2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency
= 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG
;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity
= 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency
= 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly
;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
93 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window
= 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
116 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
122 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
128 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling
) {
149 case SCHED_TUNABLESCALING_NONE
:
152 case SCHED_TUNABLESCALING_LINEAR
:
155 case SCHED_TUNABLESCALING_LOG
:
157 factor
= 1 + ilog2(cpus
);
164 static void update_sysctl(void)
166 unsigned int factor
= get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity
);
171 SET_SYSCTL(sched_latency
);
172 SET_SYSCTL(sched_wakeup_granularity
);
176 void sched_init_granularity(void)
181 #if BITS_PER_LONG == 32
182 # define WMULT_CONST (~0UL)
184 # define WMULT_CONST (1UL << 32)
187 #define WMULT_SHIFT 32
190 * Shift right and round:
192 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
195 * delta *= weight / lw
198 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
199 struct load_weight
*lw
)
204 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
205 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
206 * 2^SCHED_LOAD_RESOLUTION.
208 if (likely(weight
> (1UL << SCHED_LOAD_RESOLUTION
)))
209 tmp
= (u64
)delta_exec
* scale_load_down(weight
);
211 tmp
= (u64
)delta_exec
;
213 if (!lw
->inv_weight
) {
214 unsigned long w
= scale_load_down(lw
->weight
);
216 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
218 else if (unlikely(!w
))
219 lw
->inv_weight
= WMULT_CONST
;
221 lw
->inv_weight
= WMULT_CONST
/ w
;
225 * Check whether we'd overflow the 64-bit multiplication:
227 if (unlikely(tmp
> WMULT_CONST
))
228 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
231 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
233 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
237 const struct sched_class fair_sched_class
;
239 /**************************************************************
240 * CFS operations on generic schedulable entities:
243 #ifdef CONFIG_FAIR_GROUP_SCHED
245 /* cpu runqueue to which this cfs_rq is attached */
246 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
251 /* An entity is a task if it doesn't "own" a runqueue */
252 #define entity_is_task(se) (!se->my_q)
254 static inline struct task_struct
*task_of(struct sched_entity
*se
)
256 #ifdef CONFIG_SCHED_DEBUG
257 WARN_ON_ONCE(!entity_is_task(se
));
259 return container_of(se
, struct task_struct
, se
);
262 /* Walk up scheduling entities hierarchy */
263 #define for_each_sched_entity(se) \
264 for (; se; se = se->parent)
266 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
271 /* runqueue on which this entity is (to be) queued */
272 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
277 /* runqueue "owned" by this group */
278 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
283 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
288 if (!cfs_rq
->on_list
) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq
->tg
->parent
&&
296 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
297 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
298 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
300 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
301 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
305 /* We should have no load, but we need to update last_decay. */
306 update_cfs_rq_blocked_load(cfs_rq
, 0);
310 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
312 if (cfs_rq
->on_list
) {
313 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
318 /* Iterate thr' all leaf cfs_rq's on a runqueue */
319 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
320 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
322 /* Do the two (enqueued) entities belong to the same group ? */
324 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
326 if (se
->cfs_rq
== pse
->cfs_rq
)
332 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
337 /* return depth at which a sched entity is present in the hierarchy */
338 static inline int depth_se(struct sched_entity
*se
)
342 for_each_sched_entity(se
)
349 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
351 int se_depth
, pse_depth
;
354 * preemption test can be made between sibling entities who are in the
355 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
356 * both tasks until we find their ancestors who are siblings of common
360 /* First walk up until both entities are at same depth */
361 se_depth
= depth_se(*se
);
362 pse_depth
= depth_se(*pse
);
364 while (se_depth
> pse_depth
) {
366 *se
= parent_entity(*se
);
369 while (pse_depth
> se_depth
) {
371 *pse
= parent_entity(*pse
);
374 while (!is_same_group(*se
, *pse
)) {
375 *se
= parent_entity(*se
);
376 *pse
= parent_entity(*pse
);
380 #else /* !CONFIG_FAIR_GROUP_SCHED */
382 static inline struct task_struct
*task_of(struct sched_entity
*se
)
384 return container_of(se
, struct task_struct
, se
);
387 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
389 return container_of(cfs_rq
, struct rq
, cfs
);
392 #define entity_is_task(se) 1
394 #define for_each_sched_entity(se) \
395 for (; se; se = NULL)
397 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
399 return &task_rq(p
)->cfs
;
402 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
404 struct task_struct
*p
= task_of(se
);
405 struct rq
*rq
= task_rq(p
);
410 /* runqueue "owned" by this group */
411 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
416 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
420 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
424 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
425 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
433 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
439 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
454 s64 delta
= (s64
)(vruntime
- max_vruntime
);
456 max_vruntime
= vruntime
;
461 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
463 s64 delta
= (s64
)(vruntime
- min_vruntime
);
465 min_vruntime
= vruntime
;
470 static inline int entity_before(struct sched_entity
*a
,
471 struct sched_entity
*b
)
473 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
476 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
478 u64 vruntime
= cfs_rq
->min_vruntime
;
481 vruntime
= cfs_rq
->curr
->vruntime
;
483 if (cfs_rq
->rb_leftmost
) {
484 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
489 vruntime
= se
->vruntime
;
491 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
498 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
507 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
508 struct rb_node
*parent
= NULL
;
509 struct sched_entity
*entry
;
513 * Find the right place in the rbtree:
517 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se
, entry
)) {
523 link
= &parent
->rb_left
;
525 link
= &parent
->rb_right
;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq
->rb_leftmost
= &se
->run_node
;
537 rb_link_node(&se
->run_node
, parent
, link
);
538 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
541 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
543 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
544 struct rb_node
*next_node
;
546 next_node
= rb_next(&se
->run_node
);
547 cfs_rq
->rb_leftmost
= next_node
;
550 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
553 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
555 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
560 return rb_entry(left
, struct sched_entity
, run_node
);
563 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
565 struct rb_node
*next
= rb_next(&se
->run_node
);
570 return rb_entry(next
, struct sched_entity
, run_node
);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
576 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
581 return rb_entry(last
, struct sched_entity
, run_node
);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
589 void __user
*buffer
, size_t *lenp
,
592 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
593 int factor
= get_update_sysctl_factor();
598 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
599 sysctl_sched_min_granularity
);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity
);
604 WRT_SYSCTL(sched_latency
);
605 WRT_SYSCTL(sched_wakeup_granularity
);
615 static inline unsigned long
616 calc_delta_fair(unsigned long delta
, struct sched_entity
*se
)
618 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
619 delta
= calc_delta_mine(delta
, NICE_0_LOAD
, &se
->load
);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64
__sched_period(unsigned long nr_running
)
634 u64 period
= sysctl_sched_latency
;
635 unsigned long nr_latency
= sched_nr_latency
;
637 if (unlikely(nr_running
> nr_latency
)) {
638 period
= sysctl_sched_min_granularity
;
639 period
*= nr_running
;
646 * We calculate the wall-time slice from the period by taking a part
647 * proportional to the weight.
651 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
653 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
655 for_each_sched_entity(se
) {
656 struct load_weight
*load
;
657 struct load_weight lw
;
659 cfs_rq
= cfs_rq_of(se
);
660 load
= &cfs_rq
->load
;
662 if (unlikely(!se
->on_rq
)) {
665 update_load_add(&lw
, se
->load
.weight
);
668 slice
= calc_delta_mine(slice
, se
->load
.weight
, load
);
674 * We calculate the vruntime slice of a to-be-inserted task.
678 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
680 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
684 static inline void __update_task_entity_contrib(struct sched_entity
*se
);
686 /* Give new task start runnable values to heavy its load in infant time */
687 void init_task_runnable_average(struct task_struct
*p
)
691 p
->se
.avg
.decay_count
= 0;
692 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
693 p
->se
.avg
.runnable_avg_sum
= slice
;
694 p
->se
.avg
.runnable_avg_period
= slice
;
695 __update_task_entity_contrib(&p
->se
);
698 void init_task_runnable_average(struct task_struct
*p
)
704 * Update the current task's runtime statistics. Skip current tasks that
705 * are not in our scheduling class.
708 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
709 unsigned long delta_exec
)
711 unsigned long delta_exec_weighted
;
713 schedstat_set(curr
->statistics
.exec_max
,
714 max((u64
)delta_exec
, curr
->statistics
.exec_max
));
716 curr
->sum_exec_runtime
+= delta_exec
;
717 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
718 delta_exec_weighted
= calc_delta_fair(delta_exec
, curr
);
720 curr
->vruntime
+= delta_exec_weighted
;
721 update_min_vruntime(cfs_rq
);
724 static void update_curr(struct cfs_rq
*cfs_rq
)
726 struct sched_entity
*curr
= cfs_rq
->curr
;
727 u64 now
= rq_clock_task(rq_of(cfs_rq
));
728 unsigned long delta_exec
;
734 * Get the amount of time the current task was running
735 * since the last time we changed load (this cannot
736 * overflow on 32 bits):
738 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
742 __update_curr(cfs_rq
, curr
, delta_exec
);
743 curr
->exec_start
= now
;
745 if (entity_is_task(curr
)) {
746 struct task_struct
*curtask
= task_of(curr
);
748 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
749 cpuacct_charge(curtask
, delta_exec
);
750 account_group_exec_runtime(curtask
, delta_exec
);
753 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
757 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
759 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
763 * Task is being enqueued - update stats:
765 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
768 * Are we enqueueing a waiting task? (for current tasks
769 * a dequeue/enqueue event is a NOP)
771 if (se
!= cfs_rq
->curr
)
772 update_stats_wait_start(cfs_rq
, se
);
776 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
778 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
779 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
780 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
781 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
782 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
783 #ifdef CONFIG_SCHEDSTATS
784 if (entity_is_task(se
)) {
785 trace_sched_stat_wait(task_of(se
),
786 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
789 schedstat_set(se
->statistics
.wait_start
, 0);
793 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
796 * Mark the end of the wait period if dequeueing a
799 if (se
!= cfs_rq
->curr
)
800 update_stats_wait_end(cfs_rq
, se
);
804 * We are picking a new current task - update its stats:
807 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
810 * We are starting a new run period:
812 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
815 /**************************************************
816 * Scheduling class queueing methods:
819 #ifdef CONFIG_NUMA_BALANCING
821 * numa task sample period in ms
823 unsigned int sysctl_numa_balancing_scan_period_min
= 100;
824 unsigned int sysctl_numa_balancing_scan_period_max
= 100*50;
825 unsigned int sysctl_numa_balancing_scan_period_reset
= 100*600;
827 /* Portion of address space to scan in MB */
828 unsigned int sysctl_numa_balancing_scan_size
= 256;
830 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
831 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
833 static void task_numa_placement(struct task_struct
*p
)
837 if (!p
->mm
) /* for example, ksmd faulting in a user's mm */
839 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
840 if (p
->numa_scan_seq
== seq
)
842 p
->numa_scan_seq
= seq
;
844 /* FIXME: Scheduling placement policy hints go here */
848 * Got a PROT_NONE fault for a page on @node.
850 void task_numa_fault(int node
, int pages
, bool migrated
)
852 struct task_struct
*p
= current
;
854 if (!sched_feat_numa(NUMA
))
857 /* FIXME: Allocate task-specific structure for placement policy here */
860 * If pages are properly placed (did not migrate) then scan slower.
861 * This is reset periodically in case of phase changes
864 p
->numa_scan_period
= min(sysctl_numa_balancing_scan_period_max
,
865 p
->numa_scan_period
+ jiffies_to_msecs(10));
867 task_numa_placement(p
);
870 static void reset_ptenuma_scan(struct task_struct
*p
)
872 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
873 p
->mm
->numa_scan_offset
= 0;
877 * The expensive part of numa migration is done from task_work context.
878 * Triggered from task_tick_numa().
880 void task_numa_work(struct callback_head
*work
)
882 unsigned long migrate
, next_scan
, now
= jiffies
;
883 struct task_struct
*p
= current
;
884 struct mm_struct
*mm
= p
->mm
;
885 struct vm_area_struct
*vma
;
886 unsigned long start
, end
;
889 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
891 work
->next
= work
; /* protect against double add */
893 * Who cares about NUMA placement when they're dying.
895 * NOTE: make sure not to dereference p->mm before this check,
896 * exit_task_work() happens _after_ exit_mm() so we could be called
897 * without p->mm even though we still had it when we enqueued this
900 if (p
->flags
& PF_EXITING
)
904 * We do not care about task placement until a task runs on a node
905 * other than the first one used by the address space. This is
906 * largely because migrations are driven by what CPU the task
907 * is running on. If it's never scheduled on another node, it'll
908 * not migrate so why bother trapping the fault.
910 if (mm
->first_nid
== NUMA_PTE_SCAN_INIT
)
911 mm
->first_nid
= numa_node_id();
912 if (mm
->first_nid
!= NUMA_PTE_SCAN_ACTIVE
) {
913 /* Are we running on a new node yet? */
914 if (numa_node_id() == mm
->first_nid
&&
915 !sched_feat_numa(NUMA_FORCE
))
918 mm
->first_nid
= NUMA_PTE_SCAN_ACTIVE
;
922 * Reset the scan period if enough time has gone by. Objective is that
923 * scanning will be reduced if pages are properly placed. As tasks
924 * can enter different phases this needs to be re-examined. Lacking
925 * proper tracking of reference behaviour, this blunt hammer is used.
927 migrate
= mm
->numa_next_reset
;
928 if (time_after(now
, migrate
)) {
929 p
->numa_scan_period
= sysctl_numa_balancing_scan_period_min
;
930 next_scan
= now
+ msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset
);
931 xchg(&mm
->numa_next_reset
, next_scan
);
935 * Enforce maximal scan/migration frequency..
937 migrate
= mm
->numa_next_scan
;
938 if (time_before(now
, migrate
))
941 if (p
->numa_scan_period
== 0)
942 p
->numa_scan_period
= sysctl_numa_balancing_scan_period_min
;
944 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
945 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
949 * Do not set pte_numa if the current running node is rate-limited.
950 * This loses statistics on the fault but if we are unwilling to
951 * migrate to this node, it is less likely we can do useful work
953 if (migrate_ratelimited(numa_node_id()))
956 start
= mm
->numa_scan_offset
;
957 pages
= sysctl_numa_balancing_scan_size
;
958 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
962 down_read(&mm
->mmap_sem
);
963 vma
= find_vma(mm
, start
);
965 reset_ptenuma_scan(p
);
969 for (; vma
; vma
= vma
->vm_next
) {
970 if (!vma_migratable(vma
))
973 /* Skip small VMAs. They are not likely to be of relevance */
974 if (vma
->vm_end
- vma
->vm_start
< HPAGE_SIZE
)
978 start
= max(start
, vma
->vm_start
);
979 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
980 end
= min(end
, vma
->vm_end
);
981 pages
-= change_prot_numa(vma
, start
, end
);
986 } while (end
!= vma
->vm_end
);
991 * It is possible to reach the end of the VMA list but the last few VMAs are
992 * not guaranteed to the vma_migratable. If they are not, we would find the
993 * !migratable VMA on the next scan but not reset the scanner to the start
997 mm
->numa_scan_offset
= start
;
999 reset_ptenuma_scan(p
);
1000 up_read(&mm
->mmap_sem
);
1004 * Drive the periodic memory faults..
1006 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1008 struct callback_head
*work
= &curr
->numa_work
;
1012 * We don't care about NUMA placement if we don't have memory.
1014 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
1018 * Using runtime rather than walltime has the dual advantage that
1019 * we (mostly) drive the selection from busy threads and that the
1020 * task needs to have done some actual work before we bother with
1023 now
= curr
->se
.sum_exec_runtime
;
1024 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
1026 if (now
- curr
->node_stamp
> period
) {
1027 if (!curr
->node_stamp
)
1028 curr
->numa_scan_period
= sysctl_numa_balancing_scan_period_min
;
1029 curr
->node_stamp
= now
;
1031 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
1032 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
1033 task_work_add(curr
, work
, true);
1038 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1041 #endif /* CONFIG_NUMA_BALANCING */
1044 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1046 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
1047 if (!parent_entity(se
))
1048 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1050 if (entity_is_task(se
))
1051 list_add(&se
->group_node
, &rq_of(cfs_rq
)->cfs_tasks
);
1053 cfs_rq
->nr_running
++;
1057 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1059 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
1060 if (!parent_entity(se
))
1061 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1062 if (entity_is_task(se
))
1063 list_del_init(&se
->group_node
);
1064 cfs_rq
->nr_running
--;
1067 #ifdef CONFIG_FAIR_GROUP_SCHED
1069 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
1074 * Use this CPU's actual weight instead of the last load_contribution
1075 * to gain a more accurate current total weight. See
1076 * update_cfs_rq_load_contribution().
1078 tg_weight
= atomic_long_read(&tg
->load_avg
);
1079 tg_weight
-= cfs_rq
->tg_load_contrib
;
1080 tg_weight
+= cfs_rq
->load
.weight
;
1085 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1087 long tg_weight
, load
, shares
;
1089 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
1090 load
= cfs_rq
->load
.weight
;
1092 shares
= (tg
->shares
* load
);
1094 shares
/= tg_weight
;
1096 if (shares
< MIN_SHARES
)
1097 shares
= MIN_SHARES
;
1098 if (shares
> tg
->shares
)
1099 shares
= tg
->shares
;
1103 # else /* CONFIG_SMP */
1104 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1108 # endif /* CONFIG_SMP */
1109 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
1110 unsigned long weight
)
1113 /* commit outstanding execution time */
1114 if (cfs_rq
->curr
== se
)
1115 update_curr(cfs_rq
);
1116 account_entity_dequeue(cfs_rq
, se
);
1119 update_load_set(&se
->load
, weight
);
1122 account_entity_enqueue(cfs_rq
, se
);
1125 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
1127 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
1129 struct task_group
*tg
;
1130 struct sched_entity
*se
;
1134 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
1135 if (!se
|| throttled_hierarchy(cfs_rq
))
1138 if (likely(se
->load
.weight
== tg
->shares
))
1141 shares
= calc_cfs_shares(cfs_rq
, tg
);
1143 reweight_entity(cfs_rq_of(se
), se
, shares
);
1145 #else /* CONFIG_FAIR_GROUP_SCHED */
1146 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
1149 #endif /* CONFIG_FAIR_GROUP_SCHED */
1153 * We choose a half-life close to 1 scheduling period.
1154 * Note: The tables below are dependent on this value.
1156 #define LOAD_AVG_PERIOD 32
1157 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1158 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1160 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1161 static const u32 runnable_avg_yN_inv
[] = {
1162 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1163 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1164 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1165 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1166 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1167 0x85aac367, 0x82cd8698,
1171 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1172 * over-estimates when re-combining.
1174 static const u32 runnable_avg_yN_sum
[] = {
1175 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1176 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1177 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1182 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1184 static __always_inline u64
decay_load(u64 val
, u64 n
)
1186 unsigned int local_n
;
1190 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
1193 /* after bounds checking we can collapse to 32-bit */
1197 * As y^PERIOD = 1/2, we can combine
1198 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1199 * With a look-up table which covers k^n (n<PERIOD)
1201 * To achieve constant time decay_load.
1203 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
1204 val
>>= local_n
/ LOAD_AVG_PERIOD
;
1205 local_n
%= LOAD_AVG_PERIOD
;
1208 val
*= runnable_avg_yN_inv
[local_n
];
1209 /* We don't use SRR here since we always want to round down. */
1214 * For updates fully spanning n periods, the contribution to runnable
1215 * average will be: \Sum 1024*y^n
1217 * We can compute this reasonably efficiently by combining:
1218 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1220 static u32
__compute_runnable_contrib(u64 n
)
1224 if (likely(n
<= LOAD_AVG_PERIOD
))
1225 return runnable_avg_yN_sum
[n
];
1226 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
1227 return LOAD_AVG_MAX
;
1229 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1231 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1232 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
1234 n
-= LOAD_AVG_PERIOD
;
1235 } while (n
> LOAD_AVG_PERIOD
);
1237 contrib
= decay_load(contrib
, n
);
1238 return contrib
+ runnable_avg_yN_sum
[n
];
1242 * We can represent the historical contribution to runnable average as the
1243 * coefficients of a geometric series. To do this we sub-divide our runnable
1244 * history into segments of approximately 1ms (1024us); label the segment that
1245 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1247 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1249 * (now) (~1ms ago) (~2ms ago)
1251 * Let u_i denote the fraction of p_i that the entity was runnable.
1253 * We then designate the fractions u_i as our co-efficients, yielding the
1254 * following representation of historical load:
1255 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1257 * We choose y based on the with of a reasonably scheduling period, fixing:
1260 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1261 * approximately half as much as the contribution to load within the last ms
1264 * When a period "rolls over" and we have new u_0`, multiplying the previous
1265 * sum again by y is sufficient to update:
1266 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1267 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1269 static __always_inline
int __update_entity_runnable_avg(u64 now
,
1270 struct sched_avg
*sa
,
1274 u32 runnable_contrib
;
1275 int delta_w
, decayed
= 0;
1277 delta
= now
- sa
->last_runnable_update
;
1279 * This should only happen when time goes backwards, which it
1280 * unfortunately does during sched clock init when we swap over to TSC.
1282 if ((s64
)delta
< 0) {
1283 sa
->last_runnable_update
= now
;
1288 * Use 1024ns as the unit of measurement since it's a reasonable
1289 * approximation of 1us and fast to compute.
1294 sa
->last_runnable_update
= now
;
1296 /* delta_w is the amount already accumulated against our next period */
1297 delta_w
= sa
->runnable_avg_period
% 1024;
1298 if (delta
+ delta_w
>= 1024) {
1299 /* period roll-over */
1303 * Now that we know we're crossing a period boundary, figure
1304 * out how much from delta we need to complete the current
1305 * period and accrue it.
1307 delta_w
= 1024 - delta_w
;
1309 sa
->runnable_avg_sum
+= delta_w
;
1310 sa
->runnable_avg_period
+= delta_w
;
1314 /* Figure out how many additional periods this update spans */
1315 periods
= delta
/ 1024;
1318 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
1320 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
1323 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1324 runnable_contrib
= __compute_runnable_contrib(periods
);
1326 sa
->runnable_avg_sum
+= runnable_contrib
;
1327 sa
->runnable_avg_period
+= runnable_contrib
;
1330 /* Remainder of delta accrued against u_0` */
1332 sa
->runnable_avg_sum
+= delta
;
1333 sa
->runnable_avg_period
+= delta
;
1338 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1339 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
1341 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1342 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
1344 decays
-= se
->avg
.decay_count
;
1348 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
1349 se
->avg
.decay_count
= 0;
1354 #ifdef CONFIG_FAIR_GROUP_SCHED
1355 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
1358 struct task_group
*tg
= cfs_rq
->tg
;
1361 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
1362 tg_contrib
-= cfs_rq
->tg_load_contrib
;
1364 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
1365 atomic_long_add(tg_contrib
, &tg
->load_avg
);
1366 cfs_rq
->tg_load_contrib
+= tg_contrib
;
1371 * Aggregate cfs_rq runnable averages into an equivalent task_group
1372 * representation for computing load contributions.
1374 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
1375 struct cfs_rq
*cfs_rq
)
1377 struct task_group
*tg
= cfs_rq
->tg
;
1380 /* The fraction of a cpu used by this cfs_rq */
1381 contrib
= div_u64(sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
1382 sa
->runnable_avg_period
+ 1);
1383 contrib
-= cfs_rq
->tg_runnable_contrib
;
1385 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
1386 atomic_add(contrib
, &tg
->runnable_avg
);
1387 cfs_rq
->tg_runnable_contrib
+= contrib
;
1391 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
1393 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
1394 struct task_group
*tg
= cfs_rq
->tg
;
1399 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
1400 se
->avg
.load_avg_contrib
= div_u64(contrib
,
1401 atomic_long_read(&tg
->load_avg
) + 1);
1404 * For group entities we need to compute a correction term in the case
1405 * that they are consuming <1 cpu so that we would contribute the same
1406 * load as a task of equal weight.
1408 * Explicitly co-ordinating this measurement would be expensive, but
1409 * fortunately the sum of each cpus contribution forms a usable
1410 * lower-bound on the true value.
1412 * Consider the aggregate of 2 contributions. Either they are disjoint
1413 * (and the sum represents true value) or they are disjoint and we are
1414 * understating by the aggregate of their overlap.
1416 * Extending this to N cpus, for a given overlap, the maximum amount we
1417 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1418 * cpus that overlap for this interval and w_i is the interval width.
1420 * On a small machine; the first term is well-bounded which bounds the
1421 * total error since w_i is a subset of the period. Whereas on a
1422 * larger machine, while this first term can be larger, if w_i is the
1423 * of consequential size guaranteed to see n_i*w_i quickly converge to
1424 * our upper bound of 1-cpu.
1426 runnable_avg
= atomic_read(&tg
->runnable_avg
);
1427 if (runnable_avg
< NICE_0_LOAD
) {
1428 se
->avg
.load_avg_contrib
*= runnable_avg
;
1429 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
1433 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
1434 int force_update
) {}
1435 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
1436 struct cfs_rq
*cfs_rq
) {}
1437 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
1440 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
1444 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1445 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
1446 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
1447 se
->avg
.load_avg_contrib
= scale_load(contrib
);
1450 /* Compute the current contribution to load_avg by se, return any delta */
1451 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
1453 long old_contrib
= se
->avg
.load_avg_contrib
;
1455 if (entity_is_task(se
)) {
1456 __update_task_entity_contrib(se
);
1458 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
1459 __update_group_entity_contrib(se
);
1462 return se
->avg
.load_avg_contrib
- old_contrib
;
1465 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
1468 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
1469 cfs_rq
->blocked_load_avg
-= load_contrib
;
1471 cfs_rq
->blocked_load_avg
= 0;
1474 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
1476 /* Update a sched_entity's runnable average */
1477 static inline void update_entity_load_avg(struct sched_entity
*se
,
1480 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1485 * For a group entity we need to use their owned cfs_rq_clock_task() in
1486 * case they are the parent of a throttled hierarchy.
1488 if (entity_is_task(se
))
1489 now
= cfs_rq_clock_task(cfs_rq
);
1491 now
= cfs_rq_clock_task(group_cfs_rq(se
));
1493 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
1496 contrib_delta
= __update_entity_load_avg_contrib(se
);
1502 cfs_rq
->runnable_load_avg
+= contrib_delta
;
1504 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
1508 * Decay the load contributed by all blocked children and account this so that
1509 * their contribution may appropriately discounted when they wake up.
1511 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
1513 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
1516 decays
= now
- cfs_rq
->last_decay
;
1517 if (!decays
&& !force_update
)
1520 if (atomic64_read(&cfs_rq
->removed_load
)) {
1521 u64 removed_load
= atomic64_xchg(&cfs_rq
->removed_load
, 0);
1522 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
1526 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
1528 atomic64_add(decays
, &cfs_rq
->decay_counter
);
1529 cfs_rq
->last_decay
= now
;
1532 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
1535 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
1537 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
1538 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
1541 /* Add the load generated by se into cfs_rq's child load-average */
1542 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
1543 struct sched_entity
*se
,
1547 * We track migrations using entity decay_count <= 0, on a wake-up
1548 * migration we use a negative decay count to track the remote decays
1549 * accumulated while sleeping.
1551 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1552 * are seen by enqueue_entity_load_avg() as a migration with an already
1553 * constructed load_avg_contrib.
1555 if (unlikely(se
->avg
.decay_count
<= 0)) {
1556 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
1557 if (se
->avg
.decay_count
) {
1559 * In a wake-up migration we have to approximate the
1560 * time sleeping. This is because we can't synchronize
1561 * clock_task between the two cpus, and it is not
1562 * guaranteed to be read-safe. Instead, we can
1563 * approximate this using our carried decays, which are
1564 * explicitly atomically readable.
1566 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
1568 update_entity_load_avg(se
, 0);
1569 /* Indicate that we're now synchronized and on-rq */
1570 se
->avg
.decay_count
= 0;
1575 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1576 * would have made count negative); we must be careful to avoid
1577 * double-accounting blocked time after synchronizing decays.
1579 se
->avg
.last_runnable_update
+= __synchronize_entity_decay(se
)
1583 /* migrated tasks did not contribute to our blocked load */
1585 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
1586 update_entity_load_avg(se
, 0);
1589 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
1590 /* we force update consideration on load-balancer moves */
1591 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
1595 * Remove se's load from this cfs_rq child load-average, if the entity is
1596 * transitioning to a blocked state we track its projected decay using
1599 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
1600 struct sched_entity
*se
,
1603 update_entity_load_avg(se
, 1);
1604 /* we force update consideration on load-balancer moves */
1605 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
1607 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
1609 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
1610 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
1611 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1615 * Update the rq's load with the elapsed running time before entering
1616 * idle. if the last scheduled task is not a CFS task, idle_enter will
1617 * be the only way to update the runnable statistic.
1619 void idle_enter_fair(struct rq
*this_rq
)
1621 update_rq_runnable_avg(this_rq
, 1);
1625 * Update the rq's load with the elapsed idle time before a task is
1626 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1627 * be the only way to update the runnable statistic.
1629 void idle_exit_fair(struct rq
*this_rq
)
1631 update_rq_runnable_avg(this_rq
, 0);
1635 static inline void update_entity_load_avg(struct sched_entity
*se
,
1636 int update_cfs_rq
) {}
1637 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
1638 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
1639 struct sched_entity
*se
,
1641 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
1642 struct sched_entity
*se
,
1644 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
1645 int force_update
) {}
1648 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1650 #ifdef CONFIG_SCHEDSTATS
1651 struct task_struct
*tsk
= NULL
;
1653 if (entity_is_task(se
))
1656 if (se
->statistics
.sleep_start
) {
1657 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
1662 if (unlikely(delta
> se
->statistics
.sleep_max
))
1663 se
->statistics
.sleep_max
= delta
;
1665 se
->statistics
.sleep_start
= 0;
1666 se
->statistics
.sum_sleep_runtime
+= delta
;
1669 account_scheduler_latency(tsk
, delta
>> 10, 1);
1670 trace_sched_stat_sleep(tsk
, delta
);
1673 if (se
->statistics
.block_start
) {
1674 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
1679 if (unlikely(delta
> se
->statistics
.block_max
))
1680 se
->statistics
.block_max
= delta
;
1682 se
->statistics
.block_start
= 0;
1683 se
->statistics
.sum_sleep_runtime
+= delta
;
1686 if (tsk
->in_iowait
) {
1687 se
->statistics
.iowait_sum
+= delta
;
1688 se
->statistics
.iowait_count
++;
1689 trace_sched_stat_iowait(tsk
, delta
);
1692 trace_sched_stat_blocked(tsk
, delta
);
1695 * Blocking time is in units of nanosecs, so shift by
1696 * 20 to get a milliseconds-range estimation of the
1697 * amount of time that the task spent sleeping:
1699 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
1700 profile_hits(SLEEP_PROFILING
,
1701 (void *)get_wchan(tsk
),
1704 account_scheduler_latency(tsk
, delta
>> 10, 0);
1710 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1712 #ifdef CONFIG_SCHED_DEBUG
1713 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
1718 if (d
> 3*sysctl_sched_latency
)
1719 schedstat_inc(cfs_rq
, nr_spread_over
);
1724 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
1726 u64 vruntime
= cfs_rq
->min_vruntime
;
1729 * The 'current' period is already promised to the current tasks,
1730 * however the extra weight of the new task will slow them down a
1731 * little, place the new task so that it fits in the slot that
1732 * stays open at the end.
1734 if (initial
&& sched_feat(START_DEBIT
))
1735 vruntime
+= sched_vslice(cfs_rq
, se
);
1737 /* sleeps up to a single latency don't count. */
1739 unsigned long thresh
= sysctl_sched_latency
;
1742 * Halve their sleep time's effect, to allow
1743 * for a gentler effect of sleepers:
1745 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
1751 /* ensure we never gain time by being placed backwards. */
1752 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
1755 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
1758 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1761 * Update the normalized vruntime before updating min_vruntime
1762 * through callig update_curr().
1764 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
1765 se
->vruntime
+= cfs_rq
->min_vruntime
;
1768 * Update run-time statistics of the 'current'.
1770 update_curr(cfs_rq
);
1771 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
1772 account_entity_enqueue(cfs_rq
, se
);
1773 update_cfs_shares(cfs_rq
);
1775 if (flags
& ENQUEUE_WAKEUP
) {
1776 place_entity(cfs_rq
, se
, 0);
1777 enqueue_sleeper(cfs_rq
, se
);
1780 update_stats_enqueue(cfs_rq
, se
);
1781 check_spread(cfs_rq
, se
);
1782 if (se
!= cfs_rq
->curr
)
1783 __enqueue_entity(cfs_rq
, se
);
1786 if (cfs_rq
->nr_running
== 1) {
1787 list_add_leaf_cfs_rq(cfs_rq
);
1788 check_enqueue_throttle(cfs_rq
);
1792 static void __clear_buddies_last(struct sched_entity
*se
)
1794 for_each_sched_entity(se
) {
1795 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1796 if (cfs_rq
->last
== se
)
1797 cfs_rq
->last
= NULL
;
1803 static void __clear_buddies_next(struct sched_entity
*se
)
1805 for_each_sched_entity(se
) {
1806 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1807 if (cfs_rq
->next
== se
)
1808 cfs_rq
->next
= NULL
;
1814 static void __clear_buddies_skip(struct sched_entity
*se
)
1816 for_each_sched_entity(se
) {
1817 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1818 if (cfs_rq
->skip
== se
)
1819 cfs_rq
->skip
= NULL
;
1825 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1827 if (cfs_rq
->last
== se
)
1828 __clear_buddies_last(se
);
1830 if (cfs_rq
->next
== se
)
1831 __clear_buddies_next(se
);
1833 if (cfs_rq
->skip
== se
)
1834 __clear_buddies_skip(se
);
1837 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
1840 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1843 * Update run-time statistics of the 'current'.
1845 update_curr(cfs_rq
);
1846 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
1848 update_stats_dequeue(cfs_rq
, se
);
1849 if (flags
& DEQUEUE_SLEEP
) {
1850 #ifdef CONFIG_SCHEDSTATS
1851 if (entity_is_task(se
)) {
1852 struct task_struct
*tsk
= task_of(se
);
1854 if (tsk
->state
& TASK_INTERRUPTIBLE
)
1855 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
1856 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
1857 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
1862 clear_buddies(cfs_rq
, se
);
1864 if (se
!= cfs_rq
->curr
)
1865 __dequeue_entity(cfs_rq
, se
);
1867 account_entity_dequeue(cfs_rq
, se
);
1870 * Normalize the entity after updating the min_vruntime because the
1871 * update can refer to the ->curr item and we need to reflect this
1872 * movement in our normalized position.
1874 if (!(flags
& DEQUEUE_SLEEP
))
1875 se
->vruntime
-= cfs_rq
->min_vruntime
;
1877 /* return excess runtime on last dequeue */
1878 return_cfs_rq_runtime(cfs_rq
);
1880 update_min_vruntime(cfs_rq
);
1881 update_cfs_shares(cfs_rq
);
1885 * Preempt the current task with a newly woken task if needed:
1888 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
1890 unsigned long ideal_runtime
, delta_exec
;
1891 struct sched_entity
*se
;
1894 ideal_runtime
= sched_slice(cfs_rq
, curr
);
1895 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
1896 if (delta_exec
> ideal_runtime
) {
1897 resched_task(rq_of(cfs_rq
)->curr
);
1899 * The current task ran long enough, ensure it doesn't get
1900 * re-elected due to buddy favours.
1902 clear_buddies(cfs_rq
, curr
);
1907 * Ensure that a task that missed wakeup preemption by a
1908 * narrow margin doesn't have to wait for a full slice.
1909 * This also mitigates buddy induced latencies under load.
1911 if (delta_exec
< sysctl_sched_min_granularity
)
1914 se
= __pick_first_entity(cfs_rq
);
1915 delta
= curr
->vruntime
- se
->vruntime
;
1920 if (delta
> ideal_runtime
)
1921 resched_task(rq_of(cfs_rq
)->curr
);
1925 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1927 /* 'current' is not kept within the tree. */
1930 * Any task has to be enqueued before it get to execute on
1931 * a CPU. So account for the time it spent waiting on the
1934 update_stats_wait_end(cfs_rq
, se
);
1935 __dequeue_entity(cfs_rq
, se
);
1938 update_stats_curr_start(cfs_rq
, se
);
1940 #ifdef CONFIG_SCHEDSTATS
1942 * Track our maximum slice length, if the CPU's load is at
1943 * least twice that of our own weight (i.e. dont track it
1944 * when there are only lesser-weight tasks around):
1946 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
1947 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
1948 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
1951 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
1955 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
1958 * Pick the next process, keeping these things in mind, in this order:
1959 * 1) keep things fair between processes/task groups
1960 * 2) pick the "next" process, since someone really wants that to run
1961 * 3) pick the "last" process, for cache locality
1962 * 4) do not run the "skip" process, if something else is available
1964 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
1966 struct sched_entity
*se
= __pick_first_entity(cfs_rq
);
1967 struct sched_entity
*left
= se
;
1970 * Avoid running the skip buddy, if running something else can
1971 * be done without getting too unfair.
1973 if (cfs_rq
->skip
== se
) {
1974 struct sched_entity
*second
= __pick_next_entity(se
);
1975 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
1980 * Prefer last buddy, try to return the CPU to a preempted task.
1982 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
1986 * Someone really wants this to run. If it's not unfair, run it.
1988 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
1991 clear_buddies(cfs_rq
, se
);
1996 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
1998 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
2001 * If still on the runqueue then deactivate_task()
2002 * was not called and update_curr() has to be done:
2005 update_curr(cfs_rq
);
2007 /* throttle cfs_rqs exceeding runtime */
2008 check_cfs_rq_runtime(cfs_rq
);
2010 check_spread(cfs_rq
, prev
);
2012 update_stats_wait_start(cfs_rq
, prev
);
2013 /* Put 'current' back into the tree. */
2014 __enqueue_entity(cfs_rq
, prev
);
2015 /* in !on_rq case, update occurred at dequeue */
2016 update_entity_load_avg(prev
, 1);
2018 cfs_rq
->curr
= NULL
;
2022 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
2025 * Update run-time statistics of the 'current'.
2027 update_curr(cfs_rq
);
2030 * Ensure that runnable average is periodically updated.
2032 update_entity_load_avg(curr
, 1);
2033 update_cfs_rq_blocked_load(cfs_rq
, 1);
2035 #ifdef CONFIG_SCHED_HRTICK
2037 * queued ticks are scheduled to match the slice, so don't bother
2038 * validating it and just reschedule.
2041 resched_task(rq_of(cfs_rq
)->curr
);
2045 * don't let the period tick interfere with the hrtick preemption
2047 if (!sched_feat(DOUBLE_TICK
) &&
2048 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
2052 if (cfs_rq
->nr_running
> 1)
2053 check_preempt_tick(cfs_rq
, curr
);
2057 /**************************************************
2058 * CFS bandwidth control machinery
2061 #ifdef CONFIG_CFS_BANDWIDTH
2063 #ifdef HAVE_JUMP_LABEL
2064 static struct static_key __cfs_bandwidth_used
;
2066 static inline bool cfs_bandwidth_used(void)
2068 return static_key_false(&__cfs_bandwidth_used
);
2071 void account_cfs_bandwidth_used(int enabled
, int was_enabled
)
2073 /* only need to count groups transitioning between enabled/!enabled */
2074 if (enabled
&& !was_enabled
)
2075 static_key_slow_inc(&__cfs_bandwidth_used
);
2076 else if (!enabled
&& was_enabled
)
2077 static_key_slow_dec(&__cfs_bandwidth_used
);
2079 #else /* HAVE_JUMP_LABEL */
2080 static bool cfs_bandwidth_used(void)
2085 void account_cfs_bandwidth_used(int enabled
, int was_enabled
) {}
2086 #endif /* HAVE_JUMP_LABEL */
2089 * default period for cfs group bandwidth.
2090 * default: 0.1s, units: nanoseconds
2092 static inline u64
default_cfs_period(void)
2094 return 100000000ULL;
2097 static inline u64
sched_cfs_bandwidth_slice(void)
2099 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
2103 * Replenish runtime according to assigned quota and update expiration time.
2104 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2105 * additional synchronization around rq->lock.
2107 * requires cfs_b->lock
2109 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
2113 if (cfs_b
->quota
== RUNTIME_INF
)
2116 now
= sched_clock_cpu(smp_processor_id());
2117 cfs_b
->runtime
= cfs_b
->quota
;
2118 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
2121 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
2123 return &tg
->cfs_bandwidth
;
2126 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2127 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
2129 if (unlikely(cfs_rq
->throttle_count
))
2130 return cfs_rq
->throttled_clock_task
;
2132 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
2135 /* returns 0 on failure to allocate runtime */
2136 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2138 struct task_group
*tg
= cfs_rq
->tg
;
2139 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
2140 u64 amount
= 0, min_amount
, expires
;
2142 /* note: this is a positive sum as runtime_remaining <= 0 */
2143 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
2145 raw_spin_lock(&cfs_b
->lock
);
2146 if (cfs_b
->quota
== RUNTIME_INF
)
2147 amount
= min_amount
;
2150 * If the bandwidth pool has become inactive, then at least one
2151 * period must have elapsed since the last consumption.
2152 * Refresh the global state and ensure bandwidth timer becomes
2155 if (!cfs_b
->timer_active
) {
2156 __refill_cfs_bandwidth_runtime(cfs_b
);
2157 __start_cfs_bandwidth(cfs_b
);
2160 if (cfs_b
->runtime
> 0) {
2161 amount
= min(cfs_b
->runtime
, min_amount
);
2162 cfs_b
->runtime
-= amount
;
2166 expires
= cfs_b
->runtime_expires
;
2167 raw_spin_unlock(&cfs_b
->lock
);
2169 cfs_rq
->runtime_remaining
+= amount
;
2171 * we may have advanced our local expiration to account for allowed
2172 * spread between our sched_clock and the one on which runtime was
2175 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
2176 cfs_rq
->runtime_expires
= expires
;
2178 return cfs_rq
->runtime_remaining
> 0;
2182 * Note: This depends on the synchronization provided by sched_clock and the
2183 * fact that rq->clock snapshots this value.
2185 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2187 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2189 /* if the deadline is ahead of our clock, nothing to do */
2190 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
2193 if (cfs_rq
->runtime_remaining
< 0)
2197 * If the local deadline has passed we have to consider the
2198 * possibility that our sched_clock is 'fast' and the global deadline
2199 * has not truly expired.
2201 * Fortunately we can check determine whether this the case by checking
2202 * whether the global deadline has advanced.
2205 if ((s64
)(cfs_rq
->runtime_expires
- cfs_b
->runtime_expires
) >= 0) {
2206 /* extend local deadline, drift is bounded above by 2 ticks */
2207 cfs_rq
->runtime_expires
+= TICK_NSEC
;
2209 /* global deadline is ahead, expiration has passed */
2210 cfs_rq
->runtime_remaining
= 0;
2214 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
,
2215 unsigned long delta_exec
)
2217 /* dock delta_exec before expiring quota (as it could span periods) */
2218 cfs_rq
->runtime_remaining
-= delta_exec
;
2219 expire_cfs_rq_runtime(cfs_rq
);
2221 if (likely(cfs_rq
->runtime_remaining
> 0))
2225 * if we're unable to extend our runtime we resched so that the active
2226 * hierarchy can be throttled
2228 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
2229 resched_task(rq_of(cfs_rq
)->curr
);
2232 static __always_inline
2233 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
)
2235 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
2238 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
2241 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
2243 return cfs_bandwidth_used() && cfs_rq
->throttled
;
2246 /* check whether cfs_rq, or any parent, is throttled */
2247 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
2249 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
2253 * Ensure that neither of the group entities corresponding to src_cpu or
2254 * dest_cpu are members of a throttled hierarchy when performing group
2255 * load-balance operations.
2257 static inline int throttled_lb_pair(struct task_group
*tg
,
2258 int src_cpu
, int dest_cpu
)
2260 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
2262 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
2263 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
2265 return throttled_hierarchy(src_cfs_rq
) ||
2266 throttled_hierarchy(dest_cfs_rq
);
2269 /* updated child weight may affect parent so we have to do this bottom up */
2270 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
2272 struct rq
*rq
= data
;
2273 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
2275 cfs_rq
->throttle_count
--;
2277 if (!cfs_rq
->throttle_count
) {
2278 /* adjust cfs_rq_clock_task() */
2279 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
2280 cfs_rq
->throttled_clock_task
;
2287 static int tg_throttle_down(struct task_group
*tg
, void *data
)
2289 struct rq
*rq
= data
;
2290 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
2292 /* group is entering throttled state, stop time */
2293 if (!cfs_rq
->throttle_count
)
2294 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
2295 cfs_rq
->throttle_count
++;
2300 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
2302 struct rq
*rq
= rq_of(cfs_rq
);
2303 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2304 struct sched_entity
*se
;
2305 long task_delta
, dequeue
= 1;
2307 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
2309 /* freeze hierarchy runnable averages while throttled */
2311 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
2314 task_delta
= cfs_rq
->h_nr_running
;
2315 for_each_sched_entity(se
) {
2316 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
2317 /* throttled entity or throttle-on-deactivate */
2322 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
2323 qcfs_rq
->h_nr_running
-= task_delta
;
2325 if (qcfs_rq
->load
.weight
)
2330 rq
->nr_running
-= task_delta
;
2332 cfs_rq
->throttled
= 1;
2333 cfs_rq
->throttled_clock
= rq_clock(rq
);
2334 raw_spin_lock(&cfs_b
->lock
);
2335 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
2336 raw_spin_unlock(&cfs_b
->lock
);
2339 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
2341 struct rq
*rq
= rq_of(cfs_rq
);
2342 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2343 struct sched_entity
*se
;
2347 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
2349 cfs_rq
->throttled
= 0;
2351 update_rq_clock(rq
);
2353 raw_spin_lock(&cfs_b
->lock
);
2354 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
2355 list_del_rcu(&cfs_rq
->throttled_list
);
2356 raw_spin_unlock(&cfs_b
->lock
);
2358 /* update hierarchical throttle state */
2359 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
2361 if (!cfs_rq
->load
.weight
)
2364 task_delta
= cfs_rq
->h_nr_running
;
2365 for_each_sched_entity(se
) {
2369 cfs_rq
= cfs_rq_of(se
);
2371 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
2372 cfs_rq
->h_nr_running
+= task_delta
;
2374 if (cfs_rq_throttled(cfs_rq
))
2379 rq
->nr_running
+= task_delta
;
2381 /* determine whether we need to wake up potentially idle cpu */
2382 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
2383 resched_task(rq
->curr
);
2386 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
2387 u64 remaining
, u64 expires
)
2389 struct cfs_rq
*cfs_rq
;
2390 u64 runtime
= remaining
;
2393 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
2395 struct rq
*rq
= rq_of(cfs_rq
);
2397 raw_spin_lock(&rq
->lock
);
2398 if (!cfs_rq_throttled(cfs_rq
))
2401 runtime
= -cfs_rq
->runtime_remaining
+ 1;
2402 if (runtime
> remaining
)
2403 runtime
= remaining
;
2404 remaining
-= runtime
;
2406 cfs_rq
->runtime_remaining
+= runtime
;
2407 cfs_rq
->runtime_expires
= expires
;
2409 /* we check whether we're throttled above */
2410 if (cfs_rq
->runtime_remaining
> 0)
2411 unthrottle_cfs_rq(cfs_rq
);
2414 raw_spin_unlock(&rq
->lock
);
2425 * Responsible for refilling a task_group's bandwidth and unthrottling its
2426 * cfs_rqs as appropriate. If there has been no activity within the last
2427 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2428 * used to track this state.
2430 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
2432 u64 runtime
, runtime_expires
;
2433 int idle
= 1, throttled
;
2435 raw_spin_lock(&cfs_b
->lock
);
2436 /* no need to continue the timer with no bandwidth constraint */
2437 if (cfs_b
->quota
== RUNTIME_INF
)
2440 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
2441 /* idle depends on !throttled (for the case of a large deficit) */
2442 idle
= cfs_b
->idle
&& !throttled
;
2443 cfs_b
->nr_periods
+= overrun
;
2445 /* if we're going inactive then everything else can be deferred */
2449 __refill_cfs_bandwidth_runtime(cfs_b
);
2452 /* mark as potentially idle for the upcoming period */
2457 /* account preceding periods in which throttling occurred */
2458 cfs_b
->nr_throttled
+= overrun
;
2461 * There are throttled entities so we must first use the new bandwidth
2462 * to unthrottle them before making it generally available. This
2463 * ensures that all existing debts will be paid before a new cfs_rq is
2466 runtime
= cfs_b
->runtime
;
2467 runtime_expires
= cfs_b
->runtime_expires
;
2471 * This check is repeated as we are holding onto the new bandwidth
2472 * while we unthrottle. This can potentially race with an unthrottled
2473 * group trying to acquire new bandwidth from the global pool.
2475 while (throttled
&& runtime
> 0) {
2476 raw_spin_unlock(&cfs_b
->lock
);
2477 /* we can't nest cfs_b->lock while distributing bandwidth */
2478 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
2480 raw_spin_lock(&cfs_b
->lock
);
2482 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
2485 /* return (any) remaining runtime */
2486 cfs_b
->runtime
= runtime
;
2488 * While we are ensured activity in the period following an
2489 * unthrottle, this also covers the case in which the new bandwidth is
2490 * insufficient to cover the existing bandwidth deficit. (Forcing the
2491 * timer to remain active while there are any throttled entities.)
2496 cfs_b
->timer_active
= 0;
2497 raw_spin_unlock(&cfs_b
->lock
);
2502 /* a cfs_rq won't donate quota below this amount */
2503 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
2504 /* minimum remaining period time to redistribute slack quota */
2505 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
2506 /* how long we wait to gather additional slack before distributing */
2507 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
2509 /* are we near the end of the current quota period? */
2510 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
2512 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
2515 /* if the call-back is running a quota refresh is already occurring */
2516 if (hrtimer_callback_running(refresh_timer
))
2519 /* is a quota refresh about to occur? */
2520 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
2521 if (remaining
< min_expire
)
2527 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
2529 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
2531 /* if there's a quota refresh soon don't bother with slack */
2532 if (runtime_refresh_within(cfs_b
, min_left
))
2535 start_bandwidth_timer(&cfs_b
->slack_timer
,
2536 ns_to_ktime(cfs_bandwidth_slack_period
));
2539 /* we know any runtime found here is valid as update_curr() precedes return */
2540 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2542 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2543 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
2545 if (slack_runtime
<= 0)
2548 raw_spin_lock(&cfs_b
->lock
);
2549 if (cfs_b
->quota
!= RUNTIME_INF
&&
2550 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
2551 cfs_b
->runtime
+= slack_runtime
;
2553 /* we are under rq->lock, defer unthrottling using a timer */
2554 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
2555 !list_empty(&cfs_b
->throttled_cfs_rq
))
2556 start_cfs_slack_bandwidth(cfs_b
);
2558 raw_spin_unlock(&cfs_b
->lock
);
2560 /* even if it's not valid for return we don't want to try again */
2561 cfs_rq
->runtime_remaining
-= slack_runtime
;
2564 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2566 if (!cfs_bandwidth_used())
2569 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
2572 __return_cfs_rq_runtime(cfs_rq
);
2576 * This is done with a timer (instead of inline with bandwidth return) since
2577 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2579 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
2581 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
2584 /* confirm we're still not at a refresh boundary */
2585 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
))
2588 raw_spin_lock(&cfs_b
->lock
);
2589 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
) {
2590 runtime
= cfs_b
->runtime
;
2593 expires
= cfs_b
->runtime_expires
;
2594 raw_spin_unlock(&cfs_b
->lock
);
2599 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
2601 raw_spin_lock(&cfs_b
->lock
);
2602 if (expires
== cfs_b
->runtime_expires
)
2603 cfs_b
->runtime
= runtime
;
2604 raw_spin_unlock(&cfs_b
->lock
);
2608 * When a group wakes up we want to make sure that its quota is not already
2609 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2610 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2612 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
2614 if (!cfs_bandwidth_used())
2617 /* an active group must be handled by the update_curr()->put() path */
2618 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
2621 /* ensure the group is not already throttled */
2622 if (cfs_rq_throttled(cfs_rq
))
2625 /* update runtime allocation */
2626 account_cfs_rq_runtime(cfs_rq
, 0);
2627 if (cfs_rq
->runtime_remaining
<= 0)
2628 throttle_cfs_rq(cfs_rq
);
2631 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2632 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2634 if (!cfs_bandwidth_used())
2637 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
2641 * it's possible for a throttled entity to be forced into a running
2642 * state (e.g. set_curr_task), in this case we're finished.
2644 if (cfs_rq_throttled(cfs_rq
))
2647 throttle_cfs_rq(cfs_rq
);
2650 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
2652 struct cfs_bandwidth
*cfs_b
=
2653 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
2654 do_sched_cfs_slack_timer(cfs_b
);
2656 return HRTIMER_NORESTART
;
2659 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
2661 struct cfs_bandwidth
*cfs_b
=
2662 container_of(timer
, struct cfs_bandwidth
, period_timer
);
2668 now
= hrtimer_cb_get_time(timer
);
2669 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
2674 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
2677 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
2680 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2682 raw_spin_lock_init(&cfs_b
->lock
);
2684 cfs_b
->quota
= RUNTIME_INF
;
2685 cfs_b
->period
= ns_to_ktime(default_cfs_period());
2687 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
2688 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2689 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
2690 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2691 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
2694 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2696 cfs_rq
->runtime_enabled
= 0;
2697 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
2700 /* requires cfs_b->lock, may release to reprogram timer */
2701 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2704 * The timer may be active because we're trying to set a new bandwidth
2705 * period or because we're racing with the tear-down path
2706 * (timer_active==0 becomes visible before the hrtimer call-back
2707 * terminates). In either case we ensure that it's re-programmed
2709 while (unlikely(hrtimer_active(&cfs_b
->period_timer
))) {
2710 raw_spin_unlock(&cfs_b
->lock
);
2711 /* ensure cfs_b->lock is available while we wait */
2712 hrtimer_cancel(&cfs_b
->period_timer
);
2714 raw_spin_lock(&cfs_b
->lock
);
2715 /* if someone else restarted the timer then we're done */
2716 if (cfs_b
->timer_active
)
2720 cfs_b
->timer_active
= 1;
2721 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
2724 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2726 hrtimer_cancel(&cfs_b
->period_timer
);
2727 hrtimer_cancel(&cfs_b
->slack_timer
);
2730 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
2732 struct cfs_rq
*cfs_rq
;
2734 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
2735 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2737 if (!cfs_rq
->runtime_enabled
)
2741 * clock_task is not advancing so we just need to make sure
2742 * there's some valid quota amount
2744 cfs_rq
->runtime_remaining
= cfs_b
->quota
;
2745 if (cfs_rq_throttled(cfs_rq
))
2746 unthrottle_cfs_rq(cfs_rq
);
2750 #else /* CONFIG_CFS_BANDWIDTH */
2751 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
2753 return rq_clock_task(rq_of(cfs_rq
));
2756 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
,
2757 unsigned long delta_exec
) {}
2758 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2759 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
2760 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2762 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
2767 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
2772 static inline int throttled_lb_pair(struct task_group
*tg
,
2773 int src_cpu
, int dest_cpu
)
2778 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
2780 #ifdef CONFIG_FAIR_GROUP_SCHED
2781 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2784 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
2788 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
2789 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
2791 #endif /* CONFIG_CFS_BANDWIDTH */
2793 /**************************************************
2794 * CFS operations on tasks:
2797 #ifdef CONFIG_SCHED_HRTICK
2798 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
2800 struct sched_entity
*se
= &p
->se
;
2801 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2803 WARN_ON(task_rq(p
) != rq
);
2805 if (cfs_rq
->nr_running
> 1) {
2806 u64 slice
= sched_slice(cfs_rq
, se
);
2807 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
2808 s64 delta
= slice
- ran
;
2817 * Don't schedule slices shorter than 10000ns, that just
2818 * doesn't make sense. Rely on vruntime for fairness.
2821 delta
= max_t(s64
, 10000LL, delta
);
2823 hrtick_start(rq
, delta
);
2828 * called from enqueue/dequeue and updates the hrtick when the
2829 * current task is from our class and nr_running is low enough
2832 static void hrtick_update(struct rq
*rq
)
2834 struct task_struct
*curr
= rq
->curr
;
2836 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
2839 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
2840 hrtick_start_fair(rq
, curr
);
2842 #else /* !CONFIG_SCHED_HRTICK */
2844 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
2848 static inline void hrtick_update(struct rq
*rq
)
2854 * The enqueue_task method is called before nr_running is
2855 * increased. Here we update the fair scheduling stats and
2856 * then put the task into the rbtree:
2859 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
2861 struct cfs_rq
*cfs_rq
;
2862 struct sched_entity
*se
= &p
->se
;
2864 for_each_sched_entity(se
) {
2867 cfs_rq
= cfs_rq_of(se
);
2868 enqueue_entity(cfs_rq
, se
, flags
);
2871 * end evaluation on encountering a throttled cfs_rq
2873 * note: in the case of encountering a throttled cfs_rq we will
2874 * post the final h_nr_running increment below.
2876 if (cfs_rq_throttled(cfs_rq
))
2878 cfs_rq
->h_nr_running
++;
2880 flags
= ENQUEUE_WAKEUP
;
2883 for_each_sched_entity(se
) {
2884 cfs_rq
= cfs_rq_of(se
);
2885 cfs_rq
->h_nr_running
++;
2887 if (cfs_rq_throttled(cfs_rq
))
2890 update_cfs_shares(cfs_rq
);
2891 update_entity_load_avg(se
, 1);
2895 update_rq_runnable_avg(rq
, rq
->nr_running
);
2901 static void set_next_buddy(struct sched_entity
*se
);
2904 * The dequeue_task method is called before nr_running is
2905 * decreased. We remove the task from the rbtree and
2906 * update the fair scheduling stats:
2908 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
2910 struct cfs_rq
*cfs_rq
;
2911 struct sched_entity
*se
= &p
->se
;
2912 int task_sleep
= flags
& DEQUEUE_SLEEP
;
2914 for_each_sched_entity(se
) {
2915 cfs_rq
= cfs_rq_of(se
);
2916 dequeue_entity(cfs_rq
, se
, flags
);
2919 * end evaluation on encountering a throttled cfs_rq
2921 * note: in the case of encountering a throttled cfs_rq we will
2922 * post the final h_nr_running decrement below.
2924 if (cfs_rq_throttled(cfs_rq
))
2926 cfs_rq
->h_nr_running
--;
2928 /* Don't dequeue parent if it has other entities besides us */
2929 if (cfs_rq
->load
.weight
) {
2931 * Bias pick_next to pick a task from this cfs_rq, as
2932 * p is sleeping when it is within its sched_slice.
2934 if (task_sleep
&& parent_entity(se
))
2935 set_next_buddy(parent_entity(se
));
2937 /* avoid re-evaluating load for this entity */
2938 se
= parent_entity(se
);
2941 flags
|= DEQUEUE_SLEEP
;
2944 for_each_sched_entity(se
) {
2945 cfs_rq
= cfs_rq_of(se
);
2946 cfs_rq
->h_nr_running
--;
2948 if (cfs_rq_throttled(cfs_rq
))
2951 update_cfs_shares(cfs_rq
);
2952 update_entity_load_avg(se
, 1);
2957 update_rq_runnable_avg(rq
, 1);
2963 /* Used instead of source_load when we know the type == 0 */
2964 static unsigned long weighted_cpuload(const int cpu
)
2966 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
2970 * Return a low guess at the load of a migration-source cpu weighted
2971 * according to the scheduling class and "nice" value.
2973 * We want to under-estimate the load of migration sources, to
2974 * balance conservatively.
2976 static unsigned long source_load(int cpu
, int type
)
2978 struct rq
*rq
= cpu_rq(cpu
);
2979 unsigned long total
= weighted_cpuload(cpu
);
2981 if (type
== 0 || !sched_feat(LB_BIAS
))
2984 return min(rq
->cpu_load
[type
-1], total
);
2988 * Return a high guess at the load of a migration-target cpu weighted
2989 * according to the scheduling class and "nice" value.
2991 static unsigned long target_load(int cpu
, int type
)
2993 struct rq
*rq
= cpu_rq(cpu
);
2994 unsigned long total
= weighted_cpuload(cpu
);
2996 if (type
== 0 || !sched_feat(LB_BIAS
))
2999 return max(rq
->cpu_load
[type
-1], total
);
3002 static unsigned long power_of(int cpu
)
3004 return cpu_rq(cpu
)->cpu_power
;
3007 static unsigned long cpu_avg_load_per_task(int cpu
)
3009 struct rq
*rq
= cpu_rq(cpu
);
3010 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
3011 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
3014 return load_avg
/ nr_running
;
3020 static void task_waking_fair(struct task_struct
*p
)
3022 struct sched_entity
*se
= &p
->se
;
3023 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3026 #ifndef CONFIG_64BIT
3027 u64 min_vruntime_copy
;
3030 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
3032 min_vruntime
= cfs_rq
->min_vruntime
;
3033 } while (min_vruntime
!= min_vruntime_copy
);
3035 min_vruntime
= cfs_rq
->min_vruntime
;
3038 se
->vruntime
-= min_vruntime
;
3041 #ifdef CONFIG_FAIR_GROUP_SCHED
3043 * effective_load() calculates the load change as seen from the root_task_group
3045 * Adding load to a group doesn't make a group heavier, but can cause movement
3046 * of group shares between cpus. Assuming the shares were perfectly aligned one
3047 * can calculate the shift in shares.
3049 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3050 * on this @cpu and results in a total addition (subtraction) of @wg to the
3051 * total group weight.
3053 * Given a runqueue weight distribution (rw_i) we can compute a shares
3054 * distribution (s_i) using:
3056 * s_i = rw_i / \Sum rw_j (1)
3058 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3059 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3060 * shares distribution (s_i):
3062 * rw_i = { 2, 4, 1, 0 }
3063 * s_i = { 2/7, 4/7, 1/7, 0 }
3065 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3066 * task used to run on and the CPU the waker is running on), we need to
3067 * compute the effect of waking a task on either CPU and, in case of a sync
3068 * wakeup, compute the effect of the current task going to sleep.
3070 * So for a change of @wl to the local @cpu with an overall group weight change
3071 * of @wl we can compute the new shares distribution (s'_i) using:
3073 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3075 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3076 * differences in waking a task to CPU 0. The additional task changes the
3077 * weight and shares distributions like:
3079 * rw'_i = { 3, 4, 1, 0 }
3080 * s'_i = { 3/8, 4/8, 1/8, 0 }
3082 * We can then compute the difference in effective weight by using:
3084 * dw_i = S * (s'_i - s_i) (3)
3086 * Where 'S' is the group weight as seen by its parent.
3088 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3089 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3090 * 4/7) times the weight of the group.
3092 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
3094 struct sched_entity
*se
= tg
->se
[cpu
];
3096 if (!tg
->parent
) /* the trivial, non-cgroup case */
3099 for_each_sched_entity(se
) {
3105 * W = @wg + \Sum rw_j
3107 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
3112 w
= se
->my_q
->load
.weight
+ wl
;
3115 * wl = S * s'_i; see (2)
3118 wl
= (w
* tg
->shares
) / W
;
3123 * Per the above, wl is the new se->load.weight value; since
3124 * those are clipped to [MIN_SHARES, ...) do so now. See
3125 * calc_cfs_shares().
3127 if (wl
< MIN_SHARES
)
3131 * wl = dw_i = S * (s'_i - s_i); see (3)
3133 wl
-= se
->load
.weight
;
3136 * Recursively apply this logic to all parent groups to compute
3137 * the final effective load change on the root group. Since
3138 * only the @tg group gets extra weight, all parent groups can
3139 * only redistribute existing shares. @wl is the shift in shares
3140 * resulting from this level per the above.
3149 static inline unsigned long effective_load(struct task_group
*tg
, int cpu
,
3150 unsigned long wl
, unsigned long wg
)
3157 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
3159 s64 this_load
, load
;
3160 int idx
, this_cpu
, prev_cpu
;
3161 unsigned long tl_per_task
;
3162 struct task_group
*tg
;
3163 unsigned long weight
;
3167 this_cpu
= smp_processor_id();
3168 prev_cpu
= task_cpu(p
);
3169 load
= source_load(prev_cpu
, idx
);
3170 this_load
= target_load(this_cpu
, idx
);
3173 * If sync wakeup then subtract the (maximum possible)
3174 * effect of the currently running task from the load
3175 * of the current CPU:
3178 tg
= task_group(current
);
3179 weight
= current
->se
.load
.weight
;
3181 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
3182 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
3186 weight
= p
->se
.load
.weight
;
3189 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3190 * due to the sync cause above having dropped this_load to 0, we'll
3191 * always have an imbalance, but there's really nothing you can do
3192 * about that, so that's good too.
3194 * Otherwise check if either cpus are near enough in load to allow this
3195 * task to be woken on this_cpu.
3197 if (this_load
> 0) {
3198 s64 this_eff_load
, prev_eff_load
;
3200 this_eff_load
= 100;
3201 this_eff_load
*= power_of(prev_cpu
);
3202 this_eff_load
*= this_load
+
3203 effective_load(tg
, this_cpu
, weight
, weight
);
3205 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
3206 prev_eff_load
*= power_of(this_cpu
);
3207 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
3209 balanced
= this_eff_load
<= prev_eff_load
;
3214 * If the currently running task will sleep within
3215 * a reasonable amount of time then attract this newly
3218 if (sync
&& balanced
)
3221 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
3222 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
3225 (this_load
<= load
&&
3226 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
3228 * This domain has SD_WAKE_AFFINE and
3229 * p is cache cold in this domain, and
3230 * there is no bad imbalance.
3232 schedstat_inc(sd
, ttwu_move_affine
);
3233 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
3241 * find_idlest_group finds and returns the least busy CPU group within the
3244 static struct sched_group
*
3245 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
3246 int this_cpu
, int load_idx
)
3248 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
3249 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
3250 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
3253 unsigned long load
, avg_load
;
3257 /* Skip over this group if it has no CPUs allowed */
3258 if (!cpumask_intersects(sched_group_cpus(group
),
3259 tsk_cpus_allowed(p
)))
3262 local_group
= cpumask_test_cpu(this_cpu
,
3263 sched_group_cpus(group
));
3265 /* Tally up the load of all CPUs in the group */
3268 for_each_cpu(i
, sched_group_cpus(group
)) {
3269 /* Bias balancing toward cpus of our domain */
3271 load
= source_load(i
, load_idx
);
3273 load
= target_load(i
, load_idx
);
3278 /* Adjust by relative CPU power of the group */
3279 avg_load
= (avg_load
* SCHED_POWER_SCALE
) / group
->sgp
->power
;
3282 this_load
= avg_load
;
3283 } else if (avg_load
< min_load
) {
3284 min_load
= avg_load
;
3287 } while (group
= group
->next
, group
!= sd
->groups
);
3289 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
3295 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3298 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
3300 unsigned long load
, min_load
= ULONG_MAX
;
3304 /* Traverse only the allowed CPUs */
3305 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
3306 load
= weighted_cpuload(i
);
3308 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
3318 * Try and locate an idle CPU in the sched_domain.
3320 static int select_idle_sibling(struct task_struct
*p
, int target
)
3322 struct sched_domain
*sd
;
3323 struct sched_group
*sg
;
3324 int i
= task_cpu(p
);
3326 if (idle_cpu(target
))
3330 * If the prevous cpu is cache affine and idle, don't be stupid.
3332 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
3336 * Otherwise, iterate the domains and find an elegible idle cpu.
3338 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
3339 for_each_lower_domain(sd
) {
3342 if (!cpumask_intersects(sched_group_cpus(sg
),
3343 tsk_cpus_allowed(p
)))
3346 for_each_cpu(i
, sched_group_cpus(sg
)) {
3347 if (i
== target
|| !idle_cpu(i
))
3351 target
= cpumask_first_and(sched_group_cpus(sg
),
3352 tsk_cpus_allowed(p
));
3356 } while (sg
!= sd
->groups
);
3363 * sched_balance_self: balance the current task (running on cpu) in domains
3364 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3367 * Balance, ie. select the least loaded group.
3369 * Returns the target CPU number, or the same CPU if no balancing is needed.
3371 * preempt must be disabled.
3374 select_task_rq_fair(struct task_struct
*p
, int sd_flag
, int wake_flags
)
3376 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
3377 int cpu
= smp_processor_id();
3378 int prev_cpu
= task_cpu(p
);
3380 int want_affine
= 0;
3381 int sync
= wake_flags
& WF_SYNC
;
3383 if (p
->nr_cpus_allowed
== 1)
3386 if (sd_flag
& SD_BALANCE_WAKE
) {
3387 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
3393 for_each_domain(cpu
, tmp
) {
3394 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
3398 * If both cpu and prev_cpu are part of this domain,
3399 * cpu is a valid SD_WAKE_AFFINE target.
3401 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
3402 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
3407 if (tmp
->flags
& sd_flag
)
3412 if (cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
3415 new_cpu
= select_idle_sibling(p
, prev_cpu
);
3420 int load_idx
= sd
->forkexec_idx
;
3421 struct sched_group
*group
;
3424 if (!(sd
->flags
& sd_flag
)) {
3429 if (sd_flag
& SD_BALANCE_WAKE
)
3430 load_idx
= sd
->wake_idx
;
3432 group
= find_idlest_group(sd
, p
, cpu
, load_idx
);
3438 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
3439 if (new_cpu
== -1 || new_cpu
== cpu
) {
3440 /* Now try balancing at a lower domain level of cpu */
3445 /* Now try balancing at a lower domain level of new_cpu */
3447 weight
= sd
->span_weight
;
3449 for_each_domain(cpu
, tmp
) {
3450 if (weight
<= tmp
->span_weight
)
3452 if (tmp
->flags
& sd_flag
)
3455 /* while loop will break here if sd == NULL */
3464 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3465 * cfs_rq_of(p) references at time of call are still valid and identify the
3466 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3467 * other assumptions, including the state of rq->lock, should be made.
3470 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
3472 struct sched_entity
*se
= &p
->se
;
3473 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3476 * Load tracking: accumulate removed load so that it can be processed
3477 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3478 * to blocked load iff they have a positive decay-count. It can never
3479 * be negative here since on-rq tasks have decay-count == 0.
3481 if (se
->avg
.decay_count
) {
3482 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
3483 atomic64_add(se
->avg
.load_avg_contrib
, &cfs_rq
->removed_load
);
3486 #endif /* CONFIG_SMP */
3488 static unsigned long
3489 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
3491 unsigned long gran
= sysctl_sched_wakeup_granularity
;
3494 * Since its curr running now, convert the gran from real-time
3495 * to virtual-time in his units.
3497 * By using 'se' instead of 'curr' we penalize light tasks, so
3498 * they get preempted easier. That is, if 'se' < 'curr' then
3499 * the resulting gran will be larger, therefore penalizing the
3500 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3501 * be smaller, again penalizing the lighter task.
3503 * This is especially important for buddies when the leftmost
3504 * task is higher priority than the buddy.
3506 return calc_delta_fair(gran
, se
);
3510 * Should 'se' preempt 'curr'.
3524 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
3526 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
3531 gran
= wakeup_gran(curr
, se
);
3538 static void set_last_buddy(struct sched_entity
*se
)
3540 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
3543 for_each_sched_entity(se
)
3544 cfs_rq_of(se
)->last
= se
;
3547 static void set_next_buddy(struct sched_entity
*se
)
3549 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
3552 for_each_sched_entity(se
)
3553 cfs_rq_of(se
)->next
= se
;
3556 static void set_skip_buddy(struct sched_entity
*se
)
3558 for_each_sched_entity(se
)
3559 cfs_rq_of(se
)->skip
= se
;
3563 * Preempt the current task with a newly woken task if needed:
3565 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
3567 struct task_struct
*curr
= rq
->curr
;
3568 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
3569 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
3570 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
3571 int next_buddy_marked
= 0;
3573 if (unlikely(se
== pse
))
3577 * This is possible from callers such as move_task(), in which we
3578 * unconditionally check_prempt_curr() after an enqueue (which may have
3579 * lead to a throttle). This both saves work and prevents false
3580 * next-buddy nomination below.
3582 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
3585 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
3586 set_next_buddy(pse
);
3587 next_buddy_marked
= 1;
3591 * We can come here with TIF_NEED_RESCHED already set from new task
3594 * Note: this also catches the edge-case of curr being in a throttled
3595 * group (e.g. via set_curr_task), since update_curr() (in the
3596 * enqueue of curr) will have resulted in resched being set. This
3597 * prevents us from potentially nominating it as a false LAST_BUDDY
3600 if (test_tsk_need_resched(curr
))
3603 /* Idle tasks are by definition preempted by non-idle tasks. */
3604 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
3605 likely(p
->policy
!= SCHED_IDLE
))
3609 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3610 * is driven by the tick):
3612 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
3615 find_matching_se(&se
, &pse
);
3616 update_curr(cfs_rq_of(se
));
3618 if (wakeup_preempt_entity(se
, pse
) == 1) {
3620 * Bias pick_next to pick the sched entity that is
3621 * triggering this preemption.
3623 if (!next_buddy_marked
)
3624 set_next_buddy(pse
);
3633 * Only set the backward buddy when the current task is still
3634 * on the rq. This can happen when a wakeup gets interleaved
3635 * with schedule on the ->pre_schedule() or idle_balance()
3636 * point, either of which can * drop the rq lock.
3638 * Also, during early boot the idle thread is in the fair class,
3639 * for obvious reasons its a bad idea to schedule back to it.
3641 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
3644 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
3648 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
3650 struct task_struct
*p
;
3651 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
3652 struct sched_entity
*se
;
3654 if (!cfs_rq
->nr_running
)
3658 se
= pick_next_entity(cfs_rq
);
3659 set_next_entity(cfs_rq
, se
);
3660 cfs_rq
= group_cfs_rq(se
);
3664 if (hrtick_enabled(rq
))
3665 hrtick_start_fair(rq
, p
);
3671 * Account for a descheduled task:
3673 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
3675 struct sched_entity
*se
= &prev
->se
;
3676 struct cfs_rq
*cfs_rq
;
3678 for_each_sched_entity(se
) {
3679 cfs_rq
= cfs_rq_of(se
);
3680 put_prev_entity(cfs_rq
, se
);
3685 * sched_yield() is very simple
3687 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3689 static void yield_task_fair(struct rq
*rq
)
3691 struct task_struct
*curr
= rq
->curr
;
3692 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
3693 struct sched_entity
*se
= &curr
->se
;
3696 * Are we the only task in the tree?
3698 if (unlikely(rq
->nr_running
== 1))
3701 clear_buddies(cfs_rq
, se
);
3703 if (curr
->policy
!= SCHED_BATCH
) {
3704 update_rq_clock(rq
);
3706 * Update run-time statistics of the 'current'.
3708 update_curr(cfs_rq
);
3710 * Tell update_rq_clock() that we've just updated,
3711 * so we don't do microscopic update in schedule()
3712 * and double the fastpath cost.
3714 rq
->skip_clock_update
= 1;
3720 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
3722 struct sched_entity
*se
= &p
->se
;
3724 /* throttled hierarchies are not runnable */
3725 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
3728 /* Tell the scheduler that we'd really like pse to run next. */
3731 yield_task_fair(rq
);
3737 /**************************************************
3738 * Fair scheduling class load-balancing methods.
3742 * The purpose of load-balancing is to achieve the same basic fairness the
3743 * per-cpu scheduler provides, namely provide a proportional amount of compute
3744 * time to each task. This is expressed in the following equation:
3746 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3748 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3749 * W_i,0 is defined as:
3751 * W_i,0 = \Sum_j w_i,j (2)
3753 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3754 * is derived from the nice value as per prio_to_weight[].
3756 * The weight average is an exponential decay average of the instantaneous
3759 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3761 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3762 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3763 * can also include other factors [XXX].
3765 * To achieve this balance we define a measure of imbalance which follows
3766 * directly from (1):
3768 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3770 * We them move tasks around to minimize the imbalance. In the continuous
3771 * function space it is obvious this converges, in the discrete case we get
3772 * a few fun cases generally called infeasible weight scenarios.
3775 * - infeasible weights;
3776 * - local vs global optima in the discrete case. ]
3781 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3782 * for all i,j solution, we create a tree of cpus that follows the hardware
3783 * topology where each level pairs two lower groups (or better). This results
3784 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3785 * tree to only the first of the previous level and we decrease the frequency
3786 * of load-balance at each level inv. proportional to the number of cpus in
3792 * \Sum { --- * --- * 2^i } = O(n) (5)
3794 * `- size of each group
3795 * | | `- number of cpus doing load-balance
3797 * `- sum over all levels
3799 * Coupled with a limit on how many tasks we can migrate every balance pass,
3800 * this makes (5) the runtime complexity of the balancer.
3802 * An important property here is that each CPU is still (indirectly) connected
3803 * to every other cpu in at most O(log n) steps:
3805 * The adjacency matrix of the resulting graph is given by:
3808 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3811 * And you'll find that:
3813 * A^(log_2 n)_i,j != 0 for all i,j (7)
3815 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3816 * The task movement gives a factor of O(m), giving a convergence complexity
3819 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3824 * In order to avoid CPUs going idle while there's still work to do, new idle
3825 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3826 * tree itself instead of relying on other CPUs to bring it work.
3828 * This adds some complexity to both (5) and (8) but it reduces the total idle
3836 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3839 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3844 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3846 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3848 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3851 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3852 * rewrite all of this once again.]
3855 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
3857 #define LBF_ALL_PINNED 0x01
3858 #define LBF_NEED_BREAK 0x02
3859 #define LBF_SOME_PINNED 0x04
3862 struct sched_domain
*sd
;
3870 struct cpumask
*dst_grpmask
;
3872 enum cpu_idle_type idle
;
3874 /* The set of CPUs under consideration for load-balancing */
3875 struct cpumask
*cpus
;
3880 unsigned int loop_break
;
3881 unsigned int loop_max
;
3885 * move_task - move a task from one runqueue to another runqueue.
3886 * Both runqueues must be locked.
3888 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
3890 deactivate_task(env
->src_rq
, p
, 0);
3891 set_task_cpu(p
, env
->dst_cpu
);
3892 activate_task(env
->dst_rq
, p
, 0);
3893 check_preempt_curr(env
->dst_rq
, p
, 0);
3897 * Is this task likely cache-hot:
3900 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
3904 if (p
->sched_class
!= &fair_sched_class
)
3907 if (unlikely(p
->policy
== SCHED_IDLE
))
3911 * Buddy candidates are cache hot:
3913 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
3914 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
3915 &p
->se
== cfs_rq_of(&p
->se
)->last
))
3918 if (sysctl_sched_migration_cost
== -1)
3920 if (sysctl_sched_migration_cost
== 0)
3923 delta
= now
- p
->se
.exec_start
;
3925 return delta
< (s64
)sysctl_sched_migration_cost
;
3929 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3932 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
3934 int tsk_cache_hot
= 0;
3936 * We do not migrate tasks that are:
3937 * 1) throttled_lb_pair, or
3938 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3939 * 3) running (obviously), or
3940 * 4) are cache-hot on their current CPU.
3942 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
3945 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
3948 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
3951 * Remember if this task can be migrated to any other cpu in
3952 * our sched_group. We may want to revisit it if we couldn't
3953 * meet load balance goals by pulling other tasks on src_cpu.
3955 * Also avoid computing new_dst_cpu if we have already computed
3956 * one in current iteration.
3958 if (!env
->dst_grpmask
|| (env
->flags
& LBF_SOME_PINNED
))
3961 /* Prevent to re-select dst_cpu via env's cpus */
3962 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
3963 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
3964 env
->flags
|= LBF_SOME_PINNED
;
3965 env
->new_dst_cpu
= cpu
;
3973 /* Record that we found atleast one task that could run on dst_cpu */
3974 env
->flags
&= ~LBF_ALL_PINNED
;
3976 if (task_running(env
->src_rq
, p
)) {
3977 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
3982 * Aggressive migration if:
3983 * 1) task is cache cold, or
3984 * 2) too many balance attempts have failed.
3987 tsk_cache_hot
= task_hot(p
, rq_clock_task(env
->src_rq
), env
->sd
);
3988 if (!tsk_cache_hot
||
3989 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
3991 if (tsk_cache_hot
) {
3992 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
3993 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
3999 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
4004 * move_one_task tries to move exactly one task from busiest to this_rq, as
4005 * part of active balancing operations within "domain".
4006 * Returns 1 if successful and 0 otherwise.
4008 * Called with both runqueues locked.
4010 static int move_one_task(struct lb_env
*env
)
4012 struct task_struct
*p
, *n
;
4014 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
4015 if (!can_migrate_task(p
, env
))
4020 * Right now, this is only the second place move_task()
4021 * is called, so we can safely collect move_task()
4022 * stats here rather than inside move_task().
4024 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
4030 static unsigned long task_h_load(struct task_struct
*p
);
4032 static const unsigned int sched_nr_migrate_break
= 32;
4035 * move_tasks tries to move up to imbalance weighted load from busiest to
4036 * this_rq, as part of a balancing operation within domain "sd".
4037 * Returns 1 if successful and 0 otherwise.
4039 * Called with both runqueues locked.
4041 static int move_tasks(struct lb_env
*env
)
4043 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
4044 struct task_struct
*p
;
4048 if (env
->imbalance
<= 0)
4051 while (!list_empty(tasks
)) {
4052 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
4055 /* We've more or less seen every task there is, call it quits */
4056 if (env
->loop
> env
->loop_max
)
4059 /* take a breather every nr_migrate tasks */
4060 if (env
->loop
> env
->loop_break
) {
4061 env
->loop_break
+= sched_nr_migrate_break
;
4062 env
->flags
|= LBF_NEED_BREAK
;
4066 if (!can_migrate_task(p
, env
))
4069 load
= task_h_load(p
);
4071 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
4074 if ((load
/ 2) > env
->imbalance
)
4079 env
->imbalance
-= load
;
4081 #ifdef CONFIG_PREEMPT
4083 * NEWIDLE balancing is a source of latency, so preemptible
4084 * kernels will stop after the first task is pulled to minimize
4085 * the critical section.
4087 if (env
->idle
== CPU_NEWLY_IDLE
)
4092 * We only want to steal up to the prescribed amount of
4095 if (env
->imbalance
<= 0)
4100 list_move_tail(&p
->se
.group_node
, tasks
);
4104 * Right now, this is one of only two places move_task() is called,
4105 * so we can safely collect move_task() stats here rather than
4106 * inside move_task().
4108 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
4113 #ifdef CONFIG_FAIR_GROUP_SCHED
4115 * update tg->load_weight by folding this cpu's load_avg
4117 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
4119 struct sched_entity
*se
= tg
->se
[cpu
];
4120 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
4122 /* throttled entities do not contribute to load */
4123 if (throttled_hierarchy(cfs_rq
))
4126 update_cfs_rq_blocked_load(cfs_rq
, 1);
4129 update_entity_load_avg(se
, 1);
4131 * We pivot on our runnable average having decayed to zero for
4132 * list removal. This generally implies that all our children
4133 * have also been removed (modulo rounding error or bandwidth
4134 * control); however, such cases are rare and we can fix these
4137 * TODO: fix up out-of-order children on enqueue.
4139 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
4140 list_del_leaf_cfs_rq(cfs_rq
);
4142 struct rq
*rq
= rq_of(cfs_rq
);
4143 update_rq_runnable_avg(rq
, rq
->nr_running
);
4147 static void update_blocked_averages(int cpu
)
4149 struct rq
*rq
= cpu_rq(cpu
);
4150 struct cfs_rq
*cfs_rq
;
4151 unsigned long flags
;
4153 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4154 update_rq_clock(rq
);
4156 * Iterates the task_group tree in a bottom up fashion, see
4157 * list_add_leaf_cfs_rq() for details.
4159 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4161 * Note: We may want to consider periodically releasing
4162 * rq->lock about these updates so that creating many task
4163 * groups does not result in continually extending hold time.
4165 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
4168 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4172 * Compute the cpu's hierarchical load factor for each task group.
4173 * This needs to be done in a top-down fashion because the load of a child
4174 * group is a fraction of its parents load.
4176 static int tg_load_down(struct task_group
*tg
, void *data
)
4179 long cpu
= (long)data
;
4182 load
= cpu_rq(cpu
)->avg
.load_avg_contrib
;
4184 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
4185 load
= div64_ul(load
* tg
->se
[cpu
]->avg
.load_avg_contrib
,
4186 tg
->parent
->cfs_rq
[cpu
]->runnable_load_avg
+ 1);
4189 tg
->cfs_rq
[cpu
]->h_load
= load
;
4194 static void update_h_load(long cpu
)
4196 struct rq
*rq
= cpu_rq(cpu
);
4197 unsigned long now
= jiffies
;
4199 if (rq
->h_load_throttle
== now
)
4202 rq
->h_load_throttle
= now
;
4205 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
4209 static unsigned long task_h_load(struct task_struct
*p
)
4211 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
4213 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
4214 cfs_rq
->runnable_load_avg
+ 1);
4217 static inline void update_blocked_averages(int cpu
)
4221 static inline void update_h_load(long cpu
)
4225 static unsigned long task_h_load(struct task_struct
*p
)
4227 return p
->se
.avg
.load_avg_contrib
;
4231 /********** Helpers for find_busiest_group ************************/
4233 * sd_lb_stats - Structure to store the statistics of a sched_domain
4234 * during load balancing.
4236 struct sd_lb_stats
{
4237 struct sched_group
*busiest
; /* Busiest group in this sd */
4238 struct sched_group
*this; /* Local group in this sd */
4239 unsigned long total_load
; /* Total load of all groups in sd */
4240 unsigned long total_pwr
; /* Total power of all groups in sd */
4241 unsigned long avg_load
; /* Average load across all groups in sd */
4243 /** Statistics of this group */
4244 unsigned long this_load
;
4245 unsigned long this_load_per_task
;
4246 unsigned long this_nr_running
;
4247 unsigned long this_has_capacity
;
4248 unsigned int this_idle_cpus
;
4250 /* Statistics of the busiest group */
4251 unsigned int busiest_idle_cpus
;
4252 unsigned long max_load
;
4253 unsigned long busiest_load_per_task
;
4254 unsigned long busiest_nr_running
;
4255 unsigned long busiest_group_capacity
;
4256 unsigned long busiest_has_capacity
;
4257 unsigned int busiest_group_weight
;
4259 int group_imb
; /* Is there imbalance in this sd */
4263 * sg_lb_stats - stats of a sched_group required for load_balancing
4265 struct sg_lb_stats
{
4266 unsigned long avg_load
; /*Avg load across the CPUs of the group */
4267 unsigned long group_load
; /* Total load over the CPUs of the group */
4268 unsigned long sum_nr_running
; /* Nr tasks running in the group */
4269 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
4270 unsigned long group_capacity
;
4271 unsigned long idle_cpus
;
4272 unsigned long group_weight
;
4273 int group_imb
; /* Is there an imbalance in the group ? */
4274 int group_has_capacity
; /* Is there extra capacity in the group? */
4278 * get_sd_load_idx - Obtain the load index for a given sched domain.
4279 * @sd: The sched_domain whose load_idx is to be obtained.
4280 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4282 static inline int get_sd_load_idx(struct sched_domain
*sd
,
4283 enum cpu_idle_type idle
)
4289 load_idx
= sd
->busy_idx
;
4292 case CPU_NEWLY_IDLE
:
4293 load_idx
= sd
->newidle_idx
;
4296 load_idx
= sd
->idle_idx
;
4303 static unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
4305 return SCHED_POWER_SCALE
;
4308 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
4310 return default_scale_freq_power(sd
, cpu
);
4313 static unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
4315 unsigned long weight
= sd
->span_weight
;
4316 unsigned long smt_gain
= sd
->smt_gain
;
4323 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
4325 return default_scale_smt_power(sd
, cpu
);
4328 static unsigned long scale_rt_power(int cpu
)
4330 struct rq
*rq
= cpu_rq(cpu
);
4331 u64 total
, available
, age_stamp
, avg
;
4334 * Since we're reading these variables without serialization make sure
4335 * we read them once before doing sanity checks on them.
4337 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
4338 avg
= ACCESS_ONCE(rq
->rt_avg
);
4340 total
= sched_avg_period() + (rq_clock(rq
) - age_stamp
);
4342 if (unlikely(total
< avg
)) {
4343 /* Ensures that power won't end up being negative */
4346 available
= total
- avg
;
4349 if (unlikely((s64
)total
< SCHED_POWER_SCALE
))
4350 total
= SCHED_POWER_SCALE
;
4352 total
>>= SCHED_POWER_SHIFT
;
4354 return div_u64(available
, total
);
4357 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
4359 unsigned long weight
= sd
->span_weight
;
4360 unsigned long power
= SCHED_POWER_SCALE
;
4361 struct sched_group
*sdg
= sd
->groups
;
4363 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
4364 if (sched_feat(ARCH_POWER
))
4365 power
*= arch_scale_smt_power(sd
, cpu
);
4367 power
*= default_scale_smt_power(sd
, cpu
);
4369 power
>>= SCHED_POWER_SHIFT
;
4372 sdg
->sgp
->power_orig
= power
;
4374 if (sched_feat(ARCH_POWER
))
4375 power
*= arch_scale_freq_power(sd
, cpu
);
4377 power
*= default_scale_freq_power(sd
, cpu
);
4379 power
>>= SCHED_POWER_SHIFT
;
4381 power
*= scale_rt_power(cpu
);
4382 power
>>= SCHED_POWER_SHIFT
;
4387 cpu_rq(cpu
)->cpu_power
= power
;
4388 sdg
->sgp
->power
= power
;
4391 void update_group_power(struct sched_domain
*sd
, int cpu
)
4393 struct sched_domain
*child
= sd
->child
;
4394 struct sched_group
*group
, *sdg
= sd
->groups
;
4395 unsigned long power
;
4396 unsigned long interval
;
4398 interval
= msecs_to_jiffies(sd
->balance_interval
);
4399 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
4400 sdg
->sgp
->next_update
= jiffies
+ interval
;
4403 update_cpu_power(sd
, cpu
);
4409 if (child
->flags
& SD_OVERLAP
) {
4411 * SD_OVERLAP domains cannot assume that child groups
4412 * span the current group.
4415 for_each_cpu(cpu
, sched_group_cpus(sdg
))
4416 power
+= power_of(cpu
);
4419 * !SD_OVERLAP domains can assume that child groups
4420 * span the current group.
4423 group
= child
->groups
;
4425 power
+= group
->sgp
->power
;
4426 group
= group
->next
;
4427 } while (group
!= child
->groups
);
4430 sdg
->sgp
->power_orig
= sdg
->sgp
->power
= power
;
4434 * Try and fix up capacity for tiny siblings, this is needed when
4435 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4436 * which on its own isn't powerful enough.
4438 * See update_sd_pick_busiest() and check_asym_packing().
4441 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
4444 * Only siblings can have significantly less than SCHED_POWER_SCALE
4446 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
4450 * If ~90% of the cpu_power is still there, we're good.
4452 if (group
->sgp
->power
* 32 > group
->sgp
->power_orig
* 29)
4459 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4460 * @env: The load balancing environment.
4461 * @group: sched_group whose statistics are to be updated.
4462 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4463 * @local_group: Does group contain this_cpu.
4464 * @balance: Should we balance.
4465 * @sgs: variable to hold the statistics for this group.
4467 static inline void update_sg_lb_stats(struct lb_env
*env
,
4468 struct sched_group
*group
, int load_idx
,
4469 int local_group
, int *balance
, struct sg_lb_stats
*sgs
)
4471 unsigned long nr_running
, max_nr_running
, min_nr_running
;
4472 unsigned long load
, max_cpu_load
, min_cpu_load
;
4473 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
4474 unsigned long avg_load_per_task
= 0;
4478 balance_cpu
= group_balance_cpu(group
);
4480 /* Tally up the load of all CPUs in the group */
4482 min_cpu_load
= ~0UL;
4484 min_nr_running
= ~0UL;
4486 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
4487 struct rq
*rq
= cpu_rq(i
);
4489 nr_running
= rq
->nr_running
;
4491 /* Bias balancing toward cpus of our domain */
4493 if (idle_cpu(i
) && !first_idle_cpu
&&
4494 cpumask_test_cpu(i
, sched_group_mask(group
))) {
4499 load
= target_load(i
, load_idx
);
4501 load
= source_load(i
, load_idx
);
4502 if (load
> max_cpu_load
)
4503 max_cpu_load
= load
;
4504 if (min_cpu_load
> load
)
4505 min_cpu_load
= load
;
4507 if (nr_running
> max_nr_running
)
4508 max_nr_running
= nr_running
;
4509 if (min_nr_running
> nr_running
)
4510 min_nr_running
= nr_running
;
4513 sgs
->group_load
+= load
;
4514 sgs
->sum_nr_running
+= nr_running
;
4515 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
4521 * First idle cpu or the first cpu(busiest) in this sched group
4522 * is eligible for doing load balancing at this and above
4523 * domains. In the newly idle case, we will allow all the cpu's
4524 * to do the newly idle load balance.
4527 if (env
->idle
!= CPU_NEWLY_IDLE
) {
4528 if (balance_cpu
!= env
->dst_cpu
) {
4532 update_group_power(env
->sd
, env
->dst_cpu
);
4533 } else if (time_after_eq(jiffies
, group
->sgp
->next_update
))
4534 update_group_power(env
->sd
, env
->dst_cpu
);
4537 /* Adjust by relative CPU power of the group */
4538 sgs
->avg_load
= (sgs
->group_load
*SCHED_POWER_SCALE
) / group
->sgp
->power
;
4541 * Consider the group unbalanced when the imbalance is larger
4542 * than the average weight of a task.
4544 * APZ: with cgroup the avg task weight can vary wildly and
4545 * might not be a suitable number - should we keep a
4546 * normalized nr_running number somewhere that negates
4549 if (sgs
->sum_nr_running
)
4550 avg_load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
4552 if ((max_cpu_load
- min_cpu_load
) >= avg_load_per_task
&&
4553 (max_nr_running
- min_nr_running
) > 1)
4556 sgs
->group_capacity
= DIV_ROUND_CLOSEST(group
->sgp
->power
,
4558 if (!sgs
->group_capacity
)
4559 sgs
->group_capacity
= fix_small_capacity(env
->sd
, group
);
4560 sgs
->group_weight
= group
->group_weight
;
4562 if (sgs
->group_capacity
> sgs
->sum_nr_running
)
4563 sgs
->group_has_capacity
= 1;
4567 * update_sd_pick_busiest - return 1 on busiest group
4568 * @env: The load balancing environment.
4569 * @sds: sched_domain statistics
4570 * @sg: sched_group candidate to be checked for being the busiest
4571 * @sgs: sched_group statistics
4573 * Determine if @sg is a busier group than the previously selected
4576 static bool update_sd_pick_busiest(struct lb_env
*env
,
4577 struct sd_lb_stats
*sds
,
4578 struct sched_group
*sg
,
4579 struct sg_lb_stats
*sgs
)
4581 if (sgs
->avg_load
<= sds
->max_load
)
4584 if (sgs
->sum_nr_running
> sgs
->group_capacity
)
4591 * ASYM_PACKING needs to move all the work to the lowest
4592 * numbered CPUs in the group, therefore mark all groups
4593 * higher than ourself as busy.
4595 if ((env
->sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
4596 env
->dst_cpu
< group_first_cpu(sg
)) {
4600 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
4608 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4609 * @env: The load balancing environment.
4610 * @balance: Should we balance.
4611 * @sds: variable to hold the statistics for this sched_domain.
4613 static inline void update_sd_lb_stats(struct lb_env
*env
,
4614 int *balance
, struct sd_lb_stats
*sds
)
4616 struct sched_domain
*child
= env
->sd
->child
;
4617 struct sched_group
*sg
= env
->sd
->groups
;
4618 struct sg_lb_stats sgs
;
4619 int load_idx
, prefer_sibling
= 0;
4621 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
4624 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
4629 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
4630 memset(&sgs
, 0, sizeof(sgs
));
4631 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, balance
, &sgs
);
4633 if (local_group
&& !(*balance
))
4636 sds
->total_load
+= sgs
.group_load
;
4637 sds
->total_pwr
+= sg
->sgp
->power
;
4640 * In case the child domain prefers tasks go to siblings
4641 * first, lower the sg capacity to one so that we'll try
4642 * and move all the excess tasks away. We lower the capacity
4643 * of a group only if the local group has the capacity to fit
4644 * these excess tasks, i.e. nr_running < group_capacity. The
4645 * extra check prevents the case where you always pull from the
4646 * heaviest group when it is already under-utilized (possible
4647 * with a large weight task outweighs the tasks on the system).
4649 if (prefer_sibling
&& !local_group
&& sds
->this_has_capacity
)
4650 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
4653 sds
->this_load
= sgs
.avg_load
;
4655 sds
->this_nr_running
= sgs
.sum_nr_running
;
4656 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
4657 sds
->this_has_capacity
= sgs
.group_has_capacity
;
4658 sds
->this_idle_cpus
= sgs
.idle_cpus
;
4659 } else if (update_sd_pick_busiest(env
, sds
, sg
, &sgs
)) {
4660 sds
->max_load
= sgs
.avg_load
;
4662 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
4663 sds
->busiest_idle_cpus
= sgs
.idle_cpus
;
4664 sds
->busiest_group_capacity
= sgs
.group_capacity
;
4665 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
4666 sds
->busiest_has_capacity
= sgs
.group_has_capacity
;
4667 sds
->busiest_group_weight
= sgs
.group_weight
;
4668 sds
->group_imb
= sgs
.group_imb
;
4672 } while (sg
!= env
->sd
->groups
);
4676 * check_asym_packing - Check to see if the group is packed into the
4679 * This is primarily intended to used at the sibling level. Some
4680 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4681 * case of POWER7, it can move to lower SMT modes only when higher
4682 * threads are idle. When in lower SMT modes, the threads will
4683 * perform better since they share less core resources. Hence when we
4684 * have idle threads, we want them to be the higher ones.
4686 * This packing function is run on idle threads. It checks to see if
4687 * the busiest CPU in this domain (core in the P7 case) has a higher
4688 * CPU number than the packing function is being run on. Here we are
4689 * assuming lower CPU number will be equivalent to lower a SMT thread
4692 * Returns 1 when packing is required and a task should be moved to
4693 * this CPU. The amount of the imbalance is returned in *imbalance.
4695 * @env: The load balancing environment.
4696 * @sds: Statistics of the sched_domain which is to be packed
4698 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4702 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
4708 busiest_cpu
= group_first_cpu(sds
->busiest
);
4709 if (env
->dst_cpu
> busiest_cpu
)
4712 env
->imbalance
= DIV_ROUND_CLOSEST(
4713 sds
->max_load
* sds
->busiest
->sgp
->power
, SCHED_POWER_SCALE
);
4719 * fix_small_imbalance - Calculate the minor imbalance that exists
4720 * amongst the groups of a sched_domain, during
4722 * @env: The load balancing environment.
4723 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4726 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4728 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
4729 unsigned int imbn
= 2;
4730 unsigned long scaled_busy_load_per_task
;
4732 if (sds
->this_nr_running
) {
4733 sds
->this_load_per_task
/= sds
->this_nr_running
;
4734 if (sds
->busiest_load_per_task
>
4735 sds
->this_load_per_task
)
4738 sds
->this_load_per_task
=
4739 cpu_avg_load_per_task(env
->dst_cpu
);
4742 scaled_busy_load_per_task
= sds
->busiest_load_per_task
4743 * SCHED_POWER_SCALE
;
4744 scaled_busy_load_per_task
/= sds
->busiest
->sgp
->power
;
4746 if (sds
->max_load
- sds
->this_load
+ scaled_busy_load_per_task
>=
4747 (scaled_busy_load_per_task
* imbn
)) {
4748 env
->imbalance
= sds
->busiest_load_per_task
;
4753 * OK, we don't have enough imbalance to justify moving tasks,
4754 * however we may be able to increase total CPU power used by
4758 pwr_now
+= sds
->busiest
->sgp
->power
*
4759 min(sds
->busiest_load_per_task
, sds
->max_load
);
4760 pwr_now
+= sds
->this->sgp
->power
*
4761 min(sds
->this_load_per_task
, sds
->this_load
);
4762 pwr_now
/= SCHED_POWER_SCALE
;
4764 /* Amount of load we'd subtract */
4765 tmp
= (sds
->busiest_load_per_task
* SCHED_POWER_SCALE
) /
4766 sds
->busiest
->sgp
->power
;
4767 if (sds
->max_load
> tmp
)
4768 pwr_move
+= sds
->busiest
->sgp
->power
*
4769 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
4771 /* Amount of load we'd add */
4772 if (sds
->max_load
* sds
->busiest
->sgp
->power
<
4773 sds
->busiest_load_per_task
* SCHED_POWER_SCALE
)
4774 tmp
= (sds
->max_load
* sds
->busiest
->sgp
->power
) /
4775 sds
->this->sgp
->power
;
4777 tmp
= (sds
->busiest_load_per_task
* SCHED_POWER_SCALE
) /
4778 sds
->this->sgp
->power
;
4779 pwr_move
+= sds
->this->sgp
->power
*
4780 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
4781 pwr_move
/= SCHED_POWER_SCALE
;
4783 /* Move if we gain throughput */
4784 if (pwr_move
> pwr_now
)
4785 env
->imbalance
= sds
->busiest_load_per_task
;
4789 * calculate_imbalance - Calculate the amount of imbalance present within the
4790 * groups of a given sched_domain during load balance.
4791 * @env: load balance environment
4792 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4794 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4796 unsigned long max_pull
, load_above_capacity
= ~0UL;
4798 sds
->busiest_load_per_task
/= sds
->busiest_nr_running
;
4799 if (sds
->group_imb
) {
4800 sds
->busiest_load_per_task
=
4801 min(sds
->busiest_load_per_task
, sds
->avg_load
);
4805 * In the presence of smp nice balancing, certain scenarios can have
4806 * max load less than avg load(as we skip the groups at or below
4807 * its cpu_power, while calculating max_load..)
4809 if (sds
->max_load
< sds
->avg_load
) {
4811 return fix_small_imbalance(env
, sds
);
4814 if (!sds
->group_imb
) {
4816 * Don't want to pull so many tasks that a group would go idle.
4818 load_above_capacity
= (sds
->busiest_nr_running
-
4819 sds
->busiest_group_capacity
);
4821 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_POWER_SCALE
);
4823 load_above_capacity
/= sds
->busiest
->sgp
->power
;
4827 * We're trying to get all the cpus to the average_load, so we don't
4828 * want to push ourselves above the average load, nor do we wish to
4829 * reduce the max loaded cpu below the average load. At the same time,
4830 * we also don't want to reduce the group load below the group capacity
4831 * (so that we can implement power-savings policies etc). Thus we look
4832 * for the minimum possible imbalance.
4833 * Be careful of negative numbers as they'll appear as very large values
4834 * with unsigned longs.
4836 max_pull
= min(sds
->max_load
- sds
->avg_load
, load_above_capacity
);
4838 /* How much load to actually move to equalise the imbalance */
4839 env
->imbalance
= min(max_pull
* sds
->busiest
->sgp
->power
,
4840 (sds
->avg_load
- sds
->this_load
) * sds
->this->sgp
->power
)
4841 / SCHED_POWER_SCALE
;
4844 * if *imbalance is less than the average load per runnable task
4845 * there is no guarantee that any tasks will be moved so we'll have
4846 * a think about bumping its value to force at least one task to be
4849 if (env
->imbalance
< sds
->busiest_load_per_task
)
4850 return fix_small_imbalance(env
, sds
);
4854 /******* find_busiest_group() helpers end here *********************/
4857 * find_busiest_group - Returns the busiest group within the sched_domain
4858 * if there is an imbalance. If there isn't an imbalance, and
4859 * the user has opted for power-savings, it returns a group whose
4860 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4861 * such a group exists.
4863 * Also calculates the amount of weighted load which should be moved
4864 * to restore balance.
4866 * @env: The load balancing environment.
4867 * @balance: Pointer to a variable indicating if this_cpu
4868 * is the appropriate cpu to perform load balancing at this_level.
4870 * Returns: - the busiest group if imbalance exists.
4871 * - If no imbalance and user has opted for power-savings balance,
4872 * return the least loaded group whose CPUs can be
4873 * put to idle by rebalancing its tasks onto our group.
4875 static struct sched_group
*
4876 find_busiest_group(struct lb_env
*env
, int *balance
)
4878 struct sd_lb_stats sds
;
4880 memset(&sds
, 0, sizeof(sds
));
4883 * Compute the various statistics relavent for load balancing at
4886 update_sd_lb_stats(env
, balance
, &sds
);
4889 * this_cpu is not the appropriate cpu to perform load balancing at
4895 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
4896 check_asym_packing(env
, &sds
))
4899 /* There is no busy sibling group to pull tasks from */
4900 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
4903 sds
.avg_load
= (SCHED_POWER_SCALE
* sds
.total_load
) / sds
.total_pwr
;
4906 * If the busiest group is imbalanced the below checks don't
4907 * work because they assumes all things are equal, which typically
4908 * isn't true due to cpus_allowed constraints and the like.
4913 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4914 if (env
->idle
== CPU_NEWLY_IDLE
&& sds
.this_has_capacity
&&
4915 !sds
.busiest_has_capacity
)
4919 * If the local group is more busy than the selected busiest group
4920 * don't try and pull any tasks.
4922 if (sds
.this_load
>= sds
.max_load
)
4926 * Don't pull any tasks if this group is already above the domain
4929 if (sds
.this_load
>= sds
.avg_load
)
4932 if (env
->idle
== CPU_IDLE
) {
4934 * This cpu is idle. If the busiest group load doesn't
4935 * have more tasks than the number of available cpu's and
4936 * there is no imbalance between this and busiest group
4937 * wrt to idle cpu's, it is balanced.
4939 if ((sds
.this_idle_cpus
<= sds
.busiest_idle_cpus
+ 1) &&
4940 sds
.busiest_nr_running
<= sds
.busiest_group_weight
)
4944 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4945 * imbalance_pct to be conservative.
4947 if (100 * sds
.max_load
<= env
->sd
->imbalance_pct
* sds
.this_load
)
4952 /* Looks like there is an imbalance. Compute it */
4953 calculate_imbalance(env
, &sds
);
4963 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4965 static struct rq
*find_busiest_queue(struct lb_env
*env
,
4966 struct sched_group
*group
)
4968 struct rq
*busiest
= NULL
, *rq
;
4969 unsigned long max_load
= 0;
4972 for_each_cpu(i
, sched_group_cpus(group
)) {
4973 unsigned long power
= power_of(i
);
4974 unsigned long capacity
= DIV_ROUND_CLOSEST(power
,
4979 capacity
= fix_small_capacity(env
->sd
, group
);
4981 if (!cpumask_test_cpu(i
, env
->cpus
))
4985 wl
= weighted_cpuload(i
);
4988 * When comparing with imbalance, use weighted_cpuload()
4989 * which is not scaled with the cpu power.
4991 if (capacity
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
4995 * For the load comparisons with the other cpu's, consider
4996 * the weighted_cpuload() scaled with the cpu power, so that
4997 * the load can be moved away from the cpu that is potentially
4998 * running at a lower capacity.
5000 wl
= (wl
* SCHED_POWER_SCALE
) / power
;
5002 if (wl
> max_load
) {
5012 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5013 * so long as it is large enough.
5015 #define MAX_PINNED_INTERVAL 512
5017 /* Working cpumask for load_balance and load_balance_newidle. */
5018 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5020 static int need_active_balance(struct lb_env
*env
)
5022 struct sched_domain
*sd
= env
->sd
;
5024 if (env
->idle
== CPU_NEWLY_IDLE
) {
5027 * ASYM_PACKING needs to force migrate tasks from busy but
5028 * higher numbered CPUs in order to pack all tasks in the
5029 * lowest numbered CPUs.
5031 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
5035 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
5038 static int active_load_balance_cpu_stop(void *data
);
5041 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5042 * tasks if there is an imbalance.
5044 static int load_balance(int this_cpu
, struct rq
*this_rq
,
5045 struct sched_domain
*sd
, enum cpu_idle_type idle
,
5048 int ld_moved
, cur_ld_moved
, active_balance
= 0;
5049 struct sched_group
*group
;
5051 unsigned long flags
;
5052 struct cpumask
*cpus
= __get_cpu_var(load_balance_mask
);
5054 struct lb_env env
= {
5056 .dst_cpu
= this_cpu
,
5058 .dst_grpmask
= sched_group_cpus(sd
->groups
),
5060 .loop_break
= sched_nr_migrate_break
,
5065 * For NEWLY_IDLE load_balancing, we don't need to consider
5066 * other cpus in our group
5068 if (idle
== CPU_NEWLY_IDLE
)
5069 env
.dst_grpmask
= NULL
;
5071 cpumask_copy(cpus
, cpu_active_mask
);
5073 schedstat_inc(sd
, lb_count
[idle
]);
5076 group
= find_busiest_group(&env
, balance
);
5082 schedstat_inc(sd
, lb_nobusyg
[idle
]);
5086 busiest
= find_busiest_queue(&env
, group
);
5088 schedstat_inc(sd
, lb_nobusyq
[idle
]);
5092 BUG_ON(busiest
== env
.dst_rq
);
5094 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
5097 if (busiest
->nr_running
> 1) {
5099 * Attempt to move tasks. If find_busiest_group has found
5100 * an imbalance but busiest->nr_running <= 1, the group is
5101 * still unbalanced. ld_moved simply stays zero, so it is
5102 * correctly treated as an imbalance.
5104 env
.flags
|= LBF_ALL_PINNED
;
5105 env
.src_cpu
= busiest
->cpu
;
5106 env
.src_rq
= busiest
;
5107 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
5109 update_h_load(env
.src_cpu
);
5111 local_irq_save(flags
);
5112 double_rq_lock(env
.dst_rq
, busiest
);
5115 * cur_ld_moved - load moved in current iteration
5116 * ld_moved - cumulative load moved across iterations
5118 cur_ld_moved
= move_tasks(&env
);
5119 ld_moved
+= cur_ld_moved
;
5120 double_rq_unlock(env
.dst_rq
, busiest
);
5121 local_irq_restore(flags
);
5124 * some other cpu did the load balance for us.
5126 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
5127 resched_cpu(env
.dst_cpu
);
5129 if (env
.flags
& LBF_NEED_BREAK
) {
5130 env
.flags
&= ~LBF_NEED_BREAK
;
5135 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5136 * us and move them to an alternate dst_cpu in our sched_group
5137 * where they can run. The upper limit on how many times we
5138 * iterate on same src_cpu is dependent on number of cpus in our
5141 * This changes load balance semantics a bit on who can move
5142 * load to a given_cpu. In addition to the given_cpu itself
5143 * (or a ilb_cpu acting on its behalf where given_cpu is
5144 * nohz-idle), we now have balance_cpu in a position to move
5145 * load to given_cpu. In rare situations, this may cause
5146 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5147 * _independently_ and at _same_ time to move some load to
5148 * given_cpu) causing exceess load to be moved to given_cpu.
5149 * This however should not happen so much in practice and
5150 * moreover subsequent load balance cycles should correct the
5151 * excess load moved.
5153 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0) {
5155 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
5156 env
.dst_cpu
= env
.new_dst_cpu
;
5157 env
.flags
&= ~LBF_SOME_PINNED
;
5159 env
.loop_break
= sched_nr_migrate_break
;
5161 /* Prevent to re-select dst_cpu via env's cpus */
5162 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
5165 * Go back to "more_balance" rather than "redo" since we
5166 * need to continue with same src_cpu.
5171 /* All tasks on this runqueue were pinned by CPU affinity */
5172 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
5173 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
5174 if (!cpumask_empty(cpus
)) {
5176 env
.loop_break
= sched_nr_migrate_break
;
5184 schedstat_inc(sd
, lb_failed
[idle
]);
5186 * Increment the failure counter only on periodic balance.
5187 * We do not want newidle balance, which can be very
5188 * frequent, pollute the failure counter causing
5189 * excessive cache_hot migrations and active balances.
5191 if (idle
!= CPU_NEWLY_IDLE
)
5192 sd
->nr_balance_failed
++;
5194 if (need_active_balance(&env
)) {
5195 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
5197 /* don't kick the active_load_balance_cpu_stop,
5198 * if the curr task on busiest cpu can't be
5201 if (!cpumask_test_cpu(this_cpu
,
5202 tsk_cpus_allowed(busiest
->curr
))) {
5203 raw_spin_unlock_irqrestore(&busiest
->lock
,
5205 env
.flags
|= LBF_ALL_PINNED
;
5206 goto out_one_pinned
;
5210 * ->active_balance synchronizes accesses to
5211 * ->active_balance_work. Once set, it's cleared
5212 * only after active load balance is finished.
5214 if (!busiest
->active_balance
) {
5215 busiest
->active_balance
= 1;
5216 busiest
->push_cpu
= this_cpu
;
5219 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
5221 if (active_balance
) {
5222 stop_one_cpu_nowait(cpu_of(busiest
),
5223 active_load_balance_cpu_stop
, busiest
,
5224 &busiest
->active_balance_work
);
5228 * We've kicked active balancing, reset the failure
5231 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
5234 sd
->nr_balance_failed
= 0;
5236 if (likely(!active_balance
)) {
5237 /* We were unbalanced, so reset the balancing interval */
5238 sd
->balance_interval
= sd
->min_interval
;
5241 * If we've begun active balancing, start to back off. This
5242 * case may not be covered by the all_pinned logic if there
5243 * is only 1 task on the busy runqueue (because we don't call
5246 if (sd
->balance_interval
< sd
->max_interval
)
5247 sd
->balance_interval
*= 2;
5253 schedstat_inc(sd
, lb_balanced
[idle
]);
5255 sd
->nr_balance_failed
= 0;
5258 /* tune up the balancing interval */
5259 if (((env
.flags
& LBF_ALL_PINNED
) &&
5260 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
5261 (sd
->balance_interval
< sd
->max_interval
))
5262 sd
->balance_interval
*= 2;
5270 * idle_balance is called by schedule() if this_cpu is about to become
5271 * idle. Attempts to pull tasks from other CPUs.
5273 void idle_balance(int this_cpu
, struct rq
*this_rq
)
5275 struct sched_domain
*sd
;
5276 int pulled_task
= 0;
5277 unsigned long next_balance
= jiffies
+ HZ
;
5279 this_rq
->idle_stamp
= rq_clock(this_rq
);
5281 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
5285 * Drop the rq->lock, but keep IRQ/preempt disabled.
5287 raw_spin_unlock(&this_rq
->lock
);
5289 update_blocked_averages(this_cpu
);
5291 for_each_domain(this_cpu
, sd
) {
5292 unsigned long interval
;
5295 if (!(sd
->flags
& SD_LOAD_BALANCE
))
5298 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
5299 /* If we've pulled tasks over stop searching: */
5300 pulled_task
= load_balance(this_cpu
, this_rq
,
5301 sd
, CPU_NEWLY_IDLE
, &balance
);
5304 interval
= msecs_to_jiffies(sd
->balance_interval
);
5305 if (time_after(next_balance
, sd
->last_balance
+ interval
))
5306 next_balance
= sd
->last_balance
+ interval
;
5308 this_rq
->idle_stamp
= 0;
5314 raw_spin_lock(&this_rq
->lock
);
5316 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
5318 * We are going idle. next_balance may be set based on
5319 * a busy processor. So reset next_balance.
5321 this_rq
->next_balance
= next_balance
;
5326 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5327 * running tasks off the busiest CPU onto idle CPUs. It requires at
5328 * least 1 task to be running on each physical CPU where possible, and
5329 * avoids physical / logical imbalances.
5331 static int active_load_balance_cpu_stop(void *data
)
5333 struct rq
*busiest_rq
= data
;
5334 int busiest_cpu
= cpu_of(busiest_rq
);
5335 int target_cpu
= busiest_rq
->push_cpu
;
5336 struct rq
*target_rq
= cpu_rq(target_cpu
);
5337 struct sched_domain
*sd
;
5339 raw_spin_lock_irq(&busiest_rq
->lock
);
5341 /* make sure the requested cpu hasn't gone down in the meantime */
5342 if (unlikely(busiest_cpu
!= smp_processor_id() ||
5343 !busiest_rq
->active_balance
))
5346 /* Is there any task to move? */
5347 if (busiest_rq
->nr_running
<= 1)
5351 * This condition is "impossible", if it occurs
5352 * we need to fix it. Originally reported by
5353 * Bjorn Helgaas on a 128-cpu setup.
5355 BUG_ON(busiest_rq
== target_rq
);
5357 /* move a task from busiest_rq to target_rq */
5358 double_lock_balance(busiest_rq
, target_rq
);
5360 /* Search for an sd spanning us and the target CPU. */
5362 for_each_domain(target_cpu
, sd
) {
5363 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
5364 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
5369 struct lb_env env
= {
5371 .dst_cpu
= target_cpu
,
5372 .dst_rq
= target_rq
,
5373 .src_cpu
= busiest_rq
->cpu
,
5374 .src_rq
= busiest_rq
,
5378 schedstat_inc(sd
, alb_count
);
5380 if (move_one_task(&env
))
5381 schedstat_inc(sd
, alb_pushed
);
5383 schedstat_inc(sd
, alb_failed
);
5386 double_unlock_balance(busiest_rq
, target_rq
);
5388 busiest_rq
->active_balance
= 0;
5389 raw_spin_unlock_irq(&busiest_rq
->lock
);
5393 #ifdef CONFIG_NO_HZ_COMMON
5395 * idle load balancing details
5396 * - When one of the busy CPUs notice that there may be an idle rebalancing
5397 * needed, they will kick the idle load balancer, which then does idle
5398 * load balancing for all the idle CPUs.
5401 cpumask_var_t idle_cpus_mask
;
5403 unsigned long next_balance
; /* in jiffy units */
5404 } nohz ____cacheline_aligned
;
5406 static inline int find_new_ilb(int call_cpu
)
5408 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
5410 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
5417 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5418 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5419 * CPU (if there is one).
5421 static void nohz_balancer_kick(int cpu
)
5425 nohz
.next_balance
++;
5427 ilb_cpu
= find_new_ilb(cpu
);
5429 if (ilb_cpu
>= nr_cpu_ids
)
5432 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
5435 * Use smp_send_reschedule() instead of resched_cpu().
5436 * This way we generate a sched IPI on the target cpu which
5437 * is idle. And the softirq performing nohz idle load balance
5438 * will be run before returning from the IPI.
5440 smp_send_reschedule(ilb_cpu
);
5444 static inline void nohz_balance_exit_idle(int cpu
)
5446 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
5447 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
5448 atomic_dec(&nohz
.nr_cpus
);
5449 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
5453 static inline void set_cpu_sd_state_busy(void)
5455 struct sched_domain
*sd
;
5458 sd
= rcu_dereference_check_sched_domain(this_rq()->sd
);
5460 if (!sd
|| !sd
->nohz_idle
)
5464 for (; sd
; sd
= sd
->parent
)
5465 atomic_inc(&sd
->groups
->sgp
->nr_busy_cpus
);
5470 void set_cpu_sd_state_idle(void)
5472 struct sched_domain
*sd
;
5475 sd
= rcu_dereference_check_sched_domain(this_rq()->sd
);
5477 if (!sd
|| sd
->nohz_idle
)
5481 for (; sd
; sd
= sd
->parent
)
5482 atomic_dec(&sd
->groups
->sgp
->nr_busy_cpus
);
5488 * This routine will record that the cpu is going idle with tick stopped.
5489 * This info will be used in performing idle load balancing in the future.
5491 void nohz_balance_enter_idle(int cpu
)
5494 * If this cpu is going down, then nothing needs to be done.
5496 if (!cpu_active(cpu
))
5499 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
5502 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
5503 atomic_inc(&nohz
.nr_cpus
);
5504 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
5507 static int __cpuinit
sched_ilb_notifier(struct notifier_block
*nfb
,
5508 unsigned long action
, void *hcpu
)
5510 switch (action
& ~CPU_TASKS_FROZEN
) {
5512 nohz_balance_exit_idle(smp_processor_id());
5520 static DEFINE_SPINLOCK(balancing
);
5523 * Scale the max load_balance interval with the number of CPUs in the system.
5524 * This trades load-balance latency on larger machines for less cross talk.
5526 void update_max_interval(void)
5528 max_load_balance_interval
= HZ
*num_online_cpus()/10;
5532 * It checks each scheduling domain to see if it is due to be balanced,
5533 * and initiates a balancing operation if so.
5535 * Balancing parameters are set up in init_sched_domains.
5537 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
5540 struct rq
*rq
= cpu_rq(cpu
);
5541 unsigned long interval
;
5542 struct sched_domain
*sd
;
5543 /* Earliest time when we have to do rebalance again */
5544 unsigned long next_balance
= jiffies
+ 60*HZ
;
5545 int update_next_balance
= 0;
5548 update_blocked_averages(cpu
);
5551 for_each_domain(cpu
, sd
) {
5552 if (!(sd
->flags
& SD_LOAD_BALANCE
))
5555 interval
= sd
->balance_interval
;
5556 if (idle
!= CPU_IDLE
)
5557 interval
*= sd
->busy_factor
;
5559 /* scale ms to jiffies */
5560 interval
= msecs_to_jiffies(interval
);
5561 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5563 need_serialize
= sd
->flags
& SD_SERIALIZE
;
5565 if (need_serialize
) {
5566 if (!spin_trylock(&balancing
))
5570 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
5571 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
5573 * The LBF_SOME_PINNED logic could have changed
5574 * env->dst_cpu, so we can't know our idle
5575 * state even if we migrated tasks. Update it.
5577 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
5579 sd
->last_balance
= jiffies
;
5582 spin_unlock(&balancing
);
5584 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
5585 next_balance
= sd
->last_balance
+ interval
;
5586 update_next_balance
= 1;
5590 * Stop the load balance at this level. There is another
5591 * CPU in our sched group which is doing load balancing more
5600 * next_balance will be updated only when there is a need.
5601 * When the cpu is attached to null domain for ex, it will not be
5604 if (likely(update_next_balance
))
5605 rq
->next_balance
= next_balance
;
5608 #ifdef CONFIG_NO_HZ_COMMON
5610 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5611 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5613 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
)
5615 struct rq
*this_rq
= cpu_rq(this_cpu
);
5619 if (idle
!= CPU_IDLE
||
5620 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
5623 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
5624 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
5628 * If this cpu gets work to do, stop the load balancing
5629 * work being done for other cpus. Next load
5630 * balancing owner will pick it up.
5635 rq
= cpu_rq(balance_cpu
);
5637 raw_spin_lock_irq(&rq
->lock
);
5638 update_rq_clock(rq
);
5639 update_idle_cpu_load(rq
);
5640 raw_spin_unlock_irq(&rq
->lock
);
5642 rebalance_domains(balance_cpu
, CPU_IDLE
);
5644 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
5645 this_rq
->next_balance
= rq
->next_balance
;
5647 nohz
.next_balance
= this_rq
->next_balance
;
5649 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
5653 * Current heuristic for kicking the idle load balancer in the presence
5654 * of an idle cpu is the system.
5655 * - This rq has more than one task.
5656 * - At any scheduler domain level, this cpu's scheduler group has multiple
5657 * busy cpu's exceeding the group's power.
5658 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5659 * domain span are idle.
5661 static inline int nohz_kick_needed(struct rq
*rq
, int cpu
)
5663 unsigned long now
= jiffies
;
5664 struct sched_domain
*sd
;
5666 if (unlikely(idle_cpu(cpu
)))
5670 * We may be recently in ticked or tickless idle mode. At the first
5671 * busy tick after returning from idle, we will update the busy stats.
5673 set_cpu_sd_state_busy();
5674 nohz_balance_exit_idle(cpu
);
5677 * None are in tickless mode and hence no need for NOHZ idle load
5680 if (likely(!atomic_read(&nohz
.nr_cpus
)))
5683 if (time_before(now
, nohz
.next_balance
))
5686 if (rq
->nr_running
>= 2)
5690 for_each_domain(cpu
, sd
) {
5691 struct sched_group
*sg
= sd
->groups
;
5692 struct sched_group_power
*sgp
= sg
->sgp
;
5693 int nr_busy
= atomic_read(&sgp
->nr_busy_cpus
);
5695 if (sd
->flags
& SD_SHARE_PKG_RESOURCES
&& nr_busy
> 1)
5696 goto need_kick_unlock
;
5698 if (sd
->flags
& SD_ASYM_PACKING
&& nr_busy
!= sg
->group_weight
5699 && (cpumask_first_and(nohz
.idle_cpus_mask
,
5700 sched_domain_span(sd
)) < cpu
))
5701 goto need_kick_unlock
;
5703 if (!(sd
->flags
& (SD_SHARE_PKG_RESOURCES
| SD_ASYM_PACKING
)))
5715 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
) { }
5719 * run_rebalance_domains is triggered when needed from the scheduler tick.
5720 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5722 static void run_rebalance_domains(struct softirq_action
*h
)
5724 int this_cpu
= smp_processor_id();
5725 struct rq
*this_rq
= cpu_rq(this_cpu
);
5726 enum cpu_idle_type idle
= this_rq
->idle_balance
?
5727 CPU_IDLE
: CPU_NOT_IDLE
;
5729 rebalance_domains(this_cpu
, idle
);
5732 * If this cpu has a pending nohz_balance_kick, then do the
5733 * balancing on behalf of the other idle cpus whose ticks are
5736 nohz_idle_balance(this_cpu
, idle
);
5739 static inline int on_null_domain(int cpu
)
5741 return !rcu_dereference_sched(cpu_rq(cpu
)->sd
);
5745 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5747 void trigger_load_balance(struct rq
*rq
, int cpu
)
5749 /* Don't need to rebalance while attached to NULL domain */
5750 if (time_after_eq(jiffies
, rq
->next_balance
) &&
5751 likely(!on_null_domain(cpu
)))
5752 raise_softirq(SCHED_SOFTIRQ
);
5753 #ifdef CONFIG_NO_HZ_COMMON
5754 if (nohz_kick_needed(rq
, cpu
) && likely(!on_null_domain(cpu
)))
5755 nohz_balancer_kick(cpu
);
5759 static void rq_online_fair(struct rq
*rq
)
5764 static void rq_offline_fair(struct rq
*rq
)
5768 /* Ensure any throttled groups are reachable by pick_next_task */
5769 unthrottle_offline_cfs_rqs(rq
);
5772 #endif /* CONFIG_SMP */
5775 * scheduler tick hitting a task of our scheduling class:
5777 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
5779 struct cfs_rq
*cfs_rq
;
5780 struct sched_entity
*se
= &curr
->se
;
5782 for_each_sched_entity(se
) {
5783 cfs_rq
= cfs_rq_of(se
);
5784 entity_tick(cfs_rq
, se
, queued
);
5787 if (sched_feat_numa(NUMA
))
5788 task_tick_numa(rq
, curr
);
5790 update_rq_runnable_avg(rq
, 1);
5794 * called on fork with the child task as argument from the parent's context
5795 * - child not yet on the tasklist
5796 * - preemption disabled
5798 static void task_fork_fair(struct task_struct
*p
)
5800 struct cfs_rq
*cfs_rq
;
5801 struct sched_entity
*se
= &p
->se
, *curr
;
5802 int this_cpu
= smp_processor_id();
5803 struct rq
*rq
= this_rq();
5804 unsigned long flags
;
5806 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5808 update_rq_clock(rq
);
5810 cfs_rq
= task_cfs_rq(current
);
5811 curr
= cfs_rq
->curr
;
5813 if (unlikely(task_cpu(p
) != this_cpu
)) {
5815 __set_task_cpu(p
, this_cpu
);
5819 update_curr(cfs_rq
);
5822 se
->vruntime
= curr
->vruntime
;
5823 place_entity(cfs_rq
, se
, 1);
5825 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
5827 * Upon rescheduling, sched_class::put_prev_task() will place
5828 * 'current' within the tree based on its new key value.
5830 swap(curr
->vruntime
, se
->vruntime
);
5831 resched_task(rq
->curr
);
5834 se
->vruntime
-= cfs_rq
->min_vruntime
;
5836 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5840 * Priority of the task has changed. Check to see if we preempt
5844 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
5850 * Reschedule if we are currently running on this runqueue and
5851 * our priority decreased, or if we are not currently running on
5852 * this runqueue and our priority is higher than the current's
5854 if (rq
->curr
== p
) {
5855 if (p
->prio
> oldprio
)
5856 resched_task(rq
->curr
);
5858 check_preempt_curr(rq
, p
, 0);
5861 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
5863 struct sched_entity
*se
= &p
->se
;
5864 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5867 * Ensure the task's vruntime is normalized, so that when its
5868 * switched back to the fair class the enqueue_entity(.flags=0) will
5869 * do the right thing.
5871 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5872 * have normalized the vruntime, if it was !on_rq, then only when
5873 * the task is sleeping will it still have non-normalized vruntime.
5875 if (!se
->on_rq
&& p
->state
!= TASK_RUNNING
) {
5877 * Fix up our vruntime so that the current sleep doesn't
5878 * cause 'unlimited' sleep bonus.
5880 place_entity(cfs_rq
, se
, 0);
5881 se
->vruntime
-= cfs_rq
->min_vruntime
;
5886 * Remove our load from contribution when we leave sched_fair
5887 * and ensure we don't carry in an old decay_count if we
5890 if (p
->se
.avg
.decay_count
) {
5891 struct cfs_rq
*cfs_rq
= cfs_rq_of(&p
->se
);
5892 __synchronize_entity_decay(&p
->se
);
5893 subtract_blocked_load_contrib(cfs_rq
,
5894 p
->se
.avg
.load_avg_contrib
);
5900 * We switched to the sched_fair class.
5902 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
5908 * We were most likely switched from sched_rt, so
5909 * kick off the schedule if running, otherwise just see
5910 * if we can still preempt the current task.
5913 resched_task(rq
->curr
);
5915 check_preempt_curr(rq
, p
, 0);
5918 /* Account for a task changing its policy or group.
5920 * This routine is mostly called to set cfs_rq->curr field when a task
5921 * migrates between groups/classes.
5923 static void set_curr_task_fair(struct rq
*rq
)
5925 struct sched_entity
*se
= &rq
->curr
->se
;
5927 for_each_sched_entity(se
) {
5928 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5930 set_next_entity(cfs_rq
, se
);
5931 /* ensure bandwidth has been allocated on our new cfs_rq */
5932 account_cfs_rq_runtime(cfs_rq
, 0);
5936 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
5938 cfs_rq
->tasks_timeline
= RB_ROOT
;
5939 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
5940 #ifndef CONFIG_64BIT
5941 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
5944 atomic64_set(&cfs_rq
->decay_counter
, 1);
5945 atomic64_set(&cfs_rq
->removed_load
, 0);
5949 #ifdef CONFIG_FAIR_GROUP_SCHED
5950 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
5952 struct cfs_rq
*cfs_rq
;
5954 * If the task was not on the rq at the time of this cgroup movement
5955 * it must have been asleep, sleeping tasks keep their ->vruntime
5956 * absolute on their old rq until wakeup (needed for the fair sleeper
5957 * bonus in place_entity()).
5959 * If it was on the rq, we've just 'preempted' it, which does convert
5960 * ->vruntime to a relative base.
5962 * Make sure both cases convert their relative position when migrating
5963 * to another cgroup's rq. This does somewhat interfere with the
5964 * fair sleeper stuff for the first placement, but who cares.
5967 * When !on_rq, vruntime of the task has usually NOT been normalized.
5968 * But there are some cases where it has already been normalized:
5970 * - Moving a forked child which is waiting for being woken up by
5971 * wake_up_new_task().
5972 * - Moving a task which has been woken up by try_to_wake_up() and
5973 * waiting for actually being woken up by sched_ttwu_pending().
5975 * To prevent boost or penalty in the new cfs_rq caused by delta
5976 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5978 if (!on_rq
&& (!p
->se
.sum_exec_runtime
|| p
->state
== TASK_WAKING
))
5982 p
->se
.vruntime
-= cfs_rq_of(&p
->se
)->min_vruntime
;
5983 set_task_rq(p
, task_cpu(p
));
5985 cfs_rq
= cfs_rq_of(&p
->se
);
5986 p
->se
.vruntime
+= cfs_rq
->min_vruntime
;
5989 * migrate_task_rq_fair() will have removed our previous
5990 * contribution, but we must synchronize for ongoing future
5993 p
->se
.avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
5994 cfs_rq
->blocked_load_avg
+= p
->se
.avg
.load_avg_contrib
;
5999 void free_fair_sched_group(struct task_group
*tg
)
6003 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
6005 for_each_possible_cpu(i
) {
6007 kfree(tg
->cfs_rq
[i
]);
6016 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
6018 struct cfs_rq
*cfs_rq
;
6019 struct sched_entity
*se
;
6022 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
6025 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
6029 tg
->shares
= NICE_0_LOAD
;
6031 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
6033 for_each_possible_cpu(i
) {
6034 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
6035 GFP_KERNEL
, cpu_to_node(i
));
6039 se
= kzalloc_node(sizeof(struct sched_entity
),
6040 GFP_KERNEL
, cpu_to_node(i
));
6044 init_cfs_rq(cfs_rq
);
6045 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
6056 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
6058 struct rq
*rq
= cpu_rq(cpu
);
6059 unsigned long flags
;
6062 * Only empty task groups can be destroyed; so we can speculatively
6063 * check on_list without danger of it being re-added.
6065 if (!tg
->cfs_rq
[cpu
]->on_list
)
6068 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6069 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
6070 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6073 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
6074 struct sched_entity
*se
, int cpu
,
6075 struct sched_entity
*parent
)
6077 struct rq
*rq
= cpu_rq(cpu
);
6081 init_cfs_rq_runtime(cfs_rq
);
6083 tg
->cfs_rq
[cpu
] = cfs_rq
;
6086 /* se could be NULL for root_task_group */
6091 se
->cfs_rq
= &rq
->cfs
;
6093 se
->cfs_rq
= parent
->my_q
;
6096 update_load_set(&se
->load
, 0);
6097 se
->parent
= parent
;
6100 static DEFINE_MUTEX(shares_mutex
);
6102 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
6105 unsigned long flags
;
6108 * We can't change the weight of the root cgroup.
6113 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
6115 mutex_lock(&shares_mutex
);
6116 if (tg
->shares
== shares
)
6119 tg
->shares
= shares
;
6120 for_each_possible_cpu(i
) {
6121 struct rq
*rq
= cpu_rq(i
);
6122 struct sched_entity
*se
;
6125 /* Propagate contribution to hierarchy */
6126 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6128 /* Possible calls to update_curr() need rq clock */
6129 update_rq_clock(rq
);
6130 for_each_sched_entity(se
)
6131 update_cfs_shares(group_cfs_rq(se
));
6132 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6136 mutex_unlock(&shares_mutex
);
6139 #else /* CONFIG_FAIR_GROUP_SCHED */
6141 void free_fair_sched_group(struct task_group
*tg
) { }
6143 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
6148 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
6150 #endif /* CONFIG_FAIR_GROUP_SCHED */
6153 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
6155 struct sched_entity
*se
= &task
->se
;
6156 unsigned int rr_interval
= 0;
6159 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6162 if (rq
->cfs
.load
.weight
)
6163 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
6169 * All the scheduling class methods:
6171 const struct sched_class fair_sched_class
= {
6172 .next
= &idle_sched_class
,
6173 .enqueue_task
= enqueue_task_fair
,
6174 .dequeue_task
= dequeue_task_fair
,
6175 .yield_task
= yield_task_fair
,
6176 .yield_to_task
= yield_to_task_fair
,
6178 .check_preempt_curr
= check_preempt_wakeup
,
6180 .pick_next_task
= pick_next_task_fair
,
6181 .put_prev_task
= put_prev_task_fair
,
6184 .select_task_rq
= select_task_rq_fair
,
6185 .migrate_task_rq
= migrate_task_rq_fair
,
6187 .rq_online
= rq_online_fair
,
6188 .rq_offline
= rq_offline_fair
,
6190 .task_waking
= task_waking_fair
,
6193 .set_curr_task
= set_curr_task_fair
,
6194 .task_tick
= task_tick_fair
,
6195 .task_fork
= task_fork_fair
,
6197 .prio_changed
= prio_changed_fair
,
6198 .switched_from
= switched_from_fair
,
6199 .switched_to
= switched_to_fair
,
6201 .get_rr_interval
= get_rr_interval_fair
,
6203 #ifdef CONFIG_FAIR_GROUP_SCHED
6204 .task_move_group
= task_move_group_fair
,
6208 #ifdef CONFIG_SCHED_DEBUG
6209 void print_cfs_stats(struct seq_file
*m
, int cpu
)
6211 struct cfs_rq
*cfs_rq
;
6214 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
6215 print_cfs_rq(m
, cpu
, cfs_rq
);
6220 __init
void init_sched_fair_class(void)
6223 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
6225 #ifdef CONFIG_NO_HZ_COMMON
6226 nohz
.next_balance
= jiffies
;
6227 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
6228 cpu_notifier(sched_ilb_notifier
, 0);