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
= atomic64_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
|| abs64(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
1365 atomic64_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
= div64_u64(contrib
,
1401 atomic64_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;
1574 __synchronize_entity_decay(se
);
1577 /* migrated tasks did not contribute to our blocked load */
1579 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
1580 update_entity_load_avg(se
, 0);
1583 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
1584 /* we force update consideration on load-balancer moves */
1585 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
1589 * Remove se's load from this cfs_rq child load-average, if the entity is
1590 * transitioning to a blocked state we track its projected decay using
1593 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
1594 struct sched_entity
*se
,
1597 update_entity_load_avg(se
, 1);
1598 /* we force update consideration on load-balancer moves */
1599 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
1601 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
1603 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
1604 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
1605 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1609 * Update the rq's load with the elapsed running time before entering
1610 * idle. if the last scheduled task is not a CFS task, idle_enter will
1611 * be the only way to update the runnable statistic.
1613 void idle_enter_fair(struct rq
*this_rq
)
1615 update_rq_runnable_avg(this_rq
, 1);
1619 * Update the rq's load with the elapsed idle time before a task is
1620 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1621 * be the only way to update the runnable statistic.
1623 void idle_exit_fair(struct rq
*this_rq
)
1625 update_rq_runnable_avg(this_rq
, 0);
1629 static inline void update_entity_load_avg(struct sched_entity
*se
,
1630 int update_cfs_rq
) {}
1631 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
1632 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
1633 struct sched_entity
*se
,
1635 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
1636 struct sched_entity
*se
,
1638 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
1639 int force_update
) {}
1642 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1644 #ifdef CONFIG_SCHEDSTATS
1645 struct task_struct
*tsk
= NULL
;
1647 if (entity_is_task(se
))
1650 if (se
->statistics
.sleep_start
) {
1651 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
1656 if (unlikely(delta
> se
->statistics
.sleep_max
))
1657 se
->statistics
.sleep_max
= delta
;
1659 se
->statistics
.sleep_start
= 0;
1660 se
->statistics
.sum_sleep_runtime
+= delta
;
1663 account_scheduler_latency(tsk
, delta
>> 10, 1);
1664 trace_sched_stat_sleep(tsk
, delta
);
1667 if (se
->statistics
.block_start
) {
1668 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
1673 if (unlikely(delta
> se
->statistics
.block_max
))
1674 se
->statistics
.block_max
= delta
;
1676 se
->statistics
.block_start
= 0;
1677 se
->statistics
.sum_sleep_runtime
+= delta
;
1680 if (tsk
->in_iowait
) {
1681 se
->statistics
.iowait_sum
+= delta
;
1682 se
->statistics
.iowait_count
++;
1683 trace_sched_stat_iowait(tsk
, delta
);
1686 trace_sched_stat_blocked(tsk
, delta
);
1689 * Blocking time is in units of nanosecs, so shift by
1690 * 20 to get a milliseconds-range estimation of the
1691 * amount of time that the task spent sleeping:
1693 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
1694 profile_hits(SLEEP_PROFILING
,
1695 (void *)get_wchan(tsk
),
1698 account_scheduler_latency(tsk
, delta
>> 10, 0);
1704 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1706 #ifdef CONFIG_SCHED_DEBUG
1707 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
1712 if (d
> 3*sysctl_sched_latency
)
1713 schedstat_inc(cfs_rq
, nr_spread_over
);
1718 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
1720 u64 vruntime
= cfs_rq
->min_vruntime
;
1723 * The 'current' period is already promised to the current tasks,
1724 * however the extra weight of the new task will slow them down a
1725 * little, place the new task so that it fits in the slot that
1726 * stays open at the end.
1728 if (initial
&& sched_feat(START_DEBIT
))
1729 vruntime
+= sched_vslice(cfs_rq
, se
);
1731 /* sleeps up to a single latency don't count. */
1733 unsigned long thresh
= sysctl_sched_latency
;
1736 * Halve their sleep time's effect, to allow
1737 * for a gentler effect of sleepers:
1739 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
1745 /* ensure we never gain time by being placed backwards. */
1746 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
1749 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
1752 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1755 * Update the normalized vruntime before updating min_vruntime
1756 * through callig update_curr().
1758 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
1759 se
->vruntime
+= cfs_rq
->min_vruntime
;
1762 * Update run-time statistics of the 'current'.
1764 update_curr(cfs_rq
);
1765 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
1766 account_entity_enqueue(cfs_rq
, se
);
1767 update_cfs_shares(cfs_rq
);
1769 if (flags
& ENQUEUE_WAKEUP
) {
1770 place_entity(cfs_rq
, se
, 0);
1771 enqueue_sleeper(cfs_rq
, se
);
1774 update_stats_enqueue(cfs_rq
, se
);
1775 check_spread(cfs_rq
, se
);
1776 if (se
!= cfs_rq
->curr
)
1777 __enqueue_entity(cfs_rq
, se
);
1780 if (cfs_rq
->nr_running
== 1) {
1781 list_add_leaf_cfs_rq(cfs_rq
);
1782 check_enqueue_throttle(cfs_rq
);
1786 static void __clear_buddies_last(struct sched_entity
*se
)
1788 for_each_sched_entity(se
) {
1789 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1790 if (cfs_rq
->last
== se
)
1791 cfs_rq
->last
= NULL
;
1797 static void __clear_buddies_next(struct sched_entity
*se
)
1799 for_each_sched_entity(se
) {
1800 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1801 if (cfs_rq
->next
== se
)
1802 cfs_rq
->next
= NULL
;
1808 static void __clear_buddies_skip(struct sched_entity
*se
)
1810 for_each_sched_entity(se
) {
1811 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1812 if (cfs_rq
->skip
== se
)
1813 cfs_rq
->skip
= NULL
;
1819 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1821 if (cfs_rq
->last
== se
)
1822 __clear_buddies_last(se
);
1824 if (cfs_rq
->next
== se
)
1825 __clear_buddies_next(se
);
1827 if (cfs_rq
->skip
== se
)
1828 __clear_buddies_skip(se
);
1831 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
1834 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1837 * Update run-time statistics of the 'current'.
1839 update_curr(cfs_rq
);
1840 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
1842 update_stats_dequeue(cfs_rq
, se
);
1843 if (flags
& DEQUEUE_SLEEP
) {
1844 #ifdef CONFIG_SCHEDSTATS
1845 if (entity_is_task(se
)) {
1846 struct task_struct
*tsk
= task_of(se
);
1848 if (tsk
->state
& TASK_INTERRUPTIBLE
)
1849 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
1850 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
1851 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
1856 clear_buddies(cfs_rq
, se
);
1858 if (se
!= cfs_rq
->curr
)
1859 __dequeue_entity(cfs_rq
, se
);
1861 account_entity_dequeue(cfs_rq
, se
);
1864 * Normalize the entity after updating the min_vruntime because the
1865 * update can refer to the ->curr item and we need to reflect this
1866 * movement in our normalized position.
1868 if (!(flags
& DEQUEUE_SLEEP
))
1869 se
->vruntime
-= cfs_rq
->min_vruntime
;
1871 /* return excess runtime on last dequeue */
1872 return_cfs_rq_runtime(cfs_rq
);
1874 update_min_vruntime(cfs_rq
);
1875 update_cfs_shares(cfs_rq
);
1879 * Preempt the current task with a newly woken task if needed:
1882 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
1884 unsigned long ideal_runtime
, delta_exec
;
1885 struct sched_entity
*se
;
1888 ideal_runtime
= sched_slice(cfs_rq
, curr
);
1889 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
1890 if (delta_exec
> ideal_runtime
) {
1891 resched_task(rq_of(cfs_rq
)->curr
);
1893 * The current task ran long enough, ensure it doesn't get
1894 * re-elected due to buddy favours.
1896 clear_buddies(cfs_rq
, curr
);
1901 * Ensure that a task that missed wakeup preemption by a
1902 * narrow margin doesn't have to wait for a full slice.
1903 * This also mitigates buddy induced latencies under load.
1905 if (delta_exec
< sysctl_sched_min_granularity
)
1908 se
= __pick_first_entity(cfs_rq
);
1909 delta
= curr
->vruntime
- se
->vruntime
;
1914 if (delta
> ideal_runtime
)
1915 resched_task(rq_of(cfs_rq
)->curr
);
1919 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1921 /* 'current' is not kept within the tree. */
1924 * Any task has to be enqueued before it get to execute on
1925 * a CPU. So account for the time it spent waiting on the
1928 update_stats_wait_end(cfs_rq
, se
);
1929 __dequeue_entity(cfs_rq
, se
);
1932 update_stats_curr_start(cfs_rq
, se
);
1934 #ifdef CONFIG_SCHEDSTATS
1936 * Track our maximum slice length, if the CPU's load is at
1937 * least twice that of our own weight (i.e. dont track it
1938 * when there are only lesser-weight tasks around):
1940 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
1941 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
1942 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
1945 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
1949 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
1952 * Pick the next process, keeping these things in mind, in this order:
1953 * 1) keep things fair between processes/task groups
1954 * 2) pick the "next" process, since someone really wants that to run
1955 * 3) pick the "last" process, for cache locality
1956 * 4) do not run the "skip" process, if something else is available
1958 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
1960 struct sched_entity
*se
= __pick_first_entity(cfs_rq
);
1961 struct sched_entity
*left
= se
;
1964 * Avoid running the skip buddy, if running something else can
1965 * be done without getting too unfair.
1967 if (cfs_rq
->skip
== se
) {
1968 struct sched_entity
*second
= __pick_next_entity(se
);
1969 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
1974 * Prefer last buddy, try to return the CPU to a preempted task.
1976 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
1980 * Someone really wants this to run. If it's not unfair, run it.
1982 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
1985 clear_buddies(cfs_rq
, se
);
1990 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
1992 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
1995 * If still on the runqueue then deactivate_task()
1996 * was not called and update_curr() has to be done:
1999 update_curr(cfs_rq
);
2001 /* throttle cfs_rqs exceeding runtime */
2002 check_cfs_rq_runtime(cfs_rq
);
2004 check_spread(cfs_rq
, prev
);
2006 update_stats_wait_start(cfs_rq
, prev
);
2007 /* Put 'current' back into the tree. */
2008 __enqueue_entity(cfs_rq
, prev
);
2009 /* in !on_rq case, update occurred at dequeue */
2010 update_entity_load_avg(prev
, 1);
2012 cfs_rq
->curr
= NULL
;
2016 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
2019 * Update run-time statistics of the 'current'.
2021 update_curr(cfs_rq
);
2024 * Ensure that runnable average is periodically updated.
2026 update_entity_load_avg(curr
, 1);
2027 update_cfs_rq_blocked_load(cfs_rq
, 1);
2029 #ifdef CONFIG_SCHED_HRTICK
2031 * queued ticks are scheduled to match the slice, so don't bother
2032 * validating it and just reschedule.
2035 resched_task(rq_of(cfs_rq
)->curr
);
2039 * don't let the period tick interfere with the hrtick preemption
2041 if (!sched_feat(DOUBLE_TICK
) &&
2042 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
2046 if (cfs_rq
->nr_running
> 1)
2047 check_preempt_tick(cfs_rq
, curr
);
2051 /**************************************************
2052 * CFS bandwidth control machinery
2055 #ifdef CONFIG_CFS_BANDWIDTH
2057 #ifdef HAVE_JUMP_LABEL
2058 static struct static_key __cfs_bandwidth_used
;
2060 static inline bool cfs_bandwidth_used(void)
2062 return static_key_false(&__cfs_bandwidth_used
);
2065 void account_cfs_bandwidth_used(int enabled
, int was_enabled
)
2067 /* only need to count groups transitioning between enabled/!enabled */
2068 if (enabled
&& !was_enabled
)
2069 static_key_slow_inc(&__cfs_bandwidth_used
);
2070 else if (!enabled
&& was_enabled
)
2071 static_key_slow_dec(&__cfs_bandwidth_used
);
2073 #else /* HAVE_JUMP_LABEL */
2074 static bool cfs_bandwidth_used(void)
2079 void account_cfs_bandwidth_used(int enabled
, int was_enabled
) {}
2080 #endif /* HAVE_JUMP_LABEL */
2083 * default period for cfs group bandwidth.
2084 * default: 0.1s, units: nanoseconds
2086 static inline u64
default_cfs_period(void)
2088 return 100000000ULL;
2091 static inline u64
sched_cfs_bandwidth_slice(void)
2093 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
2097 * Replenish runtime according to assigned quota and update expiration time.
2098 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2099 * additional synchronization around rq->lock.
2101 * requires cfs_b->lock
2103 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
2107 if (cfs_b
->quota
== RUNTIME_INF
)
2110 now
= sched_clock_cpu(smp_processor_id());
2111 cfs_b
->runtime
= cfs_b
->quota
;
2112 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
2115 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
2117 return &tg
->cfs_bandwidth
;
2120 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2121 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
2123 if (unlikely(cfs_rq
->throttle_count
))
2124 return cfs_rq
->throttled_clock_task
;
2126 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
2129 /* returns 0 on failure to allocate runtime */
2130 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2132 struct task_group
*tg
= cfs_rq
->tg
;
2133 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
2134 u64 amount
= 0, min_amount
, expires
;
2136 /* note: this is a positive sum as runtime_remaining <= 0 */
2137 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
2139 raw_spin_lock(&cfs_b
->lock
);
2140 if (cfs_b
->quota
== RUNTIME_INF
)
2141 amount
= min_amount
;
2144 * If the bandwidth pool has become inactive, then at least one
2145 * period must have elapsed since the last consumption.
2146 * Refresh the global state and ensure bandwidth timer becomes
2149 if (!cfs_b
->timer_active
) {
2150 __refill_cfs_bandwidth_runtime(cfs_b
);
2151 __start_cfs_bandwidth(cfs_b
);
2154 if (cfs_b
->runtime
> 0) {
2155 amount
= min(cfs_b
->runtime
, min_amount
);
2156 cfs_b
->runtime
-= amount
;
2160 expires
= cfs_b
->runtime_expires
;
2161 raw_spin_unlock(&cfs_b
->lock
);
2163 cfs_rq
->runtime_remaining
+= amount
;
2165 * we may have advanced our local expiration to account for allowed
2166 * spread between our sched_clock and the one on which runtime was
2169 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
2170 cfs_rq
->runtime_expires
= expires
;
2172 return cfs_rq
->runtime_remaining
> 0;
2176 * Note: This depends on the synchronization provided by sched_clock and the
2177 * fact that rq->clock snapshots this value.
2179 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2181 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2183 /* if the deadline is ahead of our clock, nothing to do */
2184 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
2187 if (cfs_rq
->runtime_remaining
< 0)
2191 * If the local deadline has passed we have to consider the
2192 * possibility that our sched_clock is 'fast' and the global deadline
2193 * has not truly expired.
2195 * Fortunately we can check determine whether this the case by checking
2196 * whether the global deadline has advanced.
2199 if ((s64
)(cfs_rq
->runtime_expires
- cfs_b
->runtime_expires
) >= 0) {
2200 /* extend local deadline, drift is bounded above by 2 ticks */
2201 cfs_rq
->runtime_expires
+= TICK_NSEC
;
2203 /* global deadline is ahead, expiration has passed */
2204 cfs_rq
->runtime_remaining
= 0;
2208 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
,
2209 unsigned long delta_exec
)
2211 /* dock delta_exec before expiring quota (as it could span periods) */
2212 cfs_rq
->runtime_remaining
-= delta_exec
;
2213 expire_cfs_rq_runtime(cfs_rq
);
2215 if (likely(cfs_rq
->runtime_remaining
> 0))
2219 * if we're unable to extend our runtime we resched so that the active
2220 * hierarchy can be throttled
2222 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
2223 resched_task(rq_of(cfs_rq
)->curr
);
2226 static __always_inline
2227 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
)
2229 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
2232 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
2235 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
2237 return cfs_bandwidth_used() && cfs_rq
->throttled
;
2240 /* check whether cfs_rq, or any parent, is throttled */
2241 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
2243 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
2247 * Ensure that neither of the group entities corresponding to src_cpu or
2248 * dest_cpu are members of a throttled hierarchy when performing group
2249 * load-balance operations.
2251 static inline int throttled_lb_pair(struct task_group
*tg
,
2252 int src_cpu
, int dest_cpu
)
2254 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
2256 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
2257 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
2259 return throttled_hierarchy(src_cfs_rq
) ||
2260 throttled_hierarchy(dest_cfs_rq
);
2263 /* updated child weight may affect parent so we have to do this bottom up */
2264 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
2266 struct rq
*rq
= data
;
2267 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
2269 cfs_rq
->throttle_count
--;
2271 if (!cfs_rq
->throttle_count
) {
2272 /* adjust cfs_rq_clock_task() */
2273 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
2274 cfs_rq
->throttled_clock_task
;
2281 static int tg_throttle_down(struct task_group
*tg
, void *data
)
2283 struct rq
*rq
= data
;
2284 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
2286 /* group is entering throttled state, stop time */
2287 if (!cfs_rq
->throttle_count
)
2288 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
2289 cfs_rq
->throttle_count
++;
2294 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
2296 struct rq
*rq
= rq_of(cfs_rq
);
2297 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2298 struct sched_entity
*se
;
2299 long task_delta
, dequeue
= 1;
2301 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
2303 /* freeze hierarchy runnable averages while throttled */
2305 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
2308 task_delta
= cfs_rq
->h_nr_running
;
2309 for_each_sched_entity(se
) {
2310 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
2311 /* throttled entity or throttle-on-deactivate */
2316 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
2317 qcfs_rq
->h_nr_running
-= task_delta
;
2319 if (qcfs_rq
->load
.weight
)
2324 rq
->nr_running
-= task_delta
;
2326 cfs_rq
->throttled
= 1;
2327 cfs_rq
->throttled_clock
= rq_clock(rq
);
2328 raw_spin_lock(&cfs_b
->lock
);
2329 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
2330 raw_spin_unlock(&cfs_b
->lock
);
2333 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
2335 struct rq
*rq
= rq_of(cfs_rq
);
2336 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2337 struct sched_entity
*se
;
2341 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
2343 cfs_rq
->throttled
= 0;
2345 update_rq_clock(rq
);
2347 raw_spin_lock(&cfs_b
->lock
);
2348 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
2349 list_del_rcu(&cfs_rq
->throttled_list
);
2350 raw_spin_unlock(&cfs_b
->lock
);
2352 /* update hierarchical throttle state */
2353 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
2355 if (!cfs_rq
->load
.weight
)
2358 task_delta
= cfs_rq
->h_nr_running
;
2359 for_each_sched_entity(se
) {
2363 cfs_rq
= cfs_rq_of(se
);
2365 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
2366 cfs_rq
->h_nr_running
+= task_delta
;
2368 if (cfs_rq_throttled(cfs_rq
))
2373 rq
->nr_running
+= task_delta
;
2375 /* determine whether we need to wake up potentially idle cpu */
2376 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
2377 resched_task(rq
->curr
);
2380 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
2381 u64 remaining
, u64 expires
)
2383 struct cfs_rq
*cfs_rq
;
2384 u64 runtime
= remaining
;
2387 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
2389 struct rq
*rq
= rq_of(cfs_rq
);
2391 raw_spin_lock(&rq
->lock
);
2392 if (!cfs_rq_throttled(cfs_rq
))
2395 runtime
= -cfs_rq
->runtime_remaining
+ 1;
2396 if (runtime
> remaining
)
2397 runtime
= remaining
;
2398 remaining
-= runtime
;
2400 cfs_rq
->runtime_remaining
+= runtime
;
2401 cfs_rq
->runtime_expires
= expires
;
2403 /* we check whether we're throttled above */
2404 if (cfs_rq
->runtime_remaining
> 0)
2405 unthrottle_cfs_rq(cfs_rq
);
2408 raw_spin_unlock(&rq
->lock
);
2419 * Responsible for refilling a task_group's bandwidth and unthrottling its
2420 * cfs_rqs as appropriate. If there has been no activity within the last
2421 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2422 * used to track this state.
2424 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
2426 u64 runtime
, runtime_expires
;
2427 int idle
= 1, throttled
;
2429 raw_spin_lock(&cfs_b
->lock
);
2430 /* no need to continue the timer with no bandwidth constraint */
2431 if (cfs_b
->quota
== RUNTIME_INF
)
2434 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
2435 /* idle depends on !throttled (for the case of a large deficit) */
2436 idle
= cfs_b
->idle
&& !throttled
;
2437 cfs_b
->nr_periods
+= overrun
;
2439 /* if we're going inactive then everything else can be deferred */
2443 __refill_cfs_bandwidth_runtime(cfs_b
);
2446 /* mark as potentially idle for the upcoming period */
2451 /* account preceding periods in which throttling occurred */
2452 cfs_b
->nr_throttled
+= overrun
;
2455 * There are throttled entities so we must first use the new bandwidth
2456 * to unthrottle them before making it generally available. This
2457 * ensures that all existing debts will be paid before a new cfs_rq is
2460 runtime
= cfs_b
->runtime
;
2461 runtime_expires
= cfs_b
->runtime_expires
;
2465 * This check is repeated as we are holding onto the new bandwidth
2466 * while we unthrottle. This can potentially race with an unthrottled
2467 * group trying to acquire new bandwidth from the global pool.
2469 while (throttled
&& runtime
> 0) {
2470 raw_spin_unlock(&cfs_b
->lock
);
2471 /* we can't nest cfs_b->lock while distributing bandwidth */
2472 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
2474 raw_spin_lock(&cfs_b
->lock
);
2476 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
2479 /* return (any) remaining runtime */
2480 cfs_b
->runtime
= runtime
;
2482 * While we are ensured activity in the period following an
2483 * unthrottle, this also covers the case in which the new bandwidth is
2484 * insufficient to cover the existing bandwidth deficit. (Forcing the
2485 * timer to remain active while there are any throttled entities.)
2490 cfs_b
->timer_active
= 0;
2491 raw_spin_unlock(&cfs_b
->lock
);
2496 /* a cfs_rq won't donate quota below this amount */
2497 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
2498 /* minimum remaining period time to redistribute slack quota */
2499 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
2500 /* how long we wait to gather additional slack before distributing */
2501 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
2503 /* are we near the end of the current quota period? */
2504 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
2506 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
2509 /* if the call-back is running a quota refresh is already occurring */
2510 if (hrtimer_callback_running(refresh_timer
))
2513 /* is a quota refresh about to occur? */
2514 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
2515 if (remaining
< min_expire
)
2521 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
2523 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
2525 /* if there's a quota refresh soon don't bother with slack */
2526 if (runtime_refresh_within(cfs_b
, min_left
))
2529 start_bandwidth_timer(&cfs_b
->slack_timer
,
2530 ns_to_ktime(cfs_bandwidth_slack_period
));
2533 /* we know any runtime found here is valid as update_curr() precedes return */
2534 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2536 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2537 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
2539 if (slack_runtime
<= 0)
2542 raw_spin_lock(&cfs_b
->lock
);
2543 if (cfs_b
->quota
!= RUNTIME_INF
&&
2544 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
2545 cfs_b
->runtime
+= slack_runtime
;
2547 /* we are under rq->lock, defer unthrottling using a timer */
2548 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
2549 !list_empty(&cfs_b
->throttled_cfs_rq
))
2550 start_cfs_slack_bandwidth(cfs_b
);
2552 raw_spin_unlock(&cfs_b
->lock
);
2554 /* even if it's not valid for return we don't want to try again */
2555 cfs_rq
->runtime_remaining
-= slack_runtime
;
2558 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2560 if (!cfs_bandwidth_used())
2563 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
2566 __return_cfs_rq_runtime(cfs_rq
);
2570 * This is done with a timer (instead of inline with bandwidth return) since
2571 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2573 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
2575 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
2578 /* confirm we're still not at a refresh boundary */
2579 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
))
2582 raw_spin_lock(&cfs_b
->lock
);
2583 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
) {
2584 runtime
= cfs_b
->runtime
;
2587 expires
= cfs_b
->runtime_expires
;
2588 raw_spin_unlock(&cfs_b
->lock
);
2593 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
2595 raw_spin_lock(&cfs_b
->lock
);
2596 if (expires
== cfs_b
->runtime_expires
)
2597 cfs_b
->runtime
= runtime
;
2598 raw_spin_unlock(&cfs_b
->lock
);
2602 * When a group wakes up we want to make sure that its quota is not already
2603 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2604 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2606 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
2608 if (!cfs_bandwidth_used())
2611 /* an active group must be handled by the update_curr()->put() path */
2612 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
2615 /* ensure the group is not already throttled */
2616 if (cfs_rq_throttled(cfs_rq
))
2619 /* update runtime allocation */
2620 account_cfs_rq_runtime(cfs_rq
, 0);
2621 if (cfs_rq
->runtime_remaining
<= 0)
2622 throttle_cfs_rq(cfs_rq
);
2625 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2626 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2628 if (!cfs_bandwidth_used())
2631 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
2635 * it's possible for a throttled entity to be forced into a running
2636 * state (e.g. set_curr_task), in this case we're finished.
2638 if (cfs_rq_throttled(cfs_rq
))
2641 throttle_cfs_rq(cfs_rq
);
2644 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
2646 struct cfs_bandwidth
*cfs_b
=
2647 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
2648 do_sched_cfs_slack_timer(cfs_b
);
2650 return HRTIMER_NORESTART
;
2653 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
2655 struct cfs_bandwidth
*cfs_b
=
2656 container_of(timer
, struct cfs_bandwidth
, period_timer
);
2662 now
= hrtimer_cb_get_time(timer
);
2663 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
2668 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
2671 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
2674 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2676 raw_spin_lock_init(&cfs_b
->lock
);
2678 cfs_b
->quota
= RUNTIME_INF
;
2679 cfs_b
->period
= ns_to_ktime(default_cfs_period());
2681 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
2682 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2683 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
2684 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2685 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
2688 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2690 cfs_rq
->runtime_enabled
= 0;
2691 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
2694 /* requires cfs_b->lock, may release to reprogram timer */
2695 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2698 * The timer may be active because we're trying to set a new bandwidth
2699 * period or because we're racing with the tear-down path
2700 * (timer_active==0 becomes visible before the hrtimer call-back
2701 * terminates). In either case we ensure that it's re-programmed
2703 while (unlikely(hrtimer_active(&cfs_b
->period_timer
))) {
2704 raw_spin_unlock(&cfs_b
->lock
);
2705 /* ensure cfs_b->lock is available while we wait */
2706 hrtimer_cancel(&cfs_b
->period_timer
);
2708 raw_spin_lock(&cfs_b
->lock
);
2709 /* if someone else restarted the timer then we're done */
2710 if (cfs_b
->timer_active
)
2714 cfs_b
->timer_active
= 1;
2715 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
2718 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2720 hrtimer_cancel(&cfs_b
->period_timer
);
2721 hrtimer_cancel(&cfs_b
->slack_timer
);
2724 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
2726 struct cfs_rq
*cfs_rq
;
2728 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
2729 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2731 if (!cfs_rq
->runtime_enabled
)
2735 * clock_task is not advancing so we just need to make sure
2736 * there's some valid quota amount
2738 cfs_rq
->runtime_remaining
= cfs_b
->quota
;
2739 if (cfs_rq_throttled(cfs_rq
))
2740 unthrottle_cfs_rq(cfs_rq
);
2744 #else /* CONFIG_CFS_BANDWIDTH */
2745 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
2747 return rq_clock_task(rq_of(cfs_rq
));
2750 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
,
2751 unsigned long delta_exec
) {}
2752 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2753 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
2754 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2756 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
2761 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
2766 static inline int throttled_lb_pair(struct task_group
*tg
,
2767 int src_cpu
, int dest_cpu
)
2772 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
2774 #ifdef CONFIG_FAIR_GROUP_SCHED
2775 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2778 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
2782 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
2783 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
2785 #endif /* CONFIG_CFS_BANDWIDTH */
2787 /**************************************************
2788 * CFS operations on tasks:
2791 #ifdef CONFIG_SCHED_HRTICK
2792 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
2794 struct sched_entity
*se
= &p
->se
;
2795 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2797 WARN_ON(task_rq(p
) != rq
);
2799 if (cfs_rq
->nr_running
> 1) {
2800 u64 slice
= sched_slice(cfs_rq
, se
);
2801 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
2802 s64 delta
= slice
- ran
;
2811 * Don't schedule slices shorter than 10000ns, that just
2812 * doesn't make sense. Rely on vruntime for fairness.
2815 delta
= max_t(s64
, 10000LL, delta
);
2817 hrtick_start(rq
, delta
);
2822 * called from enqueue/dequeue and updates the hrtick when the
2823 * current task is from our class and nr_running is low enough
2826 static void hrtick_update(struct rq
*rq
)
2828 struct task_struct
*curr
= rq
->curr
;
2830 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
2833 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
2834 hrtick_start_fair(rq
, curr
);
2836 #else /* !CONFIG_SCHED_HRTICK */
2838 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
2842 static inline void hrtick_update(struct rq
*rq
)
2848 * The enqueue_task method is called before nr_running is
2849 * increased. Here we update the fair scheduling stats and
2850 * then put the task into the rbtree:
2853 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
2855 struct cfs_rq
*cfs_rq
;
2856 struct sched_entity
*se
= &p
->se
;
2858 for_each_sched_entity(se
) {
2861 cfs_rq
= cfs_rq_of(se
);
2862 enqueue_entity(cfs_rq
, se
, flags
);
2865 * end evaluation on encountering a throttled cfs_rq
2867 * note: in the case of encountering a throttled cfs_rq we will
2868 * post the final h_nr_running increment below.
2870 if (cfs_rq_throttled(cfs_rq
))
2872 cfs_rq
->h_nr_running
++;
2874 flags
= ENQUEUE_WAKEUP
;
2877 for_each_sched_entity(se
) {
2878 cfs_rq
= cfs_rq_of(se
);
2879 cfs_rq
->h_nr_running
++;
2881 if (cfs_rq_throttled(cfs_rq
))
2884 update_cfs_shares(cfs_rq
);
2885 update_entity_load_avg(se
, 1);
2889 update_rq_runnable_avg(rq
, rq
->nr_running
);
2895 static void set_next_buddy(struct sched_entity
*se
);
2898 * The dequeue_task method is called before nr_running is
2899 * decreased. We remove the task from the rbtree and
2900 * update the fair scheduling stats:
2902 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
2904 struct cfs_rq
*cfs_rq
;
2905 struct sched_entity
*se
= &p
->se
;
2906 int task_sleep
= flags
& DEQUEUE_SLEEP
;
2908 for_each_sched_entity(se
) {
2909 cfs_rq
= cfs_rq_of(se
);
2910 dequeue_entity(cfs_rq
, se
, flags
);
2913 * end evaluation on encountering a throttled cfs_rq
2915 * note: in the case of encountering a throttled cfs_rq we will
2916 * post the final h_nr_running decrement below.
2918 if (cfs_rq_throttled(cfs_rq
))
2920 cfs_rq
->h_nr_running
--;
2922 /* Don't dequeue parent if it has other entities besides us */
2923 if (cfs_rq
->load
.weight
) {
2925 * Bias pick_next to pick a task from this cfs_rq, as
2926 * p is sleeping when it is within its sched_slice.
2928 if (task_sleep
&& parent_entity(se
))
2929 set_next_buddy(parent_entity(se
));
2931 /* avoid re-evaluating load for this entity */
2932 se
= parent_entity(se
);
2935 flags
|= DEQUEUE_SLEEP
;
2938 for_each_sched_entity(se
) {
2939 cfs_rq
= cfs_rq_of(se
);
2940 cfs_rq
->h_nr_running
--;
2942 if (cfs_rq_throttled(cfs_rq
))
2945 update_cfs_shares(cfs_rq
);
2946 update_entity_load_avg(se
, 1);
2951 update_rq_runnable_avg(rq
, 1);
2957 /* Used instead of source_load when we know the type == 0 */
2958 static unsigned long weighted_cpuload(const int cpu
)
2960 return cpu_rq(cpu
)->load
.weight
;
2964 * Return a low guess at the load of a migration-source cpu weighted
2965 * according to the scheduling class and "nice" value.
2967 * We want to under-estimate the load of migration sources, to
2968 * balance conservatively.
2970 static unsigned long source_load(int cpu
, int type
)
2972 struct rq
*rq
= cpu_rq(cpu
);
2973 unsigned long total
= weighted_cpuload(cpu
);
2975 if (type
== 0 || !sched_feat(LB_BIAS
))
2978 return min(rq
->cpu_load
[type
-1], total
);
2982 * Return a high guess at the load of a migration-target cpu weighted
2983 * according to the scheduling class and "nice" value.
2985 static unsigned long target_load(int cpu
, int type
)
2987 struct rq
*rq
= cpu_rq(cpu
);
2988 unsigned long total
= weighted_cpuload(cpu
);
2990 if (type
== 0 || !sched_feat(LB_BIAS
))
2993 return max(rq
->cpu_load
[type
-1], total
);
2996 static unsigned long power_of(int cpu
)
2998 return cpu_rq(cpu
)->cpu_power
;
3001 static unsigned long cpu_avg_load_per_task(int cpu
)
3003 struct rq
*rq
= cpu_rq(cpu
);
3004 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
3007 return rq
->load
.weight
/ nr_running
;
3013 static void task_waking_fair(struct task_struct
*p
)
3015 struct sched_entity
*se
= &p
->se
;
3016 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3019 #ifndef CONFIG_64BIT
3020 u64 min_vruntime_copy
;
3023 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
3025 min_vruntime
= cfs_rq
->min_vruntime
;
3026 } while (min_vruntime
!= min_vruntime_copy
);
3028 min_vruntime
= cfs_rq
->min_vruntime
;
3031 se
->vruntime
-= min_vruntime
;
3034 #ifdef CONFIG_FAIR_GROUP_SCHED
3036 * effective_load() calculates the load change as seen from the root_task_group
3038 * Adding load to a group doesn't make a group heavier, but can cause movement
3039 * of group shares between cpus. Assuming the shares were perfectly aligned one
3040 * can calculate the shift in shares.
3042 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3043 * on this @cpu and results in a total addition (subtraction) of @wg to the
3044 * total group weight.
3046 * Given a runqueue weight distribution (rw_i) we can compute a shares
3047 * distribution (s_i) using:
3049 * s_i = rw_i / \Sum rw_j (1)
3051 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3052 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3053 * shares distribution (s_i):
3055 * rw_i = { 2, 4, 1, 0 }
3056 * s_i = { 2/7, 4/7, 1/7, 0 }
3058 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3059 * task used to run on and the CPU the waker is running on), we need to
3060 * compute the effect of waking a task on either CPU and, in case of a sync
3061 * wakeup, compute the effect of the current task going to sleep.
3063 * So for a change of @wl to the local @cpu with an overall group weight change
3064 * of @wl we can compute the new shares distribution (s'_i) using:
3066 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3068 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3069 * differences in waking a task to CPU 0. The additional task changes the
3070 * weight and shares distributions like:
3072 * rw'_i = { 3, 4, 1, 0 }
3073 * s'_i = { 3/8, 4/8, 1/8, 0 }
3075 * We can then compute the difference in effective weight by using:
3077 * dw_i = S * (s'_i - s_i) (3)
3079 * Where 'S' is the group weight as seen by its parent.
3081 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3082 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3083 * 4/7) times the weight of the group.
3085 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
3087 struct sched_entity
*se
= tg
->se
[cpu
];
3089 if (!tg
->parent
) /* the trivial, non-cgroup case */
3092 for_each_sched_entity(se
) {
3098 * W = @wg + \Sum rw_j
3100 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
3105 w
= se
->my_q
->load
.weight
+ wl
;
3108 * wl = S * s'_i; see (2)
3111 wl
= (w
* tg
->shares
) / W
;
3116 * Per the above, wl is the new se->load.weight value; since
3117 * those are clipped to [MIN_SHARES, ...) do so now. See
3118 * calc_cfs_shares().
3120 if (wl
< MIN_SHARES
)
3124 * wl = dw_i = S * (s'_i - s_i); see (3)
3126 wl
-= se
->load
.weight
;
3129 * Recursively apply this logic to all parent groups to compute
3130 * the final effective load change on the root group. Since
3131 * only the @tg group gets extra weight, all parent groups can
3132 * only redistribute existing shares. @wl is the shift in shares
3133 * resulting from this level per the above.
3142 static inline unsigned long effective_load(struct task_group
*tg
, int cpu
,
3143 unsigned long wl
, unsigned long wg
)
3150 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
3152 s64 this_load
, load
;
3153 int idx
, this_cpu
, prev_cpu
;
3154 unsigned long tl_per_task
;
3155 struct task_group
*tg
;
3156 unsigned long weight
;
3160 this_cpu
= smp_processor_id();
3161 prev_cpu
= task_cpu(p
);
3162 load
= source_load(prev_cpu
, idx
);
3163 this_load
= target_load(this_cpu
, idx
);
3166 * If sync wakeup then subtract the (maximum possible)
3167 * effect of the currently running task from the load
3168 * of the current CPU:
3171 tg
= task_group(current
);
3172 weight
= current
->se
.load
.weight
;
3174 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
3175 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
3179 weight
= p
->se
.load
.weight
;
3182 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3183 * due to the sync cause above having dropped this_load to 0, we'll
3184 * always have an imbalance, but there's really nothing you can do
3185 * about that, so that's good too.
3187 * Otherwise check if either cpus are near enough in load to allow this
3188 * task to be woken on this_cpu.
3190 if (this_load
> 0) {
3191 s64 this_eff_load
, prev_eff_load
;
3193 this_eff_load
= 100;
3194 this_eff_load
*= power_of(prev_cpu
);
3195 this_eff_load
*= this_load
+
3196 effective_load(tg
, this_cpu
, weight
, weight
);
3198 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
3199 prev_eff_load
*= power_of(this_cpu
);
3200 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
3202 balanced
= this_eff_load
<= prev_eff_load
;
3207 * If the currently running task will sleep within
3208 * a reasonable amount of time then attract this newly
3211 if (sync
&& balanced
)
3214 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
3215 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
3218 (this_load
<= load
&&
3219 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
3221 * This domain has SD_WAKE_AFFINE and
3222 * p is cache cold in this domain, and
3223 * there is no bad imbalance.
3225 schedstat_inc(sd
, ttwu_move_affine
);
3226 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
3234 * find_idlest_group finds and returns the least busy CPU group within the
3237 static struct sched_group
*
3238 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
3239 int this_cpu
, int load_idx
)
3241 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
3242 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
3243 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
3246 unsigned long load
, avg_load
;
3250 /* Skip over this group if it has no CPUs allowed */
3251 if (!cpumask_intersects(sched_group_cpus(group
),
3252 tsk_cpus_allowed(p
)))
3255 local_group
= cpumask_test_cpu(this_cpu
,
3256 sched_group_cpus(group
));
3258 /* Tally up the load of all CPUs in the group */
3261 for_each_cpu(i
, sched_group_cpus(group
)) {
3262 /* Bias balancing toward cpus of our domain */
3264 load
= source_load(i
, load_idx
);
3266 load
= target_load(i
, load_idx
);
3271 /* Adjust by relative CPU power of the group */
3272 avg_load
= (avg_load
* SCHED_POWER_SCALE
) / group
->sgp
->power
;
3275 this_load
= avg_load
;
3276 } else if (avg_load
< min_load
) {
3277 min_load
= avg_load
;
3280 } while (group
= group
->next
, group
!= sd
->groups
);
3282 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
3288 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3291 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
3293 unsigned long load
, min_load
= ULONG_MAX
;
3297 /* Traverse only the allowed CPUs */
3298 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
3299 load
= weighted_cpuload(i
);
3301 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
3311 * Try and locate an idle CPU in the sched_domain.
3313 static int select_idle_sibling(struct task_struct
*p
, int target
)
3315 struct sched_domain
*sd
;
3316 struct sched_group
*sg
;
3317 int i
= task_cpu(p
);
3319 if (idle_cpu(target
))
3323 * If the prevous cpu is cache affine and idle, don't be stupid.
3325 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
3329 * Otherwise, iterate the domains and find an elegible idle cpu.
3331 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
3332 for_each_lower_domain(sd
) {
3335 if (!cpumask_intersects(sched_group_cpus(sg
),
3336 tsk_cpus_allowed(p
)))
3339 for_each_cpu(i
, sched_group_cpus(sg
)) {
3340 if (i
== target
|| !idle_cpu(i
))
3344 target
= cpumask_first_and(sched_group_cpus(sg
),
3345 tsk_cpus_allowed(p
));
3349 } while (sg
!= sd
->groups
);
3356 * sched_balance_self: balance the current task (running on cpu) in domains
3357 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3360 * Balance, ie. select the least loaded group.
3362 * Returns the target CPU number, or the same CPU if no balancing is needed.
3364 * preempt must be disabled.
3367 select_task_rq_fair(struct task_struct
*p
, int sd_flag
, int wake_flags
)
3369 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
3370 int cpu
= smp_processor_id();
3371 int prev_cpu
= task_cpu(p
);
3373 int want_affine
= 0;
3374 int sync
= wake_flags
& WF_SYNC
;
3376 if (p
->nr_cpus_allowed
== 1)
3379 if (sd_flag
& SD_BALANCE_WAKE
) {
3380 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
3386 for_each_domain(cpu
, tmp
) {
3387 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
3391 * If both cpu and prev_cpu are part of this domain,
3392 * cpu is a valid SD_WAKE_AFFINE target.
3394 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
3395 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
3400 if (tmp
->flags
& sd_flag
)
3405 if (cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
3408 new_cpu
= select_idle_sibling(p
, prev_cpu
);
3413 int load_idx
= sd
->forkexec_idx
;
3414 struct sched_group
*group
;
3417 if (!(sd
->flags
& sd_flag
)) {
3422 if (sd_flag
& SD_BALANCE_WAKE
)
3423 load_idx
= sd
->wake_idx
;
3425 group
= find_idlest_group(sd
, p
, cpu
, load_idx
);
3431 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
3432 if (new_cpu
== -1 || new_cpu
== cpu
) {
3433 /* Now try balancing at a lower domain level of cpu */
3438 /* Now try balancing at a lower domain level of new_cpu */
3440 weight
= sd
->span_weight
;
3442 for_each_domain(cpu
, tmp
) {
3443 if (weight
<= tmp
->span_weight
)
3445 if (tmp
->flags
& sd_flag
)
3448 /* while loop will break here if sd == NULL */
3457 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3458 * cfs_rq_of(p) references at time of call are still valid and identify the
3459 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3460 * other assumptions, including the state of rq->lock, should be made.
3463 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
3465 struct sched_entity
*se
= &p
->se
;
3466 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3469 * Load tracking: accumulate removed load so that it can be processed
3470 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3471 * to blocked load iff they have a positive decay-count. It can never
3472 * be negative here since on-rq tasks have decay-count == 0.
3474 if (se
->avg
.decay_count
) {
3475 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
3476 atomic64_add(se
->avg
.load_avg_contrib
, &cfs_rq
->removed_load
);
3479 #endif /* CONFIG_SMP */
3481 static unsigned long
3482 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
3484 unsigned long gran
= sysctl_sched_wakeup_granularity
;
3487 * Since its curr running now, convert the gran from real-time
3488 * to virtual-time in his units.
3490 * By using 'se' instead of 'curr' we penalize light tasks, so
3491 * they get preempted easier. That is, if 'se' < 'curr' then
3492 * the resulting gran will be larger, therefore penalizing the
3493 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3494 * be smaller, again penalizing the lighter task.
3496 * This is especially important for buddies when the leftmost
3497 * task is higher priority than the buddy.
3499 return calc_delta_fair(gran
, se
);
3503 * Should 'se' preempt 'curr'.
3517 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
3519 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
3524 gran
= wakeup_gran(curr
, se
);
3531 static void set_last_buddy(struct sched_entity
*se
)
3533 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
3536 for_each_sched_entity(se
)
3537 cfs_rq_of(se
)->last
= se
;
3540 static void set_next_buddy(struct sched_entity
*se
)
3542 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
3545 for_each_sched_entity(se
)
3546 cfs_rq_of(se
)->next
= se
;
3549 static void set_skip_buddy(struct sched_entity
*se
)
3551 for_each_sched_entity(se
)
3552 cfs_rq_of(se
)->skip
= se
;
3556 * Preempt the current task with a newly woken task if needed:
3558 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
3560 struct task_struct
*curr
= rq
->curr
;
3561 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
3562 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
3563 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
3564 int next_buddy_marked
= 0;
3566 if (unlikely(se
== pse
))
3570 * This is possible from callers such as move_task(), in which we
3571 * unconditionally check_prempt_curr() after an enqueue (which may have
3572 * lead to a throttle). This both saves work and prevents false
3573 * next-buddy nomination below.
3575 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
3578 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
3579 set_next_buddy(pse
);
3580 next_buddy_marked
= 1;
3584 * We can come here with TIF_NEED_RESCHED already set from new task
3587 * Note: this also catches the edge-case of curr being in a throttled
3588 * group (e.g. via set_curr_task), since update_curr() (in the
3589 * enqueue of curr) will have resulted in resched being set. This
3590 * prevents us from potentially nominating it as a false LAST_BUDDY
3593 if (test_tsk_need_resched(curr
))
3596 /* Idle tasks are by definition preempted by non-idle tasks. */
3597 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
3598 likely(p
->policy
!= SCHED_IDLE
))
3602 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3603 * is driven by the tick):
3605 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
3608 find_matching_se(&se
, &pse
);
3609 update_curr(cfs_rq_of(se
));
3611 if (wakeup_preempt_entity(se
, pse
) == 1) {
3613 * Bias pick_next to pick the sched entity that is
3614 * triggering this preemption.
3616 if (!next_buddy_marked
)
3617 set_next_buddy(pse
);
3626 * Only set the backward buddy when the current task is still
3627 * on the rq. This can happen when a wakeup gets interleaved
3628 * with schedule on the ->pre_schedule() or idle_balance()
3629 * point, either of which can * drop the rq lock.
3631 * Also, during early boot the idle thread is in the fair class,
3632 * for obvious reasons its a bad idea to schedule back to it.
3634 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
3637 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
3641 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
3643 struct task_struct
*p
;
3644 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
3645 struct sched_entity
*se
;
3647 if (!cfs_rq
->nr_running
)
3651 se
= pick_next_entity(cfs_rq
);
3652 set_next_entity(cfs_rq
, se
);
3653 cfs_rq
= group_cfs_rq(se
);
3657 if (hrtick_enabled(rq
))
3658 hrtick_start_fair(rq
, p
);
3664 * Account for a descheduled task:
3666 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
3668 struct sched_entity
*se
= &prev
->se
;
3669 struct cfs_rq
*cfs_rq
;
3671 for_each_sched_entity(se
) {
3672 cfs_rq
= cfs_rq_of(se
);
3673 put_prev_entity(cfs_rq
, se
);
3678 * sched_yield() is very simple
3680 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3682 static void yield_task_fair(struct rq
*rq
)
3684 struct task_struct
*curr
= rq
->curr
;
3685 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
3686 struct sched_entity
*se
= &curr
->se
;
3689 * Are we the only task in the tree?
3691 if (unlikely(rq
->nr_running
== 1))
3694 clear_buddies(cfs_rq
, se
);
3696 if (curr
->policy
!= SCHED_BATCH
) {
3697 update_rq_clock(rq
);
3699 * Update run-time statistics of the 'current'.
3701 update_curr(cfs_rq
);
3703 * Tell update_rq_clock() that we've just updated,
3704 * so we don't do microscopic update in schedule()
3705 * and double the fastpath cost.
3707 rq
->skip_clock_update
= 1;
3713 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
3715 struct sched_entity
*se
= &p
->se
;
3717 /* throttled hierarchies are not runnable */
3718 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
3721 /* Tell the scheduler that we'd really like pse to run next. */
3724 yield_task_fair(rq
);
3730 /**************************************************
3731 * Fair scheduling class load-balancing methods.
3735 * The purpose of load-balancing is to achieve the same basic fairness the
3736 * per-cpu scheduler provides, namely provide a proportional amount of compute
3737 * time to each task. This is expressed in the following equation:
3739 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3741 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3742 * W_i,0 is defined as:
3744 * W_i,0 = \Sum_j w_i,j (2)
3746 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3747 * is derived from the nice value as per prio_to_weight[].
3749 * The weight average is an exponential decay average of the instantaneous
3752 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3754 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3755 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3756 * can also include other factors [XXX].
3758 * To achieve this balance we define a measure of imbalance which follows
3759 * directly from (1):
3761 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3763 * We them move tasks around to minimize the imbalance. In the continuous
3764 * function space it is obvious this converges, in the discrete case we get
3765 * a few fun cases generally called infeasible weight scenarios.
3768 * - infeasible weights;
3769 * - local vs global optima in the discrete case. ]
3774 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3775 * for all i,j solution, we create a tree of cpus that follows the hardware
3776 * topology where each level pairs two lower groups (or better). This results
3777 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3778 * tree to only the first of the previous level and we decrease the frequency
3779 * of load-balance at each level inv. proportional to the number of cpus in
3785 * \Sum { --- * --- * 2^i } = O(n) (5)
3787 * `- size of each group
3788 * | | `- number of cpus doing load-balance
3790 * `- sum over all levels
3792 * Coupled with a limit on how many tasks we can migrate every balance pass,
3793 * this makes (5) the runtime complexity of the balancer.
3795 * An important property here is that each CPU is still (indirectly) connected
3796 * to every other cpu in at most O(log n) steps:
3798 * The adjacency matrix of the resulting graph is given by:
3801 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3804 * And you'll find that:
3806 * A^(log_2 n)_i,j != 0 for all i,j (7)
3808 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3809 * The task movement gives a factor of O(m), giving a convergence complexity
3812 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3817 * In order to avoid CPUs going idle while there's still work to do, new idle
3818 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3819 * tree itself instead of relying on other CPUs to bring it work.
3821 * This adds some complexity to both (5) and (8) but it reduces the total idle
3829 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3832 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3837 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3839 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3841 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3844 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3845 * rewrite all of this once again.]
3848 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
3850 #define LBF_ALL_PINNED 0x01
3851 #define LBF_NEED_BREAK 0x02
3852 #define LBF_SOME_PINNED 0x04
3855 struct sched_domain
*sd
;
3863 struct cpumask
*dst_grpmask
;
3865 enum cpu_idle_type idle
;
3867 /* The set of CPUs under consideration for load-balancing */
3868 struct cpumask
*cpus
;
3873 unsigned int loop_break
;
3874 unsigned int loop_max
;
3878 * move_task - move a task from one runqueue to another runqueue.
3879 * Both runqueues must be locked.
3881 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
3883 deactivate_task(env
->src_rq
, p
, 0);
3884 set_task_cpu(p
, env
->dst_cpu
);
3885 activate_task(env
->dst_rq
, p
, 0);
3886 check_preempt_curr(env
->dst_rq
, p
, 0);
3890 * Is this task likely cache-hot:
3893 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
3897 if (p
->sched_class
!= &fair_sched_class
)
3900 if (unlikely(p
->policy
== SCHED_IDLE
))
3904 * Buddy candidates are cache hot:
3906 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
3907 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
3908 &p
->se
== cfs_rq_of(&p
->se
)->last
))
3911 if (sysctl_sched_migration_cost
== -1)
3913 if (sysctl_sched_migration_cost
== 0)
3916 delta
= now
- p
->se
.exec_start
;
3918 return delta
< (s64
)sysctl_sched_migration_cost
;
3922 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3925 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
3927 int tsk_cache_hot
= 0;
3929 * We do not migrate tasks that are:
3930 * 1) throttled_lb_pair, or
3931 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3932 * 3) running (obviously), or
3933 * 4) are cache-hot on their current CPU.
3935 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
3938 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
3941 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
3944 * Remember if this task can be migrated to any other cpu in
3945 * our sched_group. We may want to revisit it if we couldn't
3946 * meet load balance goals by pulling other tasks on src_cpu.
3948 * Also avoid computing new_dst_cpu if we have already computed
3949 * one in current iteration.
3951 if (!env
->dst_grpmask
|| (env
->flags
& LBF_SOME_PINNED
))
3954 /* Prevent to re-select dst_cpu via env's cpus */
3955 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
3956 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
3957 env
->flags
|= LBF_SOME_PINNED
;
3958 env
->new_dst_cpu
= cpu
;
3966 /* Record that we found atleast one task that could run on dst_cpu */
3967 env
->flags
&= ~LBF_ALL_PINNED
;
3969 if (task_running(env
->src_rq
, p
)) {
3970 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
3975 * Aggressive migration if:
3976 * 1) task is cache cold, or
3977 * 2) too many balance attempts have failed.
3980 tsk_cache_hot
= task_hot(p
, rq_clock_task(env
->src_rq
), env
->sd
);
3981 if (!tsk_cache_hot
||
3982 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
3984 if (tsk_cache_hot
) {
3985 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
3986 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
3992 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
3997 * move_one_task tries to move exactly one task from busiest to this_rq, as
3998 * part of active balancing operations within "domain".
3999 * Returns 1 if successful and 0 otherwise.
4001 * Called with both runqueues locked.
4003 static int move_one_task(struct lb_env
*env
)
4005 struct task_struct
*p
, *n
;
4007 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
4008 if (!can_migrate_task(p
, env
))
4013 * Right now, this is only the second place move_task()
4014 * is called, so we can safely collect move_task()
4015 * stats here rather than inside move_task().
4017 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
4023 static unsigned long task_h_load(struct task_struct
*p
);
4025 static const unsigned int sched_nr_migrate_break
= 32;
4028 * move_tasks tries to move up to imbalance weighted load from busiest to
4029 * this_rq, as part of a balancing operation within domain "sd".
4030 * Returns 1 if successful and 0 otherwise.
4032 * Called with both runqueues locked.
4034 static int move_tasks(struct lb_env
*env
)
4036 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
4037 struct task_struct
*p
;
4041 if (env
->imbalance
<= 0)
4044 while (!list_empty(tasks
)) {
4045 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
4048 /* We've more or less seen every task there is, call it quits */
4049 if (env
->loop
> env
->loop_max
)
4052 /* take a breather every nr_migrate tasks */
4053 if (env
->loop
> env
->loop_break
) {
4054 env
->loop_break
+= sched_nr_migrate_break
;
4055 env
->flags
|= LBF_NEED_BREAK
;
4059 if (!can_migrate_task(p
, env
))
4062 load
= task_h_load(p
);
4064 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
4067 if ((load
/ 2) > env
->imbalance
)
4072 env
->imbalance
-= load
;
4074 #ifdef CONFIG_PREEMPT
4076 * NEWIDLE balancing is a source of latency, so preemptible
4077 * kernels will stop after the first task is pulled to minimize
4078 * the critical section.
4080 if (env
->idle
== CPU_NEWLY_IDLE
)
4085 * We only want to steal up to the prescribed amount of
4088 if (env
->imbalance
<= 0)
4093 list_move_tail(&p
->se
.group_node
, tasks
);
4097 * Right now, this is one of only two places move_task() is called,
4098 * so we can safely collect move_task() stats here rather than
4099 * inside move_task().
4101 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
4106 #ifdef CONFIG_FAIR_GROUP_SCHED
4108 * update tg->load_weight by folding this cpu's load_avg
4110 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
4112 struct sched_entity
*se
= tg
->se
[cpu
];
4113 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
4115 /* throttled entities do not contribute to load */
4116 if (throttled_hierarchy(cfs_rq
))
4119 update_cfs_rq_blocked_load(cfs_rq
, 1);
4122 update_entity_load_avg(se
, 1);
4124 * We pivot on our runnable average having decayed to zero for
4125 * list removal. This generally implies that all our children
4126 * have also been removed (modulo rounding error or bandwidth
4127 * control); however, such cases are rare and we can fix these
4130 * TODO: fix up out-of-order children on enqueue.
4132 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
4133 list_del_leaf_cfs_rq(cfs_rq
);
4135 struct rq
*rq
= rq_of(cfs_rq
);
4136 update_rq_runnable_avg(rq
, rq
->nr_running
);
4140 static void update_blocked_averages(int cpu
)
4142 struct rq
*rq
= cpu_rq(cpu
);
4143 struct cfs_rq
*cfs_rq
;
4144 unsigned long flags
;
4146 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4147 update_rq_clock(rq
);
4149 * Iterates the task_group tree in a bottom up fashion, see
4150 * list_add_leaf_cfs_rq() for details.
4152 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4154 * Note: We may want to consider periodically releasing
4155 * rq->lock about these updates so that creating many task
4156 * groups does not result in continually extending hold time.
4158 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
4161 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4165 * Compute the cpu's hierarchical load factor for each task group.
4166 * This needs to be done in a top-down fashion because the load of a child
4167 * group is a fraction of its parents load.
4169 static int tg_load_down(struct task_group
*tg
, void *data
)
4172 long cpu
= (long)data
;
4175 load
= cpu_rq(cpu
)->load
.weight
;
4177 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
4178 load
*= tg
->se
[cpu
]->load
.weight
;
4179 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
4182 tg
->cfs_rq
[cpu
]->h_load
= load
;
4187 static void update_h_load(long cpu
)
4189 struct rq
*rq
= cpu_rq(cpu
);
4190 unsigned long now
= jiffies
;
4192 if (rq
->h_load_throttle
== now
)
4195 rq
->h_load_throttle
= now
;
4198 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
4202 static unsigned long task_h_load(struct task_struct
*p
)
4204 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
4207 load
= p
->se
.load
.weight
;
4208 load
= div_u64(load
* cfs_rq
->h_load
, cfs_rq
->load
.weight
+ 1);
4213 static inline void update_blocked_averages(int cpu
)
4217 static inline void update_h_load(long cpu
)
4221 static unsigned long task_h_load(struct task_struct
*p
)
4223 return p
->se
.load
.weight
;
4227 /********** Helpers for find_busiest_group ************************/
4229 * sd_lb_stats - Structure to store the statistics of a sched_domain
4230 * during load balancing.
4232 struct sd_lb_stats
{
4233 struct sched_group
*busiest
; /* Busiest group in this sd */
4234 struct sched_group
*this; /* Local group in this sd */
4235 unsigned long total_load
; /* Total load of all groups in sd */
4236 unsigned long total_pwr
; /* Total power of all groups in sd */
4237 unsigned long avg_load
; /* Average load across all groups in sd */
4239 /** Statistics of this group */
4240 unsigned long this_load
;
4241 unsigned long this_load_per_task
;
4242 unsigned long this_nr_running
;
4243 unsigned long this_has_capacity
;
4244 unsigned int this_idle_cpus
;
4246 /* Statistics of the busiest group */
4247 unsigned int busiest_idle_cpus
;
4248 unsigned long max_load
;
4249 unsigned long busiest_load_per_task
;
4250 unsigned long busiest_nr_running
;
4251 unsigned long busiest_group_capacity
;
4252 unsigned long busiest_has_capacity
;
4253 unsigned int busiest_group_weight
;
4255 int group_imb
; /* Is there imbalance in this sd */
4259 * sg_lb_stats - stats of a sched_group required for load_balancing
4261 struct sg_lb_stats
{
4262 unsigned long avg_load
; /*Avg load across the CPUs of the group */
4263 unsigned long group_load
; /* Total load over the CPUs of the group */
4264 unsigned long sum_nr_running
; /* Nr tasks running in the group */
4265 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
4266 unsigned long group_capacity
;
4267 unsigned long idle_cpus
;
4268 unsigned long group_weight
;
4269 int group_imb
; /* Is there an imbalance in the group ? */
4270 int group_has_capacity
; /* Is there extra capacity in the group? */
4274 * get_sd_load_idx - Obtain the load index for a given sched domain.
4275 * @sd: The sched_domain whose load_idx is to be obtained.
4276 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4278 static inline int get_sd_load_idx(struct sched_domain
*sd
,
4279 enum cpu_idle_type idle
)
4285 load_idx
= sd
->busy_idx
;
4288 case CPU_NEWLY_IDLE
:
4289 load_idx
= sd
->newidle_idx
;
4292 load_idx
= sd
->idle_idx
;
4299 static unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
4301 return SCHED_POWER_SCALE
;
4304 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
4306 return default_scale_freq_power(sd
, cpu
);
4309 static unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
4311 unsigned long weight
= sd
->span_weight
;
4312 unsigned long smt_gain
= sd
->smt_gain
;
4319 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
4321 return default_scale_smt_power(sd
, cpu
);
4324 static unsigned long scale_rt_power(int cpu
)
4326 struct rq
*rq
= cpu_rq(cpu
);
4327 u64 total
, available
, age_stamp
, avg
;
4330 * Since we're reading these variables without serialization make sure
4331 * we read them once before doing sanity checks on them.
4333 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
4334 avg
= ACCESS_ONCE(rq
->rt_avg
);
4336 total
= sched_avg_period() + (rq_clock(rq
) - age_stamp
);
4338 if (unlikely(total
< avg
)) {
4339 /* Ensures that power won't end up being negative */
4342 available
= total
- avg
;
4345 if (unlikely((s64
)total
< SCHED_POWER_SCALE
))
4346 total
= SCHED_POWER_SCALE
;
4348 total
>>= SCHED_POWER_SHIFT
;
4350 return div_u64(available
, total
);
4353 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
4355 unsigned long weight
= sd
->span_weight
;
4356 unsigned long power
= SCHED_POWER_SCALE
;
4357 struct sched_group
*sdg
= sd
->groups
;
4359 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
4360 if (sched_feat(ARCH_POWER
))
4361 power
*= arch_scale_smt_power(sd
, cpu
);
4363 power
*= default_scale_smt_power(sd
, cpu
);
4365 power
>>= SCHED_POWER_SHIFT
;
4368 sdg
->sgp
->power_orig
= power
;
4370 if (sched_feat(ARCH_POWER
))
4371 power
*= arch_scale_freq_power(sd
, cpu
);
4373 power
*= default_scale_freq_power(sd
, cpu
);
4375 power
>>= SCHED_POWER_SHIFT
;
4377 power
*= scale_rt_power(cpu
);
4378 power
>>= SCHED_POWER_SHIFT
;
4383 cpu_rq(cpu
)->cpu_power
= power
;
4384 sdg
->sgp
->power
= power
;
4387 void update_group_power(struct sched_domain
*sd
, int cpu
)
4389 struct sched_domain
*child
= sd
->child
;
4390 struct sched_group
*group
, *sdg
= sd
->groups
;
4391 unsigned long power
;
4392 unsigned long interval
;
4394 interval
= msecs_to_jiffies(sd
->balance_interval
);
4395 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
4396 sdg
->sgp
->next_update
= jiffies
+ interval
;
4399 update_cpu_power(sd
, cpu
);
4405 if (child
->flags
& SD_OVERLAP
) {
4407 * SD_OVERLAP domains cannot assume that child groups
4408 * span the current group.
4411 for_each_cpu(cpu
, sched_group_cpus(sdg
))
4412 power
+= power_of(cpu
);
4415 * !SD_OVERLAP domains can assume that child groups
4416 * span the current group.
4419 group
= child
->groups
;
4421 power
+= group
->sgp
->power
;
4422 group
= group
->next
;
4423 } while (group
!= child
->groups
);
4426 sdg
->sgp
->power_orig
= sdg
->sgp
->power
= power
;
4430 * Try and fix up capacity for tiny siblings, this is needed when
4431 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4432 * which on its own isn't powerful enough.
4434 * See update_sd_pick_busiest() and check_asym_packing().
4437 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
4440 * Only siblings can have significantly less than SCHED_POWER_SCALE
4442 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
4446 * If ~90% of the cpu_power is still there, we're good.
4448 if (group
->sgp
->power
* 32 > group
->sgp
->power_orig
* 29)
4455 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4456 * @env: The load balancing environment.
4457 * @group: sched_group whose statistics are to be updated.
4458 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4459 * @local_group: Does group contain this_cpu.
4460 * @balance: Should we balance.
4461 * @sgs: variable to hold the statistics for this group.
4463 static inline void update_sg_lb_stats(struct lb_env
*env
,
4464 struct sched_group
*group
, int load_idx
,
4465 int local_group
, int *balance
, struct sg_lb_stats
*sgs
)
4467 unsigned long nr_running
, max_nr_running
, min_nr_running
;
4468 unsigned long load
, max_cpu_load
, min_cpu_load
;
4469 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
4470 unsigned long avg_load_per_task
= 0;
4474 balance_cpu
= group_balance_cpu(group
);
4476 /* Tally up the load of all CPUs in the group */
4478 min_cpu_load
= ~0UL;
4480 min_nr_running
= ~0UL;
4482 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
4483 struct rq
*rq
= cpu_rq(i
);
4485 nr_running
= rq
->nr_running
;
4487 /* Bias balancing toward cpus of our domain */
4489 if (idle_cpu(i
) && !first_idle_cpu
&&
4490 cpumask_test_cpu(i
, sched_group_mask(group
))) {
4495 load
= target_load(i
, load_idx
);
4497 load
= source_load(i
, load_idx
);
4498 if (load
> max_cpu_load
)
4499 max_cpu_load
= load
;
4500 if (min_cpu_load
> load
)
4501 min_cpu_load
= load
;
4503 if (nr_running
> max_nr_running
)
4504 max_nr_running
= nr_running
;
4505 if (min_nr_running
> nr_running
)
4506 min_nr_running
= nr_running
;
4509 sgs
->group_load
+= load
;
4510 sgs
->sum_nr_running
+= nr_running
;
4511 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
4517 * First idle cpu or the first cpu(busiest) in this sched group
4518 * is eligible for doing load balancing at this and above
4519 * domains. In the newly idle case, we will allow all the cpu's
4520 * to do the newly idle load balance.
4523 if (env
->idle
!= CPU_NEWLY_IDLE
) {
4524 if (balance_cpu
!= env
->dst_cpu
) {
4528 update_group_power(env
->sd
, env
->dst_cpu
);
4529 } else if (time_after_eq(jiffies
, group
->sgp
->next_update
))
4530 update_group_power(env
->sd
, env
->dst_cpu
);
4533 /* Adjust by relative CPU power of the group */
4534 sgs
->avg_load
= (sgs
->group_load
*SCHED_POWER_SCALE
) / group
->sgp
->power
;
4537 * Consider the group unbalanced when the imbalance is larger
4538 * than the average weight of a task.
4540 * APZ: with cgroup the avg task weight can vary wildly and
4541 * might not be a suitable number - should we keep a
4542 * normalized nr_running number somewhere that negates
4545 if (sgs
->sum_nr_running
)
4546 avg_load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
4548 if ((max_cpu_load
- min_cpu_load
) >= avg_load_per_task
&&
4549 (max_nr_running
- min_nr_running
) > 1)
4552 sgs
->group_capacity
= DIV_ROUND_CLOSEST(group
->sgp
->power
,
4554 if (!sgs
->group_capacity
)
4555 sgs
->group_capacity
= fix_small_capacity(env
->sd
, group
);
4556 sgs
->group_weight
= group
->group_weight
;
4558 if (sgs
->group_capacity
> sgs
->sum_nr_running
)
4559 sgs
->group_has_capacity
= 1;
4563 * update_sd_pick_busiest - return 1 on busiest group
4564 * @env: The load balancing environment.
4565 * @sds: sched_domain statistics
4566 * @sg: sched_group candidate to be checked for being the busiest
4567 * @sgs: sched_group statistics
4569 * Determine if @sg is a busier group than the previously selected
4572 static bool update_sd_pick_busiest(struct lb_env
*env
,
4573 struct sd_lb_stats
*sds
,
4574 struct sched_group
*sg
,
4575 struct sg_lb_stats
*sgs
)
4577 if (sgs
->avg_load
<= sds
->max_load
)
4580 if (sgs
->sum_nr_running
> sgs
->group_capacity
)
4587 * ASYM_PACKING needs to move all the work to the lowest
4588 * numbered CPUs in the group, therefore mark all groups
4589 * higher than ourself as busy.
4591 if ((env
->sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
4592 env
->dst_cpu
< group_first_cpu(sg
)) {
4596 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
4604 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4605 * @env: The load balancing environment.
4606 * @balance: Should we balance.
4607 * @sds: variable to hold the statistics for this sched_domain.
4609 static inline void update_sd_lb_stats(struct lb_env
*env
,
4610 int *balance
, struct sd_lb_stats
*sds
)
4612 struct sched_domain
*child
= env
->sd
->child
;
4613 struct sched_group
*sg
= env
->sd
->groups
;
4614 struct sg_lb_stats sgs
;
4615 int load_idx
, prefer_sibling
= 0;
4617 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
4620 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
4625 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
4626 memset(&sgs
, 0, sizeof(sgs
));
4627 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, balance
, &sgs
);
4629 if (local_group
&& !(*balance
))
4632 sds
->total_load
+= sgs
.group_load
;
4633 sds
->total_pwr
+= sg
->sgp
->power
;
4636 * In case the child domain prefers tasks go to siblings
4637 * first, lower the sg capacity to one so that we'll try
4638 * and move all the excess tasks away. We lower the capacity
4639 * of a group only if the local group has the capacity to fit
4640 * these excess tasks, i.e. nr_running < group_capacity. The
4641 * extra check prevents the case where you always pull from the
4642 * heaviest group when it is already under-utilized (possible
4643 * with a large weight task outweighs the tasks on the system).
4645 if (prefer_sibling
&& !local_group
&& sds
->this_has_capacity
)
4646 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
4649 sds
->this_load
= sgs
.avg_load
;
4651 sds
->this_nr_running
= sgs
.sum_nr_running
;
4652 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
4653 sds
->this_has_capacity
= sgs
.group_has_capacity
;
4654 sds
->this_idle_cpus
= sgs
.idle_cpus
;
4655 } else if (update_sd_pick_busiest(env
, sds
, sg
, &sgs
)) {
4656 sds
->max_load
= sgs
.avg_load
;
4658 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
4659 sds
->busiest_idle_cpus
= sgs
.idle_cpus
;
4660 sds
->busiest_group_capacity
= sgs
.group_capacity
;
4661 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
4662 sds
->busiest_has_capacity
= sgs
.group_has_capacity
;
4663 sds
->busiest_group_weight
= sgs
.group_weight
;
4664 sds
->group_imb
= sgs
.group_imb
;
4668 } while (sg
!= env
->sd
->groups
);
4672 * check_asym_packing - Check to see if the group is packed into the
4675 * This is primarily intended to used at the sibling level. Some
4676 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4677 * case of POWER7, it can move to lower SMT modes only when higher
4678 * threads are idle. When in lower SMT modes, the threads will
4679 * perform better since they share less core resources. Hence when we
4680 * have idle threads, we want them to be the higher ones.
4682 * This packing function is run on idle threads. It checks to see if
4683 * the busiest CPU in this domain (core in the P7 case) has a higher
4684 * CPU number than the packing function is being run on. Here we are
4685 * assuming lower CPU number will be equivalent to lower a SMT thread
4688 * Returns 1 when packing is required and a task should be moved to
4689 * this CPU. The amount of the imbalance is returned in *imbalance.
4691 * @env: The load balancing environment.
4692 * @sds: Statistics of the sched_domain which is to be packed
4694 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4698 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
4704 busiest_cpu
= group_first_cpu(sds
->busiest
);
4705 if (env
->dst_cpu
> busiest_cpu
)
4708 env
->imbalance
= DIV_ROUND_CLOSEST(
4709 sds
->max_load
* sds
->busiest
->sgp
->power
, SCHED_POWER_SCALE
);
4715 * fix_small_imbalance - Calculate the minor imbalance that exists
4716 * amongst the groups of a sched_domain, during
4718 * @env: The load balancing environment.
4719 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4722 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4724 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
4725 unsigned int imbn
= 2;
4726 unsigned long scaled_busy_load_per_task
;
4728 if (sds
->this_nr_running
) {
4729 sds
->this_load_per_task
/= sds
->this_nr_running
;
4730 if (sds
->busiest_load_per_task
>
4731 sds
->this_load_per_task
)
4734 sds
->this_load_per_task
=
4735 cpu_avg_load_per_task(env
->dst_cpu
);
4738 scaled_busy_load_per_task
= sds
->busiest_load_per_task
4739 * SCHED_POWER_SCALE
;
4740 scaled_busy_load_per_task
/= sds
->busiest
->sgp
->power
;
4742 if (sds
->max_load
- sds
->this_load
+ scaled_busy_load_per_task
>=
4743 (scaled_busy_load_per_task
* imbn
)) {
4744 env
->imbalance
= sds
->busiest_load_per_task
;
4749 * OK, we don't have enough imbalance to justify moving tasks,
4750 * however we may be able to increase total CPU power used by
4754 pwr_now
+= sds
->busiest
->sgp
->power
*
4755 min(sds
->busiest_load_per_task
, sds
->max_load
);
4756 pwr_now
+= sds
->this->sgp
->power
*
4757 min(sds
->this_load_per_task
, sds
->this_load
);
4758 pwr_now
/= SCHED_POWER_SCALE
;
4760 /* Amount of load we'd subtract */
4761 tmp
= (sds
->busiest_load_per_task
* SCHED_POWER_SCALE
) /
4762 sds
->busiest
->sgp
->power
;
4763 if (sds
->max_load
> tmp
)
4764 pwr_move
+= sds
->busiest
->sgp
->power
*
4765 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
4767 /* Amount of load we'd add */
4768 if (sds
->max_load
* sds
->busiest
->sgp
->power
<
4769 sds
->busiest_load_per_task
* SCHED_POWER_SCALE
)
4770 tmp
= (sds
->max_load
* sds
->busiest
->sgp
->power
) /
4771 sds
->this->sgp
->power
;
4773 tmp
= (sds
->busiest_load_per_task
* SCHED_POWER_SCALE
) /
4774 sds
->this->sgp
->power
;
4775 pwr_move
+= sds
->this->sgp
->power
*
4776 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
4777 pwr_move
/= SCHED_POWER_SCALE
;
4779 /* Move if we gain throughput */
4780 if (pwr_move
> pwr_now
)
4781 env
->imbalance
= sds
->busiest_load_per_task
;
4785 * calculate_imbalance - Calculate the amount of imbalance present within the
4786 * groups of a given sched_domain during load balance.
4787 * @env: load balance environment
4788 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4790 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4792 unsigned long max_pull
, load_above_capacity
= ~0UL;
4794 sds
->busiest_load_per_task
/= sds
->busiest_nr_running
;
4795 if (sds
->group_imb
) {
4796 sds
->busiest_load_per_task
=
4797 min(sds
->busiest_load_per_task
, sds
->avg_load
);
4801 * In the presence of smp nice balancing, certain scenarios can have
4802 * max load less than avg load(as we skip the groups at or below
4803 * its cpu_power, while calculating max_load..)
4805 if (sds
->max_load
< sds
->avg_load
) {
4807 return fix_small_imbalance(env
, sds
);
4810 if (!sds
->group_imb
) {
4812 * Don't want to pull so many tasks that a group would go idle.
4814 load_above_capacity
= (sds
->busiest_nr_running
-
4815 sds
->busiest_group_capacity
);
4817 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_POWER_SCALE
);
4819 load_above_capacity
/= sds
->busiest
->sgp
->power
;
4823 * We're trying to get all the cpus to the average_load, so we don't
4824 * want to push ourselves above the average load, nor do we wish to
4825 * reduce the max loaded cpu below the average load. At the same time,
4826 * we also don't want to reduce the group load below the group capacity
4827 * (so that we can implement power-savings policies etc). Thus we look
4828 * for the minimum possible imbalance.
4829 * Be careful of negative numbers as they'll appear as very large values
4830 * with unsigned longs.
4832 max_pull
= min(sds
->max_load
- sds
->avg_load
, load_above_capacity
);
4834 /* How much load to actually move to equalise the imbalance */
4835 env
->imbalance
= min(max_pull
* sds
->busiest
->sgp
->power
,
4836 (sds
->avg_load
- sds
->this_load
) * sds
->this->sgp
->power
)
4837 / SCHED_POWER_SCALE
;
4840 * if *imbalance is less than the average load per runnable task
4841 * there is no guarantee that any tasks will be moved so we'll have
4842 * a think about bumping its value to force at least one task to be
4845 if (env
->imbalance
< sds
->busiest_load_per_task
)
4846 return fix_small_imbalance(env
, sds
);
4850 /******* find_busiest_group() helpers end here *********************/
4853 * find_busiest_group - Returns the busiest group within the sched_domain
4854 * if there is an imbalance. If there isn't an imbalance, and
4855 * the user has opted for power-savings, it returns a group whose
4856 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4857 * such a group exists.
4859 * Also calculates the amount of weighted load which should be moved
4860 * to restore balance.
4862 * @env: The load balancing environment.
4863 * @balance: Pointer to a variable indicating if this_cpu
4864 * is the appropriate cpu to perform load balancing at this_level.
4866 * Returns: - the busiest group if imbalance exists.
4867 * - If no imbalance and user has opted for power-savings balance,
4868 * return the least loaded group whose CPUs can be
4869 * put to idle by rebalancing its tasks onto our group.
4871 static struct sched_group
*
4872 find_busiest_group(struct lb_env
*env
, int *balance
)
4874 struct sd_lb_stats sds
;
4876 memset(&sds
, 0, sizeof(sds
));
4879 * Compute the various statistics relavent for load balancing at
4882 update_sd_lb_stats(env
, balance
, &sds
);
4885 * this_cpu is not the appropriate cpu to perform load balancing at
4891 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
4892 check_asym_packing(env
, &sds
))
4895 /* There is no busy sibling group to pull tasks from */
4896 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
4899 sds
.avg_load
= (SCHED_POWER_SCALE
* sds
.total_load
) / sds
.total_pwr
;
4902 * If the busiest group is imbalanced the below checks don't
4903 * work because they assumes all things are equal, which typically
4904 * isn't true due to cpus_allowed constraints and the like.
4909 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4910 if (env
->idle
== CPU_NEWLY_IDLE
&& sds
.this_has_capacity
&&
4911 !sds
.busiest_has_capacity
)
4915 * If the local group is more busy than the selected busiest group
4916 * don't try and pull any tasks.
4918 if (sds
.this_load
>= sds
.max_load
)
4922 * Don't pull any tasks if this group is already above the domain
4925 if (sds
.this_load
>= sds
.avg_load
)
4928 if (env
->idle
== CPU_IDLE
) {
4930 * This cpu is idle. If the busiest group load doesn't
4931 * have more tasks than the number of available cpu's and
4932 * there is no imbalance between this and busiest group
4933 * wrt to idle cpu's, it is balanced.
4935 if ((sds
.this_idle_cpus
<= sds
.busiest_idle_cpus
+ 1) &&
4936 sds
.busiest_nr_running
<= sds
.busiest_group_weight
)
4940 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4941 * imbalance_pct to be conservative.
4943 if (100 * sds
.max_load
<= env
->sd
->imbalance_pct
* sds
.this_load
)
4948 /* Looks like there is an imbalance. Compute it */
4949 calculate_imbalance(env
, &sds
);
4959 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4961 static struct rq
*find_busiest_queue(struct lb_env
*env
,
4962 struct sched_group
*group
)
4964 struct rq
*busiest
= NULL
, *rq
;
4965 unsigned long max_load
= 0;
4968 for_each_cpu(i
, sched_group_cpus(group
)) {
4969 unsigned long power
= power_of(i
);
4970 unsigned long capacity
= DIV_ROUND_CLOSEST(power
,
4975 capacity
= fix_small_capacity(env
->sd
, group
);
4977 if (!cpumask_test_cpu(i
, env
->cpus
))
4981 wl
= weighted_cpuload(i
);
4984 * When comparing with imbalance, use weighted_cpuload()
4985 * which is not scaled with the cpu power.
4987 if (capacity
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
4991 * For the load comparisons with the other cpu's, consider
4992 * the weighted_cpuload() scaled with the cpu power, so that
4993 * the load can be moved away from the cpu that is potentially
4994 * running at a lower capacity.
4996 wl
= (wl
* SCHED_POWER_SCALE
) / power
;
4998 if (wl
> max_load
) {
5008 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5009 * so long as it is large enough.
5011 #define MAX_PINNED_INTERVAL 512
5013 /* Working cpumask for load_balance and load_balance_newidle. */
5014 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5016 static int need_active_balance(struct lb_env
*env
)
5018 struct sched_domain
*sd
= env
->sd
;
5020 if (env
->idle
== CPU_NEWLY_IDLE
) {
5023 * ASYM_PACKING needs to force migrate tasks from busy but
5024 * higher numbered CPUs in order to pack all tasks in the
5025 * lowest numbered CPUs.
5027 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
5031 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
5034 static int active_load_balance_cpu_stop(void *data
);
5037 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5038 * tasks if there is an imbalance.
5040 static int load_balance(int this_cpu
, struct rq
*this_rq
,
5041 struct sched_domain
*sd
, enum cpu_idle_type idle
,
5044 int ld_moved
, cur_ld_moved
, active_balance
= 0;
5045 struct sched_group
*group
;
5047 unsigned long flags
;
5048 struct cpumask
*cpus
= __get_cpu_var(load_balance_mask
);
5050 struct lb_env env
= {
5052 .dst_cpu
= this_cpu
,
5054 .dst_grpmask
= sched_group_cpus(sd
->groups
),
5056 .loop_break
= sched_nr_migrate_break
,
5061 * For NEWLY_IDLE load_balancing, we don't need to consider
5062 * other cpus in our group
5064 if (idle
== CPU_NEWLY_IDLE
)
5065 env
.dst_grpmask
= NULL
;
5067 cpumask_copy(cpus
, cpu_active_mask
);
5069 schedstat_inc(sd
, lb_count
[idle
]);
5072 group
= find_busiest_group(&env
, balance
);
5078 schedstat_inc(sd
, lb_nobusyg
[idle
]);
5082 busiest
= find_busiest_queue(&env
, group
);
5084 schedstat_inc(sd
, lb_nobusyq
[idle
]);
5088 BUG_ON(busiest
== env
.dst_rq
);
5090 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
5093 if (busiest
->nr_running
> 1) {
5095 * Attempt to move tasks. If find_busiest_group has found
5096 * an imbalance but busiest->nr_running <= 1, the group is
5097 * still unbalanced. ld_moved simply stays zero, so it is
5098 * correctly treated as an imbalance.
5100 env
.flags
|= LBF_ALL_PINNED
;
5101 env
.src_cpu
= busiest
->cpu
;
5102 env
.src_rq
= busiest
;
5103 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
5105 update_h_load(env
.src_cpu
);
5107 local_irq_save(flags
);
5108 double_rq_lock(env
.dst_rq
, busiest
);
5111 * cur_ld_moved - load moved in current iteration
5112 * ld_moved - cumulative load moved across iterations
5114 cur_ld_moved
= move_tasks(&env
);
5115 ld_moved
+= cur_ld_moved
;
5116 double_rq_unlock(env
.dst_rq
, busiest
);
5117 local_irq_restore(flags
);
5120 * some other cpu did the load balance for us.
5122 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
5123 resched_cpu(env
.dst_cpu
);
5125 if (env
.flags
& LBF_NEED_BREAK
) {
5126 env
.flags
&= ~LBF_NEED_BREAK
;
5131 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5132 * us and move them to an alternate dst_cpu in our sched_group
5133 * where they can run. The upper limit on how many times we
5134 * iterate on same src_cpu is dependent on number of cpus in our
5137 * This changes load balance semantics a bit on who can move
5138 * load to a given_cpu. In addition to the given_cpu itself
5139 * (or a ilb_cpu acting on its behalf where given_cpu is
5140 * nohz-idle), we now have balance_cpu in a position to move
5141 * load to given_cpu. In rare situations, this may cause
5142 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5143 * _independently_ and at _same_ time to move some load to
5144 * given_cpu) causing exceess load to be moved to given_cpu.
5145 * This however should not happen so much in practice and
5146 * moreover subsequent load balance cycles should correct the
5147 * excess load moved.
5149 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0) {
5151 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
5152 env
.dst_cpu
= env
.new_dst_cpu
;
5153 env
.flags
&= ~LBF_SOME_PINNED
;
5155 env
.loop_break
= sched_nr_migrate_break
;
5157 /* Prevent to re-select dst_cpu via env's cpus */
5158 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
5161 * Go back to "more_balance" rather than "redo" since we
5162 * need to continue with same src_cpu.
5167 /* All tasks on this runqueue were pinned by CPU affinity */
5168 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
5169 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
5170 if (!cpumask_empty(cpus
)) {
5172 env
.loop_break
= sched_nr_migrate_break
;
5180 schedstat_inc(sd
, lb_failed
[idle
]);
5182 * Increment the failure counter only on periodic balance.
5183 * We do not want newidle balance, which can be very
5184 * frequent, pollute the failure counter causing
5185 * excessive cache_hot migrations and active balances.
5187 if (idle
!= CPU_NEWLY_IDLE
)
5188 sd
->nr_balance_failed
++;
5190 if (need_active_balance(&env
)) {
5191 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
5193 /* don't kick the active_load_balance_cpu_stop,
5194 * if the curr task on busiest cpu can't be
5197 if (!cpumask_test_cpu(this_cpu
,
5198 tsk_cpus_allowed(busiest
->curr
))) {
5199 raw_spin_unlock_irqrestore(&busiest
->lock
,
5201 env
.flags
|= LBF_ALL_PINNED
;
5202 goto out_one_pinned
;
5206 * ->active_balance synchronizes accesses to
5207 * ->active_balance_work. Once set, it's cleared
5208 * only after active load balance is finished.
5210 if (!busiest
->active_balance
) {
5211 busiest
->active_balance
= 1;
5212 busiest
->push_cpu
= this_cpu
;
5215 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
5217 if (active_balance
) {
5218 stop_one_cpu_nowait(cpu_of(busiest
),
5219 active_load_balance_cpu_stop
, busiest
,
5220 &busiest
->active_balance_work
);
5224 * We've kicked active balancing, reset the failure
5227 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
5230 sd
->nr_balance_failed
= 0;
5232 if (likely(!active_balance
)) {
5233 /* We were unbalanced, so reset the balancing interval */
5234 sd
->balance_interval
= sd
->min_interval
;
5237 * If we've begun active balancing, start to back off. This
5238 * case may not be covered by the all_pinned logic if there
5239 * is only 1 task on the busy runqueue (because we don't call
5242 if (sd
->balance_interval
< sd
->max_interval
)
5243 sd
->balance_interval
*= 2;
5249 schedstat_inc(sd
, lb_balanced
[idle
]);
5251 sd
->nr_balance_failed
= 0;
5254 /* tune up the balancing interval */
5255 if (((env
.flags
& LBF_ALL_PINNED
) &&
5256 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
5257 (sd
->balance_interval
< sd
->max_interval
))
5258 sd
->balance_interval
*= 2;
5266 * idle_balance is called by schedule() if this_cpu is about to become
5267 * idle. Attempts to pull tasks from other CPUs.
5269 void idle_balance(int this_cpu
, struct rq
*this_rq
)
5271 struct sched_domain
*sd
;
5272 int pulled_task
= 0;
5273 unsigned long next_balance
= jiffies
+ HZ
;
5275 this_rq
->idle_stamp
= rq_clock(this_rq
);
5277 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
5281 * Drop the rq->lock, but keep IRQ/preempt disabled.
5283 raw_spin_unlock(&this_rq
->lock
);
5285 update_blocked_averages(this_cpu
);
5287 for_each_domain(this_cpu
, sd
) {
5288 unsigned long interval
;
5291 if (!(sd
->flags
& SD_LOAD_BALANCE
))
5294 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
5295 /* If we've pulled tasks over stop searching: */
5296 pulled_task
= load_balance(this_cpu
, this_rq
,
5297 sd
, CPU_NEWLY_IDLE
, &balance
);
5300 interval
= msecs_to_jiffies(sd
->balance_interval
);
5301 if (time_after(next_balance
, sd
->last_balance
+ interval
))
5302 next_balance
= sd
->last_balance
+ interval
;
5304 this_rq
->idle_stamp
= 0;
5310 raw_spin_lock(&this_rq
->lock
);
5312 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
5314 * We are going idle. next_balance may be set based on
5315 * a busy processor. So reset next_balance.
5317 this_rq
->next_balance
= next_balance
;
5322 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5323 * running tasks off the busiest CPU onto idle CPUs. It requires at
5324 * least 1 task to be running on each physical CPU where possible, and
5325 * avoids physical / logical imbalances.
5327 static int active_load_balance_cpu_stop(void *data
)
5329 struct rq
*busiest_rq
= data
;
5330 int busiest_cpu
= cpu_of(busiest_rq
);
5331 int target_cpu
= busiest_rq
->push_cpu
;
5332 struct rq
*target_rq
= cpu_rq(target_cpu
);
5333 struct sched_domain
*sd
;
5335 raw_spin_lock_irq(&busiest_rq
->lock
);
5337 /* make sure the requested cpu hasn't gone down in the meantime */
5338 if (unlikely(busiest_cpu
!= smp_processor_id() ||
5339 !busiest_rq
->active_balance
))
5342 /* Is there any task to move? */
5343 if (busiest_rq
->nr_running
<= 1)
5347 * This condition is "impossible", if it occurs
5348 * we need to fix it. Originally reported by
5349 * Bjorn Helgaas on a 128-cpu setup.
5351 BUG_ON(busiest_rq
== target_rq
);
5353 /* move a task from busiest_rq to target_rq */
5354 double_lock_balance(busiest_rq
, target_rq
);
5356 /* Search for an sd spanning us and the target CPU. */
5358 for_each_domain(target_cpu
, sd
) {
5359 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
5360 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
5365 struct lb_env env
= {
5367 .dst_cpu
= target_cpu
,
5368 .dst_rq
= target_rq
,
5369 .src_cpu
= busiest_rq
->cpu
,
5370 .src_rq
= busiest_rq
,
5374 schedstat_inc(sd
, alb_count
);
5376 if (move_one_task(&env
))
5377 schedstat_inc(sd
, alb_pushed
);
5379 schedstat_inc(sd
, alb_failed
);
5382 double_unlock_balance(busiest_rq
, target_rq
);
5384 busiest_rq
->active_balance
= 0;
5385 raw_spin_unlock_irq(&busiest_rq
->lock
);
5389 #ifdef CONFIG_NO_HZ_COMMON
5391 * idle load balancing details
5392 * - When one of the busy CPUs notice that there may be an idle rebalancing
5393 * needed, they will kick the idle load balancer, which then does idle
5394 * load balancing for all the idle CPUs.
5397 cpumask_var_t idle_cpus_mask
;
5399 unsigned long next_balance
; /* in jiffy units */
5400 } nohz ____cacheline_aligned
;
5402 static inline int find_new_ilb(int call_cpu
)
5404 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
5406 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
5413 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5414 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5415 * CPU (if there is one).
5417 static void nohz_balancer_kick(int cpu
)
5421 nohz
.next_balance
++;
5423 ilb_cpu
= find_new_ilb(cpu
);
5425 if (ilb_cpu
>= nr_cpu_ids
)
5428 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
5431 * Use smp_send_reschedule() instead of resched_cpu().
5432 * This way we generate a sched IPI on the target cpu which
5433 * is idle. And the softirq performing nohz idle load balance
5434 * will be run before returning from the IPI.
5436 smp_send_reschedule(ilb_cpu
);
5440 static inline void nohz_balance_exit_idle(int cpu
)
5442 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
5443 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
5444 atomic_dec(&nohz
.nr_cpus
);
5445 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
5449 static inline void set_cpu_sd_state_busy(void)
5451 struct sched_domain
*sd
;
5454 sd
= rcu_dereference_check_sched_domain(this_rq()->sd
);
5456 if (!sd
|| !sd
->nohz_idle
)
5460 for (; sd
; sd
= sd
->parent
)
5461 atomic_inc(&sd
->groups
->sgp
->nr_busy_cpus
);
5466 void set_cpu_sd_state_idle(void)
5468 struct sched_domain
*sd
;
5471 sd
= rcu_dereference_check_sched_domain(this_rq()->sd
);
5473 if (!sd
|| sd
->nohz_idle
)
5477 for (; sd
; sd
= sd
->parent
)
5478 atomic_dec(&sd
->groups
->sgp
->nr_busy_cpus
);
5484 * This routine will record that the cpu is going idle with tick stopped.
5485 * This info will be used in performing idle load balancing in the future.
5487 void nohz_balance_enter_idle(int cpu
)
5490 * If this cpu is going down, then nothing needs to be done.
5492 if (!cpu_active(cpu
))
5495 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
5498 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
5499 atomic_inc(&nohz
.nr_cpus
);
5500 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
5503 static int __cpuinit
sched_ilb_notifier(struct notifier_block
*nfb
,
5504 unsigned long action
, void *hcpu
)
5506 switch (action
& ~CPU_TASKS_FROZEN
) {
5508 nohz_balance_exit_idle(smp_processor_id());
5516 static DEFINE_SPINLOCK(balancing
);
5519 * Scale the max load_balance interval with the number of CPUs in the system.
5520 * This trades load-balance latency on larger machines for less cross talk.
5522 void update_max_interval(void)
5524 max_load_balance_interval
= HZ
*num_online_cpus()/10;
5528 * It checks each scheduling domain to see if it is due to be balanced,
5529 * and initiates a balancing operation if so.
5531 * Balancing parameters are set up in init_sched_domains.
5533 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
5536 struct rq
*rq
= cpu_rq(cpu
);
5537 unsigned long interval
;
5538 struct sched_domain
*sd
;
5539 /* Earliest time when we have to do rebalance again */
5540 unsigned long next_balance
= jiffies
+ 60*HZ
;
5541 int update_next_balance
= 0;
5544 update_blocked_averages(cpu
);
5547 for_each_domain(cpu
, sd
) {
5548 if (!(sd
->flags
& SD_LOAD_BALANCE
))
5551 interval
= sd
->balance_interval
;
5552 if (idle
!= CPU_IDLE
)
5553 interval
*= sd
->busy_factor
;
5555 /* scale ms to jiffies */
5556 interval
= msecs_to_jiffies(interval
);
5557 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5559 need_serialize
= sd
->flags
& SD_SERIALIZE
;
5561 if (need_serialize
) {
5562 if (!spin_trylock(&balancing
))
5566 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
5567 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
5569 * The LBF_SOME_PINNED logic could have changed
5570 * env->dst_cpu, so we can't know our idle
5571 * state even if we migrated tasks. Update it.
5573 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
5575 sd
->last_balance
= jiffies
;
5578 spin_unlock(&balancing
);
5580 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
5581 next_balance
= sd
->last_balance
+ interval
;
5582 update_next_balance
= 1;
5586 * Stop the load balance at this level. There is another
5587 * CPU in our sched group which is doing load balancing more
5596 * next_balance will be updated only when there is a need.
5597 * When the cpu is attached to null domain for ex, it will not be
5600 if (likely(update_next_balance
))
5601 rq
->next_balance
= next_balance
;
5604 #ifdef CONFIG_NO_HZ_COMMON
5606 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5607 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5609 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
)
5611 struct rq
*this_rq
= cpu_rq(this_cpu
);
5615 if (idle
!= CPU_IDLE
||
5616 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
5619 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
5620 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
5624 * If this cpu gets work to do, stop the load balancing
5625 * work being done for other cpus. Next load
5626 * balancing owner will pick it up.
5631 rq
= cpu_rq(balance_cpu
);
5633 raw_spin_lock_irq(&rq
->lock
);
5634 update_rq_clock(rq
);
5635 update_idle_cpu_load(rq
);
5636 raw_spin_unlock_irq(&rq
->lock
);
5638 rebalance_domains(balance_cpu
, CPU_IDLE
);
5640 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
5641 this_rq
->next_balance
= rq
->next_balance
;
5643 nohz
.next_balance
= this_rq
->next_balance
;
5645 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
5649 * Current heuristic for kicking the idle load balancer in the presence
5650 * of an idle cpu is the system.
5651 * - This rq has more than one task.
5652 * - At any scheduler domain level, this cpu's scheduler group has multiple
5653 * busy cpu's exceeding the group's power.
5654 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5655 * domain span are idle.
5657 static inline int nohz_kick_needed(struct rq
*rq
, int cpu
)
5659 unsigned long now
= jiffies
;
5660 struct sched_domain
*sd
;
5662 if (unlikely(idle_cpu(cpu
)))
5666 * We may be recently in ticked or tickless idle mode. At the first
5667 * busy tick after returning from idle, we will update the busy stats.
5669 set_cpu_sd_state_busy();
5670 nohz_balance_exit_idle(cpu
);
5673 * None are in tickless mode and hence no need for NOHZ idle load
5676 if (likely(!atomic_read(&nohz
.nr_cpus
)))
5679 if (time_before(now
, nohz
.next_balance
))
5682 if (rq
->nr_running
>= 2)
5686 for_each_domain(cpu
, sd
) {
5687 struct sched_group
*sg
= sd
->groups
;
5688 struct sched_group_power
*sgp
= sg
->sgp
;
5689 int nr_busy
= atomic_read(&sgp
->nr_busy_cpus
);
5691 if (sd
->flags
& SD_SHARE_PKG_RESOURCES
&& nr_busy
> 1)
5692 goto need_kick_unlock
;
5694 if (sd
->flags
& SD_ASYM_PACKING
&& nr_busy
!= sg
->group_weight
5695 && (cpumask_first_and(nohz
.idle_cpus_mask
,
5696 sched_domain_span(sd
)) < cpu
))
5697 goto need_kick_unlock
;
5699 if (!(sd
->flags
& (SD_SHARE_PKG_RESOURCES
| SD_ASYM_PACKING
)))
5711 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
) { }
5715 * run_rebalance_domains is triggered when needed from the scheduler tick.
5716 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5718 static void run_rebalance_domains(struct softirq_action
*h
)
5720 int this_cpu
= smp_processor_id();
5721 struct rq
*this_rq
= cpu_rq(this_cpu
);
5722 enum cpu_idle_type idle
= this_rq
->idle_balance
?
5723 CPU_IDLE
: CPU_NOT_IDLE
;
5725 rebalance_domains(this_cpu
, idle
);
5728 * If this cpu has a pending nohz_balance_kick, then do the
5729 * balancing on behalf of the other idle cpus whose ticks are
5732 nohz_idle_balance(this_cpu
, idle
);
5735 static inline int on_null_domain(int cpu
)
5737 return !rcu_dereference_sched(cpu_rq(cpu
)->sd
);
5741 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5743 void trigger_load_balance(struct rq
*rq
, int cpu
)
5745 /* Don't need to rebalance while attached to NULL domain */
5746 if (time_after_eq(jiffies
, rq
->next_balance
) &&
5747 likely(!on_null_domain(cpu
)))
5748 raise_softirq(SCHED_SOFTIRQ
);
5749 #ifdef CONFIG_NO_HZ_COMMON
5750 if (nohz_kick_needed(rq
, cpu
) && likely(!on_null_domain(cpu
)))
5751 nohz_balancer_kick(cpu
);
5755 static void rq_online_fair(struct rq
*rq
)
5760 static void rq_offline_fair(struct rq
*rq
)
5764 /* Ensure any throttled groups are reachable by pick_next_task */
5765 unthrottle_offline_cfs_rqs(rq
);
5768 #endif /* CONFIG_SMP */
5771 * scheduler tick hitting a task of our scheduling class:
5773 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
5775 struct cfs_rq
*cfs_rq
;
5776 struct sched_entity
*se
= &curr
->se
;
5778 for_each_sched_entity(se
) {
5779 cfs_rq
= cfs_rq_of(se
);
5780 entity_tick(cfs_rq
, se
, queued
);
5783 if (sched_feat_numa(NUMA
))
5784 task_tick_numa(rq
, curr
);
5786 update_rq_runnable_avg(rq
, 1);
5790 * called on fork with the child task as argument from the parent's context
5791 * - child not yet on the tasklist
5792 * - preemption disabled
5794 static void task_fork_fair(struct task_struct
*p
)
5796 struct cfs_rq
*cfs_rq
;
5797 struct sched_entity
*se
= &p
->se
, *curr
;
5798 int this_cpu
= smp_processor_id();
5799 struct rq
*rq
= this_rq();
5800 unsigned long flags
;
5802 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5804 update_rq_clock(rq
);
5806 cfs_rq
= task_cfs_rq(current
);
5807 curr
= cfs_rq
->curr
;
5809 if (unlikely(task_cpu(p
) != this_cpu
)) {
5811 __set_task_cpu(p
, this_cpu
);
5815 update_curr(cfs_rq
);
5818 se
->vruntime
= curr
->vruntime
;
5819 place_entity(cfs_rq
, se
, 1);
5821 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
5823 * Upon rescheduling, sched_class::put_prev_task() will place
5824 * 'current' within the tree based on its new key value.
5826 swap(curr
->vruntime
, se
->vruntime
);
5827 resched_task(rq
->curr
);
5830 se
->vruntime
-= cfs_rq
->min_vruntime
;
5832 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5836 * Priority of the task has changed. Check to see if we preempt
5840 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
5846 * Reschedule if we are currently running on this runqueue and
5847 * our priority decreased, or if we are not currently running on
5848 * this runqueue and our priority is higher than the current's
5850 if (rq
->curr
== p
) {
5851 if (p
->prio
> oldprio
)
5852 resched_task(rq
->curr
);
5854 check_preempt_curr(rq
, p
, 0);
5857 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
5859 struct sched_entity
*se
= &p
->se
;
5860 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5863 * Ensure the task's vruntime is normalized, so that when its
5864 * switched back to the fair class the enqueue_entity(.flags=0) will
5865 * do the right thing.
5867 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5868 * have normalized the vruntime, if it was !on_rq, then only when
5869 * the task is sleeping will it still have non-normalized vruntime.
5871 if (!se
->on_rq
&& p
->state
!= TASK_RUNNING
) {
5873 * Fix up our vruntime so that the current sleep doesn't
5874 * cause 'unlimited' sleep bonus.
5876 place_entity(cfs_rq
, se
, 0);
5877 se
->vruntime
-= cfs_rq
->min_vruntime
;
5882 * Remove our load from contribution when we leave sched_fair
5883 * and ensure we don't carry in an old decay_count if we
5886 if (p
->se
.avg
.decay_count
) {
5887 struct cfs_rq
*cfs_rq
= cfs_rq_of(&p
->se
);
5888 __synchronize_entity_decay(&p
->se
);
5889 subtract_blocked_load_contrib(cfs_rq
,
5890 p
->se
.avg
.load_avg_contrib
);
5896 * We switched to the sched_fair class.
5898 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
5904 * We were most likely switched from sched_rt, so
5905 * kick off the schedule if running, otherwise just see
5906 * if we can still preempt the current task.
5909 resched_task(rq
->curr
);
5911 check_preempt_curr(rq
, p
, 0);
5914 /* Account for a task changing its policy or group.
5916 * This routine is mostly called to set cfs_rq->curr field when a task
5917 * migrates between groups/classes.
5919 static void set_curr_task_fair(struct rq
*rq
)
5921 struct sched_entity
*se
= &rq
->curr
->se
;
5923 for_each_sched_entity(se
) {
5924 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
5926 set_next_entity(cfs_rq
, se
);
5927 /* ensure bandwidth has been allocated on our new cfs_rq */
5928 account_cfs_rq_runtime(cfs_rq
, 0);
5932 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
5934 cfs_rq
->tasks_timeline
= RB_ROOT
;
5935 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
5936 #ifndef CONFIG_64BIT
5937 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
5940 atomic64_set(&cfs_rq
->decay_counter
, 1);
5941 atomic64_set(&cfs_rq
->removed_load
, 0);
5945 #ifdef CONFIG_FAIR_GROUP_SCHED
5946 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
5948 struct cfs_rq
*cfs_rq
;
5950 * If the task was not on the rq at the time of this cgroup movement
5951 * it must have been asleep, sleeping tasks keep their ->vruntime
5952 * absolute on their old rq until wakeup (needed for the fair sleeper
5953 * bonus in place_entity()).
5955 * If it was on the rq, we've just 'preempted' it, which does convert
5956 * ->vruntime to a relative base.
5958 * Make sure both cases convert their relative position when migrating
5959 * to another cgroup's rq. This does somewhat interfere with the
5960 * fair sleeper stuff for the first placement, but who cares.
5963 * When !on_rq, vruntime of the task has usually NOT been normalized.
5964 * But there are some cases where it has already been normalized:
5966 * - Moving a forked child which is waiting for being woken up by
5967 * wake_up_new_task().
5968 * - Moving a task which has been woken up by try_to_wake_up() and
5969 * waiting for actually being woken up by sched_ttwu_pending().
5971 * To prevent boost or penalty in the new cfs_rq caused by delta
5972 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5974 if (!on_rq
&& (!p
->se
.sum_exec_runtime
|| p
->state
== TASK_WAKING
))
5978 p
->se
.vruntime
-= cfs_rq_of(&p
->se
)->min_vruntime
;
5979 set_task_rq(p
, task_cpu(p
));
5981 cfs_rq
= cfs_rq_of(&p
->se
);
5982 p
->se
.vruntime
+= cfs_rq
->min_vruntime
;
5985 * migrate_task_rq_fair() will have removed our previous
5986 * contribution, but we must synchronize for ongoing future
5989 p
->se
.avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
5990 cfs_rq
->blocked_load_avg
+= p
->se
.avg
.load_avg_contrib
;
5995 void free_fair_sched_group(struct task_group
*tg
)
5999 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
6001 for_each_possible_cpu(i
) {
6003 kfree(tg
->cfs_rq
[i
]);
6012 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
6014 struct cfs_rq
*cfs_rq
;
6015 struct sched_entity
*se
;
6018 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
6021 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
6025 tg
->shares
= NICE_0_LOAD
;
6027 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
6029 for_each_possible_cpu(i
) {
6030 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
6031 GFP_KERNEL
, cpu_to_node(i
));
6035 se
= kzalloc_node(sizeof(struct sched_entity
),
6036 GFP_KERNEL
, cpu_to_node(i
));
6040 init_cfs_rq(cfs_rq
);
6041 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
6052 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
6054 struct rq
*rq
= cpu_rq(cpu
);
6055 unsigned long flags
;
6058 * Only empty task groups can be destroyed; so we can speculatively
6059 * check on_list without danger of it being re-added.
6061 if (!tg
->cfs_rq
[cpu
]->on_list
)
6064 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6065 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
6066 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6069 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
6070 struct sched_entity
*se
, int cpu
,
6071 struct sched_entity
*parent
)
6073 struct rq
*rq
= cpu_rq(cpu
);
6077 init_cfs_rq_runtime(cfs_rq
);
6079 tg
->cfs_rq
[cpu
] = cfs_rq
;
6082 /* se could be NULL for root_task_group */
6087 se
->cfs_rq
= &rq
->cfs
;
6089 se
->cfs_rq
= parent
->my_q
;
6092 update_load_set(&se
->load
, 0);
6093 se
->parent
= parent
;
6096 static DEFINE_MUTEX(shares_mutex
);
6098 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
6101 unsigned long flags
;
6104 * We can't change the weight of the root cgroup.
6109 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
6111 mutex_lock(&shares_mutex
);
6112 if (tg
->shares
== shares
)
6115 tg
->shares
= shares
;
6116 for_each_possible_cpu(i
) {
6117 struct rq
*rq
= cpu_rq(i
);
6118 struct sched_entity
*se
;
6121 /* Propagate contribution to hierarchy */
6122 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6124 /* Possible calls to update_curr() need rq clock */
6125 update_rq_clock(rq
);
6126 for_each_sched_entity(se
)
6127 update_cfs_shares(group_cfs_rq(se
));
6128 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6132 mutex_unlock(&shares_mutex
);
6135 #else /* CONFIG_FAIR_GROUP_SCHED */
6137 void free_fair_sched_group(struct task_group
*tg
) { }
6139 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
6144 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
6146 #endif /* CONFIG_FAIR_GROUP_SCHED */
6149 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
6151 struct sched_entity
*se
= &task
->se
;
6152 unsigned int rr_interval
= 0;
6155 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6158 if (rq
->cfs
.load
.weight
)
6159 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
6165 * All the scheduling class methods:
6167 const struct sched_class fair_sched_class
= {
6168 .next
= &idle_sched_class
,
6169 .enqueue_task
= enqueue_task_fair
,
6170 .dequeue_task
= dequeue_task_fair
,
6171 .yield_task
= yield_task_fair
,
6172 .yield_to_task
= yield_to_task_fair
,
6174 .check_preempt_curr
= check_preempt_wakeup
,
6176 .pick_next_task
= pick_next_task_fair
,
6177 .put_prev_task
= put_prev_task_fair
,
6180 .select_task_rq
= select_task_rq_fair
,
6181 .migrate_task_rq
= migrate_task_rq_fair
,
6183 .rq_online
= rq_online_fair
,
6184 .rq_offline
= rq_offline_fair
,
6186 .task_waking
= task_waking_fair
,
6189 .set_curr_task
= set_curr_task_fair
,
6190 .task_tick
= task_tick_fair
,
6191 .task_fork
= task_fork_fair
,
6193 .prio_changed
= prio_changed_fair
,
6194 .switched_from
= switched_from_fair
,
6195 .switched_to
= switched_to_fair
,
6197 .get_rr_interval
= get_rr_interval_fair
,
6199 #ifdef CONFIG_FAIR_GROUP_SCHED
6200 .task_move_group
= task_move_group_fair
,
6204 #ifdef CONFIG_SCHED_DEBUG
6205 void print_cfs_stats(struct seq_file
*m
, int cpu
)
6207 struct cfs_rq
*cfs_rq
;
6210 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
6211 print_cfs_rq(m
, cpu
, cfs_rq
);
6216 __init
void init_sched_fair_class(void)
6219 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
6221 #ifdef CONFIG_NO_HZ_COMMON
6222 nohz
.next_balance
= jiffies
;
6223 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
6224 cpu_notifier(sched_ilb_notifier
, 0);