_rq
->tg
->parent
&&
272 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
273 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
274 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
276 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
277 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
284 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
286 if (cfs_rq
->on_list
) {
287 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
292 /* Iterate thr' all leaf cfs_rq's on a runqueue */
293 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
294 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
296 /* Do the two (enqueued) entities belong to the same group ? */
298 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
300 if (se
->cfs_rq
== pse
->cfs_rq
)
306 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
311 /* return depth at which a sched entity is present in the hierarchy */
312 static inline int depth_se(struct sched_entity
*se
)
316 for_each_sched_entity(se
)
323 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
325 int se_depth
, pse_depth
;
328 * preemption test can be made between sibling entities who are in the
329 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
330 * both tasks until we find their ancestors who are siblings of common
334 /* First walk up until both entities are at same depth */
335 se_depth
= depth_se(*se
);
336 pse_depth
= depth_se(*pse
);
338 while (se_depth
> pse_depth
) {
340 *se
= parent_entity(*se
);
343 while (pse_depth
> se_depth
) {
345 *pse
= parent_entity(*pse
);
348 while (!is_same_group(*se
, *pse
)) {
349 *se
= parent_entity(*se
);
350 *pse
= parent_entity(*pse
);
354 #else /* !CONFIG_FAIR_GROUP_SCHED */
356 static inline struct task_struct
*task_of(struct sched_entity
*se
)
358 return container_of(se
, struct task_struct
, se
);
361 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
363 return container_of(cfs_rq
, struct rq
, cfs
);
366 #define entity_is_task(se) 1
368 #define for_each_sched_entity(se) \
369 for (; se; se = NULL)
371 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
373 return &task_rq(p
)->cfs
;
376 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
378 struct task_struct
*p
= task_of(se
);
379 struct rq
*rq
= task_rq(p
);
384 /* runqueue "owned" by this group */
385 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
390 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
394 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
398 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
399 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
402 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
407 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
413 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
417 #endif /* CONFIG_FAIR_GROUP_SCHED */
419 static __always_inline
420 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
);
422 /**************************************************************
423 * Scheduling class tree data structure manipulation methods:
426 static inline u64
max_vruntime(u64 min_vruntime
, u64 vruntime
)
428 s64 delta
= (s64
)(vruntime
- min_vruntime
);
430 min_vruntime
= vruntime
;
435 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
437 s64 delta
= (s64
)(vruntime
- min_vruntime
);
439 min_vruntime
= vruntime
;
444 static inline int entity_before(struct sched_entity
*a
,
445 struct sched_entity
*b
)
447 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
450 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
452 u64 vruntime
= cfs_rq
->min_vruntime
;
455 vruntime
= cfs_rq
->curr
->vruntime
;
457 if (cfs_rq
->rb_leftmost
) {
458 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
463 vruntime
= se
->vruntime
;
465 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
468 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
471 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
476 * Enqueue an entity into the rb-tree:
478 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
480 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
481 struct rb_node
*parent
= NULL
;
482 struct sched_entity
*entry
;
486 * Find the right place in the rbtree:
490 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
492 * We dont care about collisions. Nodes with
493 * the same key stay together.
495 if (entity_before(se
, entry
)) {
496 link
= &parent
->rb_left
;
498 link
= &parent
->rb_right
;
504 * Maintain a cache of leftmost tree entries (it is frequently
508 cfs_rq
->rb_leftmost
= &se
->run_node
;
510 rb_link_node(&se
->run_node
, parent
, link
);
511 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
514 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
516 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
517 struct rb_node
*next_node
;
519 next_node
= rb_next(&se
->run_node
);
520 cfs_rq
->rb_leftmost
= next_node
;
523 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
526 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
528 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
533 return rb_entry(left
, struct sched_entity
, run_node
);
536 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
538 struct rb_node
*next
= rb_next(&se
->run_node
);
543 return rb_entry(next
, struct sched_entity
, run_node
);
546 #ifdef CONFIG_SCHED_DEBUG
547 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
549 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
554 return rb_entry(last
, struct sched_entity
, run_node
);
557 /**************************************************************
558 * Scheduling class statistics methods:
561 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
562 void __user
*buffer
, size_t *lenp
,
565 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
566 int factor
= get_update_sysctl_factor();
571 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
572 sysctl_sched_min_granularity
);
574 #define WRT_SYSCTL(name) \
575 (normalized_sysctl_##name = sysctl_##name / (factor))
576 WRT_SYSCTL(sched_min_granularity
);
577 WRT_SYSCTL(sched_latency
);
578 WRT_SYSCTL(sched_wakeup_granularity
);
588 static inline unsigned long
589 calc_delta_fair(unsigned long delta
, struct sched_entity
*se
)
591 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
592 delta
= calc_delta_mine(delta
, NICE_0_LOAD
, &se
->load
);
598 * The idea is to set a period in which each task runs once.
600 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
601 * this period because otherwise the slices get too small.
603 * p = (nr <= nl) ? l : l*nr/nl
605 static u64
__sched_period(unsigned long nr_running
)
607 u64 period
= sysctl_sched_latency
;
608 unsigned long nr_latency
= sched_nr_latency
;
610 if (unlikely(nr_running
> nr_latency
)) {
611 period
= sysctl_sched_min_granularity
;
612 period
*= nr_running
;
619 * We calculate the wall-time slice from the period by taking a part
620 * proportional to the weight.
624 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
626 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
628 for_each_sched_entity(se
) {
629 struct load_weight
*load
;
630 struct load_weight lw
;
632 cfs_rq
= cfs_rq_of(se
);
633 load
= &cfs_rq
->load
;
635 if (unlikely(!se
->on_rq
)) {
638 update_load_add(&lw
, se
->load
.weight
);
641 slice
= calc_delta_mine(slice
, se
->load
.weight
, load
);
647 * We calculate the vruntime slice of a to be inserted task
651 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
653 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
656 static void update_cfs_load(struct cfs_rq
*cfs_rq
, int global_update
);
657 static void update_cfs_shares(struct cfs_rq
*cfs_rq
);
660 * Update the current task's runtime statistics. Skip current tasks that
661 * are not in our scheduling class.
664 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
665 unsigned long delta_exec
)
667 unsigned long delta_exec_weighted
;
669 schedstat_set(curr
->statistics
.exec_max
,
670 max((u64
)delta_exec
, curr
->statistics
.exec_max
));
672 curr
->sum_exec_runtime
+= delta_exec
;
673 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
674 delta_exec_weighted
= calc_delta_fair(delta_exec
, curr
);
676 curr
->vruntime
+= delta_exec_weighted
;
677 update_min_vruntime(cfs_rq
);
679 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
680 cfs_rq
->load_unacc_exec_time
+= delta_exec
;
684 static void update_curr(struct cfs_rq
*cfs_rq
)
686 struct sched_entity
*curr
= cfs_rq
->curr
;
687 u64 now
= rq_of(cfs_rq
)->clock_task
;
688 unsigned long delta_exec
;
694 * Get the amount of time the current task was running
695 * since the last time we changed load (this cannot
696 * overflow on 32 bits):
698 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
702 __update_curr(cfs_rq
, curr
, delta_exec
);
703 curr
->exec_start
= now
;
705 if (entity_is_task(curr
)) {
706 struct task_struct
*curtask
= task_of(curr
);
708 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
709 cpuacct_charge(curtask
, delta_exec
);
710 account_group_exec_runtime(curtask
, delta_exec
);
713 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
717 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
719 schedstat_set(se
->statistics
.wait_start
, rq_of(cfs_rq
)->clock
);
723 * Task is being enqueued - update stats:
725 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
728 * Are we enqueueing a waiting task? (for current tasks
729 * a dequeue/enqueue event is a NOP)
731 if (se
!= cfs_rq
->curr
)
732 update_stats_wait_start(cfs_rq
, se
);
736 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
738 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
739 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
));
740 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
741 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
742 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
);
743 #ifdef CONFIG_SCHEDSTATS
744 if (entity_is_task(se
)) {
745 trace_sched_stat_wait(task_of(se
),
746 rq_of(cfs_rq
)->clock
- se
->statistics
.wait_start
);
749 schedstat_set(se
->statistics
.wait_start
, 0);
753 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
756 * Mark the end of the wait period if dequeueing a
759 if (se
!= cfs_rq
->curr
)
760 update_stats_wait_end(cfs_rq
, se
);
764 * We are picking a new current task - update its stats:
767 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
770 * We are starting a new run period:
772 se
->exec_start
= rq_of(cfs_rq
)->clock_task
;
775 /**************************************************
776 * Scheduling class queueing methods:
780 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
782 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
783 if (!parent_entity(se
))
784 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
786 if (entity_is_task(se
))
787 list_add(&se
->group_node
, &rq_of(cfs_rq
)->cfs_tasks
);
789 cfs_rq
->nr_running
++;
793 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
795 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
796 if (!parent_entity(se
))
797 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
798 if (entity_is_task(se
))
799 list_del_init(&se
->group_node
);
800 cfs_rq
->nr_running
--;
803 #ifdef CONFIG_FAIR_GROUP_SCHED
804 /* we need this in update_cfs_load and load-balance functions below */
805 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
807 static void update_cfs_rq_load_contribution(struct cfs_rq
*cfs_rq
,
810 struct task_group
*tg
= cfs_rq
->tg
;
813 load_avg
= div64_u64(cfs_rq
->load_avg
, cfs_rq
->load_period
+1);
814 load_avg
-= cfs_rq
->load_contribution
;
816 if (global_update
|| abs(load_avg
) > cfs_rq
->load_contribution
/ 8) {
817 atomic_add(load_avg
, &tg
->load_weight
);
818 cfs_rq
->load_contribution
+= load_avg
;
822 static void update_cfs_load(struct cfs_rq
*cfs_rq
, int global_update
)
824 u64 period
= sysctl_sched_shares_window
;
826 unsigned long load
= cfs_rq
->load
.weight
;
828 if (cfs_rq
->tg
== &root_task_group
|| throttled_hierarchy(cfs_rq
))
831 now
= rq_of(cfs_rq
)->clock_task
;
832 delta
= now
- cfs_rq
->load_stamp
;
834 /* truncate load history at 4 idle periods */
835 if (cfs_rq
->load_stamp
> cfs_rq
->load_last
&&
836 now
- cfs_rq
->load_last
> 4 * period
) {
837 cfs_rq
->load_period
= 0;
838 cfs_rq
->load_avg
= 0;
842 cfs_rq
->load_stamp
= now
;
843 cfs_rq
->load_unacc_exec_time
= 0;
844 cfs_rq
->load_period
+= delta
;
846 cfs_rq
->load_last
= now
;
847 cfs_rq
->load_avg
+= delta
* load
;
850 /* consider updating load contribution on each fold or truncate */
851 if (global_update
|| cfs_rq
->load_period
> period
852 || !cfs_rq
->load_period
)
853 update_cfs_rq_load_contribution(cfs_rq
, global_update
);
855 while (cfs_rq
->load_period
> period
) {
857 * Inline assembly required to prevent the compiler
858 * optimising this loop into a divmod call.
859 * See __iter_div_u64_rem() for another example of this.
861 asm("" : "+rm" (cfs_rq
->load_period
));
862 cfs_rq
->load_period
/= 2;
863 cfs_rq
->load_avg
/= 2;
866 if (!cfs_rq
->curr
&& !cfs_rq
->nr_running
&& !cfs_rq
->load_avg
)
867 list_del_leaf_cfs_rq(cfs_rq
);
870 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
875 * Use this CPU's actual weight instead of the last load_contribution
876 * to gain a more accurate current total weight. See
877 * update_cfs_rq_load_contribution().
879 tg_weight
= atomic_read(&tg
->load_weight
);
880 tg_weight
-= cfs_rq
->load_contribution
;
881 tg_weight
+= cfs_rq
->load
.weight
;
886 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
888 long tg_weight
, load
, shares
;
890 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
891 load
= cfs_rq
->load
.weight
;
893 shares
= (tg
->shares
* load
);
897 if (shares
< MIN_SHARES
)
899 if (shares
> tg
->shares
)
905 static void update_entity_shares_tick(struct cfs_rq
*cfs_rq
)
907 if (cfs_rq
->load_unacc_exec_time
> sysctl_sched_shares_window
) {
908 update_cfs_load(cfs_rq
, 0);
909 update_cfs_shares(cfs_rq
);
912 # else /* CONFIG_SMP */
913 static void update_cfs_load(struct cfs_rq
*cfs_rq
, int global_update
)
917 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
922 static inline void update_entity_shares_tick(struct cfs_rq
*cfs_rq
)
925 # endif /* CONFIG_SMP */
926 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
927 unsigned long weight
)
930 /* commit outstanding execution time */
931 if (cfs_rq
->curr
== se
)
933 account_entity_dequeue(cfs_rq
, se
);
936 update_load_set(&se
->load
, weight
);
939 account_entity_enqueue(cfs_rq
, se
);
942 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
944 struct task_group
*tg
;
945 struct sched_entity
*se
;
949 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
950 if (!se
|| throttled_hierarchy(cfs_rq
))
953 if (likely(se
->load
.weight
== tg
->shares
))
956 shares
= calc_cfs_shares(cfs_rq
, tg
);
958 reweight_entity(cfs_rq_of(se
), se
, shares
);
960 #else /* CONFIG_FAIR_GROUP_SCHED */
961 static void update_cfs_load(struct cfs_rq
*cfs_rq
, int global_update
)
965 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
969 static inline void update_entity_shares_tick(struct cfs_rq
*cfs_rq
)
972 #endif /* CONFIG_FAIR_GROUP_SCHED */
974 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
976 #ifdef CONFIG_SCHEDSTATS
977 struct task_struct
*tsk
= NULL
;
979 if (entity_is_task(se
))
982 if (se
->statistics
.sleep_start
) {
983 u64 delta
= rq_of(cfs_rq
)->clock
- se
->statistics
.sleep_start
;
988 if (unlikely(delta
> se
->statistics
.sleep_max
))
989 se
->statistics
.sleep_max
= delta
;
991 se
->statistics
.sleep_start
= 0;
992 se
->statistics
.sum_sleep_runtime
+= delta
;
995 account_scheduler_latency(tsk
, delta
>> 10, 1);
996 trace_sched_stat_sleep(tsk
, delta
);
999 if (se
->statistics
.block_start
) {
1000 u64 delta
= rq_of(cfs_rq
)->clock
- se
->statistics
.block_start
;
1005 if (unlikely(delta
> se
->statistics
.block_max
))
1006 se
->statistics
.block_max
= delta
;
1008 se
->statistics
.block_start
= 0;
1009 se
->statistics
.sum_sleep_runtime
+= delta
;
1012 if (tsk
->in_iowait
) {
1013 se
->statistics
.iowait_sum
+= delta
;
1014 se
->statistics
.iowait_count
++;
1015 trace_sched_stat_iowait(tsk
, delta
);
1018 trace_sched_stat_blocked(tsk
, delta
);
1021 * Blocking time is in units of nanosecs, so shift by
1022 * 20 to get a milliseconds-range estimation of the
1023 * amount of time that the task spent sleeping:
1025 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
1026 profile_hits(SLEEP_PROFILING
,
1027 (void *)get_wchan(tsk
),
1030 account_scheduler_latency(tsk
, delta
>> 10, 0);
1036 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1038 #ifdef CONFIG_SCHED_DEBUG
1039 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
1044 if (d
> 3*sysctl_sched_latency
)
1045 schedstat_inc(cfs_rq
, nr_spread_over
);
1050 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
1052 u64 vruntime
= cfs_rq
->min_vruntime
;
1055 * The 'current' period is already promised to the current tasks,
1056 * however the extra weight of the new task will slow them down a
1057 * little, place the new task so that it fits in the slot that
1058 * stays open at the end.
1060 if (initial
&& sched_feat(START_DEBIT
))
1061 vruntime
+= sched_vslice(cfs_rq
, se
);
1063 /* sleeps up to a single latency don't count. */
1065 unsigned long thresh
= sysctl_sched_latency
;
1068 * Halve their sleep time's effect, to allow
1069 * for a gentler effect of sleepers:
1071 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
1077 /* ensure we never gain time by being placed backwards. */
1078 vruntime
= max_vruntime(se
->vruntime
, vruntime
);
1080 se
->vruntime
= vruntime
;
1083 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
1086 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1089 * Update the normalized vruntime before updating min_vruntime
1090 * through callig update_curr().
1092 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
1093 se
->vruntime
+= cfs_rq
->min_vruntime
;
1096 * Update run-time statistics of the 'current'.
1098 update_curr(cfs_rq
);
1099 update_cfs_load(cfs_rq
, 0);
1100 account_entity_enqueue(cfs_rq
, se
);
1101 update_cfs_shares(cfs_rq
);
1103 if (flags
& ENQUEUE_WAKEUP
) {
1104 place_entity(cfs_rq
, se
, 0);
1105 enqueue_sleeper(cfs_rq
, se
);
1108 update_stats_enqueue(cfs_rq
, se
);
1109 check_spread(cfs_rq
, se
);
1110 if (se
!= cfs_rq
->curr
)
1111 __enqueue_entity(cfs_rq
, se
);
1114 if (cfs_rq
->nr_running
== 1) {
1115 list_add_leaf_cfs_rq(cfs_rq
);
1116 check_enqueue_throttle(cfs_rq
);
1120 static void __clear_buddies_last(struct sched_entity
*se
)
1122 for_each_sched_entity(se
) {
1123 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1124 if (cfs_rq
->last
== se
)
1125 cfs_rq
->last
= NULL
;
1131 static void __clear_buddies_next(struct sched_entity
*se
)
1133 for_each_sched_entity(se
) {
1134 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1135 if (cfs_rq
->next
== se
)
1136 cfs_rq
->next
= NULL
;
1142 static void __clear_buddies_skip(struct sched_entity
*se
)
1144 for_each_sched_entity(se
) {
1145 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1146 if (cfs_rq
->skip
== se
)
1147 cfs_rq
->skip
= NULL
;
1153 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1155 if (cfs_rq
->last
== se
)
1156 __clear_buddies_last(se
);
1158 if (cfs_rq
->next
== se
)
1159 __clear_buddies_next(se
);
1161 if (cfs_rq
->skip
== se
)
1162 __clear_buddies_skip(se
);
1165 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
1168 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1171 * Update run-time statistics of the 'current'.
1173 update_curr(cfs_rq
);
1175 update_stats_dequeue(cfs_rq
, se
);
1176 if (flags
& DEQUEUE_SLEEP
) {
1177 #ifdef CONFIG_SCHEDSTATS
1178 if (entity_is_task(se
)) {
1179 struct task_struct
*tsk
= task_of(se
);
1181 if (tsk
->state
& TASK_INTERRUPTIBLE
)
1182 se
->statistics
.sleep_start
= rq_of(cfs_rq
)->clock
;
1183 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
1184 se
->statistics
.block_start
= rq_of(cfs_rq
)->clock
;
1189 clear_buddies(cfs_rq
, se
);
1191 if (se
!= cfs_rq
->curr
)
1192 __dequeue_entity(cfs_rq
, se
);
1194 update_cfs_load(cfs_rq
, 0);
1195 account_entity_dequeue(cfs_rq
, se
);
1198 * Normalize the entity after updating the min_vruntime because the
1199 * update can refer to the ->curr item and we need to reflect this
1200 * movement in our normalized position.
1202 if (!(flags
& DEQUEUE_SLEEP
))
1203 se
->vruntime
-= cfs_rq
->min_vruntime
;
1205 /* return excess runtime on last dequeue */
1206 return_cfs_rq_runtime(cfs_rq
);
1208 update_min_vruntime(cfs_rq
);
1209 update_cfs_shares(cfs_rq
);
1213 * Preempt the current task with a newly woken task if needed:
1216 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
1218 unsigned long ideal_runtime
, delta_exec
;
1219 struct sched_entity
*se
;
1222 ideal_runtime
= sched_slice(cfs_rq
, curr
);
1223 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
1224 if (delta_exec
> ideal_runtime
) {
1225 resched_task(rq_of(cfs_rq
)->curr
);
1227 * The current task ran long enough, ensure it doesn't get
1228 * re-elected due to buddy favours.
1230 clear_buddies(cfs_rq
, curr
);
1235 * Ensure that a task that missed wakeup preemption by a
1236 * narrow margin doesn't have to wait for a full slice.
1237 * This also mitigates buddy induced latencies under load.
1239 if (delta_exec
< sysctl_sched_min_granularity
)
1242 se
= __pick_first_entity(cfs_rq
);
1243 delta
= curr
->vruntime
- se
->vruntime
;
1248 if (delta
> ideal_runtime
)
1249 resched_task(rq_of(cfs_rq
)->curr
);
1253 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1255 /* 'current' is not kept within the tree. */
1258 * Any task has to be enqueued before it get to execute on
1259 * a CPU. So account for the time it spent waiting on the
1262 update_stats_wait_end(cfs_rq
, se
);
1263 __dequeue_entity(cfs_rq
, se
);
1266 update_stats_curr_start(cfs_rq
, se
);
1268 #ifdef CONFIG_SCHEDSTATS
1270 * Track our maximum slice length, if the CPU's load is at
1271 * least twice that of our own weight (i.e. dont track it
1272 * when there are only lesser-weight tasks around):
1274 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
1275 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
1276 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
1279 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
1283 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
1286 * Pick the next process, keeping these things in mind, in this order:
1287 * 1) keep things fair between processes/task groups
1288 * 2) pick the "next" process, since someone really wants that to run
1289 * 3) pick the "last" process, for cache locality
1290 * 4) do not run the "skip" process, if something else is available
1292 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
1294 struct sched_entity
*se
= __pick_first_entity(cfs_rq
);
1295 struct sched_entity
*left
= se
;
1298 * Avoid running the skip buddy, if running something else can
1299 * be done without getting too unfair.
1301 if (cfs_rq
->skip
== se
) {
1302 struct sched_entity
*second
= __pick_next_entity(se
);
1303 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
1308 * Prefer last buddy, try to return the CPU to a preempted task.
1310 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
1314 * Someone really wants this to run. If it's not unfair, run it.
1316 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
1319 clear_buddies(cfs_rq
, se
);
1324 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
1326 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
1329 * If still on the runqueue then deactivate_task()
1330 * was not called and update_curr() has to be done:
1333 update_curr(cfs_rq
);
1335 /* throttle cfs_rqs exceeding runtime */
1336 check_cfs_rq_runtime(cfs_rq
);
1338 check_spread(cfs_rq
, prev
);
1340 update_stats_wait_start(cfs_rq
, prev
);
1341 /* Put 'current' back into the tree. */
1342 __enqueue_entity(cfs_rq
, prev
);
1344 cfs_rq
->curr
= NULL
;
1348 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
1351 * Update run-time statistics of the 'current'.
1353 update_curr(cfs_rq
);
1356 * Update share accounting for long-running entities.
1358 update_entity_shares_tick(cfs_rq
);
1360 #ifdef CONFIG_SCHED_HRTICK
1362 * queued ticks are scheduled to match the slice, so don't bother
1363 * validating it and just reschedule.
1366 resched_task(rq_of(cfs_rq
)->curr
);
1370 * don't let the period tick interfere with the hrtick preemption
1372 if (!sched_feat(DOUBLE_TICK
) &&
1373 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
1377 if (cfs_rq
->nr_running
> 1)
1378 check_preempt_tick(cfs_rq
, curr
);
1382 /**************************************************
1383 * CFS bandwidth control machinery
1386 #ifdef CONFIG_CFS_BANDWIDTH
1388 #ifdef HAVE_JUMP_LABEL
1389 static struct static_key __cfs_bandwidth_used
;
1391 static inline bool cfs_bandwidth_used(void)
1393 return static_key_false(&__cfs_bandwidth_used
);
1396 void account_cfs_bandwidth_used(int enabled
, int was_enabled
)
1398 /* only need to count groups transitioning between enabled/!enabled */
1399 if (enabled
&& !was_enabled
)
1400 static_key_slow_inc(&__cfs_bandwidth_used
);
1401 else if (!enabled
&& was_enabled
)
1402 static_key_slow_dec(&__cfs_bandwidth_used
);
1404 #else /* HAVE_JUMP_LABEL */
1405 static bool cfs_bandwidth_used(void)
1410 void account_cfs_bandwidth_used(int enabled
, int was_enabled
) {}
1411 #endif /* HAVE_JUMP_LABEL */
1414 * default period for cfs group bandwidth.
1415 * default: 0.1s, units: nanoseconds
1417 static inline u64
default_cfs_period(void)
1419 return 100000000ULL;
1422 static inline u64
sched_cfs_bandwidth_slice(void)
1424 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
1428 * Replenish runtime according to assigned quota and update expiration time.
1429 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1430 * additional synchronization around rq->lock.
1432 * requires cfs_b->lock
1434 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
1438 if (cfs_b
->quota
== RUNTIME_INF
)
1441 now
= sched_clock_cpu(smp_processor_id());
1442 cfs_b
->runtime
= cfs_b
->quota
;
1443 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
1446 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
1448 return &tg
->cfs_bandwidth
;
1451 /* returns 0 on failure to allocate runtime */
1452 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
1454 struct task_group
*tg
= cfs_rq
->tg
;
1455 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
1456 u64 amount
= 0, min_amount
, expires
;
1458 /* note: this is a positive sum as runtime_remaining <= 0 */
1459 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
1461 raw_spin_lock(&cfs_b
->lock
);
1462 if (cfs_b
->quota
== RUNTIME_INF
)
1463 amount
= min_amount
;
1466 * If the bandwidth pool has become inactive, then at least one
1467 * period must have elapsed since the last consumption.
1468 * Refresh the global state and ensure bandwidth timer becomes
1471 if (!cfs_b
->timer_active
) {
1472 __refill_cfs_bandwidth_runtime(cfs_b
);
1473 __start_cfs_bandwidth(cfs_b
);
1476 if (cfs_b
->runtime
> 0) {
1477 amount
= min(cfs_b
->runtime
, min_amount
);
1478 cfs_b
->runtime
-= amount
;
1482 expires
= cfs_b
->runtime_expires
;
1483 raw_spin_unlock(&cfs_b
->lock
);
1485 cfs_rq
->runtime_remaining
+= amount
;
1487 * we may have advanced our local expiration to account for allowed
1488 * spread between our sched_clock and the one on which runtime was
1491 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
1492 cfs_rq
->runtime_expires
= expires
;
1494 return cfs_rq
->runtime_remaining
> 0;
1498 * Note: This depends on the synchronization provided by sched_clock and the
1499 * fact that rq->clock snapshots this value.
1501 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
1503 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
1504 struct rq
*rq
= rq_of(cfs_rq
);
1506 /* if the deadline is ahead of our clock, nothing to do */
1507 if (likely((s64
)(rq
->clock
- cfs_rq
->runtime_expires
) < 0))
1510 if (cfs_rq
->runtime_remaining
< 0)
1514 * If the local deadline has passed we have to consider the
1515 * possibility that our sched_clock is 'fast' and the global deadline
1516 * has not truly expired.
1518 * Fortunately we can check determine whether this the case by checking
1519 * whether the global deadline has advanced.
1522 if ((s64
)(cfs_rq
->runtime_expires
- cfs_b
->runtime_expires
) >= 0) {
1523 /* extend local deadline, drift is bounded above by 2 ticks */
1524 cfs_rq
->runtime_expires
+= TICK_NSEC
;
1526 /* global deadline is ahead, expiration has passed */
1527 cfs_rq
->runtime_remaining
= 0;
1531 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
,
1532 unsigned long delta_exec
)
1534 /* dock delta_exec before expiring quota (as it could span periods) */
1535 cfs_rq
->runtime_remaining
-= delta_exec
;
1536 expire_cfs_rq_runtime(cfs_rq
);
1538 if (likely(cfs_rq
->runtime_remaining
> 0))
1542 * if we're unable to extend our runtime we resched so that the active
1543 * hierarchy can be throttled
1545 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
1546 resched_task(rq_of(cfs_rq
)->curr
);
1549 static __always_inline
1550 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
)
1552 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
1555 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
1558 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
1560 return cfs_bandwidth_used() && cfs_rq
->throttled
;
1563 /* check whether cfs_rq, or any parent, is throttled */
1564 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
1566 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
1570 * Ensure that neither of the group entities corresponding to src_cpu or
1571 * dest_cpu are members of a throttled hierarchy when performing group
1572 * load-balance operations.
1574 static inline int throttled_lb_pair(struct task_group
*tg
,
1575 int src_cpu
, int dest_cpu
)
1577 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
1579 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
1580 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
1582 return throttled_hierarchy(src_cfs_rq
) ||
1583 throttled_hierarchy(dest_cfs_rq
);
1586 /* updated child weight may affect parent so we have to do this bottom up */
1587 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
1589 struct rq
*rq
= data
;
1590 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
1592 cfs_rq
->throttle_count
--;
1594 if (!cfs_rq
->throttle_count
) {
1595 u64 delta
= rq
->clock_task
- cfs_rq
->load_stamp
;
1597 /* leaving throttled state, advance shares averaging windows */
1598 cfs_rq
->load_stamp
+= delta
;
1599 cfs_rq
->load_last
+= delta
;
1601 /* update entity weight now that we are on_rq again */
1602 update_cfs_shares(cfs_rq
);
1609 static int tg_throttle_down(struct task_group
*tg
, void *data
)
1611 struct rq
*rq
= data
;
1612 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
1614 /* group is entering throttled state, record last load */
1615 if (!cfs_rq
->throttle_count
)
1616 update_cfs_load(cfs_rq
, 0);
1617 cfs_rq
->throttle_count
++;
1622 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
1624 struct rq
*rq
= rq_of(cfs_rq
);
1625 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
1626 struct sched_entity
*se
;
1627 long task_delta
, dequeue
= 1;
1629 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
1631 /* account load preceding throttle */
1633 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
1636 task_delta
= cfs_rq
->h_nr_running
;
1637 for_each_sched_entity(se
) {
1638 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
1639 /* throttled entity or throttle-on-deactivate */
1644 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
1645 qcfs_rq
->h_nr_running
-= task_delta
;
1647 if (qcfs_rq
->load
.weight
)
1652 rq
->nr_running
-= task_delta
;
1654 cfs_rq
->throttled
= 1;
1655 cfs_rq
->throttled_timestamp
= rq
->clock
;
1656 raw_spin_lock(&cfs_b
->lock
);
1657 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
1658 raw_spin_unlock(&cfs_b
->lock
);
1661 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
1663 struct rq
*rq
= rq_of(cfs_rq
);
1664 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
1665 struct sched_entity
*se
;
1669 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
1671 cfs_rq
->throttled
= 0;
1672 raw_spin_lock(&cfs_b
->lock
);
1673 cfs_b
->throttled_time
+= rq
->clock
- cfs_rq
->throttled_timestamp
;
1674 list_del_rcu(&cfs_rq
->throttled_list
);
1675 raw_spin_unlock(&cfs_b
->lock
);
1676 cfs_rq
->throttled_timestamp
= 0;
1678 update_rq_clock(rq
);
1679 /* update hierarchical throttle state */
1680 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
1682 if (!cfs_rq
->load
.weight
)
1685 task_delta
= cfs_rq
->h_nr_running
;
1686 for_each_sched_entity(se
) {
1690 cfs_rq
= cfs_rq_of(se
);
1692 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
1693 cfs_rq
->h_nr_running
+= task_delta
;
1695 if (cfs_rq_throttled(cfs_rq
))
1700 rq
->nr_running
+= task_delta
;
1702 /* determine whether we need to wake up potentially idle cpu */
1703 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
1704 resched_task(rq
->curr
);
1707 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
1708 u64 remaining
, u64 expires
)
1710 struct cfs_rq
*cfs_rq
;
1711 u64 runtime
= remaining
;
1714 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
1716 struct rq
*rq
= rq_of(cfs_rq
);
1718 raw_spin_lock(&rq
->lock
);
1719 if (!cfs_rq_throttled(cfs_rq
))
1722 runtime
= -cfs_rq
->runtime_remaining
+ 1;
1723 if (runtime
> remaining
)
1724 runtime
= remaining
;
1725 remaining
-= runtime
;
1727 cfs_rq
->runtime_remaining
+= runtime
;
1728 cfs_rq
->runtime_expires
= expires
;
1730 /* we check whether we're throttled above */
1731 if (cfs_rq
->runtime_remaining
> 0)
1732 unthrottle_cfs_rq(cfs_rq
);
1735 raw_spin_unlock(&rq
->lock
);
1746 * Responsible for refilling a task_group's bandwidth and unthrottling its
1747 * cfs_rqs as appropriate. If there has been no activity within the last
1748 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1749 * used to track this state.
1751 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
1753 u64 runtime
, runtime_expires
;
1754 int idle
= 1, throttled
;
1756 raw_spin_lock(&cfs_b
->lock
);
1757 /* no need to continue the timer with no bandwidth constraint */
1758 if (cfs_b
->quota
== RUNTIME_INF
)
1761 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
1762 /* idle depends on !throttled (for the case of a large deficit) */
1763 idle
= cfs_b
->idle
&& !throttled
;
1764 cfs_b
->nr_periods
+= overrun
;
1766 /* if we're going inactive then everything else can be deferred */
1770 __refill_cfs_bandwidth_runtime(cfs_b
);
1773 /* mark as potentially idle for the upcoming period */
1778 /* account preceding periods in which throttling occurred */
1779 cfs_b
->nr_throttled
+= overrun
;
1782 * There are throttled entities so we must first use the new bandwidth
1783 * to unthrottle them before making it generally available. This
1784 * ensures that all existing debts will be paid before a new cfs_rq is
1787 runtime
= cfs_b
->runtime
;
1788 runtime_expires
= cfs_b
->runtime_expires
;
1792 * This check is repeated as we are holding onto the new bandwidth
1793 * while we unthrottle. This can potentially race with an unthrottled
1794 * group trying to acquire new bandwidth from the global pool.
1796 while (throttled
&& runtime
> 0) {
1797 raw_spin_unlock(&cfs_b
->lock
);
1798 /* we can't nest cfs_b->lock while distributing bandwidth */
1799 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
1801 raw_spin_lock(&cfs_b
->lock
);
1803 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
1806 /* return (any) remaining runtime */
1807 cfs_b
->runtime
= runtime
;
1809 * While we are ensured activity in the period following an
1810 * unthrottle, this also covers the case in which the new bandwidth is
1811 * insufficient to cover the existing bandwidth deficit. (Forcing the
1812 * timer to remain active while there are any throttled entities.)
1817 cfs_b
->timer_active
= 0;
1818 raw_spin_unlock(&cfs_b
->lock
);
1823 /* a cfs_rq won't donate quota below this amount */
1824 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
1825 /* minimum remaining period time to redistribute slack quota */
1826 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
1827 /* how long we wait to gather additional slack before distributing */
1828 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
1830 /* are we near the end of the current quota period? */
1831 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
1833 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
1836 /* if the call-back is running a quota refresh is already occurring */
1837 if (hrtimer_callback_running(refresh_timer
))
1840 /* is a quota refresh about to occur? */
1841 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
1842 if (remaining
< min_expire
)
1848 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
1850 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
1852 /* if there's a quota refresh soon don't bother with slack */
1853 if (runtime_refresh_within(cfs_b
, min_left
))
1856 start_bandwidth_timer(&cfs_b
->slack_timer
,
1857 ns_to_ktime(cfs_bandwidth_slack_period
));
1860 /* we know any runtime found here is valid as update_curr() precedes return */
1861 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
1863 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
1864 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
1866 if (slack_runtime
<= 0)
1869 raw_spin_lock(&cfs_b
->lock
);
1870 if (cfs_b
->quota
!= RUNTIME_INF
&&
1871 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
1872 cfs_b
->runtime
+= slack_runtime
;
1874 /* we are under rq->lock, defer unthrottling using a timer */
1875 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
1876 !list_empty(&cfs_b
->throttled_cfs_rq
))
1877 start_cfs_slack_bandwidth(cfs_b
);
1879 raw_spin_unlock(&cfs_b
->lock
);
1881 /* even if it's not valid for return we don't want to try again */
1882 cfs_rq
->runtime_remaining
-= slack_runtime
;
1885 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
1887 if (!cfs_bandwidth_used())
1890 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
1893 __return_cfs_rq_runtime(cfs_rq
);
1897 * This is done with a timer (instead of inline with bandwidth return) since
1898 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
1900 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
1902 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
1905 /* confirm we're still not at a refresh boundary */
1906 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
))
1909 raw_spin_lock(&cfs_b
->lock
);
1910 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
) {
1911 runtime
= cfs_b
->runtime
;
1914 expires
= cfs_b
->runtime_expires
;
1915 raw_spin_unlock(&cfs_b
->lock
);
1920 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
1922 raw_spin_lock(&cfs_b
->lock
);
1923 if (expires
== cfs_b
->runtime_expires
)
1924 cfs_b
->runtime
= runtime
;
1925 raw_spin_unlock(&cfs_b
->lock
);
1929 * When a group wakes up we want to make sure that its quota is not already
1930 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
1931 * runtime as update_curr() throttling can not not trigger until it's on-rq.
1933 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
1935 if (!cfs_bandwidth_used())
1938 /* an active group must be handled by the update_curr()->put() path */
1939 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
1942 /* ensure the group is not already throttled */
1943 if (cfs_rq_throttled(cfs_rq
))
1946 /* update runtime allocation */
1947 account_cfs_rq_runtime(cfs_rq
, 0);
1948 if (cfs_rq
->runtime_remaining
<= 0)
1949 throttle_cfs_rq(cfs_rq
);
1952 /* conditionally throttle active cfs_rq's from put_prev_entity() */
1953 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
1955 if (!cfs_bandwidth_used())
1958 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
1962 * it's possible for a throttled entity to be forced into a running
1963 * state (e.g. set_curr_task), in this case we're finished.
1965 if (cfs_rq_throttled(cfs_rq
))
1968 throttle_cfs_rq(cfs_rq
);
1971 static inline u64
default_cfs_period(void);
1972 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
);
1973 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
);
1975 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
1977 struct cfs_bandwidth
*cfs_b
=
1978 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
1979 do_sched_cfs_slack_timer(cfs_b
);
1981 return HRTIMER_NORESTART
;
1984 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
1986 struct cfs_bandwidth
*cfs_b
=
1987 container_of(timer
, struct cfs_bandwidth
, period_timer
);
1993 now
= hrtimer_cb_get_time(timer
);
1994 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
1999 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
2002 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
2005 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2007 raw_spin_lock_init(&cfs_b
->lock
);
2009 cfs_b
->quota
= RUNTIME_INF
;
2010 cfs_b
->period
= ns_to_ktime(default_cfs_period());
2012 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
2013 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2014 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
2015 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2016 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
2019 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2021 cfs_rq
->runtime_enabled
= 0;
2022 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
2025 /* requires cfs_b->lock, may release to reprogram timer */
2026 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2029 * The timer may be active because we're trying to set a new bandwidth
2030 * period or because we're racing with the tear-down path
2031 * (timer_active==0 becomes visible before the hrtimer call-back
2032 * terminates). In either case we ensure that it's re-programmed
2034 while (unlikely(hrtimer_active(&cfs_b
->period_timer
))) {
2035 raw_spin_unlock(&cfs_b
->lock
);
2036 /* ensure cfs_b->lock is available while we wait */
2037 hrtimer_cancel(&cfs_b
->period_timer
);
2039 raw_spin_lock(&cfs_b
->lock
);
2040 /* if someone else restarted the timer then we're done */
2041 if (cfs_b
->timer_active
)
2045 cfs_b
->timer_active
= 1;
2046 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
2049 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2051 hrtimer_cancel(&cfs_b
->period_timer
);
2052 hrtimer_cancel(&cfs_b
->slack_timer
);
2055 void unthrottle_offline_cfs_rqs(struct rq
*rq
)
2057 struct cfs_rq
*cfs_rq
;
2059 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
2060 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2062 if (!cfs_rq
->runtime_enabled
)
2066 * clock_task is not advancing so we just need to make sure
2067 * there's some valid quota amount
2069 cfs_rq
->runtime_remaining
= cfs_b
->quota
;
2070 if (cfs_rq_throttled(cfs_rq
))
2071 unthrottle_cfs_rq(cfs_rq
);
2075 #else /* CONFIG_CFS_BANDWIDTH */
2076 static __always_inline
2077 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
) {}
2078 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2079 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
2080 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2082 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
2087 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
2092 static inline int throttled_lb_pair(struct task_group
*tg
,
2093 int src_cpu
, int dest_cpu
)
2098 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
2100 #ifdef CONFIG_FAIR_GROUP_SCHED
2101 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2104 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
2108 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
2109 void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
2111 #endif /* CONFIG_CFS_BANDWIDTH */
2113 /**************************************************
2114 * CFS operations on tasks:
2117 #ifdef CONFIG_SCHED_HRTICK
2118 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
2120 struct sched_entity
*se
= &p
->se
;
2121 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2123 WARN_ON(task_rq(p
) != rq
);
2125 if (cfs_rq
->nr_running
> 1) {
2126 u64 slice
= sched_slice(cfs_rq
, se
);
2127 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
2128 s64 delta
= slice
- ran
;
2137 * Don't schedule slices shorter than 10000ns, that just
2138 * doesn't make sense. Rely on vruntime for fairness.
2141 delta
= max_t(s64
, 10000LL, delta
);
2143 hrtick_start(rq
, delta
);
2148 * called from enqueue/dequeue and updates the hrtick when the
2149 * current task is from our class and nr_running is low enough
2152 static void hrtick_update(struct rq
*rq
)
2154 struct task_struct
*curr
= rq
->curr
;
2156 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
2159 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
2160 hrtick_start_fair(rq
, curr
);
2162 #else /* !CONFIG_SCHED_HRTICK */
2164 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
2168 static inline void hrtick_update(struct rq
*rq
)
2174 * The enqueue_task method is called before nr_running is
2175 * increased. Here we update the fair scheduling stats and
2176 * then put the task into the rbtree:
2179 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
2181 struct cfs_rq
*cfs_rq
;
2182 struct sched_entity
*se
= &p
->se
;
2184 for_each_sched_entity(se
) {
2187 cfs_rq
= cfs_rq_of(se
);
2188 enqueue_entity(cfs_rq
, se
, flags
);
2191 * end evaluation on encountering a throttled cfs_rq
2193 * note: in the case of encountering a throttled cfs_rq we will
2194 * post the final h_nr_running increment below.
2196 if (cfs_rq_throttled(cfs_rq
))
2198 cfs_rq
->h_nr_running
++;
2200 flags
= ENQUEUE_WAKEUP
;
2203 for_each_sched_entity(se
) {
2204 cfs_rq
= cfs_rq_of(se
);
2205 cfs_rq
->h_nr_running
++;
2207 if (cfs_rq_throttled(cfs_rq
))
2210 update_cfs_load(cfs_rq
, 0);
2211 update_cfs_shares(cfs_rq
);
2219 static void set_next_buddy(struct sched_entity
*se
);
2222 * The dequeue_task method is called before nr_running is
2223 * decreased. We remove the task from the rbtree and
2224 * update the fair scheduling stats:
2226 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
2228 struct cfs_rq
*cfs_rq
;
2229 struct sched_entity
*se
= &p
->se
;
2230 int task_sleep
= flags
& DEQUEUE_SLEEP
;
2232 for_each_sched_entity(se
) {
2233 cfs_rq
= cfs_rq_of(se
);
2234 dequeue_entity(cfs_rq
, se
, flags
);
2237 * end evaluation on encountering a throttled cfs_rq
2239 * note: in the case of encountering a throttled cfs_rq we will
2240 * post the final h_nr_running decrement below.
2242 if (cfs_rq_throttled(cfs_rq
))
2244 cfs_rq
->h_nr_running
--;
2246 /* Don't dequeue parent if it has other entities besides us */
2247 if (cfs_rq
->load
.weight
) {
2249 * Bias pick_next to pick a task from this cfs_rq, as
2250 * p is sleeping when it is within its sched_slice.
2252 if (task_sleep
&& parent_entity(se
))
2253 set_next_buddy(parent_entity(se
));
2255 /* avoid re-evaluating load for this entity */
2256 se
= parent_entity(se
);
2259 flags
|= DEQUEUE_SLEEP
;
2262 for_each_sched_entity(se
) {
2263 cfs_rq
= cfs_rq_of(se
);
2264 cfs_rq
->h_nr_running
--;
2266 if (cfs_rq_throttled(cfs_rq
))
2269 update_cfs_load(cfs_rq
, 0);
2270 update_cfs_shares(cfs_rq
);
2279 /* Used instead of source_load when we know the type == 0 */
2280 static unsigned long weighted_cpuload(const int cpu
)
2282 return cpu_rq(cpu
)->load
.weight
;
2286 * Return a low guess at the load of a migration-source cpu weighted
2287 * according to the scheduling class and "nice" value.
2289 * We want to under-estimate the load of migration sources, to
2290 * balance conservatively.
2292 static unsigned long source_load(int cpu
, int type
)
2294 struct rq
*rq
= cpu_rq(cpu
);
2295 unsigned long total
= weighted_cpuload(cpu
);
2297 if (type
== 0 || !sched_feat(LB_BIAS
))
2300 return min(rq
->cpu_load
[type
-1], total
);
2304 * Return a high guess at the load of a migration-target cpu weighted
2305 * according to the scheduling class and "nice" value.
2307 static unsigned long target_load(int cpu
, int type
)
2309 struct rq
*rq
= cpu_rq(cpu
);
2310 unsigned long total
= weighted_cpuload(cpu
);
2312 if (type
== 0 || !sched_feat(LB_BIAS
))
2315 return max(rq
->cpu_load
[type
-1], total
);
2318 static unsigned long power_of(int cpu
)
2320 return cpu_rq(cpu
)->cpu_power
;
2323 static unsigned long cpu_avg_load_per_task(int cpu
)
2325 struct rq
*rq
= cpu_rq(cpu
);
2326 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
2329 return rq
->load
.weight
/ nr_running
;
2335 static void task_waking_fair(struct task_struct
*p
)
2337 struct sched_entity
*se
= &p
->se
;
2338 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2341 #ifndef CONFIG_64BIT
2342 u64 min_vruntime_copy
;
2345 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
2347 min_vruntime
= cfs_rq
->min_vruntime
;
2348 } while (min_vruntime
!= min_vruntime_copy
);
2350 min_vruntime
= cfs_rq
->min_vruntime
;
2353 se
->vruntime
-= min_vruntime
;
2356 #ifdef CONFIG_FAIR_GROUP_SCHED
2358 * effective_load() calculates the load change as seen from the root_task_group
2360 * Adding load to a group doesn't make a group heavier, but can cause movement
2361 * of group shares between cpus. Assuming the shares were perfectly aligned one
2362 * can calculate the shift in shares.
2364 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2365 * on this @cpu and results in a total addition (subtraction) of @wg to the
2366 * total group weight.
2368 * Given a runqueue weight distribution (rw_i) we can compute a shares
2369 * distribution (s_i) using:
2371 * s_i = rw_i / \Sum rw_j (1)
2373 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2374 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2375 * shares distribution (s_i):
2377 * rw_i = { 2, 4, 1, 0 }
2378 * s_i = { 2/7, 4/7, 1/7, 0 }
2380 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2381 * task used to run on and the CPU the waker is running on), we need to
2382 * compute the effect of waking a task on either CPU and, in case of a sync
2383 * wakeup, compute the effect of the current task going to sleep.
2385 * So for a change of @wl to the local @cpu with an overall group weight change
2386 * of @wl we can compute the new shares distribution (s'_i) using:
2388 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2390 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2391 * differences in waking a task to CPU 0. The additional task changes the
2392 * weight and shares distributions like:
2394 * rw'_i = { 3, 4, 1, 0 }
2395 * s'_i = { 3/8, 4/8, 1/8, 0 }
2397 * We can then compute the difference in effective weight by using:
2399 * dw_i = S * (s'_i - s_i) (3)
2401 * Where 'S' is the group weight as seen by its parent.
2403 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2404 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2405 * 4/7) times the weight of the group.
2407 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
2409 struct sched_entity
*se
= tg
->se
[cpu
];
2411 if (!tg
->parent
) /* the trivial, non-cgroup case */
2414 for_each_sched_entity(se
) {
2420 * W = @wg + \Sum rw_j
2422 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
2427 w
= se
->my_q
->load
.weight
+ wl
;
2430 * wl = S * s'_i; see (2)
2433 wl
= (w
* tg
->shares
) / W
;
2438 * Per the above, wl is the new se->load.weight value; since
2439 * those are clipped to [MIN_SHARES, ...) do so now. See
2440 * calc_cfs_shares().
2442 if (wl
< MIN_SHARES
)
2446 * wl = dw_i = S * (s'_i - s_i); see (3)
2448 wl
-= se
->load
.weight
;
2451 * Recursively apply this logic to all parent groups to compute
2452 * the final effective load change on the root group. Since
2453 * only the @tg group gets extra weight, all parent groups can
2454 * only redistribute existing shares. @wl is the shift in shares
2455 * resulting from this level per the above.
2464 static inline unsigned long effective_load(struct task_group
*tg
, int cpu
,
2465 unsigned long wl
, unsigned long wg
)
2472 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
2474 s64 this_load
, load
;
2475 int idx
, this_cpu
, prev_cpu
;
2476 unsigned long tl_per_task
;
2477 struct task_group
*tg
;
2478 unsigned long weight
;
2482 this_cpu
= smp_processor_id();
2483 prev_cpu
= task_cpu(p
);
2484 load
= source_load(prev_cpu
, idx
);
2485 this_load
= target_load(this_cpu
, idx
);
2488 * If sync wakeup then subtract the (maximum possible)
2489 * effect of the currently running task from the load
2490 * of the current CPU:
2493 tg
= task_group(current
);
2494 weight
= current
->se
.load
.weight
;
2496 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
2497 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
2501 weight
= p
->se
.load
.weight
;
2504 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2505 * due to the sync cause above having dropped this_load to 0, we'll
2506 * always have an imbalance, but there's really nothing you can do
2507 * about that, so that's good too.
2509 * Otherwise check if either cpus are near enough in load to allow this
2510 * task to be woken on this_cpu.
2512 if (this_load
> 0) {
2513 s64 this_eff_load
, prev_eff_load
;
2515 this_eff_load
= 100;
2516 this_eff_load
*= power_of(prev_cpu
);
2517 this_eff_load
*= this_load
+
2518 effective_load(tg
, this_cpu
, weight
, weight
);
2520 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
2521 prev_eff_load
*= power_of(this_cpu
);
2522 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
2524 balanced
= this_eff_load
<= prev_eff_load
;
2529 * If the currently running task will sleep within
2530 * a reasonable amount of time then attract this newly
2533 if (sync
&& balanced
)
2536 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
2537 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
2540 (this_load
<= load
&&
2541 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
2543 * This domain has SD_WAKE_AFFINE and
2544 * p is cache cold in this domain, and
2545 * there is no bad imbalance.
2547 schedstat_inc(sd
, ttwu_move_affine
);
2548 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
2556 * find_idlest_group finds and returns the least busy CPU group within the
2559 static struct sched_group
*
2560 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
2561 int this_cpu
, int load_idx
)
2563 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
2564 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2565 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2568 unsigned long load
, avg_load
;
2572 /* Skip over this group if it has no CPUs allowed */
2573 if (!cpumask_intersects(sched_group_cpus(group
),
2574 tsk_cpus_allowed(p
)))
2577 local_group
= cpumask_test_cpu(this_cpu
,
2578 sched_group_cpus(group
));
2580 /* Tally up the load of all CPUs in the group */
2583 for_each_cpu(i
, sched_group_cpus(group
)) {
2584 /* Bias balancing toward cpus of our domain */
2586 load
= source_load(i
, load_idx
);
2588 load
= target_load(i
, load_idx
);
2593 /* Adjust by relative CPU power of the group */
2594 avg_load
= (avg_load
* SCHED_POWER_SCALE
) / group
->sgp
->power
;
2597 this_load
= avg_load
;
2598 } else if (avg_load
< min_load
) {
2599 min_load
= avg_load
;
2602 } while (group
= group
->next
, group
!= sd
->groups
);
2604 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2610 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2613 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2615 unsigned long load
, min_load
= ULONG_MAX
;
2619 /* Traverse only the allowed CPUs */
2620 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
2621 load
= weighted_cpuload(i
);
2623 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2633 * Try and locate an idle CPU in the sched_domain.
2635 static int select_idle_sibling(struct task_struct
*p
, int target
)
2637 int cpu
= smp_processor_id();
2638 int prev_cpu
= task_cpu(p
);
2639 struct sched_domain
*sd
;
2640 struct sched_group
*sg
;
2644 * If the task is going to be woken-up on this cpu and if it is
2645 * already idle, then it is the right target.
2647 if (target
== cpu
&& idle_cpu(cpu
))
2651 * If the task is going to be woken-up on the cpu where it previously
2652 * ran and if it is currently idle, then it the right target.
2654 if (target
== prev_cpu
&& idle_cpu(prev_cpu
))
2658 * Otherwise, iterate the domains and find an elegible idle cpu.
2660 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
2661 for_each_lower_domain(sd
) {
2664 if (!cpumask_intersects(sched_group_cpus(sg
),
2665 tsk_cpus_allowed(p
)))
2668 for_each_cpu(i
, sched_group_cpus(sg
)) {
2673 target
= cpumask_first_and(sched_group_cpus(sg
),
2674 tsk_cpus_allowed(p
));
2678 } while (sg
!= sd
->groups
);
2685 * sched_balance_self: balance the current task (running on cpu) in domains
2686 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2689 * Balance, ie. select the least loaded group.
2691 * Returns the target CPU number, or the same CPU if no balancing is needed.
2693 * preempt must be disabled.
2696 select_task_rq_fair(struct task_struct
*p
, int sd_flag
, int wake_flags
)
2698 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
2699 int cpu
= smp_processor_id();
2700 int prev_cpu
= task_cpu(p
);
2702 int want_affine
= 0;
2704 int sync
= wake_flags
& WF_SYNC
;
2706 if (p
->rt
.nr_cpus_allowed
== 1)
2709 if (sd_flag
& SD_BALANCE_WAKE
) {
2710 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
2716 for_each_domain(cpu
, tmp
) {
2717 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
2721 * If power savings logic is enabled for a domain, see if we
2722 * are not overloaded, if so, don't balance wider.
2724 if (tmp
->flags
& (SD_POWERSAVINGS_BALANCE
|SD_PREFER_LOCAL
)) {
2725 unsigned long power
= 0;
2726 unsigned long nr_running
= 0;
2727 unsigned long capacity
;
2730 for_each_cpu(i
, sched_domain_span(tmp
)) {
2731 power
+= power_of(i
);
2732 nr_running
+= cpu_rq(i
)->cfs
.nr_running
;
2735 capacity
= DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
);
2737 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2740 if (nr_running
< capacity
)
2745 * If both cpu and prev_cpu are part of this domain,
2746 * cpu is a valid SD_WAKE_AFFINE target.
2748 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
2749 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
2754 if (!want_sd
&& !want_affine
)
2757 if (!(tmp
->flags
& sd_flag
))
2765 if (cpu
== prev_cpu
|| wake_affine(affine_sd
, p
, sync
))
2768 new_cpu
= select_idle_sibling(p
, prev_cpu
);
2773 int load_idx
= sd
->forkexec_idx
;
2774 struct sched_group
*group
;
2777 if (!(sd
->flags
& sd_flag
)) {
2782 if (sd_flag
& SD_BALANCE_WAKE
)
2783 load_idx
= sd
->wake_idx
;
2785 group
= find_idlest_group(sd
, p
, cpu
, load_idx
);
2791 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
2792 if (new_cpu
== -1 || new_cpu
== cpu
) {
2793 /* Now try balancing at a lower domain level of cpu */
2798 /* Now try balancing at a lower domain level of new_cpu */
2800 weight
= sd
->span_weight
;
2802 for_each_domain(cpu
, tmp
) {
2803 if (weight
<= tmp
->span_weight
)
2805 if (tmp
->flags
& sd_flag
)
2808 /* while loop will break here if sd == NULL */
2815 #endif /* CONFIG_SMP */
2817 static unsigned long
2818 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
2820 unsigned long gran
= sysctl_sched_wakeup_granularity
;
2823 * Since its curr running now, convert the gran from real-time
2824 * to virtual-time in his units.
2826 * By using 'se' instead of 'curr' we penalize light tasks, so
2827 * they get preempted easier. That is, if 'se' < 'curr' then
2828 * the resulting gran will be larger, therefore penalizing the
2829 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2830 * be smaller, again penalizing the lighter task.
2832 * This is especially important for buddies when the leftmost
2833 * task is higher priority than the buddy.
2835 return calc_delta_fair(gran
, se
);
2839 * Should 'se' preempt 'curr'.
2853 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
2855 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
2860 gran
= wakeup_gran(curr
, se
);
2867 static void set_last_buddy(struct sched_entity
*se
)
2869 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
2872 for_each_sched_entity(se
)
2873 cfs_rq_of(se
)->last
= se
;
2876 static void set_next_buddy(struct sched_entity
*se
)
2878 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
2881 for_each_sched_entity(se
)
2882 cfs_rq_of(se
)->next
= se
;
2885 static void set_skip_buddy(struct sched_entity
*se
)
2887 for_each_sched_entity(se
)
2888 cfs_rq_of(se
)->skip
= se
;
2892 * Preempt the current task with a newly woken task if needed:
2894 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
2896 struct task_struct
*curr
= rq
->curr
;
2897 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
2898 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
2899 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
2900 int next_buddy_marked
= 0;
2902 if (unlikely(se
== pse
))
2906 * This is possible from callers such as move_task(), in which we
2907 * unconditionally check_prempt_curr() after an enqueue (which may have
2908 * lead to a throttle). This both saves work and prevents false
2909 * next-buddy nomination below.
2911 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
2914 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
2915 set_next_buddy(pse
);
2916 next_buddy_marked
= 1;
2920 * We can come here with TIF_NEED_RESCHED already set from new task
2923 * Note: this also catches the edge-case of curr being in a throttled
2924 * group (e.g. via set_curr_task), since update_curr() (in the
2925 * enqueue of curr) will have resulted in resched being set. This
2926 * prevents us from potentially nominating it as a false LAST_BUDDY
2929 if (test_tsk_need_resched(curr
))
2932 /* Idle tasks are by definition preempted by non-idle tasks. */
2933 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
2934 likely(p
->policy
!= SCHED_IDLE
))
2938 * Batch and idle tasks do not preempt non-idle tasks (their preemption
2939 * is driven by the tick):
2941 if (unlikely(p
->policy
!= SCHED_NORMAL
))
2944 find_matching_se(&se
, &pse
);
2945 update_curr(cfs_rq_of(se
));
2947 if (wakeup_preempt_entity(se
, pse
) == 1) {
2949 * Bias pick_next to pick the sched entity that is
2950 * triggering this preemption.
2952 if (!next_buddy_marked
)
2953 set_next_buddy(pse
);
2962 * Only set the backward buddy when the current task is still
2963 * on the rq. This can happen when a wakeup gets interleaved
2964 * with schedule on the ->pre_schedule() or idle_balance()
2965 * point, either of which can * drop the rq lock.
2967 * Also, during early boot the idle thread is in the fair class,
2968 * for obvious reasons its a bad idea to schedule back to it.
2970 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
2973 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
2977 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
2979 struct task_struct
*p
;
2980 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
2981 struct sched_entity
*se
;
2983 if (!cfs_rq
->nr_running
)
2987 se
= pick_next_entity(cfs_rq
);
2988 set_next_entity(cfs_rq
, se
);
2989 cfs_rq
= group_cfs_rq(se
);
2993 if (hrtick_enabled(rq
))
2994 hrtick_start_fair(rq
, p
);
3000 * Account for a descheduled task:
3002 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
3004 struct sched_entity
*se
= &prev
->se
;
3005 struct cfs_rq
*cfs_rq
;
3007 for_each_sched_entity(se
) {
3008 cfs_rq
= cfs_rq_of(se
);
3009 put_prev_entity(cfs_rq
, se
);
3014 * sched_yield() is very simple
3016 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3018 static void yield_task_fair(struct rq
*rq
)
3020 struct task_struct
*curr
= rq
->curr
;
3021 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
3022 struct sched_entity
*se
= &curr
->se
;
3025 * Are we the only task in the tree?
3027 if (unlikely(rq
->nr_running
== 1))
3030 clear_buddies(cfs_rq
, se
);
3032 if (curr
->policy
!= SCHED_BATCH
) {
3033 update_rq_clock(rq
);
3035 * Update run-time statistics of the 'current'.
3037 update_curr(cfs_rq
);
3039 * Tell update_rq_clock() that we've just updated,
3040 * so we don't do microscopic update in schedule()
3041 * and double the fastpath cost.
3043 rq
->skip_clock_update
= 1;
3049 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
3051 struct sched_entity
*se
= &p
->se
;
3053 /* throttled hierarchies are not runnable */
3054 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
3057 /* Tell the scheduler that we'd really like pse to run next. */
3060 yield_task_fair(rq
);
3066 /**************************************************
3067 * Fair scheduling class load-balancing methods:
3070 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
3072 #define LBF_ALL_PINNED 0x01
3073 #define LBF_NEED_BREAK 0x02
3076 struct sched_domain
*sd
;
3084 enum cpu_idle_type idle
;
3089 unsigned int loop_break
;
3090 unsigned int loop_max
;
3094 * move_task - move a task from one runqueue to another runqueue.
3095 * Both runqueues must be locked.
3097 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
3099 deactivate_task(env
->src_rq
, p
, 0);
3100 set_task_cpu(p
, env
->dst_cpu
);
3101 activate_task(env
->dst_rq
, p
, 0);
3102 check_preempt_curr(env
->dst_rq
, p
, 0);
3106 * Is this task likely cache-hot:
3109 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
3113 if (p
->sched_class
!= &fair_sched_class
)
3116 if (unlikely(p
->policy
== SCHED_IDLE
))
3120 * Buddy candidates are cache hot:
3122 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
3123 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
3124 &p
->se
== cfs_rq_of(&p
->se
)->last
))
3127 if (sysctl_sched_migration_cost
== -1)
3129 if (sysctl_sched_migration_cost
== 0)
3132 delta
= now
- p
->se
.exec_start
;
3134 return delta
< (s64
)sysctl_sched_migration_cost
;
3138 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3141 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
3143 int tsk_cache_hot
= 0;
3145 * We do not migrate tasks that are:
3146 * 1) running (obviously), or
3147 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3148 * 3) are cache-hot on their current CPU.
3150 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
3151 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
3154 env
->flags
&= ~LBF_ALL_PINNED
;
3156 if (task_running(env
->src_rq
, p
)) {
3157 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
3162 * Aggressive migration if:
3163 * 1) task is cache cold, or
3164 * 2) too many balance attempts have failed.
3167 tsk_cache_hot
= task_hot(p
, env
->src_rq
->clock_task
, env
->sd
);
3168 if (!tsk_cache_hot
||
3169 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
3170 #ifdef CONFIG_SCHEDSTATS
3171 if (tsk_cache_hot
) {
3172 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
3173 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
3179 if (tsk_cache_hot
) {
3180 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
3187 * move_one_task tries to move exactly one task from busiest to this_rq, as
3188 * part of active balancing operations within "domain".
3189 * Returns 1 if successful and 0 otherwise.
3191 * Called with both runqueues locked.
3193 static int move_one_task(struct lb_env
*env
)
3195 struct task_struct
*p
, *n
;
3197 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
3198 if (throttled_lb_pair(task_group(p
), env
->src_rq
->cpu
, env
->dst_cpu
))
3201 if (!can_migrate_task(p
, env
))
3206 * Right now, this is only the second place move_task()
3207 * is called, so we can safely collect move_task()
3208 * stats here rather than inside move_task().
3210 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
3216 static unsigned long task_h_load(struct task_struct
*p
);
3218 static const unsigned int sched_nr_migrate_break
= 32;
3221 * move_tasks tries to move up to imbalance weighted load from busiest to
3222 * this_rq, as part of a balancing operation within domain "sd".
3223 * Returns 1 if successful and 0 otherwise.
3225 * Called with both runqueues locked.
3227 static int move_tasks(struct lb_env
*env
)
3229 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
3230 struct task_struct
*p
;
3234 if (env
->imbalance
<= 0)
3237 while (!list_empty(tasks
)) {
3238 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
3241 /* We've more or less seen every task there is, call it quits */
3242 if (env
->loop
> env
->loop_max
)
3245 /* take a breather every nr_migrate tasks */
3246 if (env
->loop
> env
->loop_break
) {
3247 env
->loop_break
+= sched_nr_migrate_break
;
3248 env
->flags
|= LBF_NEED_BREAK
;
3252 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
3255 load
= task_h_load(p
);
3257 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
3260 if ((load
/ 2) > env
->imbalance
)
3263 if (!can_migrate_task(p
, env
))
3268 env
->imbalance
-= load
;
3270 #ifdef CONFIG_PREEMPT
3272 * NEWIDLE balancing is a source of latency, so preemptible
3273 * kernels will stop after the first task is pulled to minimize
3274 * the critical section.
3276 if (env
->idle
== CPU_NEWLY_IDLE
)
3281 * We only want to steal up to the prescribed amount of
3284 if (env
->imbalance
<= 0)
3289 list_move_tail(&p
->se
.group_node
, tasks
);
3293 * Right now, this is one of only two places move_task() is called,
3294 * so we can safely collect move_task() stats here rather than
3295 * inside move_task().
3297 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
3302 #ifdef CONFIG_FAIR_GROUP_SCHED
3304 * update tg->load_weight by folding this cpu's load_avg
3306 static int update_shares_cpu(struct task_group
*tg
, int cpu
)
3308 struct cfs_rq
*cfs_rq
;
3309 unsigned long flags
;
3316 cfs_rq
= tg
->cfs_rq
[cpu
];
3318 raw_spin_lock_irqsave(&rq
->lock
, flags
);
3320 update_rq_clock(rq
);
3321 update_cfs_load(cfs_rq
, 1);
3324 * We need to update shares after updating tg->load_weight in
3325 * order to adjust the weight of groups with long running tasks.
3327 update_cfs_shares(cfs_rq
);
3329 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
3334 static void update_shares(int cpu
)
3336 struct cfs_rq
*cfs_rq
;
3337 struct rq
*rq
= cpu_rq(cpu
);
3341 * Iterates the task_group tree in a bottom up fashion, see
3342 * list_add_leaf_cfs_rq() for details.
3344 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3345 /* throttled entities do not contribute to load */
3346 if (throttled_hierarchy(cfs_rq
))
3349 update_shares_cpu(cfs_rq
->tg
, cpu
);
3355 * Compute the cpu's hierarchical load factor for each task group.
3356 * This needs to be done in a top-down fashion because the load of a child
3357 * group is a fraction of its parents load.
3359 static int tg_load_down(struct task_group
*tg
, void *data
)
3362 long cpu
= (long)data
;
3365 load
= cpu_rq(cpu
)->load
.weight
;
3367 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
3368 load
*= tg
->se
[cpu
]->load
.weight
;
3369 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
3372 tg
->cfs_rq
[cpu
]->h_load
= load
;
3377 static void update_h_load(long cpu
)
3380 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
3384 static unsigned long task_h_load(struct task_struct
*p
)
3386 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
3389 load
= p
->se
.load
.weight
;
3390 load
= div_u64(load
* cfs_rq
->h_load
, cfs_rq
->load
.weight
+ 1);
3395 static inline void update_shares(int cpu
)
3399 static inline void update_h_load(long cpu
)
3403 static unsigned long task_h_load(struct task_struct
*p
)
3405 return p
->se
.load
.weight
;
3409 /********** Helpers for find_busiest_group ************************/
3411 * sd_lb_stats - Structure to store the statistics of a sched_domain
3412 * during load balancing.
3414 struct sd_lb_stats
{
3415 struct sched_group
*busiest
; /* Busiest group in this sd */
3416 struct sched_group
*this; /* Local group in this sd */
3417 unsigned long total_load
; /* Total load of all groups in sd */
3418 unsigned long total_pwr
; /* Total power of all groups in sd */
3419 unsigned long avg_load
; /* Average load across all groups in sd */
3421 /** Statistics of this group */
3422 unsigned long this_load
;
3423 unsigned long this_load_per_task
;
3424 unsigned long this_nr_running
;
3425 unsigned long this_has_capacity
;
3426 unsigned int this_idle_cpus
;
3428 /* Statistics of the busiest group */
3429 unsigned int busiest_idle_cpus
;
3430 unsigned long max_load
;
3431 unsigned long busiest_load_per_task
;
3432 unsigned long busiest_nr_running
;
3433 unsigned long busiest_group_capacity
;
3434 unsigned long busiest_has_capacity
;
3435 unsigned int busiest_group_weight
;
3437 int group_imb
; /* Is there imbalance in this sd */
3438 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3439 int power_savings_balance
; /* Is powersave balance needed for this sd */
3440 struct sched_group
*group_min
; /* Least loaded group in sd */
3441 struct sched_group
*group_leader
; /* Group which relieves group_min */
3442 unsigned long min_load_per_task
; /* load_per_task in group_min */
3443 unsigned long leader_nr_running
; /* Nr running of group_leader */
3444 unsigned long min_nr_running
; /* Nr running of group_min */
3449 * sg_lb_stats - stats of a sched_group required for load_balancing
3451 struct sg_lb_stats
{
3452 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3453 unsigned long group_load
; /* Total load over the CPUs of the group */
3454 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3455 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3456 unsigned long group_capacity
;
3457 unsigned long idle_cpus
;
3458 unsigned long group_weight
;
3459 int group_imb
; /* Is there an imbalance in the group ? */
3460 int group_has_capacity
; /* Is there extra capacity in the group? */
3464 * get_sd_load_idx - Obtain the load index for a given sched domain.
3465 * @sd: The sched_domain whose load_idx is to be obtained.
3466 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3468 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3469 enum cpu_idle_type idle
)
3475 load_idx
= sd
->busy_idx
;
3478 case CPU_NEWLY_IDLE
:
3479 load_idx
= sd
->newidle_idx
;
3482 load_idx
= sd
->idle_idx
;
3490 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3492 * init_sd_power_savings_stats - Initialize power savings statistics for
3493 * the given sched_domain, during load balancing.
3495 * @sd: Sched domain whose power-savings statistics are to be initialized.
3496 * @sds: Variable containing the statistics for sd.
3497 * @idle: Idle status of the CPU at which we're performing load-balancing.
3499 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3500 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3503 * Busy processors will not participate in power savings
3506 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3507 sds
->power_savings_balance
= 0;
3509 sds
->power_savings_balance
= 1;
3510 sds
->min_nr_running
= ULONG_MAX
;
3511 sds
->leader_nr_running
= 0;
3516 * update_sd_power_savings_stats - Update the power saving stats for a
3517 * sched_domain while performing load balancing.
3519 * @group: sched_group belonging to the sched_domain under consideration.
3520 * @sds: Variable containing the statistics of the sched_domain
3521 * @local_group: Does group contain the CPU for which we're performing
3523 * @sgs: Variable containing the statistics of the group.
3525 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3526 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3529 if (!sds
->power_savings_balance
)
3533 * If the local group is idle or completely loaded
3534 * no need to do power savings balance at this domain
3536 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3537 !sds
->this_nr_running
))
3538 sds
->power_savings_balance
= 0;
3541 * If a group is already running at full capacity or idle,
3542 * don't include that group in power savings calculations
3544 if (!sds
->power_savings_balance
||
3545 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3546 !sgs
->sum_nr_running
)
3550 * Calculate the group which has the least non-idle load.
3551 * This is the group from where we need to pick up the load
3554 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3555 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3556 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3557 sds
->group_min
= group
;
3558 sds
->min_nr_running
= sgs
->sum_nr_running
;
3559 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3560 sgs
->sum_nr_running
;
3564 * Calculate the group which is almost near its
3565 * capacity but still has some space to pick up some load
3566 * from other group and save more power
3568 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
3571 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3572 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3573 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3574 sds
->group_leader
= group
;
3575 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3580 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3581 * @env: load balance environment
3582 * @sds: Variable containing the statistics of the sched_domain
3583 * under consideration.
3586 * Check if we have potential to perform some power-savings balance.
3587 * If yes, set the busiest group to be the least loaded group in the
3588 * sched_domain, so that it's CPUs can be put to idle.
3590 * Returns 1 if there is potential to perform power-savings balance.
3594 int check_power_save_busiest_group(struct lb_env
*env
, struct sd_lb_stats
*sds
)
3596 if (!sds
->power_savings_balance
)
3599 if (sds
->this != sds
->group_leader
||
3600 sds
->group_leader
== sds
->group_min
)
3603 env
->imbalance
= sds
->min_load_per_task
;
3604 sds
->busiest
= sds
->group_min
;
3609 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3610 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3611 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3616 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3617 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3623 int check_power_save_busiest_group(struct lb_env
*env
, struct sd_lb_stats
*sds
)
3627 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3630 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3632 return SCHED_POWER_SCALE
;
3635 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3637 return default_scale_freq_power(sd
, cpu
);
3640 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3642 unsigned long weight
= sd
->span_weight
;
3643 unsigned long smt_gain
= sd
->smt_gain
;
3650 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3652 return default_scale_smt_power(sd
, cpu
);
3655 unsigned long scale_rt_power(int cpu
)
3657 struct rq
*rq
= cpu_rq(cpu
);
3658 u64 total
, available
;
3660 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
3662 if (unlikely(total
< rq
->rt_avg
)) {
3663 /* Ensures that power won't end up being negative */
3666 available
= total
- rq
->rt_avg
;
3669 if (unlikely((s64
)total
< SCHED_POWER_SCALE
))
3670 total
= SCHED_POWER_SCALE
;
3672 total
>>= SCHED_POWER_SHIFT
;
3674 return div_u64(available
, total
);
3677 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
3679 unsigned long weight
= sd
->span_weight
;
3680 unsigned long power
= SCHED_POWER_SCALE
;
3681 struct sched_group
*sdg
= sd
->groups
;
3683 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
3684 if (sched_feat(ARCH_POWER
))
3685 power
*= arch_scale_smt_power(sd
, cpu
);
3687 power
*= default_scale_smt_power(sd
, cpu
);
3689 power
>>= SCHED_POWER_SHIFT
;
3692 sdg
->sgp
->power_orig
= power
;
3694 if (sched_feat(ARCH_POWER
))
3695 power
*= arch_scale_freq_power(sd
, cpu
);
3697 power
*= default_scale_freq_power(sd
, cpu
);
3699 power
>>= SCHED_POWER_SHIFT
;
3701 power
*= scale_rt_power(cpu
);
3702 power
>>= SCHED_POWER_SHIFT
;
3707 cpu_rq(cpu
)->cpu_power
= power
;
3708 sdg
->sgp
->power
= power
;
3711 void update_group_power(struct sched_domain
*sd
, int cpu
)
3713 struct sched_domain
*child
= sd
->child
;
3714 struct sched_group
*group
, *sdg
= sd
->groups
;
3715 unsigned long power
;
3716 unsigned long interval
;
3718 interval
= msecs_to_jiffies(sd
->balance_interval
);
3719 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
3720 sdg
->sgp
->next_update
= jiffies
+ interval
;
3723 update_cpu_power(sd
, cpu
);
3729 group
= child
->groups
;
3731 power
+= group
->sgp
->power
;
3732 group
= group
->next
;
3733 } while (group
!= child
->groups
);
3735 sdg
->sgp
->power
= power
;
3739 * Try and fix up capacity for tiny siblings, this is needed when
3740 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3741 * which on its own isn't powerful enough.
3743 * See update_sd_pick_busiest() and check_asym_packing().
3746 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
3749 * Only siblings can have significantly less than SCHED_POWER_SCALE
3751 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
3755 * If ~90% of the cpu_power is still there, we're good.
3757 if (group
->sgp
->power
* 32 > group
->sgp
->power_orig
* 29)
3764 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3765 * @sd: The sched_domain whose statistics are to be updated.
3766 * @group: sched_group whose statistics are to be updated.
3767 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3768 * @local_group: Does group contain this_cpu.
3769 * @cpus: Set of cpus considered for load balancing.
3770 * @balance: Should we balance.
3771 * @sgs: variable to hold the statistics for this group.
3773 static inline void update_sg_lb_stats(struct lb_env
*env
,
3774 struct sched_group
*group
, int load_idx
,
3775 int local_group
, const struct cpumask
*cpus
,
3776 int *balance
, struct sg_lb_stats
*sgs
)
3778 unsigned long load
, max_cpu_load
, min_cpu_load
, max_nr_running
;
3779 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3780 unsigned long avg_load_per_task
= 0;
3784 balance_cpu
= group_first_cpu(group
);
3786 /* Tally up the load of all CPUs in the group */
3788 min_cpu_load
= ~0UL;
3791 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3792 struct rq
*rq
= cpu_rq(i
);
3794 /* Bias balancing toward cpus of our domain */
3796 if (idle_cpu(i
) && !first_idle_cpu
) {
3801 load
= target_load(i
, load_idx
);
3803 load
= source_load(i
, load_idx
);
3804 if (load
> max_cpu_load
) {
3805 max_cpu_load
= load
;
3806 max_nr_running
= rq
->nr_running
;
3808 if (min_cpu_load
> load
)
3809 min_cpu_load
= load
;
3812 sgs
->group_load
+= load
;
3813 sgs
->sum_nr_running
+= rq
->nr_running
;
3814 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3820 * First idle cpu or the first cpu(busiest) in this sched group
3821 * is eligible for doing load balancing at this and above
3822 * domains. In the newly idle case, we will allow all the cpu's
3823 * to do the newly idle load balance.
3826 if (env
->idle
!= CPU_NEWLY_IDLE
) {
3827 if (balance_cpu
!= env
->dst_cpu
) {
3831 update_group_power(env
->sd
, env
->dst_cpu
);
3832 } else if (time_after_eq(jiffies
, group
->sgp
->next_update
))
3833 update_group_power(env
->sd
, env
->dst_cpu
);
3836 /* Adjust by relative CPU power of the group */
3837 sgs
->avg_load
= (sgs
->group_load
*SCHED_POWER_SCALE
) / group
->sgp
->power
;
3840 * Consider the group unbalanced when the imbalance is larger
3841 * than the average weight of a task.
3843 * APZ: with cgroup the avg task weight can vary wildly and
3844 * might not be a suitable number - should we keep a
3845 * normalized nr_running number somewhere that negates
3848 if (sgs
->sum_nr_running
)
3849 avg_load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
3851 if ((max_cpu_load
- min_cpu_load
) >= avg_load_per_task
&& max_nr_running
> 1)
3854 sgs
->group_capacity
= DIV_ROUND_CLOSEST(group
->sgp
->power
,
3856 if (!sgs
->group_capacity
)
3857 sgs
->group_capacity
= fix_small_capacity(env
->sd
, group
);
3858 sgs
->group_weight
= group
->group_weight
;
3860 if (sgs
->group_capacity
> sgs
->sum_nr_running
)
3861 sgs
->group_has_capacity
= 1;
3865 * update_sd_pick_busiest - return 1 on busiest group
3866 * @sd: sched_domain whose statistics are to be checked
3867 * @sds: sched_domain statistics
3868 * @sg: sched_group candidate to be checked for being the busiest
3869 * @sgs: sched_group statistics
3870 * @this_cpu: the current cpu
3872 * Determine if @sg is a busier group than the previously selected
3875 static bool update_sd_pick_busiest(struct lb_env
*env
,
3876 struct sd_lb_stats
*sds
,
3877 struct sched_group
*sg
,
3878 struct sg_lb_stats
*sgs
)
3880 if (sgs
->avg_load
<= sds
->max_load
)
3883 if (sgs
->sum_nr_running
> sgs
->group_capacity
)
3890 * ASYM_PACKING needs to move all the work to the lowest
3891 * numbered CPUs in the group, therefore mark all groups
3892 * higher than ourself as busy.
3894 if ((env
->sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
3895 env
->dst_cpu
< group_first_cpu(sg
)) {
3899 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
3907 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
3908 * @sd: sched_domain whose statistics are to be updated.
3909 * @this_cpu: Cpu for which load balance is currently performed.
3910 * @idle: Idle status of this_cpu
3911 * @cpus: Set of cpus considered for load balancing.
3912 * @balance: Should we balance.
3913 * @sds: variable to hold the statistics for this sched_domain.
3915 static inline void update_sd_lb_stats(struct lb_env
*env
,
3916 const struct cpumask
*cpus
,
3917 int *balance
, struct sd_lb_stats
*sds
)
3919 struct sched_domain
*child
= env
->sd
->child
;
3920 struct sched_group
*sg
= env
->sd
->groups
;
3921 struct sg_lb_stats sgs
;
3922 int load_idx
, prefer_sibling
= 0;
3924 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
3927 init_sd_power_savings_stats(env
->sd
, sds
, env
->idle
);
3928 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
3933 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
3934 memset(&sgs
, 0, sizeof(sgs
));
3935 update_sg_lb_stats(env
, sg
, load_idx
, local_group
,
3936 cpus
, balance
, &sgs
);
3938 if (local_group
&& !(*balance
))
3941 sds
->total_load
+= sgs
.group_load
;
3942 sds
->total_pwr
+= sg
->sgp
->power
;
3945 * In case the child domain prefers tasks go to siblings
3946 * first, lower the sg capacity to one so that we'll try
3947 * and move all the excess tasks away. We lower the capacity
3948 * of a group only if the local group has the capacity to fit
3949 * these excess tasks, i.e. nr_running < group_capacity. The
3950 * extra check prevents the case where you always pull from the
3951 * heaviest group when it is already under-utilized (possible
3952 * with a large weight task outweighs the tasks on the system).
3954 if (prefer_sibling
&& !local_group
&& sds
->this_has_capacity
)
3955 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
3958 sds
->this_load
= sgs
.avg_load
;
3960 sds
->this_nr_running
= sgs
.sum_nr_running
;
3961 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3962 sds
->this_has_capacity
= sgs
.group_has_capacity
;
3963 sds
->this_idle_cpus
= sgs
.idle_cpus
;
3964 } else if (update_sd_pick_busiest(env
, sds
, sg
, &sgs
)) {
3965 sds
->max_load
= sgs
.avg_load
;
3967 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3968 sds
->busiest_idle_cpus
= sgs
.idle_cpus
;
3969 sds
->busiest_group_capacity
= sgs
.group_capacity
;
3970 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3971 sds
->busiest_has_capacity
= sgs
.group_has_capacity
;
3972 sds
->busiest_group_weight
= sgs
.group_weight
;
3973 sds
->group_imb
= sgs
.group_imb
;
3976 update_sd_power_savings_stats(sg
, sds
, local_group
, &sgs
);
3978 } while (sg
!= env
->sd
->groups
);
3982 * check_asym_packing - Check to see if the group is packed into the
3985 * This is primarily intended to used at the sibling level. Some
3986 * cores like POWER7 prefer to use lower numbered SMT threads. In the
3987 * case of POWER7, it can move to lower SMT modes only when higher
3988 * threads are idle. When in lower SMT modes, the threads will
3989 * perform better since they share less core resources. Hence when we
3990 * have idle threads, we want them to be the higher ones.
3992 * This packing function is run on idle threads. It checks to see if
3993 * the busiest CPU in this domain (core in the P7 case) has a higher
3994 * CPU number than the packing function is being run on. Here we are
3995 * assuming lower CPU number will be equivalent to lower a SMT thread
3998 * Returns 1 when packing is required and a task should be moved to
3999 * this CPU. The amount of the imbalance is returned in *imbalance.
4001 * @sd: The sched_domain whose packing is to be checked.
4002 * @sds: Statistics of the sched_domain which is to be packed
4003 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
4004 * @imbalance: returns amount of imbalanced due to packing.
4006 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4010 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
4016 busiest_cpu
= group_first_cpu(sds
->busiest
);
4017 if (env
->dst_cpu
> busiest_cpu
)
4020 env
->imbalance
= DIV_ROUND_CLOSEST(
4021 sds
->max_load
* sds
->busiest
->sgp
->power
, SCHED_POWER_SCALE
);
4027 * fix_small_imbalance - Calculate the minor imbalance that exists
4028 * amongst the groups of a sched_domain, during
4030 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4031 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
4032 * @imbalance: Variable to store the imbalance.
4035 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4037 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
4038 unsigned int imbn
= 2;
4039 unsigned long scaled_busy_load_per_task
;
4041 if (sds
->this_nr_running
) {
4042 sds
->this_load_per_task
/= sds
->this_nr_running
;
4043 if (sds
->busiest_load_per_task
>
4044 sds
->this_load_per_task
)
4047 sds
->this_load_per_task
=
4048 cpu_avg_load_per_task(env
->dst_cpu
);
4051 scaled_busy_load_per_task
= sds
->busiest_load_per_task
4052 * SCHED_POWER_SCALE
;
4053 scaled_busy_load_per_task
/= sds
->busiest
->sgp
->power
;
4055 if (sds
->max_load
- sds
->this_load
+ scaled_busy_load_per_task
>=
4056 (scaled_busy_load_per_task
* imbn
)) {
4057 env
->imbalance
= sds
->busiest_load_per_task
;
4062 * OK, we don't have enough imbalance to justify moving tasks,
4063 * however we may be able to increase total CPU power used by
4067 pwr_now
+= sds
->busiest
->sgp
->power
*
4068 min(sds
->busiest_load_per_task
, sds
->max_load
);
4069 pwr_now
+= sds
->this->sgp
->power
*
4070 min(sds
->this_load_per_task
, sds
->this_load
);
4071 pwr_now
/= SCHED_POWER_SCALE
;
4073 /* Amount of load we'd subtract */
4074 tmp
= (sds
->busiest_load_per_task
* SCHED_POWER_SCALE
) /
4075 sds
->busiest
->sgp
->power
;
4076 if (sds
->max_load
> tmp
)
4077 pwr_move
+= sds
->busiest
->sgp
->power
*
4078 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
4080 /* Amount of load we'd add */
4081 if (sds
->max_load
* sds
->busiest
->sgp
->power
<
4082 sds
->busiest_load_per_task
* SCHED_POWER_SCALE
)
4083 tmp
= (sds
->max_load
* sds
->busiest
->sgp
->power
) /
4084 sds
->this->sgp
->power
;
4086 tmp
= (sds
->busiest_load_per_task
* SCHED_POWER_SCALE
) /
4087 sds
->this->sgp
->power
;
4088 pwr_move
+= sds
->this->sgp
->power
*
4089 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
4090 pwr_move
/= SCHED_POWER_SCALE
;
4092 /* Move if we gain throughput */
4093 if (pwr_move
> pwr_now
)
4094 env
->imbalance
= sds
->busiest_load_per_task
;
4098 * calculate_imbalance - Calculate the amount of imbalance present within the
4099 * groups of a given sched_domain during load balance.
4100 * @env: load balance environment
4101 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4103 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4105 unsigned long max_pull
, load_above_capacity
= ~0UL;
4107 sds
->busiest_load_per_task
/= sds
->busiest_nr_running
;
4108 if (sds
->group_imb
) {
4109 sds
->busiest_load_per_task
=
4110 min(sds
->busiest_load_per_task
, sds
->avg_load
);
4114 * In the presence of smp nice balancing, certain scenarios can have
4115 * max load less than avg load(as we skip the groups at or below
4116 * its cpu_power, while calculating max_load..)
4118 if (sds
->max_load
< sds
->avg_load
) {
4120 return fix_small_imbalance(env
, sds
);
4123 if (!sds
->group_imb
) {
4125 * Don't want to pull so many tasks that a group would go idle.
4127 load_above_capacity
= (sds
->busiest_nr_running
-
4128 sds
->busiest_group_capacity
);
4130 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_POWER_SCALE
);
4132 load_above_capacity
/= sds
->busiest
->sgp
->power
;
4136 * We're trying to get all the cpus to the average_load, so we don't
4137 * want to push ourselves above the average load, nor do we wish to
4138 * reduce the max loaded cpu below the average load. At the same time,
4139 * we also don't want to reduce the group load below the group capacity
4140 * (so that we can implement power-savings policies etc). Thus we look
4141 * for the minimum possible imbalance.
4142 * Be careful of negative numbers as they'll appear as very large values
4143 * with unsigned longs.
4145 max_pull
= min(sds
->max_load
- sds
->avg_load
, load_above_capacity
);
4147 /* How much load to actually move to equalise the imbalance */
4148 env
->imbalance
= min(max_pull
* sds
->busiest
->sgp
->power
,
4149 (sds
->avg_load
- sds
->this_load
) * sds
->this->sgp
->power
)
4150 / SCHED_POWER_SCALE
;
4153 * if *imbalance is less than the average load per runnable task
4154 * there is no guarantee that any tasks will be moved so we'll have
4155 * a think about bumping its value to force at least one task to be
4158 if (env
->imbalance
< sds
->busiest_load_per_task
)
4159 return fix_small_imbalance(env
, sds
);
4163 /******* find_busiest_group() helpers end here *********************/
4166 * find_busiest_group - Returns the busiest group within the sched_domain
4167 * if there is an imbalance. If there isn't an imbalance, and
4168 * the user has opted for power-savings, it returns a group whose
4169 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4170 * such a group exists.
4172 * Also calculates the amount of weighted load which should be moved
4173 * to restore balance.
4175 * @sd: The sched_domain whose busiest group is to be returned.
4176 * @this_cpu: The cpu for which load balancing is currently being performed.
4177 * @imbalance: Variable which stores amount of weighted load which should
4178 * be moved to restore balance/put a group to idle.
4179 * @idle: The idle status of this_cpu.
4180 * @cpus: The set of CPUs under consideration for load-balancing.
4181 * @balance: Pointer to a variable indicating if this_cpu
4182 * is the appropriate cpu to perform load balancing at this_level.
4184 * Returns: - the busiest group if imbalance exists.
4185 * - If no imbalance and user has opted for power-savings balance,
4186 * return the least loaded group whose CPUs can be
4187 * put to idle by rebalancing its tasks onto our group.
4189 static struct sched_group
*
4190 find_busiest_group(struct lb_env
*env
, const struct cpumask
*cpus
, int *balance
)
4192 struct sd_lb_stats sds
;
4194 memset(&sds
, 0, sizeof(sds
));
4197 * Compute the various statistics relavent for load balancing at
4200 update_sd_lb_stats(env
, cpus
, balance
, &sds
);
4203 * this_cpu is not the appropriate cpu to perform load balancing at
4209 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
4210 check_asym_packing(env
, &sds
))
4213 /* There is no busy sibling group to pull tasks from */
4214 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
4217 sds
.avg_load
= (SCHED_POWER_SCALE
* sds
.total_load
) / sds
.total_pwr
;
4220 * If the busiest group is imbalanced the below checks don't
4221 * work because they assumes all things are equal, which typically
4222 * isn't true due to cpus_allowed constraints and the like.
4227 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4228 if (env
->idle
== CPU_NEWLY_IDLE
&& sds
.this_has_capacity
&&
4229 !sds
.busiest_has_capacity
)
4233 * If the local group is more busy than the selected busiest group
4234 * don't try and pull any tasks.
4236 if (sds
.this_load
>= sds
.max_load
)
4240 * Don't pull any tasks if this group is already above the domain
4243 if (sds
.this_load
>= sds
.avg_load
)
4246 if (env
->idle
== CPU_IDLE
) {
4248 * This cpu is idle. If the busiest group load doesn't
4249 * have more tasks than the number of available cpu's and
4250 * there is no imbalance between this and busiest group
4251 * wrt to idle cpu's, it is balanced.
4253 if ((sds
.this_idle_cpus
<= sds
.busiest_idle_cpus
+ 1) &&
4254 sds
.busiest_nr_running
<= sds
.busiest_group_weight
)
4258 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4259 * imbalance_pct to be conservative.
4261 if (100 * sds
.max_load
<= env
->sd
->imbalance_pct
* sds
.this_load
)
4266 /* Looks like there is an imbalance. Compute it */
4267 calculate_imbalance(env
, &sds
);
4272 * There is no obvious imbalance. But check if we can do some balancing
4275 if (check_power_save_busiest_group(env
, &sds
))
4283 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4285 static struct rq
*find_busiest_queue(struct lb_env
*env
,
4286 struct sched_group
*group
,
4287 const struct cpumask
*cpus
)
4289 struct rq
*busiest
= NULL
, *rq
;
4290 unsigned long max_load
= 0;
4293 for_each_cpu(i
, sched_group_cpus(group
)) {
4294 unsigned long power
= power_of(i
);
4295 unsigned long capacity
= DIV_ROUND_CLOSEST(power
,
4300 capacity
= fix_small_capacity(env
->sd
, group
);
4302 if (!cpumask_test_cpu(i
, cpus
))
4306 wl
= weighted_cpuload(i
);
4309 * When comparing with imbalance, use weighted_cpuload()
4310 * which is not scaled with the cpu power.
4312 if (capacity
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
4316 * For the load comparisons with the other cpu's, consider
4317 * the weighted_cpuload() scaled with the cpu power, so that
4318 * the load can be moved away from the cpu that is potentially
4319 * running at a lower capacity.
4321 wl
= (wl
* SCHED_POWER_SCALE
) / power
;
4323 if (wl
> max_load
) {
4333 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4334 * so long as it is large enough.
4336 #define MAX_PINNED_INTERVAL 512
4338 /* Working cpumask for load_balance and load_balance_newidle. */
4339 DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4341 static int need_active_balance(struct lb_env
*env
)
4343 struct sched_domain
*sd
= env
->sd
;
4345 if (env
->idle
== CPU_NEWLY_IDLE
) {
4348 * ASYM_PACKING needs to force migrate tasks from busy but
4349 * higher numbered CPUs in order to pack all tasks in the
4350 * lowest numbered CPUs.
4352 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
4356 * The only task running in a non-idle cpu can be moved to this
4357 * cpu in an attempt to completely freeup the other CPU
4360 * The package power saving logic comes from
4361 * find_busiest_group(). If there are no imbalance, then
4362 * f_b_g() will return NULL. However when sched_mc={1,2} then
4363 * f_b_g() will select a group from which a running task may be
4364 * pulled to this cpu in order to make the other package idle.
4365 * If there is no opportunity to make a package idle and if
4366 * there are no imbalance, then f_b_g() will return NULL and no
4367 * action will be taken in load_balance_newidle().
4369 * Under normal task pull operation due to imbalance, there
4370 * will be more than one task in the source run queue and
4371 * move_tasks() will succeed. ld_moved will be true and this
4372 * active balance code will not be triggered.
4374 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4378 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
4381 static int active_load_balance_cpu_stop(void *data
);
4384 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4385 * tasks if there is an imbalance.
4387 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4388 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4391 int ld_moved
, active_balance
= 0;
4392 struct sched_group
*group
;
4394 unsigned long flags
;
4395 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4397 struct lb_env env
= {
4399 .dst_cpu
= this_cpu
,
4402 .loop_break
= sched_nr_migrate_break
,
4405 cpumask_copy(cpus
, cpu_active_mask
);
4407 schedstat_inc(sd
, lb_count
[idle
]);
4410 group
= find_busiest_group(&env
, cpus
, balance
);
4416 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4420 busiest
= find_busiest_queue(&env
, group
, cpus
);
4422 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4426 BUG_ON(busiest
== this_rq
);
4428 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
4431 if (busiest
->nr_running
> 1) {
4433 * Attempt to move tasks. If find_busiest_group has found
4434 * an imbalance but busiest->nr_running <= 1, the group is
4435 * still unbalanced. ld_moved simply stays zero, so it is
4436 * correctly treated as an imbalance.
4438 env
.flags
|= LBF_ALL_PINNED
;
4439 env
.src_cpu
= busiest
->cpu
;
4440 env
.src_rq
= busiest
;
4441 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
4444 local_irq_save(flags
);
4445 double_rq_lock(this_rq
, busiest
);
4447 update_h_load(env
.src_cpu
);
4448 ld_moved
+= move_tasks(&env
);
4449 double_rq_unlock(this_rq
, busiest
);
4450 local_irq_restore(flags
);
4452 if (env
.flags
& LBF_NEED_BREAK
) {
4453 env
.flags
&= ~LBF_NEED_BREAK
;
4458 * some other cpu did the load balance for us.
4460 if (ld_moved
&& this_cpu
!= smp_processor_id())
4461 resched_cpu(this_cpu
);
4463 /* All tasks on this runqueue were pinned by CPU affinity */
4464 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
4465 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4466 if (!cpumask_empty(cpus
))
4473 schedstat_inc(sd
, lb_failed
[idle
]);
4475 * Increment the failure counter only on periodic balance.
4476 * We do not want newidle balance, which can be very
4477 * frequent, pollute the failure counter causing
4478 * excessive cache_hot migrations and active balances.
4480 if (idle
!= CPU_NEWLY_IDLE
)
4481 sd
->nr_balance_failed
++;
4483 if (need_active_balance(&env
)) {
4484 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
4486 /* don't kick the active_load_balance_cpu_stop,
4487 * if the curr task on busiest cpu can't be
4490 if (!cpumask_test_cpu(this_cpu
,
4491 tsk_cpus_allowed(busiest
->curr
))) {
4492 raw_spin_unlock_irqrestore(&busiest
->lock
,
4494 env
.flags
|= LBF_ALL_PINNED
;
4495 goto out_one_pinned
;
4499 * ->active_balance synchronizes accesses to
4500 * ->active_balance_work. Once set, it's cleared
4501 * only after active load balance is finished.
4503 if (!busiest
->active_balance
) {
4504 busiest
->active_balance
= 1;
4505 busiest
->push_cpu
= this_cpu
;
4508 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
4510 if (active_balance
) {
4511 stop_one_cpu_nowait(cpu_of(busiest
),
4512 active_load_balance_cpu_stop
, busiest
,
4513 &busiest
->active_balance_work
);
4517 * We've kicked active balancing, reset the failure
4520 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4523 sd
->nr_balance_failed
= 0;
4525 if (likely(!active_balance
)) {
4526 /* We were unbalanced, so reset the balancing interval */
4527 sd
->balance_interval
= sd
->min_interval
;
4530 * If we've begun active balancing, start to back off. This
4531 * case may not be covered by the all_pinned logic if there
4532 * is only 1 task on the busy runqueue (because we don't call
4535 if (sd
->balance_interval
< sd
->max_interval
)
4536 sd
->balance_interval
*= 2;
4542 schedstat_inc(sd
, lb_balanced
[idle
]);
4544 sd
->nr_balance_failed
= 0;
4547 /* tune up the balancing interval */
4548 if (((env
.flags
& LBF_ALL_PINNED
) &&
4549 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4550 (sd
->balance_interval
< sd
->max_interval
))
4551 sd
->balance_interval
*= 2;
4559 * idle_balance is called by schedule() if this_cpu is about to become
4560 * idle. Attempts to pull tasks from other CPUs.
4562 void idle_balance(int this_cpu
, struct rq
*this_rq
)
4564 struct sched_domain
*sd
;
4565 int pulled_task
= 0;
4566 unsigned long next_balance
= jiffies
+ HZ
;
4568 this_rq
->idle_stamp
= this_rq
->clock
;
4570 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
4574 * Drop the rq->lock, but keep IRQ/preempt disabled.
4576 raw_spin_unlock(&this_rq
->lock
);
4578 update_shares(this_cpu
);
4580 for_each_domain(this_cpu
, sd
) {
4581 unsigned long interval
;
4584 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4587 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
4588 /* If we've pulled tasks over stop searching: */
4589 pulled_task
= load_balance(this_cpu
, this_rq
,
4590 sd
, CPU_NEWLY_IDLE
, &balance
);
4593 interval
= msecs_to_jiffies(sd
->balance_interval
);
4594 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4595 next_balance
= sd
->last_balance
+ interval
;
4597 this_rq
->idle_stamp
= 0;
4603 raw_spin_lock(&this_rq
->lock
);
4605 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4607 * We are going idle. next_balance may be set based on
4608 * a busy processor. So reset next_balance.
4610 this_rq
->next_balance
= next_balance
;
4615 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4616 * running tasks off the busiest CPU onto idle CPUs. It requires at
4617 * least 1 task to be running on each physical CPU where possible, and
4618 * avoids physical / logical imbalances.
4620 static int active_load_balance_cpu_stop(void *data
)
4622 struct rq
*busiest_rq
= data
;
4623 int busiest_cpu
= cpu_of(busiest_rq
);
4624 int target_cpu
= busiest_rq
->push_cpu
;
4625 struct rq
*target_rq
= cpu_rq(target_cpu
);
4626 struct sched_domain
*sd
;
4628 raw_spin_lock_irq(&busiest_rq
->lock
);
4630 /* make sure the requested cpu hasn't gone down in the meantime */
4631 if (unlikely(busiest_cpu
!= smp_processor_id() ||
4632 !busiest_rq
->active_balance
))
4635 /* Is there any task to move? */
4636 if (busiest_rq
->nr_running
<= 1)
4640 * This condition is "impossible", if it occurs
4641 * we need to fix it. Originally reported by
4642 * Bjorn Helgaas on a 128-cpu setup.
4644 BUG_ON(busiest_rq
== target_rq
);
4646 /* move a task from busiest_rq to target_rq */
4647 double_lock_balance(busiest_rq
, target_rq
);
4649 /* Search for an sd spanning us and the target CPU. */
4651 for_each_domain(target_cpu
, sd
) {
4652 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4653 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4658 struct lb_env env
= {
4660 .dst_cpu
= target_cpu
,
4661 .dst_rq
= target_rq
,
4662 .src_cpu
= busiest_rq
->cpu
,
4663 .src_rq
= busiest_rq
,
4667 schedstat_inc(sd
, alb_count
);
4669 if (move_one_task(&env
))
4670 schedstat_inc(sd
, alb_pushed
);
4672 schedstat_inc(sd
, alb_failed
);
4675 double_unlock_balance(busiest_rq
, target_rq
);
4677 busiest_rq
->active_balance
= 0;
4678 raw_spin_unlock_irq(&busiest_rq
->lock
);
4684 * idle load balancing details
4685 * - When one of the busy CPUs notice that there may be an idle rebalancing
4686 * needed, they will kick the idle load balancer, which then does idle
4687 * load balancing for all the idle CPUs.
4690 cpumask_var_t idle_cpus_mask
;
4692 unsigned long next_balance
; /* in jiffy units */
4693 } nohz ____cacheline_aligned
;
4695 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4697 * lowest_flag_domain - Return lowest sched_domain containing flag.
4698 * @cpu: The cpu whose lowest level of sched domain is to
4700 * @flag: The flag to check for the lowest sched_domain
4701 * for the given cpu.
4703 * Returns the lowest sched_domain of a cpu which contains the given flag.
4705 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4707 struct sched_domain
*sd
;
4709 for_each_domain(cpu
, sd
)
4710 if (sd
->flags
& flag
)
4717 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4718 * @cpu: The cpu whose domains we're iterating over.
4719 * @sd: variable holding the value of the power_savings_sd
4721 * @flag: The flag to filter the sched_domains to be iterated.
4723 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4724 * set, starting from the lowest sched_domain to the highest.
4726 #define for_each_flag_domain(cpu, sd, flag) \
4727 for (sd = lowest_flag_domain(cpu, flag); \
4728 (sd && (sd->flags & flag)); sd = sd->parent)
4731 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4732 * @cpu: The cpu which is nominating a new idle_load_balancer.
4734 * Returns: Returns the id of the idle load balancer if it exists,
4735 * Else, returns >= nr_cpu_ids.
4737 * This algorithm picks the idle load balancer such that it belongs to a
4738 * semi-idle powersavings sched_domain. The idea is to try and avoid
4739 * completely idle packages/cores just for the purpose of idle load balancing
4740 * when there are other idle cpu's which are better suited for that job.
4742 static int find_new_ilb(int cpu
)
4744 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
4745 struct sched_group
*ilbg
;
4746 struct sched_domain
*sd
;
4749 * Have idle load balancer selection from semi-idle packages only
4750 * when power-aware load balancing is enabled
4752 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4756 * Optimize for the case when we have no idle CPUs or only one
4757 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4759 if (cpumask_weight(nohz
.idle_cpus_mask
) < 2)
4763 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4767 if (ilbg
->group_weight
!=
4768 atomic_read(&ilbg
->sgp
->nr_busy_cpus
)) {
4769 ilb
= cpumask_first_and(nohz
.idle_cpus_mask
,
4770 sched_group_cpus(ilbg
));
4776 } while (ilbg
!= sd
->groups
);
4782 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
4787 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4788 static inline int find_new_ilb(int call_cpu
)
4795 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4796 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4797 * CPU (if there is one).
4799 static void nohz_balancer_kick(int cpu
)
4803 nohz
.next_balance
++;
4805 ilb_cpu
= find_new_ilb(cpu
);
4807 if (ilb_cpu
>= nr_cpu_ids
)
4810 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
4813 * Use smp_send_reschedule() instead of resched_cpu().
4814 * This way we generate a sched IPI on the target cpu which
4815 * is idle. And the softirq performing nohz idle load balance
4816 * will be run before returning from the IPI.
4818 smp_send_reschedule(ilb_cpu
);
4822 static inline void clear_nohz_tick_stopped(int cpu
)
4824 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
4825 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
4826 atomic_dec(&nohz
.nr_cpus
);
4827 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
4831 static inline void set_cpu_sd_state_busy(void)
4833 struct sched_domain
*sd
;
4834 int cpu
= smp_processor_id();
4836 if (!test_bit(NOHZ_IDLE
, nohz_flags(cpu
)))
4838 clear_bit(NOHZ_IDLE
, nohz_flags(cpu
));
4841 for_each_domain(cpu
, sd
)
4842 atomic_inc(&sd
->groups
->sgp
->nr_busy_cpus
);
4846 void set_cpu_sd_state_idle(void)
4848 struct sched_domain
*sd
;
4849 int cpu
= smp_processor_id();
4851 if (test_bit(NOHZ_IDLE
, nohz_flags(cpu
)))
4853 set_bit(NOHZ_IDLE
, nohz_flags(cpu
));
4856 for_each_domain(cpu
, sd
)
4857 atomic_dec(&sd
->groups
->sgp
->nr_busy_cpus
);
4862 * This routine will record that this cpu is going idle with tick stopped.
4863 * This info will be used in performing idle load balancing in the future.
4865 void select_nohz_load_balancer(int stop_tick
)
4867 int cpu
= smp_processor_id();
4870 * If this cpu is going down, then nothing needs to be done.
4872 if (!cpu_active(cpu
))
4876 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
4879 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
4880 atomic_inc(&nohz
.nr_cpus
);
4881 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
4886 static int __cpuinit
sched_ilb_notifier(struct notifier_block
*nfb
,
4887 unsigned long action
, void *hcpu
)
4889 switch (action
& ~CPU_TASKS_FROZEN
) {
4891 clear_nohz_tick_stopped(smp_processor_id());
4899 static DEFINE_SPINLOCK(balancing
);
4902 * Scale the max load_balance interval with the number of CPUs in the system.
4903 * This trades load-balance latency on larger machines for less cross talk.
4905 void update_max_interval(void)
4907 max_load_balance_interval
= HZ
*num_online_cpus()/10;
4911 * It checks each scheduling domain to see if it is due to be balanced,
4912 * and initiates a balancing operation if so.
4914 * Balancing parameters are set up in arch_init_sched_domains.
4916 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4919 struct rq
*rq
= cpu_rq(cpu
);
4920 unsigned long interval
;
4921 struct sched_domain
*sd
;
4922 /* Earliest time when we have to do rebalance again */
4923 unsigned long next_balance
= jiffies
+ 60*HZ
;
4924 int update_next_balance
= 0;
4930 for_each_domain(cpu
, sd
) {
4931 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4934 interval
= sd
->balance_interval
;
4935 if (idle
!= CPU_IDLE
)
4936 interval
*= sd
->busy_factor
;
4938 /* scale ms to jiffies */
4939 interval
= msecs_to_jiffies(interval
);
4940 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
4942 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4944 if (need_serialize
) {
4945 if (!spin_trylock(&balancing
))
4949 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4950 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4952 * We've pulled tasks over so either we're no
4955 idle
= CPU_NOT_IDLE
;
4957 sd
->last_balance
= jiffies
;
4960 spin_unlock(&balancing
);
4962 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4963 next_balance
= sd
->last_balance
+ interval
;
4964 update_next_balance
= 1;
4968 * Stop the load balance at this level. There is another
4969 * CPU in our sched group which is doing load balancing more
4978 * next_balance will be updated only when there is a need.
4979 * When the cpu is attached to null domain for ex, it will not be
4982 if (likely(update_next_balance
))
4983 rq
->next_balance
= next_balance
;
4988 * In CONFIG_NO_HZ case, the idle balance kickee will do the
4989 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4991 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
)
4993 struct rq
*this_rq
= cpu_rq(this_cpu
);
4997 if (idle
!= CPU_IDLE
||
4998 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
5001 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
5002 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
5006 * If this cpu gets work to do, stop the load balancing
5007 * work being done for other cpus. Next load
5008 * balancing owner will pick it up.
5013 raw_spin_lock_irq(&this_rq
->lock
);
5014 update_rq_clock(this_rq
);
5015 update_idle_cpu_load(this_rq
);
5016 raw_spin_unlock_irq(&this_rq
->lock
);
5018 rebalance_domains(balance_cpu
, CPU_IDLE
);
5020 rq
= cpu_rq(balance_cpu
);
5021 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
5022 this_rq
->next_balance
= rq
->next_balance
;
5024 nohz
.next_balance
= this_rq
->next_balance
;
5026 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
5030 * Current heuristic for kicking the idle load balancer in the presence
5031 * of an idle cpu is the system.
5032 * - This rq has more than one task.
5033 * - At any scheduler domain level, this cpu's scheduler group has multiple
5034 * busy cpu's exceeding the group's power.
5035 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5036 * domain span are idle.
5038 static inline int nohz_kick_needed(struct rq
*rq
, int cpu
)
5040 unsigned long now
= jiffies
;
5041 struct sched_domain
*sd
;
5043 if (unlikely(idle_cpu(cpu
)))
5047 * We may be recently in ticked or tickless idle mode. At the first
5048 * busy tick after returning from idle, we will update the busy stats.
5050 set_cpu_sd_state_busy();
5051 clear_nohz_tick_stopped(cpu
);
5054 * None are in tickless mode and hence no need for NOHZ idle load
5057 if (likely(!atomic_read(&nohz
.nr_cpus
)))
5060 if (time_before(now
, nohz
.next_balance
))
5063 if (rq
->nr_running
>= 2)
5067 for_each_domain(cpu
, sd
) {
5068 struct sched_group
*sg
= sd
->groups
;
5069 struct sched_group_power
*sgp
= sg
->sgp
;
5070 int nr_busy
= atomic_read(&sgp
->nr_busy_cpus
);
5072 if (sd
->flags
& SD_SHARE_PKG_RESOURCES
&& nr_busy
> 1)
5073 goto need_kick_unlock
;
5075 if (sd
->flags
& SD_ASYM_PACKING
&& nr_busy
!= sg
->group_weight
5076 && (cpumask_first_and(nohz
.idle_cpus_mask
,
5077 sched_domain_span(sd
)) < cpu
))
5078 goto need_kick_unlock
;
5080 if (!(sd
->flags
& (SD_SHARE_PKG_RESOURCES
| SD_ASYM_PACKING
)))
5092 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
) { }
5096 * run_rebalance_domains is triggered when needed from the scheduler tick.
5097 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5099 static void run_rebalance_domains(struct softirq_action
*h
)
5101 int this_cpu
= smp_processor_id();
5102 struct rq
*this_rq
= cpu_rq(this_cpu
);
5103 enum cpu_idle_type idle
= this_rq
->idle_balance
?
5104 CPU_IDLE
: CPU_NOT_IDLE
;
5106 rebalance_domains(this_cpu
, idle
);
5109 * If this cpu has a pending nohz_balance_kick, then do the
5110 * balancing on behalf of the other idle cpus whose ticks are
5113 nohz_idle_balance(this_cpu
, idle
);
5116 static inline int on_null_domain(int cpu
)
5118 return !rcu_dereference_sched(cpu_rq(cpu
)->sd
);
5122 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5124 void trigger_load_balance(struct rq
*rq
, int cpu
)
5126 /* Don't need to rebalance while attached to NULL domain */
5127 if (time_after_eq(jiffies
, rq
->next_balance
) &&
5128 likely(!on_null_domain(cpu
)))
5129 raise_softirq(SCHED_SOFTIRQ
);
5131 if (nohz_kick_needed(rq
, cpu
) && likely(!on_null_domain(cpu
)))
5132 nohz_balancer_kick(cpu
);
5136 static void rq_online_fair(struct rq
*rq
)
5141 static void rq_offline_fair(struct rq
*rq
)
5146 #endif /* CONFIG_SMP */
5149 * scheduler tick hitting a task of our scheduling class:
5151 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
5153 struct cfs_rq
*cfs_rq
;
5154 struct sched_entity
*se
= &curr
->se
;
5156 for_each_sched_entity(se
) {
5157 cfs_rq
= cfs_rq_of(se
);
5158 entity_tick(cfs_rq
, se
, queued
);
5163 * called on fork with the child task as argument from the parent's context
5164 * - child not yet on the tasklist
5165 * - preemption disabled
5167 static void task_fork_fair(struct task_struct
*p
)
5169 struct cfs_rq
*cfs_rq
;
5170 struct sched_entity
*se
= &p
->se
, *curr
;
5171 int this_cpu
= smp_processor_id();
5172 struct rq
*rq
= this_rq();
5173 unsigned long flags
;
5175 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5177 update_rq_clock(rq
);
5179 cfs_rq
= task_cfs_rq(current
);
5180 curr
= cfs_rq
->curr
;
5182 if (unlikely(task_cpu(p
) != this_cpu
)) {
5184 __set_task_cpu(p
, this_cpu
);
5188 update_curr(cfs_rq
);
5191 se
->vruntime
= curr
->vruntime
;
5192 place_entity(cfs_rq
, se
, 1);
5194 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
5196 * Upon rescheduling, sched_class::put_prev_task() will place
5197 * 'current' within the tree based on its new key value.
5199 swap(curr
->vruntime
, se
->vruntime
);
5200 resched_task(rq
->curr
);
5203 se
->vruntime
-= cfs_rq
->min_vruntime
;
5205 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5209 * Priority of the task has changed. Check to see if we preempt
5213 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
5219 * Reschedule if we are currently running on this runqueue and
5220 * our priority decreased, or if we are not currently running on
5221 * this runqueue and our priority is higher than the current's
5223 if (rq
->curr
== p
) {
5224 if (p
->prio
> oldprio
)
5225 resched_task(rq
->curr
);
5227 check_preempt_curr(rq
, p
, 0);
5230 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
5232 struct sched_entity
*se
= &p
->se
;
5233 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5236 * Ensure the task's vruntime is normalized, so that when its
5237 * switched back to the fair class the enqueue_entity(.flags=0) will
5238 * do the right thing.
5240 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5241 * have normalized the vruntime, if it was !on_rq, then only when
5242 * the task is sleeping will it still have non-normalized vruntime.
5244 if (!se
->on_rq
&& p
->state
!= TASK_RUNNING
) {
5246 * Fix up our vruntime so that the current sleep doesn't
5247 * cause 'unlimited' sleep bonus.
5249 place_entity(cfs_rq
, se
, 0);
5250 se
->vruntime
-= cfs_rq
->min_vruntime
;
5255 * We switched to the sched_fair class.
5257 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
5263 * We were most likely switched from sched_rt, so
5264 * kick off the schedule if running, otherwise just see
5265 * if we can still preempt the current task.
5268 resched_task(rq
->curr
);
5270 check_preempt_curr(rq
, p
, 0);
5273 /* Account for a task changing its policy or group.
5275 * This routine is mostly called to set cfs_rq->curr field when a task
5276 * migrates between groups/classes.
5278 static void set_curr_task_fair(struct rq
*rq
)
5280 struct sched_entity
*se
= &rq
->curr
->se
;
5282 for_each_sched_entity(se
) {
5283 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5285 set_next_entity(cfs_rq
, se
);
5286 /* ensure bandwidth has been allocated on our new cfs_rq */
5287 account_cfs_rq_runtime(cfs_rq
, 0);
5291 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
5293 cfs_rq
->tasks_timeline
= RB_ROOT
;
5294 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
5295 #ifndef CONFIG_64BIT
5296 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
5300 #ifdef CONFIG_FAIR_GROUP_SCHED
5301 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
5304 * If the task was not on the rq at the time of this cgroup movement
5305 * it must have been asleep, sleeping tasks keep their ->vruntime
5306 * absolute on their old rq until wakeup (needed for the fair sleeper
5307 * bonus in place_entity()).
5309 * If it was on the rq, we've just 'preempted' it, which does convert
5310 * ->vruntime to a relative base.
5312 * Make sure both cases convert their relative position when migrating
5313 * to another cgroup's rq. This does somewhat interfere with the
5314 * fair sleeper stuff for the first placement, but who cares.
5317 * When !on_rq, vruntime of the task has usually NOT been normalized.
5318 * But there are some cases where it has already been normalized:
5320 * - Moving a forked child which is waiting for being woken up by
5321 * wake_up_new_task().
5322 * - Moving a task which has been woken up by try_to_wake_up() and
5323 * waiting for actually being woken up by sched_ttwu_pending().
5325 * To prevent boost or penalty in the new cfs_rq caused by delta
5326 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5328 if (!on_rq
&& (!p
->se
.sum_exec_runtime
|| p
->state
== TASK_WAKING
))
5332 p
->se
.vruntime
-= cfs_rq_of(&p
->se
)->min_vruntime
;
5333 set_task_rq(p
, task_cpu(p
));
5335 p
->se
.vruntime
+= cfs_rq_of(&p
->se
)->min_vruntime
;
5338 void free_fair_sched_group(struct task_group
*tg
)
5342 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
5344 for_each_possible_cpu(i
) {
5346 kfree(tg
->cfs_rq
[i
]);
5355 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
5357 struct cfs_rq
*cfs_rq
;
5358 struct sched_entity
*se
;
5361 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
5364 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
5368 tg
->shares
= NICE_0_LOAD
;
5370 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
5372 for_each_possible_cpu(i
) {
5373 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
5374 GFP_KERNEL
, cpu_to_node(i
));
5378 se
= kzalloc_node(sizeof(struct sched_entity
),
5379 GFP_KERNEL
, cpu_to_node(i
));
5383 init_cfs_rq(cfs_rq
);
5384 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
5395 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
5397 struct rq
*rq
= cpu_rq(cpu
);
5398 unsigned long flags
;
5401 * Only empty task groups can be destroyed; so we can speculatively
5402 * check on_list without danger of it being re-added.
5404 if (!tg
->cfs_rq
[cpu
]->on_list
)
5407 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5408 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
5409 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5412 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
5413 struct sched_entity
*se
, int cpu
,
5414 struct sched_entity
*parent
)
5416 struct rq
*rq
= cpu_rq(cpu
);
5421 /* allow initial update_cfs_load() to truncate */
5422 cfs_rq
->load_stamp
= 1;
5424 init_cfs_rq_runtime(cfs_rq
);
5426 tg
->cfs_rq
[cpu
] = cfs_rq
;
5429 /* se could be NULL for root_task_group */
5434 se
->cfs_rq
= &rq
->cfs
;
5436 se
->cfs_rq
= parent
->my_q
;
5439 update_load_set(&se
->load
, 0);
5440 se
->parent
= parent
;
5443 static DEFINE_MUTEX(shares_mutex
);
5445 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
5448 unsigned long flags
;
5451 * We can't change the weight of the root cgroup.
5456 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
5458 mutex_lock(&shares_mutex
);
5459 if (tg
->shares
== shares
)
5462 tg
->shares
= shares
;
5463 for_each_possible_cpu(i
) {
5464 struct rq
*rq
= cpu_rq(i
);
5465 struct sched_entity
*se
;
5468 /* Propagate contribution to hierarchy */
5469 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5470 for_each_sched_entity(se
)
5471 update_cfs_shares(group_cfs_rq(se
));
5472 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5476 mutex_unlock(&shares_mutex
);
5479 #else /* CONFIG_FAIR_GROUP_SCHED */
5481 void free_fair_sched_group(struct task_group
*tg
) { }
5483 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
5488 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
5490 #endif /* CONFIG_FAIR_GROUP_SCHED */
5493 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
5495 struct sched_entity
*se
= &task
->se
;
5496 unsigned int rr_interval
= 0;
5499 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5502 if (rq
->cfs
.load
.weight
)
5503 rr_interval
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5509 * All the scheduling class methods:
5511 const struct sched_class fair_sched_class
= {
5512 .next
= &idle_sched_class
,
5513 .enqueue_task
= enqueue_task_fair
,
5514 .dequeue_task
= dequeue_task_fair
,
5515 .yield_task
= yield_task_fair
,
5516 .yield_to_task
= yield_to_task_fair
,
5518 .check_preempt_curr
= check_preempt_wakeup
,
5520 .pick_next_task
= pick_next_task_fair
,
5521 .put_prev_task
= put_prev_task_fair
,
5524 .select_task_rq
= select_task_rq_fair
,
5526 .rq_online
= rq_online_fair
,
5527 .rq_offline
= rq_offline_fair
,
5529 .task_waking
= task_waking_fair
,
5532 .set_curr_task
= set_curr_task_fair
,
5533 .task_tick
= task_tick_fair
,
5534 .task_fork
= task_fork_fair
,
5536 .prio_changed
= prio_changed_fair
,
5537 .switched_from
= switched_from_fair
,
5538 .switched_to
= switched_to_fair
,
5540 .get_rr_interval
= get_rr_interval_fair
,
5542 #ifdef CONFIG_FAIR_GROUP_SCHED
5543 .task_move_group
= task_move_group_fair
,
5547 #ifdef CONFIG_SCHED_DEBUG
5548 void print_cfs_stats(struct seq_file
*m
, int cpu
)
5550 struct cfs_rq
*cfs_rq
;
5553 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
5554 print_cfs_rq(m
, cpu
, cfs_rq
);
5559 __init
void init_sched_fair_class(void)
5562 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
5565 nohz
.next_balance
= jiffies
;
5566 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
5567 cpu_notifier(sched_ilb_notifier
, 0);