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 unsigned long task_h_load(struct task_struct
*p
);
686 static inline void __update_task_entity_contrib(struct sched_entity
*se
);
688 /* Give new task start runnable values to heavy its load in infant time */
689 void init_task_runnable_average(struct task_struct
*p
)
693 p
->se
.avg
.decay_count
= 0;
694 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
695 p
->se
.avg
.runnable_avg_sum
= slice
;
696 p
->se
.avg
.runnable_avg_period
= slice
;
697 __update_task_entity_contrib(&p
->se
);
700 void init_task_runnable_average(struct task_struct
*p
)
706 * Update the current task's runtime statistics. Skip current tasks that
707 * are not in our scheduling class.
710 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
711 unsigned long delta_exec
)
713 unsigned long delta_exec_weighted
;
715 schedstat_set(curr
->statistics
.exec_max
,
716 max((u64
)delta_exec
, curr
->statistics
.exec_max
));
718 curr
->sum_exec_runtime
+= delta_exec
;
719 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
720 delta_exec_weighted
= calc_delta_fair(delta_exec
, curr
);
722 curr
->vruntime
+= delta_exec_weighted
;
723 update_min_vruntime(cfs_rq
);
726 static void update_curr(struct cfs_rq
*cfs_rq
)
728 struct sched_entity
*curr
= cfs_rq
->curr
;
729 u64 now
= rq_clock_task(rq_of(cfs_rq
));
730 unsigned long delta_exec
;
736 * Get the amount of time the current task was running
737 * since the last time we changed load (this cannot
738 * overflow on 32 bits):
740 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
744 __update_curr(cfs_rq
, curr
, delta_exec
);
745 curr
->exec_start
= now
;
747 if (entity_is_task(curr
)) {
748 struct task_struct
*curtask
= task_of(curr
);
750 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
751 cpuacct_charge(curtask
, delta_exec
);
752 account_group_exec_runtime(curtask
, delta_exec
);
755 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
759 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
761 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
765 * Task is being enqueued - update stats:
767 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
770 * Are we enqueueing a waiting task? (for current tasks
771 * a dequeue/enqueue event is a NOP)
773 if (se
!= cfs_rq
->curr
)
774 update_stats_wait_start(cfs_rq
, se
);
778 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
780 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
781 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
782 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
783 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
784 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
785 #ifdef CONFIG_SCHEDSTATS
786 if (entity_is_task(se
)) {
787 trace_sched_stat_wait(task_of(se
),
788 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
791 schedstat_set(se
->statistics
.wait_start
, 0);
795 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
798 * Mark the end of the wait period if dequeueing a
801 if (se
!= cfs_rq
->curr
)
802 update_stats_wait_end(cfs_rq
, se
);
806 * We are picking a new current task - update its stats:
809 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
812 * We are starting a new run period:
814 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
817 /**************************************************
818 * Scheduling class queueing methods:
821 #ifdef CONFIG_NUMA_BALANCING
823 * Approximate time to scan a full NUMA task in ms. The task scan period is
824 * calculated based on the tasks virtual memory size and
825 * numa_balancing_scan_size.
827 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
828 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
829 unsigned int sysctl_numa_balancing_scan_period_reset
= 60000;
831 /* Portion of address space to scan in MB */
832 unsigned int sysctl_numa_balancing_scan_size
= 256;
834 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
835 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
837 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
839 unsigned long rss
= 0;
840 unsigned long nr_scan_pages
;
843 * Calculations based on RSS as non-present and empty pages are skipped
844 * by the PTE scanner and NUMA hinting faults should be trapped based
847 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
848 rss
= get_mm_rss(p
->mm
);
852 rss
= round_up(rss
, nr_scan_pages
);
853 return rss
/ nr_scan_pages
;
856 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
857 #define MAX_SCAN_WINDOW 2560
859 static unsigned int task_scan_min(struct task_struct
*p
)
861 unsigned int scan
, floor
;
862 unsigned int windows
= 1;
864 if (sysctl_numa_balancing_scan_size
< MAX_SCAN_WINDOW
)
865 windows
= MAX_SCAN_WINDOW
/ sysctl_numa_balancing_scan_size
;
866 floor
= 1000 / windows
;
868 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
869 return max_t(unsigned int, floor
, scan
);
872 static unsigned int task_scan_max(struct task_struct
*p
)
874 unsigned int smin
= task_scan_min(p
);
877 /* Watch for min being lower than max due to floor calculations */
878 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
879 return max(smin
, smax
);
883 * Once a preferred node is selected the scheduler balancer will prefer moving
884 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
885 * scans. This will give the process the chance to accumulate more faults on
886 * the preferred node but still allow the scheduler to move the task again if
887 * the nodes CPUs are overloaded.
889 unsigned int sysctl_numa_balancing_settle_count __read_mostly
= 4;
894 spinlock_t lock
; /* nr_tasks, tasks */
897 struct list_head task_list
;
900 atomic_long_t faults
[0];
903 pid_t
task_numa_group_id(struct task_struct
*p
)
905 return p
->numa_group
? p
->numa_group
->gid
: 0;
908 static inline int task_faults_idx(int nid
, int priv
)
910 return 2 * nid
+ priv
;
913 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
918 return p
->numa_faults
[task_faults_idx(nid
, 0)] +
919 p
->numa_faults
[task_faults_idx(nid
, 1)];
922 static unsigned long weighted_cpuload(const int cpu
);
923 static unsigned long source_load(int cpu
, int type
);
924 static unsigned long target_load(int cpu
, int type
);
925 static unsigned long power_of(int cpu
);
926 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
928 /* Cached statistics for all CPUs within a node */
930 unsigned long nr_running
;
933 /* Total compute capacity of CPUs on a node */
936 /* Approximate capacity in terms of runnable tasks on a node */
937 unsigned long capacity
;
942 * XXX borrowed from update_sg_lb_stats
944 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
948 memset(ns
, 0, sizeof(*ns
));
949 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
950 struct rq
*rq
= cpu_rq(cpu
);
952 ns
->nr_running
+= rq
->nr_running
;
953 ns
->load
+= weighted_cpuload(cpu
);
954 ns
->power
+= power_of(cpu
);
957 ns
->load
= (ns
->load
* SCHED_POWER_SCALE
) / ns
->power
;
958 ns
->capacity
= DIV_ROUND_CLOSEST(ns
->power
, SCHED_POWER_SCALE
);
959 ns
->has_capacity
= (ns
->nr_running
< ns
->capacity
);
962 struct task_numa_env
{
963 struct task_struct
*p
;
965 int src_cpu
, src_nid
;
966 int dst_cpu
, dst_nid
;
968 struct numa_stats src_stats
, dst_stats
;
970 int imbalance_pct
, idx
;
972 struct task_struct
*best_task
;
977 static void task_numa_assign(struct task_numa_env
*env
,
978 struct task_struct
*p
, long imp
)
981 put_task_struct(env
->best_task
);
987 env
->best_cpu
= env
->dst_cpu
;
991 * This checks if the overall compute and NUMA accesses of the system would
992 * be improved if the source tasks was migrated to the target dst_cpu taking
993 * into account that it might be best if task running on the dst_cpu should
994 * be exchanged with the source task
996 static void task_numa_compare(struct task_numa_env
*env
, long imp
)
998 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
999 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1000 struct task_struct
*cur
;
1001 long dst_load
, src_load
;
1005 cur
= ACCESS_ONCE(dst_rq
->curr
);
1006 if (cur
->pid
== 0) /* idle */
1010 * "imp" is the fault differential for the source task between the
1011 * source and destination node. Calculate the total differential for
1012 * the source task and potential destination task. The more negative
1013 * the value is, the more rmeote accesses that would be expected to
1014 * be incurred if the tasks were swapped.
1017 /* Skip this swap candidate if cannot move to the source cpu */
1018 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1021 imp
+= task_faults(cur
, env
->src_nid
) -
1022 task_faults(cur
, env
->dst_nid
);
1025 if (imp
< env
->best_imp
)
1029 /* Is there capacity at our destination? */
1030 if (env
->src_stats
.has_capacity
&&
1031 !env
->dst_stats
.has_capacity
)
1037 /* Balance doesn't matter much if we're running a task per cpu */
1038 if (src_rq
->nr_running
== 1 && dst_rq
->nr_running
== 1)
1042 * In the overloaded case, try and keep the load balanced.
1045 dst_load
= env
->dst_stats
.load
;
1046 src_load
= env
->src_stats
.load
;
1048 /* XXX missing power terms */
1049 load
= task_h_load(env
->p
);
1054 load
= task_h_load(cur
);
1059 /* make src_load the smaller */
1060 if (dst_load
< src_load
)
1061 swap(dst_load
, src_load
);
1063 if (src_load
* env
->imbalance_pct
< dst_load
* 100)
1067 task_numa_assign(env
, cur
, imp
);
1072 static void task_numa_find_cpu(struct task_numa_env
*env
, long imp
)
1076 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1077 /* Skip this CPU if the source task cannot migrate */
1078 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1082 task_numa_compare(env
, imp
);
1086 static int task_numa_migrate(struct task_struct
*p
)
1088 struct task_numa_env env
= {
1091 .src_cpu
= task_cpu(p
),
1092 .src_nid
= cpu_to_node(task_cpu(p
)),
1094 .imbalance_pct
= 112,
1100 struct sched_domain
*sd
;
1101 unsigned long faults
;
1106 * Pick the lowest SD_NUMA domain, as that would have the smallest
1107 * imbalance and would be the first to start moving tasks about.
1109 * And we want to avoid any moving of tasks about, as that would create
1110 * random movement of tasks -- counter the numa conditions we're trying
1114 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1115 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1118 faults
= task_faults(p
, env
.src_nid
);
1119 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1120 env
.dst_nid
= p
->numa_preferred_nid
;
1121 imp
= task_faults(env
.p
, env
.dst_nid
) - faults
;
1122 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1124 /* If the preferred nid has capacity, try to use it. */
1125 if (env
.dst_stats
.has_capacity
)
1126 task_numa_find_cpu(&env
, imp
);
1128 /* No space available on the preferred nid. Look elsewhere. */
1129 if (env
.best_cpu
== -1) {
1130 for_each_online_node(nid
) {
1131 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1134 /* Only consider nodes that recorded more faults */
1135 imp
= task_faults(env
.p
, nid
) - faults
;
1140 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1141 task_numa_find_cpu(&env
, imp
);
1145 /* No better CPU than the current one was found. */
1146 if (env
.best_cpu
== -1)
1149 if (env
.best_task
== NULL
) {
1150 int ret
= migrate_task_to(p
, env
.best_cpu
);
1154 ret
= migrate_swap(p
, env
.best_task
);
1155 put_task_struct(env
.best_task
);
1159 /* Attempt to migrate a task to a CPU on the preferred node. */
1160 static void numa_migrate_preferred(struct task_struct
*p
)
1162 /* Success if task is already running on preferred CPU */
1163 p
->numa_migrate_retry
= 0;
1164 if (cpu_to_node(task_cpu(p
)) == p
->numa_preferred_nid
) {
1166 * If migration is temporarily disabled due to a task migration
1167 * then re-enable it now as the task is running on its
1168 * preferred node and memory should migrate locally
1170 if (!p
->numa_migrate_seq
)
1171 p
->numa_migrate_seq
++;
1175 /* This task has no NUMA fault statistics yet */
1176 if (unlikely(p
->numa_preferred_nid
== -1))
1179 /* Otherwise, try migrate to a CPU on the preferred node */
1180 if (task_numa_migrate(p
) != 0)
1181 p
->numa_migrate_retry
= jiffies
+ HZ
*5;
1184 static void task_numa_placement(struct task_struct
*p
)
1186 int seq
, nid
, max_nid
= -1;
1187 unsigned long max_faults
= 0;
1189 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1190 if (p
->numa_scan_seq
== seq
)
1192 p
->numa_scan_seq
= seq
;
1193 p
->numa_migrate_seq
++;
1194 p
->numa_scan_period_max
= task_scan_max(p
);
1196 /* Find the node with the highest number of faults */
1197 for_each_online_node(nid
) {
1198 unsigned long faults
= 0;
1201 for (priv
= 0; priv
< 2; priv
++) {
1204 i
= task_faults_idx(nid
, priv
);
1205 diff
= -p
->numa_faults
[i
];
1207 /* Decay existing window, copy faults since last scan */
1208 p
->numa_faults
[i
] >>= 1;
1209 p
->numa_faults
[i
] += p
->numa_faults_buffer
[i
];
1210 p
->numa_faults_buffer
[i
] = 0;
1212 faults
+= p
->numa_faults
[i
];
1213 diff
+= p
->numa_faults
[i
];
1214 if (p
->numa_group
) {
1215 /* safe because we can only change our own group */
1216 atomic_long_add(diff
, &p
->numa_group
->faults
[i
]);
1220 if (faults
> max_faults
) {
1221 max_faults
= faults
;
1226 /* Preferred node as the node with the most faults */
1227 if (max_faults
&& max_nid
!= p
->numa_preferred_nid
) {
1228 /* Update the preferred nid and migrate task if possible */
1229 p
->numa_preferred_nid
= max_nid
;
1230 p
->numa_migrate_seq
= 1;
1231 numa_migrate_preferred(p
);
1235 static inline int get_numa_group(struct numa_group
*grp
)
1237 return atomic_inc_not_zero(&grp
->refcount
);
1240 static inline void put_numa_group(struct numa_group
*grp
)
1242 if (atomic_dec_and_test(&grp
->refcount
))
1243 kfree_rcu(grp
, rcu
);
1246 static void double_lock(spinlock_t
*l1
, spinlock_t
*l2
)
1252 spin_lock_nested(l2
, SINGLE_DEPTH_NESTING
);
1255 static void task_numa_group(struct task_struct
*p
, int cpupid
)
1257 struct numa_group
*grp
, *my_grp
;
1258 struct task_struct
*tsk
;
1260 int cpu
= cpupid_to_cpu(cpupid
);
1263 if (unlikely(!p
->numa_group
)) {
1264 unsigned int size
= sizeof(struct numa_group
) +
1265 2*nr_node_ids
*sizeof(atomic_long_t
);
1267 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1271 atomic_set(&grp
->refcount
, 1);
1272 spin_lock_init(&grp
->lock
);
1273 INIT_LIST_HEAD(&grp
->task_list
);
1276 for (i
= 0; i
< 2*nr_node_ids
; i
++)
1277 atomic_long_set(&grp
->faults
[i
], p
->numa_faults
[i
]);
1279 list_add(&p
->numa_entry
, &grp
->task_list
);
1281 rcu_assign_pointer(p
->numa_group
, grp
);
1285 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1287 if (!cpupid_match_pid(tsk
, cpupid
))
1290 grp
= rcu_dereference(tsk
->numa_group
);
1294 my_grp
= p
->numa_group
;
1299 * Only join the other group if its bigger; if we're the bigger group,
1300 * the other task will join us.
1302 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1306 * Tie-break on the grp address.
1308 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1311 if (!get_numa_group(grp
))
1322 for (i
= 0; i
< 2*nr_node_ids
; i
++) {
1323 atomic_long_sub(p
->numa_faults
[i
], &my_grp
->faults
[i
]);
1324 atomic_long_add(p
->numa_faults
[i
], &grp
->faults
[i
]);
1327 double_lock(&my_grp
->lock
, &grp
->lock
);
1329 list_move(&p
->numa_entry
, &grp
->task_list
);
1333 spin_unlock(&my_grp
->lock
);
1334 spin_unlock(&grp
->lock
);
1336 rcu_assign_pointer(p
->numa_group
, grp
);
1338 put_numa_group(my_grp
);
1341 void task_numa_free(struct task_struct
*p
)
1343 struct numa_group
*grp
= p
->numa_group
;
1347 for (i
= 0; i
< 2*nr_node_ids
; i
++)
1348 atomic_long_sub(p
->numa_faults
[i
], &grp
->faults
[i
]);
1350 spin_lock(&grp
->lock
);
1351 list_del(&p
->numa_entry
);
1353 spin_unlock(&grp
->lock
);
1354 rcu_assign_pointer(p
->numa_group
, NULL
);
1355 put_numa_group(grp
);
1358 kfree(p
->numa_faults
);
1362 * Got a PROT_NONE fault for a page on @node.
1364 void task_numa_fault(int last_cpupid
, int node
, int pages
, int flags
)
1366 struct task_struct
*p
= current
;
1367 bool migrated
= flags
& TNF_MIGRATED
;
1370 if (!numabalancing_enabled
)
1373 /* for example, ksmd faulting in a user's mm */
1377 /* Allocate buffer to track faults on a per-node basis */
1378 if (unlikely(!p
->numa_faults
)) {
1379 int size
= sizeof(*p
->numa_faults
) * 2 * nr_node_ids
;
1381 /* numa_faults and numa_faults_buffer share the allocation */
1382 p
->numa_faults
= kzalloc(size
* 2, GFP_KERNEL
|__GFP_NOWARN
);
1383 if (!p
->numa_faults
)
1386 BUG_ON(p
->numa_faults_buffer
);
1387 p
->numa_faults_buffer
= p
->numa_faults
+ (2 * nr_node_ids
);
1391 * First accesses are treated as private, otherwise consider accesses
1392 * to be private if the accessing pid has not changed
1394 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
1397 priv
= cpupid_match_pid(p
, last_cpupid
);
1398 if (!priv
&& !(flags
& TNF_NO_GROUP
))
1399 task_numa_group(p
, last_cpupid
);
1403 * If pages are properly placed (did not migrate) then scan slower.
1404 * This is reset periodically in case of phase changes
1407 /* Initialise if necessary */
1408 if (!p
->numa_scan_period_max
)
1409 p
->numa_scan_period_max
= task_scan_max(p
);
1411 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1412 p
->numa_scan_period
+ 10);
1415 task_numa_placement(p
);
1417 /* Retry task to preferred node migration if it previously failed */
1418 if (p
->numa_migrate_retry
&& time_after(jiffies
, p
->numa_migrate_retry
))
1419 numa_migrate_preferred(p
);
1421 p
->numa_faults_buffer
[task_faults_idx(node
, priv
)] += pages
;
1424 static void reset_ptenuma_scan(struct task_struct
*p
)
1426 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
1427 p
->mm
->numa_scan_offset
= 0;
1431 * The expensive part of numa migration is done from task_work context.
1432 * Triggered from task_tick_numa().
1434 void task_numa_work(struct callback_head
*work
)
1436 unsigned long migrate
, next_scan
, now
= jiffies
;
1437 struct task_struct
*p
= current
;
1438 struct mm_struct
*mm
= p
->mm
;
1439 struct vm_area_struct
*vma
;
1440 unsigned long start
, end
;
1441 unsigned long nr_pte_updates
= 0;
1444 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
1446 work
->next
= work
; /* protect against double add */
1448 * Who cares about NUMA placement when they're dying.
1450 * NOTE: make sure not to dereference p->mm before this check,
1451 * exit_task_work() happens _after_ exit_mm() so we could be called
1452 * without p->mm even though we still had it when we enqueued this
1455 if (p
->flags
& PF_EXITING
)
1458 if (!mm
->numa_next_reset
|| !mm
->numa_next_scan
) {
1459 mm
->numa_next_scan
= now
+
1460 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1461 mm
->numa_next_reset
= now
+
1462 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset
);
1466 * Reset the scan period if enough time has gone by. Objective is that
1467 * scanning will be reduced if pages are properly placed. As tasks
1468 * can enter different phases this needs to be re-examined. Lacking
1469 * proper tracking of reference behaviour, this blunt hammer is used.
1471 migrate
= mm
->numa_next_reset
;
1472 if (time_after(now
, migrate
)) {
1473 p
->numa_scan_period
= task_scan_min(p
);
1474 next_scan
= now
+ msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset
);
1475 xchg(&mm
->numa_next_reset
, next_scan
);
1479 * Enforce maximal scan/migration frequency..
1481 migrate
= mm
->numa_next_scan
;
1482 if (time_before(now
, migrate
))
1485 if (p
->numa_scan_period
== 0) {
1486 p
->numa_scan_period_max
= task_scan_max(p
);
1487 p
->numa_scan_period
= task_scan_min(p
);
1490 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1491 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
1495 * Delay this task enough that another task of this mm will likely win
1496 * the next time around.
1498 p
->node_stamp
+= 2 * TICK_NSEC
;
1500 start
= mm
->numa_scan_offset
;
1501 pages
= sysctl_numa_balancing_scan_size
;
1502 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
1506 down_read(&mm
->mmap_sem
);
1507 vma
= find_vma(mm
, start
);
1509 reset_ptenuma_scan(p
);
1513 for (; vma
; vma
= vma
->vm_next
) {
1514 if (!vma_migratable(vma
) || !vma_policy_mof(p
, vma
))
1518 * Shared library pages mapped by multiple processes are not
1519 * migrated as it is expected they are cache replicated. Avoid
1520 * hinting faults in read-only file-backed mappings or the vdso
1521 * as migrating the pages will be of marginal benefit.
1524 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
1528 start
= max(start
, vma
->vm_start
);
1529 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
1530 end
= min(end
, vma
->vm_end
);
1531 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
1534 * Scan sysctl_numa_balancing_scan_size but ensure that
1535 * at least one PTE is updated so that unused virtual
1536 * address space is quickly skipped.
1539 pages
-= (end
- start
) >> PAGE_SHIFT
;
1544 } while (end
!= vma
->vm_end
);
1549 * If the whole process was scanned without updates then no NUMA
1550 * hinting faults are being recorded and scan rate should be lower.
1552 if (mm
->numa_scan_offset
== 0 && !nr_pte_updates
) {
1553 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1554 p
->numa_scan_period
<< 1);
1556 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1557 mm
->numa_next_scan
= next_scan
;
1561 * It is possible to reach the end of the VMA list but the last few
1562 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1563 * would find the !migratable VMA on the next scan but not reset the
1564 * scanner to the start so check it now.
1567 mm
->numa_scan_offset
= start
;
1569 reset_ptenuma_scan(p
);
1570 up_read(&mm
->mmap_sem
);
1574 * Drive the periodic memory faults..
1576 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1578 struct callback_head
*work
= &curr
->numa_work
;
1582 * We don't care about NUMA placement if we don't have memory.
1584 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
1588 * Using runtime rather than walltime has the dual advantage that
1589 * we (mostly) drive the selection from busy threads and that the
1590 * task needs to have done some actual work before we bother with
1593 now
= curr
->se
.sum_exec_runtime
;
1594 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
1596 if (now
- curr
->node_stamp
> period
) {
1597 if (!curr
->node_stamp
)
1598 curr
->numa_scan_period
= task_scan_min(curr
);
1599 curr
->node_stamp
+= period
;
1601 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
1602 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
1603 task_work_add(curr
, work
, true);
1608 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1611 #endif /* CONFIG_NUMA_BALANCING */
1614 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1616 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
1617 if (!parent_entity(se
))
1618 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1620 if (entity_is_task(se
))
1621 list_add(&se
->group_node
, &rq_of(cfs_rq
)->cfs_tasks
);
1623 cfs_rq
->nr_running
++;
1627 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1629 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
1630 if (!parent_entity(se
))
1631 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1632 if (entity_is_task(se
))
1633 list_del_init(&se
->group_node
);
1634 cfs_rq
->nr_running
--;
1637 #ifdef CONFIG_FAIR_GROUP_SCHED
1639 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
1644 * Use this CPU's actual weight instead of the last load_contribution
1645 * to gain a more accurate current total weight. See
1646 * update_cfs_rq_load_contribution().
1648 tg_weight
= atomic_long_read(&tg
->load_avg
);
1649 tg_weight
-= cfs_rq
->tg_load_contrib
;
1650 tg_weight
+= cfs_rq
->load
.weight
;
1655 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1657 long tg_weight
, load
, shares
;
1659 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
1660 load
= cfs_rq
->load
.weight
;
1662 shares
= (tg
->shares
* load
);
1664 shares
/= tg_weight
;
1666 if (shares
< MIN_SHARES
)
1667 shares
= MIN_SHARES
;
1668 if (shares
> tg
->shares
)
1669 shares
= tg
->shares
;
1673 # else /* CONFIG_SMP */
1674 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1678 # endif /* CONFIG_SMP */
1679 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
1680 unsigned long weight
)
1683 /* commit outstanding execution time */
1684 if (cfs_rq
->curr
== se
)
1685 update_curr(cfs_rq
);
1686 account_entity_dequeue(cfs_rq
, se
);
1689 update_load_set(&se
->load
, weight
);
1692 account_entity_enqueue(cfs_rq
, se
);
1695 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
1697 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
1699 struct task_group
*tg
;
1700 struct sched_entity
*se
;
1704 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
1705 if (!se
|| throttled_hierarchy(cfs_rq
))
1708 if (likely(se
->load
.weight
== tg
->shares
))
1711 shares
= calc_cfs_shares(cfs_rq
, tg
);
1713 reweight_entity(cfs_rq_of(se
), se
, shares
);
1715 #else /* CONFIG_FAIR_GROUP_SCHED */
1716 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
1719 #endif /* CONFIG_FAIR_GROUP_SCHED */
1723 * We choose a half-life close to 1 scheduling period.
1724 * Note: The tables below are dependent on this value.
1726 #define LOAD_AVG_PERIOD 32
1727 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1728 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1730 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1731 static const u32 runnable_avg_yN_inv
[] = {
1732 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1733 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1734 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1735 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1736 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1737 0x85aac367, 0x82cd8698,
1741 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1742 * over-estimates when re-combining.
1744 static const u32 runnable_avg_yN_sum
[] = {
1745 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1746 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1747 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1752 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1754 static __always_inline u64
decay_load(u64 val
, u64 n
)
1756 unsigned int local_n
;
1760 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
1763 /* after bounds checking we can collapse to 32-bit */
1767 * As y^PERIOD = 1/2, we can combine
1768 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1769 * With a look-up table which covers k^n (n<PERIOD)
1771 * To achieve constant time decay_load.
1773 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
1774 val
>>= local_n
/ LOAD_AVG_PERIOD
;
1775 local_n
%= LOAD_AVG_PERIOD
;
1778 val
*= runnable_avg_yN_inv
[local_n
];
1779 /* We don't use SRR here since we always want to round down. */
1784 * For updates fully spanning n periods, the contribution to runnable
1785 * average will be: \Sum 1024*y^n
1787 * We can compute this reasonably efficiently by combining:
1788 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1790 static u32
__compute_runnable_contrib(u64 n
)
1794 if (likely(n
<= LOAD_AVG_PERIOD
))
1795 return runnable_avg_yN_sum
[n
];
1796 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
1797 return LOAD_AVG_MAX
;
1799 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1801 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1802 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
1804 n
-= LOAD_AVG_PERIOD
;
1805 } while (n
> LOAD_AVG_PERIOD
);
1807 contrib
= decay_load(contrib
, n
);
1808 return contrib
+ runnable_avg_yN_sum
[n
];
1812 * We can represent the historical contribution to runnable average as the
1813 * coefficients of a geometric series. To do this we sub-divide our runnable
1814 * history into segments of approximately 1ms (1024us); label the segment that
1815 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1817 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1819 * (now) (~1ms ago) (~2ms ago)
1821 * Let u_i denote the fraction of p_i that the entity was runnable.
1823 * We then designate the fractions u_i as our co-efficients, yielding the
1824 * following representation of historical load:
1825 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1827 * We choose y based on the with of a reasonably scheduling period, fixing:
1830 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1831 * approximately half as much as the contribution to load within the last ms
1834 * When a period "rolls over" and we have new u_0`, multiplying the previous
1835 * sum again by y is sufficient to update:
1836 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1837 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1839 static __always_inline
int __update_entity_runnable_avg(u64 now
,
1840 struct sched_avg
*sa
,
1844 u32 runnable_contrib
;
1845 int delta_w
, decayed
= 0;
1847 delta
= now
- sa
->last_runnable_update
;
1849 * This should only happen when time goes backwards, which it
1850 * unfortunately does during sched clock init when we swap over to TSC.
1852 if ((s64
)delta
< 0) {
1853 sa
->last_runnable_update
= now
;
1858 * Use 1024ns as the unit of measurement since it's a reasonable
1859 * approximation of 1us and fast to compute.
1864 sa
->last_runnable_update
= now
;
1866 /* delta_w is the amount already accumulated against our next period */
1867 delta_w
= sa
->runnable_avg_period
% 1024;
1868 if (delta
+ delta_w
>= 1024) {
1869 /* period roll-over */
1873 * Now that we know we're crossing a period boundary, figure
1874 * out how much from delta we need to complete the current
1875 * period and accrue it.
1877 delta_w
= 1024 - delta_w
;
1879 sa
->runnable_avg_sum
+= delta_w
;
1880 sa
->runnable_avg_period
+= delta_w
;
1884 /* Figure out how many additional periods this update spans */
1885 periods
= delta
/ 1024;
1888 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
1890 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
1893 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1894 runnable_contrib
= __compute_runnable_contrib(periods
);
1896 sa
->runnable_avg_sum
+= runnable_contrib
;
1897 sa
->runnable_avg_period
+= runnable_contrib
;
1900 /* Remainder of delta accrued against u_0` */
1902 sa
->runnable_avg_sum
+= delta
;
1903 sa
->runnable_avg_period
+= delta
;
1908 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1909 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
1911 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1912 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
1914 decays
-= se
->avg
.decay_count
;
1918 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
1919 se
->avg
.decay_count
= 0;
1924 #ifdef CONFIG_FAIR_GROUP_SCHED
1925 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
1928 struct task_group
*tg
= cfs_rq
->tg
;
1931 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
1932 tg_contrib
-= cfs_rq
->tg_load_contrib
;
1934 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
1935 atomic_long_add(tg_contrib
, &tg
->load_avg
);
1936 cfs_rq
->tg_load_contrib
+= tg_contrib
;
1941 * Aggregate cfs_rq runnable averages into an equivalent task_group
1942 * representation for computing load contributions.
1944 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
1945 struct cfs_rq
*cfs_rq
)
1947 struct task_group
*tg
= cfs_rq
->tg
;
1950 /* The fraction of a cpu used by this cfs_rq */
1951 contrib
= div_u64(sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
1952 sa
->runnable_avg_period
+ 1);
1953 contrib
-= cfs_rq
->tg_runnable_contrib
;
1955 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
1956 atomic_add(contrib
, &tg
->runnable_avg
);
1957 cfs_rq
->tg_runnable_contrib
+= contrib
;
1961 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
1963 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
1964 struct task_group
*tg
= cfs_rq
->tg
;
1969 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
1970 se
->avg
.load_avg_contrib
= div_u64(contrib
,
1971 atomic_long_read(&tg
->load_avg
) + 1);
1974 * For group entities we need to compute a correction term in the case
1975 * that they are consuming <1 cpu so that we would contribute the same
1976 * load as a task of equal weight.
1978 * Explicitly co-ordinating this measurement would be expensive, but
1979 * fortunately the sum of each cpus contribution forms a usable
1980 * lower-bound on the true value.
1982 * Consider the aggregate of 2 contributions. Either they are disjoint
1983 * (and the sum represents true value) or they are disjoint and we are
1984 * understating by the aggregate of their overlap.
1986 * Extending this to N cpus, for a given overlap, the maximum amount we
1987 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1988 * cpus that overlap for this interval and w_i is the interval width.
1990 * On a small machine; the first term is well-bounded which bounds the
1991 * total error since w_i is a subset of the period. Whereas on a
1992 * larger machine, while this first term can be larger, if w_i is the
1993 * of consequential size guaranteed to see n_i*w_i quickly converge to
1994 * our upper bound of 1-cpu.
1996 runnable_avg
= atomic_read(&tg
->runnable_avg
);
1997 if (runnable_avg
< NICE_0_LOAD
) {
1998 se
->avg
.load_avg_contrib
*= runnable_avg
;
1999 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2003 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2004 int force_update
) {}
2005 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2006 struct cfs_rq
*cfs_rq
) {}
2007 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2010 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2014 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2015 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2016 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2017 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2020 /* Compute the current contribution to load_avg by se, return any delta */
2021 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2023 long old_contrib
= se
->avg
.load_avg_contrib
;
2025 if (entity_is_task(se
)) {
2026 __update_task_entity_contrib(se
);
2028 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2029 __update_group_entity_contrib(se
);
2032 return se
->avg
.load_avg_contrib
- old_contrib
;
2035 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2038 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2039 cfs_rq
->blocked_load_avg
-= load_contrib
;
2041 cfs_rq
->blocked_load_avg
= 0;
2044 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2046 /* Update a sched_entity's runnable average */
2047 static inline void update_entity_load_avg(struct sched_entity
*se
,
2050 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2055 * For a group entity we need to use their owned cfs_rq_clock_task() in
2056 * case they are the parent of a throttled hierarchy.
2058 if (entity_is_task(se
))
2059 now
= cfs_rq_clock_task(cfs_rq
);
2061 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2063 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2066 contrib_delta
= __update_entity_load_avg_contrib(se
);
2072 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2074 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2078 * Decay the load contributed by all blocked children and account this so that
2079 * their contribution may appropriately discounted when they wake up.
2081 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2083 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2086 decays
= now
- cfs_rq
->last_decay
;
2087 if (!decays
&& !force_update
)
2090 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2091 unsigned long removed_load
;
2092 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2093 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2097 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2099 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2100 cfs_rq
->last_decay
= now
;
2103 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2106 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2108 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2109 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2112 /* Add the load generated by se into cfs_rq's child load-average */
2113 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2114 struct sched_entity
*se
,
2118 * We track migrations using entity decay_count <= 0, on a wake-up
2119 * migration we use a negative decay count to track the remote decays
2120 * accumulated while sleeping.
2122 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2123 * are seen by enqueue_entity_load_avg() as a migration with an already
2124 * constructed load_avg_contrib.
2126 if (unlikely(se
->avg
.decay_count
<= 0)) {
2127 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2128 if (se
->avg
.decay_count
) {
2130 * In a wake-up migration we have to approximate the
2131 * time sleeping. This is because we can't synchronize
2132 * clock_task between the two cpus, and it is not
2133 * guaranteed to be read-safe. Instead, we can
2134 * approximate this using our carried decays, which are
2135 * explicitly atomically readable.
2137 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2139 update_entity_load_avg(se
, 0);
2140 /* Indicate that we're now synchronized and on-rq */
2141 se
->avg
.decay_count
= 0;
2146 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2147 * would have made count negative); we must be careful to avoid
2148 * double-accounting blocked time after synchronizing decays.
2150 se
->avg
.last_runnable_update
+= __synchronize_entity_decay(se
)
2154 /* migrated tasks did not contribute to our blocked load */
2156 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2157 update_entity_load_avg(se
, 0);
2160 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2161 /* we force update consideration on load-balancer moves */
2162 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2166 * Remove se's load from this cfs_rq child load-average, if the entity is
2167 * transitioning to a blocked state we track its projected decay using
2170 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2171 struct sched_entity
*se
,
2174 update_entity_load_avg(se
, 1);
2175 /* we force update consideration on load-balancer moves */
2176 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2178 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2180 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2181 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2182 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2186 * Update the rq's load with the elapsed running time before entering
2187 * idle. if the last scheduled task is not a CFS task, idle_enter will
2188 * be the only way to update the runnable statistic.
2190 void idle_enter_fair(struct rq
*this_rq
)
2192 update_rq_runnable_avg(this_rq
, 1);
2196 * Update the rq's load with the elapsed idle time before a task is
2197 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2198 * be the only way to update the runnable statistic.
2200 void idle_exit_fair(struct rq
*this_rq
)
2202 update_rq_runnable_avg(this_rq
, 0);
2206 static inline void update_entity_load_avg(struct sched_entity
*se
,
2207 int update_cfs_rq
) {}
2208 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2209 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2210 struct sched_entity
*se
,
2212 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2213 struct sched_entity
*se
,
2215 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2216 int force_update
) {}
2219 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2221 #ifdef CONFIG_SCHEDSTATS
2222 struct task_struct
*tsk
= NULL
;
2224 if (entity_is_task(se
))
2227 if (se
->statistics
.sleep_start
) {
2228 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2233 if (unlikely(delta
> se
->statistics
.sleep_max
))
2234 se
->statistics
.sleep_max
= delta
;
2236 se
->statistics
.sleep_start
= 0;
2237 se
->statistics
.sum_sleep_runtime
+= delta
;
2240 account_scheduler_latency(tsk
, delta
>> 10, 1);
2241 trace_sched_stat_sleep(tsk
, delta
);
2244 if (se
->statistics
.block_start
) {
2245 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2250 if (unlikely(delta
> se
->statistics
.block_max
))
2251 se
->statistics
.block_max
= delta
;
2253 se
->statistics
.block_start
= 0;
2254 se
->statistics
.sum_sleep_runtime
+= delta
;
2257 if (tsk
->in_iowait
) {
2258 se
->statistics
.iowait_sum
+= delta
;
2259 se
->statistics
.iowait_count
++;
2260 trace_sched_stat_iowait(tsk
, delta
);
2263 trace_sched_stat_blocked(tsk
, delta
);
2266 * Blocking time is in units of nanosecs, so shift by
2267 * 20 to get a milliseconds-range estimation of the
2268 * amount of time that the task spent sleeping:
2270 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2271 profile_hits(SLEEP_PROFILING
,
2272 (void *)get_wchan(tsk
),
2275 account_scheduler_latency(tsk
, delta
>> 10, 0);
2281 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2283 #ifdef CONFIG_SCHED_DEBUG
2284 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2289 if (d
> 3*sysctl_sched_latency
)
2290 schedstat_inc(cfs_rq
, nr_spread_over
);
2295 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2297 u64 vruntime
= cfs_rq
->min_vruntime
;
2300 * The 'current' period is already promised to the current tasks,
2301 * however the extra weight of the new task will slow them down a
2302 * little, place the new task so that it fits in the slot that
2303 * stays open at the end.
2305 if (initial
&& sched_feat(START_DEBIT
))
2306 vruntime
+= sched_vslice(cfs_rq
, se
);
2308 /* sleeps up to a single latency don't count. */
2310 unsigned long thresh
= sysctl_sched_latency
;
2313 * Halve their sleep time's effect, to allow
2314 * for a gentler effect of sleepers:
2316 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2322 /* ensure we never gain time by being placed backwards. */
2323 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
2326 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
2329 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2332 * Update the normalized vruntime before updating min_vruntime
2333 * through calling update_curr().
2335 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
2336 se
->vruntime
+= cfs_rq
->min_vruntime
;
2339 * Update run-time statistics of the 'current'.
2341 update_curr(cfs_rq
);
2342 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
2343 account_entity_enqueue(cfs_rq
, se
);
2344 update_cfs_shares(cfs_rq
);
2346 if (flags
& ENQUEUE_WAKEUP
) {
2347 place_entity(cfs_rq
, se
, 0);
2348 enqueue_sleeper(cfs_rq
, se
);
2351 update_stats_enqueue(cfs_rq
, se
);
2352 check_spread(cfs_rq
, se
);
2353 if (se
!= cfs_rq
->curr
)
2354 __enqueue_entity(cfs_rq
, se
);
2357 if (cfs_rq
->nr_running
== 1) {
2358 list_add_leaf_cfs_rq(cfs_rq
);
2359 check_enqueue_throttle(cfs_rq
);
2363 static void __clear_buddies_last(struct sched_entity
*se
)
2365 for_each_sched_entity(se
) {
2366 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2367 if (cfs_rq
->last
== se
)
2368 cfs_rq
->last
= NULL
;
2374 static void __clear_buddies_next(struct sched_entity
*se
)
2376 for_each_sched_entity(se
) {
2377 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2378 if (cfs_rq
->next
== se
)
2379 cfs_rq
->next
= NULL
;
2385 static void __clear_buddies_skip(struct sched_entity
*se
)
2387 for_each_sched_entity(se
) {
2388 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2389 if (cfs_rq
->skip
== se
)
2390 cfs_rq
->skip
= NULL
;
2396 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2398 if (cfs_rq
->last
== se
)
2399 __clear_buddies_last(se
);
2401 if (cfs_rq
->next
== se
)
2402 __clear_buddies_next(se
);
2404 if (cfs_rq
->skip
== se
)
2405 __clear_buddies_skip(se
);
2408 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2411 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2414 * Update run-time statistics of the 'current'.
2416 update_curr(cfs_rq
);
2417 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
2419 update_stats_dequeue(cfs_rq
, se
);
2420 if (flags
& DEQUEUE_SLEEP
) {
2421 #ifdef CONFIG_SCHEDSTATS
2422 if (entity_is_task(se
)) {
2423 struct task_struct
*tsk
= task_of(se
);
2425 if (tsk
->state
& TASK_INTERRUPTIBLE
)
2426 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
2427 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
2428 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
2433 clear_buddies(cfs_rq
, se
);
2435 if (se
!= cfs_rq
->curr
)
2436 __dequeue_entity(cfs_rq
, se
);
2438 account_entity_dequeue(cfs_rq
, se
);
2441 * Normalize the entity after updating the min_vruntime because the
2442 * update can refer to the ->curr item and we need to reflect this
2443 * movement in our normalized position.
2445 if (!(flags
& DEQUEUE_SLEEP
))
2446 se
->vruntime
-= cfs_rq
->min_vruntime
;
2448 /* return excess runtime on last dequeue */
2449 return_cfs_rq_runtime(cfs_rq
);
2451 update_min_vruntime(cfs_rq
);
2452 update_cfs_shares(cfs_rq
);
2456 * Preempt the current task with a newly woken task if needed:
2459 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2461 unsigned long ideal_runtime
, delta_exec
;
2462 struct sched_entity
*se
;
2465 ideal_runtime
= sched_slice(cfs_rq
, curr
);
2466 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
2467 if (delta_exec
> ideal_runtime
) {
2468 resched_task(rq_of(cfs_rq
)->curr
);
2470 * The current task ran long enough, ensure it doesn't get
2471 * re-elected due to buddy favours.
2473 clear_buddies(cfs_rq
, curr
);
2478 * Ensure that a task that missed wakeup preemption by a
2479 * narrow margin doesn't have to wait for a full slice.
2480 * This also mitigates buddy induced latencies under load.
2482 if (delta_exec
< sysctl_sched_min_granularity
)
2485 se
= __pick_first_entity(cfs_rq
);
2486 delta
= curr
->vruntime
- se
->vruntime
;
2491 if (delta
> ideal_runtime
)
2492 resched_task(rq_of(cfs_rq
)->curr
);
2496 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2498 /* 'current' is not kept within the tree. */
2501 * Any task has to be enqueued before it get to execute on
2502 * a CPU. So account for the time it spent waiting on the
2505 update_stats_wait_end(cfs_rq
, se
);
2506 __dequeue_entity(cfs_rq
, se
);
2509 update_stats_curr_start(cfs_rq
, se
);
2511 #ifdef CONFIG_SCHEDSTATS
2513 * Track our maximum slice length, if the CPU's load is at
2514 * least twice that of our own weight (i.e. dont track it
2515 * when there are only lesser-weight tasks around):
2517 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
2518 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
2519 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
2522 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
2526 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
2529 * Pick the next process, keeping these things in mind, in this order:
2530 * 1) keep things fair between processes/task groups
2531 * 2) pick the "next" process, since someone really wants that to run
2532 * 3) pick the "last" process, for cache locality
2533 * 4) do not run the "skip" process, if something else is available
2535 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
2537 struct sched_entity
*se
= __pick_first_entity(cfs_rq
);
2538 struct sched_entity
*left
= se
;
2541 * Avoid running the skip buddy, if running something else can
2542 * be done without getting too unfair.
2544 if (cfs_rq
->skip
== se
) {
2545 struct sched_entity
*second
= __pick_next_entity(se
);
2546 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
2551 * Prefer last buddy, try to return the CPU to a preempted task.
2553 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
2557 * Someone really wants this to run. If it's not unfair, run it.
2559 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
2562 clear_buddies(cfs_rq
, se
);
2567 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2569 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
2572 * If still on the runqueue then deactivate_task()
2573 * was not called and update_curr() has to be done:
2576 update_curr(cfs_rq
);
2578 /* throttle cfs_rqs exceeding runtime */
2579 check_cfs_rq_runtime(cfs_rq
);
2581 check_spread(cfs_rq
, prev
);
2583 update_stats_wait_start(cfs_rq
, prev
);
2584 /* Put 'current' back into the tree. */
2585 __enqueue_entity(cfs_rq
, prev
);
2586 /* in !on_rq case, update occurred at dequeue */
2587 update_entity_load_avg(prev
, 1);
2589 cfs_rq
->curr
= NULL
;
2593 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
2596 * Update run-time statistics of the 'current'.
2598 update_curr(cfs_rq
);
2601 * Ensure that runnable average is periodically updated.
2603 update_entity_load_avg(curr
, 1);
2604 update_cfs_rq_blocked_load(cfs_rq
, 1);
2605 update_cfs_shares(cfs_rq
);
2607 #ifdef CONFIG_SCHED_HRTICK
2609 * queued ticks are scheduled to match the slice, so don't bother
2610 * validating it and just reschedule.
2613 resched_task(rq_of(cfs_rq
)->curr
);
2617 * don't let the period tick interfere with the hrtick preemption
2619 if (!sched_feat(DOUBLE_TICK
) &&
2620 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
2624 if (cfs_rq
->nr_running
> 1)
2625 check_preempt_tick(cfs_rq
, curr
);
2629 /**************************************************
2630 * CFS bandwidth control machinery
2633 #ifdef CONFIG_CFS_BANDWIDTH
2635 #ifdef HAVE_JUMP_LABEL
2636 static struct static_key __cfs_bandwidth_used
;
2638 static inline bool cfs_bandwidth_used(void)
2640 return static_key_false(&__cfs_bandwidth_used
);
2643 void account_cfs_bandwidth_used(int enabled
, int was_enabled
)
2645 /* only need to count groups transitioning between enabled/!enabled */
2646 if (enabled
&& !was_enabled
)
2647 static_key_slow_inc(&__cfs_bandwidth_used
);
2648 else if (!enabled
&& was_enabled
)
2649 static_key_slow_dec(&__cfs_bandwidth_used
);
2651 #else /* HAVE_JUMP_LABEL */
2652 static bool cfs_bandwidth_used(void)
2657 void account_cfs_bandwidth_used(int enabled
, int was_enabled
) {}
2658 #endif /* HAVE_JUMP_LABEL */
2661 * default period for cfs group bandwidth.
2662 * default: 0.1s, units: nanoseconds
2664 static inline u64
default_cfs_period(void)
2666 return 100000000ULL;
2669 static inline u64
sched_cfs_bandwidth_slice(void)
2671 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
2675 * Replenish runtime according to assigned quota and update expiration time.
2676 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2677 * additional synchronization around rq->lock.
2679 * requires cfs_b->lock
2681 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
2685 if (cfs_b
->quota
== RUNTIME_INF
)
2688 now
= sched_clock_cpu(smp_processor_id());
2689 cfs_b
->runtime
= cfs_b
->quota
;
2690 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
2693 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
2695 return &tg
->cfs_bandwidth
;
2698 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2699 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
2701 if (unlikely(cfs_rq
->throttle_count
))
2702 return cfs_rq
->throttled_clock_task
;
2704 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
2707 /* returns 0 on failure to allocate runtime */
2708 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2710 struct task_group
*tg
= cfs_rq
->tg
;
2711 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
2712 u64 amount
= 0, min_amount
, expires
;
2714 /* note: this is a positive sum as runtime_remaining <= 0 */
2715 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
2717 raw_spin_lock(&cfs_b
->lock
);
2718 if (cfs_b
->quota
== RUNTIME_INF
)
2719 amount
= min_amount
;
2722 * If the bandwidth pool has become inactive, then at least one
2723 * period must have elapsed since the last consumption.
2724 * Refresh the global state and ensure bandwidth timer becomes
2727 if (!cfs_b
->timer_active
) {
2728 __refill_cfs_bandwidth_runtime(cfs_b
);
2729 __start_cfs_bandwidth(cfs_b
);
2732 if (cfs_b
->runtime
> 0) {
2733 amount
= min(cfs_b
->runtime
, min_amount
);
2734 cfs_b
->runtime
-= amount
;
2738 expires
= cfs_b
->runtime_expires
;
2739 raw_spin_unlock(&cfs_b
->lock
);
2741 cfs_rq
->runtime_remaining
+= amount
;
2743 * we may have advanced our local expiration to account for allowed
2744 * spread between our sched_clock and the one on which runtime was
2747 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
2748 cfs_rq
->runtime_expires
= expires
;
2750 return cfs_rq
->runtime_remaining
> 0;
2754 * Note: This depends on the synchronization provided by sched_clock and the
2755 * fact that rq->clock snapshots this value.
2757 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2759 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2761 /* if the deadline is ahead of our clock, nothing to do */
2762 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
2765 if (cfs_rq
->runtime_remaining
< 0)
2769 * If the local deadline has passed we have to consider the
2770 * possibility that our sched_clock is 'fast' and the global deadline
2771 * has not truly expired.
2773 * Fortunately we can check determine whether this the case by checking
2774 * whether the global deadline has advanced.
2777 if ((s64
)(cfs_rq
->runtime_expires
- cfs_b
->runtime_expires
) >= 0) {
2778 /* extend local deadline, drift is bounded above by 2 ticks */
2779 cfs_rq
->runtime_expires
+= TICK_NSEC
;
2781 /* global deadline is ahead, expiration has passed */
2782 cfs_rq
->runtime_remaining
= 0;
2786 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
,
2787 unsigned long delta_exec
)
2789 /* dock delta_exec before expiring quota (as it could span periods) */
2790 cfs_rq
->runtime_remaining
-= delta_exec
;
2791 expire_cfs_rq_runtime(cfs_rq
);
2793 if (likely(cfs_rq
->runtime_remaining
> 0))
2797 * if we're unable to extend our runtime we resched so that the active
2798 * hierarchy can be throttled
2800 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
2801 resched_task(rq_of(cfs_rq
)->curr
);
2804 static __always_inline
2805 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
)
2807 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
2810 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
2813 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
2815 return cfs_bandwidth_used() && cfs_rq
->throttled
;
2818 /* check whether cfs_rq, or any parent, is throttled */
2819 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
2821 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
2825 * Ensure that neither of the group entities corresponding to src_cpu or
2826 * dest_cpu are members of a throttled hierarchy when performing group
2827 * load-balance operations.
2829 static inline int throttled_lb_pair(struct task_group
*tg
,
2830 int src_cpu
, int dest_cpu
)
2832 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
2834 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
2835 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
2837 return throttled_hierarchy(src_cfs_rq
) ||
2838 throttled_hierarchy(dest_cfs_rq
);
2841 /* updated child weight may affect parent so we have to do this bottom up */
2842 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
2844 struct rq
*rq
= data
;
2845 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
2847 cfs_rq
->throttle_count
--;
2849 if (!cfs_rq
->throttle_count
) {
2850 /* adjust cfs_rq_clock_task() */
2851 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
2852 cfs_rq
->throttled_clock_task
;
2859 static int tg_throttle_down(struct task_group
*tg
, void *data
)
2861 struct rq
*rq
= data
;
2862 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
2864 /* group is entering throttled state, stop time */
2865 if (!cfs_rq
->throttle_count
)
2866 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
2867 cfs_rq
->throttle_count
++;
2872 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
2874 struct rq
*rq
= rq_of(cfs_rq
);
2875 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2876 struct sched_entity
*se
;
2877 long task_delta
, dequeue
= 1;
2879 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
2881 /* freeze hierarchy runnable averages while throttled */
2883 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
2886 task_delta
= cfs_rq
->h_nr_running
;
2887 for_each_sched_entity(se
) {
2888 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
2889 /* throttled entity or throttle-on-deactivate */
2894 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
2895 qcfs_rq
->h_nr_running
-= task_delta
;
2897 if (qcfs_rq
->load
.weight
)
2902 rq
->nr_running
-= task_delta
;
2904 cfs_rq
->throttled
= 1;
2905 cfs_rq
->throttled_clock
= rq_clock(rq
);
2906 raw_spin_lock(&cfs_b
->lock
);
2907 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
2908 raw_spin_unlock(&cfs_b
->lock
);
2911 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
2913 struct rq
*rq
= rq_of(cfs_rq
);
2914 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2915 struct sched_entity
*se
;
2919 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
2921 cfs_rq
->throttled
= 0;
2923 update_rq_clock(rq
);
2925 raw_spin_lock(&cfs_b
->lock
);
2926 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
2927 list_del_rcu(&cfs_rq
->throttled_list
);
2928 raw_spin_unlock(&cfs_b
->lock
);
2930 /* update hierarchical throttle state */
2931 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
2933 if (!cfs_rq
->load
.weight
)
2936 task_delta
= cfs_rq
->h_nr_running
;
2937 for_each_sched_entity(se
) {
2941 cfs_rq
= cfs_rq_of(se
);
2943 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
2944 cfs_rq
->h_nr_running
+= task_delta
;
2946 if (cfs_rq_throttled(cfs_rq
))
2951 rq
->nr_running
+= task_delta
;
2953 /* determine whether we need to wake up potentially idle cpu */
2954 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
2955 resched_task(rq
->curr
);
2958 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
2959 u64 remaining
, u64 expires
)
2961 struct cfs_rq
*cfs_rq
;
2962 u64 runtime
= remaining
;
2965 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
2967 struct rq
*rq
= rq_of(cfs_rq
);
2969 raw_spin_lock(&rq
->lock
);
2970 if (!cfs_rq_throttled(cfs_rq
))
2973 runtime
= -cfs_rq
->runtime_remaining
+ 1;
2974 if (runtime
> remaining
)
2975 runtime
= remaining
;
2976 remaining
-= runtime
;
2978 cfs_rq
->runtime_remaining
+= runtime
;
2979 cfs_rq
->runtime_expires
= expires
;
2981 /* we check whether we're throttled above */
2982 if (cfs_rq
->runtime_remaining
> 0)
2983 unthrottle_cfs_rq(cfs_rq
);
2986 raw_spin_unlock(&rq
->lock
);
2997 * Responsible for refilling a task_group's bandwidth and unthrottling its
2998 * cfs_rqs as appropriate. If there has been no activity within the last
2999 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3000 * used to track this state.
3002 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3004 u64 runtime
, runtime_expires
;
3005 int idle
= 1, throttled
;
3007 raw_spin_lock(&cfs_b
->lock
);
3008 /* no need to continue the timer with no bandwidth constraint */
3009 if (cfs_b
->quota
== RUNTIME_INF
)
3012 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3013 /* idle depends on !throttled (for the case of a large deficit) */
3014 idle
= cfs_b
->idle
&& !throttled
;
3015 cfs_b
->nr_periods
+= overrun
;
3017 /* if we're going inactive then everything else can be deferred */
3021 __refill_cfs_bandwidth_runtime(cfs_b
);
3024 /* mark as potentially idle for the upcoming period */
3029 /* account preceding periods in which throttling occurred */
3030 cfs_b
->nr_throttled
+= overrun
;
3033 * There are throttled entities so we must first use the new bandwidth
3034 * to unthrottle them before making it generally available. This
3035 * ensures that all existing debts will be paid before a new cfs_rq is
3038 runtime
= cfs_b
->runtime
;
3039 runtime_expires
= cfs_b
->runtime_expires
;
3043 * This check is repeated as we are holding onto the new bandwidth
3044 * while we unthrottle. This can potentially race with an unthrottled
3045 * group trying to acquire new bandwidth from the global pool.
3047 while (throttled
&& runtime
> 0) {
3048 raw_spin_unlock(&cfs_b
->lock
);
3049 /* we can't nest cfs_b->lock while distributing bandwidth */
3050 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3052 raw_spin_lock(&cfs_b
->lock
);
3054 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3057 /* return (any) remaining runtime */
3058 cfs_b
->runtime
= runtime
;
3060 * While we are ensured activity in the period following an
3061 * unthrottle, this also covers the case in which the new bandwidth is
3062 * insufficient to cover the existing bandwidth deficit. (Forcing the
3063 * timer to remain active while there are any throttled entities.)
3068 cfs_b
->timer_active
= 0;
3069 raw_spin_unlock(&cfs_b
->lock
);
3074 /* a cfs_rq won't donate quota below this amount */
3075 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3076 /* minimum remaining period time to redistribute slack quota */
3077 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3078 /* how long we wait to gather additional slack before distributing */
3079 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3081 /* are we near the end of the current quota period? */
3082 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3084 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3087 /* if the call-back is running a quota refresh is already occurring */
3088 if (hrtimer_callback_running(refresh_timer
))
3091 /* is a quota refresh about to occur? */
3092 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3093 if (remaining
< min_expire
)
3099 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3101 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3103 /* if there's a quota refresh soon don't bother with slack */
3104 if (runtime_refresh_within(cfs_b
, min_left
))
3107 start_bandwidth_timer(&cfs_b
->slack_timer
,
3108 ns_to_ktime(cfs_bandwidth_slack_period
));
3111 /* we know any runtime found here is valid as update_curr() precedes return */
3112 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3114 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3115 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3117 if (slack_runtime
<= 0)
3120 raw_spin_lock(&cfs_b
->lock
);
3121 if (cfs_b
->quota
!= RUNTIME_INF
&&
3122 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3123 cfs_b
->runtime
+= slack_runtime
;
3125 /* we are under rq->lock, defer unthrottling using a timer */
3126 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3127 !list_empty(&cfs_b
->throttled_cfs_rq
))
3128 start_cfs_slack_bandwidth(cfs_b
);
3130 raw_spin_unlock(&cfs_b
->lock
);
3132 /* even if it's not valid for return we don't want to try again */
3133 cfs_rq
->runtime_remaining
-= slack_runtime
;
3136 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3138 if (!cfs_bandwidth_used())
3141 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3144 __return_cfs_rq_runtime(cfs_rq
);
3148 * This is done with a timer (instead of inline with bandwidth return) since
3149 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3151 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3153 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3156 /* confirm we're still not at a refresh boundary */
3157 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
))
3160 raw_spin_lock(&cfs_b
->lock
);
3161 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
) {
3162 runtime
= cfs_b
->runtime
;
3165 expires
= cfs_b
->runtime_expires
;
3166 raw_spin_unlock(&cfs_b
->lock
);
3171 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3173 raw_spin_lock(&cfs_b
->lock
);
3174 if (expires
== cfs_b
->runtime_expires
)
3175 cfs_b
->runtime
= runtime
;
3176 raw_spin_unlock(&cfs_b
->lock
);
3180 * When a group wakes up we want to make sure that its quota is not already
3181 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3182 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3184 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3186 if (!cfs_bandwidth_used())
3189 /* an active group must be handled by the update_curr()->put() path */
3190 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3193 /* ensure the group is not already throttled */
3194 if (cfs_rq_throttled(cfs_rq
))
3197 /* update runtime allocation */
3198 account_cfs_rq_runtime(cfs_rq
, 0);
3199 if (cfs_rq
->runtime_remaining
<= 0)
3200 throttle_cfs_rq(cfs_rq
);
3203 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3204 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3206 if (!cfs_bandwidth_used())
3209 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3213 * it's possible for a throttled entity to be forced into a running
3214 * state (e.g. set_curr_task), in this case we're finished.
3216 if (cfs_rq_throttled(cfs_rq
))
3219 throttle_cfs_rq(cfs_rq
);
3222 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3224 struct cfs_bandwidth
*cfs_b
=
3225 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3226 do_sched_cfs_slack_timer(cfs_b
);
3228 return HRTIMER_NORESTART
;
3231 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3233 struct cfs_bandwidth
*cfs_b
=
3234 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3240 now
= hrtimer_cb_get_time(timer
);
3241 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3246 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3249 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3252 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3254 raw_spin_lock_init(&cfs_b
->lock
);
3256 cfs_b
->quota
= RUNTIME_INF
;
3257 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3259 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3260 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3261 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3262 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3263 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3266 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3268 cfs_rq
->runtime_enabled
= 0;
3269 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3272 /* requires cfs_b->lock, may release to reprogram timer */
3273 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3276 * The timer may be active because we're trying to set a new bandwidth
3277 * period or because we're racing with the tear-down path
3278 * (timer_active==0 becomes visible before the hrtimer call-back
3279 * terminates). In either case we ensure that it's re-programmed
3281 while (unlikely(hrtimer_active(&cfs_b
->period_timer
))) {
3282 raw_spin_unlock(&cfs_b
->lock
);
3283 /* ensure cfs_b->lock is available while we wait */
3284 hrtimer_cancel(&cfs_b
->period_timer
);
3286 raw_spin_lock(&cfs_b
->lock
);
3287 /* if someone else restarted the timer then we're done */
3288 if (cfs_b
->timer_active
)
3292 cfs_b
->timer_active
= 1;
3293 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
3296 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3298 hrtimer_cancel(&cfs_b
->period_timer
);
3299 hrtimer_cancel(&cfs_b
->slack_timer
);
3302 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
3304 struct cfs_rq
*cfs_rq
;
3306 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3307 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3309 if (!cfs_rq
->runtime_enabled
)
3313 * clock_task is not advancing so we just need to make sure
3314 * there's some valid quota amount
3316 cfs_rq
->runtime_remaining
= cfs_b
->quota
;
3317 if (cfs_rq_throttled(cfs_rq
))
3318 unthrottle_cfs_rq(cfs_rq
);
3322 #else /* CONFIG_CFS_BANDWIDTH */
3323 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3325 return rq_clock_task(rq_of(cfs_rq
));
3328 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
,
3329 unsigned long delta_exec
) {}
3330 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3331 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
3332 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3334 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3339 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3344 static inline int throttled_lb_pair(struct task_group
*tg
,
3345 int src_cpu
, int dest_cpu
)
3350 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3352 #ifdef CONFIG_FAIR_GROUP_SCHED
3353 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3356 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3360 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3361 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
3363 #endif /* CONFIG_CFS_BANDWIDTH */
3365 /**************************************************
3366 * CFS operations on tasks:
3369 #ifdef CONFIG_SCHED_HRTICK
3370 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3372 struct sched_entity
*se
= &p
->se
;
3373 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3375 WARN_ON(task_rq(p
) != rq
);
3377 if (cfs_rq
->nr_running
> 1) {
3378 u64 slice
= sched_slice(cfs_rq
, se
);
3379 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
3380 s64 delta
= slice
- ran
;
3389 * Don't schedule slices shorter than 10000ns, that just
3390 * doesn't make sense. Rely on vruntime for fairness.
3393 delta
= max_t(s64
, 10000LL, delta
);
3395 hrtick_start(rq
, delta
);
3400 * called from enqueue/dequeue and updates the hrtick when the
3401 * current task is from our class and nr_running is low enough
3404 static void hrtick_update(struct rq
*rq
)
3406 struct task_struct
*curr
= rq
->curr
;
3408 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
3411 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
3412 hrtick_start_fair(rq
, curr
);
3414 #else /* !CONFIG_SCHED_HRTICK */
3416 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3420 static inline void hrtick_update(struct rq
*rq
)
3426 * The enqueue_task method is called before nr_running is
3427 * increased. Here we update the fair scheduling stats and
3428 * then put the task into the rbtree:
3431 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3433 struct cfs_rq
*cfs_rq
;
3434 struct sched_entity
*se
= &p
->se
;
3436 for_each_sched_entity(se
) {
3439 cfs_rq
= cfs_rq_of(se
);
3440 enqueue_entity(cfs_rq
, se
, flags
);
3443 * end evaluation on encountering a throttled cfs_rq
3445 * note: in the case of encountering a throttled cfs_rq we will
3446 * post the final h_nr_running increment below.
3448 if (cfs_rq_throttled(cfs_rq
))
3450 cfs_rq
->h_nr_running
++;
3452 flags
= ENQUEUE_WAKEUP
;
3455 for_each_sched_entity(se
) {
3456 cfs_rq
= cfs_rq_of(se
);
3457 cfs_rq
->h_nr_running
++;
3459 if (cfs_rq_throttled(cfs_rq
))
3462 update_cfs_shares(cfs_rq
);
3463 update_entity_load_avg(se
, 1);
3467 update_rq_runnable_avg(rq
, rq
->nr_running
);
3473 static void set_next_buddy(struct sched_entity
*se
);
3476 * The dequeue_task method is called before nr_running is
3477 * decreased. We remove the task from the rbtree and
3478 * update the fair scheduling stats:
3480 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3482 struct cfs_rq
*cfs_rq
;
3483 struct sched_entity
*se
= &p
->se
;
3484 int task_sleep
= flags
& DEQUEUE_SLEEP
;
3486 for_each_sched_entity(se
) {
3487 cfs_rq
= cfs_rq_of(se
);
3488 dequeue_entity(cfs_rq
, se
, flags
);
3491 * end evaluation on encountering a throttled cfs_rq
3493 * note: in the case of encountering a throttled cfs_rq we will
3494 * post the final h_nr_running decrement below.
3496 if (cfs_rq_throttled(cfs_rq
))
3498 cfs_rq
->h_nr_running
--;
3500 /* Don't dequeue parent if it has other entities besides us */
3501 if (cfs_rq
->load
.weight
) {
3503 * Bias pick_next to pick a task from this cfs_rq, as
3504 * p is sleeping when it is within its sched_slice.
3506 if (task_sleep
&& parent_entity(se
))
3507 set_next_buddy(parent_entity(se
));
3509 /* avoid re-evaluating load for this entity */
3510 se
= parent_entity(se
);
3513 flags
|= DEQUEUE_SLEEP
;
3516 for_each_sched_entity(se
) {
3517 cfs_rq
= cfs_rq_of(se
);
3518 cfs_rq
->h_nr_running
--;
3520 if (cfs_rq_throttled(cfs_rq
))
3523 update_cfs_shares(cfs_rq
);
3524 update_entity_load_avg(se
, 1);
3529 update_rq_runnable_avg(rq
, 1);
3535 /* Used instead of source_load when we know the type == 0 */
3536 static unsigned long weighted_cpuload(const int cpu
)
3538 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
3542 * Return a low guess at the load of a migration-source cpu weighted
3543 * according to the scheduling class and "nice" value.
3545 * We want to under-estimate the load of migration sources, to
3546 * balance conservatively.
3548 static unsigned long source_load(int cpu
, int type
)
3550 struct rq
*rq
= cpu_rq(cpu
);
3551 unsigned long total
= weighted_cpuload(cpu
);
3553 if (type
== 0 || !sched_feat(LB_BIAS
))
3556 return min(rq
->cpu_load
[type
-1], total
);
3560 * Return a high guess at the load of a migration-target cpu weighted
3561 * according to the scheduling class and "nice" value.
3563 static unsigned long target_load(int cpu
, int type
)
3565 struct rq
*rq
= cpu_rq(cpu
);
3566 unsigned long total
= weighted_cpuload(cpu
);
3568 if (type
== 0 || !sched_feat(LB_BIAS
))
3571 return max(rq
->cpu_load
[type
-1], total
);
3574 static unsigned long power_of(int cpu
)
3576 return cpu_rq(cpu
)->cpu_power
;
3579 static unsigned long cpu_avg_load_per_task(int cpu
)
3581 struct rq
*rq
= cpu_rq(cpu
);
3582 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
3583 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
3586 return load_avg
/ nr_running
;
3591 static void record_wakee(struct task_struct
*p
)
3594 * Rough decay (wiping) for cost saving, don't worry
3595 * about the boundary, really active task won't care
3598 if (jiffies
> current
->wakee_flip_decay_ts
+ HZ
) {
3599 current
->wakee_flips
= 0;
3600 current
->wakee_flip_decay_ts
= jiffies
;
3603 if (current
->last_wakee
!= p
) {
3604 current
->last_wakee
= p
;
3605 current
->wakee_flips
++;
3609 static void task_waking_fair(struct task_struct
*p
)
3611 struct sched_entity
*se
= &p
->se
;
3612 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3615 #ifndef CONFIG_64BIT
3616 u64 min_vruntime_copy
;
3619 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
3621 min_vruntime
= cfs_rq
->min_vruntime
;
3622 } while (min_vruntime
!= min_vruntime_copy
);
3624 min_vruntime
= cfs_rq
->min_vruntime
;
3627 se
->vruntime
-= min_vruntime
;
3631 #ifdef CONFIG_FAIR_GROUP_SCHED
3633 * effective_load() calculates the load change as seen from the root_task_group
3635 * Adding load to a group doesn't make a group heavier, but can cause movement
3636 * of group shares between cpus. Assuming the shares were perfectly aligned one
3637 * can calculate the shift in shares.
3639 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3640 * on this @cpu and results in a total addition (subtraction) of @wg to the
3641 * total group weight.
3643 * Given a runqueue weight distribution (rw_i) we can compute a shares
3644 * distribution (s_i) using:
3646 * s_i = rw_i / \Sum rw_j (1)
3648 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3649 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3650 * shares distribution (s_i):
3652 * rw_i = { 2, 4, 1, 0 }
3653 * s_i = { 2/7, 4/7, 1/7, 0 }
3655 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3656 * task used to run on and the CPU the waker is running on), we need to
3657 * compute the effect of waking a task on either CPU and, in case of a sync
3658 * wakeup, compute the effect of the current task going to sleep.
3660 * So for a change of @wl to the local @cpu with an overall group weight change
3661 * of @wl we can compute the new shares distribution (s'_i) using:
3663 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3665 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3666 * differences in waking a task to CPU 0. The additional task changes the
3667 * weight and shares distributions like:
3669 * rw'_i = { 3, 4, 1, 0 }
3670 * s'_i = { 3/8, 4/8, 1/8, 0 }
3672 * We can then compute the difference in effective weight by using:
3674 * dw_i = S * (s'_i - s_i) (3)
3676 * Where 'S' is the group weight as seen by its parent.
3678 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3679 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3680 * 4/7) times the weight of the group.
3682 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
3684 struct sched_entity
*se
= tg
->se
[cpu
];
3686 if (!tg
->parent
|| !wl
) /* the trivial, non-cgroup case */
3689 for_each_sched_entity(se
) {
3695 * W = @wg + \Sum rw_j
3697 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
3702 w
= se
->my_q
->load
.weight
+ wl
;
3705 * wl = S * s'_i; see (2)
3708 wl
= (w
* tg
->shares
) / W
;
3713 * Per the above, wl is the new se->load.weight value; since
3714 * those are clipped to [MIN_SHARES, ...) do so now. See
3715 * calc_cfs_shares().
3717 if (wl
< MIN_SHARES
)
3721 * wl = dw_i = S * (s'_i - s_i); see (3)
3723 wl
-= se
->load
.weight
;
3726 * Recursively apply this logic to all parent groups to compute
3727 * the final effective load change on the root group. Since
3728 * only the @tg group gets extra weight, all parent groups can
3729 * only redistribute existing shares. @wl is the shift in shares
3730 * resulting from this level per the above.
3739 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
3746 static int wake_wide(struct task_struct
*p
)
3748 int factor
= this_cpu_read(sd_llc_size
);
3751 * Yeah, it's the switching-frequency, could means many wakee or
3752 * rapidly switch, use factor here will just help to automatically
3753 * adjust the loose-degree, so bigger node will lead to more pull.
3755 if (p
->wakee_flips
> factor
) {
3757 * wakee is somewhat hot, it needs certain amount of cpu
3758 * resource, so if waker is far more hot, prefer to leave
3761 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
3768 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
3770 s64 this_load
, load
;
3771 int idx
, this_cpu
, prev_cpu
;
3772 unsigned long tl_per_task
;
3773 struct task_group
*tg
;
3774 unsigned long weight
;
3778 * If we wake multiple tasks be careful to not bounce
3779 * ourselves around too much.
3785 this_cpu
= smp_processor_id();
3786 prev_cpu
= task_cpu(p
);
3787 load
= source_load(prev_cpu
, idx
);
3788 this_load
= target_load(this_cpu
, idx
);
3791 * If sync wakeup then subtract the (maximum possible)
3792 * effect of the currently running task from the load
3793 * of the current CPU:
3796 tg
= task_group(current
);
3797 weight
= current
->se
.load
.weight
;
3799 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
3800 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
3804 weight
= p
->se
.load
.weight
;
3807 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3808 * due to the sync cause above having dropped this_load to 0, we'll
3809 * always have an imbalance, but there's really nothing you can do
3810 * about that, so that's good too.
3812 * Otherwise check if either cpus are near enough in load to allow this
3813 * task to be woken on this_cpu.
3815 if (this_load
> 0) {
3816 s64 this_eff_load
, prev_eff_load
;
3818 this_eff_load
= 100;
3819 this_eff_load
*= power_of(prev_cpu
);
3820 this_eff_load
*= this_load
+
3821 effective_load(tg
, this_cpu
, weight
, weight
);
3823 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
3824 prev_eff_load
*= power_of(this_cpu
);
3825 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
3827 balanced
= this_eff_load
<= prev_eff_load
;
3832 * If the currently running task will sleep within
3833 * a reasonable amount of time then attract this newly
3836 if (sync
&& balanced
)
3839 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
3840 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
3843 (this_load
<= load
&&
3844 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
3846 * This domain has SD_WAKE_AFFINE and
3847 * p is cache cold in this domain, and
3848 * there is no bad imbalance.
3850 schedstat_inc(sd
, ttwu_move_affine
);
3851 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
3859 * find_idlest_group finds and returns the least busy CPU group within the
3862 static struct sched_group
*
3863 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
3864 int this_cpu
, int load_idx
)
3866 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
3867 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
3868 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
3871 unsigned long load
, avg_load
;
3875 /* Skip over this group if it has no CPUs allowed */
3876 if (!cpumask_intersects(sched_group_cpus(group
),
3877 tsk_cpus_allowed(p
)))
3880 local_group
= cpumask_test_cpu(this_cpu
,
3881 sched_group_cpus(group
));
3883 /* Tally up the load of all CPUs in the group */
3886 for_each_cpu(i
, sched_group_cpus(group
)) {
3887 /* Bias balancing toward cpus of our domain */
3889 load
= source_load(i
, load_idx
);
3891 load
= target_load(i
, load_idx
);
3896 /* Adjust by relative CPU power of the group */
3897 avg_load
= (avg_load
* SCHED_POWER_SCALE
) / group
->sgp
->power
;
3900 this_load
= avg_load
;
3901 } else if (avg_load
< min_load
) {
3902 min_load
= avg_load
;
3905 } while (group
= group
->next
, group
!= sd
->groups
);
3907 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
3913 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3916 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
3918 unsigned long load
, min_load
= ULONG_MAX
;
3922 /* Traverse only the allowed CPUs */
3923 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
3924 load
= weighted_cpuload(i
);
3926 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
3936 * Try and locate an idle CPU in the sched_domain.
3938 static int select_idle_sibling(struct task_struct
*p
, int target
)
3940 struct sched_domain
*sd
;
3941 struct sched_group
*sg
;
3942 int i
= task_cpu(p
);
3944 if (idle_cpu(target
))
3948 * If the prevous cpu is cache affine and idle, don't be stupid.
3950 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
3954 * Otherwise, iterate the domains and find an elegible idle cpu.
3956 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
3957 for_each_lower_domain(sd
) {
3960 if (!cpumask_intersects(sched_group_cpus(sg
),
3961 tsk_cpus_allowed(p
)))
3964 for_each_cpu(i
, sched_group_cpus(sg
)) {
3965 if (i
== target
|| !idle_cpu(i
))
3969 target
= cpumask_first_and(sched_group_cpus(sg
),
3970 tsk_cpus_allowed(p
));
3974 } while (sg
!= sd
->groups
);
3981 * sched_balance_self: balance the current task (running on cpu) in domains
3982 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3985 * Balance, ie. select the least loaded group.
3987 * Returns the target CPU number, or the same CPU if no balancing is needed.
3989 * preempt must be disabled.
3992 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
3994 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
3995 int cpu
= smp_processor_id();
3997 int want_affine
= 0;
3998 int sync
= wake_flags
& WF_SYNC
;
4000 if (p
->nr_cpus_allowed
== 1)
4003 if (sd_flag
& SD_BALANCE_WAKE
) {
4004 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
4010 for_each_domain(cpu
, tmp
) {
4011 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4015 * If both cpu and prev_cpu are part of this domain,
4016 * cpu is a valid SD_WAKE_AFFINE target.
4018 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4019 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4024 if (tmp
->flags
& sd_flag
)
4029 if (cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4032 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4037 int load_idx
= sd
->forkexec_idx
;
4038 struct sched_group
*group
;
4041 if (!(sd
->flags
& sd_flag
)) {
4046 if (sd_flag
& SD_BALANCE_WAKE
)
4047 load_idx
= sd
->wake_idx
;
4049 group
= find_idlest_group(sd
, p
, cpu
, load_idx
);
4055 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4056 if (new_cpu
== -1 || new_cpu
== cpu
) {
4057 /* Now try balancing at a lower domain level of cpu */
4062 /* Now try balancing at a lower domain level of new_cpu */
4064 weight
= sd
->span_weight
;
4066 for_each_domain(cpu
, tmp
) {
4067 if (weight
<= tmp
->span_weight
)
4069 if (tmp
->flags
& sd_flag
)
4072 /* while loop will break here if sd == NULL */
4081 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4082 * cfs_rq_of(p) references at time of call are still valid and identify the
4083 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4084 * other assumptions, including the state of rq->lock, should be made.
4087 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4089 struct sched_entity
*se
= &p
->se
;
4090 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4093 * Load tracking: accumulate removed load so that it can be processed
4094 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4095 * to blocked load iff they have a positive decay-count. It can never
4096 * be negative here since on-rq tasks have decay-count == 0.
4098 if (se
->avg
.decay_count
) {
4099 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4100 atomic_long_add(se
->avg
.load_avg_contrib
,
4101 &cfs_rq
->removed_load
);
4104 #endif /* CONFIG_SMP */
4106 static unsigned long
4107 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4109 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4112 * Since its curr running now, convert the gran from real-time
4113 * to virtual-time in his units.
4115 * By using 'se' instead of 'curr' we penalize light tasks, so
4116 * they get preempted easier. That is, if 'se' < 'curr' then
4117 * the resulting gran will be larger, therefore penalizing the
4118 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4119 * be smaller, again penalizing the lighter task.
4121 * This is especially important for buddies when the leftmost
4122 * task is higher priority than the buddy.
4124 return calc_delta_fair(gran
, se
);
4128 * Should 'se' preempt 'curr'.
4142 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4144 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4149 gran
= wakeup_gran(curr
, se
);
4156 static void set_last_buddy(struct sched_entity
*se
)
4158 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4161 for_each_sched_entity(se
)
4162 cfs_rq_of(se
)->last
= se
;
4165 static void set_next_buddy(struct sched_entity
*se
)
4167 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4170 for_each_sched_entity(se
)
4171 cfs_rq_of(se
)->next
= se
;
4174 static void set_skip_buddy(struct sched_entity
*se
)
4176 for_each_sched_entity(se
)
4177 cfs_rq_of(se
)->skip
= se
;
4181 * Preempt the current task with a newly woken task if needed:
4183 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4185 struct task_struct
*curr
= rq
->curr
;
4186 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4187 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4188 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4189 int next_buddy_marked
= 0;
4191 if (unlikely(se
== pse
))
4195 * This is possible from callers such as move_task(), in which we
4196 * unconditionally check_prempt_curr() after an enqueue (which may have
4197 * lead to a throttle). This both saves work and prevents false
4198 * next-buddy nomination below.
4200 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4203 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4204 set_next_buddy(pse
);
4205 next_buddy_marked
= 1;
4209 * We can come here with TIF_NEED_RESCHED already set from new task
4212 * Note: this also catches the edge-case of curr being in a throttled
4213 * group (e.g. via set_curr_task), since update_curr() (in the
4214 * enqueue of curr) will have resulted in resched being set. This
4215 * prevents us from potentially nominating it as a false LAST_BUDDY
4218 if (test_tsk_need_resched(curr
))
4221 /* Idle tasks are by definition preempted by non-idle tasks. */
4222 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4223 likely(p
->policy
!= SCHED_IDLE
))
4227 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4228 * is driven by the tick):
4230 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4233 find_matching_se(&se
, &pse
);
4234 update_curr(cfs_rq_of(se
));
4236 if (wakeup_preempt_entity(se
, pse
) == 1) {
4238 * Bias pick_next to pick the sched entity that is
4239 * triggering this preemption.
4241 if (!next_buddy_marked
)
4242 set_next_buddy(pse
);
4251 * Only set the backward buddy when the current task is still
4252 * on the rq. This can happen when a wakeup gets interleaved
4253 * with schedule on the ->pre_schedule() or idle_balance()
4254 * point, either of which can * drop the rq lock.
4256 * Also, during early boot the idle thread is in the fair class,
4257 * for obvious reasons its a bad idea to schedule back to it.
4259 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4262 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
4266 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
4268 struct task_struct
*p
;
4269 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
4270 struct sched_entity
*se
;
4272 if (!cfs_rq
->nr_running
)
4276 se
= pick_next_entity(cfs_rq
);
4277 set_next_entity(cfs_rq
, se
);
4278 cfs_rq
= group_cfs_rq(se
);
4282 if (hrtick_enabled(rq
))
4283 hrtick_start_fair(rq
, p
);
4289 * Account for a descheduled task:
4291 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4293 struct sched_entity
*se
= &prev
->se
;
4294 struct cfs_rq
*cfs_rq
;
4296 for_each_sched_entity(se
) {
4297 cfs_rq
= cfs_rq_of(se
);
4298 put_prev_entity(cfs_rq
, se
);
4303 * sched_yield() is very simple
4305 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4307 static void yield_task_fair(struct rq
*rq
)
4309 struct task_struct
*curr
= rq
->curr
;
4310 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4311 struct sched_entity
*se
= &curr
->se
;
4314 * Are we the only task in the tree?
4316 if (unlikely(rq
->nr_running
== 1))
4319 clear_buddies(cfs_rq
, se
);
4321 if (curr
->policy
!= SCHED_BATCH
) {
4322 update_rq_clock(rq
);
4324 * Update run-time statistics of the 'current'.
4326 update_curr(cfs_rq
);
4328 * Tell update_rq_clock() that we've just updated,
4329 * so we don't do microscopic update in schedule()
4330 * and double the fastpath cost.
4332 rq
->skip_clock_update
= 1;
4338 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
4340 struct sched_entity
*se
= &p
->se
;
4342 /* throttled hierarchies are not runnable */
4343 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
4346 /* Tell the scheduler that we'd really like pse to run next. */
4349 yield_task_fair(rq
);
4355 /**************************************************
4356 * Fair scheduling class load-balancing methods.
4360 * The purpose of load-balancing is to achieve the same basic fairness the
4361 * per-cpu scheduler provides, namely provide a proportional amount of compute
4362 * time to each task. This is expressed in the following equation:
4364 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4366 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4367 * W_i,0 is defined as:
4369 * W_i,0 = \Sum_j w_i,j (2)
4371 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4372 * is derived from the nice value as per prio_to_weight[].
4374 * The weight average is an exponential decay average of the instantaneous
4377 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4379 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4380 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4381 * can also include other factors [XXX].
4383 * To achieve this balance we define a measure of imbalance which follows
4384 * directly from (1):
4386 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4388 * We them move tasks around to minimize the imbalance. In the continuous
4389 * function space it is obvious this converges, in the discrete case we get
4390 * a few fun cases generally called infeasible weight scenarios.
4393 * - infeasible weights;
4394 * - local vs global optima in the discrete case. ]
4399 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4400 * for all i,j solution, we create a tree of cpus that follows the hardware
4401 * topology where each level pairs two lower groups (or better). This results
4402 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4403 * tree to only the first of the previous level and we decrease the frequency
4404 * of load-balance at each level inv. proportional to the number of cpus in
4410 * \Sum { --- * --- * 2^i } = O(n) (5)
4412 * `- size of each group
4413 * | | `- number of cpus doing load-balance
4415 * `- sum over all levels
4417 * Coupled with a limit on how many tasks we can migrate every balance pass,
4418 * this makes (5) the runtime complexity of the balancer.
4420 * An important property here is that each CPU is still (indirectly) connected
4421 * to every other cpu in at most O(log n) steps:
4423 * The adjacency matrix of the resulting graph is given by:
4426 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4429 * And you'll find that:
4431 * A^(log_2 n)_i,j != 0 for all i,j (7)
4433 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4434 * The task movement gives a factor of O(m), giving a convergence complexity
4437 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4442 * In order to avoid CPUs going idle while there's still work to do, new idle
4443 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4444 * tree itself instead of relying on other CPUs to bring it work.
4446 * This adds some complexity to both (5) and (8) but it reduces the total idle
4454 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4457 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4462 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4464 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4466 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4469 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4470 * rewrite all of this once again.]
4473 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
4475 #define LBF_ALL_PINNED 0x01
4476 #define LBF_NEED_BREAK 0x02
4477 #define LBF_DST_PINNED 0x04
4478 #define LBF_SOME_PINNED 0x08
4481 struct sched_domain
*sd
;
4489 struct cpumask
*dst_grpmask
;
4491 enum cpu_idle_type idle
;
4493 /* The set of CPUs under consideration for load-balancing */
4494 struct cpumask
*cpus
;
4499 unsigned int loop_break
;
4500 unsigned int loop_max
;
4504 * move_task - move a task from one runqueue to another runqueue.
4505 * Both runqueues must be locked.
4507 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
4509 deactivate_task(env
->src_rq
, p
, 0);
4510 set_task_cpu(p
, env
->dst_cpu
);
4511 activate_task(env
->dst_rq
, p
, 0);
4512 check_preempt_curr(env
->dst_rq
, p
, 0);
4513 #ifdef CONFIG_NUMA_BALANCING
4514 if (p
->numa_preferred_nid
!= -1) {
4515 int src_nid
= cpu_to_node(env
->src_cpu
);
4516 int dst_nid
= cpu_to_node(env
->dst_cpu
);
4519 * If the load balancer has moved the task then limit
4520 * migrations from taking place in the short term in
4521 * case this is a short-lived migration.
4523 if (src_nid
!= dst_nid
&& dst_nid
!= p
->numa_preferred_nid
)
4524 p
->numa_migrate_seq
= 0;
4530 * Is this task likely cache-hot:
4533 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
4537 if (p
->sched_class
!= &fair_sched_class
)
4540 if (unlikely(p
->policy
== SCHED_IDLE
))
4544 * Buddy candidates are cache hot:
4546 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
4547 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
4548 &p
->se
== cfs_rq_of(&p
->se
)->last
))
4551 if (sysctl_sched_migration_cost
== -1)
4553 if (sysctl_sched_migration_cost
== 0)
4556 delta
= now
- p
->se
.exec_start
;
4558 return delta
< (s64
)sysctl_sched_migration_cost
;
4561 #ifdef CONFIG_NUMA_BALANCING
4562 /* Returns true if the destination node has incurred more faults */
4563 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
4565 int src_nid
, dst_nid
;
4567 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults
||
4568 !(env
->sd
->flags
& SD_NUMA
)) {
4572 src_nid
= cpu_to_node(env
->src_cpu
);
4573 dst_nid
= cpu_to_node(env
->dst_cpu
);
4575 if (src_nid
== dst_nid
||
4576 p
->numa_migrate_seq
>= sysctl_numa_balancing_settle_count
)
4579 if (dst_nid
== p
->numa_preferred_nid
||
4580 task_faults(p
, dst_nid
) > task_faults(p
, src_nid
))
4587 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
4589 int src_nid
, dst_nid
;
4591 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
4594 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
4597 src_nid
= cpu_to_node(env
->src_cpu
);
4598 dst_nid
= cpu_to_node(env
->dst_cpu
);
4600 if (src_nid
== dst_nid
||
4601 p
->numa_migrate_seq
>= sysctl_numa_balancing_settle_count
)
4604 if (task_faults(p
, dst_nid
) < task_faults(p
, src_nid
))
4611 static inline bool migrate_improves_locality(struct task_struct
*p
,
4617 static inline bool migrate_degrades_locality(struct task_struct
*p
,
4625 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4628 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
4630 int tsk_cache_hot
= 0;
4632 * We do not migrate tasks that are:
4633 * 1) throttled_lb_pair, or
4634 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4635 * 3) running (obviously), or
4636 * 4) are cache-hot on their current CPU.
4638 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
4641 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
4644 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
4646 env
->flags
|= LBF_SOME_PINNED
;
4649 * Remember if this task can be migrated to any other cpu in
4650 * our sched_group. We may want to revisit it if we couldn't
4651 * meet load balance goals by pulling other tasks on src_cpu.
4653 * Also avoid computing new_dst_cpu if we have already computed
4654 * one in current iteration.
4656 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
4659 /* Prevent to re-select dst_cpu via env's cpus */
4660 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
4661 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
4662 env
->flags
|= LBF_DST_PINNED
;
4663 env
->new_dst_cpu
= cpu
;
4671 /* Record that we found atleast one task that could run on dst_cpu */
4672 env
->flags
&= ~LBF_ALL_PINNED
;
4674 if (task_running(env
->src_rq
, p
)) {
4675 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
4680 * Aggressive migration if:
4681 * 1) destination numa is preferred
4682 * 2) task is cache cold, or
4683 * 3) too many balance attempts have failed.
4685 tsk_cache_hot
= task_hot(p
, rq_clock_task(env
->src_rq
), env
->sd
);
4687 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
4689 if (migrate_improves_locality(p
, env
)) {
4690 #ifdef CONFIG_SCHEDSTATS
4691 if (tsk_cache_hot
) {
4692 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
4693 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
4699 if (!tsk_cache_hot
||
4700 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
4702 if (tsk_cache_hot
) {
4703 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
4704 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
4710 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
4715 * move_one_task tries to move exactly one task from busiest to this_rq, as
4716 * part of active balancing operations within "domain".
4717 * Returns 1 if successful and 0 otherwise.
4719 * Called with both runqueues locked.
4721 static int move_one_task(struct lb_env
*env
)
4723 struct task_struct
*p
, *n
;
4725 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
4726 if (!can_migrate_task(p
, env
))
4731 * Right now, this is only the second place move_task()
4732 * is called, so we can safely collect move_task()
4733 * stats here rather than inside move_task().
4735 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
4741 static const unsigned int sched_nr_migrate_break
= 32;
4744 * move_tasks tries to move up to imbalance weighted load from busiest to
4745 * this_rq, as part of a balancing operation within domain "sd".
4746 * Returns 1 if successful and 0 otherwise.
4748 * Called with both runqueues locked.
4750 static int move_tasks(struct lb_env
*env
)
4752 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
4753 struct task_struct
*p
;
4757 if (env
->imbalance
<= 0)
4760 while (!list_empty(tasks
)) {
4761 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
4764 /* We've more or less seen every task there is, call it quits */
4765 if (env
->loop
> env
->loop_max
)
4768 /* take a breather every nr_migrate tasks */
4769 if (env
->loop
> env
->loop_break
) {
4770 env
->loop_break
+= sched_nr_migrate_break
;
4771 env
->flags
|= LBF_NEED_BREAK
;
4775 if (!can_migrate_task(p
, env
))
4778 load
= task_h_load(p
);
4780 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
4783 if ((load
/ 2) > env
->imbalance
)
4788 env
->imbalance
-= load
;
4790 #ifdef CONFIG_PREEMPT
4792 * NEWIDLE balancing is a source of latency, so preemptible
4793 * kernels will stop after the first task is pulled to minimize
4794 * the critical section.
4796 if (env
->idle
== CPU_NEWLY_IDLE
)
4801 * We only want to steal up to the prescribed amount of
4804 if (env
->imbalance
<= 0)
4809 list_move_tail(&p
->se
.group_node
, tasks
);
4813 * Right now, this is one of only two places move_task() is called,
4814 * so we can safely collect move_task() stats here rather than
4815 * inside move_task().
4817 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
4822 #ifdef CONFIG_FAIR_GROUP_SCHED
4824 * update tg->load_weight by folding this cpu's load_avg
4826 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
4828 struct sched_entity
*se
= tg
->se
[cpu
];
4829 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
4831 /* throttled entities do not contribute to load */
4832 if (throttled_hierarchy(cfs_rq
))
4835 update_cfs_rq_blocked_load(cfs_rq
, 1);
4838 update_entity_load_avg(se
, 1);
4840 * We pivot on our runnable average having decayed to zero for
4841 * list removal. This generally implies that all our children
4842 * have also been removed (modulo rounding error or bandwidth
4843 * control); however, such cases are rare and we can fix these
4846 * TODO: fix up out-of-order children on enqueue.
4848 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
4849 list_del_leaf_cfs_rq(cfs_rq
);
4851 struct rq
*rq
= rq_of(cfs_rq
);
4852 update_rq_runnable_avg(rq
, rq
->nr_running
);
4856 static void update_blocked_averages(int cpu
)
4858 struct rq
*rq
= cpu_rq(cpu
);
4859 struct cfs_rq
*cfs_rq
;
4860 unsigned long flags
;
4862 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4863 update_rq_clock(rq
);
4865 * Iterates the task_group tree in a bottom up fashion, see
4866 * list_add_leaf_cfs_rq() for details.
4868 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4870 * Note: We may want to consider periodically releasing
4871 * rq->lock about these updates so that creating many task
4872 * groups does not result in continually extending hold time.
4874 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
4877 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4881 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4882 * This needs to be done in a top-down fashion because the load of a child
4883 * group is a fraction of its parents load.
4885 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
4887 struct rq
*rq
= rq_of(cfs_rq
);
4888 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
4889 unsigned long now
= jiffies
;
4892 if (cfs_rq
->last_h_load_update
== now
)
4895 cfs_rq
->h_load_next
= NULL
;
4896 for_each_sched_entity(se
) {
4897 cfs_rq
= cfs_rq_of(se
);
4898 cfs_rq
->h_load_next
= se
;
4899 if (cfs_rq
->last_h_load_update
== now
)
4904 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
4905 cfs_rq
->last_h_load_update
= now
;
4908 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
4909 load
= cfs_rq
->h_load
;
4910 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
4911 cfs_rq
->runnable_load_avg
+ 1);
4912 cfs_rq
= group_cfs_rq(se
);
4913 cfs_rq
->h_load
= load
;
4914 cfs_rq
->last_h_load_update
= now
;
4918 static unsigned long task_h_load(struct task_struct
*p
)
4920 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
4922 update_cfs_rq_h_load(cfs_rq
);
4923 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
4924 cfs_rq
->runnable_load_avg
+ 1);
4927 static inline void update_blocked_averages(int cpu
)
4931 static unsigned long task_h_load(struct task_struct
*p
)
4933 return p
->se
.avg
.load_avg_contrib
;
4937 /********** Helpers for find_busiest_group ************************/
4939 * sg_lb_stats - stats of a sched_group required for load_balancing
4941 struct sg_lb_stats
{
4942 unsigned long avg_load
; /*Avg load across the CPUs of the group */
4943 unsigned long group_load
; /* Total load over the CPUs of the group */
4944 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
4945 unsigned long load_per_task
;
4946 unsigned long group_power
;
4947 unsigned int sum_nr_running
; /* Nr tasks running in the group */
4948 unsigned int group_capacity
;
4949 unsigned int idle_cpus
;
4950 unsigned int group_weight
;
4951 int group_imb
; /* Is there an imbalance in the group ? */
4952 int group_has_capacity
; /* Is there extra capacity in the group? */
4956 * sd_lb_stats - Structure to store the statistics of a sched_domain
4957 * during load balancing.
4959 struct sd_lb_stats
{
4960 struct sched_group
*busiest
; /* Busiest group in this sd */
4961 struct sched_group
*local
; /* Local group in this sd */
4962 unsigned long total_load
; /* Total load of all groups in sd */
4963 unsigned long total_pwr
; /* Total power of all groups in sd */
4964 unsigned long avg_load
; /* Average load across all groups in sd */
4966 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
4967 struct sg_lb_stats local_stat
; /* Statistics of the local group */
4970 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
4973 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4974 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4975 * We must however clear busiest_stat::avg_load because
4976 * update_sd_pick_busiest() reads this before assignment.
4978 *sds
= (struct sd_lb_stats
){
4990 * get_sd_load_idx - Obtain the load index for a given sched domain.
4991 * @sd: The sched_domain whose load_idx is to be obtained.
4992 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4994 * Return: The load index.
4996 static inline int get_sd_load_idx(struct sched_domain
*sd
,
4997 enum cpu_idle_type idle
)
5003 load_idx
= sd
->busy_idx
;
5006 case CPU_NEWLY_IDLE
:
5007 load_idx
= sd
->newidle_idx
;
5010 load_idx
= sd
->idle_idx
;
5017 static unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
5019 return SCHED_POWER_SCALE
;
5022 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
5024 return default_scale_freq_power(sd
, cpu
);
5027 static unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
5029 unsigned long weight
= sd
->span_weight
;
5030 unsigned long smt_gain
= sd
->smt_gain
;
5037 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
5039 return default_scale_smt_power(sd
, cpu
);
5042 static unsigned long scale_rt_power(int cpu
)
5044 struct rq
*rq
= cpu_rq(cpu
);
5045 u64 total
, available
, age_stamp
, avg
;
5048 * Since we're reading these variables without serialization make sure
5049 * we read them once before doing sanity checks on them.
5051 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5052 avg
= ACCESS_ONCE(rq
->rt_avg
);
5054 total
= sched_avg_period() + (rq_clock(rq
) - age_stamp
);
5056 if (unlikely(total
< avg
)) {
5057 /* Ensures that power won't end up being negative */
5060 available
= total
- avg
;
5063 if (unlikely((s64
)total
< SCHED_POWER_SCALE
))
5064 total
= SCHED_POWER_SCALE
;
5066 total
>>= SCHED_POWER_SHIFT
;
5068 return div_u64(available
, total
);
5071 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
5073 unsigned long weight
= sd
->span_weight
;
5074 unsigned long power
= SCHED_POWER_SCALE
;
5075 struct sched_group
*sdg
= sd
->groups
;
5077 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
5078 if (sched_feat(ARCH_POWER
))
5079 power
*= arch_scale_smt_power(sd
, cpu
);
5081 power
*= default_scale_smt_power(sd
, cpu
);
5083 power
>>= SCHED_POWER_SHIFT
;
5086 sdg
->sgp
->power_orig
= power
;
5088 if (sched_feat(ARCH_POWER
))
5089 power
*= arch_scale_freq_power(sd
, cpu
);
5091 power
*= default_scale_freq_power(sd
, cpu
);
5093 power
>>= SCHED_POWER_SHIFT
;
5095 power
*= scale_rt_power(cpu
);
5096 power
>>= SCHED_POWER_SHIFT
;
5101 cpu_rq(cpu
)->cpu_power
= power
;
5102 sdg
->sgp
->power
= power
;
5105 void update_group_power(struct sched_domain
*sd
, int cpu
)
5107 struct sched_domain
*child
= sd
->child
;
5108 struct sched_group
*group
, *sdg
= sd
->groups
;
5109 unsigned long power
, power_orig
;
5110 unsigned long interval
;
5112 interval
= msecs_to_jiffies(sd
->balance_interval
);
5113 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5114 sdg
->sgp
->next_update
= jiffies
+ interval
;
5117 update_cpu_power(sd
, cpu
);
5121 power_orig
= power
= 0;
5123 if (child
->flags
& SD_OVERLAP
) {
5125 * SD_OVERLAP domains cannot assume that child groups
5126 * span the current group.
5129 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
5130 struct sched_group
*sg
= cpu_rq(cpu
)->sd
->groups
;
5132 power_orig
+= sg
->sgp
->power_orig
;
5133 power
+= sg
->sgp
->power
;
5137 * !SD_OVERLAP domains can assume that child groups
5138 * span the current group.
5141 group
= child
->groups
;
5143 power_orig
+= group
->sgp
->power_orig
;
5144 power
+= group
->sgp
->power
;
5145 group
= group
->next
;
5146 } while (group
!= child
->groups
);
5149 sdg
->sgp
->power_orig
= power_orig
;
5150 sdg
->sgp
->power
= power
;
5154 * Try and fix up capacity for tiny siblings, this is needed when
5155 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5156 * which on its own isn't powerful enough.
5158 * See update_sd_pick_busiest() and check_asym_packing().
5161 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
5164 * Only siblings can have significantly less than SCHED_POWER_SCALE
5166 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
5170 * If ~90% of the cpu_power is still there, we're good.
5172 if (group
->sgp
->power
* 32 > group
->sgp
->power_orig
* 29)
5179 * Group imbalance indicates (and tries to solve) the problem where balancing
5180 * groups is inadequate due to tsk_cpus_allowed() constraints.
5182 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5183 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5186 * { 0 1 2 3 } { 4 5 6 7 }
5189 * If we were to balance group-wise we'd place two tasks in the first group and
5190 * two tasks in the second group. Clearly this is undesired as it will overload
5191 * cpu 3 and leave one of the cpus in the second group unused.
5193 * The current solution to this issue is detecting the skew in the first group
5194 * by noticing the lower domain failed to reach balance and had difficulty
5195 * moving tasks due to affinity constraints.
5197 * When this is so detected; this group becomes a candidate for busiest; see
5198 * update_sd_pick_busiest(). And calculcate_imbalance() and
5199 * find_busiest_group() avoid some of the usual balance conditions to allow it
5200 * to create an effective group imbalance.
5202 * This is a somewhat tricky proposition since the next run might not find the
5203 * group imbalance and decide the groups need to be balanced again. A most
5204 * subtle and fragile situation.
5207 static inline int sg_imbalanced(struct sched_group
*group
)
5209 return group
->sgp
->imbalance
;
5213 * Compute the group capacity.
5215 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5216 * first dividing out the smt factor and computing the actual number of cores
5217 * and limit power unit capacity with that.
5219 static inline int sg_capacity(struct lb_env
*env
, struct sched_group
*group
)
5221 unsigned int capacity
, smt
, cpus
;
5222 unsigned int power
, power_orig
;
5224 power
= group
->sgp
->power
;
5225 power_orig
= group
->sgp
->power_orig
;
5226 cpus
= group
->group_weight
;
5228 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5229 smt
= DIV_ROUND_UP(SCHED_POWER_SCALE
* cpus
, power_orig
);
5230 capacity
= cpus
/ smt
; /* cores */
5232 capacity
= min_t(unsigned, capacity
, DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
));
5234 capacity
= fix_small_capacity(env
->sd
, group
);
5240 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5241 * @env: The load balancing environment.
5242 * @group: sched_group whose statistics are to be updated.
5243 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5244 * @local_group: Does group contain this_cpu.
5245 * @sgs: variable to hold the statistics for this group.
5247 static inline void update_sg_lb_stats(struct lb_env
*env
,
5248 struct sched_group
*group
, int load_idx
,
5249 int local_group
, struct sg_lb_stats
*sgs
)
5251 unsigned long nr_running
;
5255 memset(sgs
, 0, sizeof(*sgs
));
5257 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5258 struct rq
*rq
= cpu_rq(i
);
5260 nr_running
= rq
->nr_running
;
5262 /* Bias balancing toward cpus of our domain */
5264 load
= target_load(i
, load_idx
);
5266 load
= source_load(i
, load_idx
);
5268 sgs
->group_load
+= load
;
5269 sgs
->sum_nr_running
+= nr_running
;
5270 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
5275 /* Adjust by relative CPU power of the group */
5276 sgs
->group_power
= group
->sgp
->power
;
5277 sgs
->avg_load
= (sgs
->group_load
*SCHED_POWER_SCALE
) / sgs
->group_power
;
5279 if (sgs
->sum_nr_running
)
5280 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
5282 sgs
->group_weight
= group
->group_weight
;
5284 sgs
->group_imb
= sg_imbalanced(group
);
5285 sgs
->group_capacity
= sg_capacity(env
, group
);
5287 if (sgs
->group_capacity
> sgs
->sum_nr_running
)
5288 sgs
->group_has_capacity
= 1;
5292 * update_sd_pick_busiest - return 1 on busiest group
5293 * @env: The load balancing environment.
5294 * @sds: sched_domain statistics
5295 * @sg: sched_group candidate to be checked for being the busiest
5296 * @sgs: sched_group statistics
5298 * Determine if @sg is a busier group than the previously selected
5301 * Return: %true if @sg is a busier group than the previously selected
5302 * busiest group. %false otherwise.
5304 static bool update_sd_pick_busiest(struct lb_env
*env
,
5305 struct sd_lb_stats
*sds
,
5306 struct sched_group
*sg
,
5307 struct sg_lb_stats
*sgs
)
5309 if (sgs
->avg_load
<= sds
->busiest_stat
.avg_load
)
5312 if (sgs
->sum_nr_running
> sgs
->group_capacity
)
5319 * ASYM_PACKING needs to move all the work to the lowest
5320 * numbered CPUs in the group, therefore mark all groups
5321 * higher than ourself as busy.
5323 if ((env
->sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
5324 env
->dst_cpu
< group_first_cpu(sg
)) {
5328 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
5336 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5337 * @env: The load balancing environment.
5338 * @balance: Should we balance.
5339 * @sds: variable to hold the statistics for this sched_domain.
5341 static inline void update_sd_lb_stats(struct lb_env
*env
,
5342 struct sd_lb_stats
*sds
)
5344 struct sched_domain
*child
= env
->sd
->child
;
5345 struct sched_group
*sg
= env
->sd
->groups
;
5346 struct sg_lb_stats tmp_sgs
;
5347 int load_idx
, prefer_sibling
= 0;
5349 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
5352 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
5355 struct sg_lb_stats
*sgs
= &tmp_sgs
;
5358 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
5361 sgs
= &sds
->local_stat
;
5363 if (env
->idle
!= CPU_NEWLY_IDLE
||
5364 time_after_eq(jiffies
, sg
->sgp
->next_update
))
5365 update_group_power(env
->sd
, env
->dst_cpu
);
5368 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
);
5374 * In case the child domain prefers tasks go to siblings
5375 * first, lower the sg capacity to one so that we'll try
5376 * and move all the excess tasks away. We lower the capacity
5377 * of a group only if the local group has the capacity to fit
5378 * these excess tasks, i.e. nr_running < group_capacity. The
5379 * extra check prevents the case where you always pull from the
5380 * heaviest group when it is already under-utilized (possible
5381 * with a large weight task outweighs the tasks on the system).
5383 if (prefer_sibling
&& sds
->local
&&
5384 sds
->local_stat
.group_has_capacity
)
5385 sgs
->group_capacity
= min(sgs
->group_capacity
, 1U);
5387 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
5389 sds
->busiest_stat
= *sgs
;
5393 /* Now, start updating sd_lb_stats */
5394 sds
->total_load
+= sgs
->group_load
;
5395 sds
->total_pwr
+= sgs
->group_power
;
5398 } while (sg
!= env
->sd
->groups
);
5402 * check_asym_packing - Check to see if the group is packed into the
5405 * This is primarily intended to used at the sibling level. Some
5406 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5407 * case of POWER7, it can move to lower SMT modes only when higher
5408 * threads are idle. When in lower SMT modes, the threads will
5409 * perform better since they share less core resources. Hence when we
5410 * have idle threads, we want them to be the higher ones.
5412 * This packing function is run on idle threads. It checks to see if
5413 * the busiest CPU in this domain (core in the P7 case) has a higher
5414 * CPU number than the packing function is being run on. Here we are
5415 * assuming lower CPU number will be equivalent to lower a SMT thread
5418 * Return: 1 when packing is required and a task should be moved to
5419 * this CPU. The amount of the imbalance is returned in *imbalance.
5421 * @env: The load balancing environment.
5422 * @sds: Statistics of the sched_domain which is to be packed
5424 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5428 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
5434 busiest_cpu
= group_first_cpu(sds
->busiest
);
5435 if (env
->dst_cpu
> busiest_cpu
)
5438 env
->imbalance
= DIV_ROUND_CLOSEST(
5439 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_power
,
5446 * fix_small_imbalance - Calculate the minor imbalance that exists
5447 * amongst the groups of a sched_domain, during
5449 * @env: The load balancing environment.
5450 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5453 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5455 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
5456 unsigned int imbn
= 2;
5457 unsigned long scaled_busy_load_per_task
;
5458 struct sg_lb_stats
*local
, *busiest
;
5460 local
= &sds
->local_stat
;
5461 busiest
= &sds
->busiest_stat
;
5463 if (!local
->sum_nr_running
)
5464 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
5465 else if (busiest
->load_per_task
> local
->load_per_task
)
5468 scaled_busy_load_per_task
=
5469 (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5470 busiest
->group_power
;
5472 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
5473 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
5474 env
->imbalance
= busiest
->load_per_task
;
5479 * OK, we don't have enough imbalance to justify moving tasks,
5480 * however we may be able to increase total CPU power used by
5484 pwr_now
+= busiest
->group_power
*
5485 min(busiest
->load_per_task
, busiest
->avg_load
);
5486 pwr_now
+= local
->group_power
*
5487 min(local
->load_per_task
, local
->avg_load
);
5488 pwr_now
/= SCHED_POWER_SCALE
;
5490 /* Amount of load we'd subtract */
5491 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5492 busiest
->group_power
;
5493 if (busiest
->avg_load
> tmp
) {
5494 pwr_move
+= busiest
->group_power
*
5495 min(busiest
->load_per_task
,
5496 busiest
->avg_load
- tmp
);
5499 /* Amount of load we'd add */
5500 if (busiest
->avg_load
* busiest
->group_power
<
5501 busiest
->load_per_task
* SCHED_POWER_SCALE
) {
5502 tmp
= (busiest
->avg_load
* busiest
->group_power
) /
5505 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5508 pwr_move
+= local
->group_power
*
5509 min(local
->load_per_task
, local
->avg_load
+ tmp
);
5510 pwr_move
/= SCHED_POWER_SCALE
;
5512 /* Move if we gain throughput */
5513 if (pwr_move
> pwr_now
)
5514 env
->imbalance
= busiest
->load_per_task
;
5518 * calculate_imbalance - Calculate the amount of imbalance present within the
5519 * groups of a given sched_domain during load balance.
5520 * @env: load balance environment
5521 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5523 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5525 unsigned long max_pull
, load_above_capacity
= ~0UL;
5526 struct sg_lb_stats
*local
, *busiest
;
5528 local
= &sds
->local_stat
;
5529 busiest
= &sds
->busiest_stat
;
5531 if (busiest
->group_imb
) {
5533 * In the group_imb case we cannot rely on group-wide averages
5534 * to ensure cpu-load equilibrium, look at wider averages. XXX
5536 busiest
->load_per_task
=
5537 min(busiest
->load_per_task
, sds
->avg_load
);
5541 * In the presence of smp nice balancing, certain scenarios can have
5542 * max load less than avg load(as we skip the groups at or below
5543 * its cpu_power, while calculating max_load..)
5545 if (busiest
->avg_load
<= sds
->avg_load
||
5546 local
->avg_load
>= sds
->avg_load
) {
5548 return fix_small_imbalance(env
, sds
);
5551 if (!busiest
->group_imb
) {
5553 * Don't want to pull so many tasks that a group would go idle.
5554 * Except of course for the group_imb case, since then we might
5555 * have to drop below capacity to reach cpu-load equilibrium.
5557 load_above_capacity
=
5558 (busiest
->sum_nr_running
- busiest
->group_capacity
);
5560 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_POWER_SCALE
);
5561 load_above_capacity
/= busiest
->group_power
;
5565 * We're trying to get all the cpus to the average_load, so we don't
5566 * want to push ourselves above the average load, nor do we wish to
5567 * reduce the max loaded cpu below the average load. At the same time,
5568 * we also don't want to reduce the group load below the group capacity
5569 * (so that we can implement power-savings policies etc). Thus we look
5570 * for the minimum possible imbalance.
5572 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
5574 /* How much load to actually move to equalise the imbalance */
5575 env
->imbalance
= min(
5576 max_pull
* busiest
->group_power
,
5577 (sds
->avg_load
- local
->avg_load
) * local
->group_power
5578 ) / SCHED_POWER_SCALE
;
5581 * if *imbalance is less than the average load per runnable task
5582 * there is no guarantee that any tasks will be moved so we'll have
5583 * a think about bumping its value to force at least one task to be
5586 if (env
->imbalance
< busiest
->load_per_task
)
5587 return fix_small_imbalance(env
, sds
);
5590 /******* find_busiest_group() helpers end here *********************/
5593 * find_busiest_group - Returns the busiest group within the sched_domain
5594 * if there is an imbalance. If there isn't an imbalance, and
5595 * the user has opted for power-savings, it returns a group whose
5596 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5597 * such a group exists.
5599 * Also calculates the amount of weighted load which should be moved
5600 * to restore balance.
5602 * @env: The load balancing environment.
5604 * Return: - The busiest group if imbalance exists.
5605 * - If no imbalance and user has opted for power-savings balance,
5606 * return the least loaded group whose CPUs can be
5607 * put to idle by rebalancing its tasks onto our group.
5609 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
5611 struct sg_lb_stats
*local
, *busiest
;
5612 struct sd_lb_stats sds
;
5614 init_sd_lb_stats(&sds
);
5617 * Compute the various statistics relavent for load balancing at
5620 update_sd_lb_stats(env
, &sds
);
5621 local
= &sds
.local_stat
;
5622 busiest
= &sds
.busiest_stat
;
5624 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
5625 check_asym_packing(env
, &sds
))
5628 /* There is no busy sibling group to pull tasks from */
5629 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
5632 sds
.avg_load
= (SCHED_POWER_SCALE
* sds
.total_load
) / sds
.total_pwr
;
5635 * If the busiest group is imbalanced the below checks don't
5636 * work because they assume all things are equal, which typically
5637 * isn't true due to cpus_allowed constraints and the like.
5639 if (busiest
->group_imb
)
5642 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5643 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_capacity
&&
5644 !busiest
->group_has_capacity
)
5648 * If the local group is more busy than the selected busiest group
5649 * don't try and pull any tasks.
5651 if (local
->avg_load
>= busiest
->avg_load
)
5655 * Don't pull any tasks if this group is already above the domain
5658 if (local
->avg_load
>= sds
.avg_load
)
5661 if (env
->idle
== CPU_IDLE
) {
5663 * This cpu is idle. If the busiest group load doesn't
5664 * have more tasks than the number of available cpu's and
5665 * there is no imbalance between this and busiest group
5666 * wrt to idle cpu's, it is balanced.
5668 if ((local
->idle_cpus
< busiest
->idle_cpus
) &&
5669 busiest
->sum_nr_running
<= busiest
->group_weight
)
5673 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5674 * imbalance_pct to be conservative.
5676 if (100 * busiest
->avg_load
<=
5677 env
->sd
->imbalance_pct
* local
->avg_load
)
5682 /* Looks like there is an imbalance. Compute it */
5683 calculate_imbalance(env
, &sds
);
5692 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5694 static struct rq
*find_busiest_queue(struct lb_env
*env
,
5695 struct sched_group
*group
)
5697 struct rq
*busiest
= NULL
, *rq
;
5698 unsigned long busiest_load
= 0, busiest_power
= 1;
5701 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5702 unsigned long power
= power_of(i
);
5703 unsigned long capacity
= DIV_ROUND_CLOSEST(power
,
5708 capacity
= fix_small_capacity(env
->sd
, group
);
5711 wl
= weighted_cpuload(i
);
5714 * When comparing with imbalance, use weighted_cpuload()
5715 * which is not scaled with the cpu power.
5717 if (capacity
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
5721 * For the load comparisons with the other cpu's, consider
5722 * the weighted_cpuload() scaled with the cpu power, so that
5723 * the load can be moved away from the cpu that is potentially
5724 * running at a lower capacity.
5726 * Thus we're looking for max(wl_i / power_i), crosswise
5727 * multiplication to rid ourselves of the division works out
5728 * to: wl_i * power_j > wl_j * power_i; where j is our
5731 if (wl
* busiest_power
> busiest_load
* power
) {
5733 busiest_power
= power
;
5742 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5743 * so long as it is large enough.
5745 #define MAX_PINNED_INTERVAL 512
5747 /* Working cpumask for load_balance and load_balance_newidle. */
5748 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5750 static int need_active_balance(struct lb_env
*env
)
5752 struct sched_domain
*sd
= env
->sd
;
5754 if (env
->idle
== CPU_NEWLY_IDLE
) {
5757 * ASYM_PACKING needs to force migrate tasks from busy but
5758 * higher numbered CPUs in order to pack all tasks in the
5759 * lowest numbered CPUs.
5761 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
5765 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
5768 static int active_load_balance_cpu_stop(void *data
);
5770 static int should_we_balance(struct lb_env
*env
)
5772 struct sched_group
*sg
= env
->sd
->groups
;
5773 struct cpumask
*sg_cpus
, *sg_mask
;
5774 int cpu
, balance_cpu
= -1;
5777 * In the newly idle case, we will allow all the cpu's
5778 * to do the newly idle load balance.
5780 if (env
->idle
== CPU_NEWLY_IDLE
)
5783 sg_cpus
= sched_group_cpus(sg
);
5784 sg_mask
= sched_group_mask(sg
);
5785 /* Try to find first idle cpu */
5786 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
5787 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
5794 if (balance_cpu
== -1)
5795 balance_cpu
= group_balance_cpu(sg
);
5798 * First idle cpu or the first cpu(busiest) in this sched group
5799 * is eligible for doing load balancing at this and above domains.
5801 return balance_cpu
== env
->dst_cpu
;
5805 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5806 * tasks if there is an imbalance.
5808 static int load_balance(int this_cpu
, struct rq
*this_rq
,
5809 struct sched_domain
*sd
, enum cpu_idle_type idle
,
5810 int *continue_balancing
)
5812 int ld_moved
, cur_ld_moved
, active_balance
= 0;
5813 struct sched_domain
*sd_parent
= sd
->parent
;
5814 struct sched_group
*group
;
5816 unsigned long flags
;
5817 struct cpumask
*cpus
= __get_cpu_var(load_balance_mask
);
5819 struct lb_env env
= {
5821 .dst_cpu
= this_cpu
,
5823 .dst_grpmask
= sched_group_cpus(sd
->groups
),
5825 .loop_break
= sched_nr_migrate_break
,
5830 * For NEWLY_IDLE load_balancing, we don't need to consider
5831 * other cpus in our group
5833 if (idle
== CPU_NEWLY_IDLE
)
5834 env
.dst_grpmask
= NULL
;
5836 cpumask_copy(cpus
, cpu_active_mask
);
5838 schedstat_inc(sd
, lb_count
[idle
]);
5841 if (!should_we_balance(&env
)) {
5842 *continue_balancing
= 0;
5846 group
= find_busiest_group(&env
);
5848 schedstat_inc(sd
, lb_nobusyg
[idle
]);
5852 busiest
= find_busiest_queue(&env
, group
);
5854 schedstat_inc(sd
, lb_nobusyq
[idle
]);
5858 BUG_ON(busiest
== env
.dst_rq
);
5860 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
5863 if (busiest
->nr_running
> 1) {
5865 * Attempt to move tasks. If find_busiest_group has found
5866 * an imbalance but busiest->nr_running <= 1, the group is
5867 * still unbalanced. ld_moved simply stays zero, so it is
5868 * correctly treated as an imbalance.
5870 env
.flags
|= LBF_ALL_PINNED
;
5871 env
.src_cpu
= busiest
->cpu
;
5872 env
.src_rq
= busiest
;
5873 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
5876 local_irq_save(flags
);
5877 double_rq_lock(env
.dst_rq
, busiest
);
5880 * cur_ld_moved - load moved in current iteration
5881 * ld_moved - cumulative load moved across iterations
5883 cur_ld_moved
= move_tasks(&env
);
5884 ld_moved
+= cur_ld_moved
;
5885 double_rq_unlock(env
.dst_rq
, busiest
);
5886 local_irq_restore(flags
);
5889 * some other cpu did the load balance for us.
5891 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
5892 resched_cpu(env
.dst_cpu
);
5894 if (env
.flags
& LBF_NEED_BREAK
) {
5895 env
.flags
&= ~LBF_NEED_BREAK
;
5900 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5901 * us and move them to an alternate dst_cpu in our sched_group
5902 * where they can run. The upper limit on how many times we
5903 * iterate on same src_cpu is dependent on number of cpus in our
5906 * This changes load balance semantics a bit on who can move
5907 * load to a given_cpu. In addition to the given_cpu itself
5908 * (or a ilb_cpu acting on its behalf where given_cpu is
5909 * nohz-idle), we now have balance_cpu in a position to move
5910 * load to given_cpu. In rare situations, this may cause
5911 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5912 * _independently_ and at _same_ time to move some load to
5913 * given_cpu) causing exceess load to be moved to given_cpu.
5914 * This however should not happen so much in practice and
5915 * moreover subsequent load balance cycles should correct the
5916 * excess load moved.
5918 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
5920 /* Prevent to re-select dst_cpu via env's cpus */
5921 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
5923 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
5924 env
.dst_cpu
= env
.new_dst_cpu
;
5925 env
.flags
&= ~LBF_DST_PINNED
;
5927 env
.loop_break
= sched_nr_migrate_break
;
5930 * Go back to "more_balance" rather than "redo" since we
5931 * need to continue with same src_cpu.
5937 * We failed to reach balance because of affinity.
5940 int *group_imbalance
= &sd_parent
->groups
->sgp
->imbalance
;
5942 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0) {
5943 *group_imbalance
= 1;
5944 } else if (*group_imbalance
)
5945 *group_imbalance
= 0;
5948 /* All tasks on this runqueue were pinned by CPU affinity */
5949 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
5950 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
5951 if (!cpumask_empty(cpus
)) {
5953 env
.loop_break
= sched_nr_migrate_break
;
5961 schedstat_inc(sd
, lb_failed
[idle
]);
5963 * Increment the failure counter only on periodic balance.
5964 * We do not want newidle balance, which can be very
5965 * frequent, pollute the failure counter causing
5966 * excessive cache_hot migrations and active balances.
5968 if (idle
!= CPU_NEWLY_IDLE
)
5969 sd
->nr_balance_failed
++;
5971 if (need_active_balance(&env
)) {
5972 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
5974 /* don't kick the active_load_balance_cpu_stop,
5975 * if the curr task on busiest cpu can't be
5978 if (!cpumask_test_cpu(this_cpu
,
5979 tsk_cpus_allowed(busiest
->curr
))) {
5980 raw_spin_unlock_irqrestore(&busiest
->lock
,
5982 env
.flags
|= LBF_ALL_PINNED
;
5983 goto out_one_pinned
;
5987 * ->active_balance synchronizes accesses to
5988 * ->active_balance_work. Once set, it's cleared
5989 * only after active load balance is finished.
5991 if (!busiest
->active_balance
) {
5992 busiest
->active_balance
= 1;
5993 busiest
->push_cpu
= this_cpu
;
5996 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
5998 if (active_balance
) {
5999 stop_one_cpu_nowait(cpu_of(busiest
),
6000 active_load_balance_cpu_stop
, busiest
,
6001 &busiest
->active_balance_work
);
6005 * We've kicked active balancing, reset the failure
6008 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
6011 sd
->nr_balance_failed
= 0;
6013 if (likely(!active_balance
)) {
6014 /* We were unbalanced, so reset the balancing interval */
6015 sd
->balance_interval
= sd
->min_interval
;
6018 * If we've begun active balancing, start to back off. This
6019 * case may not be covered by the all_pinned logic if there
6020 * is only 1 task on the busy runqueue (because we don't call
6023 if (sd
->balance_interval
< sd
->max_interval
)
6024 sd
->balance_interval
*= 2;
6030 schedstat_inc(sd
, lb_balanced
[idle
]);
6032 sd
->nr_balance_failed
= 0;
6035 /* tune up the balancing interval */
6036 if (((env
.flags
& LBF_ALL_PINNED
) &&
6037 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
6038 (sd
->balance_interval
< sd
->max_interval
))
6039 sd
->balance_interval
*= 2;
6047 * idle_balance is called by schedule() if this_cpu is about to become
6048 * idle. Attempts to pull tasks from other CPUs.
6050 void idle_balance(int this_cpu
, struct rq
*this_rq
)
6052 struct sched_domain
*sd
;
6053 int pulled_task
= 0;
6054 unsigned long next_balance
= jiffies
+ HZ
;
6057 this_rq
->idle_stamp
= rq_clock(this_rq
);
6059 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
6063 * Drop the rq->lock, but keep IRQ/preempt disabled.
6065 raw_spin_unlock(&this_rq
->lock
);
6067 update_blocked_averages(this_cpu
);
6069 for_each_domain(this_cpu
, sd
) {
6070 unsigned long interval
;
6071 int continue_balancing
= 1;
6072 u64 t0
, domain_cost
;
6074 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6077 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
)
6080 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
6081 t0
= sched_clock_cpu(this_cpu
);
6083 /* If we've pulled tasks over stop searching: */
6084 pulled_task
= load_balance(this_cpu
, this_rq
,
6086 &continue_balancing
);
6088 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
6089 if (domain_cost
> sd
->max_newidle_lb_cost
)
6090 sd
->max_newidle_lb_cost
= domain_cost
;
6092 curr_cost
+= domain_cost
;
6095 interval
= msecs_to_jiffies(sd
->balance_interval
);
6096 if (time_after(next_balance
, sd
->last_balance
+ interval
))
6097 next_balance
= sd
->last_balance
+ interval
;
6099 this_rq
->idle_stamp
= 0;
6105 raw_spin_lock(&this_rq
->lock
);
6107 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
6109 * We are going idle. next_balance may be set based on
6110 * a busy processor. So reset next_balance.
6112 this_rq
->next_balance
= next_balance
;
6115 if (curr_cost
> this_rq
->max_idle_balance_cost
)
6116 this_rq
->max_idle_balance_cost
= curr_cost
;
6120 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6121 * running tasks off the busiest CPU onto idle CPUs. It requires at
6122 * least 1 task to be running on each physical CPU where possible, and
6123 * avoids physical / logical imbalances.
6125 static int active_load_balance_cpu_stop(void *data
)
6127 struct rq
*busiest_rq
= data
;
6128 int busiest_cpu
= cpu_of(busiest_rq
);
6129 int target_cpu
= busiest_rq
->push_cpu
;
6130 struct rq
*target_rq
= cpu_rq(target_cpu
);
6131 struct sched_domain
*sd
;
6133 raw_spin_lock_irq(&busiest_rq
->lock
);
6135 /* make sure the requested cpu hasn't gone down in the meantime */
6136 if (unlikely(busiest_cpu
!= smp_processor_id() ||
6137 !busiest_rq
->active_balance
))
6140 /* Is there any task to move? */
6141 if (busiest_rq
->nr_running
<= 1)
6145 * This condition is "impossible", if it occurs
6146 * we need to fix it. Originally reported by
6147 * Bjorn Helgaas on a 128-cpu setup.
6149 BUG_ON(busiest_rq
== target_rq
);
6151 /* move a task from busiest_rq to target_rq */
6152 double_lock_balance(busiest_rq
, target_rq
);
6154 /* Search for an sd spanning us and the target CPU. */
6156 for_each_domain(target_cpu
, sd
) {
6157 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
6158 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
6163 struct lb_env env
= {
6165 .dst_cpu
= target_cpu
,
6166 .dst_rq
= target_rq
,
6167 .src_cpu
= busiest_rq
->cpu
,
6168 .src_rq
= busiest_rq
,
6172 schedstat_inc(sd
, alb_count
);
6174 if (move_one_task(&env
))
6175 schedstat_inc(sd
, alb_pushed
);
6177 schedstat_inc(sd
, alb_failed
);
6180 double_unlock_balance(busiest_rq
, target_rq
);
6182 busiest_rq
->active_balance
= 0;
6183 raw_spin_unlock_irq(&busiest_rq
->lock
);
6187 #ifdef CONFIG_NO_HZ_COMMON
6189 * idle load balancing details
6190 * - When one of the busy CPUs notice that there may be an idle rebalancing
6191 * needed, they will kick the idle load balancer, which then does idle
6192 * load balancing for all the idle CPUs.
6195 cpumask_var_t idle_cpus_mask
;
6197 unsigned long next_balance
; /* in jiffy units */
6198 } nohz ____cacheline_aligned
;
6200 static inline int find_new_ilb(int call_cpu
)
6202 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
6204 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
6211 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6212 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6213 * CPU (if there is one).
6215 static void nohz_balancer_kick(int cpu
)
6219 nohz
.next_balance
++;
6221 ilb_cpu
= find_new_ilb(cpu
);
6223 if (ilb_cpu
>= nr_cpu_ids
)
6226 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
6229 * Use smp_send_reschedule() instead of resched_cpu().
6230 * This way we generate a sched IPI on the target cpu which
6231 * is idle. And the softirq performing nohz idle load balance
6232 * will be run before returning from the IPI.
6234 smp_send_reschedule(ilb_cpu
);
6238 static inline void nohz_balance_exit_idle(int cpu
)
6240 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
6241 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
6242 atomic_dec(&nohz
.nr_cpus
);
6243 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
6247 static inline void set_cpu_sd_state_busy(void)
6249 struct sched_domain
*sd
;
6252 sd
= rcu_dereference_check_sched_domain(this_rq()->sd
);
6254 if (!sd
|| !sd
->nohz_idle
)
6258 for (; sd
; sd
= sd
->parent
)
6259 atomic_inc(&sd
->groups
->sgp
->nr_busy_cpus
);
6264 void set_cpu_sd_state_idle(void)
6266 struct sched_domain
*sd
;
6269 sd
= rcu_dereference_check_sched_domain(this_rq()->sd
);
6271 if (!sd
|| sd
->nohz_idle
)
6275 for (; sd
; sd
= sd
->parent
)
6276 atomic_dec(&sd
->groups
->sgp
->nr_busy_cpus
);
6282 * This routine will record that the cpu is going idle with tick stopped.
6283 * This info will be used in performing idle load balancing in the future.
6285 void nohz_balance_enter_idle(int cpu
)
6288 * If this cpu is going down, then nothing needs to be done.
6290 if (!cpu_active(cpu
))
6293 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
6296 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
6297 atomic_inc(&nohz
.nr_cpus
);
6298 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
6301 static int sched_ilb_notifier(struct notifier_block
*nfb
,
6302 unsigned long action
, void *hcpu
)
6304 switch (action
& ~CPU_TASKS_FROZEN
) {
6306 nohz_balance_exit_idle(smp_processor_id());
6314 static DEFINE_SPINLOCK(balancing
);
6317 * Scale the max load_balance interval with the number of CPUs in the system.
6318 * This trades load-balance latency on larger machines for less cross talk.
6320 void update_max_interval(void)
6322 max_load_balance_interval
= HZ
*num_online_cpus()/10;
6326 * It checks each scheduling domain to see if it is due to be balanced,
6327 * and initiates a balancing operation if so.
6329 * Balancing parameters are set up in init_sched_domains.
6331 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
6333 int continue_balancing
= 1;
6334 struct rq
*rq
= cpu_rq(cpu
);
6335 unsigned long interval
;
6336 struct sched_domain
*sd
;
6337 /* Earliest time when we have to do rebalance again */
6338 unsigned long next_balance
= jiffies
+ 60*HZ
;
6339 int update_next_balance
= 0;
6340 int need_serialize
, need_decay
= 0;
6343 update_blocked_averages(cpu
);
6346 for_each_domain(cpu
, sd
) {
6348 * Decay the newidle max times here because this is a regular
6349 * visit to all the domains. Decay ~1% per second.
6351 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
6352 sd
->max_newidle_lb_cost
=
6353 (sd
->max_newidle_lb_cost
* 253) / 256;
6354 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
6357 max_cost
+= sd
->max_newidle_lb_cost
;
6359 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6363 * Stop the load balance at this level. There is another
6364 * CPU in our sched group which is doing load balancing more
6367 if (!continue_balancing
) {
6373 interval
= sd
->balance_interval
;
6374 if (idle
!= CPU_IDLE
)
6375 interval
*= sd
->busy_factor
;
6377 /* scale ms to jiffies */
6378 interval
= msecs_to_jiffies(interval
);
6379 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6381 need_serialize
= sd
->flags
& SD_SERIALIZE
;
6383 if (need_serialize
) {
6384 if (!spin_trylock(&balancing
))
6388 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
6389 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
6391 * The LBF_DST_PINNED logic could have changed
6392 * env->dst_cpu, so we can't know our idle
6393 * state even if we migrated tasks. Update it.
6395 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
6397 sd
->last_balance
= jiffies
;
6400 spin_unlock(&balancing
);
6402 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
6403 next_balance
= sd
->last_balance
+ interval
;
6404 update_next_balance
= 1;
6409 * Ensure the rq-wide value also decays but keep it at a
6410 * reasonable floor to avoid funnies with rq->avg_idle.
6412 rq
->max_idle_balance_cost
=
6413 max((u64
)sysctl_sched_migration_cost
, max_cost
);
6418 * next_balance will be updated only when there is a need.
6419 * When the cpu is attached to null domain for ex, it will not be
6422 if (likely(update_next_balance
))
6423 rq
->next_balance
= next_balance
;
6426 #ifdef CONFIG_NO_HZ_COMMON
6428 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6429 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6431 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
)
6433 struct rq
*this_rq
= cpu_rq(this_cpu
);
6437 if (idle
!= CPU_IDLE
||
6438 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
6441 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
6442 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
6446 * If this cpu gets work to do, stop the load balancing
6447 * work being done for other cpus. Next load
6448 * balancing owner will pick it up.
6453 rq
= cpu_rq(balance_cpu
);
6455 raw_spin_lock_irq(&rq
->lock
);
6456 update_rq_clock(rq
);
6457 update_idle_cpu_load(rq
);
6458 raw_spin_unlock_irq(&rq
->lock
);
6460 rebalance_domains(balance_cpu
, CPU_IDLE
);
6462 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
6463 this_rq
->next_balance
= rq
->next_balance
;
6465 nohz
.next_balance
= this_rq
->next_balance
;
6467 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
6471 * Current heuristic for kicking the idle load balancer in the presence
6472 * of an idle cpu is the system.
6473 * - This rq has more than one task.
6474 * - At any scheduler domain level, this cpu's scheduler group has multiple
6475 * busy cpu's exceeding the group's power.
6476 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6477 * domain span are idle.
6479 static inline int nohz_kick_needed(struct rq
*rq
, int cpu
)
6481 unsigned long now
= jiffies
;
6482 struct sched_domain
*sd
;
6484 if (unlikely(idle_cpu(cpu
)))
6488 * We may be recently in ticked or tickless idle mode. At the first
6489 * busy tick after returning from idle, we will update the busy stats.
6491 set_cpu_sd_state_busy();
6492 nohz_balance_exit_idle(cpu
);
6495 * None are in tickless mode and hence no need for NOHZ idle load
6498 if (likely(!atomic_read(&nohz
.nr_cpus
)))
6501 if (time_before(now
, nohz
.next_balance
))
6504 if (rq
->nr_running
>= 2)
6508 for_each_domain(cpu
, sd
) {
6509 struct sched_group
*sg
= sd
->groups
;
6510 struct sched_group_power
*sgp
= sg
->sgp
;
6511 int nr_busy
= atomic_read(&sgp
->nr_busy_cpus
);
6513 if (sd
->flags
& SD_SHARE_PKG_RESOURCES
&& nr_busy
> 1)
6514 goto need_kick_unlock
;
6516 if (sd
->flags
& SD_ASYM_PACKING
&& nr_busy
!= sg
->group_weight
6517 && (cpumask_first_and(nohz
.idle_cpus_mask
,
6518 sched_domain_span(sd
)) < cpu
))
6519 goto need_kick_unlock
;
6521 if (!(sd
->flags
& (SD_SHARE_PKG_RESOURCES
| SD_ASYM_PACKING
)))
6533 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
) { }
6537 * run_rebalance_domains is triggered when needed from the scheduler tick.
6538 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6540 static void run_rebalance_domains(struct softirq_action
*h
)
6542 int this_cpu
= smp_processor_id();
6543 struct rq
*this_rq
= cpu_rq(this_cpu
);
6544 enum cpu_idle_type idle
= this_rq
->idle_balance
?
6545 CPU_IDLE
: CPU_NOT_IDLE
;
6547 rebalance_domains(this_cpu
, idle
);
6550 * If this cpu has a pending nohz_balance_kick, then do the
6551 * balancing on behalf of the other idle cpus whose ticks are
6554 nohz_idle_balance(this_cpu
, idle
);
6557 static inline int on_null_domain(int cpu
)
6559 return !rcu_dereference_sched(cpu_rq(cpu
)->sd
);
6563 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6565 void trigger_load_balance(struct rq
*rq
, int cpu
)
6567 /* Don't need to rebalance while attached to NULL domain */
6568 if (time_after_eq(jiffies
, rq
->next_balance
) &&
6569 likely(!on_null_domain(cpu
)))
6570 raise_softirq(SCHED_SOFTIRQ
);
6571 #ifdef CONFIG_NO_HZ_COMMON
6572 if (nohz_kick_needed(rq
, cpu
) && likely(!on_null_domain(cpu
)))
6573 nohz_balancer_kick(cpu
);
6577 static void rq_online_fair(struct rq
*rq
)
6582 static void rq_offline_fair(struct rq
*rq
)
6586 /* Ensure any throttled groups are reachable by pick_next_task */
6587 unthrottle_offline_cfs_rqs(rq
);
6590 #endif /* CONFIG_SMP */
6593 * scheduler tick hitting a task of our scheduling class:
6595 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
6597 struct cfs_rq
*cfs_rq
;
6598 struct sched_entity
*se
= &curr
->se
;
6600 for_each_sched_entity(se
) {
6601 cfs_rq
= cfs_rq_of(se
);
6602 entity_tick(cfs_rq
, se
, queued
);
6605 if (numabalancing_enabled
)
6606 task_tick_numa(rq
, curr
);
6608 update_rq_runnable_avg(rq
, 1);
6612 * called on fork with the child task as argument from the parent's context
6613 * - child not yet on the tasklist
6614 * - preemption disabled
6616 static void task_fork_fair(struct task_struct
*p
)
6618 struct cfs_rq
*cfs_rq
;
6619 struct sched_entity
*se
= &p
->se
, *curr
;
6620 int this_cpu
= smp_processor_id();
6621 struct rq
*rq
= this_rq();
6622 unsigned long flags
;
6624 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6626 update_rq_clock(rq
);
6628 cfs_rq
= task_cfs_rq(current
);
6629 curr
= cfs_rq
->curr
;
6632 * Not only the cpu but also the task_group of the parent might have
6633 * been changed after parent->se.parent,cfs_rq were copied to
6634 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6635 * of child point to valid ones.
6638 __set_task_cpu(p
, this_cpu
);
6641 update_curr(cfs_rq
);
6644 se
->vruntime
= curr
->vruntime
;
6645 place_entity(cfs_rq
, se
, 1);
6647 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
6649 * Upon rescheduling, sched_class::put_prev_task() will place
6650 * 'current' within the tree based on its new key value.
6652 swap(curr
->vruntime
, se
->vruntime
);
6653 resched_task(rq
->curr
);
6656 se
->vruntime
-= cfs_rq
->min_vruntime
;
6658 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6662 * Priority of the task has changed. Check to see if we preempt
6666 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
6672 * Reschedule if we are currently running on this runqueue and
6673 * our priority decreased, or if we are not currently running on
6674 * this runqueue and our priority is higher than the current's
6676 if (rq
->curr
== p
) {
6677 if (p
->prio
> oldprio
)
6678 resched_task(rq
->curr
);
6680 check_preempt_curr(rq
, p
, 0);
6683 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
6685 struct sched_entity
*se
= &p
->se
;
6686 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6689 * Ensure the task's vruntime is normalized, so that when its
6690 * switched back to the fair class the enqueue_entity(.flags=0) will
6691 * do the right thing.
6693 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6694 * have normalized the vruntime, if it was !on_rq, then only when
6695 * the task is sleeping will it still have non-normalized vruntime.
6697 if (!se
->on_rq
&& p
->state
!= TASK_RUNNING
) {
6699 * Fix up our vruntime so that the current sleep doesn't
6700 * cause 'unlimited' sleep bonus.
6702 place_entity(cfs_rq
, se
, 0);
6703 se
->vruntime
-= cfs_rq
->min_vruntime
;
6708 * Remove our load from contribution when we leave sched_fair
6709 * and ensure we don't carry in an old decay_count if we
6712 if (se
->avg
.decay_count
) {
6713 __synchronize_entity_decay(se
);
6714 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
6720 * We switched to the sched_fair class.
6722 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
6728 * We were most likely switched from sched_rt, so
6729 * kick off the schedule if running, otherwise just see
6730 * if we can still preempt the current task.
6733 resched_task(rq
->curr
);
6735 check_preempt_curr(rq
, p
, 0);
6738 /* Account for a task changing its policy or group.
6740 * This routine is mostly called to set cfs_rq->curr field when a task
6741 * migrates between groups/classes.
6743 static void set_curr_task_fair(struct rq
*rq
)
6745 struct sched_entity
*se
= &rq
->curr
->se
;
6747 for_each_sched_entity(se
) {
6748 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6750 set_next_entity(cfs_rq
, se
);
6751 /* ensure bandwidth has been allocated on our new cfs_rq */
6752 account_cfs_rq_runtime(cfs_rq
, 0);
6756 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
6758 cfs_rq
->tasks_timeline
= RB_ROOT
;
6759 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6760 #ifndef CONFIG_64BIT
6761 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
6764 atomic64_set(&cfs_rq
->decay_counter
, 1);
6765 atomic_long_set(&cfs_rq
->removed_load
, 0);
6769 #ifdef CONFIG_FAIR_GROUP_SCHED
6770 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
6772 struct cfs_rq
*cfs_rq
;
6774 * If the task was not on the rq at the time of this cgroup movement
6775 * it must have been asleep, sleeping tasks keep their ->vruntime
6776 * absolute on their old rq until wakeup (needed for the fair sleeper
6777 * bonus in place_entity()).
6779 * If it was on the rq, we've just 'preempted' it, which does convert
6780 * ->vruntime to a relative base.
6782 * Make sure both cases convert their relative position when migrating
6783 * to another cgroup's rq. This does somewhat interfere with the
6784 * fair sleeper stuff for the first placement, but who cares.
6787 * When !on_rq, vruntime of the task has usually NOT been normalized.
6788 * But there are some cases where it has already been normalized:
6790 * - Moving a forked child which is waiting for being woken up by
6791 * wake_up_new_task().
6792 * - Moving a task which has been woken up by try_to_wake_up() and
6793 * waiting for actually being woken up by sched_ttwu_pending().
6795 * To prevent boost or penalty in the new cfs_rq caused by delta
6796 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6798 if (!on_rq
&& (!p
->se
.sum_exec_runtime
|| p
->state
== TASK_WAKING
))
6802 p
->se
.vruntime
-= cfs_rq_of(&p
->se
)->min_vruntime
;
6803 set_task_rq(p
, task_cpu(p
));
6805 cfs_rq
= cfs_rq_of(&p
->se
);
6806 p
->se
.vruntime
+= cfs_rq
->min_vruntime
;
6809 * migrate_task_rq_fair() will have removed our previous
6810 * contribution, but we must synchronize for ongoing future
6813 p
->se
.avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
6814 cfs_rq
->blocked_load_avg
+= p
->se
.avg
.load_avg_contrib
;
6819 void free_fair_sched_group(struct task_group
*tg
)
6823 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
6825 for_each_possible_cpu(i
) {
6827 kfree(tg
->cfs_rq
[i
]);
6836 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
6838 struct cfs_rq
*cfs_rq
;
6839 struct sched_entity
*se
;
6842 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
6845 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
6849 tg
->shares
= NICE_0_LOAD
;
6851 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
6853 for_each_possible_cpu(i
) {
6854 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
6855 GFP_KERNEL
, cpu_to_node(i
));
6859 se
= kzalloc_node(sizeof(struct sched_entity
),
6860 GFP_KERNEL
, cpu_to_node(i
));
6864 init_cfs_rq(cfs_rq
);
6865 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
6876 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
6878 struct rq
*rq
= cpu_rq(cpu
);
6879 unsigned long flags
;
6882 * Only empty task groups can be destroyed; so we can speculatively
6883 * check on_list without danger of it being re-added.
6885 if (!tg
->cfs_rq
[cpu
]->on_list
)
6888 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6889 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
6890 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6893 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
6894 struct sched_entity
*se
, int cpu
,
6895 struct sched_entity
*parent
)
6897 struct rq
*rq
= cpu_rq(cpu
);
6901 init_cfs_rq_runtime(cfs_rq
);
6903 tg
->cfs_rq
[cpu
] = cfs_rq
;
6906 /* se could be NULL for root_task_group */
6911 se
->cfs_rq
= &rq
->cfs
;
6913 se
->cfs_rq
= parent
->my_q
;
6916 update_load_set(&se
->load
, 0);
6917 se
->parent
= parent
;
6920 static DEFINE_MUTEX(shares_mutex
);
6922 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
6925 unsigned long flags
;
6928 * We can't change the weight of the root cgroup.
6933 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
6935 mutex_lock(&shares_mutex
);
6936 if (tg
->shares
== shares
)
6939 tg
->shares
= shares
;
6940 for_each_possible_cpu(i
) {
6941 struct rq
*rq
= cpu_rq(i
);
6942 struct sched_entity
*se
;
6945 /* Propagate contribution to hierarchy */
6946 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6948 /* Possible calls to update_curr() need rq clock */
6949 update_rq_clock(rq
);
6950 for_each_sched_entity(se
)
6951 update_cfs_shares(group_cfs_rq(se
));
6952 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6956 mutex_unlock(&shares_mutex
);
6959 #else /* CONFIG_FAIR_GROUP_SCHED */
6961 void free_fair_sched_group(struct task_group
*tg
) { }
6963 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
6968 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
6970 #endif /* CONFIG_FAIR_GROUP_SCHED */
6973 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
6975 struct sched_entity
*se
= &task
->se
;
6976 unsigned int rr_interval
= 0;
6979 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6982 if (rq
->cfs
.load
.weight
)
6983 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
6989 * All the scheduling class methods:
6991 const struct sched_class fair_sched_class
= {
6992 .next
= &idle_sched_class
,
6993 .enqueue_task
= enqueue_task_fair
,
6994 .dequeue_task
= dequeue_task_fair
,
6995 .yield_task
= yield_task_fair
,
6996 .yield_to_task
= yield_to_task_fair
,
6998 .check_preempt_curr
= check_preempt_wakeup
,
7000 .pick_next_task
= pick_next_task_fair
,
7001 .put_prev_task
= put_prev_task_fair
,
7004 .select_task_rq
= select_task_rq_fair
,
7005 .migrate_task_rq
= migrate_task_rq_fair
,
7007 .rq_online
= rq_online_fair
,
7008 .rq_offline
= rq_offline_fair
,
7010 .task_waking
= task_waking_fair
,
7013 .set_curr_task
= set_curr_task_fair
,
7014 .task_tick
= task_tick_fair
,
7015 .task_fork
= task_fork_fair
,
7017 .prio_changed
= prio_changed_fair
,
7018 .switched_from
= switched_from_fair
,
7019 .switched_to
= switched_to_fair
,
7021 .get_rr_interval
= get_rr_interval_fair
,
7023 #ifdef CONFIG_FAIR_GROUP_SCHED
7024 .task_move_group
= task_move_group_fair
,
7028 #ifdef CONFIG_SCHED_DEBUG
7029 void print_cfs_stats(struct seq_file
*m
, int cpu
)
7031 struct cfs_rq
*cfs_rq
;
7034 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
7035 print_cfs_rq(m
, cpu
, cfs_rq
);
7040 __init
void init_sched_fair_class(void)
7043 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7045 #ifdef CONFIG_NO_HZ_COMMON
7046 nohz
.next_balance
= jiffies
;
7047 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7048 cpu_notifier(sched_ilb_notifier
, 0);